server-parallel : add "--reverse-prompt" + compiler warning fixes

.github/ISSUE_TEMPLATE/custom.md

@@ -1,7 +1,8 @@
 ---
-name: Bug template
-about: Used to report bugs in llama.cpp
-labels: ["bug-unconfirmed"]
+name: Issue and enhancement template
+about: Used to report issues and request enhancements for llama.cpp
+title: "[User] Insert summary of your issue or enhancement.."
+labels: ''
 assignees: ''
 
 ---
@@ -45,7 +46,7 @@ $ g++ --version
 
 # Failure Information (for bugs)
 
-Please help provide information about the failure / bug.
+Please help provide information about the failure if this is a bug. If it is not a bug, please remove the rest of this template.
 
 # Steps to Reproduce
 

.github/ISSUE_TEMPLATE/enhancement.md

@@ -1,28 +0,0 @@
----
-name: Enhancement template
-about: Used to request enhancements for llama.cpp
-labels: ["enhancement"]
-assignees: ''
-
----
-
-# Prerequisites
-
-Please answer the following questions for yourself before submitting an issue.
-
-- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
-- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
-- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed).
-- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share.
-
-# Feature Description
-
-Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
-
-# Motivation
-
-Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
-
-# Possible Implementation
-
-If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.

.github/workflows/build.yml

@@ -10,10 +10,10 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
 
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
@@ -258,7 +258,7 @@ jobs:
 
     strategy:
       matrix:
-        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
+        destination: ['platform=macOS,name=Any Mac', 'platform=iOS,name=Any iOS Device', 'platform=tvOS,name=Any tvOS Device']
 
     steps:
       - name: Clone
@@ -276,11 +276,6 @@ jobs:
         run: |
           xcodebuild -scheme llama -destination "${{ matrix.destination }}"
 
-      - name: Build Swift Example
-        id: make_build_swift_example
-        run: |
-            make swift
-
   windows-latest-cmake:
     runs-on: windows-latest
 

.github/workflows/gguf-publish.yml

@@ -36,9 +36,8 @@ jobs:
         poetry install
 
     - name: Build package
-      run: cd gguf-py && poetry build
+      run: poetry build
     - name: Publish package
       uses: pypa/gh-action-pypi-publish@release/v1
       with:
         password: ${{ secrets.PYPI_API_TOKEN }}
-        packages-dir: gguf-py/dist

.github/workflows/zig-build.yml

@@ -1,25 +0,0 @@
-name: Zig CI
-
-on:
-  pull_request:
-  push:
-    branches:
-      - master
-
-jobs:
-  build:
-    strategy:
-      fail-fast: false
-      matrix:
-        runs-on: [ubuntu-latest, macos-latest, windows-latest]
-    runs-on: ${{ matrix.runs-on }}
-    steps:
-      - uses: actions/checkout@v3
-        with:
-          submodules: recursive
-          fetch-depth: 0
-      - uses: goto-bus-stop/setup-zig@v2
-        with:
-          version: 0.11.0
-      - name: Build Summary
-        run: zig build --summary all -freference-trace

.gitignore

@@ -10,8 +10,6 @@
 *.gcno
 *.gcda
 *.dot
-*.bat
-*.metallib
 .DS_Store
 .build/
 .cache/
@@ -45,7 +43,6 @@ models-mnt
 /infill
 /libllama.so
 /llama-bench
-/llava
 /main
 /metal
 /perplexity
@@ -57,7 +54,6 @@ models-mnt
 /server
 /simple
 /batched
-/batched-bench
 /export-lora
 /finetune
 /speculative

CMakeLists.txt

@@ -82,7 +82,6 @@ set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 option(LLAMA_CUBLAS                          "llama: use CUDA"                                  OFF)
 #option(LLAMA_CUDA_CUBLAS                     "llama: use cuBLAS for prompt processing"          OFF)
 option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
-option(LLAMA_CUDA_FORCE_MMQ                  "llama: use mmq kernels instead of cuBLAS"         OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
 option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
@@ -94,6 +93,7 @@ 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})
@@ -277,8 +277,13 @@ if (LLAMA_BLAS)
     endif()
 endif()
 
-if (LLAMA_QKK_64)
-    add_compile_definitions(GGML_QKK_64)
+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()
 endif()
 
 if (LLAMA_CUBLAS)
@@ -300,9 +305,6 @@ if (LLAMA_CUBLAS)
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
-        if (LLAMA_CUDA_FORCE_MMQ)
-            add_compile_definitions(GGML_CUDA_FORCE_MMQ)
-        endif()
         add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
         if (DEFINED LLAMA_CUDA_DMMV_Y)
@@ -329,7 +331,6 @@ if (LLAMA_CUBLAS)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
         else()
             set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
-            #set(CMAKE_CUDA_ARCHITECTURES "") # use this to compile much faster, but only F16 models work
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@@ -403,9 +404,6 @@ if (LLAMA_HIPBLAS)
         if (LLAMA_CUDA_FORCE_DMMV)
             target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
         endif()
-        if (LLAMA_CUDA_FORCE_MMQ)
-            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_MMQ)
-        endif()
         target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
         target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
@@ -424,7 +422,8 @@ endif()
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
         set(warning_flags -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
-        set(c_flags -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int -Werror=implicit-function-declaration)
+        set(c_flags -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int
+            -Werror=implicit-function-declaration)
         set(cxx_flags -Wmissing-declarations -Wmissing-noreturn)
         set(host_cxx_flags "")
 
@@ -456,8 +455,7 @@ if (LLAMA_ALL_WARNINGS)
     set(c_flags   ${c_flags}   ${warning_flags})
     set(cxx_flags ${cxx_flags} ${warning_flags})
     add_compile_options("$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
-                        "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>"
-                        "$<$<COMPILE_LANGUAGE:CXX>:${host_cxx_flags}>")
+                        "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags} ${host_cxx_flags}>")
 
 endif()
 
@@ -665,10 +663,6 @@ add_library(ggml OBJECT
             ggml.h
             ggml-alloc.c
             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

@@ -1,14 +1,8 @@
 # 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  \
-	speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml simple batched save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o
 
 # Binaries only useful for tests
-TEST_TARGETS = \
-	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
-	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
-	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe
+TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -178,24 +172,6 @@ else
 	MK_CPPFLAGS += -DNDEBUG
 endif
 
-ifdef LLAMA_SANITIZE_THREAD
-	MK_CFLAGS   += -fsanitize=thread -g
-	MK_CXXFLAGS += -fsanitize=thread -g
-	MK_LDFLAGS  += -fsanitize=thread -g
-endif
-
-ifdef LLAMA_SANITIZE_ADDRESS
-	MK_CFLAGS   += -fsanitize=address -fno-omit-frame-pointer -g
-	MK_CXXFLAGS += -fsanitize=address -fno-omit-frame-pointer -g
-	MK_LDFLAGS  += -fsanitize=address -fno-omit-frame-pointer -g
-endif
-
-ifdef LLAMA_SANITIZE_UNDEFINED
-	MK_CFLAGS   += -fsanitize=undefined -g
-	MK_CXXFLAGS += -fsanitize=undefined -g
-	MK_LDFLAGS  += -fsanitize=undefined -g
-endif
-
 ifdef LLAMA_SERVER_VERBOSE
 	MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
 endif
@@ -342,9 +318,13 @@ 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.
@@ -361,7 +341,7 @@ 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
@@ -378,7 +358,7 @@ 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 = $(LLAMA_CUDA_NVCC)
@@ -393,9 +373,6 @@ endif # CUDA_DOCKER_ARCH
 ifdef LLAMA_CUDA_FORCE_DMMV
 	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
-ifdef LLAMA_CUDA_FORCE_MMQ
-	NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
-endif # LLAMA_CUDA_FORCE_MMQ
 ifdef LLAMA_CUDA_DMMV_X
 	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 else
@@ -493,6 +470,11 @@ 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)
@@ -530,24 +512,12 @@ ggml.o: ggml.c ggml.h ggml-cuda.h
 ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
+OBJS += ggml-alloc.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
+llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
-COMMON_DEPS   = common.o sampling.o grammar-parser.o
-
-common.o: common/common.cpp build-info.h $(COMMON_H_DEPS)
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-sampling.o: common/sampling.cpp $(COMMON_H_DEPS)
+common.o: common/common.cpp common/common.h build-info.h common/log.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 console.o: common/console.cpp common/console.h
@@ -569,22 +539,19 @@ clean:
 # 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                                  build-info.h ggml.o llama.o common.o 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                            build-info.h ggml.o llama.o common.o 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                            build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-batched: examples/batched/batched.cpp                         build-info.h 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: examples/batched/batched.cpp                         build-info.h 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)
@@ -593,49 +560,53 @@ quantize: examples/quantize/quantize.cpp                      build-info.h ggml.
 quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.h 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                build-info.h ggml.o llama.o common.o $(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                   build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp 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)
-	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
+server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2)
+
+$(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-input/embd-input-lib.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) --shared $(CXXFLAGS) $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
+
+
+embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %$(DSO_EXT),$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
 
 gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o common.o train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 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 build-info.h ggml.o llama.o common.o $(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)
-	$(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)
+baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.o 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 build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o 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 build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o 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 build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 ifdef LLAMA_METAL
@@ -643,11 +614,6 @@ metal: examples/metal/metal.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 endif
 
-ifeq ($(UNAME_S),Darwin)
-swift: examples/batched.swift
-	(cd examples/batched.swift; make build)
-endif
-
 build-info.h: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh $(CC) > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
@@ -668,7 +634,7 @@ benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o
 run-benchmark-matmult: benchmark-matmult
 	./$@
 
-.PHONY: run-benchmark-matmult swift
+.PHONY: run-benchmark-matmult
 
 vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
@@ -676,40 +642,40 @@ 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 build-info.h ggml.o common.o 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 build-info.h ggml.o llama.o common.o 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 build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o $(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 build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-c.o: tests/test-c.c llama.h

Package.swift

@@ -1,27 +1,24 @@
-// swift-tools-version:5.5
+// swift-tools-version:5.3
 
 import PackageDescription
 
 #if arch(arm) || arch(arm64)
 let platforms: [SupportedPlatform]? = [
-    .macOS(.v12),
+    .macOS(.v11),
     .iOS(.v14),
     .watchOS(.v4),
     .tvOS(.v14)
 ]
 let exclude: [String] = []
-let resources: [Resource] = [
-    .process("ggml-metal.metal")
-]
-let additionalSources: [String] = ["ggml-metal.m"]
+let additionalSources: [String] = ["ggml-metal.m", "ggml-metal.metal"]
 let additionalSettings: [CSetting] = [
     .unsafeFlags(["-fno-objc-arc"]),
+    .define("GGML_SWIFT"),
     .define("GGML_USE_METAL")
 ]
 #else
 let platforms: [SupportedPlatform]? = nil
 let exclude: [String] = ["ggml-metal.metal"]
-let resources: [Resource] = []
 let additionalSources: [String] = []
 let additionalSettings: [CSetting] = []
 #endif
@@ -41,13 +38,12 @@ let package = Package(
                 "ggml.c",
                 "llama.cpp",
                 "ggml-alloc.c",
-                "ggml-backend.c",
-                "ggml-quants.c",
+                "k_quants.c",
             ] + additionalSources,
-            resources: resources,
             publicHeadersPath: "spm-headers",
             cSettings: [
-                .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
+                .unsafeFlags(["-Wno-shorten-64-to-32"]),
+                .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

@@ -11,8 +11,12 @@ 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)
+- ‼️ Breaking change: `rope_freq_base` and `rope_freq_scale` must be set to zero to use the model default values: [#3401](https://github.com/ggerganov/llama.cpp/pull/3401)
+- Parallel decoding + continuous batching support added: [#3228](https://github.com/ggerganov/llama.cpp/pull/3228) \
+  **Devs should become familiar with the new API**
+- Local Falcon 180B inference on Mac Studio
+
+  https://github.com/ggerganov/llama.cpp/assets/1991296/98abd4e8-7077-464c-ae89-aebabca7757e
 
 ----
 
@@ -85,23 +89,18 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
 - [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy)
-- [X] [Pygmalion/Metharme](#using-pygmalion-7b--metharme-7b)
+- [X] [Pygmalion 7B / Metharme 7B](#using-pygmalion-7b--metharme-7b)
 - [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
-- [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
-- [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
+- [X] [Baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B) and its derivations (such as [baichuan-7b-sft](https://huggingface.co/hiyouga/baichuan-7b-sft))
+- [X] [Aquila-7B](https://huggingface.co/BAAI/Aquila-7B) / [AquilaChat-7B](https://huggingface.co/BAAI/AquilaChat-7B)
 - [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
 - [X] [Mistral AI v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
-- [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
-- [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
-- [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
-- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
-
 
 **Bindings:**
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
-- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
+- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp), [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
@@ -204,7 +203,7 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8
 
 ## Usage
 
-Here are the end-to-end binary build and model conversion steps for the LLaMA-7B model.
+Here are the steps for the LLaMA-7B model.
 
 ### Get the Code
 
@@ -277,7 +276,7 @@ In order to build llama.cpp you have three different options.
 On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
 To disable the Metal build at compile time use the `LLAMA_NO_METAL=1` flag or the `LLAMA_METAL=OFF` cmake option.
 
-When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
+When built with Metal support, you can explicitly disable GPU inference with the `--gpu-layers|-ngl 0` command-line
 argument.
 
 ### MPI Build
@@ -378,7 +377,7 @@ Building the program with BLAS support may lead to some performance improvements
 
 - #### cuBLAS
 
-  This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
+  This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
   - Using `make`:
     ```bash
     make LLAMA_CUBLAS=1
@@ -571,18 +570,6 @@ python3 convert.py models/7B/
 
 When running the larger models, make sure you have enough disk space to store all the intermediate files.
 
-### Running on Windows with prebuilt binaries
-
-You will find prebuilt Windows binaries on the release page.
-
-Simply download and extract the latest zip package of choice: (e.g. `llama-b1380-bin-win-avx2-x64.zip`)
-
-From the unzipped folder, open a terminal/cmd window here and place a pre-converted `.gguf` model file. Test out the main example like so:
-
-```
-.\main -m llama-2-7b.Q4_0.gguf -n 128
-```
-
 ### Memory/Disk Requirements
 
 As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
@@ -626,18 +613,6 @@ For more information, see [https://huggingface.co/docs/transformers/perplexity](
 The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
 The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 threads.
 
-#### How to run
-
-1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
-2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
-3. Output:
-```
-perplexity : calculating perplexity over 655 chunks
-24.43 seconds per pass - ETA 4.45 hours
-[1]4.5970,[2]5.1807,[3]6.0382,...
-```
-And after 4.45 hours, you will have the final perplexity.
-
 ### Interactive mode
 
 If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
@@ -800,6 +775,18 @@ If your issue is with model generation quality, then please at least scan the fo
     - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
     - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
 
+#### How to run
+
+1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
+3. Output:
+```
+perplexity : calculating perplexity over 655 chunks
+24.43 seconds per pass - ETA 4.45 hours
+[1]4.5970,[2]5.1807,[3]6.0382,...
+```
+And after 4.45 hours, you will have the final perplexity.
+
 ### Android
 
 #### Building the Project using Android NDK
@@ -962,6 +949,7 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /
 
 - [main](./examples/main/README.md)
 - [server](./examples/server/README.md)
+- [embd-input](./examples/embd-input/README.md)
 - [jeopardy](./examples/jeopardy/README.md)
 - [BLIS](./docs/BLIS.md)
 - [Performance troubleshooting](./docs/token_generation_performance_tips.md)

build.zig

@@ -36,17 +36,14 @@ const Maker = struct {
     }
 
     fn init(builder: *std.build.Builder) !Maker {
+        // const commit_hash = @embedFile(".git/refs/heads/master");
         const target = builder.standardTargetOptions(.{});
-        const zig_version = @import("builtin").zig_version_string;
-        const commit_hash = try std.ChildProcess.exec(
-            .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
-        );
         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_COMMIT = "12345", // omit newline
+                .BUILD_COMPILER = "Zig 0.11.0",
                 .BUILD_TARGET = try target.allocDescription(builder.allocator),
             },
         );
@@ -70,20 +67,12 @@ const Maker = struct {
 
     fn obj(m: *const Maker, name: []const u8, src: []const u8) *Compile {
         const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
-        if (o.target.getAbi() != .msvc)
-            o.defineCMacro("_GNU_SOURCE", null);
-        o.addConfigHeader(m.config_header);
         if (std.mem.endsWith(u8, src, ".c")) {
             o.addCSourceFiles(&.{src}, m.cflags.items);
             o.linkLibC();
         } else {
             o.addCSourceFiles(&.{src}, m.cxxflags.items);
-            if (o.target.getAbi() == .msvc) {
-                o.linkLibC(); // need winsdk + crt
-            } else {
-                // linkLibCpp already add (libc++ + libunwind + libc)
-                o.linkLibCpp();
-            }
+            o.linkLibCpp();
         }
         o.addConfigHeader(m.config_header);
         for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
@@ -97,14 +86,8 @@ const Maker = struct {
         for (deps) |d| e.addObject(d);
         for (m.objs.items) |o| e.addObject(o);
         for (m.include_dirs.items) |i| e.addIncludePath(.{ .path = i });
-
-        // https://github.com/ziglang/zig/issues/15448
-        if (e.target.getAbi() == .msvc) {
-            e.linkLibC(); // need winsdk + crt
-        } else {
-            // linkLibCpp already add (libc++ + libunwind + libc)
-            e.linkLibCpp();
-        }
+        e.linkLibC();
+        e.linkLibCpp();
         e.addConfigHeader(m.config_header);
         m.builder.installArtifact(e);
         e.want_lto = m.enable_lto;
@@ -116,26 +99,26 @@ 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 common = make.obj("common", "common/common.cpp");
-    const console = make.obj("console", "common/console.cpp");
-    const sampling = make.obj("sampling", "common/sampling.cpp");
+    const console = make.obj("common", "common/console.cpp");
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
-    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, ggml_quants, llama, common, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, 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, train });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, llama, common, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, llama, common });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, llama, common });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, llama, common });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, llama, common });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, sampling, grammar_parser, clip });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, llama, common, grammar_parser });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/run.sh

@@ -208,8 +208,6 @@ function gg_run_open_llama_3b_v2 {
     (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-
     function check_ppl {
         qnt="$1"
         ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
@@ -298,7 +296,6 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
-    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
     gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
     gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
     gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
@@ -385,8 +382,6 @@ function gg_run_open_llama_7b_v2 {
     (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-
     function check_ppl {
         qnt="$1"
         ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
@@ -475,7 +470,6 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
-    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
     gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
     gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
     #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
@@ -502,12 +496,10 @@ test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
-    if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
-        if [ -z ${GG_BUILD_CUDA} ]; then
-            test $ret -eq 0 && gg_run open_llama_3b_v2
-        else
-            test $ret -eq 0 && gg_run open_llama_7b_v2
-        fi
+    if [ -z ${GG_BUILD_CUDA} ]; then
+        test $ret -eq 0 && gg_run open_llama_3b_v2
+    else
+        test $ret -eq 0 && gg_run open_llama_7b_v2
     fi
 fi
 

common/CMakeLists.txt

@@ -5,8 +5,6 @@ set(TARGET common)
 add_library(${TARGET} OBJECT
     common.h
     common.cpp
-    sampling.h
-    sampling.cpp
     console.h
     console.cpp
     grammar-parser.h

common/common.cpp

@@ -107,7 +107,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     std::string arg;
     gpt_params default_params;
     const std::string arg_prefix = "--";
-    llama_sampling_params & sparams = params.sparams;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -168,10 +167,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            // store the external file name in params
-            params.prompt_file = argv[i];
             std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
-            if (!params.prompt.empty() && params.prompt.back() == '\n') {
+            if (params.prompt.back() == '\n') {
                 params.prompt.pop_back();
             }
         } else if (arg == "-n" || arg == "--n-predict") {
@@ -185,7 +182,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            sparams.top_k = std::stoi(argv[i]);
+            params.top_k = std::stoi(argv[i]);
         } else if (arg == "-c" || arg == "--ctx-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -217,75 +214,73 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            sparams.top_p = std::stof(argv[i]);
+            params.top_p = std::stof(argv[i]);
         } else if (arg == "--temp") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.temp = std::stof(argv[i]);
-            sparams.temp = std::max(sparams.temp, 0.0f);
+            params.temp = std::stof(argv[i]);
         } else if (arg == "--tfs") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.tfs_z = std::stof(argv[i]);
+            params.tfs_z = std::stof(argv[i]);
         } else if (arg == "--typical") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.typical_p = std::stof(argv[i]);
+            params.typical_p = std::stof(argv[i]);
         } else if (arg == "--repeat-last-n") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.penalty_last_n = std::stoi(argv[i]);
-            sparams.n_prev = std::max(sparams.n_prev, sparams.penalty_last_n);
+            params.repeat_last_n = std::stoi(argv[i]);
         } else if (arg == "--repeat-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.penalty_repeat = std::stof(argv[i]);
+            params.repeat_penalty = std::stof(argv[i]);
         } else if (arg == "--frequency-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.penalty_freq = std::stof(argv[i]);
+            params.frequency_penalty = std::stof(argv[i]);
         } else if (arg == "--presence-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.penalty_present = std::stof(argv[i]);
+            params.presence_penalty = std::stof(argv[i]);
         } else if (arg == "--mirostat") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.mirostat = std::stoi(argv[i]);
+            params.mirostat = std::stoi(argv[i]);
         } else if (arg == "--mirostat-lr") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.mirostat_eta = std::stof(argv[i]);
+            params.mirostat_eta = std::stof(argv[i]);
         } else if (arg == "--mirostat-ent") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.mirostat_tau = std::stof(argv[i]);
+            params.mirostat_tau = std::stof(argv[i]);
         } else if (arg == "--cfg-negative-prompt") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.cfg_negative_prompt = argv[i];
+            params.cfg_negative_prompt = argv[i];
         } else if (arg == "--cfg-negative-prompt-file") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -297,16 +292,16 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(sparams.cfg_negative_prompt));
-            if (!sparams.cfg_negative_prompt.empty() && sparams.cfg_negative_prompt.back() == '\n') {
-                sparams.cfg_negative_prompt.pop_back();
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.cfg_negative_prompt));
+            if (params.cfg_negative_prompt.back() == '\n') {
+                params.cfg_negative_prompt.pop_back();
             }
         } else if (arg == "--cfg-scale") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.cfg_scale = std::stof(argv[i]);
+            params.cfg_scale = std::stof(argv[i]);
         } else if (arg == "-b" || arg == "--batch-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -386,18 +381,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.lora_base = argv[i];
-        } else if (arg == "--mmproj") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.mmproj = argv[i];
-        } else if (arg == "--image") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.image = argv[i];
         } else if (arg == "-i" || arg == "--interactive") {
             params.interactive = true;
         } else if (arg == "--embedding") {
@@ -527,7 +510,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         } else if (arg == "--ignore-eos") {
             params.ignore_eos = true;
         } else if (arg == "--no-penalize-nl") {
-            sparams.penalize_nl = false;
+            params.penalize_nl = false;
         } else if (arg == "-l" || arg == "--logit-bias") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -539,7 +522,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             std::string value_str;
             try {
                 if (ss >> key && ss >> sign && std::getline(ss, value_str) && (sign == '+' || sign == '-')) {
-                    sparams.logit_bias[key] = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f);
+                    params.logit_bias[key] = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f);
                 } else {
                     throw std::exception();
                 }
@@ -574,7 +557,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            sparams.grammar = argv[i];
+            params.grammar = argv[i];
         } else if (arg == "--grammar-file") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -589,7 +572,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             std::copy(
                 std::istreambuf_iterator<char>(file),
                 std::istreambuf_iterator<char>(),
-                std::back_inserter(sparams.grammar)
+                std::back_inserter(params.grammar)
             );
 #ifndef LOG_DISABLE_LOGS
         // Parse args for logging parameters
@@ -633,7 +616,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         process_escapes(params.prompt);
         process_escapes(params.input_prefix);
         process_escapes(params.input_suffix);
-        process_escapes(sparams.cfg_negative_prompt);
         for (auto & antiprompt : params.antiprompt) {
             process_escapes(antiprompt);
         }
@@ -643,8 +625,6 @@ 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("usage: %s [options]\n", argv[0]);
     printf("\n");
     printf("options:\n");
@@ -677,19 +657,19 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
     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("  --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);
-    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)sparams.penalty_repeat);
-    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.penalty_present);
-    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.penalty_freq);
+    printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
+    printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
+    printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
+    printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
+    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
+    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
+    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
+    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
     printf("  --mirostat N          use Mirostat sampling.\n");
     printf("                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
-    printf("                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", sparams.mirostat);
-    printf("  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)sparams.mirostat_eta);
-    printf("  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)sparams.mirostat_tau);
+    printf("                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
+    printf("  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
+    printf("  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
     printf("  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
     printf("                        modifies the likelihood of token appearing in the completion,\n");
     printf("                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
@@ -700,7 +680,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        negative prompt to use for guidance. (default: empty)\n");
     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("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
     printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale\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");
@@ -708,7 +688,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     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");
     printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
-    printf("  --temp N              temperature (default: %.1f)\n", (double)sparams.temp);
+    printf("  --temp N              temperature (default: %.1f)\n", (double)params.temp);
     printf("  --logits-all          return logits for all tokens in the batch (default: disabled)\n");
     printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
     printf("  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
@@ -718,8 +698,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     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("  -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");
     if (llama_mlock_supported()) {
         printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
@@ -744,7 +722,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
 #endif // GGML_USE_CUBLAS
 #endif
     printf("  --verbose-prompt      print prompt before generation\n");
-    printf("  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
+    fprintf(stderr, "  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
     printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
     printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
@@ -823,27 +801,6 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     return cparams;
 }
 
-void llama_batch_clear(struct llama_batch & batch) {
-    batch.n_tokens = 0;
-}
-
-void llama_batch_add(
-                 struct llama_batch & batch,
-                        llama_token   id,
-                          llama_pos   pos,
-    const std::vector<llama_seq_id> & seq_ids,
-                               bool   logits) {
-    batch.token   [batch.n_tokens] = id;
-    batch.pos     [batch.n_tokens] = pos,
-    batch.n_seq_id[batch.n_tokens] = seq_ids.size();
-    for (size_t i = 0; i < seq_ids.size(); ++i) {
-        batch.seq_id[batch.n_tokens][i] = seq_ids[i];
-    }
-    batch.logits  [batch.n_tokens] = logits;
-
-    batch.n_tokens++;
-}
-
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
     auto mparams = llama_model_params_from_gpt_params(params);
 
@@ -881,15 +838,15 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     }
 
     if (params.ignore_eos) {
-        params.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
+        params.logit_bias[llama_token_eos(lctx)] = -INFINITY;
     }
 
     {
         LOG("warming up the model with an empty run\n");
 
-        std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };
+        std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
         llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
-        llama_kv_cache_clear(lctx);
+        llama_kv_cache_tokens_rm(lctx, -1, -1);
         llama_reset_timings(lctx);
     }
 
@@ -903,23 +860,21 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 std::vector<llama_token> llama_tokenize(
   const struct llama_context * ctx,
            const std::string & text,
-                        bool   add_bos,
-                        bool   special) {
-    return llama_tokenize(llama_get_model(ctx), text, add_bos, special);
+                        bool   add_bos) {
+    return llama_tokenize(llama_get_model(ctx), text, add_bos);
 }
 
 std::vector<llama_token> llama_tokenize(
     const struct llama_model * model,
            const std::string & text,
-                        bool   add_bos,
-                        bool   special) {
+                        bool   add_bos) {
     // upper limit for the number of tokens
     int n_tokens = text.length() + add_bos;
     std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, special);
+    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, special);
+        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -942,7 +897,7 @@ std::string llama_token_to_piece(const struct llama_context * ctx, llama_token t
 }
 
 std::string llama_detokenize_spm(llama_context * ctx, const std::vector<llama_token> & tokens) {
-    const llama_token bos_id = llama_token_bos(llama_get_model(ctx));
+    const llama_token bos_id = llama_token_bos(ctx);
 
     std::string piece;
     std::string result;
@@ -975,6 +930,126 @@ std::string llama_detokenize_bpe(llama_context * ctx, const std::vector<llama_to
     return result;
 }
 
+//
+// Sampling utils
+//
+
+llama_token llama_sample_token(
+                  struct llama_context * ctx,
+                  struct llama_context * ctx_guidance,
+                  struct llama_grammar * grammar,
+               const struct gpt_params & params,
+        const std::vector<llama_token> & last_tokens,
+         std::vector<llama_token_data> & candidates,
+                                   int   idx) {
+    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+
+    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   tfs_z           = params.tfs_z;
+    const float   typical_p       = params.typical_p;
+    const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
+    const float   repeat_penalty  = params.repeat_penalty;
+    const float   alpha_presence  = params.presence_penalty;
+    const float   alpha_frequency = params.frequency_penalty;
+    const int     mirostat        = params.mirostat;
+    const float   mirostat_tau    = params.mirostat_tau;
+    const float   mirostat_eta    = params.mirostat_eta;
+    const bool    penalize_nl     = params.penalize_nl;
+
+    llama_token id = 0;
+
+    float * logits = llama_get_logits_ith(ctx, idx);
+
+    // Apply params.logit_bias map
+    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
+        logits[it->first] += it->second;
+    }
+
+    candidates.clear();
+    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 cur_p = { candidates.data(), candidates.size(), false };
+
+    if (ctx_guidance) {
+        llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale);
+    }
+
+    // apply penalties
+    if (!last_tokens.empty()) {
+        const float nl_logit = logits[llama_token_nl(ctx)];
+        const int last_n_repeat = std::min(std::min((int)last_tokens.size(), repeat_last_n), n_ctx);
+
+        llama_sample_repetition_penalty(ctx, &cur_p,
+                last_tokens.data() + last_tokens.size() - last_n_repeat,
+                last_n_repeat, repeat_penalty);
+        llama_sample_frequency_and_presence_penalties(ctx, &cur_p,
+                last_tokens.data() + last_tokens.size() - last_n_repeat,
+                last_n_repeat, alpha_frequency, alpha_presence);
+
+        if (!penalize_nl) {
+            for (size_t idx = 0; idx < cur_p.size; idx++) {
+                if (cur_p.data[idx].id == llama_token_nl(ctx)) {
+                    cur_p.data[idx].logit = nl_logit;
+                    break;
+                }
+            }
+        }
+    }
+
+    if (grammar != NULL) {
+        llama_sample_grammar(ctx, &cur_p, grammar);
+    }
+
+    if (temp <= 0) {
+        // Greedy sampling
+        id = llama_sample_token_greedy(ctx, &cur_p);
+    } else {
+        if (mirostat == 1) {
+            static float mirostat_mu = 2.0f * mirostat_tau;
+            const int mirostat_m = 100;
+            llama_sample_temp(ctx, &cur_p, temp);
+            id = llama_sample_token_mirostat(ctx, &cur_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, &cur_p, temp);
+            id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu);
+        } else {
+            // Temperature sampling
+            llama_sample_top_k      (ctx, &cur_p, top_k, 1);
+            llama_sample_tail_free  (ctx, &cur_p, tfs_z, 1);
+            llama_sample_typical    (ctx, &cur_p, typical_p, 1);
+            llama_sample_top_p      (ctx, &cur_p, top_p, 1);
+            llama_sample_temp(ctx, &cur_p, temp);
+
+            {
+                const int n_top = 10;
+                LOG("top %d candidates:\n", n_top);
+
+                for (int i = 0; i < n_top; i++) {
+                    const llama_token id = cur_p.data[i].id;
+                    LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p);
+                }
+            }
+
+            id = llama_sample_token(ctx, &cur_p);
+
+            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str());
+        }
+    }
+    // printf("`%d`", candidates_p.size);
+
+    if (grammar != NULL) {
+        llama_grammar_accept_token(ctx, grammar, id);
+    }
+
+    return id;
+}
+
 //
 // YAML utils
 //
@@ -1126,28 +1201,26 @@ std::string get_sortable_timestamp() {
 
 void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const llama_context * lctx,
                                const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
-    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, "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");
-    fprintf(stream, "cpu_has_avx512: %s\n",      ggml_cpu_has_avx512()      ? "true" : "false");
+    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");
+    fprintf(stream, "cpu_has_avx512: %s\n", ggml_cpu_has_avx512() ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
-    fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
-    fprintf(stream, "cpu_has_cublas: %s\n",      ggml_cpu_has_cublas()      ? "true" : "false");
-    fprintf(stream, "cpu_has_clblast: %s\n",     ggml_cpu_has_clblast()     ? "true" : "false");
-    fprintf(stream, "cpu_has_fma: %s\n",         ggml_cpu_has_fma()         ? "true" : "false");
-    fprintf(stream, "cpu_has_gpublas: %s\n",     ggml_cpu_has_gpublas()     ? "true" : "false");
-    fprintf(stream, "cpu_has_neon: %s\n",        ggml_cpu_has_neon()        ? "true" : "false");
-    fprintf(stream, "cpu_has_f16c: %s\n",        ggml_cpu_has_f16c()        ? "true" : "false");
-    fprintf(stream, "cpu_has_fp16_va: %s\n",     ggml_cpu_has_fp16_va()     ? "true" : "false");
-    fprintf(stream, "cpu_has_wasm_simd: %s\n",   ggml_cpu_has_wasm_simd()   ? "true" : "false");
-    fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
-    fprintf(stream, "cpu_has_sse3: %s\n",        ggml_cpu_has_sse3()        ? "true" : "false");
-    fprintf(stream, "cpu_has_vsx: %s\n",         ggml_cpu_has_vsx()         ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
+    fprintf(stream, "cpu_has_cublas: %s\n", ggml_cpu_has_cublas() ? "true" : "false");
+    fprintf(stream, "cpu_has_clblast: %s\n", ggml_cpu_has_clblast() ? "true" : "false");
+    fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false");
+    fprintf(stream, "cpu_has_gpublas: %s\n", ggml_cpu_has_gpublas() ? "true" : "false");
+    fprintf(stream, "cpu_has_neon: %s\n", ggml_cpu_has_neon() ? "true" : "false");
+    fprintf(stream, "cpu_has_f16c: %s\n", ggml_cpu_has_f16c() ? "true" : "false");
+    fprintf(stream, "cpu_has_fp16_va: %s\n", ggml_cpu_has_fp16_va() ? "true" : "false");
+    fprintf(stream, "cpu_has_wasm_simd: %s\n", ggml_cpu_has_wasm_simd() ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
+    fprintf(stream, "cpu_has_sse3: %s\n", ggml_cpu_has_sse3() ? "true" : "false");
+    fprintf(stream, "cpu_has_vsx: %s\n", ggml_cpu_has_vsx() ? "true" : "false");
 
 #ifdef NDEBUG
     fprintf(stream, "debug: false\n");
@@ -1174,21 +1247,21 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
 
     fprintf(stream, "alias: %s # default: unknown\n", params.model_alias.c_str());
     fprintf(stream, "batch_size: %d # default: 512\n", params.n_batch);
-    dump_string_yaml_multiline(stream, "cfg_negative_prompt", sparams.cfg_negative_prompt.c_str());
-    fprintf(stream, "cfg_scale: %f # default: 1.0\n", sparams.cfg_scale);
+    dump_string_yaml_multiline(stream, "cfg_negative_prompt", params.cfg_negative_prompt.c_str());
+    fprintf(stream, "cfg_scale: %f # default: 1.0\n", params.cfg_scale);
     fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
     fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
     fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
     fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
     fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
-    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
-    dump_string_yaml_multiline(stream, "grammar", sparams.grammar.c_str());
+    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", params.frequency_penalty);
+    dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str());
     fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
     fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
     fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
 
-    const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(llama_get_model(lctx)));
-    const bool ignore_eos = logit_bias_eos != sparams.logit_bias.end() && logit_bias_eos->second == -INFINITY;
+    const auto logit_bias_eos = params.logit_bias.find(llama_token_eos(lctx));
+    const bool ignore_eos = logit_bias_eos != params.logit_bias.end() && logit_bias_eos->second == -INFINITY;
     fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false");
 
     dump_string_yaml_multiline(stream, "in_prefix", params.input_prefix.c_str());
@@ -1201,7 +1274,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "logdir: %s # default: unset (no logging)\n", params.logdir.c_str());
 
     fprintf(stream, "logit_bias:\n");
-    for (std::pair<llama_token, float> lb : sparams.logit_bias) {
+    for (std::pair<llama_token, float> lb : params.logit_bias) {
         if (ignore_eos && lb.first == logit_bias_eos->first) {
             continue;
         }
@@ -1225,30 +1298,30 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
     fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
-    fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
-    fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
-    fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
+    fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", params.mirostat);
+    fprintf(stream, "mirostat_ent: %f # default: 5.0\n", params.mirostat_tau);
+    fprintf(stream, "mirostat_lr: %f # default: 0.1\n", params.mirostat_eta);
     fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false");
     fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str());
     fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
     fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
     fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
     fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
-    fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", sparams.n_probs);
+    fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", params.n_probs);
     fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
     fprintf(stream, "no_mul_mat_q: %s # default: false\n", !params.mul_mat_q ? "true" : "false");
-    fprintf(stream, "no_penalize_nl: %s # default: false\n", !sparams.penalize_nl ? "true" : "false");
+    fprintf(stream, "no_penalize_nl: %s # default: false\n", !params.penalize_nl ? "true" : "false");
     fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false");
     fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
     fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
-    fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
+    fprintf(stream, "presence_penalty: %f # default: 0.0\n", params.presence_penalty);
     dump_string_yaml_multiline(stream, "prompt", params.prompt.c_str());
     fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
     fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
     fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
     dump_vector_int_yaml(stream, "prompt_tokens", prompt_tokens);
     fprintf(stream, "random_prompt: %s # default: false\n", params.random_prompt ? "true" : "false");
-    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.penalty_repeat);
+    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", params.repeat_penalty);
 
     fprintf(stream, "reverse_prompt:\n");
     for (std::string ap : params.antiprompt) {
@@ -1266,15 +1339,15 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed);
     fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
     fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
-    fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
+    fprintf(stream, "temp: %f # default: 0.8\n", params.temp);
 
     const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + LLAMA_MAX_DEVICES);
     dump_vector_float_yaml(stream, "tensor_split", tensor_split_vector);
 
-    fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
+    fprintf(stream, "tfs: %f # default: 1.0\n", params.tfs_z);
     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, "typical_p: %f # default: 1.0\n", sparams.typical_p);
+    fprintf(stream, "top_k: %d # default: 40\n", params.top_k);
+    fprintf(stream, "top_p: %f # default: 0.95\n", params.top_p);
+    fprintf(stream, "typical_p: %f # default: 1.0\n", params.typical_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
 }

common/common.h

@@ -4,8 +4,6 @@
 
 #include "llama.h"
 
-#include "sampling.h"
-
 #define LOG_NO_FILE_LINE_FUNCTION
 #include "log.h"
 
@@ -51,25 +49,43 @@ struct gpt_params {
     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_probs                         = 0;    // if greater than 0, output the probabilities of top n_probs tokens.
     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
 
-    // // sampling parameters
-    struct llama_sampling_params sparams;
+    // sampling parameters
+    int32_t top_k             = 40;    // <= 0 to use vocab size
+    float   top_p             = 0.95f; // 1.0 = disabled
+    float   tfs_z             = 1.00f; // 1.0 = disabled
+    float   typical_p         = 1.00f; // 1.0 = disabled
+    float   temp              = 0.80f; // 1.0 = disabled
+    float   repeat_penalty    = 1.10f; // 1.0 = disabled
+    int32_t repeat_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   frequency_penalty = 0.00f; // 0.0 = disabled
+    float   presence_penalty  = 0.00f; // 0.0 = disabled
+    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   mirostat_tau      = 5.00f; // target entropy
+    float   mirostat_eta      = 0.10f; // learning rate
+
+    std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
+
+    // Classifier-Free Guidance
+    // https://arxiv.org/abs/2306.17806
+    std::string cfg_negative_prompt;       // string to help guidance
+    float       cfg_scale         = 1.f;   // How strong is guidance
 
     std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
     std::string model_draft       = "";                              // draft model for speculative decoding
     std::string model_alias       = "unknown"; // model alias
     std::string prompt            = "";
-    std::string prompt_file       = "";  // store the external prompt file name
     std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
     std::string input_prefix      = "";  // string to prefix user inputs with
     std::string input_suffix      = "";  // string to suffix user inputs with
+    std::string grammar           = "";  // optional BNF-like grammar to constrain sampling
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
     std::string logdir            = "";  // directory in which to save YAML log files
 
-    // TODO: avoid tuple, use struct
     std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
     std::string lora_base  = "";                              // base model path for the lora adapter
 
@@ -98,16 +114,13 @@ struct gpt_params {
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens
     bool instruct          = false; // instruction mode (used for Alpaca models)
+    bool penalize_nl       = true;  // consider newlines as a repeatable token
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool infill            = false; // use infill mode
-
-    // multimodal models (see examples/llava)
-    std::string mmproj = ""; // path to multimodal projector
-    std::string image = ""; // path to an image file
 };
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
@@ -124,23 +137,10 @@ void process_escapes(std::string& input);
 // Model utils
 //
 
-// TODO: avoid tuplue, use struct
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
-
-struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
+struct llama_model_params   llama_model_params_from_gpt_params(const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
-// Batch utils
-
-void llama_batch_clear(struct llama_batch & batch);
-
-void llama_batch_add(
-                 struct llama_batch & batch,
-                        llama_token   id,
-                          llama_pos   pos,
-    const std::vector<llama_seq_id> & seq_ids,
-                               bool   logits);
-
 //
 // Vocab utils
 //
@@ -150,14 +150,12 @@ void llama_batch_add(
 std::vector<llama_token> llama_tokenize(
   const struct llama_context * ctx,
            const std::string & text,
-                        bool   add_bos,
-                        bool   special = false);
+                        bool   add_bos);
 
 std::vector<llama_token> llama_tokenize(
     const struct llama_model * model,
            const std::string & text,
-                        bool   add_bos,
-                        bool   special = false);
+                        bool   add_bos);
 
 // tokenizes a token into a piece
 // should work similar to Python's `tokenizer.id_to_piece`
@@ -181,6 +179,36 @@ std::string llama_detokenize_bpe(
                          llama_context * ctx,
         const std::vector<llama_token> & tokens);
 
+//
+// Sampling utils
+//
+
+// this is a common sampling function used across the examples for convenience
+// it can serve as a starting point for implementing your own sampling function
+//
+// required:
+//  - ctx:    context to use for sampling
+//  - params: sampling parameters
+//
+// optional:
+//  - ctx_guidance:  context to use for classifier-free guidance, ignore if NULL
+//  - grammar:       grammar to use for sampling, ignore if NULL
+//  - last_tokens:   needed for repetition penalty, ignore if empty
+//  - idx:           sample from llama_get_logits_ith(ctx, idx)
+//
+// returns:
+//  - token:      sampled token
+//  - candidates: vector of candidate tokens
+//
+llama_token llama_sample_token(
+                  struct llama_context * ctx,
+                  struct llama_context * ctx_guidance,
+                  struct llama_grammar * grammar,
+               const struct gpt_params & params,
+        const std::vector<llama_token> & last_tokens,
+         std::vector<llama_token_data> & candidates,
+                                   int   idx = 0);
+
 //
 // YAML utils
 //

common/grammar-parser.cpp

@@ -399,7 +399,7 @@ namespace grammar_parser {
     void print_grammar(FILE * file, const parse_state & state) {
         try {
             std::map<uint32_t, std::string> symbol_id_names;
-            for (const auto & kv : state.symbol_ids) {
+            for (auto kv : state.symbol_ids) {
                 symbol_id_names[kv.second] = kv.first;
             }
             for (size_t i = 0, end = state.rules.size(); i < end; i++) {

common/log.h

@@ -97,23 +97,22 @@
     #define LOG_TEE_TARGET stderr
 #endif
 
-// NOTE: currently disabled as it produces too many log files
 // 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"
@@ -127,8 +126,8 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base
     std::stringstream buf;
 
     buf << log_file_basename;
-    //buf << ".";
-    //buf << log_get_pid();
+    buf << ".";
+    buf << log_get_pid();
     buf << ".";
     buf << log_file_extension;
 
@@ -580,75 +579,38 @@ inline std::string log_var_to_string_impl(const std::vector<int> & var)
     return buf.str();
 }
 
-template <typename C, typename T>
-inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
-{
-    std::stringstream buf;
-    buf << "[ ";
-
-    bool first = true;
-    for (const auto &token : tokens)
-    {
-        if (!first) {
-            buf << ", ";
-        } else {
-            first = false;
-        }
-
-        auto detokenized = llama_token_to_piece(ctx, token);
-
-        detokenized.erase(
-            std::remove_if(
-                detokenized.begin(),
-                detokenized.end(),
-                [](const unsigned char c) { return !std::isprint(c); }),
-            detokenized.end());
-
-        buf
-            << "'" << detokenized << "'"
-            << ":" << std::to_string(token);
-    }
-    buf << " ]";
-
-    return buf.str();
-}
-
-template <typename C, typename B>
-inline std::string LOG_BATCH_TOSTR_PRETTY(const C & ctx, const B & batch)
-{
-    std::stringstream buf;
-    buf << "[ ";
-
-    bool first = true;
-    for (int i = 0; i < batch.n_tokens; ++i)
-    {
-        if (!first) {
-            buf << ", ";
-        } else {
-            first = false;
-        }
-
-        auto detokenized = llama_token_to_piece(ctx, batch.token[i]);
-
-        detokenized.erase(
-            std::remove_if(
-                detokenized.begin(),
-                detokenized.end(),
-                [](const unsigned char c) { return !std::isprint(c); }),
-            detokenized.end());
-
-        buf
-            << "\n" << std::to_string(i)
-            << ":token '" << detokenized << "'"
-            << ":pos " << std::to_string(batch.pos[i])
-            << ":n_seq_id  " << std::to_string(batch.n_seq_id[i])
-            << ":seq_id " << std::to_string(batch.seq_id[i][0])
-            << ":logits " << std::to_string(batch.logits[i]);
-    }
-    buf << " ]";
-
-    return buf.str();
-}
+#define LOG_TOKENS_TOSTR_PRETTY(ctx, tokens)                                 \
+    [&tokens, &ctx]()                                                        \
+    {                                                                        \
+        std::stringstream buf;                                               \
+        buf << "[ ";                                                         \
+                                                                             \
+        bool first = true;                                                   \
+        for (const auto &token : tokens)                                     \
+        {                                                                    \
+            if (!first)                                                      \
+                buf << ", ";                                                 \
+            else                                                             \
+                first = false;                                               \
+                                                                             \
+            auto detokenized = llama_token_to_piece(ctx, token);             \
+                                                                             \
+            detokenized.erase(                                               \
+                std::remove_if(                                              \
+                    detokenized.begin(),                                     \
+                    detokenized.end(),                                       \
+                    [](const unsigned char c) { return !std::isprint(c); }), \
+                detokenized.end());                                          \
+                                                                             \
+            buf                                                              \
+                << "'" << detokenized << "'"                                 \
+                << ":" << std::to_string(token);                             \
+        }                                                                    \
+        buf << " ]";                                                         \
+                                                                             \
+        return buf.str();                                                    \
+    }()                                                                      \
+        .c_str()
 
 #ifdef LOG_DISABLE_LOGS
 

common/sampling.cpp

@@ -1,226 +0,0 @@
-#include "sampling.h"
-
-struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
-    struct llama_sampling_context * result = new llama_sampling_context();
-
-    result->params  = params;
-    result->grammar = nullptr;
-
-    // if there is a grammar, parse it
-    if (!params.grammar.empty()) {
-        result->parsed_grammar = grammar_parser::parse(params.grammar.c_str());
-
-        // will be empty (default) if there are parse errors
-        if (result->parsed_grammar.rules.empty()) {
-            fprintf(stderr, "%s: failed to parse grammar\n", __func__);
-            return nullptr;
-        }
-
-        std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
-
-        result->grammar = llama_grammar_init(
-                grammar_rules.data(),
-                grammar_rules.size(), result->parsed_grammar.symbol_ids.at("root"));
-    }
-
-    result->prev.resize(params.n_prev);
-
-    return result;
-}
-
-void llama_sampling_free(struct llama_sampling_context * ctx) {
-    if (ctx->grammar != NULL) {
-        llama_grammar_free(ctx->grammar);
-    }
-
-    delete ctx;
-}
-
-void llama_sampling_reset(llama_sampling_context * ctx) {
-    if (ctx->grammar != NULL) {
-        llama_grammar_free(ctx->grammar);
-    }
-
-    if (!ctx->parsed_grammar.rules.empty()) {
-        std::vector<const llama_grammar_element *> grammar_rules(ctx->parsed_grammar.c_rules());
-
-        ctx->grammar = llama_grammar_init(
-                grammar_rules.data(),
-                grammar_rules.size(), ctx->parsed_grammar.symbol_ids.at("root"));
-    }
-
-    std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
-    ctx->cur.clear();
-}
-
-void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst) {
-    if (dst->grammar) {
-        llama_grammar_free(dst->grammar);
-        dst->grammar = nullptr;
-    }
-
-    if (src->grammar) {
-        dst->grammar = llama_grammar_copy(src->grammar);
-    }
-
-    dst->prev = src->prev;
-}
-
-llama_token llama_sampling_last(llama_sampling_context * ctx) {
-    return ctx->prev.back();
-}
-
-std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n) {
-    const int size = ctx_sampling->prev.size();
-
-    n = std::min(n, size);
-
-    std::string result;
-
-    for (int i = size - n; i < size; i++) {
-        result += llama_token_to_piece(ctx_main, ctx_sampling->prev[i]);
-    }
-
-    return result;
-}
-
-std::string llama_sampling_print(const llama_sampling_params & params) {
-    char result[1024];
-
-    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"
-            "\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.mirostat, params.mirostat_eta, params.mirostat_tau);
-
-    return std::string(result);
-}
-
-llama_token llama_sampling_sample(
-                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_main,
-                  struct llama_context * ctx_cfg,
-                  const int idx) {
-    const llama_sampling_params & params = ctx_sampling->params;
-
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
-
-    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   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;
-    const float   penalty_repeat  = params.penalty_repeat;
-    const float   penalty_freq    = params.penalty_freq;
-    const float   penalty_present = params.penalty_present;
-    const int     mirostat        = params.mirostat;
-    const float   mirostat_tau    = params.mirostat_tau;
-    const float   mirostat_eta    = params.mirostat_eta;
-    const bool    penalize_nl     = params.penalize_nl;
-
-    auto & prev = ctx_sampling->prev;
-    auto & cur  = ctx_sampling->cur;
-
-    llama_token id = 0;
-
-    float * logits = llama_get_logits_ith(ctx_main, idx);
-
-    // apply params.logit_bias map
-    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
-        logits[it->first] += it->second;
-    }
-
-    cur.clear();
-
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-    }
-
-    llama_token_data_array cur_p = { cur.data(), cur.size(), false };
-
-    if (ctx_cfg) {
-        llama_sample_classifier_free_guidance(ctx_main, &cur_p, ctx_cfg, params.cfg_scale);
-    }
-
-    // apply penalties
-    if (!prev.empty()) {
-        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
-
-        llama_sample_repetition_penalties(ctx_main, &cur_p,
-                prev.data() + prev.size() - penalty_last_n,
-                penalty_last_n, penalty_repeat, penalty_freq, penalty_present);
-
-        if (!penalize_nl) {
-            for (size_t idx = 0; idx < cur_p.size; idx++) {
-                if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
-                    cur_p.data[idx].logit = nl_logit;
-                    break;
-                }
-            }
-        }
-    }
-
-    if (ctx_sampling->grammar != NULL) {
-        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
-    }
-
-    if (temp < 0.0) {
-        // greedy sampling, with probs
-        llama_sample_softmax(ctx_main, &cur_p);
-        id = cur_p.data[0].id;
-    } else if (temp == 0.0) {
-        // greedy sampling, no probs
-        id = llama_sample_token_greedy(ctx_main, &cur_p);
-    } else {
-        if (mirostat == 1) {
-            const int mirostat_m = 100;
-            llama_sample_temp(ctx_main, &cur_p, temp);
-            id = llama_sample_token_mirostat(ctx_main, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_sampling->mirostat_mu);
-        } else if (mirostat == 2) {
-            llama_sample_temp(ctx_main, &cur_p, temp);
-            id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
-        } else {
-            // temperature sampling
-            size_t min_keep = std::max(1, params.n_probs);
-
-            llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep);
-            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_temp     (ctx_main, &cur_p, temp);
-
-            id = llama_sample_token(ctx_main, &cur_p);
-
-            //{
-            //    const int n_top = 10;
-            //    LOG("top %d candidates:\n", n_top);
-
-            //    for (int i = 0; i < n_top; i++) {
-            //        const llama_token id = cur_p.data[i].id;
-            //        (void)id; // To avoid a warning that id is unused when logging is disabled.
-            //        LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx_main, id).c_str(), cur_p.data[i].p);
-            //    }
-            //}
-
-            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
-        }
-    }
-
-    return id;
-}
-
-void llama_sampling_accept(
-        struct llama_sampling_context * ctx_sampling,
-        struct llama_context * ctx_main,
-        llama_token id,
-        bool apply_grammar) {
-    ctx_sampling->prev.erase(ctx_sampling->prev.begin());
-    ctx_sampling->prev.push_back(id);
-
-    if (ctx_sampling->grammar != NULL && apply_grammar) {
-        llama_grammar_accept_token(ctx_main, ctx_sampling->grammar, id);
-    }
-}

common/sampling.h

@@ -1,109 +0,0 @@
-#pragma once
-
-#include "llama.h"
-
-#include "grammar-parser.h"
-
-#include <string>
-#include <vector>
-#include <unordered_map>
-
-// sampling parameters
-typedef struct llama_sampling_params {
-    int32_t n_prev            = 64;    // number of previous tokens to remember
-    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   tfs_z             = 1.00f; // 1.0 = disabled
-    float   typical_p         = 1.00f; // 1.0 = disabled
-    float   temp              = 0.80f; // 1.0 = disabled
-    int32_t penalty_last_n    = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   penalty_repeat    = 1.10f; // 1.0 = disabled
-    float   penalty_freq      = 0.00f; // 0.0 = disabled
-    float   penalty_present   = 0.00f; // 0.0 = disabled
-    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   mirostat_tau      = 5.00f; // target entropy
-    float   mirostat_eta      = 0.10f; // learning rate
-    bool    penalize_nl       = true;  // consider newlines as a repeatable token
-
-    std::string grammar;  // optional BNF-like grammar to constrain sampling
-
-    // Classifier-Free Guidance
-    // https://arxiv.org/abs/2306.17806
-    std::string cfg_negative_prompt; // string to help guidance
-    float       cfg_scale     = 1.f; // how strong is guidance
-
-    std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
-} llama_sampling_params;
-
-// general sampler context
-// TODO: move to llama.h
-struct llama_sampling_context {
-    // parameters that will be used for sampling
-    llama_sampling_params params;
-
-    // mirostat sampler state
-    float mirostat_mu;
-
-    llama_grammar * grammar;
-
-    // internal
-    grammar_parser::parse_state parsed_grammar;
-
-    // TODO: replace with ring-buffer
-    std::vector<llama_token>      prev;
-    std::vector<llama_token_data> cur;
-};
-
-#include "common.h"
-
-// Create a new sampling context instance.
-struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params);
-
-void llama_sampling_free(struct llama_sampling_context * ctx);
-
-// Reset the sampler context
-// - clear prev tokens
-// - reset grammar
-void llama_sampling_reset(llama_sampling_context * ctx);
-
-// Copy the sampler context
-void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst);
-
-// Get the last sampled token
-llama_token llama_sampling_last(llama_sampling_context * ctx);
-
-// Get a string representation of the last sampled tokens
-std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n);
-
-// Print sampling parameters into a string
-std::string llama_sampling_print(const llama_sampling_params & params);
-
-// this is a common sampling function used across the examples for convenience
-// it can serve as a starting point for implementing your own sampling function
-// Note: When using multiple sequences, it is the caller's responsibility to call
-//       llama_sampling_reset when a sequence ends
-//
-// required:
-//  - ctx_main:     context to use for sampling
-//  - ctx_sampling: sampling-specific context
-//
-// optional:
-//  - ctx_cfg:      context to use for classifier-free guidance
-//  - idx:          sample from llama_get_logits_ith(ctx, idx)
-//
-// returns:
-//  - token:      sampled token
-//  - candidates: vector of candidate tokens
-//
-llama_token llama_sampling_sample(
-        struct llama_sampling_context * ctx_sampling,
-        struct llama_context * ctx_main,
-        struct llama_context * ctx_cfg,
-        int idx = 0);
-
-void llama_sampling_accept(
-        struct llama_sampling_context * ctx_sampling,
-        struct llama_context * ctx_main,
-        llama_token id,
-        bool apply_grammar);

common/train.cpp

@@ -236,8 +236,8 @@ int64_t get_example_targets_batch(
     int64_t used_samples = 0;
 
     ggml_set_f32(target_probs, 0.0f);
-    llama_token bos = llama_token_bos(llama_get_model(lctx));
-    llama_token eos = llama_token_eos(llama_get_model(lctx));
+    llama_token bos = llama_token_bos(lctx);
+    llama_token eos = llama_token_eos(lctx);
     // printf("%s: example_id=%d n_batch=%d n_train_samples=%zu\n", __func__, example_id, n_batch, n_train_samples);
     for (int k=0; k<n_batch; ++k) {
         // printf("%s: batch %d\n", __func__, k);
@@ -863,7 +863,7 @@ size_t tokenize_file(
             (int) buf.size(),
             out_tokens.data(),
             (int) out_tokens.size(),
-            false, false);
+            false);
         if (n_tokens < 0) {
             out_tokens.resize(-n_tokens);
             n_tokens = llama_tokenize(
@@ -872,7 +872,7 @@ size_t tokenize_file(
                 (int) buf.size(),
                 out_tokens.data(),
                 (int) out_tokens.size(),
-                false, false);
+                false);
         }
         if (n_tokens >= 0) {
             out_tokens.resize(n_tokens);
@@ -924,7 +924,7 @@ size_t tokenize_file(
         for (llama_token token=0; token < n_vocab; ++token) {
             max_token_text_size = std::max(
                 max_token_text_size,
-                strlen(llama_token_get_text(llama_get_model(lctx), token)));
+                strlen(llama_token_get_text(lctx, token)));
         }
 
         // upper bound of context byte length.
@@ -966,7 +966,7 @@ size_t tokenize_file(
                     (int) buf_sample.size(),
                     tok_sample.data(),
                     (int) tok_sample.size(),
-                    false, false);
+                    false);
                 if (n_tokens < 0) {
                     tok_sample.resize(-n_tokens);
                     n_tokens = llama_tokenize(llama_get_model(lctx),
@@ -974,7 +974,7 @@ size_t tokenize_file(
                         (int) buf_sample.size(),
                         tok_sample.data(),
                         (int) tok_sample.size(),
-                        false, false);
+                        false);
                     GGML_ASSERT(n_tokens >= 0);
                 }
                 GGML_ASSERT(n_tokens <= (int) tok_sample.size());
@@ -1425,7 +1425,7 @@ void train_opt_callback(void * vdata, int accum_step, float * sched, bool * canc
 
         int impr_plot = -(int)(1 + (opt->loss_before - opt->loss_after) * 10.0f + 0.5f);
         if (impr_plot > 0) impr_plot = 0;
-        if (std::isnan(opt->loss_before) || std::isnan(opt->loss_after)) impr_plot = 0;
+        if (std::isnan(opt->loss_before) || std::isnan(opt->loss_before)) impr_plot = 0;
         printf("%s: iter=%6d sample=%zu/%zu sched=%f loss=%f",
             __func__, opt->iter, std::min(1+train->shuffle_next_sample, train->shuffle_sample_count), train->shuffle_sample_count,
             *sched, opt->loss_after);

convert-baichuan-hf-to-gguf.py

@@ -76,7 +76,6 @@ def parse_args() -> argparse.Namespace:
         "ftype", type=int, choices=[0, 1], default=1, nargs='?',
         help="output format - use 0 for float32, 1 for float16",
     )
-    parser.add_argument("--bigendian",   action="store_true",    help="model is executed on big endian machine")
     return parser.parse_args()
 
 args = parse_args()
@@ -87,11 +86,6 @@ if not dir_model.is_dir():
     print(f'Error: {args.model} is not a directory', file = sys.stderr)
     sys.exit(1)
 
-endianess = gguf.GGUFEndian.LITTLE
-if args.bigendian:
-    endianess = gguf.GGUFEndian.BIG
-endianess_str = "Big Endian" if args.bigendian else "Little Endian"
-print(f"gguf: Conversion Endianess {endianess}")
 # possible tensor data types
 #   ftype == 0 -> float32
 #   ftype == 1 -> float16
@@ -110,7 +104,7 @@ print("gguf: loading model "+dir_model.name)
 with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 print("hello print: ",hparams["architectures"][0])
-if hparams["architectures"][0] != "BaichuanForCausalLM" and hparams["architectures"][0] != "BaiChuanForCausalLM":
+if hparams["architectures"][0] != "BaichuanForCausalLM":
     print("Model architecture not supported: " + hparams["architectures"][0])
 
     sys.exit()
@@ -119,7 +113,7 @@ if hparams["architectures"][0] != "BaichuanForCausalLM" and hparams["architectur
 num_parts = count_model_parts(dir_model)
 print(f"num_parts:{num_parts}\n")
 ARCH=gguf.MODEL_ARCH.BAICHUAN
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 
 print("gguf: get model metadata")
 
@@ -230,7 +224,7 @@ gguf_writer.add_token_list(tokens)
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, n_vocab = len(tokens))
+special_vocab = gguf.SpecialVocab(dir_model)
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

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

@@ -78,7 +78,7 @@ 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"):
+if hparams["architectures"][0] != "FalconForCausalLM":
     print("Model architecture not supported: " + hparams["architectures"][0])
 
     sys.exit(1)
@@ -97,17 +97,7 @@ 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
+block_count = hparams["num_hidden_layers"]
 
 gguf_writer.add_name("Falcon")
 gguf_writer.add_context_length(2048) # not in config.json
@@ -115,8 +105,11 @@ 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_head_count(hparams["num_attention_heads"])
+if "num_kv_heads" in hparams:
+    gguf_writer.add_head_count_kv(hparams["num_kv_heads"])
+else:
+    gguf_writer.add_head_count_kv(1)
 gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
 gguf_writer.add_file_type(ftype)
 
@@ -152,13 +145,17 @@ 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 = gguf.SpecialVocab(dir_model, load_merges = True)
 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["num_attention_heads"]
+n_head_kv = hparams["num_kv_heads"] if "num_kv_heads" in hparams else 1
+
 head_dim = hparams["hidden_size"] // n_head
 
 # tensor info

convert-gptneox-hf-to-gguf.py

@@ -123,27 +123,18 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 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)
+    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{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 = gguf.SpecialVocab(dir_model, load_merges = True)
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-llama-ggml-to-gguf.py

@@ -388,9 +388,7 @@ def handle_metadata(cfg, hp):
         cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir,
         cfg.vocabtype )
     # FIXME: Respect cfg.vocab_dir?
-    svocab = gguf.SpecialVocab(cfg.model_metadata_dir,
-        load_merges = cfg.vocabtype == 'bpe',
-        n_vocab = vocab.vocab_size)
+    svocab = gguf.SpecialVocab(cfg.model_metadata_dir)
     convert.check_vocab_size(params, vocab)
     return (params, vocab, svocab)
 

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-persimmon-to-gguf.py

@@ -1,130 +0,0 @@
-import torch
-import os
-from pprint import pprint
-import sys
-import argparse
-from pathlib import Path
-from sentencepiece import SentencePieceProcessor
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-def _flatten_dict(dct, tensors, prefix=None):
-    assert isinstance(dct, dict)
-    for key in dct.keys():
-        new_prefix = prefix + '.' + key if prefix is not None else key
-        if isinstance(dct[key], torch.Tensor):
-            tensors[new_prefix] = dct[key]
-        elif isinstance(dct[key], dict):
-            _flatten_dict(dct[key], tensors, new_prefix)
-        else:
-            raise ValueError(type(dct[key]))
-    return None
-
-def _get_sentencepiece_tokenizer_info(dir_model: Path):
-    tokenizer_path = dir_model / 'adept_vocab.model'
-    print('gguf: getting sentencepiece tokenizer from', tokenizer_path)
-    tokenizer = SentencePieceProcessor(str(tokenizer_path))
-    print('gguf: adding tokens')
-    tokens: list[bytes] = []
-    scores: list[float] = []
-    toktypes: list[int] = []
-
-    for i in range(tokenizer.vocab_size()):
-        text: bytes
-        score: float
-
-        piece = tokenizer.id_to_piece(i)
-        text = piece.encode("utf-8")
-        score = tokenizer.get_score(i)
-
-        toktype = 1
-        if tokenizer.is_unknown(i):
-            toktype = 2
-        if tokenizer.is_control(i):
-            toktype = 3
-        if tokenizer.is_unused(i):
-            toktype = 5
-        if tokenizer.is_byte(i):
-            toktype = 6
-
-        tokens.append(text)
-        scores.append(score)
-        toktypes.append(toktype)
-        pass
-    return tokens, scores, toktypes
-
-def main():
-    parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file")
-    parser.add_argument("--outfile",             type=Path, help="path to write to; default: based on input")
-    parser.add_argument("--ckpt-path",           type=Path, help="path to persimmon checkpoint .pt file")
-    parser.add_argument("--model-dir",           type=Path, help="directory containing model e.g. 8b_chat_model_release")
-    parser.add_argument("--adept-inference-dir", type=str, help="path to adept-inference code directory")
-    args = parser.parse_args()
-    sys.path.append(str(args.adept_inference_dir))
-    persimmon_model = torch.load(args.ckpt_path)
-    hparams = persimmon_model['args']
-    pprint(hparams)
-    tensors = {}
-    _flatten_dict(persimmon_model['model'], tensors, None)
-
-    arch = gguf.MODEL_ARCH.PERSIMMON
-    gguf_writer = gguf.GGUFWriter(args.outfile, gguf.MODEL_ARCH_NAMES[arch])
-
-    block_count = hparams.num_layers
-    head_count = hparams.num_attention_heads
-    head_count_kv = head_count
-    ctx_length = hparams.seq_length
-    hidden_size = hparams.hidden_size
-
-    gguf_writer.add_name('persimmon-8b-chat')
-    gguf_writer.add_context_length(ctx_length)
-    gguf_writer.add_embedding_length(hidden_size)
-    gguf_writer.add_block_count(block_count)
-    gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size)
-    gguf_writer.add_rope_dimension_count(hidden_size // head_count)
-    gguf_writer.add_head_count(head_count)
-    gguf_writer.add_head_count_kv(head_count_kv)
-    gguf_writer.add_rope_freq_base(hparams.rotary_emb_base)
-    gguf_writer.add_layer_norm_eps(hparams.layernorm_epsilon)
-
-    tokens, scores, toktypes = _get_sentencepiece_tokenizer_info(args.model_dir)
-    gguf_writer.add_tokenizer_model('llama')
-    gguf_writer.add_token_list(tokens)
-    gguf_writer.add_token_scores(scores)
-    gguf_writer.add_token_types(toktypes)
-    gguf_writer.add_bos_token_id(71013)
-    gguf_writer.add_eos_token_id(71013)
-
-    tensor_map = gguf.get_tensor_name_map(arch, block_count)
-    print(tensor_map)
-    for name in tensors.keys():
-        data = tensors[name]
-        if name.endswith(".self_attention.rotary_emb.inv_freq"):
-            continue
-        old_dtype = data.dtype
-        # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
-        data = data.to(torch.float32).squeeze().numpy()
-        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)
-        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()
-    print("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-    gguf_writer.close()
-
-    print(f"gguf: model successfully exported to '{args.outfile}'")
-    print("")
-
-
-
-if __name__ == '__main__':
-    main()

convert-refact-hf-to-gguf.py

@@ -17,6 +17,33 @@ if "NO_LOCAL_GGUF" not in os.environ:
     sys.path.insert(1, str(Path(__file__).parent / "gguf-py" / "gguf"))
 import gguf
 
+
+def bytes_to_unicode():
+    # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    return dict(zip(bs, (chr(n) for n in cs)))
+
+
 def count_model_parts(dir_model: Path) -> int:
     num_parts = 0
     for filename in os.listdir(dir_model):
@@ -126,40 +153,59 @@ tokens: list[bytearray] = []
 scores: list[float] = []
 toktypes: list[int] = []
 
+tokenizer_json_file = dir_model / "tokenizer.json"
+if not tokenizer_json_file.is_file():
+    print(f"Error: Missing {tokenizer_json_file}", file=sys.stderr)
+    sys.exit(1)
+
 # gpt2 tokenizer
 gguf_writer.add_tokenizer_model("gpt2")
 
-print("gguf: get gpt2 tokenizer vocab")
+with open(tokenizer_json_file, "r", encoding="utf-8") as f:
+    tokenizer_json = json.load(f)
 
-# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-tokenizer = AutoTokenizer.from_pretrained(dir_model)
+print("gguf: get gpt2 tokenizer vocab")
 
 # 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
+vocab_size = (
+    hparams["vocab_size"]
+    if "vocab_size" in hparams
+    else len(tokenizer_json["model"]["vocab"])
+)
+
+tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
 
-added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v: k for k, v in byte_encoder.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)
+    if i in reverse_vocab:
+        text = reverse_vocab[i]
+        try:
+            text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
+        except KeyError:
+            text = bytearray()
+            for c in reverse_vocab[i]:
+                if ord(c) < 256:  # single byte character
+                    text.append(byte_decoder[ord(c)])
+                else:  # multibyte special token character
+                    text.extend(c.encode("utf-8"))
     else:
-        tokens.append(reverse_vocab[i])
-        toktypes.append(gguf.TokenType.NORMAL)
+        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
+        pad_token = f"[PAD{i}]".encode("utf8")
+        text = bytearray(pad_token)
+
+    tokens.append(text)
+    scores.append(0.0)  # dymmy
+    toktypes.append(gguf.TokenType.NORMAL)  # dummy
 
 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 = gguf.SpecialVocab(dir_model, load_merges=True)
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-starcoder-hf-to-gguf.py

@@ -111,26 +111,18 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 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)
+    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{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 = gguf.SpecialVocab(dir_model, load_merges = True)
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert.py

@@ -366,19 +366,16 @@ class SentencePieceVocab:
             added_tokens = {}
 
         vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
+        expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
+        actual_ids   = sorted(added_tokens.values())
+        if expected_ids != actual_ids:
+            raise Exception(f"Expected added token IDs to be sequential and start at {len(added_tokens)}; got {actual_ids}")
 
-        new_tokens       = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
-        expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
-        actual_new_ids   = sorted(new_tokens.keys())
-
-        if expected_new_ids != actual_new_ids:
-            raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
-
-        # Token pieces that were added to the base vocabulary.
-        self.added_tokens_list  = [new_tokens[id] for id in actual_new_ids]
-        self.vocab_size_base    = vocab_size
-        self.vocab_size         = self.vocab_size_base + len(self.added_tokens_list)
-        self.fname_tokenizer    = fname_tokenizer
+        items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
+        self.added_tokens_list = [text for (text, idx) in items]
+        self.vocab_size_base: int = vocab_size
+        self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list)
+        self.fname_tokenizer = fname_tokenizer
         self.fname_added_tokens = fname_added_tokens
 
     def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
@@ -806,8 +803,8 @@ def check_vocab_size(params: Params, vocab: Vocab) -> None:
 
 
 class OutputFile:
-    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian=gguf.GGUFEndian.LITTLE) -> None:
-        self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
+    def __init__(self, fname_out: Path) -> None:
+        self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 
     def add_meta_arch(self, params: Params) -> None:
         name = "LLaMA"
@@ -878,10 +875,10 @@ class OutputFile:
         self.gguf.close()
 
     @staticmethod
-    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, endianess:gguf.GGUFEndian=gguf.GGUFEndian.LITTLE) -> None:
+    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab) -> None:
         check_vocab_size(params, vocab)
 
-        of = OutputFile(fname_out, endianess=endianess)
+        of = OutputFile(fname_out)
 
         # meta data
         of.add_meta_arch(params)
@@ -906,10 +903,10 @@ class OutputFile:
         return dt.quantize(arr)
 
     @staticmethod
-    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY, endianess=gguf.GGUFEndian.LITTLE) -> None:
+    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
         check_vocab_size(params, vocab)
 
-        of = OutputFile(fname_out, endianess=endianess)
+        of = OutputFile(fname_out)
 
         # meta data
         of.add_meta_arch(params)
@@ -1126,9 +1123,8 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--vocabtype",   choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm")
     parser.add_argument("--ctx",         type=int,               help="model training context (default: based on input)")
     parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY)
-    parser.add_argument("--bigendian",   action="store_true",    help="model is executed on big endian machine")
-
     args = parser.parse_args(args_in)
+
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
         do_dump_model(model_plus)
@@ -1142,9 +1138,6 @@ def main(args_in: list[str] | None = None) -> None:
     if args.dump:
         do_dump_model(model_plus)
         return
-    endianess = gguf.GGUFEndian.LITTLE
-    if args.bigendian:
-        endianess = gguf.GGUFEndian.BIG
 
     params = Params.load(model_plus)
     if params.n_ctx == -1:
@@ -1166,13 +1159,10 @@ def main(args_in: list[str] | None = None) -> None:
 
     vocab: Vocab
     if args.vocab_only:
-        if not args.outfile:
-            raise ValueError("need --outfile if using --vocab-only")
+        assert args.outfile, "need --outfile if using --vocab-only"
         # FIXME: Try to respect vocab_dir somehow?
         vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
-        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
-            load_merges = args.vocabtype == 'bpe',
-            n_vocab = vocab.vocab_size)
+        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
         outfile = args.outfile
         OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
         print(f"Wrote {outfile}")
@@ -1184,9 +1174,7 @@ def main(args_in: list[str] | None = None) -> None:
         vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
         vocab = load_vocab(vocab_dir, args.vocabtype)
     # FIXME: Try to respect vocab_dir somehow?
-    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
-        load_merges = args.vocabtype == 'bpe',
-        n_vocab = vocab.vocab_size)
+    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
 
     model   = model_plus.model
     model   = convert_model_names(model, params)
@@ -1197,7 +1185,7 @@ def main(args_in: list[str] | None = None) -> None:
     params.ftype = ftype
     print(f"Writing {outfile}, format {ftype}")
 
-    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency, endianess=endianess)
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency)
     print(f"Wrote {outfile}")
 
 

docs/BLIS.md

@@ -49,7 +49,7 @@ According to the BLIS documentation, we could set the following
 environment variables to modify the behavior of openmp:
 
 ```bash
-export GOMP_CPU_AFFINITY="0-19"
+export GOMP_GPU_AFFINITY="0-19"
 export BLIS_NUM_THREADS=14
 ```
 

examples/CMakeLists.txt

@@ -12,31 +12,30 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
 else()
-    add_subdirectory(baby-llama)
-    add_subdirectory(batched)
-    add_subdirectory(batched-bench)
-    add_subdirectory(beam-search)
-    add_subdirectory(benchmark)
-    add_subdirectory(convert-llama2c-to-ggml)
-    add_subdirectory(embedding)
-    add_subdirectory(finetune)
-    add_subdirectory(infill)
-    add_subdirectory(llama-bench)
-    add_subdirectory(llava)
     add_subdirectory(main)
-    add_subdirectory(parallel)
-    add_subdirectory(perplexity)
     add_subdirectory(quantize)
     add_subdirectory(quantize-stats)
+    add_subdirectory(perplexity)
+    add_subdirectory(embedding)
     add_subdirectory(save-load-state)
-    add_subdirectory(simple)
-    add_subdirectory(speculative)
+    add_subdirectory(benchmark)
+    add_subdirectory(baby-llama)
     add_subdirectory(train-text-from-scratch)
+    add_subdirectory(finetune)
+    add_subdirectory(convert-llama2c-to-ggml)
+    add_subdirectory(simple)
+    add_subdirectory(batched)
+    add_subdirectory(speculative)
+    add_subdirectory(parallel)
+    add_subdirectory(embd-input)
+    add_subdirectory(llama-bench)
+    add_subdirectory(beam-search)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()
     if (LLAMA_BUILD_SERVER)
         add_subdirectory(server)
+        add_subdirectory(server-parallel)
     endif()
     add_subdirectory(export-lora)
 endif()

examples/batched-bench/CMakeLists.txt

@@ -1,5 +0,0 @@
-set(TARGET batched-bench)
-add_executable(${TARGET} batched-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)

examples/batched-bench/README.md

@@ -1,51 +0,0 @@
-# llama.cpp/example/batched-bench
-
-Benchmark the batched decoding performance of `llama.cpp`
-
-## Usage
-
-There are 2 modes of operation:
-
-- `prompt not shared` - each batch has a separate prompt of size `PP` (i.e. `N_KV = B*(PP + TG)`)
-- `prompt is shared` - there is a common prompt of size `PP` used by all batches (i.e. `N_KV = PP + B*TG`)
-
-```bash
-./batched-bench MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] [MMQ] <PP> <TG> <PL>
-
-# LLaMA 7B, F16, N_KV_MAX = 16384 (8GB), prompt not shared
-./batched-bench ./models/llama-7b/ggml-model-f16.gguf 16384 0 99
-
-# LLaMA 7B, Q8_0, N_KV_MAX = 16384 (8GB), prompt is shared
-./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 16384 1 99
-
-# custom set of batches
-./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 2048 0 999 0 128,256,512 128,256 1,2,4,8,16,32
-```
-
-## Sample results
-
-- `PP` - prompt tokens per batch
-- `TG` - generated tokens per batch
-- `B` - number of batches
-- `N_KV` - required KV cache size
-- `T_PP` - prompt processing time (i.e. time to first token)
-- `S_PP` - prompt processing speed (`(B*PP)/T_PP` or `PP/T_PP`)
-- `T_TG` - time to generate all batches
-- `S_TG` - text generation speed (`(B*TG)/T_TG`)
-- `T` - total time
-- `S` - total speed (i.e. all tokens / total time)
-
-|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
-|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   128 |    128 |    1 |    256 |    0.108 |  1186.64 |    3.079 |    41.57 |    3.187 |    80.32 |
-|   128 |    128 |    2 |    512 |    0.198 |  1295.19 |    5.029 |    50.90 |    5.227 |    97.95 |
-|   128 |    128 |    4 |   1024 |    0.373 |  1373.96 |    6.878 |    74.44 |    7.251 |   141.23 |
-|   128 |    128 |    8 |   2048 |    0.751 |  1363.27 |    7.344 |   139.43 |    8.095 |   252.99 |
-|   128 |    128 |   16 |   4096 |    1.570 |  1304.68 |    8.455 |   242.23 |   10.024 |   408.60 |
-|   128 |    128 |   32 |   8192 |    3.408 |  1201.73 |    8.801 |   465.40 |   12.209 |   670.96 |
-|   128 |    256 |    1 |    384 |    0.107 |  1196.70 |    6.329 |    40.45 |    6.436 |    59.67 |
-|   128 |    256 |    2 |    768 |    0.194 |  1317.45 |   10.239 |    50.00 |   10.433 |    73.61 |
-|   128 |    256 |    4 |   1536 |    0.366 |  1399.03 |   13.960 |    73.35 |   14.326 |   107.22 |
-|   128 |    256 |    8 |   3072 |    0.751 |  1363.92 |   15.110 |   135.54 |   15.861 |   193.69 |
-|   128 |    256 |   16 |   6144 |    1.569 |  1304.93 |   18.073 |   226.64 |   19.642 |   312.80 |
-|   128 |    256 |   32 |  12288 |    3.409 |  1201.35 |   19.223 |   426.15 |   22.633 |   542.93 |

examples/batched-bench/batched-bench.cpp

@@ -1,247 +0,0 @@
-#include "common.h"
-#include "llama.h"
-
-#include <algorithm>
-#include <cmath>
-#include <cstdio>
-#include <string>
-#include <vector>
-
-// mutates the input string
-static std::vector<int> parse_list(char * p) {
-    std::vector<int> ret;
-
-    char * q = p;
-
-    while (*p) {
-        if (*p == ',') {
-            *p = '\0';
-            ret.push_back(std::atoi(q));
-            q = p + 1;
-        }
-
-        ++p;
-    }
-
-    ret.push_back(std::atoi(q));
-
-    return ret;
-}
-
-int main(int argc, char ** argv) {
-    gpt_params params;
-
-    if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] [MMQ] <PP> <TG> <PL>\n" , argv[0]);
-        printf("  <PP>, <TG> and PL are comma-separated lists of numbers without spaces\n\n");
-        printf("  example: %s ggml-model-f16.gguf 2048 0 999 0 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
-        return 1 ;
-    }
-
-    int n_kv_max     = 2048;
-    int is_pp_shared = 0;
-    int n_gpu_layers = 0;
-    int mmq          = 0;
-
-    std::vector<int> n_pp = { 128, 256, 512, 1024, 2048, 3584, 7680, };
-    std::vector<int> n_tg = { 128, 256, };
-    std::vector<int> n_pl = { 1, 2, 4, 8, 16, 32, };
-    //std::vector<int> n_pl = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 32, };
-
-    if (argc >= 2) {
-        params.model = argv[1];
-    }
-
-    if (argc >= 3) {
-        n_kv_max = std::atoi(argv[2]);
-    }
-
-    if (argc >= 4) {
-        is_pp_shared = std::atoi(argv[3]);
-    }
-
-    if (argc >= 5) {
-        n_gpu_layers = std::atoi(argv[4]);
-    }
-
-    if (argc >= 6) {
-        mmq = std::atoi(argv[5]);
-    }
-
-    if (argc >= 7) {
-        n_pp = parse_list(argv[6]);
-    }
-
-    if (argc >= 8) {
-        n_tg = parse_list(argv[7]);
-    }
-
-    if (argc >= 9) {
-        n_pl = parse_list(argv[8]);
-    }
-
-    // init LLM
-
-    llama_backend_init(params.numa);
-
-    // initialize the model
-
-    llama_model_params model_params = llama_model_default_params();
-
-    model_params.n_gpu_layers = n_gpu_layers;
-
-    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.seed      = 1234;
-    ctx_params.n_ctx     = n_kv_max;
-    ctx_params.n_batch   = 512;
-    ctx_params.mul_mat_q = mmq;
-
-    ctx_params.n_threads       = params.n_threads;
-    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
-
-    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
-
-    if (ctx == NULL) {
-        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
-        return 1;
-    }
-
-    llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
-
-    // decode in batches of ctx_params.n_batch tokens
-    auto decode_helper = [](llama_context * ctx, llama_batch & batch, int32_t n_batch) {
-        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
-            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
-
-            llama_batch batch_view = {
-                n_tokens,
-                batch.token    + i,
-                nullptr,
-                batch.pos      + i,
-                batch.n_seq_id + i,
-                batch.seq_id   + i,
-                batch.logits   + i,
-                0, 0, 0, // unused
-            };
-
-            const int ret = llama_decode(ctx, batch_view);
-            if (ret != 0) {
-                LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
-                return false;
-            }
-        }
-
-        return true;
-    };
-
-    // warm up
-    {
-        for (int i = 0; i < 16; ++i) {
-            llama_batch_add(batch, 0, i, { 0 }, false);
-        }
-
-        if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
-            LOG_TEE("%s: llama_decode() failed\n", __func__);
-            return 1;
-        }
-    }
-
-    LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq);
-    LOG_TEE("\n");
-
-    LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");
-    LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
-
-    for (        int i_pp = 0; i_pp < (int) n_pp.size(); ++i_pp) {
-        for (    int i_tg = 0; i_tg < (int) n_tg.size(); ++i_tg) {
-            for (int i_pl = 0; i_pl < (int) n_pl.size(); ++i_pl) {
-                const int pp = n_pp[i_pp];
-                const int tg = n_tg[i_tg];
-                const int pl = n_pl[i_pl];
-
-                const int n_ctx_req = is_pp_shared ? pp + pl*tg : pl*(pp + tg);
-
-                if (n_ctx_req > n_kv_max) {
-                    continue;
-                }
-
-                llama_batch_clear(batch);
-
-                const int n_tokens = is_pp_shared ? pp : pl*pp;
-
-                for (int i = 0; i < n_tokens; ++i) {
-                    llama_batch_add(batch, 0, i, { 0 }, false);
-                }
-                batch.logits[batch.n_tokens - 1] = true;
-
-                const auto t_pp_start = ggml_time_us();
-
-                llama_kv_cache_clear(ctx);
-
-                if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
-                    LOG_TEE("%s: llama_decode() failed\n", __func__);
-                    return 1;
-                }
-
-                if (is_pp_shared) {
-                    for (int32_t i = 1; i < pl; ++i) {
-                        llama_kv_cache_seq_cp(ctx, 0, i, 0, pp);
-                    }
-                }
-
-                const auto t_pp_end = ggml_time_us();
-
-                const auto t_tg_start = ggml_time_us();
-
-                for (int i = 0; i < tg; ++i) {
-                    llama_batch_clear(batch);
-
-                    for (int j = 0; j < pl; ++j) {
-                        llama_batch_add(batch, 0, pp + i, { j }, true);
-                    }
-
-                    if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
-                        LOG_TEE("%s: llama_decode() failed\n", __func__);
-                        return 1;
-                    }
-                }
-
-                const auto t_tg_end = ggml_time_us();
-
-                const int32_t n_kv = n_ctx_req;
-
-                const float t_pp = (t_pp_end - t_pp_start) / 1000000.0f;
-                const float t_tg = (t_tg_end - t_tg_start) / 1000000.0f;
-                const float t    = t_pp + t_tg;
-
-                const float speed_pp = is_pp_shared ? pp / t_pp : pl*pp / t_pp;
-                const float speed_tg = pl*tg / t_tg;
-                const float speed    = n_kv / t;
-
-                LOG_TEE("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
-            }
-        }
-    }
-
-    llama_print_timings(ctx);
-
-    llama_batch_free(batch);
-
-    llama_free(ctx);
-    llama_free_model(model);
-
-    llama_backend_free();
-
-    fprintf(stderr, "\n\n");
-
-    return 0;
-}

examples/batched.swift/.gitignore

@@ -1,9 +0,0 @@
-.DS_Store
-/.build
-/Packages
-xcuserdata/
-DerivedData/
-.swiftpm/configuration/registries.json
-.swiftpm/xcode/package.xcworkspace/contents.xcworkspacedata
-.netrc
-batched_swift

examples/batched.swift/Makefile

@@ -1,6 +0,0 @@
-.PHONY: build
-
-build:
-	xcodebuild -scheme batched_swift -destination "generic/platform=macOS" -derivedDataPath build
-	rm -f ./batched_swift
-	ln -s ./build/Build/Products/Debug/batched_swift ./batched_swift

examples/batched.swift/Package.swift

@@ -1,22 +0,0 @@
-// swift-tools-version: 5.5
-// The swift-tools-version declares the minimum version of Swift required to build this package.
-
-import PackageDescription
-
-let package = Package(
-    name: "batched_swift",
-    platforms: [.macOS(.v12)],
-    dependencies: [
-        .package(name: "llama", path: "../../"),
-    ],
-    targets: [
-        // Targets are the basic building blocks of a package, defining a module or a test suite.
-        // Targets can depend on other targets in this package and products from dependencies.
-        .executableTarget(
-            name: "batched_swift",
-            dependencies: ["llama"],
-            path: "Sources",
-            linkerSettings: [.linkedFramework("Foundation"), .linkedFramework("AppKit")]
-        ),
-    ]
-)

examples/batched.swift/README.md

@@ -1,4 +0,0 @@
-This is a swift clone of `examples/batched`.
-
-$ `make`
-$ `./swift MODEL_PATH [PROMPT] [PARALLEL]`

examples/batched.swift/Sources/main.swift

@@ -1,263 +0,0 @@
-import Foundation
-import llama
-
-let arguments = CommandLine.arguments
-
-// Check that we have at least one argument (the model path)
-guard arguments.count > 1 else {
-    print("Usage: swift MODEL_PATH [PROMPT] [PARALLEL]")
-    exit(1)
-}
-
-let modelPath: String = arguments[1]
-let prompt: String = arguments.count > 2 ? arguments[2] : "Hello my name is"
-let n_parallel: Int = arguments.count > 3 && Int(arguments[3]) != nil ? Int(arguments[3])! : 1
-
-// total length of the sequences including the prompt
-let n_len: Int = 32
-
-// init LLM
-llama_backend_init(false)
-defer {
-    llama_backend_free()
-}
-
-let model_params = llama_model_default_params()
-guard let model = llama_load_model_from_file(modelPath.cString(using: .utf8), model_params) else {
-    print("Failed to load model")
-    exit(1)
-}
-
-defer {
-    llama_free_model(model)
-}
-
-var tokens = tokenize(text: prompt, add_bos: true)
-
-let n_kv_req = UInt32(tokens.count) + UInt32((n_len - Int(tokens.count)) * n_parallel)
-
-var context_params = llama_context_default_params()
-context_params.seed = 1234
-context_params.n_ctx = n_kv_req
-context_params.n_batch = UInt32(max(n_len, n_parallel))
-context_params.n_threads = 8
-context_params.n_threads_batch = 8
-
-let context = llama_new_context_with_model(model, context_params)
-guard context != nil else {
-    print("Failed to initialize context")
-    exit(1)
-}
-
-defer {
-    llama_free(context)
-}
-
-let n_ctx = llama_n_ctx(context)
-
-print("\nn_len = \(n_len), n_ctx = \(n_ctx), n_batch = \(context_params.n_batch), n_parallel = \(n_parallel), n_kv_req = \(n_kv_req)\n")
-
-if n_kv_req > n_ctx {
-    print("error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", n_kv_req)
-    exit(1)
-}
-
-var buffer: [CChar] = []
-for id: llama_token in tokens {
-    print(token_to_piece(token: id, buffer: &buffer) ?? "", terminator: "")
-}
-
-print("\n")
-
-var batch = llama_batch_init(max(Int32(tokens.count), Int32(n_parallel)), 0, 1)
-defer {
-    llama_batch_free(batch)
-}
-
-// evaluate the initial prompt
-batch.n_tokens = Int32(tokens.count)
-
-for (i, token) in tokens.enumerated() {
-    batch.token[i] = token
-    batch.pos[i] = Int32(i)
-    batch.n_seq_id[i] = 1
-    // batch.seq_id[i][0] = 0
-    // TODO: is this the proper way to do this?
-    if let seq_id = batch.seq_id[i] {
-        seq_id[0] = 0
-    }
-    batch.logits[i] = 0
-}
-
-// llama_decode will output logits only for the last token of the prompt
-batch.logits[Int(batch.n_tokens) - 1] = 1
-
-if llama_decode(context, batch) != 0 {
-    print("llama_decode() failed")
-    exit(1)
-}
-
-for i in 1 ..< n_parallel {
-    llama_kv_cache_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
-}
-
-if n_parallel > 1 {
-    print("generating \(n_parallel) sequences ...\n")
-}
-
-var streams: [String] = .init(repeating: "", count: n_parallel)
-var streamBuffers: [[CChar]] = .init(repeating: [], count: n_parallel)
-var i_batch = [Int32](repeating: batch.n_tokens - 1, count: n_parallel)
-
-var n_cur = batch.n_tokens
-var n_decode = 0
-
-let t_main_start = ggml_time_us()
-
-while n_cur <= n_len {
-    // prepare the next batch
-    batch.n_tokens = 0
-
-    // sample the next token for each parallel sequence / stream
-    for i in 0 ..< n_parallel {
-        if i_batch[i] < 0 {
-            // the stream has already finished
-            continue
-        }
-
-        var n_vocab = llama_n_vocab(model)
-        var logits = llama_get_logits_ith(context, i_batch[i])
-
-        var candidates: [llama_token_data] = .init(repeating: llama_token_data(), count: Int(n_vocab))
-
-        for token_id in 0 ..< n_vocab {
-            candidates.append(llama_token_data(id: token_id, logit: logits![Int(token_id)], p: 0.0))
-        }
-
-        var candidates_p: llama_token_data_array = .init(
-            data: &candidates,
-            size: candidates.count,
-            sorted: false
-        )
-
-        let top_k: Int32 = 40
-        let top_p: Float = 0.9
-        let temp: Float = 0.4
-
-        llama_sample_top_k(context, &candidates_p, top_k, 1)
-        llama_sample_top_p(context, &candidates_p, top_p, 1)
-        llama_sample_temp(context, &candidates_p, temp)
-
-        let new_token_id = llama_sample_token(context, &candidates_p)
-
-        // const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
-
-        // is it an end of stream? -> mark the stream as finished
-        if new_token_id == llama_token_eos(context) || n_cur == n_len {
-            i_batch[i] = -1
-            // print("")
-            if n_parallel > 1 {
-                print("stream \(i) finished at n_cur = \(n_cur)")
-            }
-
-            continue
-        }
-
-        let nextStringPiece = token_to_piece(token: new_token_id, buffer: &streamBuffers[i]) ?? ""
-
-        // if there is only one stream, we print immediately to stdout
-        if n_parallel == 1 {
-            print(nextStringPiece, terminator: "")
-        }
-        streams[i] += nextStringPiece
-
-        // push this new token for next evaluation
-        batch.token[Int(batch.n_tokens)] = new_token_id
-        batch.pos[Int(batch.n_tokens)] = n_cur
-        batch.n_seq_id[Int(batch.n_tokens)] = 1
-        if let seq_id = batch.seq_id[Int(batch.n_tokens)] {
-            seq_id[0] = Int32(i)
-        }
-        batch.logits[Int(batch.n_tokens)] = 1
-
-        i_batch[i] = batch.n_tokens
-
-        batch.n_tokens += 1
-
-        n_decode += 1
-    }
-
-    // all streams are finished
-    if batch.n_tokens == 0 {
-        break
-    }
-
-    n_cur += 1
-
-    // evaluate the current batch with the transformer model
-    if llama_decode(context, batch) != 0 {
-        print("llama_decode() failed")
-        exit(1)
-    }
-}
-
-if n_parallel > 1 {
-    print("\n")
-    for (i, stream) in streams.enumerated() {
-        print("sequence \(i):\n\n\(prompt)\(stream)\n")
-    }
-}
-
-let t_main_end = ggml_time_us()
-
-print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end - t_main_start) / 1_000_000.0)) s, speed: \(String(format: "%.2f", Double(n_decode) / (Double(t_main_end - t_main_start) / 1_000_000.0))) t/s\n")
-
-llama_print_timings(context)
-
-private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
-    let n_tokens = text.count + (add_bos ? 1 : 0)
-    let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
-    let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
-    var swiftTokens: [llama_token] = []
-    for i in 0 ..< tokenCount {
-        swiftTokens.append(tokens[Int(i)])
-    }
-    tokens.deallocate()
-    return swiftTokens
-}
-
-private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String? {
-    var result = [CChar](repeating: 0, count: 8)
-    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count))
-    if nTokens < 0 {
-        if result.count >= -Int(nTokens) {
-            result.removeLast(-Int(nTokens))
-        } else {
-            result.removeAll()
-        }
-        let check = llama_token_to_piece(
-            model,
-            token,
-            &result,
-            Int32(result.count)
-        )
-        assert(check == nTokens)
-    } else {
-        result.removeLast(result.count - Int(nTokens))
-    }
-    if buffer.isEmpty, let utfString = String(cString: result + [0], encoding: .utf8) {
-        return utfString
-    } else {
-        buffer.append(contentsOf: result)
-        let data = Data(buffer.map { UInt8(bitPattern: $0) })
-        if buffer.count >= 4 { // 4 bytes is the max length of a utf8 character so if we're here we need to reset the buffer
-            buffer = []
-        }
-        guard let bufferString = String(data: data, encoding: .utf8) else {
-            return nil
-        }
-        buffer = []
-        return bufferString
-    }
-    return nil
-}

examples/batched/batched.cpp

@@ -11,19 +11,12 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL] [LEN] [NGL]\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL]\n" , argv[0]);
         return 1 ;
     }
 
-    // number of parallel batches
     int n_parallel = 1;
 
-    // total length of the sequences including the prompt
-    int n_len = 32;
-
-    // number of layers to offload to the GPU
-    int n_gpu_layers = 0;
-
     if (argc >= 2) {
         params.model = argv[1];
     }
@@ -36,18 +29,13 @@ int main(int argc, char ** argv) {
         n_parallel = std::atoi(argv[3]);
     }
 
-    if (argc >= 5) {
-        n_len = std::atoi(argv[4]);
-    }
-
-    if (argc >= 6) {
-        n_gpu_layers = std::atoi(argv[5]);
-    }
-
     if (params.prompt.empty()) {
         params.prompt = "Hello my name is";
     }
 
+    // total length of the sequences including the prompt
+    const int n_len = 32;
+
     // init LLM
 
     llama_backend_init(params.numa);
@@ -56,7 +44,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = llama_model_default_params();
 
-    model_params.n_gpu_layers = n_gpu_layers;
+    // model_params.n_gpu_layers = 99; // offload all layers to the GPU
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -78,7 +66,7 @@ int main(int argc, char ** argv) {
     ctx_params.seed  = 1234;
     ctx_params.n_ctx = n_kv_req;
     ctx_params.n_batch = std::max(n_len, n_parallel);
-    ctx_params.n_threads       = params.n_threads;
+    ctx_params.n_threads = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
@@ -109,15 +97,20 @@ int main(int argc, char ** argv) {
 
     fflush(stderr);
 
-    // create a llama_batch
+    // create a llama_batch with size 512
     // we use this object to submit token data for decoding
-    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0, 1);
+
+    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0);
 
     // evaluate the initial prompt
-    for (size_t i = 0; i < tokens_list.size(); ++i) {
-        llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
+    batch.n_tokens = tokens_list.size();
+
+    for (int32_t i = 0; i < batch.n_tokens; i++) {
+        batch.token[i]  = tokens_list[i];
+        batch.pos[i]    = i;
+        batch.seq_id[i] = 0;
+        batch.logits[i] = false;
     }
-    GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
 
     // llama_decode will output logits only for the last token of the prompt
     batch.logits[batch.n_tokens - 1] = true;
@@ -153,7 +146,7 @@ int main(int argc, char ** argv) {
 
     while (n_cur <= n_len) {
         // prepare the next batch
-        llama_batch_clear(batch);
+        batch.n_tokens = 0;
 
         // sample the next token for each parallel sequence / stream
         for (int32_t i = 0; i < n_parallel; ++i) {
@@ -187,7 +180,7 @@ int main(int argc, char ** argv) {
             //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of stream? -> mark the stream as finished
-            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
+            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
                 i_batch[i] = -1;
                 LOG_TEE("\n");
                 if (n_parallel > 1) {
@@ -205,10 +198,15 @@ int main(int argc, char ** argv) {
 
             streams[i] += llama_token_to_piece(ctx, new_token_id);
 
+            // push this new token for next evaluation
+            batch.token [batch.n_tokens] = new_token_id;
+            batch.pos   [batch.n_tokens] = n_cur;
+            batch.seq_id[batch.n_tokens] = i;
+            batch.logits[batch.n_tokens] = true;
+
             i_batch[i] = batch.n_tokens;
 
-            // push this new token for next evaluation
-            llama_batch_add(batch, new_token_id, n_cur, { i }, true);
+            batch.n_tokens += 1;
 
             n_decode += 1;
         }

examples/beam-search/beam-search.cpp

@@ -47,7 +47,7 @@ struct beam_search_callback_data {
 // In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
 // For example, eob can be flagged due to maximum token length, stop words, etc.
 static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
-    return n_tokens && tokens[n_tokens-1] == llama_token_eos(llama_get_model(callback_data.ctx));
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx);
 }
 
 // Function matching type llama_beam_search_callback_fn_t.

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -536,7 +536,7 @@ static bool is_ggml_file(const char * filename) {
     if (file.size < 4) {
         return false;
     }
-    std::string magic = file.read_string(4);
+    uint32_t magic = file.read_u32();
     return magic == GGUF_MAGIC;
 }
 

examples/embd-input/.gitignore

@@ -0,0 +1,4 @@
+PandaGPT
+MiniGPT-4
+*.pth
+

examples/embd-input/CMakeLists.txt

@@ -0,0 +1,17 @@
+set(TARGET embdinput)
+add_library(${TARGET} embd-input-lib.cpp embd-input.h)
+install(TARGETS ${TARGET} LIBRARY)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()
+
+set(TARGET embd-input-test)
+add_executable(${TARGET} embd-input-test.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama embdinput ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/embd-input/README.md

@@ -0,0 +1,63 @@
+### Examples for input embedding directly
+
+## Requirement
+build  `libembdinput.so`
+run the following comman in main dir (../../).
+```
+make
+```
+
+## [LLaVA](https://github.com/haotian-liu/LLaVA/) example  (llava.py)
+
+1. Obtian LLaVA model (following https://github.com/haotian-liu/LLaVA/ , use https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/).
+2. Convert it to ggml format.
+3. `llava_projection.pth` is [pytorch_model-00003-of-00003.bin](https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/blob/main/pytorch_model-00003-of-00003.bin).
+
+```
+import torch
+
+bin_path = "../LLaVA-13b-delta-v1-1/pytorch_model-00003-of-00003.bin"
+pth_path = "./examples/embd-input/llava_projection.pth"
+
+dic = torch.load(bin_path)
+used_key = ["model.mm_projector.weight","model.mm_projector.bias"]
+torch.save({k: dic[k] for k in used_key}, pth_path)
+```
+4. Check the path of LLaVA model and `llava_projection.pth` in `llava.py`.
+
+
+## [PandaGPT](https://github.com/yxuansu/PandaGPT) example (panda_gpt.py)
+
+1. Obtian PandaGPT lora model from https://github.com/yxuansu/PandaGPT. Rename the file to `adapter_model.bin`. Use [convert-lora-to-ggml.py](../../convert-lora-to-ggml.py) to convert it to ggml format.
+The `adapter_config.json` is
+```
+{
+  "peft_type": "LORA",
+  "fan_in_fan_out": false,
+  "bias": null,
+  "modules_to_save": null,
+  "r": 32,
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj"]
+}
+```
+2. Papare the `vicuna` v0 model.
+3. Obtain the [ImageBind](https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth) model.
+4. Clone the PandaGPT source.
+```
+git clone https://github.com/yxuansu/PandaGPT
+```
+5. Install the requirement of PandaGPT.
+6. Check the path of PandaGPT source, ImageBind model, lora model and vicuna model in panda_gpt.py.
+
+## [MiniGPT-4](https://github.com/Vision-CAIR/MiniGPT-4/) example (minigpt4.py)
+
+1. Obtain MiniGPT-4 model from https://github.com/Vision-CAIR/MiniGPT-4/ and put it in `embd-input`.
+2. Clone the MiniGPT-4 source.
+```
+git clone https://github.com/Vision-CAIR/MiniGPT-4/
+```
+3. Install the requirement of PandaGPT.
+4. Papare the `vicuna` v0 model.
+5. Check the path of MiniGPT-4 source, MiniGPT-4 model and vicuna model in `minigpt4.py`.

examples/embd-input/embd-input-lib.cpp

@@ -0,0 +1,220 @@
+#include "build-info.h"
+#include "common.h"
+#include "embd-input.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+static llama_context ** g_ctx;
+
+extern "C" {
+
+struct MyModel* create_mymodel(int argc, char ** argv) {
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return nullptr;
+    }
+
+    print_build_info();
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = uint32_t(time(NULL));
+    }
+    fprintf(stderr, "%s: seed  = %d\n", __func__, params.seed);
+
+    llama_backend_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+
+    g_ctx = &ctx;
+
+    // load the model and apply lora adapter, if any
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        return nullptr;
+    }
+
+    // print system information
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+    }
+    struct MyModel * ret = new MyModel();
+    ret->ctx = ctx;
+    ret->params = params;
+    ret->n_past = 0;
+    // printf("ctx: %d\n", ret->ctx);
+    return ret;
+}
+
+void free_mymodel(struct MyModel * mymodel) {
+    llama_context * ctx = mymodel->ctx;
+    llama_print_timings(ctx);
+    llama_free(ctx);
+    delete mymodel;
+}
+
+
+bool eval_float(void * model, float * input, int N){
+    MyModel * mymodel = (MyModel*)model;
+    llama_context * ctx = mymodel->ctx;
+    gpt_params params = mymodel->params;
+    int n_emb = llama_n_embd(llama_get_model(ctx));
+    int n_past = mymodel->n_past;
+    int n_batch = N; // params.n_batch;
+
+    for (int i = 0; i < (int) N; i += n_batch) {
+        int n_eval = (int) N - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        llama_batch batch = {  int32_t(n_eval), nullptr, (input+i*n_emb), nullptr, nullptr, nullptr, n_past, 1, 0, };
+        if (llama_decode(ctx, batch)) {
+            fprintf(stderr, "%s : failed to eval\n", __func__);
+            return false;
+        }
+        n_past += n_eval;
+    }
+    mymodel->n_past = n_past;
+    return true;
+}
+
+bool eval_tokens(void * model, std::vector<llama_token> tokens) {
+    MyModel * mymodel = (MyModel* )model;
+    llama_context * ctx;
+    ctx = mymodel->ctx;
+    gpt_params params = mymodel->params;
+    int n_past = mymodel->n_past;
+    for (int i = 0; i < (int) tokens.size(); i += params.n_batch) {
+        int n_eval = (int) tokens.size() - i;
+        if (n_eval > params.n_batch) {
+            n_eval = params.n_batch;
+        }
+        if (llama_decode(ctx, 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;
+    }
+    mymodel->n_past = n_past;
+    return true;
+}
+
+bool eval_id(struct MyModel* mymodel, int id) {
+    std::vector<llama_token> tokens;
+    tokens.push_back(id);
+    return eval_tokens(mymodel, tokens);
+}
+
+bool eval_string(struct MyModel * mymodel,const char* str){
+    llama_context * ctx = mymodel->ctx;
+    std::string str2 = str;
+    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx, str2, true);
+    eval_tokens(mymodel, embd_inp);
+    return true;
+}
+
+llama_token sampling_id(struct MyModel* mymodel) {
+    llama_context* ctx = mymodel->ctx;
+    gpt_params params = mymodel->params;
+    // int n_ctx = llama_n_ctx(ctx);
+
+    // out of user input, sample next token
+    const float   temp            = params.temp;
+    const int32_t top_k           = params.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx)) : params.top_k;
+    const float   top_p           = params.top_p;
+    const float   tfs_z           = params.tfs_z;
+    const float   typical_p       = params.typical_p;
+    // const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
+    // const float   repeat_penalty  = params.repeat_penalty;
+    // const float   alpha_presence  = params.presence_penalty;
+    // const float   alpha_frequency = params.frequency_penalty;
+    const int     mirostat        = params.mirostat;
+    const float   mirostat_tau    = params.mirostat_tau;
+    const float   mirostat_eta    = params.mirostat_eta;
+    // const bool    penalize_nl     = params.penalize_nl;
+
+    llama_token id = 0;
+    {
+        auto logits  = llama_get_logits(ctx);
+        auto n_vocab = llama_n_vocab(llama_get_model(ctx));
+
+        // Apply params.logit_bias map
+        for (auto it = params.logit_bias.begin(); it != params.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, &candidates_p);
+        } else {
+            if (mirostat == 1) {
+                static float mirostat_mu = 2.0f * mirostat_tau;
+                const int mirostat_m = 100;
+                llama_sample_temp(ctx, &candidates_p, temp);
+                id = llama_sample_token_mirostat(ctx, &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, &candidates_p, temp);
+                id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
+            } else {
+                // Temperature sampling
+                llama_sample_top_k(ctx, &candidates_p, top_k, 1);
+                llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
+                llama_sample_typical(ctx, &candidates_p, typical_p, 1);
+                llama_sample_top_p(ctx, &candidates_p, top_p, 1);
+                llama_sample_temp(ctx, &candidates_p, temp);
+                id = llama_sample_token(ctx, &candidates_p);
+            }
+        }
+    }
+
+    return id;
+}
+
+const char * sampling(struct MyModel * mymodel) {
+    llama_context * ctx = mymodel->ctx;
+    int id = sampling_id(mymodel);
+    static std::string ret;
+    if (id == llama_token_eos(ctx)) {
+        ret = "</s>";
+    } else {
+        ret = llama_token_to_piece(ctx, id);
+    }
+    eval_id(mymodel, id);
+    return ret.c_str();
+}
+
+}

examples/embd-input/embd-input-test.cpp

@@ -0,0 +1,35 @@
+#include "embd-input.h"
+#include <stdlib.h>
+#include <random>
+#include <string.h>
+
+int main(int argc, char** argv) {
+
+    auto mymodel = create_mymodel(argc, argv);
+    int N = 10;
+    int max_tgt_len = 500;
+    int n_embd = llama_n_embd(llama_get_model(mymodel->ctx));
+
+    // add random float embd to test evaluation
+    float * data = new float[N*n_embd];
+    std::default_random_engine e;
+    std::uniform_real_distribution<float>  u(0,1);
+    for (int i=0;i<N*n_embd;i++) {
+        data[i] = u(e);
+    }
+
+    eval_string(mymodel, "user: what is the color of the flag of UN?");
+    eval_float(mymodel, data, N);
+    eval_string(mymodel, "assistant:");
+    eval_string(mymodel, mymodel->params.prompt.c_str());
+    const char* tmp;
+    for (int i=0; i<max_tgt_len; i++) {
+        tmp = sampling(mymodel);
+        if (strcmp(tmp, "</s>")==0) break;
+        printf("%s", tmp);
+        fflush(stdout);
+    }
+    printf("\n");
+    free_mymodel(mymodel);
+    return 0;
+}

examples/embd-input/embd-input.h

@@ -0,0 +1,27 @@
+#ifndef _EMBD_INPUT_H_
+#define _EMBD_INPUT_H_ 1
+
+#include "common.h"
+#include "llama.h"
+
+extern "C" {
+
+typedef struct MyModel {
+    llama_context* ctx;
+    gpt_params params;
+    int n_past = 0;
+} MyModel;
+
+struct MyModel* create_mymodel(int argc, char ** argv);
+
+bool eval_float(void* model, float* input, int N);
+bool eval_tokens(void* model, std::vector<llama_token> tokens);
+bool eval_id(struct MyModel* mymodel, int id);
+bool eval_string(struct MyModel* mymodel, const char* str);
+const char * sampling(struct MyModel* mymodel);
+llama_token sampling_id(struct MyModel* mymodel);
+void free_mymodel(struct MyModel* mymodel);
+
+}
+
+#endif

examples/embd-input/embd_input.py

@@ -0,0 +1,72 @@
+#!/usr/bin/env python3
+import ctypes
+from ctypes import cdll, c_char_p, c_void_p, POINTER, c_float, c_int
+import numpy as np
+import os
+
+libc = cdll.LoadLibrary("./libembdinput.so")
+libc.sampling.restype=c_char_p
+libc.create_mymodel.restype=c_void_p
+libc.eval_string.argtypes=[c_void_p, c_char_p]
+libc.sampling.argtypes=[c_void_p]
+libc.eval_float.argtypes=[c_void_p, POINTER(c_float), c_int]
+
+
+class MyModel:
+    def __init__(self, args):
+        argc = len(args)
+        c_str = [c_char_p(i.encode()) for i in args]
+        args_c = (c_char_p * argc)(*c_str)
+        self.model = c_void_p(libc.create_mymodel(argc, args_c))
+        self.max_tgt_len = 512
+        self.print_string_eval = True
+
+    def __del__(self):
+        libc.free_mymodel(self.model)
+
+    def eval_float(self, x):
+        libc.eval_float(self.model, x.astype(np.float32).ctypes.data_as(POINTER(c_float)), x.shape[1])
+
+    def eval_string(self, x):
+        libc.eval_string(self.model, x.encode()) # c_char_p(x.encode()))
+        if self.print_string_eval:
+            print(x)
+
+    def eval_token(self, x):
+        libc.eval_id(self.model, x)
+
+    def sampling(self):
+        s = libc.sampling(self.model)
+        return s
+
+    def stream_generate(self, end="</s>"):
+        ret = b""
+        end = end.encode()
+        for _ in range(self.max_tgt_len):
+            tmp = self.sampling()
+            ret += tmp
+            yield tmp
+            if ret.endswith(end):
+                break
+
+    def generate_with_print(self, end="</s>"):
+        ret = b""
+        for i in self.stream_generate(end=end):
+            ret += i
+            print(i.decode(errors="replace"), end="", flush=True)
+        print("")
+        return ret.decode(errors="replace")
+
+
+    def generate(self, end="</s>"):
+        text = b"".join(self.stream_generate(end=end))
+        return text.decode(errors="replace")
+
+if __name__ == "__main__":
+    model = MyModel(["main", "--model", "../llama.cpp/models/ggml-vic13b-q4_1.bin", "-c", "2048"])
+    model.eval_string("""user: what is the color of the flag of UN?""")
+    x = np.random.random((5120,10))# , dtype=np.float32)
+    model.eval_float(x)
+    model.eval_string("""assistant:""")
+    for i in model.generate():
+        print(i.decode(errors="replace"), end="", flush=True)

examples/embd-input/llava.py

@@ -0,0 +1,71 @@
+#!/usr/bin/env python3
+import sys
+import os
+sys.path.insert(0, os.path.dirname(__file__))
+from embd_input import MyModel
+import numpy as np
+from torch import nn
+import torch
+from transformers import CLIPVisionModel,  CLIPImageProcessor
+from PIL import Image
+
+# model parameters from 'liuhaotian/LLaVA-13b-delta-v1-1'
+vision_tower = "openai/clip-vit-large-patch14"
+select_hidden_state_layer = -2
+# (vision_config.image_size // vision_config.patch_size) ** 2
+image_token_len = (224//14)**2
+
+class Llava:
+    def __init__(self, args):
+        self.image_processor = CLIPImageProcessor.from_pretrained(vision_tower)
+        self.vision_tower = CLIPVisionModel.from_pretrained(vision_tower)
+        self.mm_projector = nn.Linear(1024, 5120)
+        self.model = MyModel(["main", *args])
+
+    def load_projection(self, path):
+        state = torch.load(path)
+        self.mm_projector.load_state_dict({
+            "weight": state["model.mm_projector.weight"],
+            "bias": state["model.mm_projector.bias"]})
+
+    def chat(self, question):
+        self.model.eval_string("user: ")
+        self.model.eval_string(question)
+        self.model.eval_string("\nassistant: ")
+        return self.model.generate_with_print()
+
+    def chat_with_image(self, image, question):
+        with torch.no_grad():
+            embd_image = self.image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+            image_forward_out = self.vision_tower(embd_image.unsqueeze(0), output_hidden_states=True)
+            select_hidden_state = image_forward_out.hidden_states[select_hidden_state_layer]
+            image_feature = select_hidden_state[:, 1:]
+            embd_image = self.mm_projector(image_feature)
+            embd_image = embd_image.cpu().numpy()[0]
+        self.model.eval_string("user: ")
+        self.model.eval_token(32003-2) # im_start
+        self.model.eval_float(embd_image.T)
+        for i in range(image_token_len-embd_image.shape[0]):
+            self.model.eval_token(32003-3) # im_patch
+        self.model.eval_token(32003-1) # im_end
+        self.model.eval_string(question)
+        self.model.eval_string("\nassistant: ")
+        return self.model.generate_with_print()
+
+
+if __name__=="__main__":
+    # model form liuhaotian/LLaVA-13b-delta-v1-1
+    a = Llava(["--model", "./models/ggml-llava-13b-v1.1.bin", "-c", "2048"])
+    # Extract from https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/blob/main/pytorch_model-00003-of-00003.bin.
+    # Also here can use pytorch_model-00003-of-00003.bin directly.
+    a.load_projection(os.path.join(
+        os.path.dirname(__file__) ,
+        "llava_projection.pth"))
+    respose = a.chat_with_image(
+        Image.open("./media/llama1-logo.png").convert('RGB'),
+        "what is the text in the picture?")
+    respose
+    a.chat("what is the color of it?")
+
+
+

examples/embd-input/minigpt4.py

@@ -0,0 +1,129 @@
+#!/usr/bin/env python3
+import sys
+import os
+sys.path.insert(0, os.path.dirname(__file__))
+from embd_input import MyModel
+import numpy as np
+from torch import nn
+import torch
+from PIL import Image
+
+minigpt4_path = os.path.join(os.path.dirname(__file__), "MiniGPT-4")
+sys.path.insert(0, minigpt4_path)
+from minigpt4.models.blip2 import Blip2Base
+from minigpt4.processors.blip_processors import Blip2ImageEvalProcessor
+
+
+class MiniGPT4(Blip2Base):
+    """
+    MiniGPT4 model from https://github.com/Vision-CAIR/MiniGPT-4
+    """
+    def __init__(self,
+        args,
+        vit_model="eva_clip_g",
+        q_former_model="https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/blip2_pretrained_flant5xxl.pth",
+        img_size=224,
+        drop_path_rate=0,
+        use_grad_checkpoint=False,
+        vit_precision="fp32",
+        freeze_vit=True,
+        freeze_qformer=True,
+        num_query_token=32,
+        llama_model="",
+        prompt_path="",
+        prompt_template="",
+        max_txt_len=32,
+        end_sym='\n',
+        low_resource=False,  # use 8 bit and put vit in cpu
+        device_8bit=0
+    ):
+        super().__init__()
+        self.img_size = img_size
+        self.low_resource = low_resource
+        self.preprocessor = Blip2ImageEvalProcessor(img_size)
+
+        print('Loading VIT')
+        self.visual_encoder, self.ln_vision = self.init_vision_encoder(
+            vit_model, img_size, drop_path_rate, use_grad_checkpoint, vit_precision
+        )
+        print('Loading VIT Done')
+        print('Loading Q-Former')
+        self.Qformer, self.query_tokens = self.init_Qformer(
+            num_query_token, self.visual_encoder.num_features
+        )
+        self.Qformer.cls = None
+        self.Qformer.bert.embeddings.word_embeddings = None
+        self.Qformer.bert.embeddings.position_embeddings = None
+        for layer in self.Qformer.bert.encoder.layer:
+            layer.output = None
+            layer.intermediate = None
+        self.load_from_pretrained(url_or_filename=q_former_model)
+        print('Loading Q-Former Done')
+        self.llama_proj = nn.Linear(
+            self.Qformer.config.hidden_size, 5120 # self.llama_model.config.hidden_size
+        )
+        self.max_txt_len = max_txt_len
+        self.end_sym = end_sym
+        self.model = MyModel(["main", *args])
+        # system prompt
+        self.model.eval_string("Give the following image: <Img>ImageContent</Img>. "
+           "You will be able to see the image once I provide it to you. Please answer my questions."
+           "###")
+
+    def encode_img(self, image):
+        image = self.preprocessor(image)
+        image = image.unsqueeze(0)
+        device = image.device
+        if self.low_resource:
+            self.vit_to_cpu()
+            image = image.to("cpu")
+
+        with self.maybe_autocast():
+            image_embeds = self.ln_vision(self.visual_encoder(image)).to(device)
+            image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(device)
+
+            query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
+            query_output = self.Qformer.bert(
+                query_embeds=query_tokens,
+                encoder_hidden_states=image_embeds,
+                encoder_attention_mask=image_atts,
+                return_dict=True,
+            )
+
+            inputs_llama = self.llama_proj(query_output.last_hidden_state)
+            # atts_llama = torch.ones(inputs_llama.size()[:-1], dtype=torch.long).to(image.device)
+        return inputs_llama
+
+    def load_projection(self, path):
+        state = torch.load(path)["model"]
+        self.llama_proj.load_state_dict({
+            "weight": state["llama_proj.weight"],
+            "bias": state["llama_proj.bias"]})
+
+    def chat(self, question):
+        self.model.eval_string("Human: ")
+        self.model.eval_string(question)
+        self.model.eval_string("\n### Assistant:")
+        return self.model.generate_with_print(end="###")
+
+    def chat_with_image(self, image, question):
+        with torch.no_grad():
+            embd_image = self.encode_img(image)
+        embd_image = embd_image.cpu().numpy()[0]
+        self.model.eval_string("Human: <Img>")
+        self.model.eval_float(embd_image.T)
+        self.model.eval_string("</Img> ")
+        self.model.eval_string(question)
+        self.model.eval_string("\n### Assistant:")
+        return self.model.generate_with_print(end="###")
+
+
+if __name__=="__main__":
+    a = MiniGPT4(["--model", "./models/ggml-vicuna-13b-v0-q4_1.bin", "-c", "2048"])
+    a.load_projection(os.path.join(
+        os.path.dirname(__file__) ,
+        "pretrained_minigpt4.pth"))
+    respose = a.chat_with_image(
+        Image.open("./media/llama1-logo.png").convert('RGB'),
+        "what is the text in the picture?")
+    a.chat("what is the color of it?")

examples/embd-input/panda_gpt.py

@@ -0,0 +1,99 @@
+#!/usr/bin/env python3
+import sys
+import os
+sys.path.insert(0, os.path.dirname(__file__))
+from embd_input import MyModel
+import numpy as np
+from torch import nn
+import torch
+
+# use PandaGPT path
+panda_gpt_path = os.path.join(os.path.dirname(__file__), "PandaGPT")
+imagebind_ckpt_path = "./models/panda_gpt/"
+
+sys.path.insert(0, os.path.join(panda_gpt_path,"code","model"))
+from ImageBind.models import imagebind_model
+from ImageBind import data
+
+ModalityType = imagebind_model.ModalityType
+max_tgt_len = 400
+
+class PandaGPT:
+    def __init__(self, args):
+        self.visual_encoder,_ = imagebind_model.imagebind_huge(pretrained=True, store_path=imagebind_ckpt_path)
+        self.visual_encoder.eval()
+        self.llama_proj = nn.Linear(1024, 5120) # self.visual_hidden_size, 5120)
+        self.max_tgt_len = max_tgt_len
+        self.model = MyModel(["main", *args])
+        self.generated_text = ""
+        self.device = "cpu"
+
+    def load_projection(self, path):
+        state = torch.load(path, map_location="cpu")
+        self.llama_proj.load_state_dict({
+            "weight": state["llama_proj.weight"],
+            "bias": state["llama_proj.bias"]})
+
+    def eval_inputs(self, inputs):
+        self.model.eval_string("<Img>")
+        embds = self.extract_multimoal_feature(inputs)
+        for i in embds:
+            self.model.eval_float(i.T)
+        self.model.eval_string("</Img> ")
+
+    def chat(self, question):
+        return self.chat_with_image(None, question)
+
+    def chat_with_image(self, inputs, question):
+        if self.generated_text == "":
+            self.model.eval_string("###")
+        self.model.eval_string(" Human: ")
+        if inputs:
+            self.eval_inputs(inputs)
+        self.model.eval_string(question)
+        self.model.eval_string("\n### Assistant:")
+        ret = self.model.generate_with_print(end="###")
+        self.generated_text += ret
+        return ret
+
+    def extract_multimoal_feature(self, inputs):
+        features = []
+        for key in ["image", "audio", "video", "thermal"]:
+            if key + "_paths" in inputs:
+                embeds = self.encode_data(key, inputs[key+"_paths"])
+                features.append(embeds)
+        return features
+
+    def encode_data(self, data_type, data_paths):
+
+        type_map = {
+            "image": ModalityType.VISION,
+            "audio": ModalityType.AUDIO,
+            "video": ModalityType.VISION,
+            "thermal": ModalityType.THERMAL,
+        }
+        load_map = {
+            "image": data.load_and_transform_vision_data,
+            "audio": data.load_and_transform_audio_data,
+            "video": data.load_and_transform_video_data,
+            "thermal": data.load_and_transform_thermal_data
+        }
+
+        load_function = load_map[data_type]
+        key = type_map[data_type]
+
+        inputs = {key: load_function(data_paths, self.device)}
+        with torch.no_grad():
+            embeddings = self.visual_encoder(inputs)
+            embeds = embeddings[key]
+            embeds = self.llama_proj(embeds).cpu().numpy()
+        return embeds
+
+
+if __name__=="__main__":
+    a = PandaGPT(["--model", "./models/ggml-vicuna-13b-v0-q4_1.bin", "-c", "2048", "--lora", "./models/panda_gpt/ggml-adapter-model.bin","--temp", "0"])
+    a.load_projection("./models/panda_gpt/adapter_model.bin")
+    a.chat_with_image(
+        {"image_paths": ["./media/llama1-logo.png"]},
+        "what is the text in the picture? 'llama' or 'lambda'?")
+    a.chat("what is the color of it?")

examples/finetune/finetune.cpp

@@ -529,14 +529,13 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
     set_param_lora(lora);
 
     // measure data size
-    size_t size = 0;
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        size += GGML_PAD(ggml_nbytes(t), tensor_alignment);
-    }
+    struct ggml_allocr * alloc = NULL;
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    alloc_lora(alloc, lora);
 
     // allocate data
-    struct ggml_allocr * alloc = NULL;
-    lora->data.resize(size + tensor_alignment);
+    lora->data.resize(ggml_allocr_max_size(alloc) + tensor_alignment);
+    ggml_allocr_free(alloc);
     alloc = ggml_allocr_new(lora->data.data(), lora->data.size(), tensor_alignment);
     alloc_lora(alloc, lora);
     ggml_allocr_free(alloc);
@@ -1715,9 +1714,11 @@ int main(int argc, char ** argv) {
     struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
 
     // measure required memory for input tensors
-    size_t max_input_size = GGML_PAD(ggml_nbytes(tokens_input), tensor_alignment) +
-                            GGML_PAD(ggml_nbytes(target_probs), tensor_alignment) +
-                            tensor_alignment;
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    ggml_allocr_alloc(alloc, tokens_input);
+    ggml_allocr_alloc(alloc, target_probs);
+    size_t max_input_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+    ggml_allocr_free(alloc);
     printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
 
     // allocate input tensors

examples/infill/CMakeLists.txt

@@ -4,5 +4,5 @@ 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(${TARGET} BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
 endif()

examples/infill/infill.cpp

@@ -39,9 +39,9 @@ static gpt_params               * g_params;
 static std::vector<llama_token> * g_input_tokens;
 static std::ostringstream       * g_output_ss;
 static std::vector<llama_token> * g_output_tokens;
-
 static bool is_interacting = false;
 
+
 static void write_logfile(
     const llama_context * ctx, const gpt_params & params, const llama_model * model,
     const std::vector<llama_token> & input_tokens, const std::string & output,
@@ -104,7 +104,6 @@ static void sigint_handler(int signo) {
 
 int main(int argc, char ** argv) {
     gpt_params params;
-    llama_sampling_params & sparams = params.sparams;
     g_params = &params;
 
     if (!gpt_params_parse(argc, argv, params)) {
@@ -207,7 +206,7 @@ int main(int argc, char ** argv) {
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    if (sparams.cfg_scale > 1.f) {
+    if (params.cfg_scale > 1.f) {
         struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
         ctx_guidance = llama_new_context_with_model(model, lparams);
     }
@@ -234,35 +233,23 @@ int main(int argc, char ** argv) {
     const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
     LOG("add_bos: %d\n", add_bos);
 
-    bool suff_rm_leading_spc = params.escape;
-    if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
-        params.input_suffix.erase(0, 1);
-        suff_rm_leading_spc = false;
-    }
     std::vector<llama_token> embd_inp;
-    std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
-    std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
-    const int space_token = 29871;
-    if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
-        inp_sfx.erase(inp_sfx.begin());
-    }
-    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
-    if (add_bos) {
-        inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
-    }
-    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
+    std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, add_bos);
+    std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, add_bos);
+    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
     embd_inp = inp_pfx;
     embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-    embd_inp.push_back(llama_token_middle(model));
+    embd_inp.push_back(llama_token_middle(ctx));
 
     LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
     LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
-    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
-        embd_inp.push_back(llama_token_bos(model));
-        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+        embd_inp.push_back(llama_token_bos(ctx));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
     }
 
     // Tokenize negative prompt
@@ -270,13 +257,13 @@ int main(int argc, char ** argv) {
     int guidance_offset = 0;
     int original_prompt_len = 0;
     if (ctx_guidance) {
-        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
+        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(params.cfg_negative_prompt));
 
-        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos);
-        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
+        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
 
         std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
-        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
 
         original_prompt_len = original_inp.size();
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
@@ -294,8 +281,8 @@ int main(int argc, char ** argv) {
         params.n_keep = (int)embd_inp.size();
     }
 
-    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
-    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
 
 
     // enable interactive mode if interactive start is specified
@@ -313,7 +300,7 @@ int main(int argc, char ** argv) {
 
         if (ctx_guidance) {
             LOG_TEE("\n");
-            LOG_TEE("%s: negative prompt: '%s'\n", __func__, sparams.cfg_negative_prompt.c_str());
+            LOG_TEE("%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
             LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
             for (int i = 0; i < (int) guidance_inp.size(); i++) {
                 LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
@@ -358,10 +345,39 @@ int main(int argc, char ** argv) {
             LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
         }
     }
-    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
+    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
+            params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau);
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     LOG_TEE("\n\n");
 
+    struct llama_grammar * grammar = NULL;
+    grammar_parser::parse_state parsed_grammar;
+
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
+        LOG_TEE("%s: grammar:\n", __func__);
+        grammar_parser::print_grammar(stderr, parsed_grammar);
+        LOG_TEE("\n");
+
+        {
+            auto it = params.logit_bias.find(llama_token_eos(ctx));
+            if (it != params.logit_bias.end() && it->second == -INFINITY) {
+                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
+            }
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar = llama_grammar_init(
+            grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    }
+
+    // TODO: replace with ring-buffer
+    std::vector<llama_token> last_tokens(n_ctx);
+    std::fill(last_tokens.begin(), last_tokens.end(), 0);
     LOG_TEE("\n#####  Infill mode  #####\n\n");
     if (params.infill) {
         printf("\n************\n");
@@ -404,7 +420,10 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
+    const int n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
 
     while (n_remain != 0 || params.interactive) {
         // predict
@@ -451,7 +470,7 @@ int main(int argc, char ** argv) {
 
                 LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
 
-                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
 
             }
 
@@ -479,7 +498,7 @@ int main(int argc, char ** argv) {
                     input_buf  = embd_guidance.data();
                     input_size = embd_guidance.size();
 
-                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance).c_str());
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
                 } else {
                     input_buf  = embd.data();
                     input_size = embd.size();
@@ -502,7 +521,7 @@ int main(int argc, char ** argv) {
                     n_eval = params.n_batch;
                 }
 
-                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
 
                 if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
                     LOG_TEE("%s : failed to eval\n", __func__);
@@ -521,11 +540,12 @@ int main(int argc, char ** argv) {
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
 
-            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
+            const llama_token id = llama_sample_token(ctx, ctx_guidance, grammar, params, last_tokens, candidates);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            last_tokens.erase(last_tokens.begin());
+            last_tokens.push_back(id);
 
-            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
 
             embd.push_back(id);
 
@@ -541,11 +561,8 @@ int main(int argc, char ** argv) {
             LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
             while ((int) embd_inp.size() > n_consumed) {
                 embd.push_back(embd_inp[n_consumed]);
-
-                // push the prompt in the sampling context in order to apply repetition penalties later
-                // for the prompt, we don't apply grammar rules
-                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
-
+                last_tokens.erase(last_tokens.begin());
+                last_tokens.push_back(embd_inp[n_consumed]);
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
                     break;
@@ -577,10 +594,10 @@ int main(int argc, char ** argv) {
         if ((int) embd_inp.size() <= n_consumed) {
 
             // deal with eot token in infill mode
-            if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
+            if ((last_tokens.back() == llama_token_eot(ctx) || is_interacting) && params.interactive){
                 if(is_interacting && !params.interactive_first) {
                     // print an eot token
-                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
+                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
                 }
                 fflush(stdout);
                 printf("\n");
@@ -594,7 +611,7 @@ int main(int argc, char ** argv) {
                     buffer += line;
                 } while (another_line);
                 // check if we got an empty line, if so we use the old input
-                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
                     params.input_prefix = buffer;
                 }
                 buffer.clear();
@@ -604,37 +621,20 @@ int main(int argc, char ** argv) {
                     buffer += line;
                 } while (another_line);
                 // check if we got an empty line
-                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
                     params.input_suffix = buffer;
                 }
                 buffer.clear();
                 // done taking input, reset color
                 console::set_display(console::reset);
-
-                if (params.escape) {
-                    //process escape sequences, for the initial prompt this is done in common.cpp when we load the params, but for the interactive mode we need to do it here
-                    process_escapes(params.input_prefix);
-                    process_escapes(params.input_suffix);
-                }
-                suff_rm_leading_spc = params.escape;
-                if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
-                    params.input_suffix.erase(0, 1);
-                    suff_rm_leading_spc = false;
-                }
                 // tokenize new prefix and suffix
-                std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
-                std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
-                if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
-                    inp_sfx.erase(inp_sfx.begin());
-                }
-                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
-                if (add_bos) {
-                    inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
-                }
-                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
+                std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, add_bos);
+                std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, add_bos);
+                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
                 embd_inp = inp_pfx;
                 embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-                embd_inp.push_back(llama_token_middle(model));
+                embd_inp.push_back(llama_token_middle(ctx));
                 embd.clear();
                 embd_guidance.clear();
                 n_remain = params.n_predict;
@@ -644,7 +644,7 @@ int main(int argc, char ** argv) {
                 is_interacting = false;
             }
             // deal with end of text token in interactive mode
-            else if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
+            else if (last_tokens.back() == llama_token_eos(ctx)) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -661,7 +661,7 @@ int main(int argc, char ** argv) {
 
                 if (params.input_prefix_bos) {
                     LOG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_token_bos(model));
+                    embd_inp.push_back(llama_token_bos(ctx));
                 }
 
                 std::string buffer;
@@ -696,7 +696,7 @@ int main(int argc, char ** argv) {
                     const size_t original_size = embd_inp.size();
 
                     const auto line_inp = ::llama_tokenize(ctx, buffer, false);
-                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
 
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
 
@@ -717,14 +717,22 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    llama_sampling_reset(ctx_sampling);
+                    // reset grammar state if we're restarting generation
+                    if (grammar != NULL) {
+                        llama_grammar_free(grammar);
+
+                        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+                        grammar = llama_grammar_init(
+                            grammar_rules.data(), grammar_rules.size(),
+                            parsed_grammar.symbol_ids.at("root"));
+                    }
                 }
                 is_interacting = false;
             }
         }
 
         // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos(model) && !params.interactive) {
+        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !params.interactive) {
             break;
         }
 
@@ -736,7 +744,7 @@ int main(int argc, char ** argv) {
         }
     }
     if (!params.interactive && n_remain <= 0) {
-        printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
+        printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
         fflush(stdout);
     }
 
@@ -747,7 +755,9 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
     llama_free_model(model);
 
-    llama_sampling_free(ctx_sampling);
+    if (grammar != NULL) {
+        llama_grammar_free(grammar);
+    }
     llama_backend_free();
 
 #ifndef LOG_DISABLE_LOGS

examples/jeopardy/README.md

@@ -2,7 +2,7 @@
 
 This is pretty much just a straight port of aigoopy/llm-jeopardy/ with an added graph viewer.
 
-The jeopardy test can be used to compare the fact knowledge of different models and compare them to each other. This is in contrast to some other tests, which test logical deduction, creativity, writing skills, etc.
+The jeopardy test can be used to compare the fact knowledge of different models and compare them to eachother. This is in contrast to some other tests, which test logical deduction, creativity, writing skills, etc.
 
 
 Step 1: Open jeopardy.sh and modify the following:

examples/llama-bench/llama-bench.cpp

@@ -933,7 +933,7 @@ struct sql_printer : public printer {
 };
 
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
-    std::vector<llama_token> tokens(n_batch, llama_token_bos(llama_get_model(ctx)));
+    std::vector<llama_token> tokens(n_batch, llama_token_bos(ctx));
     int n_processed = 0;
 
     llama_set_n_threads(ctx, n_threads, n_threads);
@@ -946,7 +946,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
 }
 
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
-    llama_token token = llama_token_bos(llama_get_model(ctx));
+    llama_token token = llama_token_bos(ctx);
 
     llama_set_n_threads(ctx, n_threads, n_threads);
 
@@ -1037,7 +1037,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
 
         // warmup run
         if (t.n_prompt > 0) {
@@ -1048,7 +1048,7 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_cache_clear(ctx);
+            llama_kv_cache_tokens_rm(ctx, -1, -1);
 
             uint64_t t_start = get_time_ns();
             if (t.n_prompt > 0) {

examples/llava/CMakeLists.txt

@@ -1,20 +0,0 @@
-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)
-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(TARGET BUILD_INFO)
-    add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/llava/README.md

@@ -1,57 +0,0 @@
-# LLaVA
-
-Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants.
-
-The pre-converted [7b](https://huggingface.co/mys/ggml_llava-v1.5-7b)
-and [13b](https://huggingface.co/mys/ggml_llava-v1.5-13b)
-models are available.
-
-After API is confirmed, more models will be supported / uploaded.
-
-## Usage
-Build with cmake or run `make llava` to build it.
-
-After building, run: `./llava` 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
-```
-
-**note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
-
-## Model conversion
-
-- Clone `llava-v15-7b`` and `clip-vit-large-patch14-336`` locally:
-
-```sh
-git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
-
-git clone https://huggingface.co/openai/clip-vit-large-patch14-336
-```
-
-2. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
-
-```sh
-python ./examples/llava/llava-surgery.py -m ../llava-v1.5-7b
-```
-
-3. Use `convert-image-encoder-to-gguf.py` to convert the LLaVA image encoder to GGUF:
-
-```sh
-python ./examples/llava/convert-image-encoder-to-gguf -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
-```
-
-4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
-
-```sh
-python ./convert.py ../llava-v1.5-7b
-```
-
-Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` directory.
-
-## TODO
-
-- [ ] Support server mode.
-- [ ] Support non-CPU backend for the image encoding part.
-- [ ] Support different sampling methods.
-- [ ] Support more model variants.

examples/llava/clip.h

@@ -1,73 +0,0 @@
-#ifndef CLIP_H
-#define CLIP_H
-
-#include "ggml.h"
-
-struct clip_ctx;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-struct clip_vision_hparams {
-    int32_t image_size;
-    int32_t patch_size;
-    int32_t hidden_size;
-    int32_t n_intermediate;
-    int32_t projection_dim;
-    int32_t n_head;
-    int32_t n_layer;
-    float eps;
-};
-
-struct clip_ctx * clip_model_load(const char * fname, const int verbosity);
-
-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);
-
-// RGB uint8 image
-struct clip_image_u8 {
-    int nx;
-    int ny;
-    uint8_t * data;
-    size_t size;
-};
-
-// RGB float32 image (NHWC)
-// Memory layout: RGBRGBRGB...
-struct clip_image_f32 {
-    int nx;
-    int ny;
-    float * data;
-    size_t size;
-};
-
-struct clip_image_u8_batch {
-    struct clip_image_u8 * data;
-    size_t size;
-};
-
-struct clip_image_f32_batch {
-    struct clip_image_f32 * data;
-    size_t size;
-};
-
-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);
-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);
-
-bool clip_image_batch_encode(const struct clip_ctx * ctx, const int n_threads, const struct clip_image_f32_batch * imgs,
-                             float * vec);
-
-bool clip_model_quantize(const char * fname_inp, const char * fname_out, const int itype);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // CLIP_H

examples/llava/convert-image-encoder-to-gguf.py

@@ -1,250 +0,0 @@
-import argparse
-import os
-import json
-
-import torch
-import numpy as np
-from gguf import *
-from transformers import CLIPModel, CLIPProcessor
-
-TEXT = "clip.text"
-VISION = "clip.vision"
-
-
-def k(raw_key: str, arch: str) -> str:
-    return raw_key.format(arch=arch)
-
-
-def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: bool) -> bool:
-    if name in (
-        "logit_scale",
-        "text_model.embeddings.position_ids",
-        "vision_model.embeddings.position_ids",
-    ):
-        return True
-
-    if has_llava and name in ["visual_projection.weight", "vision_model.post_layernorm.weight", "vision_model.post_layernorm.bias"]:
-        return True
-
-    if name.startswith("v") and not has_vision:
-        return True
-
-    if name.startswith("t") and not has_text:
-        return True
-
-    return False
-
-
-def get_tensor_name(name: str) -> str:
-    if "projection" in name:
-        return name
-
-    if "mm_projector" in name:
-        return name.replace("model.mm_projector", "mm")
-
-    return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
-
-
-def bytes_to_unicode():
-    """
-    Returns list of utf-8 byte and a corresponding list of unicode strings.
-    The reversible bpe codes work on unicode strings.
-    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
-    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a signficant percentage of your normal, say, 32K bpe vocab.
-    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
-    And avoids mapping to whitespace/control characters the bpe code barfs on.
-    """
-    bs = (
-        list(range(ord("!"), ord("~") + 1))
-        + list(range(ord("¡"), ord("¬") + 1))
-        + list(range(ord("®"), ord("ÿ") + 1))
-    )
-    cs = bs[:]
-    n = 0
-    for b in range(2**8):
-        if b not in bs:
-            bs.append(b)
-            cs.append(2**8 + n)
-            n += 1
-    cs = [chr(n) for n in cs]
-    return dict(zip(bs, cs))
-
-
-ap = argparse.ArgumentParser(prog="convert_hf_to_gguf.py")
-ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
-ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
-ap.add_argument("--text-only", action="store_true", required=False,
-                help="Save a text-only model. It can't be used to encode images")
-ap.add_argument("--vision-only", action="store_true", required=False,
-                help="Save a vision-only model. It can't be used to encode texts")
-ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
-ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values")
-ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values")
-ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
-
-args = ap.parse_args()
-
-
-if args.text_only and args.vision_only:
-    print("--text-only and --image-only arguments cannot be specified at the same time.")
-    exit(1)
-
-if args.use_f32:
-    print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
-
-# output in the same directory as the model if output_dir is None
-dir_model = args.model_dir
-
-
-with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
-    vocab = json.load(f)
-    tokens = [key for key in vocab]
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
-    config = json.load(f)
-    v_hparams = config["vision_config"]
-    t_hparams = config["text_config"]
-
-# possible data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-#
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-ftype = 1
-if args.use_f32:
-    ftype = 0
-
-
-model = CLIPModel.from_pretrained(dir_model)
-processor = CLIPProcessor.from_pretrained(dir_model)
-
-fname_middle = None
-has_text_encoder = True
-has_vision_encoder = True
-has_llava_projector = False
-if args.text_only:
-    fname_middle = "text-"
-    has_vision_encoder = False
-elif args.vision_only:
-    fname_middle = "vision-"
-    has_text_encoder = False
-elif args.llava_projector is not None:
-    fname_middle = "mmproj-"
-    has_text_encoder = False
-    has_llava_projector = True
-else:
-    fname_middle = ""
-
-output_dir = args.output_dir if args.output_dir is not None else dir_model
-os.makedirs(output_dir, exist_ok=True)
-output_prefix = os.path.basename(output_dir).replace("ggml_", "")
-fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
-fout = GGUFWriter(path=fname_out, arch="clip")
-
-fout.add_bool("clip.has_text_encoder", has_text_encoder)
-fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
-fout.add_bool("clip.has_llava_projector", has_llava_projector)
-fout.add_file_type(ftype)
-model_name = config["_name_or_path"] if "_name_or_path" in config else os.path.basename(dir_model)
-fout.add_name(model_name)
-if args.text_only:
-    fout.add_description("text-only CLIP model")
-elif args.vision_only and not has_llava_projector:
-    fout.add_description("vision-only CLIP model")
-elif has_llava_projector:
-    fout.add_description("image encoder for LLaVA")
-else:
-    fout.add_description("two-tower CLIP model")
-
-if has_text_encoder:
-    # text_model hparams
-    fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"])
-    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"])
-    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, TEXT), t_hparams["intermediate_size"])
-    fout.add_uint32("clip.text.projection_dim", t_hparams.get("projection_dim", config["projection_dim"]))
-    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, TEXT), t_hparams["num_attention_heads"])
-    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, TEXT), t_hparams["layer_norm_eps"])
-    fout.add_uint32(k(KEY_BLOCK_COUNT, TEXT), t_hparams["num_hidden_layers"])
-    fout.add_token_list(tokens)
-
-if has_vision_encoder:
-    # vision_model hparams
-    fout.add_uint32("clip.vision.image_size", v_hparams["image_size"])
-    fout.add_uint32("clip.vision.patch_size", v_hparams["patch_size"])
-    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), v_hparams["hidden_size"])
-    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), v_hparams["intermediate_size"])
-    fout.add_uint32("clip.vision.projection_dim", v_hparams.get("projection_dim", config["projection_dim"]))
-    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"])
-    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
-    block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
-    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
-
-    image_mean = processor.image_processor.image_mean if args.image_mean is None else args.image_mean
-    image_std = processor.image_processor.image_std if args.image_std is None else args.image_std
-    fout.add_array("clip.vision.image_mean", image_mean)
-    fout.add_array("clip.vision.image_std", image_std)
-
-use_gelu = v_hparams["hidden_act"] == "gelu"
-fout.add_bool("clip.use_gelu", use_gelu)
-
-
-if has_llava_projector:
-    model.vision_model.encoder.layers.pop(-1)
-    projector = torch.load(args.llava_projector)
-    for name, data in projector.items():
-        name = get_tensor_name(name)
-        if data.ndim == 2:
-            data = data.squeeze().numpy().astype(np.float16)
-        else:
-            data = data.squeeze().numpy().astype(np.float32)
-
-        fout.add_tensor(name, data)
-
-    print("Projector tensors added\n")
-
-state_dict = model.state_dict()
-for name, data in state_dict.items():
-    if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_llava_projector):
-        # we don't need this
-        print(f"skipping parameter: {name}")
-        continue
-
-    name = get_tensor_name(name)
-    data = data.squeeze().numpy()
-
-    n_dims = len(data.shape)
-
-    # ftype == 0 -> float32, ftype == 1 -> float16
-    ftype_cur = 0
-    if n_dims == 4:
-        print(f"tensor {name} is always saved in f16")
-        data = data.astype(np.float16)
-        ftype_cur = 1
-    elif ftype == 1:
-        if name[-7:] == ".weight" and n_dims == 2:
-            print("  Converting to float16")
-            data = data.astype(np.float16)
-            ftype_cur = 1
-        else:
-            print("  Converting to float32")
-            data = data.astype(np.float32)
-            ftype_cur = 0
-    else:
-        if data.dtype != np.float32:
-            print("  Converting to float32")
-            data = data.astype(np.float32)
-            ftype_cur = 0
-
-    print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
-    fout.add_tensor(name, data)
-
-
-fout.write_header_to_file()
-fout.write_kv_data_to_file()
-fout.write_tensors_to_file()
-fout.close()
-
-print("Done. Output file: " + fname_out)

examples/llava/llava-surgery.py

@@ -1,46 +0,0 @@
-import argparse
-import glob
-import os
-import torch
-
-
-ap = argparse.ArgumentParser()
-ap.add_argument("-m", "--model", help="Path to LLaVA v1.5 model")
-args = ap.parse_args()
-
-# find the model part that includes the the multimodal projector weights
-path = sorted(glob.glob(f"{args.model}/pytorch_model*.bin"))[-1]
-checkpoint = torch.load(path)
-
-# get a list of mm tensor names
-mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_projector")]
-
-# store these tensors in a new dictionary and torch.save them
-projector = {name: checkpoint[name].float() for name in mm_tensors}
-torch.save(projector, f"{args.model}/llava.projector")
-
-# remove these tensors from the checkpoint and save it again
-for name in mm_tensors:
-    del checkpoint[name]
-
-# BakLLaVA models contain CLIP tensors in it
-clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
-if len(clip_tensors) > 0:
-    clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
-    torch.save(clip, f"{args.model}/llava.clip")
-
-    # remove these tensors
-    for name in clip_tensors:
-        del checkpoint[name]
-
-    # added tokens should be removed to be able to convert Mistral models
-    if os.path.exists(f"{args.model}/added_tokens.json"):
-        with open(f"{args.model}/added_tokens.json", "w") as f:
-            f.write("{}\n")
-
-
-torch.save(checkpoint, path)
-
-print("Done!")
-print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
-print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

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 +0,0 @@
-#include "clip.h"
-#include "llava-utils.h"
-#include "common.h"
-#include "llama.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");
-}
-
-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()) {
-        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;
-    }
-
-    const int64_t t_img_enc_start_us = ggml_time_us();
-    if (!clip_image_encode(ctx_clip, params.n_threads, &img_res, image_embd)) {
-        fprintf(stderr, "Unable to encode image\n");
-
-        return 1;
-    }
-    const int64_t t_img_enc_end_us = ggml_time_us();
-
-    // we get the embeddings, free up the memory required for CLIP
-    clip_free(ctx_clip);
-
-    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;
-}

examples/main-cmake-pkg/CMakeLists.txt

@@ -16,8 +16,6 @@ add_library(common OBJECT
     ${_common_path}/console.cpp
     ${_common_path}/grammar-parser.h
     ${_common_path}/grammar-parser.cpp
-    ${_common_path}/sampling.h
-    ${_common_path}/sampling.cpp
     )
 
 # WARNING: because build-info.h is auto-generated, it will only

examples/main/main.cpp

@@ -3,6 +3,7 @@
 #include "console.h"
 #include "llama.h"
 #include "build-info.h"
+#include "grammar-parser.h"
 
 #include <cassert>
 #include <cinttypes>
@@ -108,7 +109,6 @@ int main(int argc, char ** argv) {
     if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
-    llama_sampling_params & sparams = params.sparams;
 
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("main", "log"));
@@ -179,7 +179,7 @@ int main(int argc, char ** argv) {
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    if (sparams.cfg_scale > 1.f) {
+    if (params.cfg_scale > 1.f) {
         struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
         ctx_guidance = llama_new_context_with_model(model, lparams);
     }
@@ -237,19 +237,19 @@ int main(int argc, char ** argv) {
 
     if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
         LOG("tokenize the prompt\n");
-        embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+        embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
     } else {
         LOG("use session tokens\n");
         embd_inp = session_tokens;
     }
 
     LOG("prompt: \"%s\"\n", log_tostr(params.prompt));
-    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
-        embd_inp.push_back(llama_token_bos(model));
-        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+        embd_inp.push_back(llama_token_bos(ctx));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
     }
 
     // Tokenize negative prompt
@@ -257,13 +257,13 @@ int main(int argc, char ** argv) {
     int guidance_offset = 0;
     int original_prompt_len = 0;
     if (ctx_guidance) {
-        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
+        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(params.cfg_negative_prompt));
 
-        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos, true);
-        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
+        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
 
-        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
-        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
 
         original_prompt_len = original_inp.size();
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
@@ -296,9 +296,6 @@ int main(int argc, char ** argv) {
             LOG_TEE("%s: session file matches %zu / %zu tokens of prompt\n",
                 __func__, n_matching_session_tokens, embd_inp.size());
         }
-
-        // remove any "future" tokens that we might have inherited from the previous session
-        llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);
     }
 
     LOGLN(
@@ -319,11 +316,11 @@ int main(int argc, char ** argv) {
     }
 
     // prefix & suffix for instruct mode
-    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos, true);
-    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false,   true);
+    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
+    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false);
 
-    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
-    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
 
     // in instruct mode, we inject a prefix and a suffix to each input by the user
     if (params.instruct) {
@@ -346,7 +343,7 @@ int main(int argc, char ** argv) {
 
         if (ctx_guidance) {
             LOG_TEE("\n");
-            LOG_TEE("%s: negative prompt: '%s'\n", __func__, sparams.cfg_negative_prompt.c_str());
+            LOG_TEE("%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
             LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
             for (int i = 0; i < (int) guidance_inp.size(); i++) {
                 LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
@@ -382,12 +379,6 @@ int main(int argc, char ** argv) {
         if (!params.antiprompt.empty()) {
             for (const auto & antiprompt : params.antiprompt) {
                 LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str());
-                if (params.verbose_prompt) {
-                    auto tmp = ::llama_tokenize(ctx, antiprompt, false, true);
-                    for (int i = 0; i < (int) tmp.size(); i++) {
-                        LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
-                    }
-                }
             }
         }
 
@@ -397,28 +388,46 @@ int main(int argc, char ** argv) {
 
         if (!params.input_prefix.empty()) {
             LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
-            if (params.verbose_prompt) {
-                auto tmp = ::llama_tokenize(ctx, params.input_prefix, true, true);
-                for (int i = 0; i < (int) tmp.size(); i++) {
-                    LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
-                }
-            }
         }
 
         if (!params.input_suffix.empty()) {
             LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
-            if (params.verbose_prompt) {
-                auto tmp = ::llama_tokenize(ctx, params.input_suffix, false, true);
-                for (int i = 0; i < (int) tmp.size(); i++) {
-                    LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
-                }
-            }
         }
     }
-    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
+    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
+            params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau);
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     LOG_TEE("\n\n");
 
+    struct llama_grammar * grammar = NULL;
+    grammar_parser::parse_state parsed_grammar;
+
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
+        LOG_TEE("%s: grammar:\n", __func__);
+        grammar_parser::print_grammar(stderr, parsed_grammar);
+        LOG_TEE("\n");
+
+        {
+            auto it = params.logit_bias.find(llama_token_eos(ctx));
+            if (it != params.logit_bias.end() && it->second == -INFINITY) {
+                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
+            }
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar = llama_grammar_init(
+            grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    }
+
+    // TODO: replace with ring-buffer
+    std::vector<llama_token> last_tokens(n_ctx);
+    std::fill(last_tokens.begin(), last_tokens.end(), 0);
+
     if (params.interactive) {
         const char *control_message;
         if (params.multiline_input) {
@@ -458,7 +467,10 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
+    const int n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
 
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
@@ -505,7 +517,7 @@ int main(int argc, char ** argv) {
 
                 LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
 
-                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
 
                 LOG("clear session path\n");
                 path_session.clear();
@@ -531,10 +543,14 @@ int main(int argc, char ** argv) {
                 if (i > 0) {
                     embd.erase(embd.begin(), embd.begin() + i);
                 }
+
+                // remove any "future" tokens that we might have inherited from the session from the KV cache
+                llama_kv_cache_tokens_rm(ctx, n_past, -1);
             }
 
             // evaluate tokens in batches
             // embd is typically prepared beforehand to fit within a batch, but not always
+
             if (ctx_guidance) {
                 int input_size = 0;
                 llama_token * input_buf = NULL;
@@ -556,7 +572,7 @@ int main(int argc, char ** argv) {
                     input_buf  = embd_guidance.data();
                     input_size = embd_guidance.size();
 
-                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance).c_str());
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
                 } else {
                     input_buf  = embd.data();
                     input_size = embd.size();
@@ -579,7 +595,7 @@ int main(int argc, char ** argv) {
                     n_eval = params.n_batch;
                 }
 
-                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
 
                 if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
                     LOG_TEE("%s : failed to eval\n", __func__);
@@ -609,11 +625,12 @@ int main(int argc, char ** argv) {
                 LOG("saved session to %s\n", path_session.c_str());
             }
 
-            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
+            const llama_token id = llama_sample_token(ctx, ctx_guidance, grammar, params, last_tokens, candidates);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            last_tokens.erase(last_tokens.begin());
+            last_tokens.push_back(id);
 
-            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
 
             embd.push_back(id);
 
@@ -629,11 +646,8 @@ int main(int argc, char ** argv) {
             LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
             while ((int) embd_inp.size() > n_consumed) {
                 embd.push_back(embd_inp[n_consumed]);
-
-                // push the prompt in the sampling context in order to apply repetition penalties later
-                // for the prompt, we don't apply grammar rules
-                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
-
+                last_tokens.erase(last_tokens.begin());
+                last_tokens.push_back(embd_inp[n_consumed]);
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
                     break;
@@ -663,10 +677,12 @@ int main(int argc, char ** argv) {
 
         // if not currently processing queued inputs;
         if ((int) embd_inp.size() <= n_consumed) {
-            // check for reverse prompt in the last n_prev tokens
+            // check for reverse prompt
             if (!params.antiprompt.empty()) {
-                const int n_prev = 32;
-                const std::string last_output = llama_sampling_prev_str(ctx_sampling, ctx, n_prev);
+                std::string last_output;
+                for (auto id : last_tokens) {
+                    last_output += llama_token_to_piece(ctx, id);
+                }
 
                 is_antiprompt = false;
                 // Check if each of the reverse prompts appears at the end of the output.
@@ -693,13 +709,13 @@ int main(int argc, char ** argv) {
             }
 
             // deal with end of text token in interactive mode
-            if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
+            if (last_tokens.back() == llama_token_eos(ctx)) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
                     if (!params.antiprompt.empty()) {
                         // tokenize and inject first reverse prompt
-                        const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false, true);
+                        const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false);
                         embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
                         is_antiprompt = true;
                     }
@@ -720,13 +736,14 @@ int main(int argc, char ** argv) {
 
                 if (params.input_prefix_bos) {
                     LOG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_token_bos(model));
+                    embd_inp.push_back(llama_token_bos(ctx));
                 }
 
                 std::string buffer;
                 if (!params.input_prefix.empty()) {
                     LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
-                    printf("%s", params.input_prefix.c_str());
+                    buffer += params.input_prefix;
+                    printf("%s", buffer.c_str());
                 }
 
                 // color user input only
@@ -748,6 +765,7 @@ int main(int argc, char ** argv) {
                     // append input suffix if any
                     if (!params.input_suffix.empty()) {
                         LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
+                        buffer += params.input_suffix;
                         printf("%s", params.input_suffix.c_str());
                     }
 
@@ -761,18 +779,11 @@ int main(int argc, char ** argv) {
                         n_consumed = embd_inp.size();
                         embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
                     }
-                    if (params.escape) {
-                        process_escapes(buffer);
-                    }
 
-                    const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
-                    const auto line_inp = ::llama_tokenize(ctx, buffer,              false, false);
-                    const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
-                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
+                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
 
-                    embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end());
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
-                    embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end());
 
                     // instruct mode: insert response suffix
                     if (params.instruct) {
@@ -797,14 +808,22 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    llama_sampling_reset(ctx_sampling);
+                    // reset grammar state if we're restarting generation
+                    if (grammar != NULL) {
+                        llama_grammar_free(grammar);
+
+                        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+                        grammar = llama_grammar_init(
+                            grammar_rules.data(), grammar_rules.size(),
+                            parsed_grammar.symbol_ids.at("root"));
+                    }
                 }
                 is_interacting = false;
             }
         }
 
         // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive)) {
+        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !(params.instruct || params.interactive)) {
             LOG_TEE(" [end of text]\n");
             break;
         }
@@ -829,7 +848,9 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
     llama_free_model(model);
 
-    llama_sampling_free(ctx_sampling);
+    if (grammar != NULL) {
+        llama_grammar_free(grammar);
+    }
     llama_backend_free();
 
 #ifndef LOG_DISABLE_LOGS

examples/parallel/parallel.cpp

@@ -10,7 +10,6 @@
 #include <cstdio>
 #include <string>
 #include <vector>
-#include <ctime>
 
 // trim whitespace from the beginning and end of a string
 static std::string trim(const std::string & str) {
@@ -51,12 +50,6 @@ static std::vector<std::string> k_prompts = {
 };
 
 struct client {
-    ~client() {
-        if (ctx_sampling) {
-            llama_sampling_free(ctx_sampling);
-        }
-    }
-
     int32_t id = 0;
 
     llama_seq_id seq_id = -1;
@@ -74,29 +67,9 @@ struct client {
     std::string prompt;
     std::string response;
 
-    struct llama_sampling_context * ctx_sampling = nullptr;
+    std::vector<llama_token> tokens_prev;
 };
 
-static void print_date_time() {
-    std::time_t current_time = std::time(nullptr);
-    std::tm* local_time = std::localtime(&current_time);
-    char buffer[80];
-    strftime(buffer, sizeof(buffer), "%Y-%m-%d %H:%M:%S", local_time);
-
-    printf("\n\033[35mrun parameters as at %s\033[0m\n", buffer);
-}
-
-// Define a split string function to ...
-static std::vector<std::string> split_string(const std::string& input, char delimiter) {
-    std::vector<std::string> tokens;
-    std::istringstream stream(input);
-    std::string token;
-    while (std::getline(stream, token, delimiter)) {
-        tokens.push_back(token);
-    }
-    return tokens;
-}
-
 int main(int argc, char ** argv) {
     srand(1234);
 
@@ -131,35 +104,23 @@ int main(int argc, char ** argv) {
     params.logits_all = true;
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
-    // load the prompts from an external file if there are any
-    if (params.prompt.empty()) {
-        printf("\n\033[32mNo new questions so proceed with build-in defaults.\033[0m\n");
-    } else {
-        // Output each line of the input params.prompts vector and copy to k_prompts
-        int index = 0;
-        printf("\n\033[32mNow printing the external prompt file %s\033[0m\n\n", params.prompt_file.c_str());
-
-        std::vector<std::string> prompts = split_string(params.prompt, '\n');
-        for (const auto& prompt : prompts) {
-            k_prompts.resize(index + 1);
-            k_prompts[index] = prompt;
-            index++;
-            printf("%3d prompt: %s\n", index, prompt.c_str());
-        }
-    }
-
     fprintf(stderr, "\n\n");
     fflush(stderr);
 
-    const int n_ctx = llama_n_ctx(ctx);
+    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_vocab = llama_n_vocab(model);
 
     std::vector<client> clients(n_clients);
     for (size_t i = 0; i < clients.size(); ++i) {
         auto & client = clients[i];
         client.id = i;
-        client.ctx_sampling = llama_sampling_init(params.sparams);
+        client.tokens_prev.resize(std::max(256, params.n_predict));
+        std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
     }
 
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
     std::vector<llama_token> tokens_system;
     tokens_system = ::llama_tokenize(ctx, k_system, true);
     const int32_t n_tokens_system = tokens_system.size();
@@ -168,7 +129,7 @@ int main(int argc, char ** argv) {
 
     // the max batch size is as large as the context to handle cases where we get very long input prompt from multiple
     // users. regardless of the size, the main loop will chunk the batch into a maximum of params.n_batch tokens at a time
-    llama_batch batch = llama_batch_init(n_ctx, 0, 1);
+    llama_batch batch = llama_batch_init(params.n_ctx, 0);
 
     int32_t n_total_prompt = 0;
     int32_t n_total_gen    = 0;
@@ -183,8 +144,13 @@ int main(int argc, char ** argv) {
     {
         LOG_TEE("%s: Evaluating the system prompt ...\n", __func__);
 
-        for (int32_t i = 0; i < n_tokens_system; ++i) {
-            llama_batch_add(batch, tokens_system[i], i, { 0 }, false);
+        batch.n_tokens = n_tokens_system;
+
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            batch.token[i]  = tokens_system[i];
+            batch.pos[i]    = i;
+            batch.seq_id[i] = 0;
+            batch.logits[i] = false;
         }
 
         if (llama_decode(ctx, batch) != 0) {
@@ -203,7 +169,7 @@ int main(int argc, char ** argv) {
     LOG_TEE("Processing requests ...\n\n");
 
     while (true) {
-        llama_batch_clear(batch);
+        batch.n_tokens = 0;
 
         // decode any currently ongoing sequences
         for (auto & client : clients) {
@@ -211,11 +177,15 @@ int main(int argc, char ** argv) {
                 continue;
             }
 
-            client.i_batch = batch.n_tokens;
-
-            llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id }, true);
+            batch.token [batch.n_tokens] = client.sampled;
+            batch.pos   [batch.n_tokens] = n_tokens_system + client.n_prompt + client.n_decoded;
+            batch.seq_id[batch.n_tokens] = client.id;
+            batch.logits[batch.n_tokens] = true;
 
             client.n_decoded += 1;
+            client.i_batch = batch.n_tokens;
+
+            batch.n_tokens += 1;
         }
 
         if (batch.n_tokens == 0) {
@@ -240,14 +210,18 @@ int main(int argc, char ** argv) {
                     client.prompt   = client.input + "\nAssistant:";
                     client.response = "";
 
-                    llama_sampling_reset(client.ctx_sampling);
+                    std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
 
                     // do not prepend BOS because we have a system prompt!
                     std::vector<llama_token> tokens_prompt;
                     tokens_prompt = ::llama_tokenize(ctx, client.prompt, false);
 
                     for (size_t i = 0; i < tokens_prompt.size(); ++i) {
-                        llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id }, false);
+                        batch.token [batch.n_tokens] = tokens_prompt[i];
+                        batch.pos   [batch.n_tokens] = i + n_tokens_system;
+                        batch.seq_id[batch.n_tokens] = client.id;
+                        batch.logits[batch.n_tokens] = false;
+                        batch.n_tokens += 1;
                     }
 
                     // extract the logits only for the last token
@@ -259,7 +233,7 @@ int main(int argc, char ** argv) {
                     client.n_decoded = 0;
                     client.i_batch   = batch.n_tokens - 1;
 
-                    LOG_TEE("\033[31mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
+                    LOG_TEE("\033[1mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
 
                     g_seq_id += 1;
 
@@ -290,12 +264,11 @@ int main(int argc, char ** argv) {
 
             llama_batch batch_view = {
                 n_tokens,
-                batch.token    + i,
+                batch.token  + i,
                 nullptr,
-                batch.pos      + i,
-                batch.n_seq_id + i,
-                batch.seq_id   + i,
-                batch.logits   + i,
+                batch.pos    + i,
+                batch.seq_id + i,
+                batch.logits + i,
                 0, 0, 0, // unused
             };
 
@@ -328,9 +301,7 @@ int main(int argc, char ** argv) {
                 //printf("client %d, seq %d, token %d, pos %d, batch %d\n",
                 //        client.id, client.seq_id, client.sampled, client.n_decoded, client.i_batch);
 
-                const llama_token id = llama_sampling_sample(client.ctx_sampling, ctx, NULL, client.i_batch - i);
-
-                llama_sampling_accept(client.ctx_sampling, ctx, id, true);
+                const llama_token id = llama_sample_token(ctx, NULL, NULL, params, client.tokens_prev, candidates, client.i_batch - i);
 
                 if (client.n_decoded == 1) {
                     // start measuring generation time after the first token to make sure all concurrent clients
@@ -338,8 +309,11 @@ int main(int argc, char ** argv) {
                     client.t_start_gen = ggml_time_us();
                 }
 
-                const std::string token_str = llama_token_to_piece(ctx, id);
+                // remember which tokens were sampled - used for repetition penalties during sampling
+                client.tokens_prev.erase(client.tokens_prev.begin());
+                client.tokens_prev.push_back(id);
 
+                const std::string token_str = llama_token_to_piece(ctx, id);
                 client.response += token_str;
                 client.sampled = id;
 
@@ -347,7 +321,7 @@ int main(int argc, char ** argv) {
                 //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
 
                 if (client.n_decoded > 2 &&
-                        (id == llama_token_eos(model) ||
+                        (id == llama_token_eos(ctx) ||
                          (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
                          client.response.find("User:") != std::string::npos ||
                          client.response.find('\n') != std::string::npos)) {
@@ -362,8 +336,8 @@ int main(int argc, char ** argv) {
 
                     const auto t_main_end = ggml_time_us();
 
-                    LOG_TEE("\033[31mClient %3d, seq %3d/%3d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \nInput:    %s\n\033[35mResponse: %s\033[0m\n\n",
-                            client.id, client.seq_id, n_seq, client.n_prompt, client.n_decoded,
+                    LOG_TEE("\033[1mClient %3d, seq %4d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \n\nInput:    %s\nResponse: %s\n\n",
+                            client.id, client.seq_id, client.n_prompt, client.n_decoded,
                             (t_main_end - client.t_start_prompt) / 1e6,
                             (double) (client.n_prompt + client.n_decoded) / (t_main_end - client.t_start_prompt) * 1e6,
                             n_cache_miss,
@@ -383,21 +357,13 @@ int main(int argc, char ** argv) {
 
     const auto t_main_end = ggml_time_us();
 
-    print_date_time();
-
-    LOG_TEE("\n%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
-    if (params.prompt_file.empty()) {
-        params.prompt_file = "used built-in defaults";
-    }
-    LOG_TEE("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
-    LOG_TEE("Model and path used:  \033[32m%s\033[0m\n\n", params.model.c_str());
-
+    LOG_TEE("\n\n");
     LOG_TEE("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt              ) / (t_main_end - t_main_start) * 1e6);
     LOG_TEE("Total gen tokens:    %6d, speed: %5.2f t/s\n", n_total_gen,    (double) (n_total_gen                 ) / (t_main_end - t_main_start) * 1e6);
     LOG_TEE("Total speed (AVG):   %6s  speed: %5.2f t/s\n", "",             (double) (n_total_prompt + n_total_gen) / (t_main_end - t_main_start) * 1e6);
     LOG_TEE("Cache misses:        %6d\n", n_cache_miss);
 
-    LOG_TEE("\n");
+    LOG_TEE("\n\n");
 
     llama_print_timings(ctx);
 

examples/perplexity/perplexity.cpp

@@ -210,7 +210,7 @@ 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_clear(ctx);
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -227,7 +227,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
 
             // add BOS token for the first batch of each chunk
             if (add_bos && j == 0) {
-                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
+                tokens[batch_start] = llama_token_bos(ctx);
             }
 
             const auto batch_logits = llama_get_logits(ctx);
@@ -339,7 +339,7 @@ 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_clear(ctx);
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -350,7 +350,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 
             // add BOS token for the first batch of each chunk
             if (add_bos && j == 0) {
-                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
+                tokens[batch_start] = llama_token_bos(ctx);
             }
 
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
@@ -573,7 +573,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         }
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
 
         auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab);
         if (logits.empty()) {

examples/quantize/quantize.cpp

@@ -18,6 +18,7 @@ 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", },
@@ -30,6 +31,7 @@ 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", },
@@ -68,14 +70,13 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 }
 
 // usage:
-//  ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
+//  ./quantize [--allow-requantize] [--leave-output-tensor] 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] [--pure] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] 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") {
@@ -102,8 +103,6 @@ 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/save-load-state.cpp

@@ -8,7 +8,9 @@
 
 int main(int argc, char ** argv) {
     gpt_params params;
-
+    params.seed = 42;
+    params.n_threads = 4;
+    params.repeat_last_n = 64;
     params.prompt = "The quick brown fox";
 
     if (!gpt_params_parse(argc, argv, params)) {
@@ -22,49 +24,56 @@ int main(int argc, char ** argv) {
     }
 
     auto n_past = 0;
-
-    std::string result0;
-    std::string result1;
+    auto last_n_tokens_data = std::vector<llama_token>(params.repeat_last_n, 0);
 
     // init
     llama_model * model;
     llama_context * ctx;
 
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    if (model == nullptr || ctx == nullptr) {
-        fprintf(stderr, "%s : failed to init\n", __func__);
+    std::tie(model, ctx) = llama_init_from_gpt_params( params );
+    if (model == nullptr) {
+        return 1;
+    }
+    if (ctx == nullptr) {
+        llama_free_model(model);
+        return 1;
+    }
+    auto tokens = llama_tokenize(ctx, params.prompt, true);
+    auto n_prompt_tokens = tokens.size();
+    if (n_prompt_tokens < 1) {
+        fprintf(stderr, "%s : failed to tokenize prompt\n", __func__);
+        llama_free(ctx);
+        llama_free_model(model);
         return 1;
     }
 
-    // tokenize prompt
-    auto tokens = llama_tokenize(ctx, params.prompt, true);
-
     // evaluate prompt
-    llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), n_past, 0));
-    n_past += tokens.size();
+    llama_decode(ctx, llama_batch_get_one(tokens.data(), n_prompt_tokens, n_past, 0));
 
-    // save state (rng, logits, embedding and kv_cache) to file
+    last_n_tokens_data.insert(last_n_tokens_data.end(), tokens.data(), tokens.data() + n_prompt_tokens);
+    n_past += n_prompt_tokens;
+
+    const size_t state_size = llama_get_state_size(ctx);
+    uint8_t * state_mem = new uint8_t[state_size];
+
+    // Save state (rng, logits, embedding and kv_cache) to file
     {
-        std::vector<uint8_t> state_mem(llama_get_state_size(ctx));
-
-        {
-            FILE *fp_write = fopen("dump_state.bin", "wb");
-            llama_copy_state_data(ctx, state_mem.data()); // could also copy directly to memory mapped file
-            fwrite(state_mem.data(), 1, state_mem.size(), fp_write);
-            fclose(fp_write);
-        }
+        FILE *fp_write = fopen("dump_state.bin", "wb");
+        llama_copy_state_data(ctx, state_mem); // could also copy directly to memory mapped file
+        fwrite(state_mem, 1, state_size, fp_write);
+        fclose(fp_write);
     }
 
     // save state (last tokens)
+    const auto last_n_tokens_data_saved = std::vector<llama_token>(last_n_tokens_data);
     const auto n_past_saved = n_past;
 
     // first run
-    printf("\nfirst run: %s", params.prompt.c_str());
+    printf("\n%s", params.prompt.c_str());
 
     for (auto i = 0; i < params.n_predict; i++) {
         auto * logits = llama_get_logits(ctx);
         auto n_vocab = llama_n_vocab(model);
-
         std::vector<llama_token_data> candidates;
         candidates.reserve(n_vocab);
         for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
@@ -73,10 +82,9 @@ int main(int argc, char ** argv) {
         llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
         auto next_token = llama_sample_token(ctx, &candidates_p);
         auto next_token_str = llama_token_to_piece(ctx, next_token);
+        last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());
-        result0 += next_token_str;
-
         if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_free(ctx);
@@ -94,28 +102,32 @@ int main(int argc, char ** argv) {
     // make new context
     auto * ctx2 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
 
-    printf("\nsecond run: %s", params.prompt.c_str());
-
-    // load state (rng, logits, embedding and kv_cache) from file
+    // Load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector<uint8_t> state_mem(llama_get_state_size(ctx2));
+        FILE *fp_read = fopen("dump_state.bin", "rb");
+        if (state_size != llama_get_state_size(ctx2)) {
+            fprintf(stderr, "\n%s : failed to validate state size\n", __func__);
+            llama_free(ctx2);
+            llama_free_model(model);
+            return 1;
+        }
 
-        FILE * fp_read = fopen("dump_state.bin", "rb");
-
-        const size_t ret = fread(state_mem.data(), 1, state_mem.size(), fp_read);
-        if (ret != state_mem.size()) {
+        const size_t ret = fread(state_mem, 1, state_size, fp_read);
+        if (ret != state_size) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
             return 1;
         }
 
-        llama_set_state_data(ctx2, state_mem.data());
-
+        llama_set_state_data(ctx2, state_mem);  // could also read directly from memory mapped file
         fclose(fp_read);
     }
 
+    delete[] state_mem;
+
     // restore state (last tokens)
+    last_n_tokens_data = last_n_tokens_data_saved;
     n_past = n_past_saved;
 
     // second run
@@ -130,11 +142,10 @@ int main(int argc, char ** argv) {
         llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
         auto next_token = llama_sample_token(ctx2, &candidates_p);
         auto next_token_str = llama_token_to_piece(ctx2, next_token);
+        last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());
-        result1 += next_token_str;
-
-        if (llama_decode(ctx2, llama_batch_get_one(&next_token, 1, n_past, 0))) {
+        if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
@@ -143,17 +154,10 @@ int main(int argc, char ** argv) {
         n_past += 1;
     }
 
-    printf("\n");
+    printf("\n\n");
 
     llama_free(ctx2);
     llama_free_model(model);
 
-    if (result0 != result1) {
-        fprintf(stderr, "\n%s : error : the 2 generations are different\n", __func__);
-        return 1;
-    }
-
-    fprintf(stderr, "\n%s : success\n", __func__);
-
     return 0;
 }

examples/server-parallel/CMakeLists.txt

@@ -0,0 +1,15 @@
+set(TARGET server-parallel)
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+add_executable(${TARGET} server.cpp ../server/json.hpp ../server/httplib.h)
+install(TARGETS ${TARGET} RUNTIME)
+target_compile_definitions(${TARGET} PRIVATE
+    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
+)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+if (WIN32)
+    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
+endif()
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/server-parallel/README.md

@@ -0,0 +1,66 @@
+# llama.cpp/example/server-parallel
+
+This example demonstrates a PoC HTTP API server that handles simulataneus requests. Long prompts are not supported.
+
+## Quick Start
+
+To get started right away, run the following command, making sure to use the correct path for the model you have:
+
+### Unix-based systems (Linux, macOS, etc.):
+
+```bash
+./server-parallel -m models/7B/ggml-model.gguf --ctx_size 2048 -t 4 -ngl 33 --batch-size 512 --parallel 3 -n 512 --cont-batching
+```
+
+### Windows:
+
+```powershell
+server-parallel.exe -m models\7B\ggml-model.gguf --ctx_size 2048 -t 4 -ngl 33 --batch-size 512 --parallel 3 -n 512 --cont-batching
+```
+The above command will start a server that by default listens on `127.0.0.1:8080`.
+
+## API Endpoints
+
+-   **GET** `/props`: Return the user and assistant name for generate the prompt.
+
+*Response:*
+```json
+{
+    "user_name": "User:",
+    "assistant_name": "Assistant:"
+}
+```
+
+-   **POST** `/completion`: Given a prompt, it returns the predicted completion, just streaming mode.
+
+    *Options:*
+
+    `temperature`: Adjust the randomness of the generated text (default: 0.1).
+
+    `prompt`: Provide a prompt as a string, It should be a coherent continuation of the system prompt.
+
+    `system_prompt`: Provide a system prompt as a string.
+
+    `anti_prompt`: Provide the name of the user coherent with the system prompt.
+
+    `assistant_name`: Provide the name of the assistant coherent with the system prompt.
+
+*Example request:*
+```json
+{
+    "system_prompt": "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\n\nHuman: Hello\nAssistant: Hi, how may I help you?\nHuman:",
+    "anti_prompt": "Human:",
+    "assistant_name": "Assistant:",
+    "prompt": "When is the day of independency of US?",
+    "temperature": 0.2
+}
+```
+
+*Response:*
+```json
+{
+    "content": "<token_str>"
+}
+```
+
+# This example is a Proof of Concept, have some bugs and unexpected behaivors, this not supports long prompts.

examples/server-parallel/frontend.h

@@ -0,0 +1,263 @@
+const char* system_prompt_default =
+R"(Transcript of a never ending dialog, where the User interacts with an Assistant.
+The Assistant is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.
+User: Recommend a nice restaurant in the area.
+Assistant: I recommend the restaurant "The Golden Duck". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.
+User: Who is Richard Feynman?
+Assistant: Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including "Surely You're Joking, Mr. Feynman!" and "What Do You Care What Other People Think?".
+User:)";
+
+const char* index_html_ = R"(
+<!DOCTYPE html>
+<html>
+<head>
+  <title>llama.cpp - server parallel PoC</title>
+  <script src="index.js"></script>
+</head>
+<body>
+  <div style="width: 90%;margin: auto;">
+    <h2>Server parallel - PoC</h2>
+    <form id="myForm">
+      <input type="checkbox" id="system_promt_cb" name="myCheckbox" onchange="toggleSP() ">
+      <label for="system_promt_cb">Use custom system prompt</label>
+      <br>
+      <div id="system_prompt_view" style="display: none;">
+        <textarea id="sp_text" name="systemPrompt" style="width: 100%;height: 4rem;" placeholder="System Prompt"></textarea>
+        <label for="user_name">User name</label>
+        <input type="text" id="user_name" value=""  placeholder="Anti prompt" required>
+        <label for="assistant_name">Assistant name</label>
+        <input type="text" id="assistant_name" value="" placeholder="Assistant:" required>
+        <button type="button" id="btn_reset" onclick="clearSP() " >Clear all</button>
+      </div>
+      <br>
+      <label for="slot_id">Slot ID (-1 load in a idle slot)</label>
+      <input type="number" id="slot_id" value="-1" required>
+      <br>
+      <label for="temperature">Temperature</label>
+      <input type="number" id="temperature" value="0.1" required>
+      <br>
+      <label for="message">Message</label>
+      <input id="message" style="width: 80%;" required>
+      <br><br>
+      <button type="button" id="btn_send" onclick="perform() " >Send</button>
+      <br>
+      <br>
+      <button type="button" id="btn_reset" onclick="resetBtn() " >Reset</button>
+    </form>
+    <div id="conversation_view">
+    </div>
+  </div>
+</body>
+</html>
+)";
+
+const char* index_js_ = R"(
+ let conversation = [];
+ let current_message = -1;
+
+const questions = [
+  "Who is Elon Musk?",
+  "Who is Jeff Bezos?",
+  "How to get a job at google?",
+  "What are you?",
+  "When was born Abraham Lincoln?",
+];
+
+let user_name = "";
+let assistant_name = "";
+
+function toggleSP() {
+    if(document.getElementById("system_promt_cb").checked) {
+        document.getElementById("system_prompt_view").style.display = "block";
+    } else {
+        document.getElementById("system_prompt_view").style.display = "none";
+    }
+}
+
+function clearSP() {
+    document.getElementById("sp_text").value = "";
+    document.getElementById("anti_prompt").value = "";
+    document.getElementById("assistant_name").value = "";
+}
+
+docReady(async () => {
+  document.getElementById("message").value =
+    questions[Math.floor(Math.random() * questions.length)];
+
+    // to keep the same prompt format in all clients
+    const response = await fetch("/props");
+    if (!response.ok) {
+      alert(`HTTP error! Status: ${response.status}`);
+    }
+    const data = await response.json();
+    user_name = data.user_name;
+    assistant_name = data.assistant_name;
+});
+
+function docReady(fn) {
+  // see if DOM is already available
+  if (
+    document.readyState === "complete" ||
+    document.readyState === "interactive"
+  ) {
+    // call on next available tick
+    setTimeout(fn, 1);
+  } else {
+    document.addEventListener("DOMContentLoaded", fn);
+  }
+}
+
+function updateView() {
+  let conv_view = document.getElementById("conversation_view");
+  // build view
+  conv_view.innerHTML = "";
+  for (let index in conversation) {
+    conversation[index].assistant = conversation[index].assistant.replace(
+        user_name,
+      ""
+    );
+    conv_view.innerHTML += `
+          <p><span style="font-weight: bold">User:</span> ${conversation[index].user}<p>
+          <p style="white-space: pre-line;"><span style="font-weight: bold">Assistant:</span> ${conversation[index].assistant}<p>`;
+  }
+}
+
+async function call_llama(options) {
+  const response = await fetch("/completion", {
+    method: "POST",
+    body: JSON.stringify(options),
+    headers: {
+      Connection: "keep-alive",
+      "Content-Type": "application/json",
+      Accept: "text/event-stream",
+    },
+  });
+
+  const reader = response.body.getReader();
+  let cont = true;
+  const decoder = new TextDecoder();
+  let leftover = ""; // Buffer for partially read lines
+
+  try {
+    let cont = true;
+
+    while (cont) {
+      const result = await reader.read();
+      if (result.done) {
+        document.getElementById("btn_send").disabled = false;
+        break;
+      }
+
+      // Add any leftover data to the current chunk of data
+      const text = leftover + decoder.decode(result.value);
+
+      // Check if the last character is a line break
+      const endsWithLineBreak = text.endsWith("\n");
+
+      // Split the text into lines
+      let lines = text.split("\n");
+
+      // If the text doesn't end with a line break, then the last line is incomplete
+      // Store it in leftover to be added to the next chunk of data
+      if (!endsWithLineBreak) {
+        leftover = lines.pop();
+      } else {
+        leftover = ""; // Reset leftover if we have a line break at the end
+      }
+
+      // Parse all sse events and add them to result
+      const regex = /^(\S+):\s(.*)$/gm;
+      for (const line of lines) {
+        const match = regex.exec(line);
+        if (match) {
+          result[match[1]] = match[2];
+          // since we know this is llama.cpp, let's just decode the json in data
+          if (result.data) {
+            result.data = JSON.parse(result.data);
+            conversation[current_message].assistant += result.data.content;
+            updateView();
+          }
+        }
+      }
+    }
+  } catch (e) {
+    if (e.name !== "AbortError") {
+      console.error("llama error: ", e);
+    }
+    throw e;
+  }
+}
+
+function generatePrompt() {
+  // generate a good prompt to have coherence
+  let prompt = "";
+  for (let index in conversation) {
+    if (index == 0) {
+      prompt += conversation[index].user + "\n";
+    } else {
+      prompt += user_name + conversation[index].user + "\n";
+    }
+    if (index == current_message) {
+      prompt += assistant_name;
+    } else {
+      prompt += assistant_name + conversation[index].assistant;
+    }
+  }
+  return prompt;
+}
+
+function resetBtn() {
+  document.getElementById("slot_id").value = "-1";
+  document.getElementById("temperature").value = "0.1";
+  document.getElementById("message").value =
+    questions[Math.floor(Math.random() * questions.length)];
+  document.getElementById("conversation_view").innerHTML = "";
+  conversation = [];
+  current_message = -1;
+}
+
+async function perform() {
+  var slot_id = parseInt(document.getElementById("slot_id").value);
+  var temperature = parseFloat(document.getElementById("temperature").value);
+  var prompt = " " + document.getElementById("message").value;
+  if (!isNaN(slot_id) && !isNaN(temperature) && prompt.length > 0) {
+    let options = {
+        slot_id,
+        temperature
+    };
+    if(document.getElementById("system_promt_cb").checked) {
+        let system_prompt = document.getElementById("sp_text").value;
+        let anti_prompt = document.getElementById("user_name").value;
+        let assistant_name_ = document.getElementById("assistant_name").value;
+        if(!system_prompt || !anti_prompt || !assistant_name_) {
+          document.getElementById("conversation_view").innerText =
+                "please, insert valid props.";
+          return;
+        }
+        conversation = [];
+        current_message = -1;
+        document.getElementById("system_promt_cb").checked = false;
+        document.getElementById("system_promt_cb").dispatchEvent(new Event("change"));
+        options.system_prompt = system_prompt;
+        options.anti_prompt = anti_prompt;
+        options.assistant_name = assistant_name_;
+        user_name = anti_prompt;
+        assistant_name = assistant_name_;
+    }
+    current_message++;
+    conversation.push({
+      user: prompt,
+      assistant: "",
+    });
+    updateView();
+    document.getElementById("message").value = "";
+    document.getElementById("btn_send").disabled = true;
+    options.prompt = generatePrompt();
+    await call_llama(options);
+  } else {
+    document.getElementById("conversation_view").innerText =
+      "please, insert valid props.";
+  }
+}
+
+)";

examples/server-parallel/server.cpp

@@ -0,0 +1,884 @@
+#include "frontend.h"
+#include "common.h"
+#include "llama.h"
+
+#include "../server/httplib.h"
+#include "../server/json.hpp"
+
+#include <iostream>
+#include <sstream>
+#include <thread>
+#include <vector>
+#include <chrono>
+
+using namespace httplib;
+using namespace std;
+using namespace nlohmann;
+
+struct server_params
+{
+    std::string hostname = "127.0.0.1";
+    std::string public_path = "examples/server/public";
+    int32_t port = 8080;
+    int32_t read_timeout = 600;
+    int32_t write_timeout = 600;
+};
+
+// utils functions taken of examples/server
+
+static bool ends_with(const std::string &str, const std::string &suffix)
+{
+    return str.size() >= suffix.size() &&
+           0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
+}
+
+static size_t find_partial_stop_string(const std::string &stop,
+                                       const std::string &text)
+{
+    if (!text.empty() && !stop.empty())
+    {
+        const char text_last_char = text.back();
+        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--)
+        {
+            if (stop[char_index] == text_last_char)
+            {
+                const std::string current_partial = stop.substr(0, char_index + 1);
+                if (ends_with(text, current_partial))
+                {
+                    return text.size() - char_index - 1;
+                }
+            }
+        }
+    }
+    return std::string::npos;
+}
+
+enum stop_type
+{
+    STOP_FULL,
+    STOP_PARTIAL,
+};
+
+enum slot_state
+{
+    IDLE,
+    PROCESSING
+};
+
+enum slot_command {
+    NONE,
+    LOAD_PROMPT,
+    RELEASE
+};
+
+
+struct llama_client_slot
+{
+    int id;
+    int32_t n_prompt  = 0;
+    int32_t n_decoded = 0;
+    int32_t i_batch   = -1;
+    string prompt = "";
+    string sampled_token_str;
+    string generated_text = "";
+    llama_token sampled;
+    std::vector<llama_token> tokens_prev;
+    slot_state state = IDLE;
+    slot_command command = NONE;
+    bool newToken = false;
+    float temperature = 0.1f;
+
+    void start(string prompt_, float temp_) {
+        prompt = prompt_;
+        command = LOAD_PROMPT;
+        temperature = temp_;
+        newToken = false;
+    }
+
+    bool hasNewToken() {
+        if(newToken) {
+            newToken = false;
+            return true;
+        }
+        return false;
+    }
+
+    bool available() {
+        return state == IDLE && command == NONE;
+    }
+
+    void nofity() {
+        newToken = !newToken;
+    }
+
+    void release() {
+        if(state == PROCESSING) {
+            command = RELEASE;
+        }
+    }
+};
+
+struct server_parallel_context {
+    // example props
+    vector<llama_client_slot> slots;
+    std::string system_prompt = "";
+    bool update_system_prompt = true;
+
+    // broadcast to all clients to keep the same prompt format
+    std::string user_name = ""; // this should be the anti prompt
+    std::string assistant_name = ""; // this is for generate the prompt
+
+    // llama native props
+    gpt_params params;
+    llama_model *model = NULL;
+    llama_context *ctx = NULL;
+    int n_ctx;
+    int n_vocab;
+    std::vector<llama_token_data> candidates;
+    std::vector<llama_token> tokens_system;
+    int32_t n_tokens_system = 0;
+    llama_batch batch;
+
+    bool loadModel(gpt_params params_) {
+        params = params_;
+        std::tie(model, ctx) = llama_init_from_gpt_params(params);
+        if (model == nullptr)
+        {
+            LOG_TEE("unable to load model: %s", params.model.c_str());
+            return false;
+        }
+        n_ctx = llama_n_ctx(ctx);
+        n_vocab = llama_n_vocab(model);
+        candidates.reserve(n_vocab);
+        return true;
+    }
+
+    void initialize() {
+        // create slots
+        LOG_TEE("Available slots:\n");
+        for (int i = 0; i < params.n_parallel; i++)
+        {
+            llama_client_slot slot;
+            slot.id = i;
+            slot.prompt = "default";
+            slot.state = IDLE;
+            slot.tokens_prev.resize(std::max(256, params.n_predict));
+            std::fill(slot.tokens_prev.begin(), slot.tokens_prev.end(), 0);
+            LOG_TEE(" - slot %i\n", slot.id);
+            slots.push_back(slot);
+        }
+        batch = llama_batch_init(params.n_ctx, 0);
+
+        // always assign a default system prompt
+        system_prompt = system_prompt_default;
+        user_name = "User:";
+        assistant_name = "Assistant:";
+        params.antiprompt.push_back(user_name);
+    }
+
+    void updateSystemPrompt() {
+        tokens_system = ::llama_tokenize(ctx, system_prompt, true);
+        n_tokens_system = tokens_system.size();
+
+        batch.n_tokens = n_tokens_system;
+
+        // clear the entire KV cache
+        for (int i = 0; i < params.n_parallel; ++i)
+        {
+            llama_kv_cache_seq_rm(ctx, i, 0, -1);
+        }
+
+        for (int32_t i = 0; i < batch.n_tokens; ++i)
+        {
+            batch.token[i] = tokens_system[i];
+            batch.pos[i] = i;
+            batch.seq_id[i] = 0;
+            batch.logits[i] = false;
+        }
+
+        if (llama_decode(ctx, batch) != 0)
+        {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return;
+        }
+
+        // assign the system KV cache to all parallel sequences
+        for (int32_t i = 1; i < params.n_parallel; ++i)
+        {
+            llama_kv_cache_seq_cp(ctx, 0, i, 0, n_tokens_system);
+        }
+
+        LOG_TEE("system prompt updated\n");
+        update_system_prompt = false;
+    }
+
+    void notifySystemPromptChanged() {
+        // release all slots
+        for (llama_client_slot &slot : slots)
+        {
+            slot.release();
+        }
+        waitAllAreIdle();
+        // wait until system prompt load
+        update_system_prompt = true;
+        while(update_system_prompt) {
+            this_thread::sleep_for(chrono::milliseconds(5));
+        }
+        // system prompt loaded, continue
+    }
+
+    llama_client_slot* requestCompletion(json data) {
+        if(data.contains("system_prompt") &&
+            data.contains("anti_prompt") &&
+            data.contains("assistant_name")) {
+            system_prompt = data.value("system_prompt", "");
+            user_name = data.value("anti_prompt", "");
+            assistant_name = data.value("assistant_name", "");
+            params.antiprompt.clear();
+            params.antiprompt.push_back(user_name);
+            notifySystemPromptChanged();
+        }
+        int slot_id = data.value("slot_id", -1);
+        float temperature = data.value("temperature", 0.1f);
+        string prompt = data.value("prompt", "");
+        for (llama_client_slot & slot : slots)
+        {
+            if ((slot_id == -1 && slot.available()) || slot.id == slot_id)
+            {
+                slot.start(prompt, temperature);
+                LOG_TEE("slot %i is processing\n", slot.id);
+                return &slot; // return a pointer to slot (thread safe?)
+            }
+        }
+        return nullptr;
+    }
+
+    size_t findStoppingStrings(const std::string &text, const size_t last_token_size,
+                               const stop_type type)
+    {
+        size_t stop_pos = std::string::npos;
+        for (const std::string &word : params.antiprompt)
+        {
+            size_t pos;
+            if (type == STOP_FULL)
+            {
+                const size_t tmp = word.size() + last_token_size;
+                const size_t from_pos = text.size() > tmp ? text.size() - tmp : 0;
+                pos = text.find(word, from_pos);
+            }
+            else
+            {
+                pos = find_partial_stop_string(word, text);
+            }
+            if (pos != std::string::npos &&
+                (stop_pos == std::string::npos || pos < stop_pos))
+            {
+                stop_pos = pos;
+            }
+        }
+        return stop_pos;
+    }
+
+    void waitAllAreIdle() {
+        bool wait = true;
+        while(wait) {
+            wait = false;
+            for (auto &slot : slots)
+            {
+                if (!slot.available())
+                {
+                    wait = true;
+                    break;
+                }
+            }
+        }
+    }
+
+    bool updateSlots() {
+        // update the system prompt wait until all slots are idle state
+        if(update_system_prompt) {
+            updateSystemPrompt();
+        }
+
+        batch.n_tokens = 0;
+
+        // decode any currently ongoing sequences
+        for (auto & slot : slots) {
+            if (slot.state == PROCESSING && slot.command == RELEASE)
+            {
+                LOG_TEE("slot %i released\n", slot.id);
+                llama_kv_cache_seq_rm(ctx, slot.id, n_tokens_system, n_ctx);
+                slot.state = IDLE;
+                slot.command = NONE;
+                continue;
+            }
+
+            // no decode wait until the token had been send to client
+            // improves performance and avoid decoherence?
+
+            if (slot.state == IDLE || slot.newToken) {
+                continue;
+            }
+
+            batch.token [batch.n_tokens] = slot.sampled;
+            batch.pos   [batch.n_tokens] = n_tokens_system + slot.n_prompt + slot.n_decoded;
+            batch.seq_id[batch.n_tokens] = slot.id;
+            batch.logits[batch.n_tokens] = true;
+
+            slot.n_decoded += 1;
+            slot.i_batch = batch.n_tokens;
+
+            batch.n_tokens += 1;
+        }
+
+        // assign workload to the slots
+        if (params.cont_batching || batch.n_tokens == 0) {
+            for (llama_client_slot & slot : slots) {
+                // need process the prompt
+                if (slot.state == IDLE && slot.command == LOAD_PROMPT) {
+                    slot.state = PROCESSING;
+                    slot.command = NONE;
+                    //LOG_TEE("slot %i process prompt:\n%s%s'------------------------------\n", slot.id, system_prompt.c_str(), slot.prompt.c_str());
+                    std::fill(slot.tokens_prev.begin(), slot.tokens_prev.end(), 0);
+
+                    // do not prepend BOS because we have a system prompt!
+                    std::vector<llama_token> tokens_prompt;
+                    tokens_prompt = ::llama_tokenize(ctx, slot.prompt, false);
+
+                    for (size_t i = 0; i < tokens_prompt.size(); ++i) {
+                        batch.token [batch.n_tokens] = tokens_prompt[i];
+                        batch.pos   [batch.n_tokens] = i + n_tokens_system;
+                        batch.seq_id[batch.n_tokens] = slot.id;
+                        batch.logits[batch.n_tokens] = false;
+                        batch.n_tokens += 1;
+                    }
+
+                    // extract the logits only for the last token
+                    if (batch.n_tokens > 0) {
+                        batch.logits[batch.n_tokens - 1] = true;
+                    }
+
+                    slot.n_prompt  = tokens_prompt.size();
+                    slot.n_decoded = 0;
+                    slot.i_batch   = batch.n_tokens - 1;
+
+                    // insert new requests one-by-one
+                    //if (cont_batching) {
+                    //    break;
+                    //}
+                }
+            }
+        }
+
+        if (batch.n_tokens == 0) {
+            return true;
+        }
+
+        // process in chunks of params.n_batch
+        int32_t n_batch = params.n_batch;
+
+        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
+            // experiment: process in powers of 2
+            //if (i + n_batch > (int32_t) batch.n_tokens && n_batch > 32) {
+            //    n_batch /= 2;
+            //    i -= n_batch;
+            //    continue;
+            //}
+
+            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+
+            llama_batch batch_view = {
+                n_tokens,
+                batch.token  + i,
+                nullptr,
+                batch.pos    + i,
+                batch.seq_id + i,
+                batch.logits + i,
+                0, 0, 0, // unused
+            };
+
+            const int ret = llama_decode(ctx, batch_view);
+            if (ret != 0) {
+                if (n_batch == 1 || ret < 0) {
+                    // if you get here, it means the KV cache is full - try increasing it via the context size
+                    LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
+                    return false;
+                }
+
+                LOG("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
+
+                // retry with half the batch size to try to find a free slot in the KV cache
+                n_batch /= 2;
+                i -= n_batch;
+                continue;
+            }
+
+            for (auto & slot : slots) {
+                if (slot.i_batch < (int) i || slot.i_batch >= (int) (i + n_tokens)) {
+                    continue;
+                }
+
+                params.temp = slot.temperature;
+                const llama_token id = llama_sample_token(ctx, NULL, NULL, params, slot.tokens_prev, candidates, slot.i_batch - i);
+
+                // remember which tokens were sampled - used for repetition penalties during sampling
+                slot.tokens_prev.erase(slot.tokens_prev.begin());
+                slot.tokens_prev.push_back(id);
+
+                const std::string token_str = llama_token_to_piece(ctx, id);
+                slot.generated_text += token_str;
+                slot.sampled = id;
+
+                size_t stop_pos =
+                        findStoppingStrings(slot.generated_text, token_str.size(), STOP_FULL);
+
+                slot.sampled_token_str = token_str;
+                // notify new token
+                slot.nofity();
+
+                if (slot.n_decoded > 2 &&
+                        (id == llama_token_eos(ctx) ||
+                         (params.n_predict > 0 &&
+                        slot.n_decoded + slot.n_prompt >=
+                        params.n_predict) ||
+                         stop_pos != std::string::npos)) {
+                    //LOG_TEE("slot %i generated text:\n%s'------------------------------\n", slot.id, slot.generated_text.c_str());
+                    slot.generated_text.clear();
+                    slot.release();
+                }
+                slot.i_batch = -1;
+            }
+        }
+        return true;
+    }
+};
+
+static void server_print_usage(const char *argv0, const gpt_params &params,
+                               const server_params &sparams)
+{
+    printf("usage: %s [options]\n", argv0);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help            show this help message and exit\n");
+    printf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    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("  --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())
+    {
+        printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+    }
+    if (llama_mmap_supported())
+    {
+        printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+    }
+    printf("  --numa                attempt optimizations that help on some NUMA systems\n");
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+    printf("  -ngl N, --n-gpu-layers N\n");
+    printf("                        number of layers to store in VRAM\n");
+    printf("  -ts SPLIT --tensor-split SPLIT\n");
+    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    printf("  -nommq, --no-mul-mat-q\n");
+    printf("                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
+    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
+#endif
+    printf("  -m FNAME, --model FNAME\n");
+    printf("                        model path (default: %s)\n", params.model.c_str());
+    printf("  -a ALIAS, --alias ALIAS\n");
+    printf("                        set an alias for the model, will be added as `model` field in completion response\n");
+    printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    printf("  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
+    printf("  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
+    printf("  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
+    printf("  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
+
+    // new arguments
+    printf("  -np N, --parallel N   number of parallel sequences to decode (default: %d)\n", params.n_parallel);
+    printf("  -cb, --cont-batching  enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
+    printf("  -f FNAME, --file FNAME\n");
+    printf("                        load a system prompt from a file.\n");
+    printf("\n");
+}
+
+static void server_params_parse(int argc, char **argv, server_params &sparams,
+                                gpt_params &params)
+{
+    gpt_params default_params;
+    server_params default_sparams;
+    std::string arg;
+    bool invalid_param = false;
+
+    for (int i = 1; i < argc; i++)
+    {
+        arg = argv[i];
+        if (arg == "--port")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            sparams.port = std::stoi(argv[i]);
+        }
+        else if (arg == "--host")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            sparams.hostname = argv[i];
+        }
+        else if (arg == "--path")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            sparams.public_path = argv[i];
+        }
+        else if (arg == "--timeout" || arg == "-to")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            sparams.read_timeout = std::stoi(argv[i]);
+            sparams.write_timeout = std::stoi(argv[i]);
+        }
+        else if (arg == "-m" || arg == "--model")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.model = argv[i];
+        }
+        else if (arg == "-a" || arg == "--alias")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.model_alias = argv[i];
+        }
+        else if (arg == "-h" || arg == "--help")
+        {
+            server_print_usage(argv[0], default_params, default_sparams);
+            exit(0);
+        }
+        else if (arg == "-c" || arg == "--ctx-size" || arg == "--ctx_size")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.n_ctx = std::stoi(argv[i]);
+        }
+        else if (arg == "--rope-freq-base")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_base = std::stof(argv[i]);
+        }
+        else if (arg == "--rope-freq-scale")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_scale = std::stof(argv[i]);
+        }
+        else if (arg == "--memory-f32" || arg == "--memory_f32")
+        {
+            params.memory_f16 = false;
+        }
+        else if (arg == "--threads" || arg == "-t")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.n_threads = std::stoi(argv[i]);
+        }
+        else if (arg == "-b" || arg == "--batch-size")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.n_batch = std::stoi(argv[i]);
+            params.n_batch = std::min(512, params.n_batch);
+        }
+        else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+            params.n_gpu_layers = std::stoi(argv[i]);
+#else
+            LOG_TEE("Not compiled with GPU offload support, --n-gpu-layers option will be ignored. "
+                        "See main README.md for information on enabling GPU BLAS support\n");
+#endif
+        }
+        else if (arg == "--tensor-split" || arg == "-ts")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+#ifdef GGML_USE_CUBLAS
+            std::string arg_next = argv[i];
+
+            // split string by , and /
+            const std::regex regex{R"([,/]+)"};
+            std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
+            std::vector<std::string> split_arg{it, {}};
+            GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
+
+            for (size_t i_device = 0; i_device < LLAMA_MAX_DEVICES; ++i_device)
+            {
+                if (i_device < split_arg.size())
+                {
+                    params.tensor_split[i_device] = std::stof(split_arg[i_device]);
+                }
+                else
+                {
+                    params.tensor_split[i_device] = 0.0f;
+                }
+            }
+#else
+            LOG_TEE("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+        }
+        else if (arg == "--no-mul-mat-q" || arg == "-nommq")
+        {
+#ifdef GGML_USE_CUBLAS
+            params.mul_mat_q = false;
+#else
+            LOG_TEE("warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
+#endif // GGML_USE_CUBLAS
+        }
+        else if (arg == "--main-gpu" || arg == "-mg")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+#ifdef GGML_USE_CUBLAS
+            params.main_gpu = std::stoi(argv[i]);
+#else
+            LOG_TEE("llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.");
+#endif
+        }
+        else if (arg == "--lora")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter.push_back({argv[i], 1.0f});
+            params.use_mmap = false;
+        }
+        else if (arg == "--lora-scaled")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            const char * lora_adapter = argv[i];
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter.push_back(make_tuple(lora_adapter, std::stof(argv[i])));
+            params.use_mmap = false;
+        }
+        else if (arg == "--lora-base")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.lora_base = argv[i];
+        }
+        else if (arg == "--mlock")
+        {
+            params.use_mlock = true;
+        }
+        else if (arg == "--no-mmap")
+        {
+            params.use_mmap = false;
+        }
+        else if (arg == "--numa")
+        {
+            params.numa = true;
+        } else if (arg == "-cb" || arg == "--cont-batching")
+        {
+            params.cont_batching = true;
+        }
+        else if (arg == "-np" || arg == "--parallel")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.n_parallel = std::stoi(argv[i]);
+        } else if (arg == "-n" || arg == "--n-predict")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.n_predict = std::stoi(argv[i]);
+        } else if (arg == "-r" || arg == "--reverse-prompt")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.antiprompt.push_back(argv[i]);
+        }
+        else
+        {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            server_print_usage(argv[0], default_params, default_sparams);
+            exit(1);
+        }
+    }
+
+    if (invalid_param)
+    {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        server_print_usage(argv[0], default_params, default_sparams);
+        exit(1);
+    }
+}
+
+int main(int argc, char **argv)
+{
+    gpt_params params;
+
+    server_params sparams;
+
+    server_params_parse(argc, argv, sparams, params);
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("server-parallel", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    llama_backend_init(params.numa);
+
+    // load the target model
+    params.logits_all = true;
+    server_parallel_context llama;
+
+    if(!llama.loadModel(params)) {
+        return 1;
+    }
+
+    llama.initialize();
+
+    Server svr;
+
+    svr.Get("/", [&](const Request & /*req*/, Response &res)
+            { res.set_content(index_html_, "text/html"); });
+
+    svr.Get("/index.js", [&](const Request & /*req*/, Response &res)
+            { res.set_content(index_js_, "text/html"); });
+
+    svr.Get("/props", [&llama](const Request & /*req*/, Response &res)
+            {
+                json data = {
+                    { "user_name", llama.user_name.c_str() },
+                    { "assistant_name", llama.assistant_name.c_str() }
+                };
+                res.set_content(data.dump(), "application/json"); });
+
+    svr.Post("/completion", [&llama](const Request &req, Response &res)
+             {
+        llama_client_slot* slot = llama.requestCompletion(json::parse(req.body));
+        // Verify if the slot exist
+        if (slot) {
+            res.set_chunked_content_provider("text/event-stream",
+                [slot](size_t /*offset*/, DataSink &sink) {
+                    if(slot->available()) { // slot has been released
+                        sink.done();
+                        return false;
+                    }
+                    if(slot->hasNewToken()) { // new token notification
+                        stringstream ss;
+                        json res_d = {{ "content", slot->sampled_token_str }};
+                        ss << "data: " << res_d.dump() << "\n\n";
+                        string result = ss.str();
+                        if(!sink.write(result.c_str(), result.size())) {
+                            slot->release();
+                            return false;
+                        }
+                    }
+                    return true;
+                });
+        } else {
+            LOG_TEE("slot unavailable\n");
+            res.status = 404;
+            res.set_content("slot_error", "text/plain");
+        } });
+
+    thread t([&llama]()
+             {
+            bool running = true;
+            while (running)
+            {
+                running = llama.updateSlots();
+            } });
+
+    svr.set_read_timeout(sparams.read_timeout);
+    svr.set_write_timeout(sparams.write_timeout);
+
+    if (!svr.bind_to_port(sparams.hostname, sparams.port))
+    {
+        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port);
+        return 1;
+    }
+
+    // Set the base directory for serving static files
+    svr.set_base_dir(sparams.public_path);
+
+    // to make it ctrl+clickable:
+    printf("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
+
+    if (!svr.listen_after_bind())
+    {
+        return 1;
+    }
+}

examples/server/CMakeLists.txt

@@ -6,7 +6,7 @@ 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 ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()

examples/server/README.md

@@ -24,10 +24,6 @@ Command line options:
 -   `--port`: Set the port to listen. Default: `8080`.
 -   `--path`: path from which to serve static files (default examples/server/public)
 -   `--embedding`: Enable embedding extraction, Default: disabled.
--   `-np N`, `--parallel N`: Set the number of slots for process requests (default: 1)
--   `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled)
--   `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
--   `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
 
 ## Build
 
@@ -110,25 +106,25 @@ node index.js
 
 ## API Endpoints
 
--   **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
+-   **POST** `/completion`: Given a prompt, it returns the predicted completion.
 
     *Options:*
 
-    `prompt`: Provide the prompt for this completion as a string or as an array of strings or numbers representing tokens. Internally, the prompt is compared to the previous completion and only the "unseen" suffix is evaluated. If the prompt is a string or an array with the first element given as a string, a `bos` token is inserted in the front like `main` does.
-
     `temperature`: Adjust the randomness of the generated text (default: 0.8).
 
     `top_k`: Limit the next token selection to the K most probable tokens (default: 40).
 
-    `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.95).
+    `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
 
-    `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_predict`: Set the 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: 128, -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.
-    By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the prompt.
+    `n_keep`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context.
+    By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
 
     `stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
 
+    `prompt`: Provide a prompt as a string, or as an array of strings and numbers representing tokens. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate. If the prompt is a string, or an array with the first element given as a string, a space is inserted in the front like main.cpp does.
+
     `stop`: Specify a JSON array of stopping strings.
     These words will not be included in the completion, so make sure to add them to the prompt for the next iteration (default: []).
 
@@ -160,46 +156,6 @@ node index.js
 
     `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced (default: []).
 
-    `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
-
-    `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:` In this case, `[img-12]` will be replaced by the embeddings of the image id 12 in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
-
-    *Result JSON:*
-
-    Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
-
-    `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
-
-    `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
-
-    `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
-
-    `model`: The path to the model loaded with `-m`
-
-    `prompt`: The provided `prompt`
-
-    `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
-
-    `stopped_limit`: Indicating whether the completion stopped because `n_predict` tokens were generated before stop words or EOS was encountered
-
-    `stopped_word`: Indicating whether the completion stopped due to encountering a stopping word from `stop` JSON array provided
-
-    `stopping_word`: The stopping word encountered which stopped the generation (or "" if not stopped due to a stopping word)
-
-    `timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second`
-
-    `tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`)
-
-    `tokens_evaluated`: Number of tokens evaluated in total from the prompt
-
-    `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
-
-    `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
-
-    `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
-
-    `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
-
 -   **POST** `/tokenize`: Tokenize a given text.
 
     *Options:*
@@ -230,32 +186,8 @@ node index.js
 
     It also accepts all the options of `/completion` except `stream` and `prompt`.
 
--   **GET** `/props`: Return the required assistant name and anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
-
 ## More examples
 
-### Change system prompt on runtime
-
-To use the server example to serve multiple chat-type clients while keeping the same system prompt, you can utilize the option `system_prompt` to achieve that. This only needs to be done once to establish it.
-
-`prompt`: Specify a context that you want all connecting clients to respect.
-
-`anti_prompt`: Specify the word you want to use to instruct the model to stop. This must be sent to each client through the `/props` endpoint.
-
-`assistant_name`: The bot's name is necessary for each customer to generate the prompt. This must be sent to each client through the `/props` endpoint.
-
-```json
-{
-    "system_prompt": {
-        "prompt": "Transcript of a never ending dialog, where the User interacts with an Assistant.\nThe Assistant is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.\nUser: Recommend a nice restaurant in the area.\nAssistant: I recommend the restaurant \"The Golden Duck\". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.\nUser: Who is Richard Feynman?\nAssistant: Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including \"Surely You're Joking, Mr. Feynman!\" and \"What Do You Care What Other People Think?\".\nUser:",
-        "anti_prompt": "User:",
-        "assistant_name": "Assistant:"
-    }
-}
-```
-
-**NOTE**: You can do this automatically when starting the server by simply creating a .json file with these options and using the CLI option `-spf FNAME` or `--system-prompt-file FNAME`.
-
 ### Interactive mode
 
 Check the sample in [chat.mjs](chat.mjs).

examples/server/api_like_OAI.py

@@ -8,7 +8,6 @@ import json
 
 
 app = Flask(__name__)
-slot_id = -1
 
 parser = argparse.ArgumentParser(description="An example of using server.cpp with a similar API to OAI. It must be used together with server.cpp.")
 parser.add_argument("--chat-prompt", type=str, help="the top prompt in chat completions(default: 'A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n')", default='A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n')
@@ -28,10 +27,10 @@ def is_present(json, key):
         buf = json[key]
     except KeyError:
         return False
-    if json[key] == None:
-        return False
     return True
 
+
+
 #convert chat to prompt
 def convert_chat(messages):
     prompt = "" + args.chat_prompt.replace("\\n", "\n")
@@ -78,8 +77,7 @@ def make_postData(body, chat=False, stream=False):
     if(is_present(body, "stop")): postData["stop"] += body["stop"]
     postData["n_keep"] = -1
     postData["stream"] = stream
-    postData["cache_prompt"] = True
-    postData["slot_id"] = slot_id
+
     return postData
 
 def make_resData(data, chat=False, promptToken=[]):
@@ -130,7 +128,6 @@ def make_resData_stream(data, chat=False, time_now = 0, start=False):
             }
         ]
     }
-    slot_id = data["slot_id"]
     if (chat):
         if (start):
             resData["choices"][0]["delta"] =  {

examples/server/chat.mjs

@@ -7,11 +7,6 @@ const args = process.argv.slice(2);
 const grammarJsonSchemaFile = args.find(
     (_, index) => args[index - 1] === "--grammar-json-schema"
 );
-
-const no_cached_prompt = args.find(
-    (_, index) => args[index - 1] === "--no-cache-prompt"
-) ?? "false";
-
 const grammarFile = args.find((_, index) => args[index - 1] === "--grammar");
 
 // Example usage: function,arguments
@@ -35,9 +30,6 @@ if (grammarFile) {
     grammar = readFileSync(grammarFile, 'utf-8')
 }
 
-// for cached prompt
-let slot_id = -1;
-
 const API_URL = 'http://127.0.0.1:8080'
 
 const chat = [
@@ -84,8 +76,6 @@ async function chat_completion(question) {
             top_p: 0.9,
             n_keep: n_keep,
             n_predict: 256,
-            cache_prompt: no_cached_prompt === "false",
-            slot_id: slot_id,
             stop: ["\n### Human:"], // stop completion after generating this
             grammar,
             stream: true,
@@ -102,7 +92,6 @@ async function chat_completion(question) {
         const t = Buffer.from(chunk).toString('utf8')
         if (t.startsWith('data: ')) {
             const message = JSON.parse(t.substring(6))
-            slot_id = message.slot_id
             answer += message.content
             process.stdout.write(message.content)
             if (message.stop) {

examples/server/public/index.html

@@ -125,7 +125,6 @@
       background-color: #222;
       color: #ddd;
     }
-
     code {
       font-family: monospace;
       padding: 0.1em 0.3em;
@@ -137,13 +136,7 @@
       display: block;
     }
 
-    fieldset label.slim {
-      margin: 0 0.5em;
-      display: inline;
-    }
-
-    header,
-    footer {
+    header, footer {
       text-align: center;
     }
 
@@ -152,20 +145,11 @@
       color: #888;
     }
 
-    .mode-chat textarea[name=prompt] {
-      height: 4.5em;
-    }
-
-    .mode-completion textarea[name=prompt] {
-      height: 10em;
-    }
-
 
     @keyframes loading-bg-wipe {
       0% {
         background-position: 0%;
       }
-
       100% {
         background-position: 100%;
       }
@@ -184,7 +168,6 @@
         --loading-color-1: #22222200;
         --loading-color-2: #222222ff;
       }
-
       .popover-content {
         background-color: black;
       }
@@ -198,18 +181,15 @@
 
     import { llama } from '/completion.js';
     import { SchemaConverter } from '/json-schema-to-grammar.mjs';
-    let selected_image = false;
-    var slot_id = -1;
 
     const session = signal({
       prompt: "This is a conversation between User and Llama, a friendly chatbot. Llama is helpful, kind, honest, good at writing, and never fails to answer any requests immediately and with precision.",
       template: "{{prompt}}\n\n{{history}}\n{{char}}:",
       historyTemplate: "{{name}}: {{message}}",
       transcript: [],
-      type: "chat",  // "chat" | "completion"
+      type: "chat",
       char: "Llama",
       user: "User",
-      image_selected: ''
     })
 
     const params = signal({
@@ -227,9 +207,7 @@
       mirostat_tau: 5, // target entropy
       mirostat_eta: 0.1, // learning rate
       grammar: '',
-      n_probs: 0, // no completion_probabilities,
-      image_data: [],
-      cache_prompt: true
+      n_probs: 0, // no completion_probabilities
     })
 
     /* START: Support for storing prompt templates and parameters in borwser LocalStorage */
@@ -279,7 +257,6 @@
       // saved templates were successfuly imported.
 
       console.log('Processing saved templates and updating default template')
-      params.value = { ...params.value, image_data: [] };
 
       //console.log(importedTemplates);
       savedUserTemplates.value = importedTemplates;
@@ -304,9 +281,7 @@
 
     function userTemplateApply(t) {
       session.value = t.data.session;
-      session.value = { ...session.value, image_selected: '' };
       params.value = t.data.params;
-      params.value = { ...params.value, image_data: [] };
     }
 
     function userTemplateResetToDefaultAndApply() {
@@ -390,53 +365,17 @@
       return String(str).replaceAll(/\{\{(.*?)\}\}/g, (_, key) => template(settings[key]));
     }
 
-    async function runLlama(prompt, llamaParams, char) {
-      const currentMessages = [];
-      const history = session.value.transcript;
-      if (controller.value) {
-        throw new Error("already running");
-      }
-      controller.value = new AbortController();
-      for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) {
-        const data = chunk.data;
-
-        if (data.stop) {
-          while (
-            currentMessages.length > 0 &&
-            currentMessages[currentMessages.length - 1].content.match(/\n$/) != null
-          ) {
-            currentMessages.pop();
-          }
-          transcriptUpdate([...history, [char, currentMessages]])
-          console.log("Completion finished: '", currentMessages.map(msg => msg.content).join(''), "', summary: ", data);
-        } else {
-          currentMessages.push(data);
-          slot_id = data.slot_id;
-          if (selected_image && !data.multimodal) {
-            alert("The server was not compiled for multimodal or the model projector can't be loaded.");
-            return;
-          }
-          transcriptUpdate([...history, [char, currentMessages]])
-        }
-
-        if (data.timings) {
-          llamaStats.value = data.timings;
-        }
-      }
-
-      controller.value = null;
-    }
-
     // send message to server
     const chat = async (msg) => {
       if (controller.value) {
         console.log('already running...');
         return;
       }
+      controller.value = new AbortController();
 
       transcriptUpdate([...session.value.transcript, ["{{user}}", msg]])
 
-      let prompt = template(session.value.template, {
+      const prompt = template(session.value.template, {
         message: msg,
         history: session.value.transcript.flatMap(
           ([name, data]) =>
@@ -451,67 +390,56 @@
             )
         ).join("\n"),
       });
-      if (selected_image) {
-        prompt = `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:[img-10]${msg}\nASSISTANT:`;
-      }
-      await runLlama(prompt, {
+
+      const currentMessages = [];
+      const history = session.value.transcript
+
+      const llamaParams = {
         ...params.value,
-        slot_id: slot_id,
         stop: ["</s>", template("{{char}}:"), template("{{user}}:")],
-      }, "{{char}}");
-    }
-
-    const runCompletion = async () => {
-      if (controller.value) {
-        console.log('already running...');
-        return;
       }
-      const { prompt } = session.value;
-      transcriptUpdate([...session.value.transcript, ["", prompt]]);
-      await runLlama(prompt, {
-        ...params.value,
-        slot_id: slot_id,
-        stop: [],
-      }, "");
-    }
 
-    const stop = (e) => {
-      e.preventDefault();
-      if (controller.value) {
-        controller.value.abort();
-        controller.value = null;
-      }
-    }
+      for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) {
+        const data = chunk.data;
 
-    const reset = (e) => {
-      stop(e);
-      transcriptUpdate([]);
-    }
-
-    const uploadImage = (e) => {
-      e.preventDefault();
-      document.getElementById("fileInput").click();
-      document.getElementById("fileInput").addEventListener("change", function (event) {
-        const selectedFile = event.target.files[0];
-        if (selectedFile) {
-          const reader = new FileReader();
-          reader.onload = function () {
-            const image_data = reader.result;
-            session.value = { ...session.value, image_selected: image_data };
-            params.value = {
-              ...params.value, image_data: [
-                { data: image_data.replace(/data:image\/[^;]+;base64,/, ''), id: 10 }]
-            }
-          };
-          selected_image = true;
-          reader.readAsDataURL(selectedFile);
+        if (data.stop) {
+          while (
+            currentMessages.length > 0 &&
+            currentMessages[currentMessages.length - 1].content.match(/\n$/) != null
+          ) {
+            currentMessages.pop();
+          }
+          transcriptUpdate([...history, ["{{char}}", currentMessages]])
+          console.log("Completion finished: '", currentMessages.map(msg => msg.content).join(''), "', summary: ", data);
+        } else {
+          currentMessages.push(data);
+          transcriptUpdate([...history, ["{{char}}", currentMessages]])
         }
-      });
+
+        if (data.timings) {
+          llamaStats.value = data.timings;
+        }
+      }
+
+      controller.value = null;
     }
 
     function MessageInput() {
       const message = useSignal("")
 
+      const stop = (e) => {
+        e.preventDefault();
+        if (controller.value) {
+          controller.value.abort();
+          controller.value = null;
+        }
+      }
+
+      const reset = (e) => {
+        stop(e);
+        transcriptUpdate([]);
+      }
+
       const submit = (e) => {
         stop(e);
         chat(message.value);
@@ -539,7 +467,6 @@
           </div>
           <div class="right">
             <button type="submit" disabled=${generating.value}>Send</button>
-            <button onclick=${uploadImage}>Upload Image</button>
             <button onclick=${stop} disabled=${!generating.value}>Stop</button>
             <button onclick=${reset}>Reset</button>
           </div>
@@ -547,19 +474,6 @@
       `
     }
 
-    function CompletionControls() {
-      const submit = (e) => {
-        stop(e);
-        runCompletion();
-      }
-      return html`
-        <div>
-          <button onclick=${submit} type="button" disabled=${generating.value}>Start</button>
-          <button onclick=${stop} disabled=${!generating.value}>Stop</button>
-          <button onclick=${reset}>Reset</button>
-        </div>`;
-    }
-
     const ChatLog = (props) => {
       const messages = session.value.transcript;
       const container = useRef(null)
@@ -583,16 +497,11 @@
             data;
           message = html`<${Markdownish} text=${template(text)} />`
         }
-        if (user) {
-          return html`<p key=${index}><strong>${template(user)}:</strong> ${message}</p>`
-        } else {
-          return html`<p key=${index}>${message}</p>`
-        }
+        return html`<p key=${index}><strong>${template(user)}:</strong> ${message}</p>`
       };
 
       return html`
         <section id="chat" ref=${container}>
-          <img style="width: 60%;${!session.value.image_selected ? `display: none;` : ``}" src="${session.value.image_selected}"/>
           ${messages.flatMap(chatLine)}
         </section>`;
     };
@@ -611,7 +520,7 @@
           const converter = new SchemaConverter(
             grammarJsonSchemaPropOrder.value
               .split(',')
-              .reduce((acc, cur, i) => ({ ...acc, [cur.trim()]: i }), {})
+              .reduce((acc, cur, i) => ({...acc, [cur.trim()]: i}), {})
           )
           converter.visit(schema, '')
           params.value = {
@@ -623,7 +532,7 @@
         }
       }
 
-      const FloatField = ({ label, max, min, name, step, value }) => {
+      const FloatField = ({label, max, min, name, step, value}) => {
         return html`
           <div>
             <label for="${name}">${label}</label>
@@ -633,7 +542,7 @@
         `
       };
 
-      const IntField = ({ label, max, min, name, value }) => {
+      const IntField = ({label, max, min, name, value}) => {
         return html`
           <div>
             <label for="${name}">${label}</label>
@@ -665,31 +574,18 @@
         userTemplateAutosave()
       }, [session.value, params.value])
 
-      const GrammarControl = () => (
-        html`
-          <div>
-            <label for="template">Grammar</label>
-            <textarea id="grammar" name="grammar" placeholder="Use gbnf or JSON Schema+convert" value="${params.value.grammar}" rows=4 oninput=${updateParams}/>
-            <input type="text" name="prop-order" placeholder="order: prop1,prop2,prop3" oninput=${updateGrammarJsonSchemaPropOrder} />
-            <button type="button" onclick=${convertJSONSchemaGrammar}>Convert JSON Schema</button>
-          </div>
-          `
-      );
+      return html`
+        <form>
+          <fieldset>
+            <${UserTemplateResetButton}/>
+          </fieldset>
 
-      const PromptControlFieldSet = () => (
-        html`
-        <fieldset>
-          <div>
-            <label htmlFor="prompt">Prompt</label>
-            <textarea type="text" name="prompt" value="${session.value.prompt}" oninput=${updateSession}/>
-          </div>
-        </fieldset>
-        `
-      );
-
-      const ChatConfigForm = () => (
-        html`
-          ${PromptControlFieldSet()}
+          <fieldset>
+            <div>
+              <label for="prompt">Prompt</label>
+              <textarea type="text" name="prompt" value="${session.value.prompt}" rows=4 oninput=${updateSession}/>
+            </div>
+          </fieldset>
 
           <fieldset class="two">
             <div>
@@ -713,45 +609,30 @@
               <label for="template">Chat history template</label>
               <textarea id="template" name="historyTemplate" value="${session.value.historyTemplate}" rows=1 oninput=${updateSession}/>
             </div>
-            ${GrammarControl()}
-          </fieldset>
-      `
-      );
 
-      const CompletionConfigForm = () => (
-        html`
-          ${PromptControlFieldSet()}
-          <fieldset>${GrammarControl()}</fieldset>
-        `
-      );
-
-      return html`
-        <form>
-          <fieldset class="two">
-            <${UserTemplateResetButton}/>
             <div>
-              <label class="slim"><input type="radio" name="type" value="chat" checked=${session.value.type === "chat"} oninput=${updateSession} /> Chat</label>
-              <label class="slim"><input type="radio" name="type" value="completion" checked=${session.value.type === "completion"} oninput=${updateSession} /> Completion</label>
+              <label for="template">Grammar</label>
+              <textarea id="grammar" name="grammar" placeholder="Use gbnf or JSON Schema+convert" value="${params.value.grammar}" rows=4 oninput=${updateParams}/>
+              <input type="text" name="prop-order" placeholder="order: prop1,prop2,prop3" oninput=${updateGrammarJsonSchemaPropOrder} />
+              <button type="button" onclick=${convertJSONSchemaGrammar}>Convert JSON Schema</button>
             </div>
           </fieldset>
 
-          ${session.value.type === 'chat' ? ChatConfigForm() : CompletionConfigForm()}
-
           <fieldset class="two">
-            ${IntField({ label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict })}
-            ${FloatField({ label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })}
-            ${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })}
-            ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })}
-            ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })}
-            ${FloatField({ label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
+            ${IntField({label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict})}
+            ${FloatField({label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature})}
+            ${FloatField({label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty})}
+            ${IntField({label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n})}
+            ${IntField({label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k})}
+            ${FloatField({label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p})}
           </fieldset>
           <details>
             <summary>More options</summary>
             <fieldset class="two">
-              ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
-              ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
-              ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
-              ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
+              ${FloatField({label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z})}
+              ${FloatField({label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p})}
+              ${FloatField({label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty})}
+              ${FloatField({label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty})}
             </fieldset>
             <hr />
             <fieldset class="three">
@@ -760,11 +641,11 @@
                 <label><input type="radio" name="mirostat" value="1" checked=${params.value.mirostat == 1} oninput=${updateParamsInt} /> Mirostat v1</label>
                 <label><input type="radio" name="mirostat" value="2" checked=${params.value.mirostat == 2} oninput=${updateParamsInt} /> Mirostat v2</label>
               </div>
-              ${FloatField({ label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau })}
-              ${FloatField({ label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta })}
+              ${FloatField({label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau})}
+              ${FloatField({label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta})}
             </fieldset>
             <fieldset>
-              ${IntField({ label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs })}
+              ${IntField({label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs})}
             </fieldset>
           </details>
         </form>
@@ -803,20 +684,20 @@
         const popoverChildren = html`
           <div class="prob-set">
             ${probs.map((p, index) => {
-          return html`
+              return html`
                 <div
                   key=${index}
                   title=${`prob: ${p.prob}`}
                   style=${{
-              padding: '0.3em',
-              backgroundColor: p.tok_str === content ? probColor(p.prob) : 'transparent'
-            }}
+                    padding: '0.3em',
+                    backgroundColor: p.tok_str === content ? probColor(p.prob) : 'transparent'
+                  }}
                 >
                   <span>${p.tok_str}: </span>
                   <span>${Math.floor(p.prob * 100)}%</span>
                 </div>
               `
-        })}
+            })}
           </div>
         `
 
@@ -895,9 +776,9 @@
               ref=${popoverRef}
               class="popover-content"
               style=${{
-            top: position.value.top,
-            left: position.value.left,
-          }}
+                top: position.value.top,
+                left: position.value.left,
+              }}
             >
               ${props.popoverChildren}
             </div>
@@ -970,7 +851,7 @@
     function App(props) {
 
       return html`
-        <div class="mode-${session.value.type}">
+        <div>
           <header>
             <h1>llama.cpp</h1>
           </header>
@@ -980,7 +861,7 @@
           </main>
 
           <section id="write">
-            <${session.value.type === 'chat' ? MessageInput : CompletionControls} />
+            <${MessageInput} />
           </section>
 
           <footer>
@@ -996,11 +877,8 @@
 </head>
 
 <body>
-  <div id="container">
-    <input type="file" id="fileInput" accept="image/*" style="display: none;">
-  </div>
+  <div id="container"></div>
   <div id="portal"></div>
 </body>
 
 </html>
-

examples/simple/simple.cpp

@@ -92,11 +92,16 @@ int main(int argc, char ** argv) {
     // create a llama_batch with size 512
     // we use this object to submit token data for decoding
 
-    llama_batch batch = llama_batch_init(512, 0, 1);
+    llama_batch batch = llama_batch_init(512, 0);
 
     // evaluate the initial prompt
-    for (size_t i = 0; i < tokens_list.size(); i++) {
-        llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
+    batch.n_tokens = tokens_list.size();
+
+    for (int32_t i = 0; i < batch.n_tokens; i++) {
+        batch.token[i]  = tokens_list[i];
+        batch.pos[i]    = i;
+        batch.seq_id[i] = 0;
+        batch.logits[i] = false;
     }
 
     // llama_decode will output logits only for the last token of the prompt
@@ -133,7 +138,7 @@ int main(int argc, char ** argv) {
             const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of stream?
-            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
+            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
                 LOG_TEE("\n");
 
                 break;
@@ -143,10 +148,15 @@ int main(int argc, char ** argv) {
             fflush(stdout);
 
             // prepare the next batch
-            llama_batch_clear(batch);
+            batch.n_tokens = 0;
 
             // push this new token for next evaluation
-            llama_batch_add(batch, new_token_id, n_cur, { 0 }, true);
+            batch.token [batch.n_tokens] = new_token_id;
+            batch.pos   [batch.n_tokens] = n_cur;
+            batch.seq_id[batch.n_tokens] = 0;
+            batch.logits[batch.n_tokens] = true;
+
+            batch.n_tokens += 1;
 
             n_decode += 1;
         }

examples/speculative/speculative.cpp

@@ -2,28 +2,13 @@
 
 #include "common.h"
 #include "llama.h"
+#include "grammar-parser.h"
 
 #include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  100
-#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
-
-struct seq_draft {
-    bool active   = false;
-    bool drafting = false;
-    bool skip     = false;
-
-    int i_batch_dft = 0;
-    std::vector<int> i_batch_tgt;
-
-    std::vector<llama_token> tokens;
-
-    struct llama_sampling_context * ctx_sampling;
-};
-
 int main(int argc, char ** argv) {
     gpt_params params;
 
@@ -36,13 +21,6 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    // 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;
-
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("speculative", "log"));
     LOG_TEE("Log start\n");
@@ -67,33 +45,6 @@ int main(int argc, char ** argv) {
     params.n_gpu_layers = params.n_gpu_layers_draft;
     std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
 
-    {
-        const int n_vocab_tgt = llama_n_vocab(model_tgt);
-        const int n_vocab_dft = llama_n_vocab(model_dft);
-        const int vocab_diff  = n_vocab_tgt > n_vocab_dft
-            ? n_vocab_tgt - n_vocab_dft
-            : n_vocab_dft - n_vocab_tgt;
-
-        if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
-            fprintf(stderr, "%s: error: draft model vocab must closely match target model to use speculation but ", __func__);
-            fprintf(stderr, "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
-                    n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
-            return 1;
-        }
-
-        for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
-            const char * token_text_tgt = llama_token_get_text(model_tgt, i);
-            const char * token_text_dft = llama_token_get_text(model_dft, i);
-            if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
-                fprintf(stderr, "%s: error: draft model vocab must match target model to use speculation but ", __func__);
-                fprintf(stderr, "token %d content differs - target '%s', draft '%s'\n", i,
-                        llama_token_to_piece(ctx_tgt, i).c_str(),
-                        llama_token_to_piece(ctx_dft, i).c_str());
-                return 1;
-            }
-        }
-    }
-
     // tokenize the prompt
     std::vector<llama_token> inp;
     inp = ::llama_tokenize(ctx_tgt, params.prompt, true);
@@ -126,6 +77,8 @@ int main(int argc, char ** argv) {
     const auto t_enc_end = ggml_time_us();
 
     // the 2 models should have the same vocab
+    const int n_ctx   = llama_n_ctx(ctx_tgt);
+    const int n_vocab = llama_n_vocab(model_tgt);
     //GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
 
     // how many tokens to draft each time
@@ -138,130 +91,115 @@ int main(int argc, char ** argv) {
     int n_past_tgt = inp.size();
     int n_past_dft = inp.size();
 
+    std::vector<llama_token> drafted;
+
+    std::vector<llama_token> last_tokens(n_ctx);
+    std::fill(last_tokens.begin(), last_tokens.end(), 0);
+
+    for (auto & id : inp) {
+        last_tokens.erase(last_tokens.begin());
+        last_tokens.push_back(id);
+    }
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
     // used to determine end of generation
     bool has_eos = false;
 
-    // target model sampling context
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
+    // grammar stuff
+    struct llama_grammar * grammar_dft = NULL;
+    struct llama_grammar * grammar_tgt = NULL;
 
-    // draft sequence data
-    std::vector<seq_draft> drafts(n_seq_dft);
+    grammar_parser::parse_state parsed_grammar;
 
-    params.sparams.grammar.clear(); // the draft samplers will copy the target sampler's grammar
-    params.sparams.temp = -1.0f;    // force greedy sampling with probs for the draft model
+    // if requested - load the grammar, error checking is omitted for brevity
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
 
-    for (int s = 0; s < n_seq_dft; ++s) {
-        drafts[s].ctx_sampling = llama_sampling_init(params.sparams);
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar_tgt = llama_grammar_init(grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
     }
 
-    llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1);
-    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, n_seq_dft);
-
     const auto t_dec_start = ggml_time_us();
 
-    // sample from the last token of the prompt
-    drafts[0].i_batch_tgt.resize(1);
-    drafts[0].i_batch_tgt[0] = 0;
-
     while (true) {
-        // print current draft sequences
-        for (int s = 0; s < n_seq_dft; ++s) {
-            if (!drafts[s].active) {
-                continue;
-            }
+        LOG("drafted: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_dft, drafted));
 
-            const auto & tokens = drafts[s].tokens;
-
-            LOG("draft %d: %s\n", s, LOG_TOKENS_TOSTR_PRETTY(ctx_dft, tokens).c_str());
-        }
-
-        int i_dft  = 0;
-        int s_keep = 0;
+        int i_dft = 0;
 
         while (true) {
-            LOG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
-
             // sample from the target model
-            llama_token id = llama_sampling_sample(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
+            llama_token id = llama_sample_token(ctx_tgt, NULL, grammar_tgt, params, last_tokens, candidates, i_dft);
 
-            llama_sampling_accept(ctx_sampling, ctx_tgt, id, true);
+            // remember which tokens were sampled - used for repetition penalties during sampling
+            last_tokens.erase(last_tokens.begin());
+            last_tokens.push_back(id);
 
-            //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, ctx_sampling->prev).c_str());
+            //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, last_tokens));
 
             const std::string token_str = llama_token_to_piece(ctx_tgt, id);
-
             printf("%s", token_str.c_str());
             fflush(stdout);
 
-            if (id == llama_token_eos(model_tgt)) {
+            if (id == llama_token_eos(ctx_tgt)) {
                 has_eos = true;
             }
 
             ++n_predict;
 
-            // check if the target token matches any of the drafts
-            {
-                bool matches = false;
+            // check if the draft matches the target
+            if (i_dft < (int) drafted.size() && id == drafted[i_dft]) {
+                LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
+                ++n_accept;
+                ++n_past_tgt;
+                ++n_past_dft;
+                ++i_dft;
 
-                for (int s = 0; s < n_seq_dft; ++s) {
-                    if (!drafts[s].active) {
-                        continue;
-                    }
-
-                    if (i_dft < (int) drafts[s].tokens.size() && id == drafts[s].tokens[i_dft]) {
-                        LOG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, id, token_str.c_str());
-
-                        s_keep = s;
-                        matches = true;
-                    } else {
-                        drafts[s].active = false;
-                    }
-                }
-
-                if (matches) {
-                    ++n_accept;
-                    ++n_past_tgt;
-                    ++n_past_dft;
-                    ++i_dft;
-
-                    continue;
-                }
+                continue;
             }
 
-            LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
+            // the drafted token was rejected or we are out of drafted tokens
 
-            // TODO: simplify
-            {
-                LOG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
-
-                llama_kv_cache_seq_keep(ctx_dft, s_keep);
-                llama_kv_cache_seq_cp  (ctx_dft, s_keep, 0, -1, -1);
-                llama_kv_cache_seq_keep(ctx_dft, 0);
-
-                llama_kv_cache_seq_rm  (ctx_tgt, s_keep, n_past_tgt, -1);
-                llama_kv_cache_seq_keep(ctx_tgt, s_keep);
-                llama_kv_cache_seq_cp  (ctx_tgt, s_keep, 0, -1, -1);
-                llama_kv_cache_seq_keep(ctx_tgt, 0);
+            if (i_dft < (int) drafted.size()) {
+                LOG("the %dth drafted token (%d, '%s') does not match the sampled target token (%d, '%s') - rejected\n",
+                        i_dft, drafted[i_dft], llama_token_to_piece(ctx_dft, drafted[i_dft]).c_str(), id, token_str.c_str());
+            } else {
+                LOG("out of drafted tokens\n");
             }
 
-            for (int s = 0; s < n_seq_dft; ++s) {
-                drafts[s].active = false;
-                drafts[s].tokens.clear();
-                drafts[s].i_batch_tgt.clear();
-            }
-            // note: will be erased after the speculation phase
-            drafts[0].tokens.push_back(id);
-            drafts[0].i_batch_tgt.push_back(0);
-
-            llama_batch_clear(batch_dft);
-            llama_batch_add  (batch_dft, id, n_past_dft, { 0 }, true);
-
             llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
-            // LOG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
-            llama_decode         (ctx_dft, batch_dft);
-
+            llama_decode(ctx_dft, llama_batch_get_one(&id, 1, n_past_dft, 0));
             ++n_past_dft;
 
+            // heuristic for n_draft
+            {
+                const int  n_draft_cur  = (int) drafted.size();
+                const bool all_accepted = i_dft == n_draft_cur;
+
+                LOG("n_draft      = %d\n", n_draft);
+                LOG("n_draft_cur  = %d\n", n_draft_cur);
+                LOG("i_dft        = %d\n", i_dft);
+                LOG("all_accepted = %d\n", all_accepted);
+
+                if (all_accepted && n_draft == n_draft_cur) {
+                    LOG(" - max drafted tokens accepted - n_draft += 8\n");
+                    n_draft = std::min(30, n_draft + 8);
+                } else if (all_accepted) {
+                    LOG(" - partially drafted tokens accepted - no change\n");
+                } else {
+                    LOG(" - drafted token rejected - n_draft -= 1\n");
+                    n_draft = std::max(2, n_draft - 1);
+                }
+            }
+
+            drafted.clear();
+            drafted.push_back(id);
+
             break;
         }
 
@@ -269,151 +207,72 @@ int main(int argc, char ** argv) {
             break;
         }
 
-        llama_sampling_cp(ctx_sampling, drafts[0].ctx_sampling);
+        if (grammar_tgt) {
+            if (grammar_dft) {
+                llama_grammar_free(grammar_dft);
+            }
+            grammar_dft = llama_grammar_copy(grammar_tgt);
 
-        int n_seq_cur  = 1;
-        int n_past_cur = n_past_dft;
-
-        for (int s = 0; s < n_seq_dft; ++s) {
-            drafts[s].active   = false;
-            drafts[s].drafting = false;
+            LOG("copied target grammar to draft grammar\n");
         }
-        drafts[0].active      = true;
-        drafts[0].drafting    = true;
-        drafts[0].i_batch_dft = 0;
 
-        llama_batch_clear(batch_tgt);
-        llama_batch_add  (batch_tgt, drafts[0].tokens[0], n_past_tgt, { 0 }, true);
-
-        // sample n_draft tokens from the draft model using tree-based sampling
+        // sample n_draft tokens from the draft model using greedy decoding
+        int n_past_cur = n_past_dft;
         for (int i = 0; i < n_draft; ++i) {
-            batch_dft.n_tokens = 0;
+            float * logits = llama_get_logits(ctx_dft);
 
-            for (int s = 0; s < n_seq_dft; ++s) {
-                drafts[s].skip = false;
+            candidates.clear();
+            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});
             }
 
-            for (int s = 0; s < n_seq_dft; ++s) {
-                if (!drafts[s].drafting || drafts[s].skip) {
-                    continue;
-                }
+            llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
 
-                llama_sampling_sample(drafts[s].ctx_sampling, ctx_dft, NULL, drafts[s].i_batch_dft);
-
-                const auto & cur_p = drafts[s].ctx_sampling->cur;
-
-                for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p.size()); ++k) {
-                    LOG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
-                            k, s, i, cur_p[k].id, cur_p[k].p, llama_token_to_piece(ctx_dft, cur_p[k].id).c_str());
-                }
-
-                if (cur_p[0].p < p_accept) {
-                    LOG("stopping drafting for seq %3d, probability too low: %.3f < %.3f\n", s, cur_p[0].p, p_accept);
-                    drafts[s].drafting = false;
-                    continue;
-                }
-
-                std::vector<int> sa(1, s);
-
-                // attempt to split the branch if the probability is high enough
-                for (int f = 1; f < 8; ++f) {
-                    if (n_seq_cur < n_seq_dft && cur_p[f].p > p_split) {
-                        LOG("splitting seq %3d into %3d\n", s, n_seq_cur);
-
-                        llama_kv_cache_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
-                        llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
-
-                        // all previous tokens from this branch are now also part of the new branch
-                        for (int t = 0; t < batch_tgt.n_tokens; ++t) {
-                            for (int p = 0; p < batch_tgt.n_seq_id[t]; ++p) {
-                                if (batch_tgt.seq_id[t][p] == s) {
-                                    batch_tgt.seq_id[t][batch_tgt.n_seq_id[t]] = n_seq_cur;
-                                    batch_tgt.n_seq_id[t]++;
-                                    break;
-                                }
-                            }
-                        }
-
-                        // copy the draft state
-                        drafts[n_seq_cur].active   = true;
-                        drafts[n_seq_cur].drafting = true;
-                        drafts[n_seq_cur].skip     = true;
-
-                        drafts[n_seq_cur].tokens      = drafts[s].tokens;
-                        drafts[n_seq_cur].i_batch_dft = drafts[s].i_batch_dft;
-                        drafts[n_seq_cur].i_batch_tgt = drafts[s].i_batch_tgt;
-
-                        llama_sampling_cp(drafts[s].ctx_sampling, drafts[n_seq_cur].ctx_sampling);
-
-                        sa.push_back(n_seq_cur);
-
-                        n_seq_cur++;
-                    } else {
-                        break;
-                    }
-                }
-
-                // add drafted token for each sequence
-                for (int is = 0; is < (int) sa.size(); ++is) {
-                    const llama_token id = cur_p[is].id;
-
-                    const int s = sa[is];
-
-                    llama_sampling_accept(drafts[s].ctx_sampling, ctx_dft, id, true);
-
-                    drafts[s].tokens.push_back(id);
-
-                    // add unique drafted tokens to the target batch
-                    drafts[s].i_batch_tgt.push_back(batch_tgt.n_tokens);
-
-                    llama_batch_add(batch_tgt, id, n_past_tgt + i + 1, { s }, true);
-
-                    // add the token to the batch for batched decoding with the draft model
-                    drafts[s].i_batch_dft = batch_dft.n_tokens;
-
-                    llama_batch_add(batch_dft, id, n_past_cur, { s }, true);
-
-                    if (batch_tgt.n_tokens > n_draft) {
-                        drafts[s].drafting = false;
-                    }
-                }
+            if (grammar_dft != NULL) {
+                llama_sample_grammar(ctx_dft, &cur_p, grammar_dft);
             }
 
-            // no sequence is drafting anymore
-            if (batch_dft.n_tokens == 0) {
+            // computes softmax and sorts the candidates
+            llama_sample_softmax(ctx_dft, &cur_p);
+
+            for (int i = 0; i < 3; ++i) {
+                LOG(" - draft candidate %3d: %6d (%8.3f) '%s'\n", i, cur_p.data[i].id, cur_p.data[i].p, llama_token_to_piece(ctx_dft, cur_p.data[i].id).c_str());
+            }
+
+            // TODO: better logic?
+            if (cur_p.data[0].p < 2*cur_p.data[1].p) {
+                LOG("stopping drafting, probability too low: %.3f < 2*%.3f\n", cur_p.data[0].p, cur_p.data[1].p);
                 break;
             }
 
-            // evaluate the drafted tokens on the draft model
-            llama_decode(ctx_dft, batch_dft);
-            ++n_past_cur;
+            // drafted token
+            const llama_token id = cur_p.data[0].id;
+
+            drafted.push_back(id);
             ++n_drafted;
 
-            if (batch_tgt.n_tokens > n_draft) {
+            // no need to evaluate the last drafted token, since we won't use the result
+            if (i == n_draft - 1) {
                 break;
             }
+
+            // evaluate the drafted token on the draft model
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_cur, -1);
+            llama_decode(ctx_dft, llama_batch_get_one(&drafted.back(), 1, n_past_cur, 0));
+            ++n_past_cur;
+
+            if (grammar_dft != NULL) {
+                llama_grammar_accept_token(ctx_dft, grammar_dft, id);
+            }
         }
 
         // evaluate the target model on the drafted tokens
-        {
-            llama_kv_cache_seq_keep(ctx_tgt, 0);
-            for (int s = 1; s < n_seq_dft; ++s) {
-                llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1);
-            }
+        llama_kv_cache_seq_rm(ctx_tgt, 0, n_past_tgt, -1);
+        llama_decode(ctx_tgt, llama_batch_get_one(drafted.data(), drafted.size(), n_past_tgt, 0));
+        ++n_past_tgt;
 
-            // LOG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());
-            llama_decode(ctx_tgt, batch_tgt);
-            ++n_past_tgt;
-        }
-
-        // the first token is always proposed by the traget model before the speculation loop so we erase it here
-        for (int s = 0; s < n_seq_dft; ++s) {
-            if (!drafts[s].active) {
-                continue;
-            }
-
-            drafts[s].tokens.erase(drafts[s].tokens.begin());
-        }
+        // the first token is always proposed by the traget model before the speculation loop
+        drafted.erase(drafted.begin());
     }
 
     auto t_dec_end = ggml_time_us();
@@ -421,8 +280,9 @@ int main(int argc, char ** argv) {
     LOG_TEE("\n\n");
 
     LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
-    LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
+    LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
 
+    // TODO: make sure these numbers are computed correctly
     LOG_TEE("\n");
     LOG_TEE("n_draft   = %d\n", n_draft);
     LOG_TEE("n_predict = %d\n", n_predict);
@@ -436,19 +296,16 @@ int main(int argc, char ** argv) {
     LOG_TEE("\ntarget:\n");
     llama_print_timings(ctx_tgt);
 
-    llama_sampling_free(ctx_sampling);
-    for (int s = 0; s < n_seq_dft; ++s) {
-        llama_sampling_free(drafts[s].ctx_sampling);
-    }
-
-    llama_batch_free(batch_dft);
-
     llama_free(ctx_tgt);
     llama_free_model(model_tgt);
 
     llama_free(ctx_dft);
     llama_free_model(model_dft);
 
+    if (grammar_dft != NULL) {
+        llama_grammar_free(grammar_dft);
+        llama_grammar_free(grammar_tgt);
+    }
     llama_backend_free();
 
     fprintf(stderr, "\n\n");

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

@@ -253,14 +253,13 @@ static void init_model(struct my_llama_model * model) {
     set_param_model(model);
 
     // measure data size
-    size_t size = 0;
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        size += GGML_PAD(ggml_nbytes(t), tensor_alignment);
-    }
+    struct ggml_allocr * alloc = NULL;
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    alloc_model(alloc, model);
 
     // allocate data
-    struct ggml_allocr * alloc = NULL;
-    model->data.resize(size + tensor_alignment);
+    model->data.resize(ggml_allocr_max_size(alloc) + tensor_alignment);
+    ggml_allocr_free(alloc);
     alloc = ggml_allocr_new(model->data.data(), model->data.size(), tensor_alignment);
     alloc_model(alloc, model);
     ggml_allocr_free(alloc);
@@ -1095,9 +1094,11 @@ int main(int argc, char ** argv) {
     struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
 
     // measure required memory for input tensors
-    size_t max_input_size = GGML_PAD(ggml_nbytes(tokens_input), tensor_alignment) +
-                            GGML_PAD(ggml_nbytes(target_probs), tensor_alignment) +
-                            tensor_alignment;
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    ggml_allocr_alloc(alloc, tokens_input);
+    ggml_allocr_alloc(alloc, target_probs);
+    size_t max_input_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+    ggml_allocr_free(alloc);
     printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
 
     // allocate input tensors

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1698134075,
-        "narHash": "sha256-foCD+nuKzfh49bIoiCBur4+Fx1nozo+4C/6k8BYk4sg=",
+        "lastModified": 1692913444,
+        "narHash": "sha256-1SvMQm2DwofNxXVtNWWtIcTh7GctEVrS/Xel/mdc6iY=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "8efd5d1e283604f75a808a20e6cde0ef313d07d4",
+        "rev": "18324978d632ffc55ef1d928e81630c620f4f447",
         "type": "github"
       },
       "original": {

flake.nix

@@ -51,9 +51,6 @@
         };
         llama-python =
           pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
-        # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
-        llama-python-extra =
-          pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece torchWithoutCuda transformers ]);
         postPatch = ''
           substituteInPlace ./ggml-metal.m \
             --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
@@ -129,9 +126,5 @@
           buildInputs = [ llama-python ];
           packages = nativeBuildInputs ++ osSpecific;
         };
-        devShells.extra = pkgs.mkShell {
-          buildInputs = [ llama-python-extra ];
-          packages = nativeBuildInputs ++ osSpecific;
-        };
       });
 }

ggml-alloc.c

@@ -1,5 +1,4 @@
 #include "ggml-alloc.h"
-#include "ggml-backend.h"
 #include "ggml.h"
 #include <assert.h>
 #include <stdarg.h>
@@ -7,6 +6,25 @@
 #include <stdlib.h>
 #include <string.h>
 
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/types.h>
+            #include <sys/mman.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #include <memoryapi.h>
+#endif
+
 
 #define UNUSED(x) (void)(x)
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
@@ -62,9 +80,8 @@ struct free_block {
 #define MAX_FREE_BLOCKS 256
 
 struct ggml_allocr {
-    struct ggml_backend_buffer * buffer;
-    bool buffer_owned;
     void * data;
+    size_t size;
     size_t alignment;
     int n_free_blocks;
     struct free_block free_blocks[MAX_FREE_BLOCKS];
@@ -102,9 +119,16 @@ static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tens
 }
 #endif
 
+static size_t ggml_allocr_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    return ggml_nbytes(tensor);
+
+    UNUSED(alloc);
+}
+
 // check if a tensor is allocated by this buffer
 static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_tensor * tensor) {
-    return tensor->buffer == alloc->buffer;
+    void * ptr = tensor->data;
+    return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size;
 }
 
 static bool ggml_is_view(struct ggml_tensor * t) {
@@ -112,10 +136,11 @@ static bool ggml_is_view(struct ggml_tensor * t) {
 }
 
 void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+#ifdef GGML_ALLOCATOR_DEBUG
     GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources
     GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated
-
-    size_t size = ggml_backend_buffer_get_alloc_size(alloc->buffer, tensor);
+#endif
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
 
     AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
@@ -163,8 +188,6 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
 
     tensor->data = addr;
     AT_PRINTF("%s: allocated data at %p\n", __func__, tensor->data);
-    tensor->buffer = alloc->buffer;
-    ggml_backend_buffer_init_tensor(alloc->buffer, tensor);
 
 #ifdef GGML_ALLOCATOR_DEBUG
     add_allocated_tensor(alloc, tensor);
@@ -185,21 +208,19 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
 
 // this is a very naive implementation, but for our case the number of free blocks should be very small
 static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    void * ptr = tensor->data;
+
     if (ggml_allocr_is_own(alloc, tensor) == false) {
         // the tensor was not allocated in this buffer
         // this can happen because the graph allocator will try to free weights and other tensors from different buffers
         // the easiest way to deal with this is just to ignore it
-        AT_PRINTF("ignoring %s (their buffer: %p, our buffer: %p)\n", tensor->name, (void *)tensor->buffer, (void *)alloc->buffer);
         return;
     }
 
-    void * ptr = tensor->data;
-
-    size_t size = ggml_backend_buffer_get_alloc_size(alloc->buffer, tensor);
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
     AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks);
-
-    ggml_backend_buffer_free_tensor(alloc->buffer, tensor);
+    AT_PRINTF("%s: alloc->data = %p alloc->data+alloc->size = %p alloc->data+alloc->max_size = %p\n", __func__, alloc->data, (char*)alloc->data + alloc->size, (char*)alloc->data + alloc->max_size);
 
 #ifdef GGML_ALLOCATOR_DEBUG
     remove_allocated_tensor(alloc, tensor);
@@ -264,18 +285,15 @@ void ggml_allocr_reset(struct ggml_allocr * alloc) {
     alloc->n_free_blocks = 1;
     size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
     alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
-    alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset;
+    alloc->free_blocks[0].size = alloc->size - align_offset;
 }
 
 struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
-    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(NULL, data, size);
-
-    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr));
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
 
     *alloc = (struct ggml_allocr){
-        /*.buffer        = */ buffer,
-        /*.buffer_owned  = */ true,
-        /*.base          = */ ggml_backend_buffer_get_base(buffer),
+        /*.data          = */ data,
+        /*.size          = */ size,
         /*.alignment     = */ alignment,
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
@@ -294,26 +312,74 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
     return alloc;
 }
 
-struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
-    struct ggml_allocr * alloc = ggml_allocr_new((void *)0x1000, (size_t)-0x1001, alignment);
-    alloc->measure = true;
-
-    return alloc;
+// OS specific functions to allocate and free uncommitted virtual memory
+static void * alloc_vmem(size_t size) {
+#if defined(_WIN32)
+    return VirtualAlloc(NULL, size, MEM_RESERVE, PAGE_NOACCESS);
+#elif defined(_POSIX_MAPPED_FILES)
+    void * ptr = mmap(NULL, size, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0);
+    if (ptr == MAP_FAILED) {
+        return NULL;
+    }
+    return ptr;
+#else
+    // use a fixed address for other platforms
+    uintptr_t base_addr = (uintptr_t)-size - 0x100;
+    return (void *)base_addr;
+#endif
 }
 
-struct ggml_allocr * ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer) {
-    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr));
+static void free_vmem(void * base_addr, size_t size) {
+#if defined(_WIN32)
+    VirtualFree(base_addr, 0, MEM_RELEASE);
+    UNUSED(size);
+#elif defined(_POSIX_MAPPED_FILES)
+    munmap(base_addr, size);
+#else
+    // nothing to do
+    UNUSED(base_addr);
+    UNUSED(size);
+#endif
+}
+
+// allocate uncommitted virtual memory to measure the size of the graph
+static void alloc_measure_vmem(void ** base_addr, size_t * size) {
+    // 128GB for 64-bit, 1GB for 32-bit
+    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<37;
+    do {
+        *base_addr = alloc_vmem(*size);
+        if (*base_addr != NULL) {
+            AT_PRINTF("allocated %.2f GB of virtual memory for measure buffer at %p\n", *size / 1024.0 / 1024.0 / 1024.0, *base_addr);
+            return;
+        }
+        // try again with half the size
+        *size /= 2;
+    } while (*size > 0);
+
+    GGML_ASSERT(!"failed to allocate virtual memory for measure buffer");
+}
+
+static void free_measure_vmem(void * base_addr, size_t size) {
+    free_vmem(base_addr, size);
+}
+
+struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+
+    void * base_addr;
+    size_t size;
+
+    alloc_measure_vmem(&base_addr, &size);
 
     *alloc = (struct ggml_allocr){
-        /*.buffer        = */ buffer,
-        /*.buffer_owned  = */ false,
-        /*.base          = */ ggml_backend_buffer_get_base(buffer),
-        /*.alignment     = */ ggml_backend_buffer_get_alignment(buffer),
+        /*.data          = */ base_addr,
+        /*.size          = */ size,
+        /*.alignment     = */ alignment,
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
         /*.hash_table    = */ {{0}},
         /*.max_size      = */ 0,
-        /*.measure       = */ false,
+        /*.measure       = */ true,
         /*.parse_seq     = */ {0},
         /*.parse_seq_len = */ 0,
 #ifdef GGML_ALLOCATOR_DEBUG
@@ -327,8 +393,8 @@ struct ggml_allocr * ggml_allocr_new_from_buffer(struct ggml_backend_buffer * bu
 }
 
 void ggml_allocr_free(struct ggml_allocr * alloc) {
-    if (alloc->buffer_owned) {
-        ggml_backend_buffer_free(alloc->buffer);
+    if (alloc->measure) {
+        free_measure_vmem(alloc->data, alloc->size);
     }
     free(alloc);
 }
@@ -371,6 +437,7 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_ROPE:
         case GGML_OP_RMS_NORM:
         case GGML_OP_SOFT_MAX:
+        case GGML_OP_CONT:
             return true;
 
         default:
@@ -378,23 +445,12 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
     }
 }
 
-static void init_view(struct ggml_allocr * alloc, struct ggml_tensor * view) {
-    assert(view->view_src != NULL && view->view_src->data != NULL);
-    view->backend = view->view_src->backend;
-    view->buffer  = view->view_src->buffer;
-    view->data    = (char *)view->view_src->data + view->view_offs;
-
-    // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
-    // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
-    assert(ggml_allocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend);
-    ggml_backend_buffer_init_tensor(alloc->buffer, view);
-}
-
 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);
+            assert(node->view_src->data != NULL);
+            node->data = (char *)node->view_src->data + node->view_offs;
         } else {
             // see if we can reuse a parent's buffer (inplace)
             if (ggml_op_can_inplace(node->op)) {
@@ -422,17 +478,13 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                                 // adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
                                 // for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
                                 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);
+                                node->data = parent->data;
                                 return;
                             }
                         }
                         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);
+                            node->data = parent->data;
                             return;
                         }
                     }
@@ -443,7 +495,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     }
 }
 
-size_t ggml_allocr_alloc_graph_n(
+static size_t ggml_allocr_alloc_graph_tensors_n(
     struct ggml_allocr * alloc,
     struct ggml_cgraph ** graphs, int n_graphs,
     struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
@@ -461,10 +513,6 @@ size_t ggml_allocr_alloc_graph_n(
             if (ggml_is_view(node)) {
                 struct ggml_tensor * view_src = node->view_src;
                 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);
-                }
             }
 
             for (int j = 0; j < GGML_MAX_SRC; j++) {
@@ -473,9 +521,6 @@ size_t ggml_allocr_alloc_graph_n(
                     break;
                 }
                 hash_get(ht, parent)->n_children += 1;
-                if (ggml_is_view(parent) && parent->buffer == NULL && parent->data != NULL) {
-                    init_view(alloc, parent);
-                }
             }
         }
     }
@@ -586,7 +631,7 @@ size_t ggml_allocr_alloc_graph_n(
 }
 
 size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
-    return ggml_allocr_alloc_graph_n(alloc, &graph, 1, NULL, NULL);
+    return ggml_allocr_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
 }
 
 size_t ggml_allocr_max_size(struct ggml_allocr * alloc) {

ggml-alloc.h

@@ -6,27 +6,21 @@
 extern "C" {
 #endif
 
-struct ggml_backend_buffer;
 
 GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
 GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
-GGML_API struct ggml_allocr * ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer);
 
 // tell the allocator to parse nodes following the order described in the list
 // you should call this if your graph are optimized to execute out-of-order
 GGML_API void   ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n);
 
-GGML_API void   ggml_allocr_free       (struct ggml_allocr * alloc);
-GGML_API bool   ggml_allocr_is_measure (struct ggml_allocr * alloc);
-GGML_API void   ggml_allocr_reset      (struct ggml_allocr * alloc);
-GGML_API void   ggml_allocr_alloc      (struct ggml_allocr * alloc, struct ggml_tensor * tensor);
+GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
+GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
 GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
-GGML_API size_t ggml_allocr_max_size   (struct ggml_allocr * alloc);
+GGML_API size_t ggml_allocr_max_size(struct ggml_allocr * alloc);
 
-GGML_API size_t ggml_allocr_alloc_graph_n(
-                    struct ggml_allocr * alloc,
-                    struct ggml_cgraph ** graphs, int n_graphs,
-                    struct ggml_tensor *** inputs, struct ggml_tensor *** outputs);
 
 #ifdef  __cplusplus
 }

ggml-backend.c

@@ -1,385 +0,0 @@
-#include "ggml-backend.h"
-#include "ggml-alloc.h"
-
-#include <assert.h>
-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-
-#define UNUSED GGML_UNUSED
-
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-
-// backend buffer
-
-ggml_backend_buffer_t ggml_backend_buffer_init(
-        struct ggml_backend                  * backend,
-        struct ggml_backend_buffer_i           iface,
-               ggml_backend_buffer_context_t   context,
-               size_t                          size) {
-    ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
-
-    GGML_ASSERT(iface.get_base != NULL);
-
-    (*buffer) = (struct ggml_backend_buffer) {
-        /* .interface = */ iface,
-        /* .backend   = */ backend,
-        /* .context   = */ context,
-        /* .size      = */ size,
-    };
-
-    return buffer;
-}
-
-void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
-    if (buffer->iface.free_buffer != NULL) {
-        buffer->iface.free_buffer(buffer);
-    }
-    free(buffer);
-}
-
-size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
-    return ggml_backend_get_alignment(buffer->backend);
-}
-
-void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return buffer->iface.get_base(buffer);
-}
-
-size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
-    return buffer->size;
-}
-
-size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    if (buffer->iface.get_alloc_size) {
-        return buffer->iface.get_alloc_size(buffer, tensor);
-    }
-    return ggml_nbytes(tensor);
-}
-
-void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    if (buffer->iface.init_tensor) {
-        buffer->iface.init_tensor(buffer, tensor);
-    }
-}
-
-void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    if (buffer->iface.free_tensor) {
-        buffer->iface.free_tensor(buffer, tensor);
-    }
-}
-
-// backend
-
-ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) {
-    return tensor->buffer->backend;
-}
-
-const char * ggml_backend_name(ggml_backend_t backend) {
-    return backend->iface.get_name(backend);
-}
-
-void ggml_backend_free(ggml_backend_t backend) {
-    backend->iface.free(backend);
-}
-
-ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
-    return backend->iface.alloc_buffer(backend, size);
-}
-
-size_t ggml_backend_get_alignment(ggml_backend_t backend) {
-    return backend->iface.get_alignment(backend);
-}
-
-void ggml_backend_tensor_set_async(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
-}
-
-void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
-}
-
-void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
-    ggml_get_backend(tensor)->iface.synchronize(ggml_get_backend(tensor));
-}
-
-void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
-    ggml_get_backend(tensor)->iface.synchronize(ggml_get_backend(tensor));
-}
-
-void ggml_backend_synchronize(ggml_backend_t backend) {
-    backend->iface.synchronize(backend);
-}
-
-ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    return backend->iface.graph_plan_create(backend, cgraph);
-}
-
-void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    backend->iface.graph_plan_free(backend, plan);
-}
-
-void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    backend->iface.graph_plan_compute(backend, plan);
-}
-
-void ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    backend->iface.graph_compute(backend, cgraph);
-}
-
-bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return backend->iface.supports_op(backend, op);
-}
-
-// backend copy
-
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
-void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
-    //printf("src: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", src->name, (int)src->ne[0], (int)src->ne[1], (int)src->ne[2], (int)src->ne[3], (int)src->nb[0], (int)src->nb[1], (int)src->nb[2], (int)src->nb[3]);
-    //printf("dst: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", dst->name, (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], (int)dst->nb[0], (int)dst->nb[1], (int)dst->nb[2], (int)dst->nb[3]);
-    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
-
-    // printf("cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src));
-
-    if (src == dst) {
-        return;
-    }
-
-    // TODO: allow backends to support copy to/from same backend
-
-    if (ggml_get_backend(dst)->iface.cpy_tensor_from != NULL) {
-        ggml_get_backend(dst)->iface.cpy_tensor_from(ggml_get_backend(dst)->context, src, dst);
-    } else if (ggml_get_backend(src)->iface.cpy_tensor_to != NULL) {
-        ggml_get_backend(src)->iface.cpy_tensor_to(ggml_get_backend(src)->context, src, dst);
-    } else {
-        // shouldn't be hit when copying from/to CPU
-        #ifndef NDEBUG
-        fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to are implemented for backends %s and %s, falling back to get/set\n", ggml_backend_name(src->buffer->backend), ggml_backend_name(dst->buffer->backend));
-        #endif
-        size_t nbytes = ggml_nbytes(src);
-        void * data = malloc(nbytes);
-        ggml_backend_tensor_get(src, data, 0, nbytes);
-        ggml_backend_tensor_set(dst, data, 0, nbytes);
-        free(data);
-    }
-}
-
-// backend CPU
-
-struct ggml_backend_cpu_context {
-    int n_threads;
-    void * work_data;
-    size_t work_size;
-};
-
-static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
-    return "CPU";
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_free(ggml_backend_t backend) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-    free(cpu_ctx->work_data);
-    free(cpu_ctx);
-    free(backend);
-}
-
-static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return (void *)buffer->context;
-}
-
-static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    free(buffer->context);
-    UNUSED(buffer);
-}
-
-static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
-    /* .free_buffer    = */ ggml_backend_cpu_buffer_free_buffer,
-    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
-    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
-    /* .init_tensor    = */ NULL, // no initialization required
-    /* .free_tensor    = */ NULL, // no cleanup required
-};
-
-// for buffers from ptr, free is not called
-static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
-    /* .free_buffer    = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
-    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
-    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
-    /* .init_tensor    = */ NULL,
-    /* .free_tensor    = */ NULL,
-};
-
-static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
-
-static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backend, size_t size) {
-    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
-    void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC?
-
-    return ggml_backend_buffer_init(backend, cpu_backend_buffer_i, data, size);
-}
-
-static size_t ggml_backend_cpu_get_alignment(ggml_backend_t backend) {
-    return TENSOR_ALIGNMENT;
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_set_tensor_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
-    memcpy((char *)tensor->data + offset, data, size);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_get_tensor_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
-    memcpy(data, (const char *)tensor->data + offset, size);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_synchronize(ggml_backend_t backend) {
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    // for a backend such as CUDA that can queue async calls, it is ok to do this asynchronously, but it may not be the case for other backends
-    ggml_backend_tensor_set_async(dst, src->data, 0, ggml_nbytes(src));
-
-    UNUSED(backend);
-}
-
-struct ggml_backend_plan_cpu {
-    struct ggml_cplan cplan;
-    struct ggml_cgraph cgraph;
-};
-
-static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
-
-    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-    cpu_plan->cgraph = *cgraph;
-
-    if (cpu_plan->cplan.work_size > 0) {
-        cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
-    }
-
-    return cpu_plan;
-}
-
-static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    free(cpu_plan->cplan.work_data);
-    free(cpu_plan);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-
-    if (cpu_ctx->work_size < cplan.work_size) {
-        // TODO: may be faster to free and use malloc to avoid the copy
-        cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size);
-        cpu_ctx->work_size = cplan.work_size;
-    }
-
-    cplan.work_data = cpu_ctx->work_data;
-
-    ggml_graph_compute(cgraph, &cplan);
-}
-
-static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return true;
-    UNUSED(backend);
-    UNUSED(op);
-}
-
-static struct ggml_backend_i cpu_backend_i = {
-    /* .get_name            = */ ggml_backend_cpu_name,
-    /* .free                = */ ggml_backend_cpu_free,
-    /* .alloc_buffer        = */ ggml_backend_cpu_alloc_buffer,
-    /* .get_alignment       = */ ggml_backend_cpu_get_alignment,
-    /* .set_tensor_async    = */ ggml_backend_cpu_set_tensor_async,
-    /* .get_tensor_async    = */ ggml_backend_cpu_get_tensor_async,
-    /* .synchronize         = */ ggml_backend_cpu_synchronize,
-    /* .cpy_tensor_from     = */ ggml_backend_cpu_cpy_tensor_from,
-    /* .cpy_tensor_to       = */ ggml_backend_cpu_cpy_tensor_to,
-    /* .graph_plan_create   = */ ggml_backend_cpu_graph_plan_create,
-    /* .graph_plan_free     = */ ggml_backend_cpu_graph_plan_free,
-    /* .graph_plan_compute  = */ ggml_backend_cpu_graph_plan_compute,
-    /* .graph_compute       = */ ggml_backend_cpu_graph_compute,
-    /* .supports_op         = */ ggml_backend_cpu_supports_op,
-};
-
-ggml_backend_t ggml_backend_cpu_init(void) {
-    struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
-
-    ctx->n_threads = GGML_DEFAULT_N_THREADS;
-    ctx->work_data = NULL;
-    ctx->work_size = 0;
-
-    ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
-
-    *cpu_backend = (struct ggml_backend) {
-        /* .interface = */ cpu_backend_i,
-        /* .context   = */ ctx
-    };
-    return cpu_backend;
-}
-
-bool ggml_backend_is_cpu(ggml_backend_t backend) {
-    return backend->iface.get_name == ggml_backend_cpu_name;
-}
-
-void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-    ctx->n_threads = n_threads;
-}
-
-ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size) {
-    return ggml_backend_buffer_init(backend_cpu, cpu_backend_buffer_i_from_ptr, ptr, size);
-}

ggml-backend.h

@@ -1,143 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-    struct ggml_backend;
-    struct ggml_backend_buffer;
-
-    // type-erased backend-specific types / wrappers
-    typedef void * ggml_backend_context_t;
-    typedef void * ggml_backend_graph_plan_t;
-    typedef void * ggml_backend_buffer_context_t;
-
-    // avoid accessing internals of these types
-    typedef struct ggml_backend        * ggml_backend_t;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-
-    //
-    // backend buffer
-    //
-
-    struct ggml_backend_buffer_i {
-        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
-        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
-        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
-        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
-        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
-    };
-
-    // TODO: hide behind API
-    struct ggml_backend_buffer {
-        struct ggml_backend_buffer_i iface;
-
-        ggml_backend_t                backend;
-        ggml_backend_buffer_context_t context;
-
-        size_t size;
-    };
-
-    // backend buffer functions
-    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
-            struct ggml_backend                  * backend,
-            struct ggml_backend_buffer_i           iface,
-                   ggml_backend_buffer_context_t   context,
-                   size_t                          size);
-
-    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-
-    //
-    // backend
-    //
-
-    struct ggml_backend_i {
-        const char * (*get_name)(ggml_backend_t backend);
-
-        void (*free)(ggml_backend_t backend);
-
-        // buffer allocation
-        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
-
-        // get buffer alignment
-        size_t (*get_alignment)(ggml_backend_t backend);
-
-        // tensor data access
-        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
-        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        void (*synchronize)     (ggml_backend_t backend);
-
-        // (optional) copy tensor between different backends, allow for single-copy tranfers
-        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-
-        // compute graph with a plan
-        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-        void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-
-        // compute graph without a plan
-        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-        // check if the backend supports an operation
-        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-    };
-
-    // TODO: hide behind API
-    struct ggml_backend {
-        struct ggml_backend_i iface;
-
-        ggml_backend_context_t context;
-    };
-
-    // backend helper functions
-    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
-
-    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
-    GGML_API void         ggml_backend_free(ggml_backend_t backend);
-
-    GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
-
-    GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
-
-    GGML_API void ggml_backend_tensor_set_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
-
-    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
-
-    // tensor copy between different backends
-    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    //
-    // CPU backend
-    //
-
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
-
-    GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend);
-
-    GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
-
-    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
-
-#ifdef  __cplusplus
-}
-#endif

ggml-cuda.h

@@ -1,7 +1,6 @@
 #pragma once
 
 #include "ggml.h"
-#include "ggml-backend.h"
 
 #ifdef GGML_USE_HIPBLAS
 #define GGML_CUDA_NAME "ROCm"
@@ -43,9 +42,6 @@ GGML_API bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, s
 GGML_API int    ggml_cuda_get_device_count(void);
 GGML_API void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
 
-// backend API
-GGML_API ggml_backend_t ggml_backend_cuda_init(void); // TODO: take a list of devices to use
-
 #ifdef  __cplusplus
 }
 #endif

ggml-impl.h

@@ -1,237 +0,0 @@
-#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.h

@@ -20,7 +20,6 @@
 #pragma once
 
 #include "ggml.h"
-#include "ggml-backend.h"
 
 #include <stddef.h>
 #include <stdbool.h>
@@ -36,15 +35,10 @@ struct ggml_cgraph;
 extern "C" {
 #endif
 
-//
-// internal API
-// temporary exposed to user-code
-//
+void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
 
 struct ggml_metal_context;
 
-void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
-
 // number of command buffers to use
 struct ggml_metal_context * ggml_metal_init(int n_cb);
 void ggml_metal_free(struct ggml_metal_context * ctx);
@@ -89,17 +83,6 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
 // creates gf->n_threads command buffers in parallel
 void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
 
-//
-// backend API
-// user-code should use only these functions
-//
-
-GGML_API ggml_backend_t ggml_backend_metal_init(void);
-
-GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
-
-GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
-
 #ifdef __cplusplus
 }
 #endif