Upload folder using huggingface_hub

.gitignore

@@ -12,6 +12,9 @@
 .vs/
 .vscode/
 
+lcov-report/
+gcovr-report/
+
 build*/
 out/
 tmp/
@@ -19,24 +22,34 @@ tmp/
 models/*
 models-mnt
 
-/main
-/quantize
-/quantize-stats
-/result
-/perplexity
-/embedding
-/train-text-from-scratch
-/convert-llama2c-to-ggml
-/simple
-/benchmark-matmult
-/vdot
-/server
 /Pipfile
+/baby-llama
+/beam-search
+/benchmark-matmult
+/convert-llama2c-to-ggml
 /embd-input-test
+/embedding
 /gguf
 /gguf-llama-simple
 /libllama.so
 /llama-bench
+/main
+/metal
+/perplexity
+/q8dot
+/quantize
+/quantize-stats
+/result
+/save-load-state
+/server
+/simple
+/batched
+/export-lora
+/finetune
+/speculative
+/parallel
+/train-text-from-scratch
+/vdot
 build-info.h
 arm_neon.h
 compile_commands.json

CMakeLists.txt

@@ -197,37 +197,38 @@ endif()
 
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
-        set(c_flags
-            -Wall
-            -Wextra
-            -Wpedantic
-            -Wcast-qual
-            -Wdouble-promotion
-            -Wshadow
-            -Wstrict-prototypes
-            -Wpointer-arith
-            -Wmissing-prototypes
-            -Werror=implicit-int
-            -Wno-unused-function
-        )
-        set(cxx_flags
-            -Wall
-            -Wextra
-            -Wpedantic
-            -Wcast-qual
-            -Wmissing-declarations
-            -Wno-unused-function
-            -Wno-multichar
-        )
-        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
-            # g++ only
-            set(cxx_flags ${cxx_flags} -Wno-format-truncation -Wno-array-bounds)
+        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(cxx_flags -Wmissing-declarations -Wmissing-noreturn)
+
+        if (CMAKE_C_COMPILER_ID MATCHES "Clang")
+            set(warning_flags ${warning_flags} -Wunreachable-code-break -Wunreachable-code-return)
+            set(cxx_flags ${cxx_flags} -Wmissing-prototypes -Wextra-semi)
+
+            if (
+                (CMAKE_C_COMPILER_ID STREQUAL "Clang"      AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 3.8.0) OR
+                (CMAKE_C_COMPILER_ID STREQUAL "AppleClang" AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 7.3.0)
+            )
+                set(c_flags ${c_flags} -Wdouble-promotion)
+            endif()
+        elseif (CMAKE_C_COMPILER_ID STREQUAL "GNU")
+            set(c_flags ${c_flags} -Wdouble-promotion)
+            set(cxx_flags ${cxx_flags} -Wno-array-bounds)
+
+            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 7.1.0)
+                set(cxx_flags ${cxx_flags} -Wno-format-truncation)
+            endif()
+            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 8.1.0)
+                set(cxx_flags ${cxx_flags} -Wextra-semi)
+            endif()
         endif()
     else()
         # todo : msvc
     endif()
 
     add_compile_options(
+            ${warning_flags}
             "$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
             "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>"
     )

Dockerfile

@@ -4,8 +4,8 @@ COPY . .
 RUN apt update \
  && apt install build-essential wget libopenblas-dev make -y \
  && make LLAMA_OPENBLAS=1 \
- && wget https://huggingface.co/TheBloke/Pygmalion-2-7B-GGUF/resolve/main/pygmalion-2-7b.Q4_0.gguf \
+ && wget https://huggingface.co/notstoic/pygmalion-13b-ggml/resolve/main/pygmalion-13b-ggml-q4_0.bin \
  && apt remove build-essential wget make -y \
- && rm -fr *.bat convert-* ci docs examples otherarchs tests 
+ && rm -fr *.bat convert-* ci docs examples otherarchs tests
 
- ENTRYPOINT ["python", "koboldcpp.py", "pygmalion-2-7b.Q4_0.gguf", "--port", "7860"]
+ ENTRYPOINT ["python", "koboldcpp.py", "pygmalion-13b-ggml-q4_0.bin", "--port", "7860", "--smartcontext", "--stream"]

Makefile

@@ -39,10 +39,15 @@ endif
 #
 
 # keep standard at C11 and C++11
-CFLAGS   = -I.              -I./include -I./include/CL -I./otherarch -I./otherarch/tools -Ofast -DNDEBUG -std=c11   -fPIC -DGGML_USE_K_QUANTS -DLOG_DISABLE_LOGS -D_GNU_SOURCE
-CXXFLAGS = -I. -I./common -I./include -I./include/CL -I./otherarch -I./otherarch/tools -Ofast -DNDEBUG -std=c++11 -fPIC -DGGML_USE_K_QUANTS -DLOG_DISABLE_LOGS -D_GNU_SOURCE
+CFLAGS   = -I.            -I./include -I./include/CL -I./otherarch -I./otherarch/tools -Ofast -DNDEBUG -std=c11   -fPIC -DLOG_DISABLE_LOGS -D_GNU_SOURCE
+CXXFLAGS = -I. -I./common -I./include -I./include/CL -I./otherarch -I./otherarch/tools -Ofast -DNDEBUG -std=c++11 -fPIC -DLOG_DISABLE_LOGS -D_GNU_SOURCE
 LDFLAGS  =
 
+ifndef LLAMA_NO_K_QUANTS
+CFLAGS   += -DGGML_USE_K_QUANTS
+CXXFLAGS += -DGGML_USE_K_QUANTS
+endif
+
 # these are used on windows, to build some libraries with extra old device compatibility
 SIMPLECFLAGS =
 FULLCFLAGS =
@@ -285,19 +290,17 @@ ifeq ($(OS),Windows_NT)
 	endif
 else
 	DEFAULT_BUILD = $(CXX) $(CXXFLAGS)  $^ -shared -o $@.so $(LDFLAGS)
-	FAILSAFE_BUILD = $(CXX) $(CXXFLAGS) $^ -shared -o $@.so $(LDFLAGS)
+
 	ifdef LLAMA_OPENBLAS
 	OPENBLAS_BUILD = $(CXX) $(CXXFLAGS) $^ $(ARCH_ADD) -lopenblas -shared -o $@.so $(LDFLAGS)
-	NOAVX2_BUILD = $(CXX) $(CXXFLAGS) $^ $(ARCH_ADD) -lopenblas -shared -o $@.so $(LDFLAGS)
 	endif
 	ifdef LLAMA_CLBLAST
 		ifeq ($(UNAME_S),Darwin)
-				CLBLAST_BUILD = $(CXX) $(CXXFLAGS) $^ -lclblast -framework OpenCL $(ARCH_ADD) -lopenblas -shared -o $@.so $(LDFLAGS)
+			CLBLAST_BUILD = $(CXX) $(CXXFLAGS) $^ -lclblast -framework OpenCL $(ARCH_ADD) -lopenblas -shared -o $@.so $(LDFLAGS)
 		else
-				CLBLAST_BUILD = $(CXX) $(CXXFLAGS) $^ -lclblast -lOpenCL $(ARCH_ADD) -lopenblas -shared -o $@.so $(LDFLAGS)
+			CLBLAST_BUILD = $(CXX) $(CXXFLAGS) $^ -lclblast -lOpenCL $(ARCH_ADD) -lopenblas -shared -o $@.so $(LDFLAGS)
 		endif
 	endif
-
 	ifdef LLAMA_CUBLAS
 		CUBLAS_BUILD = $(CXX) $(CXXFLAGS) $(CUBLAS_FLAGS) $^ -shared -o $@.so $(CUBLASLD_FLAGS) $(LDFLAGS)
 	endif
@@ -351,12 +354,20 @@ ggml_cublas.o: ggml.c ggml.h ggml-cuda.h k_quants.h
 	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(CUBLAS_FLAGS) $(HIPFLAGS) -c $< -o $@
 
 #quants K
+KQ1 =
+KQ2 =
+KQ3 =
+ifndef LLAMA_NO_K_QUANTS
+KQ1 = k_quants.o
+KQ2 = k_quants_noavx2.o
+KQ3 = k_quants_failsafe.o
 k_quants.o: k_quants.c k_quants.h ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS) $(FULLCFLAGS) -c $< -o $@
 k_quants_noavx2.o: k_quants.c k_quants.h ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS) $(SIMPLECFLAGS) -c $< -o $@
 k_quants_failsafe.o: k_quants.c k_quants.h ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS) $(NONECFLAGS) -c $< -o $@
+endif # LLAMA_NO_K_QUANTS
 
 #there's no intrinsics or special gpu ops used here, so we can have a universal object
 ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
@@ -416,7 +427,7 @@ gpttype_adapter_cublas.o: $(GPTTYPE_ADAPTER)
 clean:
 	rm -vf *.o main quantize_llama quantize_gpt2 quantize_gptj quantize_neox quantize_mpt quantize-stats perplexity embedding benchmark-matmult save-load-state gguf gguf.exe main.exe quantize_llama.exe quantize_gptj.exe quantize_gpt2.exe quantize_neox.exe quantize_mpt.exe koboldcpp_default.dll koboldcpp_openblas.dll koboldcpp_failsafe.dll koboldcpp_noavx2.dll koboldcpp_clblast.dll koboldcpp_cublas.dll koboldcpp_hipblas.dll koboldcpp_default.so koboldcpp_openblas.so koboldcpp_failsafe.so koboldcpp_noavx2.so koboldcpp_clblast.so koboldcpp_cublas.so koboldcpp_hipblas.so
 
-main: examples/main/main.cpp build-info.h ggml.o k_quants.o ggml-alloc.o llama.o common.o console.o grammar-parser.o $(OBJS)
+main: examples/main/main.cpp build-info.h ggml.o $(KQ1) ggml-alloc.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.  ===='
@@ -425,31 +436,69 @@ main: examples/main/main.cpp build-info.h ggml.o k_quants.o ggml-alloc.o llama.o
 gguf: examples/gguf/gguf.cpp build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+
 #generated libraries
-koboldcpp_default: ggml.o ggml_v2.o ggml_v1.o expose.o common.o gpttype_adapter.o k_quants.o ggml-alloc.o grammar-parser.o $(OBJS)
+koboldcpp_default: ggml.o ggml_v2.o ggml_v1.o expose.o common.o gpttype_adapter.o $(KQ1) ggml-alloc.o grammar-parser.o $(OBJS)
 	$(DEFAULT_BUILD)
-koboldcpp_openblas: ggml_openblas.o ggml_v2_openblas.o ggml_v1.o expose.o common.o gpttype_adapter.o k_quants.o ggml-alloc.o grammar-parser.o $(OBJS)
+
+ifdef OPENBLAS_BUILD
+koboldcpp_openblas: ggml_openblas.o ggml_v2_openblas.o ggml_v1.o expose.o common.o gpttype_adapter.o $(KQ1) ggml-alloc.o grammar-parser.o $(OBJS)
 	$(OPENBLAS_BUILD)
-koboldcpp_failsafe: ggml_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o k_quants_failsafe.o ggml-alloc.o grammar-parser.o $(OBJS)
+else
+koboldcpp_openblas:
+	$(DONOTHING)
+endif
+
+ifdef FAILSAFE_BUILD
+koboldcpp_failsafe: ggml_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o $(KQ3) ggml-alloc.o grammar-parser.o $(OBJS)
 	$(FAILSAFE_BUILD)
-koboldcpp_noavx2: ggml_noavx2.o ggml_v2_noavx2.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o k_quants_noavx2.o ggml-alloc.o grammar-parser.o $(OBJS)
+else
+koboldcpp_failsafe:
+	$(DONOTHING)
+endif
+
+ifdef NOAVX2_BUILD
+koboldcpp_noavx2: ggml_noavx2.o ggml_v2_noavx2.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o $(KQ2) ggml-alloc.o grammar-parser.o $(OBJS)
 	$(NOAVX2_BUILD)
-koboldcpp_clblast: ggml_clblast.o ggml_v2_clblast.o ggml_v1.o expose.o common.o gpttype_adapter_clblast.o ggml-opencl.o ggml_v2-opencl.o ggml_v2-opencl-legacy.o k_quants.o ggml-alloc.o grammar-parser.o $(OBJS)
+else
+koboldcpp_noavx2:
+	$(DONOTHING)
+endif
+
+ifdef CLBLAST_BUILD
+koboldcpp_clblast: ggml_clblast.o ggml_v2_clblast.o ggml_v1.o expose.o common.o gpttype_adapter_clblast.o ggml-opencl.o ggml_v2-opencl.o ggml_v2-opencl-legacy.o $(KQ1) ggml-alloc.o grammar-parser.o $(OBJS)
 	$(CLBLAST_BUILD)
-koboldcpp_cublas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o k_quants.o ggml-alloc.o grammar-parser.o $(CUBLAS_OBJS) $(OBJS)
+else
+koboldcpp_clblast:
+	$(DONOTHING)
+endif
+
+ifdef CUBLAS_BUILD
+koboldcpp_cublas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o $(KQ1) ggml-alloc.o grammar-parser.o $(CUBLAS_OBJS) $(OBJS)
 	$(CUBLAS_BUILD)
-koboldcpp_hipblas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o k_quants.o ggml-alloc.o grammar-parser.o $(HIP_OBJS) $(OBJS)
+else
+koboldcpp_cublas:
+	$(DONOTHING)
+endif
+
+ifdef HIPBLAS_BUILD
+koboldcpp_hipblas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o $(KQ1) ggml-alloc.o grammar-parser.o $(HIP_OBJS) $(OBJS)
 	$(HIPBLAS_BUILD)
+else
+koboldcpp_hipblas:
+	$(DONOTHING)
+endif
 
-quantize_llama: examples/quantize/quantize.cpp ggml.o llama.o k_quants.o ggml-alloc.o
+# tools
+quantize_llama: examples/quantize/quantize.cpp ggml.o llama.o $(KQ1) ggml-alloc.o
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_gptj: ggml.o llama.o k_quants.o ggml-alloc.o otherarch/tools/gptj_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_gptj: ggml.o llama.o $(KQ1) ggml-alloc.o otherarch/tools/gptj_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_gpt2: ggml.o llama.o k_quants.o ggml-alloc.o otherarch/tools/gpt2_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_gpt2: ggml.o llama.o $(KQ1) ggml-alloc.o otherarch/tools/gpt2_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_neox: ggml.o llama.o k_quants.o ggml-alloc.o otherarch/tools/neox_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_neox: ggml.o llama.o $(KQ1) ggml-alloc.o otherarch/tools/neox_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_mpt: ggml.o llama.o k_quants.o ggml-alloc.o otherarch/tools/mpt_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_mpt: ggml.o llama.o $(KQ1) ggml-alloc.o otherarch/tools/mpt_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
 

Package.swift

@@ -10,7 +10,7 @@ let platforms: [SupportedPlatform]? = [
     .tvOS(.v14)
 ]
 let exclude: [String] = []
-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"),
@@ -44,7 +44,9 @@ let package = Package(
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32"]),
                 .define("GGML_USE_K_QUANTS"),
-                .define("GGML_USE_ACCELERATE")
+                .define("GGML_USE_ACCELERATE"),
+                .define("ACCELERATE_NEW_LAPACK"),
+                .define("ACCELERATE_LAPACK_ILP64")
             ] + additionalSettings,
             linkerSettings: [
                 .linkedFramework("Accelerate")

README.md

@@ -1,6 +1,84 @@
----
-sdk: docker
-emoji: 🚀
-colorFrom: yellow
-colorTo: blue
----
+# koboldcpp
+
+KoboldCpp is an easy-to-use AI text-generation software for GGML models. It's a single self contained distributable from Concedo, that builds off llama.cpp, and adds a versatile Kobold API endpoint, additional format support, backward compatibility, as well as a fancy UI with persistent stories, editing tools, save formats, memory, world info, author's note, characters, scenarios and everything Kobold and Kobold Lite have to offer.
+
+![Preview](media/preview.png)
+
+## Usage
+- **[Download the latest .exe release here](https://github.com/LostRuins/koboldcpp/releases/latest)** or clone the git repo.
+- Windows binaries are provided in the form of **koboldcpp.exe**, which is a pyinstaller wrapper for a few **.dll** files and **koboldcpp.py**. If you feel concerned, you may prefer to rebuild it yourself with the provided makefiles and scripts.
+- Weights are not included, you can use the official llama.cpp `quantize.exe` to generate them from your official weight files (or download them from other places such as [TheBloke's Huggingface](https://huggingface.co/TheBloke).
+- To run, execute **koboldcpp.exe** or drag and drop your quantized `ggml_model.bin` file onto the .exe, and then connect with Kobold or Kobold Lite. If you're not on windows, then run the script **KoboldCpp.py** after compiling the libraries.
+- Launching with no command line arguments displays a GUI containing a subset of configurable settings. Generally you dont have to change much besides the `Presets` and `GPU Layers`. Read the `--help` for more info about each settings.
+- By default, you can connect to http://localhost:5001
+- You can also run it using the command line `koboldcpp.exe [ggml_model.bin] [port]`. For info, please check `koboldcpp.exe --help`
+- Default context size to small? Try `--contextsize 3072` to 1.5x your context size! without much perplexity gain. Note that you'll have to increase the max context in the Kobold Lite UI as well (click and edit the number text field).
+- Big context too slow? Try the `--smartcontext` flag to reduce prompt processing frequency. Also, you can try to run with your GPU using CLBlast, with `--useclblast` flag for a speedup
+- Want even more speedup? Combine `--useclblast` with `--gpulayers` to offload entire layers to the GPU! **Much faster, but uses more VRAM**. Experiment to determine number of layers to offload, and reduce by a few if you run out of memory.
+- If you are having crashes or issues, you can try turning off BLAS with the `--noblas` flag. You can also try running in a non-avx2 compatibility mode with `--noavx2`. Lastly, you can try turning off mmap with `--nommap`.
+
+For more information, be sure to run the program with the `--help` flag.
+
+## OSX and Linux
+- You will have to compile your binaries from source. A makefile is provided, simply run `make`
+- If you want you can also link your own install of OpenBLAS manually with `make LLAMA_OPENBLAS=1`
+- Alternatively, if you want you can also link your own install of CLBlast manually with `make LLAMA_CLBLAST=1`, for this you will need to obtain and link OpenCL and CLBlast libraries.
+  - For Arch Linux: Install `cblas` `openblas` and `clblast`.
+  - For Debian: Install `libclblast-dev` and `libopenblas-dev`.
+- For a full featured build, do `make LLAMA_OPENBLAS=1 LLAMA_CLBLAST=1 LLAMA_CUBLAS=1`
+- After all binaries are built, you can run the python script with the command `koboldcpp.py [ggml_model.bin] [port]`
+- Note: Many OSX users have found that the using Accelerate is actually faster than OpenBLAS. To try, you may wish to run with `--noblas` and compare speeds.
+
+## Compiling on Windows
+- You're encouraged to use the .exe released, but if you want to compile your binaries from source at Windows, the easiest way is:
+  - Use the latest release of w64devkit (https://github.com/skeeto/w64devkit). Be sure to use the "vanilla one", not i686 or other different stuff. If you try they will conflit with the precompiled libs!
+  - Make sure you are using the w64devkit integrated terminal, then run 'make' at the KoboldCpp source folder. This will create the .dll files.
+  - If you want to generate the .exe file, make sure you have the python module PyInstaller installed with pip ('pip install PyInstaller').
+  - Run the script make_pyinstaller.bat at a regular terminal (or Windows Explorer).
+  - The koboldcpp.exe file will be at your dist folder.
+- If you wish to use your own version of the additional Windows libraries (OpenCL, CLBlast and OpenBLAS), you can do it with:
+  - OpenCL - tested with https://github.com/KhronosGroup/OpenCL-SDK . If you wish to compile it, follow the repository instructions. You will need vcpkg.
+  - CLBlast - tested with https://github.com/CNugteren/CLBlast . If you wish to compile it you will need to reference the OpenCL files. It will only generate the ".lib" file if you compile using MSVC.
+  - OpenBLAS - tested with https://github.com/xianyi/OpenBLAS .
+  - Move the respectives .lib files to the /lib folder of your project, overwriting the older files.
+  - Also, replace the existing versions of the corresponding .dll files located in the project directory root (e.g. libopenblas.dll).
+  - Make the KoboldCPP project using the instructions above.
+
+## Android (Termux) Alternative method
+- See https://github.com/ggerganov/llama.cpp/pull/1828/files
+
+## Using CuBLAS
+- If you're on Windows with an Nvidia GPU you can get CUDA support out of the box using the `--usecublas` flag, make sure you select the correct .exe with CUDA support.
+- You can attempt a CuBLAS build with `LLAMA_CUBLAS=1` or using the provided CMake file (best for visual studio users). If you use the CMake file to build, copy the `koboldcpp_cublas.dll` generated into the same directory as the `koboldcpp.py` file. If you are bundling executables, you may need to include CUDA dynamic libraries (such as `cublasLt64_11.dll` and `cublas64_11.dll`) in order for the executable to work correctly on a different PC.
+
+## AMD
+- Please check out https://github.com/YellowRoseCx/koboldcpp-rocm
+
+## Questions and Help
+- **First, please check out [The KoboldCpp FAQ and Knowledgebase](https://github.com/LostRuins/koboldcpp/wiki) which may already have answers to your questions! Also please search through past issues and discussions.**
+- If you cannot find an answer, open an issue on this github, or find us on the [KoboldAI Discord](https://koboldai.org/discord).
+
+## Considerations
+- For Windows: No installation, single file executable, (It Just Works)
+- Since v1.0.6, requires libopenblas, the prebuilt windows binaries are included in this repo. If not found, it will fall back to a mode without BLAS.
+- Since v1.15, requires CLBlast if enabled, the prebuilt windows binaries are included in this repo. If not found, it will fall back to a mode without CLBlast.
+- Since v1.33, you can set the context size to be above what the model supports officially. It does increases perplexity but should still work well below 4096 even on untuned models. (For GPT-NeoX, GPT-J, and LLAMA models) Customize this with `--ropeconfig`.
+- **I plan to keep backwards compatibility with ALL past llama.cpp AND alpaca.cpp models**. But you are also encouraged to reconvert/update your models if possible for best results.
+
+## License
+- The original GGML library and llama.cpp by ggerganov are licensed under the MIT License
+- However, Kobold Lite is licensed under the AGPL v3.0 License
+- The other files are also under the AGPL v3.0 License unless otherwise stated
+
+## Notes
+- Generation delay scales linearly with original prompt length. If OpenBLAS is enabled then prompt ingestion becomes about 2-3x faster. This is automatic on windows, but will require linking on OSX and Linux. CLBlast speeds this up even further, and `--gpulayers` + `--useclblast` or `--usecublas` more so.
+- I have heard of someone claiming a false AV positive report. The exe is a simple pyinstaller bundle that includes the necessary python scripts and dlls to run. If this still concerns you, you might wish to rebuild everything from source code using the makefile, and you can rebuild the exe yourself with pyinstaller by using `make_pyinstaller.bat`
+- Supported GGML models (Includes backward compatibility for older versions/legacy GGML models, though some newer features might be unavailable):
+  - LLAMA and LLAMA2 (LLaMA / Alpaca / GPT4All / Vicuna / Koala / Pygmalion 7B / Metharme 7B / WizardLM and many more)
+  - GPT-2 / Cerebras
+  - GPT-J
+  - RWKV
+  - GPT-NeoX / Pythia / StableLM / Dolly / RedPajama
+  - MPT models
+  - Falcon (GGUF only)
+

Remote-Link.cmd

@@ -1,2 +1,18 @@
-curl -L https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-windows-amd64.exe -o cloudflared.exe
-cloudflared.exe tunnel --url localhost:5001
+: # This script will help setup a cloudflared tunnel for accessing KoboldCpp over the internet
+: # It should work out of the box on both linux and windows
+: # ======
+: # WINDOWS PORTION
+:<<BATCH
+    @echo off
+    echo Starting Cloudflare Tunnel for Windows
+    curl -L https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-windows-amd64.exe -o cloudflared.exe
+    cloudflared.exe tunnel --url localhost:5001
+    GOTO ENDING
+BATCH
+: # LINUX PORTION
+echo 'Starting Cloudflare Tunnel for Linux'
+curl -L https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-linux-amd64 -o 'cloudflared-linux-amd64' # 
+chmod +x 'cloudflared-linux-amd64' # 
+./cloudflared-linux-amd64 tunnel --url http://localhost:5001 # 
+exit # 
+:ENDING

class.py

@@ -188,6 +188,18 @@ class model_backend(InferenceModel):
                                     "extra_classes": "",
                                     'children': [{'text': 'False', 'value': False}, {'text': 'True', 'value': True}],
                                     })
+        requested_parameters.append({
+                                    "uitype": "text",
+                                    "unit": "text",
+                                    "label": "GPU ID",
+                                    "id": "kcpp_tensor_split_str",
+                                    "default": "1",
+                                    "check": {"value": "", 'check': "!="},
+                                    "tooltip": "Which GPU's do we use? For example:1 2",
+                                    "menu_path": "",
+                                    "refresh_model_inputs": False,
+                                    "extra_classes": ""
+                                    })
         requested_parameters.append({
                                     "uitype": "dropdown",
                                     "unit": "int",
@@ -202,18 +214,6 @@ class model_backend(InferenceModel):
                                     "extra_classes": "",
                                     'children': [{'text': 'False', 'value': 0}, {'text': 'True', 'value': 1}],
                                     })
-        requested_parameters.append({
-                                    "uitype": "text",
-                                    "unit": "text",
-                                    "label": "Tensor Split",
-                                    "id": "kcpp_tensor_split_str",
-                                    "default": self.kcpp_tensor_split_str,
-                                    "check": {"value": "", 'check': "!="},
-                                    "tooltip": "Tensor Split, values are space separated",
-                                    "menu_path": "",
-                                    "refresh_model_inputs": False,
-                                    "extra_classes": ""
-                                    })
         return requested_parameters
 
     def set_input_parameters(self, parameters):
@@ -232,6 +232,7 @@ class model_backend(InferenceModel):
             self.kcpp_tensor_split = []
             for s in splits:
                 self.kcpp_tensor_split.append(int(s))
+                print(self.kcpp_tensor_split)
 
         accel = parameters["kcpp_accelerator"]
         if accel==0:
@@ -271,7 +272,8 @@ class model_backend(InferenceModel):
             blasbatchsize=self.kcpp_blasbatchsize, ropeconfig=[self.kcpp_ropescale, self.kcpp_ropebase], stream=False, smartcontext=self.kcpp_smartcontext,
             unbantokens=False, bantokens=None, usemirostat=None, forceversion=0, nommap=self.kcpp_nommap,
             usemlock=False, noavx2=self.kcpp_noavx2, debugmode=self.kcpp_debugmode, skiplauncher=True, hordeconfig=None, noblas=self.kcpp_noblas,
-            useclblast=self.kcpp_useclblast, usecublas=self.kcpp_usecublas, gpulayers=self.kcpp_gpulayers, tensor_split=self.kcpp_tensor_split, config=None, onready='', multiuser=False)
+            useclblast=self.kcpp_useclblast, usecublas=self.kcpp_usecublas, gpulayers=self.kcpp_gpulayers, tensor_split=self.kcpp_tensor_split, config=None,
+            onready='', multiuser=False, foreground=False)
 
             koboldcpp.main(kcppargs,False) #initialize library without enabling Lite http server
             kcpp_backend_loaded = True

colab.ipynb

@@ -6,7 +6,7 @@
       "private_outputs": true,
       "provenance": [],
       "gpuType": "T4",
-      "authorship_tag": "ABX9TyOv14c2MWENhO6RJ3uy6vD7",
+      "authorship_tag": "",
       "include_colab_link": true
     },
     "kernelspec": {
@@ -25,9 +25,7 @@
         "id": "view-in-github",
         "colab_type": "text"
       },
-      "source": [
-        "<a href=\"https://colab.research.google.com/github/henk717/koboldcpp/blob/concedo/colab.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
-      ]
+      "source": []
     },
     {
       "cell_type": "code",
@@ -40,7 +38,7 @@
       "source": [
         "#@title <b>v-- Enter your model below and then click this to start Koboldcpp</b>\n",
         "\n",
-        "Model = \"https://huggingface.co/TheBloke/airoboros-l2-13B-gpt4-1.4.1-GGML/resolve/main/airoboros-l2-13b-gpt4-1.4.1.ggmlv3.q4_0.bin\" #@param [\"\"]{allow-input: true}\n",
+        "Model = \"https://huggingface.co/TheBloke/Airoboros-L2-13B-2.2-GGUF/resolve/main/airoboros-l2-13b-2.2.Q4_K_M.gguf\" #@param [\"\"]{allow-input: true}\n",
         "Layers = 43 #@param [43]{allow-input: true}\n",
         "\n",
         "%cd /content\n",
@@ -54,7 +52,7 @@
         "!nohup ./cloudflared-linux-amd64 tunnel --url http://localhost:5001 &\n",
         "!sleep 10\n",
         "!cat nohup.out\n",
-        "!python koboldcpp.py model.ggml --stream --usecublas 0 --gpulayers $Layers --hordeconfig concedo\n"
+        "!python koboldcpp.py model.ggml --stream --usecublas 0 mmq --gpulayers $Layers --hordeconfig concedo\n"
       ]
     }
   ]

common/CMakeLists.txt

@@ -9,6 +9,8 @@ add_library(${TARGET} OBJECT
     console.cpp
     grammar-parser.h
     grammar-parser.cpp
+    train.h
+    train.cpp
     )
 
 if (BUILD_SHARED_LIBS)

common/common.cpp

@@ -78,7 +78,7 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
-static void process_escapes(std::string& input) {
+void process_escapes(std::string& input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
 
@@ -129,6 +129,15 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             if (params.n_threads <= 0) {
                 params.n_threads = std::thread::hardware_concurrency();
             }
+        } else if (arg == "-tb" || arg == "--threads-batch") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_threads_batch = std::stoi(argv[i]);
+            if (params.n_threads_batch <= 0) {
+                params.n_threads_batch = std::thread::hardware_concurrency();
+            }
         } else if (arg == "-p" || arg == "--prompt") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -317,6 +326,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_chunks = std::stoi(argv[i]);
+        } else if (arg == "-np" || arg == "--parallel") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_parallel = std::stoi(argv[i]);
+        } else if (arg == "-ns" || arg == "--sequences") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_sequences = std::stoi(argv[i]);
         } else if (arg == "-m" || arg == "--model") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -340,7 +361,19 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter = argv[i];
+            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({lora_adapter, std::stof(argv[i])});
             params.use_mmap = false;
         } else if (arg == "--lora-base") {
             if (++i >= argc) {
@@ -360,6 +393,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.multiline_input = true;
         } else if (arg == "--simple-io") {
             params.simple_io = true;
+        } else if (arg == "-cb" || arg == "--cont-batching") {
+            params.cont_batching = true;
         } else if (arg == "--color") {
             params.use_color = true;
         } else if (arg == "--mlock") {
@@ -425,19 +460,11 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.mul_mat_q = false;
 #else
             fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
-#endif // GGML_USE_CUBLAS
-        } else if (arg == "--low-vram" || arg == "-lv") {
-#ifdef GGML_USE_CUBLAS
-            params.low_vram = true;
-#else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--numa") {
             params.numa = true;
-        } else if (arg == "--export") {
-            params.export_cgraph = true;
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
         } else if (arg == "-r" || arg == "--reverse-prompt") {
@@ -456,8 +483,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             if (params.logdir.back() != DIRECTORY_SEPARATOR) {
                 params.logdir += DIRECTORY_SEPARATOR;
             }
-        } else if (arg == "--perplexity") {
-            params.perplexity = true;
+        } else if (arg == "--perplexity" || arg == "--all-logits") {
+            params.logits_all = true;
         } else if (arg == "--ppl-stride") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -606,7 +633,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        (can be specified more than once for multiple prompts).\n");
     printf("  --color               colorise output to distinguish prompt and user input from generations\n");
     printf("  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
-    printf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -t N, --threads N     number of threads to use during generation (default: %d)\n", params.n_threads);
+    printf("  -tb N, --threads-batch N\n");
+    printf("                        number of threads to use during batch and prompt processing (default: same as --threads)\n");
     printf("  -p PROMPT, --prompt PROMPT\n");
     printf("                        prompt to start generation with (default: empty)\n");
     printf("  -e, --escape          process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
@@ -621,7 +650,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -f FNAME, --file FNAME\n");
     printf("                        prompt file to start generation.\n");
     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)\n", params.n_ctx);
+    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", params.top_k);
     printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
@@ -655,12 +684,15 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     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)params.temp);
-    printf("  --perplexity          compute perplexity over each ctx window of the prompt\n");
+    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);
     printf("  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
     printf("  --draft N             number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
     printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
+    printf("  -np N, --parallel N   number of parallel sequences to decode (default: %d)\n", params.n_parallel);
+    printf("  -ns N, --sequences N  number of sequences to decode (default: %d)\n", params.n_sequences);
+    printf("  -cb, --cont-batching  enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
     if (llama_mlock_supported()) {
         printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
@@ -678,17 +710,16 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     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("  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
 #ifdef GGML_USE_CUBLAS
     printf("  -nommq, --no-mul-mat-q\n");
     printf("                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
     printf("                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif // GGML_USE_CUBLAS
 #endif
-    printf("  --export              export the computation graph to 'llama.ggml'\n");
     printf("  --verbose-prompt      print prompt before generation\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");
     printf("  -m FNAME, --model FNAME\n");
     printf("                        model path (default: %s)\n", params.model.c_str());
@@ -699,6 +730,18 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("\n");
 }
 
+std::string get_system_info(const gpt_params & params) {
+    std::ostringstream os;
+
+    os << "system_info: n_threads = " << params.n_threads;
+    if (params.n_threads_batch != -1) {
+        os << " (n_threads_batch = " << params.n_threads_batch << ")";
+    }
+    os << " / " << std::thread::hardware_concurrency() << " | " << llama_print_system_info();
+
+    return os.str();
+}
+
 std::string gpt_random_prompt(std::mt19937 & rng) {
     const int r = rng() % 10;
     switch (r) {
@@ -712,60 +755,74 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
         case 7: return "He";
         case 8: return "She";
         case 9: return "They";
-        default: return "To";
     }
 
-    return "The";
+    GGML_UNREACHABLE();
 }
 
 //
 // Model utils
 //
 
-struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
-    auto lparams = llama_context_default_params();
+struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
+    auto mparams = llama_model_default_params();
 
-    lparams.n_ctx           = params.n_ctx;
-    lparams.n_batch         = params.n_batch;
     if (params.n_gpu_layers != -1) {
-        lparams.n_gpu_layers = params.n_gpu_layers;
+        mparams.n_gpu_layers = params.n_gpu_layers;
     }
-    lparams.main_gpu        = params.main_gpu;
-    lparams.tensor_split    = params.tensor_split;
-    lparams.low_vram        = params.low_vram;
-    lparams.mul_mat_q       = params.mul_mat_q;
-    lparams.seed            = params.seed;
-    lparams.f16_kv          = params.memory_f16;
-    lparams.use_mmap        = params.use_mmap;
-    lparams.use_mlock       = params.use_mlock;
-    lparams.logits_all      = params.perplexity;
-    lparams.embedding       = params.embedding;
-    lparams.rope_freq_base  = params.rope_freq_base;
-    lparams.rope_freq_scale = params.rope_freq_scale;
-
-    return lparams;
+    mparams.main_gpu        = params.main_gpu;
+    mparams.tensor_split    = params.tensor_split;
+    mparams.use_mmap        = params.use_mmap;
+    mparams.use_mlock       = params.use_mlock;
+
+    return mparams;
+}
+
+struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
+    auto cparams = llama_context_default_params();
+
+    cparams.n_ctx           = params.n_ctx;
+    cparams.n_batch         = params.n_batch;
+    cparams.n_threads       = params.n_threads;
+    cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    cparams.mul_mat_q       = params.mul_mat_q;
+    cparams.seed            = params.seed;
+    cparams.f16_kv          = params.memory_f16;
+    cparams.logits_all      = params.logits_all;
+    cparams.embedding       = params.embedding;
+    cparams.rope_freq_base  = params.rope_freq_base;
+    cparams.rope_freq_scale = params.rope_freq_scale;
+
+    return cparams;
 }
 
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
-    auto lparams = llama_context_params_from_gpt_params(params);
+    auto mparams = llama_model_params_from_gpt_params(params);
 
-    llama_model * model  = llama_load_model_from_file(params.model.c_str(), lparams);
+    llama_model * model  = llama_load_model_from_file(params.model.c_str(), mparams);
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
         return std::make_tuple(nullptr, nullptr);
     }
 
-    llama_context * lctx = llama_new_context_with_model(model, lparams);
+    auto cparams = llama_context_params_from_gpt_params(params);
+
+    llama_context * lctx = llama_new_context_with_model(model, cparams);
     if (lctx == NULL) {
         fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
         llama_free_model(model);
         return std::make_tuple(nullptr, nullptr);
     }
 
-    if (!params.lora_adapter.empty()) {
+    for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
+        const std::string& lora_adapter = std::get<0>(params.lora_adapter[i]);
+        float lora_scale = std::get<1>(params.lora_adapter[i]);
         int err = llama_model_apply_lora_from_file(model,
-                                             params.lora_adapter.c_str(),
-                                             params.lora_base.empty() ? NULL : params.lora_base.c_str(),
+                                             lora_adapter.c_str(),
+                                             lora_scale,
+                                             ((i > 0) || params.lora_base.empty())
+                                                ? NULL
+                                                : params.lora_base.c_str(),
                                              params.n_threads);
         if (err != 0) {
             fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
@@ -782,8 +839,9 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     {
         LOG("warming up the model with an empty run\n");
 
-        const std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
-        llama_eval(lctx, tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, params.n_threads);
+        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_tokens_rm(lctx, -1, -1);
         llama_reset_timings(lctx);
     }
 
@@ -795,16 +853,23 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 //
 
 std::vector<llama_token> llama_tokenize(
-        struct llama_context * ctx,
+  const struct llama_context * ctx,
+           const std::string & text,
+                        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) {
     // 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(ctx, text.data(), text.length(), result.data(), result.size(), add_bos);
+    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(ctx, text.data(), text.length(), result.data(), result.size(), add_bos);
+        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);
@@ -814,10 +879,10 @@ std::vector<llama_token> llama_tokenize(
 
 std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_piece(ctx, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_piece(ctx, token, result.data(), result.size());
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -872,7 +937,7 @@ llama_token llama_sample_token(
          std::vector<llama_token_data> & candidates,
                                    int   idx) {
     const int n_ctx   = llama_n_ctx(ctx);
-    const int n_vocab = llama_n_vocab(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;
@@ -890,7 +955,7 @@ llama_token llama_sample_token(
 
     llama_token id = 0;
 
-    float * logits = llama_get_logits(ctx) + idx * n_vocab;
+    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++) {
@@ -941,11 +1006,11 @@ llama_token llama_sample_token(
         if (mirostat == 1) {
             static float mirostat_mu = 2.0f * mirostat_tau;
             const int mirostat_m = 100;
-            llama_sample_temperature(ctx, &cur_p, temp);
+            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_temperature(ctx, &cur_p, temp);
+            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
@@ -953,7 +1018,7 @@ llama_token llama_sample_token(
             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_temperature(ctx, &cur_p, temp);
+            llama_sample_temp(ctx, &cur_p, temp);
 
             {
                 const int n_top = 10;
@@ -1158,7 +1223,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
 #endif // NDEBUG
 
     fprintf(stream, "model_desc: %s\n", model_desc);
-    fprintf(stream, "n_vocab: %d  # output size of the final layer, 32001 for some models\n", llama_n_vocab(lctx));
+    fprintf(stream, "n_vocab: %d  # output size of the final layer, 32001 for some models\n", llama_n_vocab(llama_get_model(lctx)));
 
 #ifdef __OPTIMIZE__
     fprintf(stream, "optimize: true\n");
@@ -1182,7 +1247,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     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, "export: %s # default: false\n", params.export_cgraph ? "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", params.frequency_penalty);
     dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str());
@@ -1211,9 +1275,21 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
         fprintf(stream, "  %d: %f", lb.first, lb.second);
     }
 
-    fprintf(stream, "lora: %s\n", params.lora_adapter.c_str());
+    fprintf(stream, "lora:\n");
+    for (std::tuple<std::string, float> la : params.lora_adapter) {
+        if (std::get<1>(la) != 1.0f) {
+            continue;
+        }
+        fprintf(stream, "  - %s\n", std::get<0>(la).c_str());
+    }
+    fprintf(stream, "lora_scaled:\n");
+    for (std::tuple<std::string, float> la : params.lora_adapter) {
+        if (std::get<1>(la) == 1.0f) {
+            continue;
+        }
+        fprintf(stream, "  - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la));
+    }
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
-    fprintf(stream, "low_vram: %s # default: false\n", params.low_vram ? "true" : "false");
     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", params.mirostat);
@@ -1256,6 +1332,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
     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", params.temp);
 
     const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + LLAMA_MAX_DEVICES);

common/common.h

@@ -36,20 +36,23 @@ int32_t get_num_physical_cores();
 struct gpt_params {
     uint32_t seed                           = -1;   // RNG seed
     int32_t n_threads                       = get_num_physical_cores();
+    int32_t n_threads_batch                 = -1;   // number of threads to use for batch processing (-1 = use n_threads)
     int32_t n_predict                       = -1;   // new tokens to predict
     int32_t n_ctx                           = 512;  // context size
     int32_t n_batch                         = 512;  // batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                          = 0;    // number of tokens to keep from initial prompt
     int32_t n_draft                         = 16;   // number of tokens to draft during speculative decoding
     int32_t n_chunks                        = -1;   // max number of chunks to process (-1 = unlimited)
+    int32_t n_parallel                      = 1;    // number of parallel sequences to decode
+    int32_t n_sequences                     = 1;    // number of sequences to decode
     int32_t n_gpu_layers                    = -1;   // number of layers to store in VRAM (-1 - use default)
     int32_t n_gpu_layers_draft              = -1;   // number of layers to store in VRAM for the draft model (-1 - use default)
     int32_t main_gpu                        = 0;    // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0};  // how split tensors should be distributed across GPUs
     int32_t n_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                  = 10000.0f; // RoPE base frequency
-    float   rope_freq_scale                 = 1.0f;     // RoPE frequency scaling factor
+    float   rope_freq_base                  = 0.0f; // RoPE base frequency
+    float   rope_freq_scale                 = 0.0f; // RoPE frequency scaling factor
 
     // sampling parameters
     int32_t top_k             = 40;    // <= 0 to use vocab size
@@ -83,8 +86,8 @@ struct gpt_params {
     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
 
-    std::string lora_adapter = "";  // lora adapter path
-    std::string lora_base    = "";  // base model path for the lora adapter
+    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
 
     int  ppl_stride        = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
     int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
@@ -93,7 +96,6 @@ struct gpt_params {
     bool hellaswag         = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
-    bool low_vram          = false; // if true, reduce VRAM usage at the cost of performance
     bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
@@ -107,16 +109,16 @@ struct gpt_params {
     bool interactive_first = false; // wait for user input immediately
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
+    bool cont_batching     = false; // insert new sequences for decoding on-the-fly
 
     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 perplexity        = false; // compute perplexity over the prompt
+    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 export_cgraph     = false; // export the computation graph
     bool verbose_prompt    = false; // print prompt tokens before generation
 };
 
@@ -124,13 +126,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
 
 void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
 
+std::string get_system_info(const gpt_params & params);
+
 std::string gpt_random_prompt(std::mt19937 & rng);
 
+void process_escapes(std::string& input);
+
 //
 // Model utils
 //
 
 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_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
 //
@@ -140,7 +147,12 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
 // tokenizes a string into a vector of tokens
 // should work similar to Python's `tokenizer.encode`
 std::vector<llama_token> llama_tokenize(
-        struct llama_context * ctx,
+  const struct llama_context * ctx,
+           const std::string & text,
+                        bool   add_bos);
+
+std::vector<llama_token> llama_tokenize(
+    const struct llama_model * model,
            const std::string & text,
                         bool   add_bos);
 
@@ -181,7 +193,7 @@ std::string llama_detokenize_bpe(
 //  - 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(ctx) + idx * n_vocab
+//  - idx:           sample from llama_get_logits_ith(ctx, idx)
 //
 // returns:
 //  - token:      sampled token

common/log.h

@@ -225,31 +225,31 @@ enum LogTriState
 //  USE LOG() INSTEAD
 //
 #ifndef _MSC_VER
-    #define LOG_IMPL(str, ...)                                                                                          \
-    {                                                                                                               \
+    #define LOG_IMPL(str, ...)                                                                                      \
+    do {                                                                                                            \
         if (LOG_TARGET != nullptr)                                                                                  \
         {                                                                                                           \
             fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \
             fflush(LOG_TARGET);                                                                                     \
         }                                                                                                           \
-    }
+    } while (0)
 #else
-    #define LOG_IMPL(str, ...)                                                                                               \
-    {                                                                                                                    \
+    #define LOG_IMPL(str, ...)                                                                                           \
+    do {                                                                                                                 \
         if (LOG_TARGET != nullptr)                                                                                       \
         {                                                                                                                \
             fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \
             fflush(LOG_TARGET);                                                                                          \
         }                                                                                                                \
-    }
+    } while (0)
 #endif
 
 // INTERNAL, DO NOT USE
 //  USE LOG_TEE() INSTEAD
 //
 #ifndef _MSC_VER
-    #define LOG_TEE_IMPL(str, ...)                                                                                                          \
-    {                                                                                                                                   \
+    #define LOG_TEE_IMPL(str, ...)                                                                                                      \
+    do {                                                                                                                                \
         if (LOG_TARGET != nullptr)                                                                                                      \
         {                                                                                                                               \
             fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__);                     \
@@ -260,10 +260,10 @@ enum LogTriState
             fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL, __VA_ARGS__); \
             fflush(LOG_TEE_TARGET);                                                                                                     \
         }                                                                                                                               \
-    }
+    } while (0)
 #else
-    #define LOG_TEE_IMPL(str, ...)                                                                                                               \
-    {                                                                                                                                        \
+    #define LOG_TEE_IMPL(str, ...)                                                                                                           \
+    do {                                                                                                                                     \
         if (LOG_TARGET != nullptr)                                                                                                           \
         {                                                                                                                                    \
             fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__);                     \
@@ -274,7 +274,7 @@ enum LogTriState
             fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL "", ##__VA_ARGS__); \
             fflush(LOG_TEE_TARGET);                                                                                                          \
         }                                                                                                                                    \
-    }
+    } while (0)
 #endif
 
 // The '\0' as a last argument, is a trick to bypass the silly
@@ -435,41 +435,41 @@ inline FILE *log_handler() { return log_handler1_impl(); }
 inline void log_test()
 {
     log_disable();
-    LOG("01 Hello World to nobody, because logs are disabled!\n")
+    LOG("01 Hello World to nobody, because logs are disabled!\n");
     log_enable();
-    LOG("02 Hello World to default output, which is \"%s\" ( Yaaay, arguments! )!\n", LOG_STRINGIZE(LOG_TARGET))
-    LOG_TEE("03 Hello World to **both** default output and " LOG_TEE_TARGET_STRING "!\n")
+    LOG("02 Hello World to default output, which is \"%s\" ( Yaaay, arguments! )!\n", LOG_STRINGIZE(LOG_TARGET));
+    LOG_TEE("03 Hello World to **both** default output and " LOG_TEE_TARGET_STRING "!\n");
     log_set_target(stderr);
-    LOG("04 Hello World to stderr!\n")
-    LOG_TEE("05 Hello World TEE with double printing to stderr prevented!\n")
+    LOG("04 Hello World to stderr!\n");
+    LOG_TEE("05 Hello World TEE with double printing to stderr prevented!\n");
     log_set_target(LOG_DEFAULT_FILE_NAME);
-    LOG("06 Hello World to default log file!\n")
+    LOG("06 Hello World to default log file!\n");
     log_set_target(stdout);
-    LOG("07 Hello World to stdout!\n")
+    LOG("07 Hello World to stdout!\n");
     log_set_target(LOG_DEFAULT_FILE_NAME);
-    LOG("08 Hello World to default log file again!\n")
+    LOG("08 Hello World to default log file again!\n");
     log_disable();
-    LOG("09 Hello World _1_ into the void!\n")
+    LOG("09 Hello World _1_ into the void!\n");
     log_enable();
-    LOG("10 Hello World back from the void ( you should not see _1_ in the log or the output )!\n")
+    LOG("10 Hello World back from the void ( you should not see _1_ in the log or the output )!\n");
     log_disable();
     log_set_target("llama.anotherlog.log");
-    LOG("11 Hello World _2_ to nobody, new target was selected but logs are still disabled!\n")
+    LOG("11 Hello World _2_ to nobody, new target was selected but logs are still disabled!\n");
     log_enable();
-    LOG("12 Hello World this time in a new file ( you should not see _2_ in the log or the output )?\n")
+    LOG("12 Hello World this time in a new file ( you should not see _2_ in the log or the output )?\n");
     log_set_target("llama.yetanotherlog.log");
-    LOG("13 Hello World this time in yet new file?\n")
+    LOG("13 Hello World this time in yet new file?\n");
     log_set_target(log_filename_generator("llama_autonamed", "log"));
-    LOG("14 Hello World in log with generated filename!\n")
+    LOG("14 Hello World in log with generated filename!\n");
 #ifdef _MSC_VER
-    LOG_TEE("15 Hello msvc TEE without arguments\n")
-    LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test")
-    LOG_TEELN("17 Hello msvc TEELN without arguments\n")
-    LOG_TEELN("18 Hello msvc TEELN with (%d)(%s) arguments\n", 1, "test")
-    LOG("19 Hello msvc LOG without arguments\n")
-    LOG("20 Hello msvc LOG with (%d)(%s) arguments\n", 1, "test")
-    LOGLN("21 Hello msvc LOGLN without arguments\n")
-    LOGLN("22 Hello msvc LOGLN with (%d)(%s) arguments\n", 1, "test")
+    LOG_TEE("15 Hello msvc TEE without arguments\n");
+    LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test");
+    LOG_TEELN("17 Hello msvc TEELN without arguments\n");
+    LOG_TEELN("18 Hello msvc TEELN with (%d)(%s) arguments\n", 1, "test");
+    LOG("19 Hello msvc LOG without arguments\n");
+    LOG("20 Hello msvc LOG with (%d)(%s) arguments\n", 1, "test");
+    LOGLN("21 Hello msvc LOGLN without arguments\n");
+    LOGLN("22 Hello msvc LOGLN with (%d)(%s) arguments\n", 1, "test");
 #endif
 }
 
@@ -542,7 +542,7 @@ inline void log_dump_cmdline_impl(int argc, char **argv)
             buf << " " << argv[i];
         }
     }
-    LOGLN("Cmd:%s", buf.str().c_str())
+    LOGLN("Cmd:%s", buf.str().c_str());
 }
 
 #define log_tostr(var) log_var_to_string_impl(var).c_str()
@@ -620,10 +620,10 @@ inline std::string log_var_to_string_impl(const std::vector<int> & var)
 #define LOGLN(...) // dummy stub
 
 #undef LOG_TEE
-#define LOG_TEE(...) fprintf(stderr, __VA_ARGS__); // convert to normal fprintf
+#define LOG_TEE(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
 
 #undef LOG_TEELN
-#define LOG_TEELN(...) fprintf(stderr, __VA_ARGS__); // convert to normal fprintf
+#define LOG_TEELN(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
 
 #undef LOG_DISABLE
 #define LOG_DISABLE() // dummy stub

common/train.cpp

@@ -0,0 +1,1496 @@
+#include "train.h"
+#include "common.h"
+
+#include <random>
+#include <sstream>
+#include <functional>
+
+struct random_normal_distribution {
+    std::mt19937 gen;
+    std::normal_distribution<float> rd;
+    float min;
+    float max;
+};
+
+struct random_uniform_distribution {
+    std::mt19937 gen;
+    std::uniform_real_distribution<float> rd;
+};
+
+struct train_state  * init_train_state() {
+    struct train_state * state = new struct train_state;
+    state->train_its     = 0;
+    state->train_samples = 0;
+    state->train_tokens  = 0;
+    state->train_epochs  = 0;
+    state->shuffle_samples_hash  = 0;
+    state->shuffle_sample_count  = 0;
+    state->shuffle_next_sample   = 0;
+    state->shuffle_rng_state_current = "";
+    state->shuffle_rng_state_next    = "";
+
+    state->opt = new struct ggml_opt_context;
+    state->opt->ctx = NULL;
+    state->opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
+    state->opt->loss_after = 0.0f;
+
+    return state;
+}
+
+void free_train_state(struct train_state  * state) {
+    delete state->opt;
+    delete state;
+}
+
+struct random_normal_distribution * init_random_normal_distribution(
+    int seed, float mean, float std, float min, float max
+) {
+    struct random_normal_distribution * rnd = (struct random_normal_distribution *) malloc(sizeof(struct random_normal_distribution));
+    rnd->gen = std::mt19937(seed);
+    rnd->rd = std::normal_distribution<float>{mean, std};
+    rnd->min = min;
+    rnd->max = max;
+    return rnd;
+}
+
+struct random_uniform_distribution * init_random_uniform_distribution(int seed, float min, float max) {
+    struct random_uniform_distribution * rnd = (struct random_uniform_distribution *) malloc(sizeof(struct random_uniform_distribution));
+    rnd->gen = std::mt19937(seed);
+    rnd->rd = std::uniform_real_distribution<float>{min, max};
+    return rnd;
+}
+
+void free_random_normal_distribution (struct random_normal_distribution  * rnd) {
+    free(rnd);
+}
+
+void free_random_uniform_distribution(struct random_uniform_distribution * rnd) {
+    free(rnd);
+}
+
+struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct random_normal_distribution * rnd) {
+    float scale = 1.0f; // xavier
+    switch (tensor->n_dims) {
+        case 1:
+            scale /= sqrtf((float) tensor->ne[0]);
+            for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
+                *dst = scale * frand_normal(rnd);
+            }
+            break;
+        case 2:
+            scale /= sqrtf((float) tensor->ne[0]+tensor->ne[1]);
+            for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                    float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
+                    *dst = scale * frand_normal(rnd);
+                }
+            }
+            break;
+        case 3:
+            scale /= sqrtf((float) tensor->ne[0]+tensor->ne[1]);
+            for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                    for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                        float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2]);
+                        *dst = scale * frand_normal(rnd);
+                    }
+                }
+            }
+            break;
+        case 4:
+            scale /= sqrtf((float) tensor->ne[0]+tensor->ne[1]);
+            for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
+                for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                    for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                        for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                            float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3]);
+                            *dst = scale * frand_normal(rnd);
+                        }
+                    }
+                }
+            }
+            break;
+        default:
+            die("Unsupported tensor->n_dims");
+    };
+    return tensor;
+}
+
+struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd) {
+    switch (tensor->n_dims) {
+        case 1:
+            for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
+                *dst = frand_uniform(rnd);
+            }
+            break;
+        case 2:
+            for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                    float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
+                    *dst = frand_uniform(rnd);
+                }
+            }
+            break;
+        case 3:
+            for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                    for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                        float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2]);
+                        *dst = frand_uniform(rnd);
+                    }
+                }
+            }
+            break;
+        case 4:
+            for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
+                for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                    for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                        for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                            float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3]);
+                            *dst = frand_uniform(rnd);
+                        }
+                    }
+                }
+            }
+            break;
+        default:
+            die("Unsupported tensor->n_dims");
+    };
+    return tensor;
+}
+
+float frand() {
+    return (float)rand()/((float)(RAND_MAX) + 1.0f);
+}
+
+float frand_normal(struct random_normal_distribution * rnd) {
+    return fclamp(rnd->rd(rnd->gen), rnd->min, rnd->max);
+}
+
+float frand_uniform(struct random_uniform_distribution * rnd) {
+    return rnd->rd(rnd->gen);
+}
+
+int clamp(const int v, const int min, const int max) {
+    return ((v < min) ? (min) : (v > max) ? (max) : v);
+}
+
+float fclamp(const float v, const float min, const float max) {
+    return ((v < min) ? (min) : (v > max) ? (max) : v);
+}
+
+void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
+    GGML_ASSERT(tensor->n_dims == 1);
+    GGML_ASSERT(tensor->ne[0] == ne0);
+}
+
+void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
+    GGML_ASSERT(tensor->n_dims == 2);
+    GGML_ASSERT(tensor->ne[0] == ne0);
+    GGML_ASSERT(tensor->ne[1] == ne1);
+}
+
+void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
+    GGML_ASSERT(tensor->n_dims == 3);
+    GGML_ASSERT(tensor->ne[0] == ne0);
+    GGML_ASSERT(tensor->ne[1] == ne1);
+    GGML_ASSERT(tensor->ne[2] == ne2);
+}
+
+void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
+    GGML_ASSERT(tensor->n_dims == 4);
+    GGML_ASSERT(tensor->ne[0] == ne0);
+    GGML_ASSERT(tensor->ne[1] == ne1);
+    GGML_ASSERT(tensor->ne[2] == ne2);
+    GGML_ASSERT(tensor->ne[3] == ne3);
+}
+
+int64_t get_example_targets_batch(
+    struct llama_context * lctx,
+    struct ggml_tensor   * tokens_input,
+    struct ggml_tensor   * target_probs,
+    int64_t                example_id,
+    const size_t         * samples_offs,
+    const size_t         * samples_begin,
+    const size_t         * samples_size,
+          size_t           samples_count,
+    const llama_token    * train_data,
+    size_t                 n_train_data,
+    bool                   separate_with_eos,
+    bool                   separate_with_bos,
+    bool                   fill_with_next_samples,
+    bool                   sample_random_offsets
+) {
+    GGML_ASSERT(samples_count > 0);
+    GGML_ASSERT(tokens_input->n_dims  == 2);
+    GGML_ASSERT(target_probs->n_dims  == 3);
+    int64_t n_vocab  = target_probs->ne[0];
+    int64_t n_tokens = tokens_input->ne[0];
+    int64_t n_batch  = tokens_input->ne[1];
+    GGML_ASSERT(n_vocab  == target_probs->ne[0]);
+    GGML_ASSERT(n_tokens == target_probs->ne[1]);
+    GGML_ASSERT(n_batch  == target_probs->ne[2]);
+
+    int64_t used_samples = 0;
+
+    ggml_set_f32(target_probs, 0.0f);
+    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);
+        size_t sample_idx   = (example_id + used_samples) % samples_count;
+        size_t sample_offs  = sample_random_offsets ? samples_offs[sample_idx] : 0;
+        size_t sample_begin = samples_begin[sample_idx];
+        size_t sample_size  = samples_size[sample_idx];
+        ++used_samples;
+
+        // printf("%s: sample_idx=%zu sample=%zu\n", __func__, sample_idx, sample);
+        GGML_ASSERT(sample_begin+sample_size-1 < n_train_data);
+
+        ggml_set_i32_nd(tokens_input, 0, k, 0, 0, bos);
+        bool sample_separation_eos = !separate_with_eos;
+        bool sample_separation_bos = !separate_with_bos;
+        for (int64_t i=0; i<n_tokens; ++i) {
+            llama_token token = eos;
+            if (sample_offs >= sample_size && fill_with_next_samples) {
+                if (!sample_separation_eos) {
+                    // insert eos token to separate samples
+                    sample_separation_eos = true;
+                } else if (!sample_separation_bos) {
+                    // insert bos token to separate samples
+                    sample_separation_bos = true;
+                    token = bos;
+                } else {
+                    // sample separation is done, continue with next sample
+                    sample_separation_eos = !separate_with_eos;
+                    sample_separation_bos = !separate_with_bos;
+                    sample_offs  = 0;
+                    sample_idx   = (example_id + used_samples) % samples_count;
+                    sample_begin = samples_begin[sample_idx];
+                    sample_size  = samples_size[sample_idx];
+                    ++used_samples;
+                }
+            }
+            // note: no else-if here
+            if (sample_offs < sample_size) {
+                token = clamp(train_data[sample_begin+sample_offs], 0, (llama_token) (n_vocab - 1));
+                ++sample_offs;
+            }
+            ggml_set_f32_nd(target_probs,  token, (int) i, (int) k, 0, +1.0f);
+            if (i+1<n_tokens) {
+                ggml_set_i32_nd(tokens_input, (int) (i + 1), (int) k, 0, 0, token);
+            }
+        }
+    }
+
+    return used_samples;
+}
+
+void mt19937_set_state(std::mt19937& rng, const std::string& rng_state) {
+    std::stringstream s_rng_state;
+    s_rng_state.imbue(std::locale::classic());
+    s_rng_state.exceptions(std::stringstream::failbit);
+    s_rng_state.str(rng_state);
+    s_rng_state >> rng;
+}
+
+std::string mt19937_get_state(const std::mt19937& rng) {
+    std::stringstream s_rng_state;
+    s_rng_state.imbue(std::locale::classic());
+    s_rng_state << rng;
+    return s_rng_state.str();
+}
+
+std::string mt19937_seed_to_state(unsigned seed) {
+    std::mt19937 rng(seed);
+    return mt19937_get_state(rng);
+}
+
+std::string shuffle_samples(
+        const std::string & rng_state,
+        size_t            * shuffled_offs,
+        size_t            * shuffled_begins,
+        size_t            * shuffled_sizes,
+        const size_t      * begins,
+        const size_t      * sizes,
+        size_t              count) {
+    if (count == 0) return rng_state;
+
+    std::mt19937 rng;
+    mt19937_set_state(rng, rng_state);
+
+    // sort indices by random value for each index
+    std::vector<size_t> idcs;
+    {
+        std::vector<unsigned> rnd;
+        idcs.resize(count);
+        rnd.resize(count);
+        for (unsigned i=0; i<count; ++i) {
+            idcs[i] = i;
+            rnd[i]  = rng();
+        }
+
+        std::sort(idcs.begin(), idcs.end(), [&rnd](size_t a, size_t b){
+            // stable sort for reproducibility
+            return (rnd[a] == rnd[b]) ? (a < b) : (rnd[a] < rnd[b]);
+        });
+    }
+
+    // create random offsets
+    for (unsigned i=0; i<count; ++i) {
+        shuffled_offs[i] = (size_t) ((sizes[idcs[i]] - 1) * ((double) rng() / (double) (rng.max()-1)));
+    }
+
+    // reorder begins and sizes by sorted indices
+    for (unsigned i=0; i<count; ++i) {
+        shuffled_begins[i] = begins[idcs[i]];
+    }
+
+    for (unsigned i=0; i<count; ++i) {
+        shuffled_sizes[i] = sizes[idcs[i]];
+    }
+
+    return mt19937_get_state(rng);
+}
+
+size_t hash_combine(size_t h1, size_t h2) {
+    return h1 ^ (h2 << 1);
+}
+
+size_t compute_samples_hash(const char* fn, const size_t* samples_begin, const size_t* samples_size, size_t sample_count) {
+    std::hash<std::string> h_string;
+    std::hash<unsigned long long> h_ull;
+    size_t h = h_string(std::string(fn));
+    h = hash_combine(h, h_ull((unsigned long long) sample_count));
+    for (size_t i=0; i< sample_count; ++i) {
+        h = hash_combine(h, h_ull((unsigned long long) samples_begin[i]));
+        h = hash_combine(h, h_ull((unsigned long long) samples_size[i]));
+    }
+    return h;
+}
+
+std::string replace_str(const char * s, const char * needle, const char * replacement) {
+    std::string str = s;
+    size_t pos = str.find(needle);
+    if (pos != std::string::npos) {
+        str.replace(pos, strlen(needle), replacement);
+    }
+    return str;
+}
+
+void print_duration(double fmillis) {
+    if (fmillis < 1000.0f) {
+        printf("%.1fms", (float) fmillis);
+        return;
+    }
+    const int64_t one_sec  = 1000;
+    const int64_t one_min  = one_sec  * 60;
+    const int64_t one_hour = one_min  * 60;
+    const int64_t one_day  = one_hour * 24;
+
+    int64_t millis  = (int64_t) fmillis;
+    int64_t days    = millis/one_day;
+    int64_t hours   = (millis - days*one_day)/one_hour;
+    int64_t minutes = (millis - days*one_day - hours*one_hour)/one_min;
+    int64_t seconds = (millis - days*one_day - hours*one_hour - minutes*one_min)/one_sec;
+
+    // to print int64_t either cast to (long long int) or use macro PRId64 from <inttypes.h>
+    if (days > 0) {
+        printf("%lldd ", (long long int) days);
+    }
+    printf("%02lld:%02lld:%02lld", (long long int) hours, (long long int) minutes, (long long int) seconds);
+}
+
+float cosine_decay(int64_t step, int64_t decay_steps, float minimum) {
+    if (step > decay_steps) {
+        step = decay_steps;
+    }
+    const float cosine_decay = 0.50f*(1.0f + cosf(3.14159265359f*step/decay_steps));
+    const float decay = (1 - minimum)*cosine_decay + minimum;
+    return decay;
+}
+
+float cosine_decay_restart(int64_t step, int64_t decay_steps, float minimum, float restart_step_mult) {
+    while (step > decay_steps) {
+        step -= decay_steps;
+        decay_steps = (int64_t) (restart_step_mult * decay_steps);
+    }
+    return cosine_decay(step, decay_steps, minimum);
+}
+
+float learning_schedule(
+    int64_t step,
+    int64_t warmup_steps,
+    int64_t cos_decay_steps,
+    float   learning_rate,
+    float   overall_minimum,
+    float   cos_decay_minimum,
+    float   cos_decay_restart_step_mult,
+    bool    enable_restart) {
+
+    float result =
+        (step < warmup_steps)
+            ? (float) step / (float) warmup_steps
+            : enable_restart
+                ? cosine_decay_restart(
+                    step - warmup_steps,
+                    cos_decay_steps,
+                    cos_decay_minimum,
+                    cos_decay_restart_step_mult)
+                : cosine_decay(
+                    step,
+                    cos_decay_steps,
+                    cos_decay_minimum);
+
+    float min = overall_minimum / learning_rate;
+    result = min + result * (1.0f - min);
+    return result;
+}
+
+static bool are_same_layout(struct ggml_tensor * a, struct ggml_tensor * b) {
+    GGML_ASSERT(a != NULL);
+    GGML_ASSERT(b != NULL);
+    GGML_ASSERT(a->type == b->type);
+    GGML_ASSERT(ggml_are_same_shape(a, b));
+    GGML_ASSERT(ggml_is_contiguous(a) && ggml_is_contiguous(b));
+
+    return true;
+}
+
+void copy_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name) {
+    if (dst == NULL) {
+        return;
+    }
+    struct ggml_tensor * t  = ggml_get_tensor(ctx, name);
+    GGML_ASSERT(are_same_layout(dst, t));
+    memcpy(dst->data, t->data, ggml_nbytes(t));
+
+    if (strlen(ggml_get_name(dst)) == 0) {
+        ggml_set_name(dst, name);
+    }
+}
+
+// gguf constants
+static const char * LLM_KV_OPTIMIZER_TYPE = "optimizer.type";
+static const char * LLM_KV_OPTIMIZER_TYPE_ADAM  = "adam";
+static const char * LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs";
+static const char * LLM_KV_OPTIMIZER_FILE_VERSION               = "optimizer.file_version";
+static const char * LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT     = "optimizer.convergence_past_count";
+static const char * LLM_KV_OPTIMIZER_PARAMETER_COUNT            = "optimizer.parameter_count";
+static const char * LLM_KV_OPTIMIZER_ITERATION_COUNT            = "optimizer.iteration_count";
+static const char * LLM_KV_OPTIMIZER_JUST_INITIALIZED           = "optimizer.just_initialized";
+static const char * LLM_KV_OPTIMIZER_ADAM_BEST_LOSS             = "optimizer.adam.best_loss";
+static const char * LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS         = "optimizer.adam.previous_loss";
+static const char * LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT  = "optimizer.adam.no_improvement_count";
+static const char * LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count";
+static const char * LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS            = "optimizer.lbfgs.best_loss";
+static const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP     = "optimizer.lbfgs.line_search_step";
+static const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J        = "optimizer.lbfgs.line_search_j";
+static const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K        = "optimizer.lbfgs.line_search_k";
+static const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END      = "optimizer.lbfgs.line_search_end";
+static const char * LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count";
+
+static const char * LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS    = "optimizer.adam.first_moments";
+static const char * LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS   = "optimizer.adam.second_moments";
+static const char * LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values";
+
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS  = "optimizer.lbfgs.current_parameters";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS   = "optimizer.lbfgs.current_gradients";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS  = "optimizer.lbfgs.previous_gradients";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION    = "optimizer.lbfgs.search_direction";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES    = "optimizer.lbfgs.past_loss_values";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA        = "optimizer.lbfgs.memory_alpha";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS           = "optimizer.lbfgs.memory_ys";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S            = "optimizer.lbfgs.memory_s";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y            = "optimizer.lbfgs.memory_y";
+
+static const char * LLM_KV_TRAINING_FILE_VERSION         = "training.file_version";
+static const char * LLM_KV_TRAINING_ITERATION_COUNT      = "training.iteration_count";
+static const char * LLM_KV_TRAINING_SAMPLE_COUNT         = "training.sample_count";
+static const char * LLM_KV_TRAINING_TOKEN_COUNT          = "training.token_count";
+static const char * LLM_KV_TRAINING_EPOCH_COUNT          = "training.epoch_count";
+static const char * LLM_KV_TRAINING_SHUFFLE_SAMPLES_HASH = "training.shuffle.samples_hash";
+static const char * LLM_KV_TRAINING_SHUFFLE_RNG_STATE    = "training.shuffle.rng_state";
+static const char * LLM_KV_TRAINING_SHUFFLE_SAMPLE_COUNT = "training.shuffle.sample_count";
+static const char * LLM_KV_TRAINING_SHUFFLE_NEXT_SAMPLE  = "training.shuffle.next_sample";
+
+#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \
+{ \
+    const std::string skey(key); \
+    const int kid = gguf_find_key(ctx, skey.c_str()); \
+    if (kid >= 0) { \
+        enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \
+        if (ktype != (type)) { \
+            die_fmt("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype)); \
+        } \
+        (dst) = func(ctx, kid); \
+    } else if (req) { \
+        die_fmt("key not found in model: %s", skey.c_str()); \
+    } \
+}
+
+void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt) {
+    // NOTE: gguf_context must be initialized with f_ggml_ctx and no_alloc=false, otherwise tensor data can not be read
+
+    uint32_t file_version;
+    GGUF_GET_KEY(fctx, file_version, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_FILE_VERSION);
+    GGML_ASSERT(file_version == 0);
+
+    GGUF_GET_KEY(fctx, opt->params.past, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT);
+    GGUF_GET_KEY(fctx, opt->iter, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_ITERATION_COUNT);
+    GGUF_GET_KEY(fctx, opt->just_initialized, gguf_get_val_bool, GGUF_TYPE_BOOL, true, LLM_KV_OPTIMIZER_JUST_INITIALIZED);
+
+    uint64_t nx;
+    GGUF_GET_KEY(fctx, nx, gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_OPTIMIZER_PARAMETER_COUNT);
+    opt->nx = (size_t) nx;
+
+    // don't call ggml_opt_init until optimizer type and optimizer specific parameters are know
+
+    std::string opt_type;
+    GGUF_GET_KEY(fctx, opt_type, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_OPTIMIZER_TYPE);
+    if (opt_type == LLM_KV_OPTIMIZER_TYPE_ADAM) {
+        opt->params.type = GGML_OPT_ADAM;
+
+        GGUF_GET_KEY(fctx, opt->adam.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS);
+        GGUF_GET_KEY(fctx, opt->adam.fx_prev,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS);
+        GGUF_GET_KEY(fctx, opt->adam.n_no_improvement, gguf_get_val_u32, GGUF_TYPE_UINT32,  true, LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT);
+
+        ggml_opt_init(opt->ctx, opt, opt->params, opt->nx);
+
+        copy_tensor_by_name(opt->adam.m,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS);
+        copy_tensor_by_name(opt->adam.v,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
+        copy_tensor_by_name(opt->adam.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
+    } else if (opt_type == LLM_KV_OPTIMIZER_TYPE_LBFGS) {
+        opt->params.type = GGML_OPT_LBFGS;
+
+        GGUF_GET_KEY(fctx, opt->params.lbfgs.m,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT);
+        GGUF_GET_KEY(fctx, opt->lbfgs.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS);
+        GGUF_GET_KEY(fctx, opt->lbfgs.step,             gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP);
+        GGUF_GET_KEY(fctx, opt->lbfgs.j,                gguf_get_val_i32, GGUF_TYPE_INT32,   true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J);
+        GGUF_GET_KEY(fctx, opt->lbfgs.k,                gguf_get_val_i32, GGUF_TYPE_INT32,   true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K);
+        GGUF_GET_KEY(fctx, opt->lbfgs.end,              gguf_get_val_i32, GGUF_TYPE_INT32,   true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END);
+        GGUF_GET_KEY(fctx, opt->lbfgs.n_no_improvement, gguf_get_val_u32, GGUF_TYPE_UINT32,  true, LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT);
+
+        ggml_opt_init(opt->ctx, opt, opt->params, opt->nx);
+
+        copy_tensor_by_name(opt->lbfgs.x,    f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS);
+        copy_tensor_by_name(opt->lbfgs.xp,   f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS);
+        copy_tensor_by_name(opt->lbfgs.g,    f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS);
+        copy_tensor_by_name(opt->lbfgs.gp,   f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS);
+        copy_tensor_by_name(opt->lbfgs.d,    f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION);
+        copy_tensor_by_name(opt->lbfgs.pf,   f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES);
+        copy_tensor_by_name(opt->lbfgs.lmal, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA);
+        copy_tensor_by_name(opt->lbfgs.lmys, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS);
+        copy_tensor_by_name(opt->lbfgs.lms,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S);
+        copy_tensor_by_name(opt->lbfgs.lmy,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y);
+    } else {
+        die("unknown optimizer type\n");
+    }
+}
+
+void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt) {
+    gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_FILE_VERSION, 0);
+    gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT, opt->params.past);
+    gguf_set_val_u64(fctx, LLM_KV_OPTIMIZER_PARAMETER_COUNT, (uint64_t) opt->nx);
+    gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_ITERATION_COUNT, opt->iter);
+    gguf_set_val_bool(fctx, LLM_KV_OPTIMIZER_JUST_INITIALIZED, opt->just_initialized);
+
+    switch (opt->params.type) {
+        case GGML_OPT_ADAM:
+            {
+                gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM);
+                gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS,            opt->adam.fx_best);
+                gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS,        opt->adam.fx_prev);
+                gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT, opt->adam.n_no_improvement);
+
+                ggml_set_name(opt->adam.m, LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS);
+                ggml_set_name(opt->adam.v, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
+                if (opt->adam.pf) {
+                    ggml_set_name(opt->adam.pf, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
+                }
+
+                gguf_add_tensor(fctx, opt->adam.m);
+                gguf_add_tensor(fctx, opt->adam.v);
+                if (opt->adam.pf) {
+                    gguf_add_tensor(fctx, opt->adam.pf);
+                }
+            } break;
+        case GGML_OPT_LBFGS:
+            {
+                gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS);
+                gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, opt->params.lbfgs.m);
+                gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS,            opt->lbfgs.fx_best);
+                gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP,     opt->lbfgs.step);
+                gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J,        opt->lbfgs.j);
+                gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K,        opt->lbfgs.k);
+                gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END,      opt->lbfgs.end);
+                gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, opt->lbfgs.n_no_improvement);
+
+                ggml_set_name(opt->lbfgs.x,    LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS);
+                ggml_set_name(opt->lbfgs.xp,   LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS);
+                ggml_set_name(opt->lbfgs.g,    LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS);
+                ggml_set_name(opt->lbfgs.gp,   LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS);
+                ggml_set_name(opt->lbfgs.d,    LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION);
+                if (opt->lbfgs.pf) {
+                    ggml_set_name(opt->lbfgs.pf, LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES);
+                }
+                ggml_set_name(opt->lbfgs.lmal, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA);
+                ggml_set_name(opt->lbfgs.lmys, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS);
+                ggml_set_name(opt->lbfgs.lms,  LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S);
+                ggml_set_name(opt->lbfgs.lmy,  LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y);
+
+                gguf_add_tensor(fctx, opt->lbfgs.x);
+                gguf_add_tensor(fctx, opt->lbfgs.xp);
+                gguf_add_tensor(fctx, opt->lbfgs.g);
+                gguf_add_tensor(fctx, opt->lbfgs.gp);
+                gguf_add_tensor(fctx, opt->lbfgs.d);
+                if (opt->lbfgs.pf) {
+                    gguf_add_tensor(fctx, opt->lbfgs.pf);
+                }
+                gguf_add_tensor(fctx, opt->lbfgs.lmal);
+                gguf_add_tensor(fctx, opt->lbfgs.lmys);
+                gguf_add_tensor(fctx, opt->lbfgs.lms);
+                gguf_add_tensor(fctx, opt->lbfgs.lmy);
+            } break;
+    }
+}
+
+bool load_train_state_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct train_state * train) {
+    if (gguf_find_key(fctx, LLM_KV_TRAINING_FILE_VERSION) < 0) {
+        return false;
+    }
+
+    uint32_t file_version;
+    GGUF_GET_KEY(fctx, file_version,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_FILE_VERSION);
+    GGML_ASSERT(file_version <= 1);
+
+    if (file_version == 0) {
+
+        GGUF_GET_KEY(fctx, train->train_its,     gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_ITERATION_COUNT);
+        GGUF_GET_KEY(fctx, train->train_samples, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_SAMPLE_COUNT);
+        GGUF_GET_KEY(fctx, train->train_tokens,  gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_TOKEN_COUNT);
+
+    } else if (file_version == 1) {
+
+        GGUF_GET_KEY(fctx, train->train_its,     gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_TRAINING_ITERATION_COUNT);
+        GGUF_GET_KEY(fctx, train->train_samples, gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_TRAINING_SAMPLE_COUNT);
+        GGUF_GET_KEY(fctx, train->train_tokens,  gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_TRAINING_TOKEN_COUNT);
+        GGUF_GET_KEY(fctx, train->train_epochs,  gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_TRAINING_EPOCH_COUNT);
+
+        GGUF_GET_KEY(fctx, train->shuffle_samples_hash,      gguf_get_val_u64, GGUF_TYPE_UINT64, false, LLM_KV_TRAINING_SHUFFLE_SAMPLES_HASH);
+        GGUF_GET_KEY(fctx, train->shuffle_rng_state_current, gguf_get_val_str, GGUF_TYPE_STRING, false, LLM_KV_TRAINING_SHUFFLE_RNG_STATE);
+        GGUF_GET_KEY(fctx, train->shuffle_sample_count,      gguf_get_val_u64, GGUF_TYPE_UINT64, false, LLM_KV_TRAINING_SHUFFLE_SAMPLE_COUNT);
+        GGUF_GET_KEY(fctx, train->shuffle_next_sample,       gguf_get_val_u64, GGUF_TYPE_UINT64, false, LLM_KV_TRAINING_SHUFFLE_NEXT_SAMPLE);
+    }
+
+    load_opt_context_gguf(fctx, f_ggml_ctx, train->opt);
+    return true;
+}
+
+void save_train_state_gguf(struct gguf_context * fctx, struct train_state * train) {
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_FILE_VERSION,    1);
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_ITERATION_COUNT, train->train_its);
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_SAMPLE_COUNT,    train->train_samples);
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_TOKEN_COUNT,     train->train_tokens);
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_EPOCH_COUNT,     train->train_epochs);
+
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_SHUFFLE_SAMPLES_HASH, (uint64_t) train->shuffle_samples_hash);
+    gguf_set_val_str(fctx, LLM_KV_TRAINING_SHUFFLE_RNG_STATE,    train->shuffle_rng_state_current.c_str());
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_SHUFFLE_SAMPLE_COUNT, (uint64_t) train->shuffle_sample_count);
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_SHUFFLE_NEXT_SAMPLE,  (uint64_t) train->shuffle_next_sample);
+
+    save_opt_context_gguf(fctx, train->opt);
+}
+
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            size = 0;
+        } else {
+            seek(0, SEEK_END);
+            size = tell();
+            seek(0, SEEK_SET);
+        }
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        GGML_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            die_fmt("read error: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            die("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            die_fmt("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+static size_t utf8_len(char src) {
+    const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+    uint8_t highbits = static_cast<uint8_t>(src) >> 4;
+    return lookup[highbits];
+}
+
+// mark each byte with its utf8 unit number.
+// returns the number of utf8 characters.
+// e.g. when bytes == '\x61\xD0\xB0\x62',
+// then utf8_units will become [0,0,1,0]
+// utf8_nunits will become [1,2,2,1] and 3 is returned.
+// bytes where utf8_units is zero, are the begin of an utf8 character.
+static size_t mark_utf8_units(const char* bytes, int * utf8_units, int * utf8_nunits, size_t count) {
+    size_t offs = 0;
+    size_t count_utf8 = 0;
+    while(offs < count) {
+        int len = (int) utf8_len(bytes[offs]);
+        for (int i=0; i<len; ++i) {
+            utf8_units[offs+i]  = i;
+            utf8_nunits[offs+i] = len;
+        }
+        offs += len;
+        ++count_utf8;
+    }
+    return count_utf8;
+}
+
+size_t tokenize_file(
+        struct llama_context     * lctx,
+        const char               * filename,
+        const std::string        & sample_start,
+        bool                       include_sample_start,
+        bool                       overlapping_samples,
+        unsigned                   context_length,
+        std::vector<llama_token> & out_tokens,
+        std::vector<size_t>      & out_samples_begin,
+        std::vector<size_t>      & out_samples_size) {
+    struct llama_file f(filename, "rb");
+
+    if (f.size == 0) {
+        out_tokens.clear();
+        out_samples_begin.clear();
+        out_samples_size.clear();
+        printf("%s: warning: empty or not existing training data file '%s'\n",
+            __func__, filename);
+        return out_tokens.size();
+    }
+
+    // account for possible leading whitespace that will be added by tokenizer
+    // e.g. '\t' will be tokenized by llama spm tokenizer to [29871, 12]
+    const int n_max_tokens_overhead = 1;
+
+    std::vector<char> buf;
+    buf.resize(f.size);
+
+    f.read_raw(buf.data(), f.size);
+
+    std::vector<int> utf8_units;
+    std::vector<int> utf8_nunits;
+    utf8_units.resize(buf.size());
+    utf8_nunits.resize(buf.size());
+    mark_utf8_units(buf.data(), utf8_units.data(), utf8_nunits.data(), buf.size());
+
+    if (sample_start.size() == 0) {
+        // tokenize all data at once
+        out_tokens.resize(buf.size() + n_max_tokens_overhead);
+
+        int n_tokens = llama_tokenize(
+            llama_get_model(lctx),
+            buf.data(),
+            (int) buf.size(),
+            out_tokens.data(),
+            (int) out_tokens.size(),
+            false);
+        if (n_tokens < 0) {
+            out_tokens.resize(-n_tokens);
+            n_tokens = llama_tokenize(
+                llama_get_model(lctx),
+                buf.data(),
+                (int) buf.size(),
+                out_tokens.data(),
+                (int) out_tokens.size(),
+                false);
+        }
+        if (n_tokens >= 0) {
+            out_tokens.resize(n_tokens);
+        }
+
+        // generate sample starts at all token positions
+        out_samples_begin.clear();
+        out_samples_begin.push_back(0);
+        out_samples_size.push_back(std::min((size_t) context_length, out_tokens.size()));
+        size_t end = (out_tokens.size() >= context_length) ? (out_tokens.size() - context_length) : 0;
+        for (size_t sample_begin = 1; sample_begin < end; ++sample_begin) {
+            out_samples_begin.push_back(sample_begin);
+            out_samples_size.push_back(context_length);
+        }
+    } else {
+        // split data into samples and tokenize each sample
+        std::string data_str(buf.data(), buf.size());
+        out_samples_begin.clear();
+        out_samples_size.clear();
+        out_tokens.clear();
+
+        // find all positions of pattern sample_start
+        size_t sample_begin = data_str.find(sample_start, 0);
+        while (sample_begin != std::string::npos) {
+            out_samples_begin.push_back(sample_begin);
+            const size_t search_start = sample_begin + sample_start.size();
+            sample_begin = data_str.find(sample_start, search_start);
+        }
+        if (out_samples_begin.size() == 0) {
+            printf("%s: warning: sample start pattern '%s' not found. inserting single sample at data begin\n",
+                __func__, sample_start.c_str());
+            out_samples_begin.push_back(0);
+        }
+
+        out_samples_size.resize(out_samples_begin.size(), 0);
+
+        std::vector<char>        buf_sample;
+        std::vector<llama_token> tok_sample;
+
+        const size_t sample_begin_offset = (include_sample_start ? 0 : sample_start.size());
+        size_t found_too_big_sample   = 0;
+        size_t found_too_small_sample = 0;
+        size_t found_empty_sample     = 0;
+        size_t found_min_sample_size  = SIZE_MAX;
+        size_t found_max_sample_size  = 0;
+
+        size_t max_token_text_size = 0;
+        int n_vocab = llama_n_vocab(llama_get_model(lctx));
+        for (llama_token token=0; token < n_vocab; ++token) {
+            max_token_text_size = std::max(
+                max_token_text_size,
+                strlen(llama_token_get_text(lctx, token)));
+        }
+
+        // upper bound of context byte length.
+        // strings with this byte length should always tokenize to at least context_length tokens.
+        size_t context_byte_len = max_token_text_size*context_length;
+
+        for (unsigned i=0; i<out_samples_begin.size(); ++i) {
+            // determine sample begin and end from pattern positions
+            size_t sample_begin = out_samples_begin[i] + sample_begin_offset;
+            size_t sample_end   = overlapping_samples
+                                    ? std::min(
+                                        data_str.size(),
+                                        sample_begin + context_byte_len)
+                                    : (i+1 < out_samples_begin.size()
+                                        ? out_samples_begin[i+1]
+                                        : data_str.size());
+            if (sample_end < utf8_units.size() && utf8_units[sample_end] > 0) {
+                // sample end is in the middle of an utf8 character.
+                // advance sample_end to the begin of the next utf8 character.
+                sample_end += utf8_nunits[sample_end] - utf8_units[sample_end];
+            }
+            size_t sample_size = sample_end - sample_begin;
+            if (sample_size == 0) {
+                ++found_empty_sample;
+            }
+
+            if (sample_size > 0) {
+                // llama_tokenize expects zero terminated string,
+                // copy sample into buffer and zero terminate it.
+                buf_sample.resize(sample_size);
+                memcpy(buf_sample.data(), data_str.data() + sample_begin, sample_size);
+
+                // printf("sample: '%s'\n", buf_sample.data());
+
+                // tokenize the sample
+                tok_sample.resize(buf_sample.size() + n_max_tokens_overhead);
+                int n_tokens = llama_tokenize(llama_get_model(lctx),
+                    buf_sample.data(),
+                    (int) buf_sample.size(),
+                    tok_sample.data(),
+                    (int) tok_sample.size(),
+                    false);
+                if (n_tokens < 0) {
+                    tok_sample.resize(-n_tokens);
+                    n_tokens = llama_tokenize(llama_get_model(lctx),
+                        buf_sample.data(),
+                        (int) buf_sample.size(),
+                        tok_sample.data(),
+                        (int) tok_sample.size(),
+                        false);
+                    GGML_ASSERT(n_tokens >= 0);
+                }
+                GGML_ASSERT(n_tokens <= (int) tok_sample.size());
+
+                if ((size_t) n_tokens > context_length) {
+                    ++found_too_big_sample;
+                } else if ((size_t) n_tokens < context_length) {
+                    ++found_too_small_sample;
+                }
+                found_max_sample_size = std::max(found_max_sample_size, (size_t) n_tokens);
+                found_min_sample_size = std::min(found_min_sample_size, (size_t) n_tokens);
+
+                // write out tokens, start and size of sample
+                // overwrite the string start position with the token start position
+                out_samples_begin[i] = out_tokens.size();
+                out_samples_size[i] = (size_t) n_tokens;
+                out_tokens.insert(out_tokens.end(), tok_sample.begin(), tok_sample.begin() + n_tokens);
+            } else {
+                out_samples_begin[i] = out_tokens.size();
+                out_samples_size[i] = 0;
+            }
+
+        }
+        if (found_too_big_sample > 0) {
+            printf("%s: warning: found %zu samples (max length %zu) that exceed context length of %u. samples will be cut off.\n",
+                __func__, found_too_big_sample, found_max_sample_size, context_length);
+        }
+
+        if (found_too_small_sample > 0) {
+            printf("%s: warning: found %zu samples (min length %zu) that are shorter than context length of %u.\n",
+                __func__, found_too_small_sample, found_min_sample_size, context_length);
+        }
+
+        if (found_empty_sample) {
+            printf("%s: warning: found %zu empty samples.\n",
+                __func__, found_empty_sample);
+        }
+    }
+    printf("%s: total number of samples: %zu\n",
+        __func__, out_samples_begin.size());
+
+    GGML_ASSERT(out_samples_begin.size() == out_samples_size.size());
+
+    return out_tokens.size();
+}
+
+std::string get_train_filename(const char * filename, const char * pattern_it, const char * latest, int64_t iteration) {
+    std::string sit = (iteration >= 0) ? std::to_string(iteration) : std::string(latest);
+    return replace_str(filename, pattern_it, sit.c_str());
+}
+
+struct train_params_common get_default_train_params_common() {
+    struct train_params_common params;
+    params.fn_train_data     = "shakespeare.txt";
+    params.fn_checkpoint_in  = "checkpoint.gguf";
+    params.fn_checkpoint_out = "checkpoint-ITERATION.gguf";
+    params.pattern_fn_it     = "ITERATION";
+    params.fn_latest         = "LATEST";
+
+    params.print_usage = false;
+
+    params.save_every = 10;
+
+    params.seed       =   -1;
+
+    params.n_ctx      =  128;
+    params.n_threads  =    6;
+    params.n_batch    =    8;
+    params.n_gradient_accumulation = 1;
+    params.n_epochs   = -1;
+
+    params.custom_n_ctx = false;
+
+    params.use_flash              = true;
+    params.use_checkpointing      = true;
+
+    params.sample_start           = "";
+    params.include_sample_start   = false;
+    params.escape                 = false;
+    params.overlapping_samples    = false;
+    params.fill_with_next_samples = false;
+    params.separate_with_eos      = false;
+    params.separate_with_bos      = true;
+    params.sample_random_offsets  = false;
+    params.force_reshuffle        = false;
+
+    params.opt_past               = 0;
+    params.opt_delta              = 1e-5f;
+    params.opt_max_no_improvement = 0;
+
+    params.warmup            =  100;
+    params.cos_decay_steps   = 1000;
+    params.cos_decay_restart = 1.1f;
+    params.cos_decay_min     = 0.1f;
+    params.enable_restart    = false;
+
+    params.adam_n_iter         = 256;
+    params.adam_alpha          = 1e-3f;
+    params.adam_min_alpha      = 0;
+    params.adam_decay          = 1e-1f;
+    params.adam_decay_min_ndim = 2;
+    params.adam_beta1          = 0.9f;
+    params.adam_beta2          = 0.999f;
+    params.adam_gclip          = 1.0f;
+    params.adam_eps_f          = 0.0f;
+    return params;
+}
+
+void print_common_train_usage(int /*argc*/, char ** /*argv*/, const struct train_params_common * params) {
+    // fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    // fprintf(stderr, "\n");
+    // fprintf(stderr, "options:\n");
+    // fprintf(stderr, "  -h, --help                 show this help message and exit\n");
+    fprintf(stderr, "  --train-data FNAME         path from which to load training data (default '%s')\n", params->fn_train_data);
+    fprintf(stderr, "  --checkpoint-in FNAME      path from which to load training checkpoint (default '%s')\n", params->fn_checkpoint_in);
+    fprintf(stderr, "  --checkpoint-out FNAME     path to save training checkpoint (default '%s')\n", params->fn_checkpoint_out);
+    fprintf(stderr, "  --pattern-fn-it STR        pattern in output filenames to be replaced by iteration number (default '%s')\n", params->pattern_fn_it);
+    fprintf(stderr, "  --fn-latest STR            string to use instead of iteration number for saving latest output (default '%s')\n", params->fn_latest);
+    fprintf(stderr, "  --save-every N             save checkpoint and lora every N iterations. Disabled when N <= 0. (default '%d')\n", params->save_every);
+    fprintf(stderr, "  -s SEED, --seed SEED       RNG seed (default: -1, use random seed for -1)\n");
+    fprintf(stderr, "  -c N, --ctx N              Context size used during training (default %d)\n", params->n_ctx);
+    fprintf(stderr, "  -t N, --threads N          Number of threads (default %d)\n", params->n_threads);
+    fprintf(stderr, "  -b N, --batch N            Parallel batch size (default %d)\n", params->n_batch);
+    fprintf(stderr, "  --grad-acc N               Number of gradient accumulation steps (simulates larger batch size of batch*gradacc) (default %d)\n", params->n_gradient_accumulation);
+    fprintf(stderr, "  --sample-start STR         Sets the starting point for samples after the specified pattern. If empty use every token position as sample start. (default '%s')\n", params->sample_start.c_str());
+    fprintf(stderr, "  --include-sample-start     Include the sample start in the samples. (default off)\n");
+    fprintf(stderr, "  --escape                   process sample start escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
+    fprintf(stderr, "  --overlapping-samples      Samples my overlap, will include sample-start of second and following samples. When off, samples will end at begin of next sample. (default off)\n");
+    fprintf(stderr, "  --fill-with-next-samples   Samples shorter than context length will be followed by the next (shuffled) samples. (default off)\n");
+    fprintf(stderr, "  --separate-with-eos        When fill-with-next-samples, insert end-of-sequence token between samples.%s\n", params->separate_with_eos ? " (default)" : "");
+    fprintf(stderr, "  --separate-with-bos        When fill-with-next-samples, insert begin-of-sequence token between samples.%s\n", params->separate_with_bos ? " (default)" : "");
+    fprintf(stderr, "  --no-separate-with-eos     When fill-with-next-samples, don't insert end-of-sequence token between samples.%s\n", !params->separate_with_eos ? " (default)" : "");
+    fprintf(stderr, "  --no-separate-with-bos     When fill-with-next-samples, don't insert begin-of-sequence token between samples.%s\n", !params->separate_with_bos ? " (default)" : "");
+    fprintf(stderr, "  --sample-random-offsets    Use samples beginning at random offsets. Together with fill-with-next-samples this may help for training endless text generation.%s\n", params->sample_random_offsets ? " (default)" : "");
+    fprintf(stderr, "  --force-reshuffle          Force a reshuffling of data at program start, otherwise the shuffling of loaded checkpoint is resumed.\n");
+    fprintf(stderr, "  --no-flash                 Don't use flash attention \n");
+    fprintf(stderr, "  --use-flash                Use flash attention (default)\n");
+    fprintf(stderr, "  --no-checkpointing         Don't use gradient checkpointing\n");
+    fprintf(stderr, "  --use-checkpointing        Use gradient checkpointing (default)\n");
+    fprintf(stderr, "  --warmup N                 Only for Adam optimizer. Number of warmup steps (default %d)\n", params->warmup);
+    fprintf(stderr, "  --cos-decay-steps N        Only for Adam optimizer. Number of cosine decay steps (default %d)\n", params->cos_decay_steps);
+    fprintf(stderr, "  --cos-decay-restart N      Only for Adam optimizer. Increase of cosine decay steps after restart (default %f)\n", params->cos_decay_restart);
+    fprintf(stderr, "  --cos-decay-min N          Only for Adam optimizer. Cosine decay minimum (default %f)\n", params->cos_decay_min);
+    fprintf(stderr, "  --enable-restart N         Only for Adam optimizer. Enable restarts of cos-decay %s\n", params->enable_restart ? "(default)" : "");
+    fprintf(stderr, "  --disable-restart N        Only for Adam optimizer. Disable restarts of cos-decay %s\n", !params->enable_restart ? "(default)" : "");
+    fprintf(stderr, "  --opt-past N               Number of optimization iterations to track for delta convergence test. Disabled when zero. (default %d)\n", params->opt_past);
+    fprintf(stderr, "  --opt-delta N              Maximum delta for delta convergence test. Disabled when <= zero. (default %f)\n", params->opt_delta);
+    fprintf(stderr, "  --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. (default %d)\n", params->opt_max_no_improvement);
+    fprintf(stderr, "  --epochs N                 Maximum number epochs to process. (default %d)\n", params->n_epochs);
+    fprintf(stderr, "  --adam-iter N              Maximum number of Adam optimization iterations for each batch (default %d)\n", params->adam_n_iter);
+    fprintf(stderr, "  --adam-alpha N             Adam learning rate alpha (default %f)\n", params->adam_alpha);
+    fprintf(stderr, "  --adam-min-alpha N         Adam minimum learning rate alpha - including warmup phase (default %f)\n", params->adam_min_alpha);
+    fprintf(stderr, "  --adam-decay N             AdamW weight decay. Values greater zero enable AdamW instead of regular Adam. (default %f)\n", params->adam_decay);
+    fprintf(stderr, "  --adam-decay-min-ndim N    Minimum number of tensor dimensions to apply AdamW weight decay. Weight decay is not applied to tensors with less n_dims. (default %d)\n", params->adam_decay_min_ndim);
+    fprintf(stderr, "  --adam-beta1 N             AdamW beta1 in interval [0,1). How much to smooth the first moment of gradients. (default %f)\n", params->adam_beta1);
+    fprintf(stderr, "  --adam-beta2 N             AdamW beta2 in interval [0,1). How much to smooth the second moment of gradients. (default %f)\n", params->adam_beta2);
+    fprintf(stderr, "  --adam-gclip N             AdamW gradient clipping. Disabled when zero. (default %f)\n", params->adam_gclip);
+    fprintf(stderr, "  --adam-epsf N              AdamW epsilon for convergence test. Disabled when <= zero. (default %f)\n", params->adam_eps_f);
+    fprintf(stderr, "\n");
+}
+
+bool consume_common_train_arg(
+    int argc, char ** argv, int * idx, struct train_params_common * params, bool * invalid_param
+) {
+    int& i = *idx;
+    std::string arg = argv[i];
+    const std::string arg_prefix = "--";
+    if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+        std::replace(arg.begin(), arg.end(), '_', '-');
+    }
+    if (arg == "--train-data") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->fn_train_data = argv[i];
+    } else if (arg == "--checkpoint-in") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->fn_checkpoint_in = argv[i];
+    } else if (arg == "--checkpoint-out") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->fn_checkpoint_out = argv[i];
+    } else if (arg == "--pattern-fn-it") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->pattern_fn_it = argv[i];
+    } else if (arg == "--fn-latest") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->fn_latest = argv[i];
+    } else if (arg == "--save-every") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->save_every = std::stoi(argv[i]);
+    } else if (arg == "-s" || arg == "--seed") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->seed = std::stoi(argv[i]);
+    } else if (arg == "-c" || arg == "--ctx") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->n_ctx = std::stoi(argv[i]);
+        params->custom_n_ctx = true;
+    } else if (arg == "-t" || arg == "--threads") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->n_threads = std::stoi(argv[i]);
+    } else if (arg == "-b" || arg == "--batch") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->n_batch = std::stoi(argv[i]);
+    } else if (arg == "--grad-acc") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->n_gradient_accumulation = std::max(1, std::stoi(argv[i]));
+    } else if (arg == "--sample-start") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->sample_start = std::string(argv[i]);
+    } else if (arg == "--escape") {
+        params->escape = true;
+    } else if (arg == "--include-sample-start") {
+        params->include_sample_start = true;
+    } else if (arg == "--overlapping-samples") {
+        params->overlapping_samples = true;
+    } else if (arg == "--fill-with-next-samples") {
+        params->fill_with_next_samples = true;
+    } else if (arg == "--separate-with-eos") {
+        params->separate_with_eos = true;
+    } else if (arg == "--separate-with-bos") {
+        params->separate_with_bos = true;
+    } else if (arg == "--no-separate-with-eos") {
+        params->separate_with_eos = false;
+    } else if (arg == "--no-separate-with-bos") {
+        params->separate_with_bos = false;
+    } else if (arg == "--sample-random-offsets") {
+        params->sample_random_offsets = true;
+    } else if (arg == "--force-reshuffle") {
+        params->force_reshuffle = true;
+    } else if (arg == "--no-flash") {
+        params->use_flash = false;
+    } else if (arg == "--use-flash") {
+        params->use_flash = true;
+    } else if (arg == "--no-checkpointing") {
+        params->use_checkpointing = false;
+    } else if (arg == "--use-checkpointing") {
+        params->use_checkpointing = true;
+    } else if (arg == "--warmup") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->warmup = std::stoi(argv[i]);
+    } else if (arg == "--cos-decay-steps") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->cos_decay_steps = std::stoi(argv[i]);
+    } else if (arg == "--cos-decay-restart") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->cos_decay_restart = std::stof(argv[i]);
+    } else if (arg == "--cos-decay-min") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->cos_decay_min = std::stof(argv[i]);
+    } else if (arg == "--enable-restart") {
+        params->enable_restart = true;
+    } else if (arg == "--disable-restart") {
+        params->enable_restart = false;
+    } else if (arg == "--opt-past") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->opt_past = std::stoi(argv[i]);
+    } else if (arg == "--opt-delta") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->opt_delta = std::stof(argv[i]);
+    } else if (arg == "--opt-max-no-improvement") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->opt_max_no_improvement = std::stoi(argv[i]);
+    } else if (arg == "--adam-epsf") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_eps_f = std::stof(argv[i]);
+    } else if (arg == "--epochs") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->n_epochs = std::stoi(argv[i]);
+    } else if (arg == "--adam-iter") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_n_iter = std::stoi(argv[i]);
+    } else if (arg == "--adam-alpha") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_alpha = std::stof(argv[i]);
+    } else if (arg == "--adam-min-alpha") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_min_alpha = std::stof(argv[i]);
+    } else if (arg == "--adam-decay") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_decay = std::stof(argv[i]);
+    } else if (arg == "--adam-decay-min-ndim") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_decay_min_ndim = std::stoi(argv[i]);
+    } else if (arg == "--adam-beta1") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_beta1 = std::stof(argv[i]);
+    } else if (arg == "--adam-beta2") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_beta2 = std::stof(argv[i]);
+    } else if (arg == "--adam-gclip") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_gclip = std::stof(argv[i]);
+    } else if (arg == "-h" || arg == "--help") {
+        params->print_usage = true;
+        return true;
+    } else {
+        return false;
+    }
+    return true;
+}
+
+void finish_processing_train_args(struct train_params_common * params) {
+    if (params->escape) {
+        process_escapes(params->sample_start);
+    }
+}
+
+void train_opt_callback(void * vdata, int accum_step, float * sched, bool * cancel) {
+    struct train_opt_callback_data * data   = (struct train_opt_callback_data *) vdata;
+    struct train_params_common     * params = data->params;
+    struct train_state             * train  = data->train;
+    struct ggml_opt_context        * opt    = train->opt;
+    int n_batch = params->n_batch;
+    int n_ctx = params->n_ctx;
+
+    if (accum_step == 0) {
+        // time measurement
+        int64_t now = ggml_time_ms();
+        if (now > data->last_time && opt->iter > data->first_iter) {
+            double dt = (double) (now - data->last_time);
+            if (data->millis_per_iter == 0.0) {
+                data->millis_per_iter = dt;
+            } else {
+                const double gain = 0.7;
+                data->millis_per_iter = data->millis_per_iter*(1.0-gain) + dt*gain;
+            }
+        }
+
+        double remaining_millis = 0.0;
+        if (data->millis_per_iter > 0.0) {
+            const int n_iter = params->adam_n_iter;
+            const int done_iter = opt->iter - data->first_iter;
+            const int remaining_iter = n_iter - done_iter;
+            remaining_millis = remaining_iter * data->millis_per_iter;
+        }
+
+        // file saving
+        const bool save_now = (params->save_every > 0) && (opt->iter - data->last_save_iter >= params->save_every);
+        if (save_now) {
+            int new_iters = opt->iter - data->last_save_iter;
+            train->train_its    += new_iters;
+            train->train_tokens += new_iters * opt->params.n_gradient_accumulation * n_batch * n_ctx;
+
+            if (data->save_cb) {
+                data->save_cb(data->save_data, train);
+            }
+
+            data->last_save_iter = opt->iter;
+        }
+
+        // exclude file saving from time measurement, by measuring last_time after saving
+        data->last_time = ggml_time_ms();
+
+        *sched = learning_schedule(
+            opt->iter,
+            params->warmup,
+            params->cos_decay_steps,
+            params->adam_alpha,
+            params->adam_min_alpha,
+            params->cos_decay_min,
+            params->cos_decay_restart,
+            params->enable_restart);
+
+        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_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);
+
+
+        if (data->millis_per_iter > 0) {
+            printf(" dt=");
+            print_duration(data->millis_per_iter);
+            printf(" eta=");
+            print_duration(remaining_millis);
+        }
+
+        float improvement = opt->loss_before - opt->loss_after;
+        const float plot_scale = 10.0f;
+        int bar_len = (int)(1 + improvement*plot_scale + 0.5);
+        printf(" |");
+        for (int i=0; i<bar_len; ++i) {
+            printf("-");
+        }
+        printf(">");
+        printf("\n");
+    }
+
+    int64_t used_samples = get_example_targets_batch(
+        data->lctx,
+        data->tokens_input,
+        data->target_probs,
+        train->shuffle_next_sample,
+        data->shuffled_samples_offs,
+        data->shuffled_samples_begin,
+        data->shuffled_samples_size,
+        data->samples_count,
+        data->tokens_data,
+        data->tokens_size,
+        params->separate_with_eos,
+        params->separate_with_bos,
+        params->fill_with_next_samples,
+        params->sample_random_offsets);
+
+    train->train_samples += used_samples;
+    train->shuffle_next_sample += used_samples;
+
+    if (train->shuffle_next_sample >= train->shuffle_sample_count) {
+        ++train->train_epochs;
+        printf("%s: reshuffle samples. completed epochs: %llu\n", __func__, (long long unsigned) train->train_epochs);
+        // note: we may have used some samples from the current shuffling more than once
+        train->shuffle_rng_state_current = train->shuffle_rng_state_next;
+        train->shuffle_rng_state_next = shuffle_samples(
+            train->shuffle_rng_state_current,
+            data->shuffled_samples_offs,
+            data->shuffled_samples_begin,
+            data->shuffled_samples_size,
+            data->samples_begin,
+            data->samples_size,
+            data->samples_count);
+        train->shuffle_next_sample = 0;
+    }
+
+    const bool last_epoch_reached = (params->n_epochs > 0 && (int64_t) train->train_epochs - data->first_epoch >= params->n_epochs);
+    if (last_epoch_reached) {
+        // allow optimization iteration at last epoch to be completed before canceling
+        if (data->iter_at_last_epoch < 0) {
+            data->iter_at_last_epoch = opt->iter;
+        } else if (opt->iter > data->iter_at_last_epoch) {
+            *cancel = true;
+        }
+    }
+}

common/train.h

@@ -0,0 +1,230 @@
+// Various helper functions and utilities for training
+
+#pragma once
+
+#include <string>
+#include <random>
+#include <vector>
+
+#include "ggml.h"
+#include "llama.h"
+
+typedef std::string mt19937_state;
+
+struct train_state {
+    struct ggml_opt_context * opt;
+
+    uint64_t train_its;
+    uint64_t train_samples;
+    uint64_t train_tokens;
+    uint64_t train_epochs;
+
+    size_t        shuffle_samples_hash; // fn, sample_count, *zip(sample_begins, sample_sizes)
+    mt19937_state shuffle_rng_state_current;
+    mt19937_state shuffle_rng_state_next;
+    size_t        shuffle_sample_count;
+    size_t        shuffle_next_sample;
+};
+
+struct train_params_common {
+    const char * fn_train_data;
+    const char * fn_checkpoint_in;
+    const char * fn_checkpoint_out;
+    const char * pattern_fn_it;
+    const char * fn_latest;
+
+    bool print_usage;
+
+    int save_every;
+
+    uint32_t seed;
+
+    int n_ctx;
+    int n_threads;
+    int n_batch;
+    int n_gradient_accumulation;
+    int n_epochs;
+
+    bool custom_n_ctx;
+
+    bool use_flash;
+    bool use_checkpointing;
+
+    std::string sample_start;
+    bool include_sample_start;
+    bool escape;
+    bool overlapping_samples;
+    bool fill_with_next_samples;
+    bool separate_with_eos;
+    bool separate_with_bos;
+    bool sample_random_offsets;
+
+    bool force_reshuffle;
+
+    int   warmup;
+    int   cos_decay_steps;
+    float cos_decay_restart;
+    float cos_decay_min;
+    bool  enable_restart;
+
+    int   opt_past;
+    float opt_delta;
+    int   opt_max_no_improvement;
+
+    int   adam_n_iter;
+    float adam_alpha;
+    float adam_min_alpha;
+    float adam_decay;
+    int   adam_decay_min_ndim;
+    float adam_beta1;
+    float adam_beta2;
+    float adam_gclip;
+    float adam_eps_f;
+};
+
+typedef void (*save_train_files_callback)(void * data, struct train_state * train);
+
+struct train_opt_callback_data {
+    struct train_params_common * params;
+    struct train_state         * train;
+    save_train_files_callback    save_cb;
+    void                       * save_data;
+    struct llama_context       * lctx;
+    int                          last_save_iter;
+    llama_token                * tokens_data;
+    size_t                       tokens_size;
+    size_t                     * samples_begin;
+    size_t                     * samples_size;
+    size_t                     * shuffled_samples_offs;
+    size_t                     * shuffled_samples_begin;
+    size_t                     * shuffled_samples_size;
+    size_t                       samples_count;
+    struct ggml_tensor         * tokens_input;
+    struct ggml_tensor         * target_probs;
+    int                          first_iter;
+    int                          first_epoch;
+    int                          iter_at_last_epoch;
+    int64_t                      last_time;
+    double                       millis_per_iter;
+};
+
+struct train_state * init_train_state();
+void free_train_state(struct train_state  * state);
+
+struct train_params_common get_default_train_params_common();
+void print_common_train_usage(int /*argc*/, char ** argv, const struct train_params_common * params);
+
+bool consume_common_train_arg(int argc, char ** argv, int * idx, struct train_params_common * params, bool * invalid_param);
+void finish_processing_train_args(struct train_params_common * params);
+
+struct random_normal_distribution;
+struct random_uniform_distribution;
+
+struct random_normal_distribution  * init_random_normal_distribution (int seed, float mean, float std, float min, float max);
+struct random_uniform_distribution * init_random_uniform_distribution(int seed, float min, float max);
+
+void free_random_normal_distribution (struct random_normal_distribution  * rnd);
+void free_random_uniform_distribution(struct random_uniform_distribution * rnd);
+
+struct ggml_tensor * randomize_tensor_normal (struct ggml_tensor * tensor, struct random_normal_distribution * rnd);
+struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd);
+
+// generate random float in interval [0,1)
+float frand();
+float frand_normal (struct random_normal_distribution * rnd);
+float frand_uniform(struct random_uniform_distribution * rnd);
+
+int   clamp (const int v, const int min, const int max);
+float fclamp(const float v, const float min, const float max);
+
+void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0);
+void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1);
+void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2);
+void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3);
+
+size_t tokenize_file(
+        struct llama_context     * lctx,
+        const char               * filename,
+        const std::string        & sample_start,
+        bool                       include_sample_start,
+        bool                       overlapping_samples,
+        unsigned                   context_length,
+        std::vector<llama_token> & out_tokens,
+        std::vector<size_t>      & out_samples_begin,
+        std::vector<size_t>      & out_samples_size);
+
+int64_t get_example_targets_batch(
+        struct llama_context * lctx,
+        struct ggml_tensor   * tokens_input,
+        struct ggml_tensor   * target_probs,
+        int64_t                example_id,
+        const size_t         * samples_offs,
+        const size_t         * samples_begin,
+        const size_t         * samples_size,
+              size_t           samples_count,
+        const llama_token    * train_data,
+        size_t                 n_train_data,
+        bool                   separate_with_eos,
+        bool                   separate_with_bos,
+        bool                   fill_with_next_samples,
+        bool                   sample_random_offsets);
+
+
+void          mt19937_set_state(std::mt19937& rng, const mt19937_state& rng_state);
+mt19937_state mt19937_get_state(const std::mt19937& rng);
+mt19937_state mt19937_seed_to_state(unsigned seed);
+
+mt19937_state shuffle_samples(
+        const mt19937_state & rng_state,
+        size_t              * shuffled_offs,
+        size_t              * shuffled_begins,
+        size_t              * shuffled_sizes,
+        const size_t        * begins,
+        const size_t        * sizes,
+        size_t                count);
+
+size_t hash_combine(size_t h1, size_t h2);
+
+size_t compute_samples_hash(
+    const char* fn,
+    const size_t* samples_begin,
+    const size_t* samples_size,
+    size_t sample_count);
+
+
+std::string replace_str(const char * s, const char * needle, const char * replacement);
+
+void print_duration(double milliseconds);
+
+float cosine_decay(
+    int64_t step,
+    int64_t decay_steps,
+    float   minimum);
+
+float cosine_decay_restart(
+    int64_t step,
+    int64_t decay_steps,
+    float   minimum,
+    float   restart_step_mult);
+
+float learning_schedule(
+    int64_t step,
+    int64_t warmup_steps,
+    int64_t decay_steps,
+    float   learning_rate,
+    float   overall_minimum,
+    float   cos_decay_minimum,
+    float   cos_decay_restart_step_mult,
+    bool    enable_restart);
+
+void copy_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name);
+
+void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt);
+void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt);
+
+bool load_train_state_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct train_state * train);
+void save_train_state_gguf(struct gguf_context * fctx, struct train_state * train);
+
+std::string get_train_filename(const char * filename, const char * pattern_it, const char * latest, int64_t iteration);
+
+void train_opt_callback(void * vdata, int accum_step, float * sched, bool * cancel);

convert-falcon-hf-to-gguf.py

@@ -133,8 +133,6 @@ gguf_writer.add_file_type(ftype)
 print("gguf: get tokenizer metadata")
 
 tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
 
 tokenizer_json_file = dir_model / 'tokenizer.json'
 if not tokenizer_json_file.is_file():
@@ -177,12 +175,8 @@ for i in range(vocab_size):
         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)
 special_vocab.add_to_gguf(gguf_writer)

convert-starcoder-hf-to-gguf.py

@@ -117,8 +117,6 @@ gguf_writer.add_file_type(ftype)
 print("gguf: get tokenizer metadata")
 
 tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
 
 tokenizer_json_file = dir_model / 'tokenizer.json'
 if not tokenizer_json_file.is_file():
@@ -161,12 +159,8 @@ for i in range(vocab_size):
         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)
 special_vocab.add_to_gguf(gguf_writer)

convert.py

@@ -439,7 +439,7 @@ Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab'
 def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
     #print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) )
     if n_head_kv is not None and n_head != n_head_kv:
-        n_head //= n_head_kv
+        n_head = n_head_kv
     return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
                 .swapaxes(1, 2)
                 .reshape(weights.shape))

examples/CMakeLists.txt

@@ -21,9 +21,12 @@ else()
     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)
@@ -33,4 +36,5 @@ else()
     if (LLAMA_BUILD_SERVER)
         add_subdirectory(server)
     endif()
+    add_subdirectory(export-lora)
 endif()

examples/baby-llama/baby-llama.cpp

@@ -1,8 +1,12 @@
 #include "ggml.h"
+#include "train.h"
+
 #include <vector>
 #include <cassert>
-#include <random>
+#include <cstdlib>
 #include <cstring>
+#include <random>
+#include <vector>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -14,31 +18,6 @@ constexpr float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
 constexpr float rms_norm_eps = 5e-6f;
 #endif
 
-static float frand() {
-    return (float)rand()/(float)RAND_MAX;
-}
-
-struct random_normal_distribution {
-    std::mt19937 gen;
-    std::normal_distribution<float> nd;
-    float min;
-    float max;
-};
-
-static void init_random_normal_distribution(
-    struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max
-) {
-    rnd->gen = std::mt19937(seed);
-    rnd->nd = std::normal_distribution<float>{mean, std};
-    rnd->min = min;
-    rnd->max = max;
-}
-
-static float frand_normal(struct random_normal_distribution * rnd) {
-    const float r = rnd->nd(rnd->gen);
-    return ((r < rnd->min) ? (rnd->min) : (r > rnd->max) ? (rnd->max) : r);
-}
-
 static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
     struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
 
@@ -88,55 +67,7 @@ static struct ggml_tensor * randomize_tensor(
             break;
         default:
             assert(false);
-    };
-
-    return tensor;
-}
-
-static struct ggml_tensor * randomize_tensor_normal(
-    struct ggml_tensor * tensor, int ndims, const int64_t ne[], struct random_normal_distribution * rnd
-) {
-    float scale = 1.0; // xavier
-    switch (ndims) {
-        case 1:
-            scale /= sqrtf(ne[0]);
-            for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((float *)tensor->data)[i0] = scale * frand_normal(rnd);
-            }
-            break;
-        case 2:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i1 = 0; i1 < ne[1]; i1++) {
-                for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((float *)tensor->data)[i1*ne[0] + i0] = scale * frand_normal(rnd);
-                }
-            }
-            break;
-        case 3:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i2 = 0; i2 < ne[2]; i2++) {
-                for (int i1 = 0; i1 < ne[1]; i1++) {
-                    for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
-                    }
-                }
-            }
-            break;
-        case 4:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i3 = 0; i3 < ne[3]; i3++) {
-                for (int i2 = 0; i2 < ne[2]; i2++) {
-                    for (int i1 = 0; i1 < ne[1]; i1++) {
-                        for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
-                        }
-                    }
-                }
-            }
-            break;
-        default:
-            assert(false);
-    };
+    }
 
     return tensor;
 }
@@ -398,27 +329,29 @@ static void randomize_model(struct llama_model * model, int seed, float mean, fl
 
     const uint32_t n_layer = hparams.n_layer;
 
-    struct random_normal_distribution rnd;
-    init_random_normal_distribution(&rnd, seed, mean, std, min, max);
-    randomize_tensor_normal(model->tok_embeddings, model->tok_embeddings->n_dims, model->tok_embeddings->ne, &rnd);
-    randomize_tensor_normal(model->norm,           model->norm->n_dims,           model->norm->ne,           &rnd);
-    randomize_tensor_normal(model->output,         model->output->n_dims,         model->output->ne,         &rnd);
+    struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
+
+    randomize_tensor_normal(model->tok_embeddings , rnd);
+    randomize_tensor_normal(model->norm           , rnd);
+    randomize_tensor_normal(model->output         , rnd);
 
     for (uint32_t i = 0; i < n_layer; ++i) {
         auto & layer = model->layers[i];
-        randomize_tensor_normal(layer.attention_norm, layer.attention_norm->n_dims, layer.attention_norm->ne, &rnd);
+        randomize_tensor_normal(layer.attention_norm, rnd);
 
-        randomize_tensor_normal(layer.wq, layer.wq->n_dims, layer.wq->ne, &rnd);
-        randomize_tensor_normal(layer.wk, layer.wk->n_dims, layer.wk->ne, &rnd);
-        randomize_tensor_normal(layer.wv, layer.wv->n_dims, layer.wv->ne, &rnd);
-        randomize_tensor_normal(layer.wo, layer.wo->n_dims, layer.wo->ne, &rnd);
+        randomize_tensor_normal(layer.wq, rnd);
+        randomize_tensor_normal(layer.wk, rnd);
+        randomize_tensor_normal(layer.wv, rnd);
+        randomize_tensor_normal(layer.wo, rnd);
 
-        randomize_tensor_normal(layer.ffn_norm, layer.ffn_norm->n_dims, layer.ffn_norm->ne, &rnd);
+        randomize_tensor_normal(layer.ffn_norm, rnd);
 
-        randomize_tensor_normal(layer.w1, layer.w1->n_dims, layer.w1->ne, &rnd);
-        randomize_tensor_normal(layer.w2, layer.w2->n_dims, layer.w2->ne, &rnd);
-        randomize_tensor_normal(layer.w3, layer.w3->n_dims, layer.w3->ne, &rnd);
+        randomize_tensor_normal(layer.w1, rnd);
+        randomize_tensor_normal(layer.w2, rnd);
+        randomize_tensor_normal(layer.w3, rnd);
     }
+
+    free_random_normal_distribution(rnd);
 }
 
 
@@ -429,35 +362,37 @@ static void randomize_model_lora(
 
     const uint32_t n_layer = hparams.n_layer;
 
-    struct random_normal_distribution rnd;
-    init_random_normal_distribution(&rnd, seed, mean, std, min, max);
-    randomize_tensor_normal(model->tok_embeddings, model->tok_embeddings->n_dims, model->tok_embeddings->ne, &rnd);
-    randomize_tensor_normal(model->norm,           model->norm->n_dims,           model->norm->ne,           &rnd);
-    randomize_tensor_normal(model->outputa,        model->outputa->n_dims,        model->outputa->ne,         &rnd);
-    randomize_tensor_normal(model->outputb,        model->outputb->n_dims,        model->outputb->ne,         &rnd);
+    struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
+
+    randomize_tensor_normal(model->tok_embeddings, rnd);
+    randomize_tensor_normal(model->norm          , rnd);
+    randomize_tensor_normal(model->outputa       , rnd);
+    randomize_tensor_normal(model->outputb       , rnd);
 
     for (uint32_t i = 0; i < n_layer; ++i) {
         auto & layer = model->layers[i];
-        randomize_tensor_normal(layer.attention_norm, layer.attention_norm->n_dims, layer.attention_norm->ne, &rnd);
-
-        randomize_tensor_normal(layer.wqa, layer.wqa->n_dims, layer.wqa->ne, &rnd);
-        randomize_tensor_normal(layer.wqb, layer.wqb->n_dims, layer.wqb->ne, &rnd);
-        randomize_tensor_normal(layer.wka, layer.wka->n_dims, layer.wka->ne, &rnd);
-        randomize_tensor_normal(layer.wkb, layer.wkb->n_dims, layer.wkb->ne, &rnd);
-        randomize_tensor_normal(layer.wva, layer.wva->n_dims, layer.wva->ne, &rnd);
-        randomize_tensor_normal(layer.wvb, layer.wvb->n_dims, layer.wvb->ne, &rnd);
-        randomize_tensor_normal(layer.woa, layer.woa->n_dims, layer.woa->ne, &rnd);
-        randomize_tensor_normal(layer.wob, layer.wob->n_dims, layer.wob->ne, &rnd);
-
-        randomize_tensor_normal(layer.ffn_norm, layer.ffn_norm->n_dims, layer.ffn_norm->ne, &rnd);
-
-        randomize_tensor_normal(layer.w1, layer.w1->n_dims, layer.w1->ne, &rnd);
-        randomize_tensor_normal(layer.w2, layer.w2->n_dims, layer.w2->ne, &rnd);
-        randomize_tensor_normal(layer.w3, layer.w3->n_dims, layer.w3->ne, &rnd);
+        randomize_tensor_normal(layer.attention_norm, rnd);
+
+        randomize_tensor_normal(layer.wqa, rnd);
+        randomize_tensor_normal(layer.wqb, rnd);
+        randomize_tensor_normal(layer.wka, rnd);
+        randomize_tensor_normal(layer.wkb, rnd);
+        randomize_tensor_normal(layer.wva, rnd);
+        randomize_tensor_normal(layer.wvb, rnd);
+        randomize_tensor_normal(layer.woa, rnd);
+        randomize_tensor_normal(layer.wob, rnd);
+
+        randomize_tensor_normal(layer.ffn_norm, rnd);
+
+        randomize_tensor_normal(layer.w1, rnd);
+        randomize_tensor_normal(layer.w2, rnd);
+        randomize_tensor_normal(layer.w3, rnd);
     }
+
+    free_random_normal_distribution(rnd);
 }
 
-static bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
+static void init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_ctx   = hparams.n_ctx;
@@ -483,14 +418,12 @@ static bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * mod
 
         if (!cache->ctx) {
             fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
-            return false;
+            exit(1);
         }
     }
 
     cache->k = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
     cache->v = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
-
-    return true;
 }
 
 static bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) {
@@ -554,6 +487,14 @@ static struct ggml_tensor * forward(
     struct ggml_tensor * kc = kv_self.k;
     struct ggml_tensor * vc = kv_self.v;
 
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
     // inpL shape [n_embd,N,1,1]
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
     for (int il = 0; il < n_layer; ++il) {
@@ -581,8 +522,8 @@ static struct ggml_tensor * forward(
             // wk   shape [n_embd, n_embd, 1, 1]
             // Qcur shape [n_embd/n_head, n_head, N, 1]
             // Kcur shape [n_embd/n_head, n_head, N, 1]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
 
             // store key and value to memory
             {
@@ -754,32 +695,6 @@ static struct ggml_tensor * forward(
     return inpL;
 }
 
-static void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
-    GGML_ASSERT(tensor->n_dims == 1);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-}
-
-static void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
-    GGML_ASSERT(tensor->n_dims == 2);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-}
-
-static void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
-    GGML_ASSERT(tensor->n_dims == 3);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == ne2);
-}
-
-static void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
-    GGML_ASSERT(tensor->n_dims == 4);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == ne2);
-    GGML_ASSERT(tensor->ne[3] == ne3);
-}
-
 static struct ggml_tensor * forward_batch(
     struct llama_model    * model,
     struct llama_kv_cache * cache,
@@ -808,9 +723,18 @@ static struct ggml_tensor * forward_batch(
     struct ggml_tensor * kc = kv_self.k;
     struct ggml_tensor * vc = kv_self.v;
 
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
     // inpL shape [n_embd,N*n_batch,1]
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
     assert_shape_2d(inpL, n_embd, N*n_batch);
+
     for (int il = 0; il < n_layer; ++il) {
         struct ggml_tensor * inpSA = inpL;
 
@@ -838,8 +762,8 @@ static struct ggml_tensor * forward_batch(
             // wk   shape [n_embd, n_embd, 1, 1]
             // Qcur shape [n_embd/n_head, n_head, N, n_batch]
             // Kcur shape [n_embd/n_head, n_head, N, n_batch]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
             assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
             assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
 
@@ -1097,6 +1021,14 @@ static struct ggml_tensor * forward_lora(
     struct ggml_tensor * kc = kv_self.k;
     struct ggml_tensor * vc = kv_self.v;
 
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
     // inpL shape [n_embd,N,1,1]
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
     for (int il = 0; il < n_layer; ++il) {
@@ -1130,7 +1062,7 @@ static struct ggml_tensor * forward_lora(
                                                         model->layers[il].wqb,
                                                         cur)),
                                                 n_embd/n_head, n_head, N),
-                                            n_past, n_rot, 0, 0);
+                                            KQ_pos, n_rot, 0, 0);
             struct ggml_tensor * Kcur = ggml_rope(ctx0,
                                             ggml_reshape_3d(ctx0,
                                                 ggml_mul_mat(ctx0,
@@ -1139,7 +1071,7 @@ static struct ggml_tensor * forward_lora(
                                                         model->layers[il].wkb,
                                                         cur)),
                                                 n_embd/n_head, n_head, N),
-                                            n_past, n_rot, 0, 0);
+                                            KQ_pos, n_rot, 0, 0);
 
             // store key and value to memory
             {

examples/batched/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET batched)
+add_executable(${TARGET} batched.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/README.md

@@ -0,0 +1,44 @@
+# llama.cpp/example/batched
+
+The example demonstrates batched generation from a given prompt
+
+```bash
+./batched ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is" 4
+
+...
+
+main: n_len = 32, n_ctx = 2048, n_parallel = 4, n_kv_req = 113
+
+ Hello my name is
+
+main: generating 4 sequences ...
+
+main: stream 0 finished
+main: stream 1 finished
+main: stream 2 finished
+main: stream 3 finished
+
+sequence 0:
+
+Hello my name is Shirley. I am a 25-year-old female who has been working for over 5 years as a b
+
+sequence 1:
+
+Hello my name is Renee and I'm a 32 year old female from the United States. I'm looking for a man between
+
+sequence 2:
+
+Hello my name is Diana. I am looking for a housekeeping job. I have experience with children and have my own transportation. I am
+
+sequence 3:
+
+Hello my name is Cody. I am a 3 year old neutered male. I am a very friendly cat. I am very playful and
+
+main: decoded 108 tokens in 3.57 s, speed: 30.26 t/s
+
+llama_print_timings:        load time =   587.00 ms
+llama_print_timings:      sample time =     2.56 ms /   112 runs   (    0.02 ms per token, 43664.72 tokens per second)
+llama_print_timings: prompt eval time =  4089.11 ms /   118 tokens (   34.65 ms per token,    28.86 tokens per second)
+llama_print_timings:        eval time =     0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_print_timings:       total time =  4156.04 ms
+```

examples/batched/batched.cpp

@@ -0,0 +1,255 @@
+#include "common.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (argc == 1 || argv[1][0] == '-') {
+        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL]\n" , argv[0]);
+        return 1 ;
+    }
+
+    int n_parallel = 1;
+
+    if (argc >= 2) {
+        params.model = argv[1];
+    }
+
+    if (argc >= 3) {
+        params.prompt = argv[2];
+    }
+
+    if (argc >= 4) {
+        n_parallel = std::atoi(argv[3]);
+    }
+
+    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);
+
+    // initialize the model
+
+    llama_model_params model_params = llama_model_default_params();
+
+    // 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);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    // tokenize the prompt
+
+    std::vector<llama_token> tokens_list;
+    tokens_list = ::llama_tokenize(model, params.prompt, true);
+    const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size())*n_parallel;
+
+    // initialize the context
+
+    llama_context_params ctx_params = llama_context_default_params();
+
+    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_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;
+    }
+
+    const int n_ctx    = llama_n_ctx(ctx);
+
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+
+    // make sure the KV cache is big enough to hold all the prompt and generated tokens
+    if (n_kv_req > n_ctx) {
+        LOG_TEE("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__,  n_kv_req);
+        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
+        return 1;
+    }
+
+    // print the prompt token-by-token
+
+    fprintf(stderr, "\n");
+
+    for (auto id : tokens_list) {
+        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
+    }
+
+    fflush(stderr);
+
+    // 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);
+
+    // evaluate the initial prompt
+    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
+    batch.logits[batch.n_tokens - 1] = true;
+
+    if (llama_decode(ctx, batch) != 0) {
+        LOG_TEE("%s: llama_decode() failed\n", __func__);
+        return 1;
+    }
+
+    // assign the system KV cache to all parallel sequences
+    // this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
+    for (int32_t i = 1; i < n_parallel; ++i) {
+        llama_kv_cache_seq_cp(ctx, 0, i, 0, batch.n_tokens);
+    }
+
+    if (n_parallel > 1) {
+        LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
+    }
+
+    // main loop
+
+    // we will store the parallel decoded sequences in this vector
+    std::vector<std::string> streams(n_parallel);
+
+    // remember the batch index of the last token for each parallel sequence
+    // we need this to determine which logits to sample from
+    std::vector<int32_t> i_batch(n_parallel, batch.n_tokens - 1);
+
+    int n_cur    = batch.n_tokens;
+    int n_decode = 0;
+
+    const auto 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 (int32_t i = 0; i < n_parallel; ++i) {
+            if (i_batch[i] < 0) {
+                // the stream has already finished
+                continue;
+            }
+
+            auto   n_vocab = llama_n_vocab(model);
+            auto * logits  = llama_get_logits_ith(ctx, i_batch[i]);
+
+            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 };
+
+            const int   top_k = 40;
+            const float top_p = 0.9f;
+            const float temp  = 0.4f;
+
+            llama_sample_top_k(ctx, &candidates_p, top_k, 1);
+            llama_sample_top_p(ctx, &candidates_p, top_p, 1);
+            llama_sample_temp (ctx, &candidates_p, temp);
+
+            const llama_token new_token_id = llama_sample_token(ctx, &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(ctx) || n_cur == n_len) {
+                i_batch[i] = -1;
+                LOG_TEE("\n");
+                if (n_parallel > 1) {
+                    LOG_TEE("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
+                }
+
+                continue;
+            }
+
+            // if there is only one stream, we print immediately to stdout
+            if (n_parallel == 1) {
+                LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+                fflush(stdout);
+            }
+
+            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;
+
+            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(ctx, batch)) {
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
+            return 1;
+        }
+    }
+
+    LOG_TEE("\n");
+
+    if (n_parallel > 1) {
+        LOG_TEE("\n");
+
+        for (int32_t i = 0; i < n_parallel; ++i) {
+            LOG_TEE("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
+        }
+    }
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+            __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
+
+    llama_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/beam-search/beam-search.cpp

@@ -158,8 +158,9 @@ int main(int argc, char ** argv)
     }
     std::cout << std::flush;
 
-    int n_past = llama_get_kv_cache_token_count(ctx);
-    if (llama_eval(ctx, tokens_list.data(), tokens_list.size(), n_past, params.n_threads))
+    int n_past = 0;
+
+    if (llama_decode(ctx, llama_batch_get_one(tokens_list.data(), tokens_list.size(), n_past, 0)))
     {
         fprintf(stderr, "%s : failed to eval prompt.\n" , __func__ );
         return 1;
@@ -169,7 +170,7 @@ int main(int argc, char ** argv)
     beam_search_callback_data callback_data{ctx, {}};
     size_t const beam_width = static_cast<size_t>(params.n_beams);
     int const n_predict = 256;
-    llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict, params.n_threads);
+    llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict);
 
     std::cout << "\n\n";
     for (llama_token const token_id : callback_data.response) {

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

@@ -48,8 +48,7 @@ struct MyModel* create_mymodel(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
     struct MyModel * ret = new MyModel();
     ret->ctx = ctx;
@@ -71,7 +70,7 @@ 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(ctx);
+    int n_emb = llama_n_embd(llama_get_model(ctx));
     int n_past = mymodel->n_past;
     int n_batch = N; // params.n_batch;
 
@@ -80,7 +79,8 @@ bool eval_float(void * model, float * input, int N){
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        if (llama_eval_embd(ctx, (input+i*n_emb), n_eval, n_past, params.n_threads)) {
+        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;
         }
@@ -101,7 +101,7 @@ bool eval_tokens(void * model, std::vector<llama_token> tokens) {
         if (n_eval > params.n_batch) {
             n_eval = params.n_batch;
         }
-        if (llama_eval(ctx, &tokens[i], n_eval, n_past, params.n_threads)) {
+        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;
         }
@@ -132,7 +132,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
 
     // out of user input, sample next token
     const float   temp            = params.temp;
-    const int32_t top_k           = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k;
+    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;
@@ -148,7 +148,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
     llama_token id = 0;
     {
         auto logits  = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(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++) {
@@ -183,11 +183,11 @@ llama_token sampling_id(struct MyModel* mymodel) {
             if (mirostat == 1) {
                 static float mirostat_mu = 2.0f * mirostat_tau;
                 const int mirostat_m = 100;
-                llama_sample_temperature(ctx, &candidates_p, temp);
+                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_temperature(ctx, &candidates_p, temp);
+                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
@@ -195,7 +195,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
                 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_temperature(ctx, &candidates_p, temp);
+                llama_sample_temp(ctx, &candidates_p, temp);
                 id = llama_sample_token(ctx, &candidates_p);
             }
         }

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

@@ -8,7 +8,7 @@ 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(mymodel->ctx);
+    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];

examples/embedding/embedding.cpp

@@ -42,17 +42,18 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    const int n_ctx_train = llama_n_ctx_train(ctx);
-    if (params.n_ctx > n_ctx_train) {
+    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx = llama_n_ctx(ctx);
+
+    if (n_ctx > n_ctx_train) {
         fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
-                __func__, n_ctx_train, params.n_ctx);
+                __func__, n_ctx_train, n_ctx);
     }
 
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
     int n_past = 0;
@@ -70,15 +71,15 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "\n");
     }
 
-    if (embd_inp.size() > (size_t)params.n_ctx) {
+    if (embd_inp.size() > (size_t)n_ctx) {
         fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
-                __func__, embd_inp.size(), params.n_ctx);
+                __func__, embd_inp.size(), n_ctx);
         return 1;
     }
 
     while (!embd_inp.empty()) {
         int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
-        if (llama_eval(ctx, embd_inp.data(), n_tokens, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(embd_inp.data(), n_tokens, n_past, 0))) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return 1;
         }
@@ -86,8 +87,8 @@ int main(int argc, char ** argv) {
         embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
     }
 
-    const int n_embd = llama_n_embd(ctx);
-    const auto embeddings = llama_get_embeddings(ctx);
+    const int n_embd = llama_n_embd(model);
+    const auto * embeddings = llama_get_embeddings(ctx);
 
     for (int i = 0; i < n_embd; i++) {
         printf("%f ", embeddings[i]);

examples/export-lora/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET export-lora)
+add_executable(${TARGET} export-lora.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/export-lora/README.md

@@ -0,0 +1,26 @@
+# export-lora
+
+Apply LORA adapters to base model and export the resulting model.
+
+```
+usage: export-lora [options]
+
+options:
+  -h, --help                         show this help message and exit
+  -m FNAME, --model-base FNAME       model path from which to load base model (default '')
+  -o FNAME, --model-out FNAME        path to save exported model (default '')
+  -l FNAME, --lora FNAME             apply LoRA adapter
+  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S
+  -t N, --threads N                  number of threads to use during computation (default: 4)
+```
+
+For example:
+
+```bash
+./bin/export-lora \
+    -m open-llama-3b-v2-q8_0.gguf \
+    -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
+    -l lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.bin
+```
+
+Multiple LORA adapters can be applied by passing multiple `-l FN` or `-s FN S` command line parameters.

examples/export-lora/export-lora.cpp

@@ -0,0 +1,474 @@
+
+#include "common.h"
+#include "ggml.h"
+#include "ggml-alloc.h"
+
+#include <vector>
+#include <string>
+#include <thread>
+
+static const size_t tensor_alignment = 32;
+
+struct lora_info {
+    std::string filename;
+    float scale;
+};
+
+struct export_lora_params {
+    std::string fn_model_base;
+    std::string fn_model_out;
+    std::vector<struct lora_info> lora;
+    int n_threads;
+};
+
+struct lora_data {
+    struct lora_info     info;
+    std::vector<uint8_t> data;
+    struct ggml_context * ctx;
+
+    uint32_t lora_r;
+    uint32_t lora_alpha;
+};
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            size = 0;
+        } else {
+            seek(0, SEEK_END);
+            size = tell();
+            seek(0, SEEK_SET);
+        }
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        GGML_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            die_fmt("read error: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            die("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            die_fmt("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    bool eof() {
+        return tell() >= size;
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+static struct export_lora_params get_default_export_lora_params() {
+    struct export_lora_params result;
+    result.fn_model_base = "";
+    result.fn_model_out  = "";
+    result.n_threads = GGML_DEFAULT_N_THREADS;
+    return result;
+}
+
+static void export_lora_print_usage(int /*argc*/, char ** argv, const struct export_lora_params * params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help                         show this help message and exit\n");
+    fprintf(stderr, "  -m FNAME, --model-base FNAME       model path from which to load base model (default '%s')\n", params->fn_model_base.c_str());
+    fprintf(stderr, "  -o FNAME, --model-out FNAME        path to save exported model (default '%s')\n", params->fn_model_out.c_str());
+    fprintf(stderr, "  -l FNAME, --lora FNAME             apply LoRA adapter\n");
+    fprintf(stderr, "  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S\n");
+    fprintf(stderr, "  -t N, --threads N                  number of threads to use during computation (default: %d)\n", params->n_threads);
+}
+
+static bool export_lora_params_parse(int argc, char ** argv, struct export_lora_params * params) {
+    bool invalid_param = false;
+    std::string arg;
+    struct export_lora_params default_params = get_default_export_lora_params();
+    const std::string arg_prefix = "--";
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (arg == "-m" || arg == "--model-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_base = argv[i];
+        } else if (arg == "-o" || arg == "--model-out") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_out = argv[i];
+        } else if (arg == "-l" || arg == "--lora") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            struct lora_info lora;
+            lora.filename = argv[i];
+            lora.scale = 1.0f;
+            params->lora.push_back(lora);
+        } else if (arg == "-s" || arg == "--lora-scaled") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            struct lora_info lora;
+            lora.filename = argv[i];
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            lora.scale = std::stof(argv[i]);
+            params->lora.push_back(lora);
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_threads = std::stoi(argv[i]);
+            if (params->n_threads <= 0) {
+                params->n_threads = std::thread::hardware_concurrency();
+            }
+        } else {
+            fprintf(stderr, "error: unknown argument: '%s'\n", arg.c_str());
+            export_lora_print_usage(argc, argv, &default_params);
+            exit(1);
+        }
+    }
+
+    if (params->fn_model_base == default_params.fn_model_base) {
+        fprintf(stderr, "error: please specify a filename for model-base.\n");
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (params->fn_model_out == default_params.fn_model_out) {
+        fprintf(stderr, "error: please specify a filename for model-out.\n");
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: '%s'\n", arg.c_str());
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    return true;
+}
+
+static void free_lora(struct lora_data * lora) {
+    if (lora->ctx != NULL) {
+        ggml_free(lora->ctx);
+    }
+    delete lora;
+}
+
+static struct lora_data * load_lora(struct lora_info * info) {
+    struct lora_data * result = new struct lora_data;
+    result->info = *info;
+    result->ctx = NULL;
+    result->lora_r     = 1;
+    result->lora_alpha = 1;
+
+    struct llama_file file(info->filename.c_str(), "rb");
+    if (file.fp == NULL) {
+        fprintf(stderr, "warning: Could not open lora adapter '%s'. Ignoring this adapter.\n",
+            info->filename.c_str());
+        free_lora(result);
+        return NULL;
+    }
+
+    struct ggml_init_params params_ggml;
+    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_MAX_NODES;
+    params_ggml.mem_buffer = NULL;
+    params_ggml.no_alloc   = true;
+    result->ctx = ggml_init(params_ggml);
+
+    uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
+    uint32_t magic   = file.read_u32();
+    if (magic != LLAMA_FILE_MAGIC_LORA) {
+        die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
+    }
+    uint32_t version = file.read_u32();
+    if (version != 1) {
+        die_fmt("unexpected lora file version '%u' in '%s'", (unsigned) version, info->filename.c_str());
+    }
+    result->lora_r     = file.read_u32();
+    result->lora_alpha = file.read_u32();
+    // read tensor infos from file
+    std::vector<char> name_buf;
+    std::vector<struct ggml_tensor *> tensors;
+    std::vector<size_t> tensors_offset;
+    size_t total_nbytes_pad = 0;
+    while(!file.eof()) {
+        int64_t ne[4]   = {1,1,1,1};
+        uint32_t n_dims  = file.read_u32();
+        uint32_t namelen = file.read_u32();
+        uint32_t type    = file.read_u32();
+        for (uint32_t k = 0; k < n_dims; ++k) {
+            ne[k] = (int64_t)file.read_u32();
+        }
+        name_buf.clear();
+        name_buf.resize(namelen + 1, '\0');
+        file.read_raw(name_buf.data(), namelen);
+        file.seek((0-file.tell()) & 31, SEEK_CUR);
+        size_t offset = file.tell();
+        struct ggml_tensor * tensor = ggml_new_tensor(result->ctx, (enum ggml_type) type, n_dims, ne);
+        ggml_set_name(tensor, name_buf.data());
+        size_t nbytes     = ggml_nbytes(tensor);
+        size_t nbytes_pad = ggml_nbytes_pad(tensor);
+        total_nbytes_pad += nbytes_pad;
+        tensors.push_back(tensor);
+        tensors_offset.push_back(offset);
+        file.seek(nbytes, SEEK_CUR);
+    }
+    // read tensor data
+    result->data.resize(total_nbytes_pad);
+    size_t data_offset = 0;
+    for (size_t i = 0; i < tensors.size(); ++i) {
+        struct ggml_tensor * tensor = tensors[i];
+        size_t offset     = tensors_offset[i];
+        size_t nbytes     = ggml_nbytes(tensor);
+        size_t nbytes_pad = ggml_nbytes_pad(tensor);
+        file.seek(offset, SEEK_SET);
+        tensor->data = result->data.data() + data_offset;
+        file.read_raw(tensor->data, nbytes);
+        data_offset += nbytes_pad;
+    }
+    return result;
+}
+
+
+static struct ggml_cgraph * build_graph_lora(
+    struct ggml_context * ctx,
+    struct ggml_tensor * tensor,
+    struct ggml_tensor * lora_a,
+    struct ggml_tensor * lora_b,
+    float scaling
+) {
+    struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
+    if (scaling != 1.0f) {
+        ab = ggml_scale(ctx, ab, ggml_new_f32(ctx, scaling));
+    }
+    struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
+
+    struct ggml_cgraph * gf = ggml_new_graph(ctx);
+    ggml_build_forward_expand (gf, res);
+    return gf;
+}
+
+static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int n_threads) {
+    if (lora->ctx == NULL) {
+        return false;
+    }
+    std::string name = ggml_get_name(tensor);
+    std::string name_a = name + std::string(".loraA");
+    std::string name_b = name + std::string(".loraB");
+    struct ggml_tensor * lora_a = ggml_get_tensor(lora->ctx, name_a.c_str());
+    struct ggml_tensor * lora_b = ggml_get_tensor(lora->ctx, name_b.c_str());
+    if (lora_a == NULL || lora_b == NULL) {
+        return false;
+    }
+
+    float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r;
+
+    struct ggml_init_params params;
+    params.mem_size   = GGML_OBJECT_SIZE + GGML_GRAPH_SIZE + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
+    params.mem_buffer = NULL;
+    params.no_alloc   = true;
+    struct ggml_context * ctx = NULL;
+    struct ggml_allocr * alloc = NULL;
+    struct ggml_cgraph * gf = NULL;
+
+    ctx   = ggml_init(params);
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
+    size_t alloc_size = ggml_allocr_alloc_graph(alloc, gf);
+    ggml_allocr_free(alloc);
+    ggml_free(ctx);
+
+    static std::vector<uint8_t> data_compute;
+    data_compute.resize(alloc_size + tensor_alignment);
+
+    ctx   = ggml_init(params);
+    alloc = ggml_allocr_new(data_compute.data(), data_compute.size(), tensor_alignment);
+    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
+    ggml_allocr_alloc_graph(alloc, gf);
+    ggml_allocr_free(alloc);
+
+    struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads);
+    static std::vector<uint8_t> data_work;
+    data_work.resize(cplan.work_size);
+    cplan.work_data = data_work.data();
+
+    ggml_graph_compute(gf, &cplan);
+
+    ggml_free(ctx);
+    return true;
+}
+
+static void export_lora(struct export_lora_params * params) {
+    // load all loras
+    std::vector<struct lora_data *> loras;
+    for (size_t i = 0; i < params->lora.size(); ++i) {
+        struct lora_data * lora = load_lora(&params->lora[i]);
+        if (lora != NULL) {
+            loras.push_back(lora);
+        }
+    }
+    if (loras.size() == 0) {
+        fprintf(stderr, "warning: no lora adapters will be applied.\n");
+    }
+
+    // open input file
+    struct llama_file fin(params->fn_model_base.c_str(), "rb");
+    if (!fin.fp) {
+        die_fmt("Could not open file '%s'\n", params->fn_model_base.c_str());
+    }
+
+    // open base model gguf, read tensors without their data
+    struct ggml_context * ctx_in;
+    struct gguf_init_params params_gguf;
+    params_gguf.no_alloc = true;
+    params_gguf.ctx      = &ctx_in;
+    struct gguf_context * gguf_in = gguf_init_from_file(params->fn_model_base.c_str(), params_gguf);
+
+    // create new gguf
+    struct gguf_context * gguf_out = gguf_init_empty();
+
+    // copy meta data from base model: kv and tensors
+    gguf_set_kv(gguf_out, gguf_in);
+    int n_tensors = gguf_get_n_tensors(gguf_in);
+    for (int i=0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(gguf_in, i);
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
+        gguf_add_tensor(gguf_out, tensor);
+    }
+
+    // create output file
+    struct llama_file fout(params->fn_model_out.c_str(), "wb");
+    if (!fout.fp) {
+        die_fmt("Could not create file '%s'\n", params->fn_model_out.c_str());
+    }
+
+    // write gguf meta data
+    std::vector<uint8_t> meta;
+    meta.resize(gguf_get_meta_size(gguf_out));
+    gguf_get_meta_data(gguf_out, meta.data());
+    fout.write_raw(meta.data(), meta.size());
+
+    std::vector<uint8_t> data;
+    std::vector<uint8_t> padding;
+    for (int i=0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(gguf_in, i);
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
+
+        // read tensor data
+        data.resize(ggml_nbytes(tensor));
+        tensor->data = data.data();
+        size_t offset = gguf_get_tensor_offset(gguf_in, i);
+        fin.seek(offset + meta.size(), SEEK_SET);
+        fin.read_raw(data.data(), data.size());
+
+        // apply all loras
+        for (size_t k = 0; k < loras.size(); ++k) {
+            apply_lora(tensor, loras[k], params->n_threads);
+        }
+
+        // write tensor data + padding
+        padding.clear();
+        padding.resize(GGML_PAD(data.size(), gguf_get_alignment(gguf_out)) - data.size(), 0);
+
+        GGML_ASSERT(fout.tell() == offset + meta.size());
+        // fout.seek(offset + meta.size(), SEEK_SET);
+        fout.write_raw(data.data(), data.size());
+        fout.write_raw(padding.data(), padding.size());
+
+        if (i % 2 == 0) {
+            printf(".");
+        }
+    }
+    printf("\n");
+
+    // close gguf
+    gguf_free(gguf_out);
+    gguf_free(gguf_in);
+
+    // free loras
+    for (size_t i = 0; i < loras.size(); ++i) {
+        free_lora(loras[i]);
+    }
+}
+
+int main(int argc, char ** argv) {
+    struct export_lora_params params = get_default_export_lora_params();
+
+    if (!export_lora_params_parse(argc, argv, &params)) {
+        return 1;
+    }
+
+    export_lora(&params);
+
+    return 0;
+}

examples/finetune/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET finetune)
+add_executable(${TARGET} finetune.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/finetune/README.md

@@ -0,0 +1,90 @@
+# finetune
+
+Basic usage instructions:
+
+```bash
+# get training data
+wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt
+
+# finetune LORA adapter
+./bin/finetune \
+        --model-base open-llama-3b-v2-q8_0.gguf \
+        --checkpoint-in  chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf \
+        --checkpoint-out chk-lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.gguf \
+        --lora-out lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.bin \
+        --train-data "shakespeare.txt" \
+        --save-every 10 \
+        --threads 6 --adam-iter 30 --batch 4 --ctx 64 \
+        --use-checkpointing
+
+# predict
+./bin/main -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+```
+
+Finetune output files will be saved every N iterations (config with `--save-every N`).
+The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output.
+So in above example after 10 iterations these files will be written:
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf
+- lora-open-llama-3b-v2-q8_0-shakespeare-10.bin
+- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+
+After 10 more iterations:
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-20.gguf
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf
+- lora-open-llama-3b-v2-q8_0-shakespeare-20.bin
+- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+
+Checkpoint files (`--checkpoint-in FN`, `--checkpoint-out FN`) store the training process. When the input checkpoint file does not exist, it will begin finetuning a new randomly initialized adapter.
+
+llama.cpp compatible LORA adapters will be saved with filename specified by `--lora-out FN`.
+These LORA adapters can then be used by `main` together with the base model, like in the 'predict' example command above.
+
+In `main` you can also load multiple LORA adapters, which will then be mixed together.
+
+For example if you have two LORA adapters `lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin` and `lora-open-llama-3b-v2-q8_0-bible-LATEST.bin`, you can mix them together like this:
+
+```bash
+./bin/main -m open-llama-3b-v2-q8_0.gguf \
+  --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin \
+  --lora lora-open-llama-3b-v2-q8_0-bible-LATEST.bin
+```
+
+You can change how strong each LORA adapter is applied to the base model by using `--lora-scaled FN SCALE` instead of `--lora FN`.
+
+For example to apply 40% of the 'shakespeare' LORA adapter, 80% of the 'bible' LORA adapter and 100% of yet another one:
+
+```bash
+./bin/main -m open-llama-3b-v2-q8_0.gguf \
+  --lora-scaled lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin 0.4 \
+  --lora-scaled lora-open-llama-3b-v2-q8_0-bible-LATEST.bin 0.8 \
+  --lora lora-open-llama-3b-v2-q8_0-yet-another-one-LATEST.bin
+```
+
+The scale numbers don't need to add up to one, and you can also use numbers creater than 1 to further increase the influence of an adapter. But making the values to big will sometimes result in worse output. Play around to find good values.
+
+Gradient checkpointing reduces the memory requirements by ~50% but increases the runtime.
+If you have enough RAM, you can make finetuning a bit faster by disabling checkpointing with `--no-checkpointing`.
+
+The default LORA rank can be specified with `--lora-r N`.
+The LORA rank can be configured for each model tensor type separately with these command line options:
+
+```bash
+  --lora-r N                 LORA r: default rank. Also specifies resulting scaling together with lora-alpha. (default 4)
+  --rank-att-norm N          LORA rank for attention norm tensor (default 1)
+  --rank-ffn-norm N          LORA rank for feed-forward norm tensor (default 1)
+  --rank-out-norm N          LORA rank for output norm tensor (default 1)
+  --rank-tok-embd N          LORA rank for token embeddings tensor (default 4)
+  --rank-out N               LORA rank for output tensor (default 4)
+  --rank-wq N                LORA rank for wq tensor (default 4)
+  --rank-wk N                LORA rank for wk tensor (default 4)
+  --rank-wv N                LORA rank for wv tensor (default 4)
+  --rank-wo N                LORA rank for wo tensor (default 4)
+  --rank-w1 N                LORA rank for w1 tensor (default 4)
+  --rank-w2 N                LORA rank for w2 tensor (default 4)
+  --rank-w3 N                LORA rank for w3 tensor (default 4)
+```
+
+The LORA rank of 'norm' tensors should always be 1.
+
+To see all available options use `finetune --help`.

examples/finetune/convert-finetune-checkpoint-to-gguf.py

@@ -0,0 +1,489 @@
+#!/usr/bin/env python3
+# finetune checkpoint --> gguf conversion
+
+import argparse
+import gguf
+import os
+import struct
+import sys
+import numpy as np
+from pathlib import Path
+
+# gguf constants
+LLM_KV_OPTIMIZER_TYPE = "optimizer.type"
+LLM_KV_OPTIMIZER_TYPE_ADAM  = "adam"
+LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs"
+LLM_KV_OPTIMIZER_FILE_VERSION               = "optimizer.file_version"
+LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT     = "optimizer.convergence_past_count"
+LLM_KV_OPTIMIZER_PARAMETER_COUNT            = "optimizer.parameter_count"
+LLM_KV_OPTIMIZER_ITERATION_COUNT            = "optimizer.iteration_count"
+LLM_KV_OPTIMIZER_JUST_INITIALIZED           = "optimizer.just_initialized"
+LLM_KV_OPTIMIZER_ADAM_BEST_LOSS             = "optimizer.adam.best_loss"
+LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS         = "optimizer.adam.previous_loss"
+LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT  = "optimizer.adam.no_improvement_count"
+LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count"
+LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS            = "optimizer.lbfgs.best_loss"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP     = "optimizer.lbfgs.line_search_step"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J        = "optimizer.lbfgs.line_search_j"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K        = "optimizer.lbfgs.line_search_k"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END      = "optimizer.lbfgs.line_search_end"
+LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count"
+
+LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS    = "optimizer.adam.first_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS   = "optimizer.adam.second_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values"
+
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS  = "optimizer.lbfgs.current_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS   = "optimizer.lbfgs.current_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS  = "optimizer.lbfgs.previous_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION    = "optimizer.lbfgs.search_direction"
+LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES    = "optimizer.lbfgs.past_loss_values"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA        = "optimizer.lbfgs.memory_alpha"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS           = "optimizer.lbfgs.memory_ys"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S            = "optimizer.lbfgs.memory_s"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y            = "optimizer.lbfgs.memory_y"
+
+LLM_KV_TRAINING_TYPE_TRAIN_MODEL   = "train_model"
+LLM_KV_TRAINING_TYPE_FINETUNE_LORA = "finetune_lora"
+LLM_KV_TRAINING_TYPE               = "training.type"
+LLM_KV_TRAINING_FILE_VERSION       = "training.file_version"
+LLM_KV_TRAINING_ITERATION_COUNT    = "training.iteration_count"
+LLM_KV_TRAINING_SAMPLE_COUNT       = "training.sample_count"
+LLM_KV_TRAINING_TOKEN_COUNT        = "training.token_count"
+
+LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD  = "training.lora.rank.token_embd"
+LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM = "training.lora.rank.output_norm"
+LLM_KV_TRAINING_LORA_RANK_OUTPUT      = "training.lora.rank.output"
+LLM_KV_TRAINING_LORA_RANK_ATTN_NORM   = "training.lora.rank.attn_norm"
+LLM_KV_TRAINING_LORA_RANK_ATTN_Q      = "training.lora.rank.attn_q"
+LLM_KV_TRAINING_LORA_RANK_ATTN_K      = "training.lora.rank.attn_k"
+LLM_KV_TRAINING_LORA_RANK_ATTN_V      = "training.lora.rank.attn_v"
+LLM_KV_TRAINING_LORA_RANK_ATTN_OUT    = "training.lora.rank.attn_output"
+LLM_KV_TRAINING_LORA_RANK_FFN_NORM    = "training.lora.rank.ffn_norm"
+LLM_KV_TRAINING_LORA_RANK_FFN_GATE    = "training.lora.rank.ffn_gate"
+LLM_KV_TRAINING_LORA_RANK_FFN_DOWN    = "training.lora.rank.ffn_down"
+LLM_KV_TRAINING_LORA_RANK_FFN_UP      = "training.lora.rank.ffn_up"
+
+class Tensor:
+    def __init__(self, dtype='f', ne=None):
+        if ne is None:
+            ne = []
+        self.dtype = dtype
+        self.ne = ne
+        self.nbytes = 0
+        if self.dtype == 'f':
+            if len(self.ne) == 0:
+                self.nbytes = 0
+            else:
+                self.nbytes = int(np.product(self.ne)) * 4
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+    def load(self, data, offset):
+        nd = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        namelen = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        dtype = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        assert(nd == len(self.ne))
+        ne = []
+        for d in range(nd):
+            n = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+            ne.append(n)
+
+        if tuple(ne) != tuple(self.ne):
+            raise ValueError(f"Tensor.load: Expected number of elements {str(self.ne)} does not match what is read from file {str(ne)}")
+
+        if self.dtype == 'f':
+            assert(dtype == 0)
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+        self.name = bytes(data[offset:offset+namelen]); offset += namelen
+        # 32-byte alignment
+        offset += (0 - offset) & 31
+        self.data = data[offset:offset+self.nbytes]
+        offset += self.nbytes
+        return offset
+
+    def max_storage_size(self):
+        result = 0
+        result += 4 # nd
+        result += 4 # namelen
+        result += 4 # dtype
+        result += len(self.ne)*8 # ne
+        result += 48 # name (maximum as of commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9)
+        result += 31 # 32-byte alignment
+        result += self.nbytes
+        return result
+
+    def save_gguf(self, gguf_writer, name):
+        gguf_writer.add_tensor(
+            name=name,
+            tensor=self.data,
+            raw_shape=np.array(list(reversed(self.ne))),
+            raw_dtype=gguf.GGMLQuantizationType.F32)
+
+class OptimizationContext:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]
+        offset += 4
+
+        if self.version != 1:
+            raise ValueError('Invalid version of optimization context in checkpoint file')
+
+        self.past    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_m = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.nx      = struct.unpack('N',  bytes(data[offset:offset + 8]))[0];  offset += 8
+        self.iter    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.just_initialized = bool(struct.unpack('<i', bytes(data[offset:offset + 4]))[0]);  offset += 4
+
+        self.adam_m  = Tensor('f', [self.nx])
+        self.adam_v  = Tensor('f', [self.nx])
+        self.adam_pf = Tensor('f', [self.past] if self.past > 0 else [])
+
+        self.lbfgs_x    = Tensor('f', [self.nx])
+        self.lbfgs_xp   = Tensor('f', [self.nx])
+        self.lbfgs_g    = Tensor('f', [self.nx])
+        self.lbfgs_gp   = Tensor('f', [self.nx])
+        self.lbfgs_d    = Tensor('f', [self.nx])
+        self.lbfgs_pf   = Tensor('f', [self.past] if self.past > 0 else [])
+        self.lbfgs_lmal = Tensor('f', [self.lbfgs_m])
+        self.lbfgs_lmys = Tensor('f', [self.lbfgs_m])
+        self.lbfgs_lms  = Tensor('f', [self.nx, self.lbfgs_m])
+        self.lbfgs_lmy  = Tensor('f', [self.nx, self.lbfgs_m])
+
+        # forgot to save type in version 1:
+        # guess self.type from number of remaining bytes
+        size_type_0 = 12 + sum([t.max_storage_size() for t in
+                                [self.adam_m, self.adam_v]
+                                +([self.adam_pf] if (self.past > 0) else [])])
+        size_type_1 = 24 + sum([t.max_storage_size() for t in
+                                [self.lbfgs_x, self.lbfgs_xp, self.lbfgs_g,
+                                 self.lbfgs_gp, self.lbfgs_d, self.lbfgs_pf,
+                                 self.lbfgs_lmal, self.lbfgs_lmys,
+                                 self.lbfgs_lms, self.lbfgs_lmy]
+                                 +([self.lbfgs_pf] if (self.past > 0) else [])])
+        # due to alignment padding the size might not by exact
+        # but the difference in size for both types is significant,
+        # so we can just use whichever is closest
+        remaining = len(data) - offset
+        if abs(remaining - size_type_0) < abs(remaining - size_type_1):
+            self.type = 0
+        else:
+            self.type = 1
+
+        if self.type == 0:
+            offset = self.adam_m.load(data, offset)
+            offset = self.adam_v.load(data, offset)
+            offset = self.adam_pf.load(data,offset)
+
+            self.adam_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.adam_fx_prev          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.adam_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+        elif self.type == 1:
+            offset = self.lbfgs_x.load(data, offset)
+            offset = self.lbfgs_xp.load(data, offset)
+            offset = self.lbfgs_g.load(data, offset)
+            offset = self.lbfgs_gp.load(data, offset)
+            offset = self.lbfgs_d.load(data, offset)
+            offset = self.lbfgs_pf.load(data, offset)
+            offset = self.lbfgs_lmal.load(data, offset)
+            offset = self.lbfgs_lmys.load(data, offset)
+            offset = self.lbfgs_lms.load(data, offset)
+            offset = self.lbfgs_lmy.load(data, offset)
+
+            self.lbfgs_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_step             = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_j                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_k                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_end              = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+        else:
+            raise ValueError(f"Invalid optimizer type '{self.type}'")
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_FILE_VERSION, 0)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT, self.past)
+        gguf_writer.add_uint64(LLM_KV_OPTIMIZER_PARAMETER_COUNT, self.nx)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ITERATION_COUNT, self.iter)
+        gguf_writer.add_bool(LLM_KV_OPTIMIZER_JUST_INITIALIZED, self.just_initialized)
+
+        if self.type == 0:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_BEST_LOSS, self.adam_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS, self.adam_fx_prev)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT, self.adam_n_no_improvement)
+
+            self.adam_m.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS)
+            self.adam_v.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS)
+            if self.past > 0:
+                self.adam_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES)
+
+        elif self.type == 1:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, self.lbfgs_m)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, self.lbfgs_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, self.lbfgs_step)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, self.lbfgs_j)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, self.lbfgs_k)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, self.lbfgs_end)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, self.lbfgs_n_no_improvement)
+
+            self.lbfgs_x.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS)
+            self.lbfgs_xp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS)
+            self.lbfgs_g.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS)
+            self.lbfgs_gp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS)
+            self.lbfgs_d.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION)
+            if self.past > 0:
+                self.lbfgs_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES)
+            self.lbfgs_lmal.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA)
+            self.lbfgs_lmys.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS)
+            self.lbfgs_lms.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S)
+            self.lbfgs_lmy.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y)
+        else:
+            raise ValueError('Unknown optimizer type')
+
+class LoraParams:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.n_rank_attention_norm  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wq              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wk              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wv              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wo              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_ffn_norm        = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w1              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w2              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w3              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_tok_embeddings  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_norm            = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_output          = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD,  self.n_rank_tok_embeddings)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM, self.n_rank_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_OUTPUT,      self.n_rank_output)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_NORM,   self.n_rank_attention_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_Q,      self.n_rank_wq)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_K,      self.n_rank_wk)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_V,      self.n_rank_wv)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,    self.n_rank_wo)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_NORM,    self.n_rank_ffn_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_GATE,    self.n_rank_w1)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,    self.n_rank_w2)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_UP,      self.n_rank_w3)
+
+class ModelParams:
+    def __init__(self, n_ff = None):
+        self.n_ff = n_ff
+
+    def load(self, data, offset):
+        self.n_vocab = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_embd  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_mult  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_head  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_layer = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rot   = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+    def get_n_ff(self):
+        if self.n_ff is None:
+            # struct my_llama_model::get_n_ff in train-text-from-scratch.cpp commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9
+            return ((2*(4*self.n_embd)//3 + self.n_mult - 1)//self.n_mult)*self.n_mult
+        else:
+            return self.n_ff
+
+    def save_gguf(self, gguf_writer):
+        # self.n_vocab not saved
+        gguf_writer.add_embedding_length(self.n_embd)
+        gguf_writer.add_head_count(self.n_head)
+        gguf_writer.add_block_count(self.n_layer)
+        gguf_writer.add_rope_dimension_count(self.n_rot)
+        gguf_writer.add_feed_forward_length(self.get_n_ff())
+
+def tensor_name(key, bid=None, suffix=".weight"):
+    return gguf.MODEL_TENSOR_NAMES[gguf.MODEL_ARCH.LLAMA][key].format(bid=bid) + suffix
+
+class Layer:
+    def __init__(self, params, lora_params, bid):
+        self.bid = bid
+        self.att_norm_a = Tensor('f', [lora_params.n_rank_attention_norm, params.n_embd])
+        self.att_norm_b = Tensor('f', [lora_params.n_rank_attention_norm, 1])
+        self.wq_a       = Tensor('f', [lora_params.n_rank_wq, params.n_embd])
+        self.wq_b       = Tensor('f', [lora_params.n_rank_wq, params.n_embd])
+        self.wk_a       = Tensor('f', [lora_params.n_rank_wk, params.n_embd])
+        self.wk_b       = Tensor('f', [lora_params.n_rank_wk, params.n_embd])
+        self.wv_a       = Tensor('f', [lora_params.n_rank_wv, params.n_embd])
+        self.wv_b       = Tensor('f', [lora_params.n_rank_wv, params.n_embd])
+        self.wo_a       = Tensor('f', [lora_params.n_rank_wo, params.n_embd])
+        self.wo_b       = Tensor('f', [lora_params.n_rank_wo, params.n_embd])
+        self.ffn_norm_a = Tensor('f', [lora_params.n_rank_ffn_norm, params.n_embd])
+        self.ffn_norm_b = Tensor('f', [lora_params.n_rank_ffn_norm, 1])
+        self.w1_a       = Tensor('f', [lora_params.n_rank_w1, params.n_embd])
+        self.w1_b       = Tensor('f', [lora_params.n_rank_w1, params.get_n_ff()])
+        self.w2_a       = Tensor('f', [lora_params.n_rank_w2, params.get_n_ff()])
+        self.w2_b       = Tensor('f', [lora_params.n_rank_w2, params.n_embd])
+        self.w3_a       = Tensor('f', [lora_params.n_rank_w3, params.n_embd])
+        self.w3_b       = Tensor('f', [lora_params.n_rank_w3, params.get_n_ff()])
+
+    def load(self, data, offset):
+        offset = self.att_norm_a.load(data, offset)
+        offset = self.att_norm_b.load(data, offset)
+        offset = self.wq_a.load(data, offset)
+        offset = self.wq_b.load(data, offset)
+        offset = self.wk_a.load(data, offset)
+        offset = self.wk_b.load(data, offset)
+        offset = self.wv_a.load(data, offset)
+        offset = self.wv_b.load(data, offset)
+        offset = self.wo_a.load(data, offset)
+        offset = self.wo_b.load(data, offset)
+        offset = self.ffn_norm_a.load(data, offset)
+        offset = self.ffn_norm_b.load(data, offset)
+        offset = self.w1_a.load(data, offset)
+        offset = self.w1_b.load(data, offset)
+        offset = self.w2_a.load(data, offset)
+        offset = self.w2_b.load(data, offset)
+        offset = self.w3_a.load(data, offset)
+        offset = self.w3_b.load(data, offset)
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.att_norm_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid, ".weight.lora_a"))
+        self.att_norm_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid, ".weight.lora_b"))
+        self.wq_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q,    self.bid, ".weight.lora_a"))
+        self.wq_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q,    self.bid, ".weight.lora_b"))
+        self.wk_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K,    self.bid, ".weight.lora_a"))
+        self.wk_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K,    self.bid, ".weight.lora_b"))
+        self.wv_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V,    self.bid, ".weight.lora_a"))
+        self.wv_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V,    self.bid, ".weight.lora_b"))
+        self.wo_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT,  self.bid, ".weight.lora_a"))
+        self.wo_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT,  self.bid, ".weight.lora_b"))
+        self.ffn_norm_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM,  self.bid, ".weight.lora_a"))
+        self.ffn_norm_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM,  self.bid, ".weight.lora_b"))
+        self.w1_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE,  self.bid, ".weight.lora_a"))
+        self.w1_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE,  self.bid, ".weight.lora_b"))
+        self.w2_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN,  self.bid, ".weight.lora_a"))
+        self.w2_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN,  self.bid, ".weight.lora_b"))
+        self.w3_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP,    self.bid, ".weight.lora_a"))
+        self.w3_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP,    self.bid, ".weight.lora_b"))
+
+class LoraModel:
+    def __init__(self, n_ff = None):
+        self.params = ModelParams(n_ff = n_ff)
+        self.lora_params = LoraParams()
+        self.layers = []
+
+    def load(self, data, offset):
+        offset = self.params.load(data, offset)
+        offset = self.lora_params.load(data, offset)
+
+        self.tok_embd_a = Tensor('f', [self.lora_params.n_rank_tok_embeddings, self.params.n_embd])
+        self.tok_embd_b = Tensor('f', [self.lora_params.n_rank_tok_embeddings, self.params.n_vocab])
+        self.norm_a     = Tensor('f', [self.lora_params.n_rank_norm, self.params.n_embd])
+        self.norm_b     = Tensor('f', [self.lora_params.n_rank_norm, 1])
+        self.output_a   = Tensor('f', [self.lora_params.n_rank_output, self.params.n_embd])
+        self.output_b   = Tensor('f', [self.lora_params.n_rank_output, self.params.n_vocab])
+
+        offset = self.tok_embd_a.load(data, offset)
+        offset = self.tok_embd_b.load(data, offset)
+        offset = self.norm_a.load(data, offset)
+        offset = self.norm_b.load(data, offset)
+        offset = self.output_a.load(data, offset)
+        offset = self.output_b.load(data, offset)
+
+        self.layers.clear()
+        for bid in range(self.params.n_layer):
+            layer = Layer(self.params, self.lora_params, bid)
+            offset = layer.load(data, offset)
+            self.layers.append(layer)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.params.save_gguf(gguf_writer)
+        self.lora_params.save_gguf(gguf_writer)
+
+        self.tok_embd_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD,  suffix=".weight.lora_a"))
+        self.tok_embd_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD,  suffix=".weight.lora_b"))
+        self.norm_a.save_gguf    (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=".weight.lora_a"))
+        self.norm_b.save_gguf    (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=".weight.lora_b"))
+        self.output_a.save_gguf  (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT,      suffix=".weight.lora_a"))
+        self.output_b.save_gguf  (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT,      suffix=".weight.lora_b"))
+
+        for layer in self.layers:
+            layer.save_gguf(gguf_writer)
+
+class LoraCheckpoint:
+    def __init__(self, n_ff = None):
+        self.model = LoraModel(n_ff = n_ff)
+        self.opt_ctx = OptimizationContext()
+
+    def load(self, data, offset):
+        magic   = bytes(reversed(data[offset:offset + 4])); offset += 4
+        if magic != b'ggcl':
+            raise ValueError(f"File header magic indicates, that this is no finetune-lora checkpoint file. Expected 'ggcl', Got '{str(magic)}'")
+
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        if self.version != 0:
+            raise ValueError('Invalid version of checkpoint file')
+
+        self.train_its     = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_samples = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_tokens  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        offset = self.model.load(data, offset)
+        offset = self.opt_ctx.load(data, offset)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_file_type(gguf.GGMLQuantizationType.F32)
+        gguf_writer.add_layer_norm_rms_eps(1e-5)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_FILE_VERSION,    0)
+        gguf_writer.add_string(LLM_KV_TRAINING_TYPE,            LLM_KV_TRAINING_TYPE_FINETUNE_LORA)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_ITERATION_COUNT, self.train_its)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_SAMPLE_COUNT,    self.train_samples)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_TOKEN_COUNT,     self.train_tokens)
+        self.model.save_gguf(gguf_writer)
+        self.opt_ctx.save_gguf(gguf_writer)
+
+def handle_args():
+    parser = argparse.ArgumentParser(description = 'Convert finetune checkpoints to GGUF')
+    parser.add_argument('--input',  '-i', type = Path, help = 'Input finetune checkpoint filename', required=True)
+    parser.add_argument('--output', '-o', type = Path, help = 'Output GGUF filename', required=True)
+    parser.add_argument('--ff', type = int, help = "Feedforward size, if not provided compute from n_mult. Provide this if you get 'ValueError: Tensor.load: Expected number of elements does not match what is read from file'", required=False)
+    return parser.parse_args()
+
+def main():
+    cfg = handle_args()
+    print(cfg)
+    data = np.memmap(cfg.input, mode = 'r')
+    chk = LoraCheckpoint(n_ff = cfg.ff)
+    offset = 0
+    offset = chk.load(data, offset)
+    # we should have read all available data
+    assert(offset == len(data))
+
+    gguf_writer = gguf.GGUFWriter(cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
+    chk.save_gguf(gguf_writer)
+    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()
+
+if __name__ == '__main__':
+    main()

examples/finetune/finetune.cpp

@@ -0,0 +1,1940 @@
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "llama.h"
+#include "common.h"
+#include "train.h"
+#include <unordered_map>
+#include <vector>
+#include <cassert>
+#include <climits>
+#include <cstring>
+#include <cstdarg>
+#include <ctime>
+#include <random>
+#include <stdexcept>
+#include <algorithm>
+#include <string>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+static const size_t tensor_alignment = 32;
+
+struct my_llama_hparams {
+    uint32_t n_vocab    = 32000;
+    uint32_t n_ctx      = 512;
+    uint32_t n_embd     = 4096;
+    uint32_t n_ff       = 11008;
+    uint32_t n_head     = 32;
+    uint32_t n_head_kv  = 32;
+    uint32_t n_layer    = 32;
+
+    // float f_norm_eps     = 1e-5f; // falcon
+    float f_norm_rms_eps = 1e-5f; // llama
+
+    float rope_freq_base  = 10000.0f;
+    float rope_freq_scale = 1.0f;
+
+    uint32_t n_gqa() const {
+        return n_head/n_head_kv;
+    }
+
+    uint32_t n_embd_head() const {
+        return n_embd/n_head;
+    }
+
+    uint32_t n_embd_gqa() const {
+        return n_embd/n_gqa();
+    }
+
+    bool operator!=(const my_llama_hparams& other) const {
+        return memcmp(this, &other, sizeof(other));
+    }
+};
+
+struct my_llama_layer {
+    // normalization
+    struct ggml_tensor * attention_norm;
+
+    // attention
+    struct ggml_tensor * wq;
+    struct ggml_tensor * wk;
+    struct ggml_tensor * wv;
+    struct ggml_tensor * wo;
+
+    // normalization
+    struct ggml_tensor * ffn_norm;
+
+    // ff
+    struct ggml_tensor * w1;
+    struct ggml_tensor * w2;
+    struct ggml_tensor * w3;
+};
+
+struct my_llama_model {
+    struct my_llama_hparams hparams;
+
+    struct ggml_tensor * tok_embeddings;
+
+    struct ggml_tensor * norm;
+    struct ggml_tensor * output;
+
+    std::vector<my_llama_layer> layers;
+};
+
+struct my_llama_lora_hparams {
+    uint32_t lora_r = 1;
+    uint32_t lora_alpha = 1;
+    uint32_t n_rank_attention_norm = 1;
+    uint32_t n_rank_wq = 4;
+    uint32_t n_rank_wk = 4;
+    uint32_t n_rank_wv = 4;
+    uint32_t n_rank_wo = 4;
+    uint32_t n_rank_ffn_norm = 1;
+    uint32_t n_rank_w1 = 4;
+    uint32_t n_rank_w2 = 4;
+    uint32_t n_rank_w3 = 4;
+    uint32_t n_rank_tok_embeddings = 4;
+    uint32_t n_rank_norm = 1;
+    uint32_t n_rank_output = 4;
+
+    bool operator!=(const my_llama_lora_hparams& other) const {
+        return memcmp(this, &other, sizeof(other));
+    }
+};
+
+struct my_llama_lora_layer {
+    // normalization
+    struct ggml_tensor * attention_norm_a;
+    struct ggml_tensor * attention_norm_b;
+
+    // attention
+    struct ggml_tensor * wq_a;
+    struct ggml_tensor * wq_b;
+    struct ggml_tensor * wk_a;
+    struct ggml_tensor * wk_b;
+    struct ggml_tensor * wv_a;
+    struct ggml_tensor * wv_b;
+    struct ggml_tensor * wo_a;
+    struct ggml_tensor * wo_b;
+
+    // normalization
+    struct ggml_tensor * ffn_norm_a;
+    struct ggml_tensor * ffn_norm_b;
+
+    // ff
+    struct ggml_tensor * w1_a;
+    struct ggml_tensor * w1_b;
+    struct ggml_tensor * w2_a;
+    struct ggml_tensor * w2_b;
+    struct ggml_tensor * w3_a;
+    struct ggml_tensor * w3_b;
+};
+
+struct my_llama_lora {
+    struct ggml_context * ctx = NULL;
+    std::vector<uint8_t> data;
+
+    my_llama_lora_hparams hparams;
+
+    struct ggml_tensor * tok_embeddings_a;
+    struct ggml_tensor * tok_embeddings_b;
+
+    struct ggml_tensor * norm_a;
+    struct ggml_tensor * norm_b;
+    struct ggml_tensor * output_a;
+    struct ggml_tensor * output_b;
+
+    std::vector<my_llama_lora_layer> layers;
+};
+
+// gguf constants
+static const char * LLM_KV_TRAINING_TYPE_FINETUNE_LORA   = "finetune_lora";
+static const char * LLM_KV_TRAINING_TYPE                 = "training.type";
+
+static const char * LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD  = "training.lora.rank.token_embd";
+static const char * LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM = "training.lora.rank.output_norm";
+static const char * LLM_KV_TRAINING_LORA_RANK_OUTPUT      = "training.lora.rank.output";
+static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_NORM   = "training.lora.rank.attn_norm";
+static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_Q      = "training.lora.rank.attn_q";
+static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_K      = "training.lora.rank.attn_k";
+static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_V      = "training.lora.rank.attn_v";
+static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_OUT    = "training.lora.rank.attn_output";
+static const char * LLM_KV_TRAINING_LORA_RANK_FFN_NORM    = "training.lora.rank.ffn_norm";
+static const char * LLM_KV_TRAINING_LORA_RANK_FFN_GATE    = "training.lora.rank.ffn_gate";
+static const char * LLM_KV_TRAINING_LORA_RANK_FFN_DOWN    = "training.lora.rank.ffn_down";
+static const char * LLM_KV_TRAINING_LORA_RANK_FFN_UP      = "training.lora.rank.ffn_up";
+
+// gguf constants (sync with gguf.py)
+
+static const char * LLM_KV_GENERAL_ARCHITECTURE        = "general.architecture";
+static const char * LLM_KV_GENERAL_FILE_TYPE           = "general.file_type";
+
+static const char * LLM_KV_CONTEXT_LENGTH              = "%s.context_length";
+static const char * LLM_KV_EMBEDDING_LENGTH            = "%s.embedding_length";
+static const char * LLM_KV_BLOCK_COUNT                 = "%s.block_count";
+static const char * LLM_KV_FEED_FORWARD_LENGTH         = "%s.feed_forward_length";
+static const char * LLM_KV_ATTENTION_HEAD_COUNT        = "%s.attention.head_count";
+static const char * LLM_KV_ATTENTION_HEAD_COUNT_KV     = "%s.attention.head_count_kv";
+static const char * LLM_KV_ATTENTION_LAYERNORM_RMS_EPS = "%s.attention.layer_norm_rms_epsilon";
+static const char * LLM_KV_ROPE_DIMENSION_COUNT        = "%s.rope.dimension_count";
+static const char * LLM_KV_ROPE_FREQ_BASE              = "%s.rope.freq_base"; // TODO load in llama.cpp
+static const char * LLM_KV_ROPE_SCALE_LINEAR           = "%s.rope.scale_linear";
+
+static const char * LLM_TENSOR_TOKEN_EMBD    = "token_embd";
+static const char * LLM_TENSOR_OUTPUT_NORM   = "output_norm";
+static const char * LLM_TENSOR_OUTPUT        = "output";
+static const char * LLM_TENSOR_ATTN_NORM     = "blk.%d.attn_norm";
+static const char * LLM_TENSOR_ATTN_Q        = "blk.%d.attn_q";
+static const char * LLM_TENSOR_ATTN_K        = "blk.%d.attn_k";
+static const char * LLM_TENSOR_ATTN_V        = "blk.%d.attn_v";
+static const char * LLM_TENSOR_ATTN_OUT      = "blk.%d.attn_output";
+static const char * LLM_TENSOR_FFN_NORM      = "blk.%d.ffn_norm";
+static const char * LLM_TENSOR_FFN_GATE      = "blk.%d.ffn_gate";
+static const char * LLM_TENSOR_FFN_DOWN      = "blk.%d.ffn_down";
+static const char * LLM_TENSOR_FFN_UP        = "blk.%d.ffn_up";
+
+static void print_params(struct my_llama_hparams * params) {
+    printf("%s: n_vocab:   %u\n", __func__, params->n_vocab);
+    printf("%s: n_ctx:     %u\n", __func__, params->n_ctx);
+    printf("%s: n_embd:    %u\n", __func__, params->n_embd);
+    printf("%s: n_ff:      %u\n", __func__, params->n_ff);
+    printf("%s: n_head:    %u\n", __func__, params->n_head);
+    printf("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
+    printf("%s: n_layer:   %u\n", __func__, params->n_layer);
+    printf("%s: norm_rms_eps          : %f\n", __func__, params->f_norm_rms_eps);
+    printf("%s: rope_freq_base        : %f\n", __func__, params->rope_freq_base);
+    printf("%s: rope_freq_scale       : %f\n", __func__, params->rope_freq_scale);
+}
+
+static void print_lora_params(struct my_llama_lora_hparams * params) {
+    printf("%s: n_rank_attention_norm : %u\n", __func__, params->n_rank_attention_norm);
+    printf("%s: n_rank_wq             : %u\n", __func__, params->n_rank_wq);
+    printf("%s: n_rank_wk             : %u\n", __func__, params->n_rank_wk);
+    printf("%s: n_rank_wv             : %u\n", __func__, params->n_rank_wv);
+    printf("%s: n_rank_wo             : %u\n", __func__, params->n_rank_wo);
+    printf("%s: n_rank_ffn_norm       : %u\n", __func__, params->n_rank_ffn_norm);
+    printf("%s: n_rank_w1             : %u\n", __func__, params->n_rank_w1);
+    printf("%s: n_rank_w2             : %u\n", __func__, params->n_rank_w2);
+    printf("%s: n_rank_w3             : %u\n", __func__, params->n_rank_w3);
+    printf("%s: n_rank_tok_embeddings : %u\n", __func__, params->n_rank_tok_embeddings);
+    printf("%s: n_rank_norm           : %u\n", __func__, params->n_rank_norm);
+    printf("%s: n_rank_output         : %u\n", __func__, params->n_rank_output);
+}
+
+#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \
+{ \
+    const std::string skey(key); \
+    const int kid = gguf_find_key(ctx, skey.c_str()); \
+    if (kid >= 0) { \
+        enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \
+        if (ktype != (type)) { \
+            die_fmt("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype)); \
+        } \
+        (dst) = func(ctx, kid); \
+    } else if (req) { \
+        die_fmt("key not found in model: %s", skey.c_str()); \
+    } \
+}
+
+static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_hparams * hparams, const char * expected_arch) {
+    std::string arch;
+
+    GGUF_GET_KEY(ctx, arch, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_GENERAL_ARCHITECTURE);
+    if (expected_arch != NULL) {
+        if (arch != expected_arch) {
+            printf("%s: arch=%s expected_arch=%s\n", __func__, arch.c_str(), expected_arch);
+        }
+        GGML_ASSERT(arch == expected_arch);
+    }
+
+    std::vector<char> keybuf;
+    keybuf.resize(512);
+    auto kv = [&arch, &keybuf](const char * key) -> const char * {
+        snprintf(keybuf.data(), keybuf.size(), key, arch.c_str());
+        return keybuf.data();
+    };
+
+    GGUF_GET_KEY(ctx, hparams->n_embd,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_EMBEDDING_LENGTH));
+    GGUF_GET_KEY(ctx, hparams->n_ctx,          gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_CONTEXT_LENGTH));
+    GGUF_GET_KEY(ctx, hparams->n_ff,           gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_FEED_FORWARD_LENGTH));
+    GGUF_GET_KEY(ctx, hparams->n_head,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_ATTENTION_HEAD_COUNT));
+    GGUF_GET_KEY(ctx, hparams->n_layer,        gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_BLOCK_COUNT));
+
+    // n_head_kv is optional, default to n_head
+    hparams->n_head_kv = hparams->n_head;
+    GGUF_GET_KEY(ctx, hparams->n_head_kv,      gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV));
+
+    float rope_freq_scale = 1.0f;
+    GGUF_GET_KEY(ctx, hparams->f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
+    GGUF_GET_KEY(ctx, hparams->rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
+    GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
+    if (rope_freq_scale != 1.0f) {
+        hparams->rope_freq_scale = 1.0f / rope_freq_scale;
+    }
+}
+
+static void init_model(struct llama_model * input, struct my_llama_model * model, const char * fn_model, uint32_t n_ctx) {
+    auto & hparams = model->hparams;
+
+    std::vector<char> tn_buf;
+    tn_buf.resize(GGML_MAX_NAME);
+    auto tn = [&tn_buf](const char * key) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", key);
+        return tn_buf.data();
+    };
+    auto tni = [&tn_buf](const char * key, int bid) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), key, bid);
+        std::string s = tn_buf.data();
+        snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", s.c_str());
+        return tn_buf.data();
+    };
+
+
+    // get parameters directly from gguf file
+    {
+        struct gguf_init_params params = {
+            /*.no_alloc = */ false,
+            /*.ctx      = */ NULL,
+        };
+        struct gguf_context * mctx = gguf_init_from_file(fn_model, params);
+
+        load_model_hparams_gguf(mctx, &hparams, "llama");
+
+        gguf_free(mctx);
+    }
+    hparams.n_vocab = llama_n_vocab(input);
+    hparams.n_ctx = n_ctx;
+
+    // get tensors from llama_model (possibly mmapped)
+    model->tok_embeddings = llama_get_model_tensor(input, tn(LLM_TENSOR_TOKEN_EMBD));
+    model->norm           = llama_get_model_tensor(input, tn(LLM_TENSOR_OUTPUT_NORM));
+    model->output         = llama_get_model_tensor(input, tn(LLM_TENSOR_OUTPUT));
+
+    assert_shape_2d(model->tok_embeddings, hparams.n_embd, hparams.n_vocab);
+    assert_shape_1d(model->norm,           hparams.n_embd);
+    assert_shape_2d(model->output,         hparams.n_embd, hparams.n_vocab);
+
+    model->layers.resize(hparams.n_layer);
+    for (uint32_t i = 0; i < hparams.n_layer; ++i) {
+        auto & layer = model->layers[i];
+
+        layer.attention_norm = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_NORM, i));
+        layer.wq             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_Q, i));
+        layer.wk             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_K, i));
+        layer.wv             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_V, i));
+        layer.wo             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_OUT, i));
+        layer.ffn_norm       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_NORM, i));
+        layer.w1             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
+        layer.w2             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
+        layer.w3             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
+
+        assert_shape_1d(layer.attention_norm, hparams.n_embd);
+        assert_shape_2d(layer.wq,             hparams.n_embd, hparams.n_embd);
+        assert_shape_2d(layer.wk,             hparams.n_embd, hparams.n_embd);
+        assert_shape_2d(layer.wv,             hparams.n_embd, hparams.n_embd);
+        assert_shape_2d(layer.wo,             hparams.n_embd, hparams.n_embd);
+        assert_shape_1d(layer.ffn_norm,       hparams.n_embd);
+        assert_shape_2d(layer.w1,             hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.w2,             hparams.n_ff,   hparams.n_embd);
+        assert_shape_2d(layer.w3,             hparams.n_embd, hparams.n_ff);
+    }
+}
+
+static void set_param_lora(struct my_llama_lora * lora) {
+    const uint32_t n_layer = lora->layers.size();
+
+    struct ggml_context* ctx = lora->ctx;
+
+    ggml_set_param(ctx, lora->tok_embeddings_a);
+    ggml_set_param(ctx, lora->tok_embeddings_b);
+    ggml_set_param(ctx, lora->norm_a);
+    ggml_set_param(ctx, lora->norm_b);
+    ggml_set_param(ctx, lora->output_a);
+    ggml_set_param(ctx, lora->output_b);
+
+    for (uint32_t i = 0; i < n_layer; ++i) {
+        auto & layer = lora->layers[i];
+
+        ggml_set_param(ctx, layer.attention_norm_a);
+        ggml_set_param(ctx, layer.attention_norm_b);
+        ggml_set_param(ctx, layer.wq_a);
+        ggml_set_param(ctx, layer.wq_b);
+        ggml_set_param(ctx, layer.wk_a);
+        ggml_set_param(ctx, layer.wk_b);
+        ggml_set_param(ctx, layer.wv_a);
+        ggml_set_param(ctx, layer.wv_b);
+        ggml_set_param(ctx, layer.wo_a);
+        ggml_set_param(ctx, layer.wo_b);
+        ggml_set_param(ctx, layer.ffn_norm_a);
+        ggml_set_param(ctx, layer.ffn_norm_b);
+        ggml_set_param(ctx, layer.w1_a);
+        ggml_set_param(ctx, layer.w1_b);
+        ggml_set_param(ctx, layer.w2_a);
+        ggml_set_param(ctx, layer.w2_b);
+        ggml_set_param(ctx, layer.w3_a);
+        ggml_set_param(ctx, layer.w3_b);
+    }
+}
+
+static void alloc_lora(struct ggml_allocr * alloc, struct my_llama_lora * lora) {
+    ggml_allocr_alloc(alloc, lora->tok_embeddings_a);
+    ggml_allocr_alloc(alloc, lora->tok_embeddings_b);
+    ggml_allocr_alloc(alloc, lora->norm_a);
+    ggml_allocr_alloc(alloc, lora->norm_b);
+    ggml_allocr_alloc(alloc, lora->output_a);
+    ggml_allocr_alloc(alloc, lora->output_b);
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+        ggml_allocr_alloc(alloc, layer.attention_norm_a);
+        ggml_allocr_alloc(alloc, layer.attention_norm_b);
+        ggml_allocr_alloc(alloc, layer.wq_a);
+        ggml_allocr_alloc(alloc, layer.wq_b);
+        ggml_allocr_alloc(alloc, layer.wk_a);
+        ggml_allocr_alloc(alloc, layer.wk_b);
+        ggml_allocr_alloc(alloc, layer.wv_a);
+        ggml_allocr_alloc(alloc, layer.wv_b);
+        ggml_allocr_alloc(alloc, layer.wo_a);
+        ggml_allocr_alloc(alloc, layer.wo_b);
+        ggml_allocr_alloc(alloc, layer.ffn_norm_a);
+        ggml_allocr_alloc(alloc, layer.ffn_norm_b);
+        ggml_allocr_alloc(alloc, layer.w1_a);
+        ggml_allocr_alloc(alloc, layer.w1_b);
+        ggml_allocr_alloc(alloc, layer.w2_a);
+        ggml_allocr_alloc(alloc, layer.w2_b);
+        ggml_allocr_alloc(alloc, layer.w3_a);
+        ggml_allocr_alloc(alloc, layer.w3_b);
+    }
+    ggml_allocr_alloc(alloc, lora->tok_embeddings_a->grad);
+    ggml_allocr_alloc(alloc, lora->tok_embeddings_b->grad);
+    ggml_allocr_alloc(alloc, lora->norm_a->grad);
+    ggml_allocr_alloc(alloc, lora->norm_b->grad);
+    ggml_allocr_alloc(alloc, lora->output_a->grad);
+    ggml_allocr_alloc(alloc, lora->output_b->grad);
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+        ggml_allocr_alloc(alloc, layer.attention_norm_a->grad);
+        ggml_allocr_alloc(alloc, layer.attention_norm_b->grad);
+        ggml_allocr_alloc(alloc, layer.wq_a->grad);
+        ggml_allocr_alloc(alloc, layer.wq_b->grad);
+        ggml_allocr_alloc(alloc, layer.wk_a->grad);
+        ggml_allocr_alloc(alloc, layer.wk_b->grad);
+        ggml_allocr_alloc(alloc, layer.wv_a->grad);
+        ggml_allocr_alloc(alloc, layer.wv_b->grad);
+        ggml_allocr_alloc(alloc, layer.wo_a->grad);
+        ggml_allocr_alloc(alloc, layer.wo_b->grad);
+        ggml_allocr_alloc(alloc, layer.ffn_norm_a->grad);
+        ggml_allocr_alloc(alloc, layer.ffn_norm_b->grad);
+        ggml_allocr_alloc(alloc, layer.w1_a->grad);
+        ggml_allocr_alloc(alloc, layer.w1_b->grad);
+        ggml_allocr_alloc(alloc, layer.w2_a->grad);
+        ggml_allocr_alloc(alloc, layer.w2_b->grad);
+        ggml_allocr_alloc(alloc, layer.w3_a->grad);
+        ggml_allocr_alloc(alloc, layer.w3_b->grad);
+    }
+}
+
+static void init_lora(const struct my_llama_model * model, struct my_llama_lora * lora) {
+    const auto & lparams = lora->hparams;
+
+    const uint32_t n_embd     = model->hparams.n_embd;
+    const uint32_t n_embd_gqa = model->hparams.n_embd_gqa();
+    const uint32_t n_layer    = model->hparams.n_layer;
+    const uint32_t n_vocab    = model->hparams.n_vocab;
+    const uint32_t n_ff       = model->hparams.n_ff;
+
+    std::vector<char> tn_buf;
+    tn_buf.resize(GGML_MAX_NAME);
+    auto tn = [&tn_buf](const char * key, const char * suffix) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), "%s%s", key, suffix);
+        return tn_buf.data();
+    };
+    auto tni = [&tn_buf](const char * key, const char * suffix, int bid) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), key, bid);
+        std::string s = tn_buf.data();
+        snprintf(tn_buf.data(), tn_buf.size(), "%s%s", s.c_str(), suffix);
+        return tn_buf.data();
+    };
+
+    // context for lora tensors without their data
+    struct ggml_init_params ctx_lora_params;
+    ctx_lora_params.mem_size   = ggml_tensor_overhead()*2*(6 + n_layer*18);
+    ctx_lora_params.mem_buffer = NULL;
+    ctx_lora_params.no_alloc   = true;
+
+    struct ggml_context * ctx = ggml_init(ctx_lora_params);
+    lora->ctx = ctx;
+
+    lora->tok_embeddings_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_tok_embeddings, n_embd);
+    lora->tok_embeddings_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_tok_embeddings, n_vocab);
+    lora->norm_a           = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_norm, n_embd);
+    lora->norm_b           = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_norm, 1);
+    lora->output_a         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_output, n_embd);
+    lora->output_b         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_output, n_vocab);
+
+    ggml_set_name(lora->tok_embeddings_a, tn(LLM_TENSOR_TOKEN_EMBD,  ".weight.lora_a"));
+    ggml_set_name(lora->tok_embeddings_b, tn(LLM_TENSOR_TOKEN_EMBD,  ".weight.lora_b"));
+    ggml_set_name(lora->norm_a,           tn(LLM_TENSOR_OUTPUT_NORM, ".weight.lora_a"));
+    ggml_set_name(lora->norm_b,           tn(LLM_TENSOR_OUTPUT_NORM, ".weight.lora_b"));
+    ggml_set_name(lora->output_a,         tn(LLM_TENSOR_OUTPUT,      ".weight.lora_a"));
+    ggml_set_name(lora->output_b,         tn(LLM_TENSOR_OUTPUT,      ".weight.lora_b"));
+
+    lora->layers.resize(n_layer);
+    for (uint32_t i = 0; i < n_layer; ++i) {
+        auto & layer = lora->layers[i];
+
+        layer.attention_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_attention_norm, n_embd);
+        layer.attention_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_attention_norm, 1);
+
+        layer.wq_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wq, n_embd);
+        layer.wq_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wq, n_embd);
+        layer.wk_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wk, n_embd);
+        layer.wk_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wk, n_embd_gqa);
+        layer.wv_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wv, n_embd);
+        layer.wv_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wv, n_embd_gqa);
+        layer.wo_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wo, n_embd);
+        layer.wo_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wo, n_embd);
+
+        layer.ffn_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, n_embd);
+        layer.ffn_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, 1);
+
+        layer.w1_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_embd);
+        layer.w1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_ff);
+        layer.w2_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_ff);
+        layer.w2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_embd);
+        layer.w3_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_embd);
+        layer.w3_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_ff);
+
+        ggml_set_name(layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_a", i));
+        ggml_set_name(layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_b", i));
+        ggml_set_name(layer.wq_a,             tni(LLM_TENSOR_ATTN_Q,    ".weight.lora_a", i));
+        ggml_set_name(layer.wq_b,             tni(LLM_TENSOR_ATTN_Q,    ".weight.lora_b", i));
+        ggml_set_name(layer.wk_a,             tni(LLM_TENSOR_ATTN_K,    ".weight.lora_a", i));
+        ggml_set_name(layer.wk_b,             tni(LLM_TENSOR_ATTN_K,    ".weight.lora_b", i));
+        ggml_set_name(layer.wv_a,             tni(LLM_TENSOR_ATTN_V,    ".weight.lora_a", i));
+        ggml_set_name(layer.wv_b,             tni(LLM_TENSOR_ATTN_V,    ".weight.lora_b", i));
+        ggml_set_name(layer.wo_a,             tni(LLM_TENSOR_ATTN_OUT,  ".weight.lora_a", i));
+        ggml_set_name(layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_b", i));
+        ggml_set_name(layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
+        ggml_set_name(layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
+        ggml_set_name(layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
+        ggml_set_name(layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
+        ggml_set_name(layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
+        ggml_set_name(layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
+    }
+
+    set_param_lora(lora);
+
+    // measure data size
+    struct ggml_allocr * alloc = NULL;
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    alloc_lora(alloc, lora);
+
+    // allocate data
+    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);
+}
+
+static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, float std, float min, float max) {
+    const uint32_t n_layer = lora->layers.size();
+
+    struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
+
+    randomize_tensor_normal(lora->tok_embeddings_a, rnd);
+    randomize_tensor_normal(lora->tok_embeddings_b, rnd);
+    randomize_tensor_normal(lora->norm_a,           rnd);
+    randomize_tensor_normal(lora->norm_b,           rnd);
+    randomize_tensor_normal(lora->output_a,         rnd);
+    randomize_tensor_normal(lora->output_b,         rnd);
+
+    for (uint32_t i = 0; i < n_layer; ++i) {
+        auto & layer = lora->layers[i];
+        randomize_tensor_normal(layer.attention_norm_a, rnd);
+        randomize_tensor_normal(layer.attention_norm_b, rnd);
+
+        randomize_tensor_normal(layer.wq_a, rnd);
+        randomize_tensor_normal(layer.wq_b, rnd);
+        randomize_tensor_normal(layer.wk_a, rnd);
+        randomize_tensor_normal(layer.wk_b, rnd);
+        randomize_tensor_normal(layer.wv_a, rnd);
+        randomize_tensor_normal(layer.wv_b, rnd);
+        randomize_tensor_normal(layer.wo_a, rnd);
+        randomize_tensor_normal(layer.wo_b, rnd);
+
+        randomize_tensor_normal(layer.ffn_norm_a, rnd);
+        randomize_tensor_normal(layer.ffn_norm_b, rnd);
+
+        randomize_tensor_normal(layer.w1_a, rnd);
+        randomize_tensor_normal(layer.w1_b, rnd);
+        randomize_tensor_normal(layer.w2_a, rnd);
+        randomize_tensor_normal(layer.w2_b, rnd);
+        randomize_tensor_normal(layer.w3_a, rnd);
+        randomize_tensor_normal(layer.w3_b, rnd);
+    }
+
+    free_random_normal_distribution(rnd);
+}
+
+static struct ggml_tensor * llama_build_lora_finetune_graphs(
+        struct my_llama_model * model,
+        struct my_llama_lora  * lora,
+        struct ggml_allocr    * alloc,
+        struct ggml_context   * ctx,
+        struct ggml_cgraph    * gf,
+        struct ggml_cgraph    * gb,
+        struct ggml_cgraph    * gb_tmp,
+        struct ggml_tensor  * * logits,
+        struct ggml_tensor    * tokens_input,
+        struct ggml_tensor    * targets,
+        const  int              n_tokens,
+        const  int              n_batch,
+        const  bool             enable_flash_attn,
+        const  bool             enable_checkpointing) {
+
+    ggml_set_scratch(ctx, { 0, 0, nullptr, });
+    const int n_past = 0;
+    const int N = n_tokens;
+    const auto & hparams  = model->hparams;
+    const int n_ctx       = hparams.n_ctx;
+    const int n_vocab     = hparams.n_vocab;
+    const int n_embd      = hparams.n_embd;
+    const int n_layer     = hparams.n_layer;
+    const int n_head      = hparams.n_head;
+    const int n_head_kv   = hparams.n_head_kv;
+    const int n_ff        = hparams.n_ff;
+    const int n_rot       = hparams.n_embd_head();
+    const int n_embd_head = hparams.n_embd_head();
+    const int n_embd_gqa  = hparams.n_embd_gqa();
+    const float rms_norm_eps    = hparams.f_norm_rms_eps;
+    const float rope_freq_base  = hparams.rope_freq_base;
+    const float rope_freq_scale = hparams.rope_freq_scale;
+
+    GGML_ASSERT((size_t) n_layer == lora->layers.size());
+
+    auto set_name = [](struct ggml_tensor * t, const char * n) {
+        ggml_set_name(t, n);
+        if (t->grad) {
+            ggml_format_name(t->grad, "%s->grad", n);
+        }
+    };
+
+    // KQ_pos - contains the positions
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
+    ggml_allocr_alloc(alloc, KQ_pos);
+    if (!ggml_allocr_is_measure(alloc)) {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
+    // rope has so much parameters that we make a custom function for it
+    auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
+                (struct ggml_tensor * t) -> struct ggml_tensor * {
+        // not capturing these, to silcence warnings
+        const int rope_mode = 0;
+
+        return ggml_rope_custom(ctx,
+            t, KQ_pos, n_rot, rope_mode, n_ctx,
+            rope_freq_base, rope_freq_scale);
+    };
+
+    set_name(tokens_input, "tokens_input");
+    set_name(targets,      "targets");
+
+    GGML_ASSERT(tokens_input->type == GGML_TYPE_I32);
+
+    auto add_to_f32 = [] (struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
+        if (ggml_is_quantized(a->type)) {
+            return ggml_add_cast(ctx, a, b, GGML_TYPE_F32);
+        } else if (a->type == GGML_TYPE_F32) {
+            return ggml_add(ctx, a, b);
+        } else {
+            die_fmt("%s: Finetuning on tensors with type '%s' is not yet supported.\n",
+                __func__, ggml_type_name(a->type));
+        }
+    };
+
+    struct ggml_tensor * tok_embeddings = add_to_f32(ctx, model->tok_embeddings, ggml_mul_mat(ctx, lora->tok_embeddings_a, lora->tok_embeddings_b));
+    struct ggml_tensor * norm           = add_to_f32(ctx, model->norm, ggml_mul_mat(ctx, lora->norm_a, lora->norm_b));
+    struct ggml_tensor * output         = add_to_f32(ctx, model->output, ggml_mul_mat(ctx, lora->output_a, lora->output_b));
+
+    struct ggml_tensor * t00 = ggml_reshape_1d(ctx, tokens_input, N*n_batch);  set_name(t00, "t00"); assert_shape_1d(t00, N*n_batch);
+    struct ggml_tensor * t01 = ggml_get_rows(ctx, tok_embeddings, t00);        set_name(t01, "t01"); assert_shape_2d(t01, n_embd, N*n_batch);
+
+    struct ggml_tensor * cur = t01;
+
+    std::vector<struct ggml_tensor *> checkpoints;
+    if (enable_checkpointing) {
+        checkpoints.push_back(tokens_input);
+        checkpoints.push_back(targets);
+        checkpoints.push_back(t00);
+        checkpoints.push_back(t01);
+    }
+
+    struct ggml_tensor * kv_scale = NULL;
+    if (!enable_flash_attn) {
+        kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct my_llama_layer & layer = model->layers[il];
+        struct my_llama_lora_layer & llayer = lora->layers[il];
+
+        struct ggml_tensor * attention_norm = add_to_f32(ctx, layer.attention_norm, ggml_mul_mat(ctx, llayer.attention_norm_a, llayer.attention_norm_b));
+        struct ggml_tensor * ffn_norm = add_to_f32(ctx, layer.ffn_norm, ggml_mul_mat(ctx, llayer.ffn_norm_a, llayer.ffn_norm_b));
+        struct ggml_tensor * wq = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
+        struct ggml_tensor * wk = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
+        struct ggml_tensor * wv = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
+        struct ggml_tensor * wo = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
+        struct ggml_tensor * w1 = add_to_f32(ctx, layer.w1, ggml_mul_mat(ctx, llayer.w1_a, llayer.w1_b));
+        struct ggml_tensor * w2 = add_to_f32(ctx, layer.w2, ggml_mul_mat(ctx, llayer.w2_a, llayer.w2_b));
+        struct ggml_tensor * w3 = add_to_f32(ctx, layer.w3, ggml_mul_mat(ctx, llayer.w3_a, llayer.w3_b));
+
+        struct ggml_tensor * t02 = ggml_rms_norm     (ctx, cur, rms_norm_eps);                       set_name(t02, "t02");     assert_shape_2d(t02, n_embd, N*n_batch);
+        struct ggml_tensor * t03 = ggml_repeat       (ctx, attention_norm, t02);                     set_name(t03, "t03");     assert_shape_2d(t03, n_embd, N*n_batch);
+        struct ggml_tensor * t04 = ggml_mul          (ctx, t03, t02);                                set_name(t04, "t04");     assert_shape_2d(t04, n_embd, N*n_batch);
+        struct ggml_tensor * t05 = ggml_mul_mat      (ctx, wq, t04);                                 set_name(t05, "t05");     assert_shape_2d(t05, n_embd, N*n_batch);
+        struct ggml_tensor * t06 = ggml_reshape_4d   (ctx, t05, n_embd_head, n_head, N, n_batch);    set_name(t06, "t06");     assert_shape_4d(t06, n_embd_head, n_head, N, n_batch);
+        struct ggml_tensor * t07 = rope              (t06);                                          set_name(t07, "t07");     assert_shape_4d(t07, n_embd_head, n_head, N, n_batch);
+        struct ggml_tensor * t08 = ggml_mul_mat      (ctx, wk, t04);                                 set_name(t08, "t08");     assert_shape_2d(t08, n_embd_gqa, N*n_batch);
+        struct ggml_tensor * t09 = ggml_reshape_4d   (ctx, t08, n_embd_head, n_head_kv, N, n_batch); set_name(t09, "t09");     assert_shape_4d(t09, n_embd_head, n_head_kv, N, n_batch);
+        struct ggml_tensor * t10 = rope              (t09);                                          set_name(t10, "t10");     assert_shape_4d(t10, n_embd_head, n_head_kv, N, n_batch);
+
+        struct ggml_tensor * t11;
+        if (ggml_is_quantized(wv->type)) {
+            struct ggml_tensor * t11_1 = ggml_mul_mat  (ctx, wv, t04);                               set_name(t11_1, "t11_1"); assert_shape_2d(t11_1, n_embd_gqa, N*n_batch);
+            struct ggml_tensor * t11_2 = ggml_transpose(ctx, t11_1);                                 set_name(t11_2, "t11_2"); assert_shape_2d(t11_2, N*n_batch, n_embd_gqa);
+                                 t11   = ggml_cont     (ctx, t11_2);                                 set_name(t11, "t11");     assert_shape_2d(t11, N*n_batch, n_embd_gqa);
+        } else {
+                                 t11   = ggml_mul_mat  (ctx, t04, wv);                               set_name(t11, "t11");     assert_shape_2d(t11, N*n_batch, n_embd_gqa);
+        }
+
+        struct ggml_tensor * t12 = ggml_reshape_4d   (ctx, t11, N, n_batch, n_embd_head, n_head_kv); set_name(t12, "t12");     assert_shape_4d(t12, N, n_batch, n_embd_head, n_head_kv);
+        struct ggml_tensor * t13 = ggml_permute      (ctx, t07, 0, 2, 1, 3);                         set_name(t13, "t13");     assert_shape_4d(t13, n_embd_head, N, n_head, n_batch);
+        struct ggml_tensor * t14 = ggml_permute      (ctx, t10, 0, 2, 1, 3);                         set_name(t14, "t14");     assert_shape_4d(t14, n_embd_head, N, n_head_kv, n_batch);
+        struct ggml_tensor * t15 = ggml_permute      (ctx, t12, 0, 3, 1, 2);                         set_name(t15, "t15");     assert_shape_4d(t15, N, n_embd_head, n_head_kv, n_batch);
+        struct ggml_tensor * t16;
+        if (enable_flash_attn) {
+            t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                         set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
+        } else {
+            struct ggml_tensor * t16_0 = ggml_mul_mat              (ctx, t14, t13);                  set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
+            struct ggml_tensor * t16_1 = ggml_scale_inplace        (ctx, t16_0, kv_scale);           set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);
+            struct ggml_tensor * t16_2 = ggml_diag_mask_inf_inplace(ctx, t16_1, n_past);             set_name(t16_2, "t16_2"); assert_shape_4d(t16_2, N, N, n_head, n_batch);
+            struct ggml_tensor * t16_3 = ggml_soft_max_inplace     (ctx, t16_2);                     set_name(t16_3, "t16_3"); assert_shape_4d(t16_3, N, N, n_head, n_batch);
+            t16 = ggml_mul_mat(ctx, t15, t16_3);                                                     set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
+        }
+        struct ggml_tensor * t17 = ggml_permute      (ctx, t16, 0, 2, 1, 3);                         set_name(t17, "t17");     assert_shape_4d(t17, n_embd_head, n_head, N, n_batch);
+        struct ggml_tensor * t18 = ggml_cont         (ctx, t17);                                     set_name(t18, "t18");     assert_shape_4d(t18, n_embd_head, n_head, N, n_batch);
+        struct ggml_tensor * t19 = ggml_reshape_2d   (ctx, t18, n_embd, N*n_batch);                  set_name(t19, "t19");     assert_shape_2d(t19, n_embd, N*n_batch);
+        struct ggml_tensor * t20 = ggml_mul_mat      (ctx, wo, t19);                                 set_name(t20, "t20");     assert_shape_2d(t20, n_embd, N*n_batch);
+        struct ggml_tensor * t21 = ggml_add          (ctx, t20, cur);                                set_name(t21, "t21");     assert_shape_2d(t21, n_embd, N*n_batch);
+        struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, rms_norm_eps);                       set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
+        struct ggml_tensor * t23 = ggml_repeat       (ctx, ffn_norm, t22);                           set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
+        struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                                set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, w3, t24);                                 set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, w1, t24);                                 set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                     set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
+        struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                                set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, w2, t28);                                 set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                                set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
+        cur = t30;
+        if (enable_checkpointing) {
+            checkpoints.push_back(cur);
+        }
+    }
+    struct ggml_tensor * t31   = ggml_rms_norm          (ctx, cur, rms_norm_eps);                    set_name(t31, "t31");     assert_shape_2d(t31, n_embd, N*n_batch);
+    struct ggml_tensor * t32   = ggml_repeat            (ctx, norm, t31);                            set_name(t32, "t32");     assert_shape_2d(t32, n_embd, N*n_batch);
+    struct ggml_tensor * t33   = ggml_mul               (ctx, t32, t31);                             set_name(t33, "t33");     assert_shape_2d(t33, n_embd, N*n_batch);
+    struct ggml_tensor * t34   = ggml_mul_mat           (ctx, output, t33);                          set_name(t34, "t34");     assert_shape_2d(t34, n_vocab, N*n_batch);
+    struct ggml_tensor * t35   = ggml_reshape_3d        (ctx, t34, n_vocab, N, n_batch);             set_name(t35, "t35");     assert_shape_3d(t35, n_vocab, N, n_batch);
+    struct ggml_tensor * t36   = ggml_cross_entropy_loss(ctx, t35, targets);                         set_name(t36, "t36");     assert_shape_1d(t36, 1);
+
+    if (enable_checkpointing) {
+        checkpoints.push_back(t31);
+        checkpoints.push_back(t32);
+        checkpoints.push_back(t33);
+        checkpoints.push_back(t34);
+        checkpoints.push_back(t35);
+        checkpoints.push_back(t36);
+    }
+
+    ggml_build_forward_expand(gf, t36);
+
+    if (enable_checkpointing) {
+        ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size());
+    } else {
+        *gb = *gf;
+        ggml_build_backward_expand(ctx, gf, gb, true);
+    }
+
+    GGML_ASSERT(alloc != NULL);
+
+    // make sure some tensors are not reallocated by inserting new temporary nodes depending on them
+    int n_leafs_before = gb->n_leafs;
+    int n_nodes_before = gb->n_nodes;
+    struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
+    // output tensors
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
+    // input gradient
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
+    GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
+    ggml_allocr_alloc(alloc, t36->grad);
+    // KQ_pos
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
+
+    // make sure base model tensors data cannot be used in viewable operations
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, one));
+    for (int il = 0; il < n_layer; ++il) {
+        struct my_llama_layer & layer = model->layers[il];
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, one));
+    }
+
+    // allocating checkpoints in one block to reduce memory fragmentation
+    // note: they will be freed in reverse order
+    for (unsigned int i = 0; i < checkpoints.size(); ++i) {
+        if (checkpoints[i]->data == NULL && checkpoints[i]->view_src == NULL) {
+            ggml_allocr_alloc(alloc, checkpoints[i]);
+        }
+    }
+
+    ggml_allocr_alloc_graph(alloc, gb);
+
+    // remove the additional nodes and leafs
+    for (int i = n_leafs_before; i < gb->n_leafs; ++i) {
+        gb->leafs[i] = NULL;
+    }
+    for (int i = n_nodes_before; i < gb->n_nodes; ++i) {
+        gb->nodes[i] = NULL;
+    }
+    gb->n_leafs = n_leafs_before;
+    gb->n_nodes = n_nodes_before;
+
+    *logits = t35;
+    return t36;
+}
+
+static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct my_llama_lora * lora) {
+    // NOTE: gguf_context must be initialized with f_ggml_ctx and no_alloc=false, otherwise tensor data can not be read
+
+    std::string arch;
+
+    std::vector<char> keybuf;
+    keybuf.resize(512);
+
+    GGUF_GET_KEY(fctx, arch, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_GENERAL_ARCHITECTURE);
+    GGML_ASSERT(arch == "llama");
+
+    uint32_t ftype_u;
+    GGUF_GET_KEY(fctx, ftype_u, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_GENERAL_FILE_TYPE);
+    GGML_ASSERT((enum llama_ftype) ftype_u == LLAMA_FTYPE_ALL_F32);
+
+    struct my_llama_hparams hparams;
+    load_model_hparams_gguf(fctx, &hparams, arch.c_str());
+
+    // parameters that define tensor shapes must match
+    GGML_ASSERT(hparams.n_embd    == model->hparams.n_embd);
+    GGML_ASSERT(hparams.n_ff      == model->hparams.n_ff);
+    GGML_ASSERT(hparams.n_head    == model->hparams.n_head);
+    GGML_ASSERT(hparams.n_head_kv == model->hparams.n_head_kv);
+    GGML_ASSERT(hparams.n_layer   == model->hparams.n_layer);
+
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_tok_embeddings, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_norm,           gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_output,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_OUTPUT);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_attention_norm, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_NORM);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_wq,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_Q);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_wk,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_K);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_wv,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_V);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_wo,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_norm,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_NORM);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w1,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w2,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w3,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
+
+    init_lora(model, lora);
+
+    copy_tensor_by_name(lora->tok_embeddings_a, f_ggml_ctx, ggml_get_name(lora->tok_embeddings_a));
+    copy_tensor_by_name(lora->tok_embeddings_b, f_ggml_ctx, ggml_get_name(lora->tok_embeddings_b));
+    copy_tensor_by_name(lora->norm_a,           f_ggml_ctx, ggml_get_name(lora->norm_a));
+    copy_tensor_by_name(lora->norm_b,           f_ggml_ctx, ggml_get_name(lora->norm_b));
+    copy_tensor_by_name(lora->output_a,         f_ggml_ctx, ggml_get_name(lora->output_a));
+    copy_tensor_by_name(lora->output_b,         f_ggml_ctx, ggml_get_name(lora->output_b));
+
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+        copy_tensor_by_name(layer.attention_norm_a, f_ggml_ctx, ggml_get_name(layer.attention_norm_a));
+        copy_tensor_by_name(layer.attention_norm_b, f_ggml_ctx, ggml_get_name(layer.attention_norm_b));
+        copy_tensor_by_name(layer.wq_a,             f_ggml_ctx, ggml_get_name(layer.wq_a));
+        copy_tensor_by_name(layer.wq_b,             f_ggml_ctx, ggml_get_name(layer.wq_b));
+        copy_tensor_by_name(layer.wk_a,             f_ggml_ctx, ggml_get_name(layer.wk_a));
+        copy_tensor_by_name(layer.wk_b,             f_ggml_ctx, ggml_get_name(layer.wk_b));
+        copy_tensor_by_name(layer.wv_a,             f_ggml_ctx, ggml_get_name(layer.wv_a));
+        copy_tensor_by_name(layer.wv_b,             f_ggml_ctx, ggml_get_name(layer.wv_b));
+        copy_tensor_by_name(layer.wo_a,             f_ggml_ctx, ggml_get_name(layer.wo_a));
+        copy_tensor_by_name(layer.wo_b,             f_ggml_ctx, ggml_get_name(layer.wo_b));
+        copy_tensor_by_name(layer.ffn_norm_a,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_a));
+        copy_tensor_by_name(layer.ffn_norm_b,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_b));
+        copy_tensor_by_name(layer.w1_a,             f_ggml_ctx, ggml_get_name(layer.w1_a));
+        copy_tensor_by_name(layer.w1_b,             f_ggml_ctx, ggml_get_name(layer.w1_b));
+        copy_tensor_by_name(layer.w2_a,             f_ggml_ctx, ggml_get_name(layer.w2_a));
+        copy_tensor_by_name(layer.w2_b,             f_ggml_ctx, ggml_get_name(layer.w2_b));
+        copy_tensor_by_name(layer.w3_a,             f_ggml_ctx, ggml_get_name(layer.w3_a));
+        copy_tensor_by_name(layer.w3_b,             f_ggml_ctx, ggml_get_name(layer.w3_b));
+    }
+}
+
+static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_model * model, struct my_llama_lora * lora) {
+    const char * arch = "llama";
+    enum llama_ftype ftype = LLAMA_FTYPE_ALL_F32;
+
+    std::vector<char> keybuf;
+    keybuf.resize(512);
+    auto kv = [arch, &keybuf](const char * key) -> const char * {
+        snprintf(keybuf.data(), keybuf.size(), key, arch);
+        return keybuf.data();
+    };
+
+    gguf_set_val_str(fctx, LLM_KV_GENERAL_ARCHITECTURE, arch);
+    gguf_set_val_u32(fctx, LLM_KV_GENERAL_FILE_TYPE, ftype);
+
+    gguf_set_val_u32(fctx, kv(LLM_KV_CONTEXT_LENGTH),              model->hparams.n_ctx);
+    gguf_set_val_u32(fctx, kv(LLM_KV_EMBEDDING_LENGTH),            model->hparams.n_embd);
+    gguf_set_val_u32(fctx, kv(LLM_KV_FEED_FORWARD_LENGTH),         model->hparams.n_ff);
+    gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT),        model->hparams.n_head);
+    gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV),     model->hparams.n_head_kv);
+    gguf_set_val_u32(fctx, kv(LLM_KV_BLOCK_COUNT),                 model->hparams.n_layer);
+    gguf_set_val_u32(fctx, kv(LLM_KV_ROPE_DIMENSION_COUNT),        model->hparams.n_embd_head());
+    gguf_set_val_f32(fctx, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS), model->hparams.f_norm_rms_eps);
+    gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_FREQ_BASE),              model->hparams.rope_freq_base);
+    gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_SCALE_LINEAR),           model->hparams.rope_freq_scale);
+
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD,   lora->hparams.n_rank_tok_embeddings);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM,  lora->hparams.n_rank_norm);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_OUTPUT,       lora->hparams.n_rank_output);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_NORM,    lora->hparams.n_rank_attention_norm);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_Q,       lora->hparams.n_rank_wq);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_K,       lora->hparams.n_rank_wk);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_V,       lora->hparams.n_rank_wv);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,     lora->hparams.n_rank_wo);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_NORM,     lora->hparams.n_rank_ffn_norm);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_w1);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_w2);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_w3);
+
+    gguf_add_tensor(fctx, lora->tok_embeddings_a);
+    gguf_add_tensor(fctx, lora->tok_embeddings_b);
+    gguf_add_tensor(fctx, lora->norm_a);
+    gguf_add_tensor(fctx, lora->norm_b);
+    gguf_add_tensor(fctx, lora->output_a);
+    gguf_add_tensor(fctx, lora->output_b);
+
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+
+        gguf_add_tensor(fctx, layer.attention_norm_a);
+        gguf_add_tensor(fctx, layer.attention_norm_b);
+        gguf_add_tensor(fctx, layer.wq_a);
+        gguf_add_tensor(fctx, layer.wq_b);
+        gguf_add_tensor(fctx, layer.wk_a);
+        gguf_add_tensor(fctx, layer.wk_b);
+        gguf_add_tensor(fctx, layer.wv_a);
+        gguf_add_tensor(fctx, layer.wv_b);
+        gguf_add_tensor(fctx, layer.wo_a);
+        gguf_add_tensor(fctx, layer.wo_b);
+        gguf_add_tensor(fctx, layer.ffn_norm_a);
+        gguf_add_tensor(fctx, layer.ffn_norm_b);
+        gguf_add_tensor(fctx, layer.w1_a);
+        gguf_add_tensor(fctx, layer.w1_b);
+        gguf_add_tensor(fctx, layer.w2_a);
+        gguf_add_tensor(fctx, layer.w2_b);
+        gguf_add_tensor(fctx, layer.w3_a);
+        gguf_add_tensor(fctx, layer.w3_b);
+    }
+}
+
+static void load_checkpoint_lora_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) {
+    std::string train_type = LLM_KV_TRAINING_TYPE_FINETUNE_LORA;
+    GGUF_GET_KEY(fctx, train_type, gguf_get_val_str, GGUF_TYPE_STRING, false, LLM_KV_TRAINING_TYPE);
+    GGML_ASSERT(train_type == LLM_KV_TRAINING_TYPE_FINETUNE_LORA);
+
+    load_train_state_gguf(fctx, f_ggml_ctx, train);
+    load_llama_lora_gguf(fctx, f_ggml_ctx, model, lora);
+}
+
+static void save_checkpoint_lora_gguf(struct gguf_context * fctx, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) {
+    gguf_set_val_str(fctx, LLM_KV_TRAINING_TYPE, LLM_KV_TRAINING_TYPE_FINETUNE_LORA);
+    save_llama_lora_gguf(fctx, model, lora);
+    save_train_state_gguf(fctx, train);
+}
+
+static bool load_checkpoint_lora_file(const char * filename, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) {
+    struct ggml_context * f_ggml_ctx;
+    struct gguf_init_params params;
+    params.no_alloc = false;
+    params.ctx = &f_ggml_ctx;
+    struct gguf_context * fctx = gguf_init_from_file(filename, params);
+    if (fctx == NULL) {
+        return false;
+    }
+
+    load_checkpoint_lora_gguf(fctx, f_ggml_ctx, model, lora, train);
+
+    gguf_free(fctx);
+    return true;
+}
+
+static void save_checkpoint_lora_file(const char * filename, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) {
+    printf("%s: saving to %s\n", __func__, filename);
+    struct gguf_context * fctx = gguf_init_empty();
+
+    save_checkpoint_lora_gguf(fctx, model, lora, train);
+
+    // write file
+    const bool only_meta = false;
+    gguf_write_to_file(fctx, filename, only_meta);
+    gguf_free(fctx);
+}
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            size = 0;
+        } else {
+            seek(0, SEEK_END);
+            size = tell();
+            seek(0, SEEK_SET);
+        }
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        GGML_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            die_fmt("read error: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            die("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            die_fmt("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+static void write_tensor(struct llama_file * file, struct ggml_tensor * tensor, const char * name) {
+    if (tensor == NULL) {
+        file->write_u32(0);
+        file->write_u32(0);
+        file->write_u32(GGML_TYPE_F32);
+        file->seek((0-file->tell()) & 31, SEEK_CUR);
+        return;
+    }
+    if (name == NULL) {
+        name = ggml_get_name(tensor);
+    }
+    uint32_t name_len = strlen(name);
+    uint32_t nd = tensor->n_dims;
+    uint32_t ne[4] = { (uint32_t)tensor->ne[0],
+                       (uint32_t)tensor->ne[1],
+                       (uint32_t)tensor->ne[2],
+                       (uint32_t)tensor->ne[3] };
+    file->write_u32(nd);
+    file->write_u32(name_len);
+    file->write_u32(tensor->type);
+    file->write_raw(ne, sizeof(ne[0]) * nd);
+    file->write_raw(name, name_len);
+    file->seek((0-file->tell()) & 31, SEEK_CUR);
+    file->write_raw(tensor->data, ggml_nbytes(tensor));
+}
+
+static void save_as_llama_lora(const char * filename, struct my_llama_lora * lora) {
+    printf("%s: saving to %s\n", __func__, filename);
+    struct llama_file file(filename, "wb");
+    if (file.fp == NULL) {
+        return;
+    }
+
+    std::vector<char> tn_buf;
+    tn_buf.resize(GGML_MAX_NAME);
+
+    auto tn = [&tn_buf](const char * key, const char * suffix) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), "%s%s", key, suffix);
+        return tn_buf.data();
+    };
+
+    auto tni = [&tn_buf](const char * key, int bid, const char * suffix) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), key, bid);
+        std::string s = tn_buf.data();
+        snprintf(tn_buf.data(), tn_buf.size(), "%s%s", s.c_str(), suffix);
+        return tn_buf.data();
+    };
+
+    uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
+    // write_magic
+    file.write_u32(LLAMA_FILE_MAGIC_LORA);   // magic
+    file.write_u32(1); // version
+    // write_hparams
+    file.write_u32(lora->hparams.lora_r);
+    file.write_u32(lora->hparams.lora_alpha);
+    // write tensors
+    write_tensor(&file, lora->tok_embeddings_a, tn(LLM_TENSOR_TOKEN_EMBD,  ".weight.loraA"));
+    write_tensor(&file, lora->tok_embeddings_b, tn(LLM_TENSOR_TOKEN_EMBD,  ".weight.loraB"));
+    write_tensor(&file, lora->norm_a,           tn(LLM_TENSOR_OUTPUT_NORM, ".weight.loraA"));
+    write_tensor(&file, lora->norm_b,           tn(LLM_TENSOR_OUTPUT_NORM, ".weight.loraB"));
+    write_tensor(&file, lora->output_a,         tn(LLM_TENSOR_OUTPUT,      ".weight.loraA"));
+    write_tensor(&file, lora->output_b,         tn(LLM_TENSOR_OUTPUT,      ".weight.loraB"));
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+        write_tensor(&file, layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, i, ".weight.loraA"));
+        write_tensor(&file, layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, i, ".weight.loraB"));
+        write_tensor(&file, layer.wq_a,             tni(LLM_TENSOR_ATTN_Q,    i, ".weight.loraA"));
+        write_tensor(&file, layer.wq_b,             tni(LLM_TENSOR_ATTN_Q,    i, ".weight.loraB"));
+        write_tensor(&file, layer.wk_a,             tni(LLM_TENSOR_ATTN_K,    i, ".weight.loraA"));
+        write_tensor(&file, layer.wk_b,             tni(LLM_TENSOR_ATTN_K,    i, ".weight.loraB"));
+        write_tensor(&file, layer.wv_a,             tni(LLM_TENSOR_ATTN_V,    i, ".weight.loraA"));
+        write_tensor(&file, layer.wv_b,             tni(LLM_TENSOR_ATTN_V,    i, ".weight.loraB"));
+        write_tensor(&file, layer.wo_a,             tni(LLM_TENSOR_ATTN_OUT,  i, ".weight.loraA"));
+        write_tensor(&file, layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraB"));
+        write_tensor(&file, layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
+        write_tensor(&file, layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
+        write_tensor(&file, layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
+        write_tensor(&file, layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
+        write_tensor(&file, layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
+        write_tensor(&file, layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
+    }
+}
+
+struct train_params {
+    struct train_params_common common;
+
+    const char * fn_model_base;
+    const char * fn_lora_out;
+
+    bool only_write_lora;
+
+    float f_norm_rms_eps;
+    float rope_freq_base;
+    float rope_freq_scale;
+
+    bool custom_f_norm_rms_eps;
+    bool custom_rope_freq_base;
+    bool custom_rope_freq_scale;
+
+    int32_t lora_r;
+    int32_t lora_alpha;
+    bool custom_lora_alpha;
+
+    uint32_t n_rank_attention_norm;
+    uint32_t n_rank_wq;
+    uint32_t n_rank_wk;
+    uint32_t n_rank_wv;
+    uint32_t n_rank_wo;
+    uint32_t n_rank_ffn_norm;
+    uint32_t n_rank_w1;
+    uint32_t n_rank_w2;
+    uint32_t n_rank_w3;
+    uint32_t n_rank_tok_embeddings;
+    uint32_t n_rank_norm;
+    uint32_t n_rank_output;
+
+    bool custom_n_rank_attention_norm;
+    bool custom_n_rank_wq;
+    bool custom_n_rank_wk;
+    bool custom_n_rank_wv;
+    bool custom_n_rank_wo;
+    bool custom_n_rank_ffn_norm;
+    bool custom_n_rank_w1;
+    bool custom_n_rank_w2;
+    bool custom_n_rank_w3;
+    bool custom_n_rank_tok_embeddings;
+    bool custom_n_rank_norm;
+    bool custom_n_rank_output;
+};
+
+static struct train_params get_default_train_params() {
+    struct train_params params;
+    params.common = get_default_train_params_common();
+    params.fn_model_base     = "";
+    params.fn_lora_out       = "ggml-lora-ITERATION-f32.gguf";
+
+    params.only_write_lora = false;
+
+    params.f_norm_rms_eps  = 1e-5f;
+    params.rope_freq_base  = 10000.0f;
+    params.rope_freq_scale = 1.0f;
+
+    params.custom_f_norm_rms_eps  = false;
+    params.custom_rope_freq_base  = false;
+    params.custom_rope_freq_scale = false;
+
+    params.lora_r      = 4;
+    params.lora_alpha  = 4;
+    params.custom_lora_alpha = false;
+
+    params.n_rank_attention_norm = 1;
+    params.n_rank_wq             = 4;
+    params.n_rank_wk             = 4;
+    params.n_rank_wv             = 4;
+    params.n_rank_wo             = 4;
+    params.n_rank_ffn_norm       = 1;
+    params.n_rank_w1             = 4;
+    params.n_rank_w2             = 4;
+    params.n_rank_w3             = 4;
+    params.n_rank_tok_embeddings = 4;
+    params.n_rank_norm           = 1;
+    params.n_rank_output         = 4;
+
+    params.custom_n_rank_attention_norm = false;
+    params.custom_n_rank_wq             = false;
+    params.custom_n_rank_wk             = false;
+    params.custom_n_rank_wv             = false;
+    params.custom_n_rank_wo             = false;
+    params.custom_n_rank_ffn_norm       = false;
+    params.custom_n_rank_w1             = false;
+    params.custom_n_rank_w2             = false;
+    params.custom_n_rank_w3             = false;
+    params.custom_n_rank_tok_embeddings = false;
+    params.custom_n_rank_norm           = false;
+    params.custom_n_rank_output         = false;
+
+    return params;
+}
+
+static void train_print_usage(int argc, char ** argv, const struct train_params * params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help                 show this help message and exit\n");
+
+    fprintf(stderr, "  --model-base FNAME         model path from which to load base model (default '%s')\n", params->fn_model_base);
+    fprintf(stderr, "  --lora-out FNAME           path to save llama lora (default '%s')\n", params->fn_lora_out);
+    fprintf(stderr, "  --only-write-lora          only save llama lora, don't do any training.  use this if you only want to convert a checkpoint to a lora adapter.\n");
+    fprintf(stderr, "  --norm-rms-eps F           RMS-Norm epsilon value (default %f)\n", params->f_norm_rms_eps);
+    fprintf(stderr, "  --rope-freq-base F         Frequency base for ROPE (default %f)\n", params->rope_freq_base);
+    fprintf(stderr, "  --rope-freq-scale F        Frequency scale for ROPE (default %f)\n", params->rope_freq_scale);
+    fprintf(stderr, "  --lora-alpha N             LORA alpha : resulting LORA scaling is alpha/r. (default %d)\n", params->lora_alpha);
+    fprintf(stderr, "  --lora-r N                 LORA r: default rank. Also specifies resulting scaling together with lora-alpha. (default %d)\n", params->lora_r);
+    fprintf(stderr, "  --rank-att-norm N          LORA rank for attention norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n");
+    fprintf(stderr, "  --rank-ffn-norm N          LORA rank for feed-forward norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n");
+    fprintf(stderr, "  --rank-out-norm N          LORA rank for output norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n");
+    fprintf(stderr, "  --rank-tok-embd N          LORA rank for token embeddings tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-out N               LORA rank for output tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-wq N                LORA rank for wq tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-wk N                LORA rank for wk tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-wv N                LORA rank for wv tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-wo N                LORA rank for wo tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-w1 N                LORA rank for w1 tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-w2 N                LORA rank for w2 tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-w3 N                LORA rank for w3 tensor, overrides default rank.\n");
+
+    print_common_train_usage(argc, argv, &params->common);
+}
+
+static bool train_params_parse(int argc, char ** argv, struct train_params * params) {
+    bool invalid_param = false;
+    std::string arg;
+    struct train_params default_params = get_default_train_params();
+    const std::string arg_prefix = "--";
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (consume_common_train_arg(argc, argv, &i, &params->common, &invalid_param)) {
+            if (invalid_param) {
+                break;
+            } else if (params->common.print_usage) {
+                train_print_usage(argc, argv, &default_params);
+                exit(0);
+            }
+        } else if (arg == "--model-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_base = argv[i];
+        } else if (arg == "--lora-out") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_lora_out = argv[i];
+        } else if (arg == "--only-write-lora") {
+            params->only_write_lora = true;
+        } else if (arg == "--norm-rms-eps") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->f_norm_rms_eps = std::stof(argv[i]);
+            params->custom_f_norm_rms_eps = true;
+        } else if (arg == "--rope-freq-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->rope_freq_base = std::stof(argv[i]);
+            params->custom_rope_freq_base = true;
+        } else if (arg == "--rope-freq-scale") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->rope_freq_scale = std::stof(argv[i]);
+            params->custom_rope_freq_scale = true;
+        } else if (arg == "--lora-alpha") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->lora_alpha = std::stoi(argv[i]);
+            params->custom_lora_alpha = true;
+        } else if (arg == "--lora-r") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->lora_r = std::stoi(argv[i]);
+        } else if (arg == "--rank-att-norm") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_attention_norm = std::stoi(argv[i]);
+            params->custom_n_rank_attention_norm = true;
+        } else if (arg == "--rank-ffn-norm") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_ffn_norm = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_norm = true;
+        } else if (arg == "--rank-out-norm") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_norm = std::stoi(argv[i]);
+            params->custom_n_rank_norm = true;
+        } else if (arg == "--rank-tok-embd") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_tok_embeddings = std::stoi(argv[i]);
+            params->custom_n_rank_tok_embeddings = true;
+        } else if (arg == "--rank-out") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_output = std::stoi(argv[i]);
+            params->custom_n_rank_output = true;
+        } else if (arg == "--rank-wq") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_wq = std::stoi(argv[i]);
+            params->custom_n_rank_wq = true;
+        } else if (arg == "--rank-wk") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_wk = std::stoi(argv[i]);
+            params->custom_n_rank_wk = true;
+        } else if (arg == "--rank-wv") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_wv = std::stoi(argv[i]);
+            params->custom_n_rank_wv = true;
+        } else if (arg == "--rank-wo") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_wo = std::stoi(argv[i]);
+            params->custom_n_rank_wo = true;
+        } else if (arg == "--rank-w1") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_w1 = std::stoi(argv[i]);
+            params->custom_n_rank_w1 = true;
+        } else if (arg == "--rank-w2") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_w2 = std::stoi(argv[i]);
+            params->custom_n_rank_w2 = true;
+        } else if (arg == "--rank-w3") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_w3 = std::stoi(argv[i]);
+            params->custom_n_rank_w3 = true;
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            train_print_usage(argc, argv, &default_params);
+            exit(1);
+        }
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        train_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    finish_processing_train_args(&params->common);
+    return true;
+}
+
+struct save_train_files_data {
+    const char            * fn_checkpoint_out;
+    const char            * fn_lora_out;
+    const char            * pattern_fn_it;
+    const char            * fn_latest;
+    struct my_llama_model * model;
+    struct my_llama_lora  * lora;
+};
+
+static void save_train_files(void * vdata, struct train_state * train) {
+    struct save_train_files_data * data   = (struct save_train_files_data *) vdata;
+
+    int64_t iter = train->opt->iter;
+
+    if (strlen(data->fn_checkpoint_out) > 0) {
+        save_checkpoint_lora_file(get_train_filename(data->fn_checkpoint_out, data->pattern_fn_it, data->fn_latest, iter).c_str(), data->model, data->lora, train);
+        save_checkpoint_lora_file(get_train_filename(data->fn_checkpoint_out, data->pattern_fn_it, data->fn_latest, -1  ).c_str(), data->model, data->lora, train);
+    }
+    if (strlen(data->fn_lora_out) > 0) {
+        save_as_llama_lora(get_train_filename(data->fn_lora_out, data->pattern_fn_it, data->fn_latest, iter).c_str(), data->lora);
+        save_as_llama_lora(get_train_filename(data->fn_lora_out, data->pattern_fn_it, data->fn_latest, -1  ).c_str(), data->lora);
+    }
+}
+
+static int64_t get_parameter_count(struct my_llama_lora* lora) {
+    int64_t nx = 0;
+    nx += ggml_nelements(lora->tok_embeddings_a);
+    nx += ggml_nelements(lora->tok_embeddings_b);
+    nx += ggml_nelements(lora->norm_a);
+    nx += ggml_nelements(lora->norm_b);
+    nx += ggml_nelements(lora->output_a);
+    nx += ggml_nelements(lora->output_b);
+
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+        nx += ggml_nelements(layer.attention_norm_a);
+        nx += ggml_nelements(layer.attention_norm_b);
+        nx += ggml_nelements(layer.wq_a);
+        nx += ggml_nelements(layer.wq_b);
+        nx += ggml_nelements(layer.wk_a);
+        nx += ggml_nelements(layer.wk_b);
+        nx += ggml_nelements(layer.wv_a);
+        nx += ggml_nelements(layer.wv_b);
+        nx += ggml_nelements(layer.wo_a);
+        nx += ggml_nelements(layer.wo_b);
+        nx += ggml_nelements(layer.ffn_norm_a);
+        nx += ggml_nelements(layer.ffn_norm_b);
+        nx += ggml_nelements(layer.w1_a);
+        nx += ggml_nelements(layer.w1_b);
+        nx += ggml_nelements(layer.w2_a);
+        nx += ggml_nelements(layer.w2_b);
+        nx += ggml_nelements(layer.w3_a);
+        nx += ggml_nelements(layer.w3_b);
+    }
+    return nx;
+}
+
+int main(int argc, char ** argv) {
+    struct train_params params = get_default_train_params();
+
+    if (!train_params_parse(argc, argv, &params)) {
+        return 1;
+    }
+
+    if (params.common.seed == LLAMA_DEFAULT_SEED) {
+        params.common.seed = time(NULL);
+    }
+    printf("%s: seed: %u\n", __func__, params.common.seed);
+    srand(params.common.seed);
+
+    struct llama_model_params llama_mparams = llama_model_default_params();
+    llama_mparams.vocab_only = false;
+
+    printf("%s: model base = '%s'\n", __func__, params.fn_model_base);
+    struct llama_model * lmodel = llama_load_model_from_file(params.fn_model_base, llama_mparams);
+
+    struct llama_context_params llama_cparams = llama_context_default_params();
+    struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_cparams);
+
+    struct my_llama_model model;
+    init_model(lmodel, &model, params.fn_model_base, params.common.n_ctx);
+
+    struct my_llama_lora lora;
+
+    struct train_state      * train = init_train_state();
+    struct ggml_opt_context * opt   = train->opt;
+
+    // set params from command line
+    if (params.custom_f_norm_rms_eps) {
+        model.hparams.f_norm_rms_eps  = params.f_norm_rms_eps;
+    }
+    if (params.custom_rope_freq_base) {
+        model.hparams.rope_freq_base  = params.rope_freq_base;
+    }
+    if (params.custom_rope_freq_scale) {
+        model.hparams.rope_freq_scale = params.rope_freq_scale;
+    }
+    lora.hparams.lora_r                = params.lora_r;
+    lora.hparams.lora_alpha            = params.custom_lora_alpha            ? params.lora_alpha            : params.lora_r;
+    uint32_t n_rank_attention_norm     = params.custom_n_rank_attention_norm ? params.n_rank_attention_norm : 1;
+    uint32_t n_rank_wq                 = params.custom_n_rank_wq             ? params.n_rank_wq             : params.lora_r;
+    uint32_t n_rank_wk                 = params.custom_n_rank_wk             ? params.n_rank_wk             : params.lora_r;
+    uint32_t n_rank_wv                 = params.custom_n_rank_wv             ? params.n_rank_wv             : params.lora_r;
+    uint32_t n_rank_wo                 = params.custom_n_rank_wo             ? params.n_rank_wo             : params.lora_r;
+    uint32_t n_rank_ffn_norm           = params.custom_n_rank_ffn_norm       ? params.n_rank_ffn_norm       : 1;
+    uint32_t n_rank_w1                 = params.custom_n_rank_w1             ? params.n_rank_w1             : params.lora_r;
+    uint32_t n_rank_w2                 = params.custom_n_rank_w2             ? params.n_rank_w2             : params.lora_r;
+    uint32_t n_rank_w3                 = params.custom_n_rank_w3             ? params.n_rank_w3             : params.lora_r;
+    uint32_t n_rank_tok_embeddings     = params.custom_n_rank_tok_embeddings ? params.n_rank_tok_embeddings : params.lora_r;
+    uint32_t n_rank_norm               = params.custom_n_rank_norm           ? params.n_rank_norm           : 1;
+    uint32_t n_rank_output             = params.custom_n_rank_output         ? params.n_rank_output         : params.lora_r;
+    lora.hparams.n_rank_attention_norm = n_rank_attention_norm;
+    lora.hparams.n_rank_wq             = n_rank_wq;
+    lora.hparams.n_rank_wk             = n_rank_wk;
+    lora.hparams.n_rank_wv             = n_rank_wv;
+    lora.hparams.n_rank_wo             = n_rank_wo;
+    lora.hparams.n_rank_ffn_norm       = n_rank_ffn_norm;
+    lora.hparams.n_rank_w1             = n_rank_w1;
+    lora.hparams.n_rank_w2             = n_rank_w2;
+    lora.hparams.n_rank_w3             = n_rank_w3;
+    lora.hparams.n_rank_tok_embeddings = n_rank_tok_embeddings;
+    lora.hparams.n_rank_norm           = n_rank_norm;
+    lora.hparams.n_rank_output         = n_rank_output;
+
+    // set opt params from command line
+    opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
+    opt->params.print_forward_graph     = false;
+    opt->params.print_backward_graph    = false;
+    opt->params.n_threads               = params.common.n_threads;
+    opt->params.past                    = params.common.opt_past;
+    opt->params.delta                   = params.common.opt_delta;
+    opt->params.max_no_improvement      = params.common.opt_max_no_improvement;
+    opt->params.n_gradient_accumulation = params.common.n_gradient_accumulation;
+    opt->params.adam.n_iter             = params.common.adam_n_iter;
+    opt->params.adam.sched              = 1.0f;
+    opt->params.adam.alpha              = params.common.adam_alpha;
+    opt->params.adam.decay              = params.common.adam_decay;
+    opt->params.adam.decay_min_ndim     = params.common.adam_decay_min_ndim;
+    opt->params.adam.beta1              = params.common.adam_beta1;
+    opt->params.adam.beta2              = params.common.adam_beta2;
+    opt->params.adam.gclip              = params.common.adam_gclip;
+    opt->params.adam.eps_f              = params.common.adam_eps_f;
+
+    ggml_allocr * alloc = NULL;
+
+    printf("%s: init model\n", __func__);
+    bool existed = load_checkpoint_lora_file(params.common.fn_checkpoint_in, &model, &lora, train);
+
+    if (existed) {
+        // overwrite last n_ctx with user provided n_ctx
+        if (params.common.custom_n_ctx) {
+            model.hparams.n_ctx = params.common.n_ctx;
+        }
+
+        const bool opt_param_count_changed = (
+           (lora.hparams.n_rank_attention_norm != n_rank_attention_norm)
+        || (lora.hparams.n_rank_wq             != n_rank_wq)
+        || (lora.hparams.n_rank_wk             != n_rank_wk)
+        || (lora.hparams.n_rank_wv             != n_rank_wv)
+        || (lora.hparams.n_rank_wo             != n_rank_wo)
+        || (lora.hparams.n_rank_ffn_norm       != n_rank_ffn_norm)
+        || (lora.hparams.n_rank_w1             != n_rank_w1)
+        || (lora.hparams.n_rank_w2             != n_rank_w2)
+        || (lora.hparams.n_rank_w3             != n_rank_w3)
+        || (lora.hparams.n_rank_tok_embeddings != n_rank_tok_embeddings)
+        || (lora.hparams.n_rank_norm           != n_rank_norm)
+        || (lora.hparams.n_rank_output         != n_rank_output)
+        );
+
+        const bool opt_past_changed = opt->params.past != params.common.opt_past;
+
+        if (opt_param_count_changed) {
+            print_lora_params(&lora.hparams);
+            die("Provided rank differs from checkpoint file. To use different rank start finetune from scratch with empty input checkpoint, e.g --checkpoint-in ''. Aborting.");
+            // need to discard previous optimizer gradient statistics and opt_init with new shapes
+            // TODO
+        }
+        if (opt_past_changed) {
+            die("Optimizer parameter '--opt-past N' differs from checkpoint file. To use different value finetune from scratch with empty input checkpoint, e.g --checkpoint-in ''. Aborting");
+            // need to discard previous optimizer past function value statistics and opt_init with new shapes
+            // TODO
+        }
+    } else { // existed == false
+        init_lora(&model, &lora);
+        randomize_lora(&lora, params.common.seed, 0.0f, 1.0f, -1.0f, +1.0f);
+        if (!params.only_write_lora) {
+            ggml_opt_init(opt->ctx, opt, opt->params, get_parameter_count(&lora));
+        }
+    }
+    opt->iter = train->train_its;
+
+    print_params(&model.hparams);
+    print_lora_params(&lora.hparams);
+    printf("%s: total train_iterations %llu\n", __func__, (long long unsigned) train->train_its);
+    printf("%s: seen train_samples     %llu\n", __func__, (long long unsigned) train->train_samples);
+    printf("%s: seen train_tokens      %llu\n", __func__, (long long unsigned) train->train_tokens);
+    printf("%s: completed train_epochs %llu\n", __func__, (long long unsigned) train->train_epochs);
+    printf("%s: lora_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(lora.ctx) + lora.data.size()), (float) (ggml_used_mem(lora.ctx) + lora.data.size()) / (1024.0f*1024.0f));
+
+    if (params.only_write_lora) {
+        save_train_files_data save_data;
+        save_data.fn_checkpoint_out = "";
+        save_data.fn_lora_out       = params.fn_lora_out;
+        save_data.pattern_fn_it     = params.common.pattern_fn_it;
+        save_data.fn_latest         = params.common.fn_latest;
+        save_data.model             = &model;
+        save_data.lora              = &lora;
+
+        save_train_files(&save_data, train);
+
+        free_train_state(train);
+        ggml_free(lora.ctx);
+        llama_free(lctx);
+        llama_free_model(lmodel);
+        return 0;
+    }
+
+    printf("%s: opt_size  = %zu bytes (%.1f MB)\n", __func__, ggml_get_mem_size(opt->ctx), (float) ggml_get_mem_size(opt->ctx) / (1024.0f*1024.0f));
+    printf("%s: opt iter %d\n", __func__, opt->iter);
+
+    int n_tokens = model.hparams.n_ctx;
+    int n_vocab  = model.hparams.n_vocab;
+    int n_batch  = params.common.n_batch;
+
+
+    std::vector<uint8_t> mem_input_data;
+    std::vector<uint8_t> mem_compute_data;
+
+    // context for input tensors without their data
+    struct ggml_init_params ctx_input_params = {
+        ggml_tensor_overhead() * 2, // mem_size
+        NULL,                       // mem_buffer
+        true,                       // no_alloc
+    };
+    struct ggml_context * ctx_input = ggml_init(ctx_input_params);
+
+    // the input tensors
+    struct ggml_tensor * tokens_input  = ggml_new_tensor_2d(ctx_input, GGML_TYPE_I32, n_tokens, n_batch);
+    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
+    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
+    mem_input_data.resize(max_input_size);
+    alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
+    ggml_allocr_alloc(alloc, tokens_input);
+    ggml_allocr_alloc(alloc, target_probs);
+    ggml_allocr_free(alloc);
+
+    // context for compute tensors without their data
+    size_t estimated_compute_size_wo_data = (
+        ggml_tensor_overhead()*GGML_MAX_NODES*2
+      + (GGML_OBJECT_SIZE+GGML_GRAPH_SIZE)*(
+            params.common.use_checkpointing ? 3 : 2
+        )
+    );
+    struct ggml_init_params ctx_compute_params = {
+        estimated_compute_size_wo_data, // mem_size
+        NULL,                           // mem_buffer
+        true,                           // no_alloc
+    };
+    struct ggml_context * ctx_compute = NULL;
+
+    struct ggml_tensor * loss   = NULL;
+    struct ggml_tensor * logits = NULL;
+
+    struct ggml_cgraph * gf     = NULL;
+    struct ggml_cgraph * gb     = NULL;
+    struct ggml_cgraph * gb_tmp = NULL;
+
+    // measure required memory for compute tensors
+    size_t best_compute_size = SIZE_MAX;
+    enum ggml_cgraph_eval_order best_order = GGML_CGRAPH_EVAL_ORDER_COUNT;
+    // find best evaluation order
+    for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
+        ctx_compute = ggml_init(ctx_compute_params);
+        alloc = ggml_allocr_new_measure(tensor_alignment);
+        gf = ggml_new_graph(ctx_compute);
+        gf->order = (enum ggml_cgraph_eval_order) order;
+        gb = ggml_new_graph(ctx_compute);
+        gb_tmp = params.common.use_checkpointing
+            ? ggml_new_graph(ctx_compute)
+            : NULL;
+        loss = llama_build_lora_finetune_graphs(
+            &model, &lora, alloc, ctx_compute,
+            gf, gb, gb_tmp,
+            &logits, tokens_input, target_probs,
+            n_tokens, n_batch,
+            params.common.use_flash,
+            params.common.use_checkpointing
+        );
+        size_t max_compute_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+        if (max_compute_size < best_compute_size) {
+            best_compute_size = max_compute_size;
+            best_order = gf->order;
+        }
+        ggml_allocr_free(alloc);
+        ggml_free(ctx_compute);
+    }
+    size_t max_compute_size = best_compute_size;
+    printf("%s: compute_size = %zu bytes (%.1f MB)\n", __func__, max_compute_size, (float) max_compute_size / (1024.0f*1024.0f));
+    printf("%s: evaluation order = %s\n", __func__,
+        (best_order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? "LEFT_TO_RIGHT" :
+        (best_order == GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT) ? "RIGHT_TO_LEFT" :
+        "invalid");
+
+    // allocate compute tensors
+    mem_compute_data.resize(max_compute_size);
+    ctx_compute = ggml_init(ctx_compute_params);
+    alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
+    gf = ggml_new_graph(ctx_compute);
+    gf->order = best_order;
+    gb = ggml_new_graph(ctx_compute);
+    gb_tmp = params.common.use_checkpointing
+        ? ggml_new_graph(ctx_compute)
+        : NULL;
+    loss = llama_build_lora_finetune_graphs(
+        &model, &lora, alloc, ctx_compute,
+        gf, gb, gb_tmp,
+        &logits, tokens_input, target_probs,
+        n_tokens, n_batch,
+        params.common.use_flash,
+        params.common.use_checkpointing
+    );
+    ggml_allocr_free(alloc);
+
+    // tokenize data
+    std::vector<llama_token> train_tokens;
+    std::vector<size_t> train_samples_begin;
+    std::vector<size_t> train_samples_size;
+    printf("%s: tokenize training data\n", __func__);
+    tokenize_file(lctx,
+            params.common.fn_train_data,
+            params.common.sample_start,
+            params.common.include_sample_start,
+            params.common.overlapping_samples,
+            n_tokens,
+            train_tokens,
+            train_samples_begin,
+            train_samples_size);
+    GGML_ASSERT(train_samples_begin.size() == train_samples_size.size());
+
+    printf("%s: number of training tokens: %zu\n", __func__, train_tokens.size());
+
+    std::vector<size_t> token_noccurs;
+    token_noccurs.resize(model.hparams.n_vocab, 0);
+    for (unsigned int i = 0; i < train_tokens.size(); ++i) {
+        ++token_noccurs[train_tokens[i]];
+    }
+    int n_unique_tokens = 0;
+    for (unsigned int i = 0; i < token_noccurs.size(); ++i) {
+        if (token_noccurs[i] == 0) continue;
+        ++n_unique_tokens;
+    }
+    printf("%s: number of unique tokens: %d\n", __func__, n_unique_tokens);
+
+    size_t shuffle_samples_hash = compute_samples_hash(params.common.fn_train_data, train_samples_begin.data(), train_samples_size.data(), train_samples_size.size());
+    const bool changed_train_data = (shuffle_samples_hash != train->shuffle_samples_hash) || (train->shuffle_sample_count != train_samples_size.size());
+    if (changed_train_data) {
+        printf("%s: train data seems to have changed. restarting shuffled epoch.\n", __func__);
+    }
+    if (params.common.force_reshuffle) {
+        printf("%s: forced reshuffling of data. restarting with newly shuffled epoch.\n", __func__);
+    }
+    if ((train->shuffle_rng_state_current == "") || changed_train_data || params.common.force_reshuffle) {
+        train->shuffle_rng_state_current = mt19937_seed_to_state(params.common.seed);
+        train->shuffle_sample_count = train_samples_size.size();
+        train->shuffle_next_sample = 0;
+        train->shuffle_samples_hash = shuffle_samples_hash;
+    }
+    std::vector<size_t> train_shuffled_samples_offs;
+    std::vector<size_t> train_shuffled_samples_begin;
+    std::vector<size_t> train_shuffled_samples_size;
+    train_shuffled_samples_offs.resize(train_samples_begin.size());
+    train_shuffled_samples_begin.resize(train_samples_begin.size());
+    train_shuffled_samples_size.resize(train_samples_size.size());
+    train->shuffle_rng_state_next = shuffle_samples(
+        train->shuffle_rng_state_current,
+        train_shuffled_samples_offs.data(),
+        train_shuffled_samples_begin.data(),
+        train_shuffled_samples_size.data(),
+        train_samples_begin.data(),
+        train_samples_size.data(),
+        train_samples_size.size());
+
+    printf("%s: begin training\n", __func__);
+
+    save_train_files_data save_data;
+    save_data.fn_checkpoint_out = params.common.fn_checkpoint_out;
+    save_data.fn_lora_out       = params.fn_lora_out;
+    save_data.pattern_fn_it     = params.common.pattern_fn_it;
+    save_data.fn_latest         = params.common.fn_latest;
+    save_data.model             = &model;
+    save_data.lora              = &lora;
+
+    struct train_opt_callback_data opt_cb_data;
+    opt_cb_data.params                 = &params.common;
+    opt_cb_data.train                  = train;
+    opt_cb_data.save_cb                = &save_train_files;
+    opt_cb_data.save_data              = &save_data;
+    opt_cb_data.lctx                   = lctx;
+    opt_cb_data.last_save_iter         = opt->iter;
+    opt_cb_data.tokens_data            = train_tokens.data();
+    opt_cb_data.tokens_size            = train_tokens.size();
+    opt_cb_data.samples_begin          = train_samples_begin.data();
+    opt_cb_data.samples_size           = train_samples_size.data();
+    opt_cb_data.shuffled_samples_offs  = train_shuffled_samples_offs.data();
+    opt_cb_data.shuffled_samples_begin = train_shuffled_samples_begin.data();
+    opt_cb_data.shuffled_samples_size  = train_shuffled_samples_size.data();
+    opt_cb_data.samples_count          = train_samples_size.size();
+    opt_cb_data.tokens_input           = tokens_input;
+    opt_cb_data.target_probs           = target_probs;
+    opt_cb_data.first_iter             = opt->iter;
+    opt_cb_data.first_epoch            = train->train_epochs;
+    opt_cb_data.iter_at_last_epoch     = -1;
+    opt_cb_data.last_time              = ggml_time_ms();
+    opt_cb_data.millis_per_iter        = 0.0;
+
+    // measure required memory for work buffer
+    size_t max_work_size = ggml_graph_plan(gb, params.common.n_threads).work_size + GGML_OBJECT_SIZE;
+    printf("%s: work_size = %zu bytes (%.1f MB)\n", __func__, max_work_size, (float) max_work_size / (1024.0f*1024.0f));
+
+    // context for work buffer
+    struct ggml_init_params ctx_work_params = {
+        max_work_size, // mem_size
+        NULL,          // mem_buffer
+        false,         // no_alloc
+    };
+    struct ggml_context * ctx_work = ggml_init(ctx_work_params);
+
+    int64_t t0 = ggml_time_ms();
+
+    ggml_opt_resume_g(ctx_work, opt, loss, gf, gb, &train_opt_callback, (void *) &opt_cb_data);
+
+    ggml_free(ctx_work);
+    ggml_free(ctx_compute);
+    ggml_free(ctx_input);
+
+    int64_t t1 = ggml_time_ms();
+    printf("%s: total training time: ", __func__);
+    print_duration((double) (t1 - t0));
+    printf("\n");
+
+    int new_iters = opt->iter - opt_cb_data.last_save_iter;
+    if (new_iters > 0) {
+        train->train_its     += new_iters;
+        train->train_tokens  += new_iters * opt->params.n_gradient_accumulation * n_batch * n_tokens;
+
+        save_train_files(&save_data, train);
+        opt_cb_data.last_save_iter = opt->iter;
+    }
+
+    ggml_free(opt->ctx);
+    free_train_state(train);
+    ggml_free(lora.ctx);
+    llama_free(lctx);
+    llama_free_model(lmodel);
+    return 0;
+}

examples/gptneox-wip/falcon-main.cpp

@@ -367,10 +367,10 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         keyidx = gguf_find_key(ggufctx, "general.architecture");
         if (keyidx != -1) { printf("%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.file_type");
-        if (keyidx != -1) { printf("%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model file type      = %" PRIu32 "\n", __func__, gguf_get_val_u32(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "gptneox.tensor_data_layout");
         if (keyidx != -1) { printf("%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.source.hugginface.repository");
+        keyidx = gguf_find_key(ggufctx, "general.source.huggingface.repository");
         if (keyidx != -1) { printf("%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
     }
 

examples/gptneox-wip/gptneox-main.cpp

@@ -380,10 +380,10 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
         keyidx = gguf_find_key(ggufctx, "general.architecture");
         if (keyidx != -1) { printf("%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.file_type");
-        if (keyidx != -1) { printf("%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model file type      = %" PRIu32 "\n", __func__, gguf_get_val_u32(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "gptneox.tensor_data_layout");
         if (keyidx != -1) { printf("%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.source.hugginface.repository");
+        keyidx = gguf_find_key(ggufctx, "general.source.huggingface.repository");
         if (keyidx != -1) { printf("%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
     }
 

examples/llama-bench/README.md

@@ -0,0 +1,271 @@
+# llama.cpp/example/llama-bench
+
+Performance testing tool for llama.cpp.
+
+## Table of contents
+
+1. [Syntax](#syntax)
+2. [Examples](#examples)
+    1. [Text generation with different models](#text-generation-with-different-models)
+    2. [Prompt processing with different batch sizes](#prompt-processing-with-different-batch-sizes)
+    3. [Different numbers of threads](#different-numbers-of-threads)
+    4. [Different numbers of layers offloaded to the GPU](#different-numbers-of-layers-offloaded-to-the-gpu)
+3. [Output formats](#output-formats)
+    1. [Markdown](#markdown)
+    2. [CSV](#csv)
+    3. [JSON](#json)
+    4. [SQL](#sql)
+
+## Syntax
+
+```
+usage: ./llama-bench [options]
+
+options:
+  -h, --help
+  -m, --model <filename>            (default: models/7B/ggml-model-q4_0.gguf)
+  -p, --n-prompt <n>                (default: 512)
+  -n, --n-gen <n>                   (default: 128)
+  -b, --batch-size <n>              (default: 512)
+  --memory-f32 <0|1>                (default: 0)
+  -t, --threads <n>                 (default: 16)
+  -ngl N, --n-gpu-layers <n>        (default: 99)
+  -mg i, --main-gpu <i>             (default: 0)
+  -mmq, --mul-mat-q <0|1>           (default: 1)
+  -ts, --tensor_split <ts0/ts1/..>
+  -r, --repetitions <n>             (default: 5)
+  -o, --output <csv|json|md|sql>    (default: md)
+  -v, --verbose                     (default: 0)
+
+Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
+```
+
+llama-bench can perform two types of tests:
+
+- Prompt processing (pp): processing a prompt in batches (`-p`)
+- Text generation (tg): generating a sequence of tokens (`-n`)
+
+With the exception of `-r`, `-o` and `-v`, all options can be specified multiple times to run multiple tests. Each pp and tg test is run with all combinations of the specified options. To specify multiple values for an option, the values can be separated by commas (e.g. `-n 16,32`), or the option can be specified multiple times (e.g. `-n 16 -n 32`).
+
+Each test is repeated the number of times given by `-r`, and the results are averaged. The results are given in average tokens per second (t/s) and standard deviation. Some output formats (e.g. json) also include the individual results of each repetition.
+
+For a description of the other options, see the [main example](../main/README.md).
+
+## Examples
+
+### Text generation with different models
+
+```sh
+$ ./llama-bench -m models/7B/ggml-model-q4_0.gguf -m models/13B/ggml-model-q4_0.gguf -p 0 -n 128,256,512
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    132.19 ± 0.55 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 256     |    129.37 ± 0.54 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 512     |    123.83 ± 0.25 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 128     |     82.17 ± 0.31 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 256     |     80.74 ± 0.23 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 512     |     78.08 ± 0.07 |
+
+### Prompt processing with different batch sizes
+
+```sh
+$ ./llama-bench -n 0 -p 1024 -b 128,256,512,1024
+```
+
+| model                          |       size |     params | backend    | ngl |    n_batch | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        128 | pp 1024    |   1436.51 ± 3.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        256 | pp 1024    |  1932.43 ± 23.48 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        512 | pp 1024    |  2254.45 ± 15.59 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |       1024 | pp 1024    |  2498.61 ± 13.58 |
+
+### Different numbers of threads
+
+```sh
+$ ./llama-bench -n 0 -n 16 -p 64 -t 1,2,4,8,16,32
+```
+
+| model                          |       size |     params | backend    |    threads | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | pp 64      |      6.17 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | tg 16      |      4.05 ± 0.02 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | pp 64      |     12.31 ± 0.13 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | tg 16      |      7.80 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | pp 64      |     23.18 ± 0.06 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | tg 16      |     12.22 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | pp 64      |     32.29 ± 1.21 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | tg 16      |     16.71 ± 0.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | pp 64      |     33.52 ± 0.03 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | tg 16      |     15.32 ± 0.05 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | pp 64      |     59.00 ± 1.11 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | tg 16      |     16.41 ± 0.79 ||
+
+### Different numbers of layers offloaded to the GPU
+
+```sh
+$ ./llama-bench -ngl 10,20,30,31,32,33,34,35
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | pp 512     |    373.36 ± 2.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | tg 128     |     13.45 ± 0.93 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | pp 512     |    472.65 ± 1.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | tg 128     |     21.36 ± 1.94 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | pp 512     |   631.87 ± 11.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | tg 128     |     40.04 ± 1.82 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | pp 512     |    657.89 ± 5.08 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | tg 128     |     48.19 ± 0.81 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | pp 512     |    688.26 ± 3.29 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | tg 128     |     54.78 ± 0.65 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | pp 512     |    704.27 ± 2.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | tg 128     |     60.62 ± 1.76 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | pp 512     |    881.34 ± 5.40 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | tg 128     |     71.76 ± 0.23 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | pp 512     |   2400.01 ± 7.72 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | tg 128     |    131.66 ± 0.49 |
+
+## Output formats
+
+By default, llama-bench outputs the results in markdown format. The results can be output in other formats by using the `-o` option.
+
+### Markdown
+
+```sh
+$ ./llama-bench -o md
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | pp 512     |  2368.80 ± 93.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    131.42 ± 0.59 |
+
+### CSV
+
+```sh
+$ ./llama-bench -o csv
+```
+
+```csv
+build_commit,build_number,cuda,opencl,metal,gpu_blas,blas,cpu_info,gpu_info,model_filename,model_type,model_size,model_n_params,n_batch,n_threads,f16_kv,n_gpu_layers,main_gpu,mul_mat_q,tensor_split,n_prompt,n_gen,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","512","0","2023-09-23T12:09:01Z","212155977","732372","2413.341687","8.305961"
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","0","128","2023-09-23T12:09:02Z","969320879","2728399","132.052051","0.371342"
+```
+
+### JSON
+
+```sh
+$ ./llama-bench -o json
+```
+
+```json
+[
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "opencl": false,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 512,
+    "n_gen": 0,
+    "test_time": "2023-09-23T12:09:57Z",
+    "avg_ns": 212365953,
+    "stddev_ns": 985423,
+    "avg_ts": 2410.974041,
+    "stddev_ts": 11.163766,
+    "samples_ns": [ 213837238, 211635853, 212328053, 211329715, 212698907 ],
+    "samples_ts": [ 2394.34, 2419.25, 2411.36, 2422.75, 2407.16 ]
+  },
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "opencl": false,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 0,
+    "n_gen": 128,
+    "test_time": "2023-09-23T12:09:59Z",
+    "avg_ns": 977425219,
+    "stddev_ns": 9268593,
+    "avg_ts": 130.965708,
+    "stddev_ts": 1.238924,
+    "samples_ns": [ 984472709, 974901233, 989474741, 970729355, 967548060 ],
+    "samples_ts": [ 130.019, 131.295, 129.362, 131.86, 132.293 ]
+  }
+]
+```
+
+### SQL
+
+SQL output is suitable for importing into a SQLite database. The output can be piped into the `sqlite3` command line tool to add the results to a database.
+
+```sh
+$ ./llama-bench -o sql
+```
+
+```sql
+CREATE TABLE IF NOT EXISTS test (
+  build_commit TEXT,
+  build_number INTEGER,
+  cuda INTEGER,
+  opencl INTEGER,
+  metal INTEGER,
+  gpu_blas INTEGER,
+  blas INTEGER,
+  cpu_info TEXT,
+  gpu_info TEXT,
+  model_filename TEXT,
+  model_type TEXT,
+  model_size INTEGER,
+  model_n_params INTEGER,
+  n_batch INTEGER,
+  n_threads INTEGER,
+  f16_kv INTEGER,
+  n_gpu_layers INTEGER,
+  main_gpu INTEGER,
+  mul_mat_q INTEGER,
+  tensor_split TEXT,
+  n_prompt INTEGER,
+  n_gen INTEGER,
+  test_time TEXT,
+  avg_ns INTEGER,
+  stddev_ns INTEGER,
+  avg_ts REAL,
+  stddev_ts REAL
+);
+
+INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '512', '0', '2023-09-23T12:10:30Z', '212693772', '743623', '2407.240204', '8.409634');
+INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '0', '128', '2023-09-23T12:10:31Z', '977925003', '4037361', '130.891159', '0.537692');
+```

examples/llama-bench/llama-bench.cpp

@@ -132,7 +132,6 @@ struct cmd_params {
     std::vector<int> n_gpu_layers;
     std::vector<int> main_gpu;
     std::vector<bool> mul_mat_q;
-    std::vector<bool> low_vram;
     std::vector<std::array<float, LLAMA_MAX_DEVICES>> tensor_split;
     int reps;
     bool verbose;
@@ -149,7 +148,6 @@ static const cmd_params cmd_params_defaults = {
     /* n_gpu_layers  */ {99},
     /* main_gpu      */ {0},
     /* mul_mat_q     */ {true},
-    /* low_vram      */ {false},
     /* tensor_split  */ {{}},
     /* reps          */ 5,
     /* verbose       */ false,
@@ -167,9 +165,8 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
     printf("  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
     printf("  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
-    printf("  -ngl N, --n-gpu-layers <n>        (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
-    printf("  -mg i, --main-gpu <n>             (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
-    printf("  -lv, --low-vram <0|1>             (default: %s)\n", join(cmd_params_defaults.low_vram, ",").c_str());
+    printf("  -ngl, --n-gpu-layers <n>          (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    printf("  -mg, --main-gpu <i>               (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
     printf("  -mmq, --mul-mat-q <0|1>           (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
     printf("  -ts, --tensor_split <ts0/ts1/..>               \n");
     printf("  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
@@ -255,13 +252,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 break;
             }
             params.main_gpu = split<int>(argv[i], split_delim);
-        } else if (arg == "-lv" || arg == "--low-vram") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = split<bool>(argv[i], split_delim);
-            params.low_vram.insert(params.low_vram.end(), p.begin(), p.end());
         } else if (arg == "-mmq" || arg == "--mul-mat-q") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -336,7 +326,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
     if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
-    if (params.low_vram.empty())     { params.low_vram = cmd_params_defaults.low_vram; }
     if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
     if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
 
@@ -353,21 +342,34 @@ struct cmd_params_instance {
     int n_gpu_layers;
     int main_gpu;
     bool mul_mat_q;
-    bool low_vram;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
 
-    llama_context_params to_llama_params() const {
-        llama_context_params lparams = llama_context_default_params();
-        lparams.n_ctx = n_prompt + n_gen;
-        lparams.n_batch = n_batch;
-        lparams.f16_kv = !f32_kv;
-        lparams.n_gpu_layers = n_gpu_layers;
-        lparams.main_gpu = main_gpu;
-        lparams.mul_mat_q = mul_mat_q;
-        lparams.low_vram = low_vram;
-        lparams.tensor_split = tensor_split.data();
+    llama_model_params to_llama_mparams() const {
+        llama_model_params mparams = llama_model_default_params();
+
+        mparams.n_gpu_layers = n_gpu_layers;
+        mparams.main_gpu = main_gpu;
+        mparams.tensor_split = tensor_split.data();
+
+        return mparams;
+    }
+
+    bool equal_mparams(const cmd_params_instance & other) const {
+        return model == other.model &&
+               n_gpu_layers == other.n_gpu_layers &&
+               main_gpu == other.main_gpu &&
+               tensor_split == other.tensor_split;
+    }
+
+    llama_context_params to_llama_cparams() const {
+        llama_context_params cparams = llama_context_default_params();
+
+        cparams.n_ctx = n_prompt + n_gen;
+        cparams.n_batch = n_batch;
+        cparams.f16_kv = !f32_kv;
+        cparams.mul_mat_q = mul_mat_q;
 
-        return lparams;
+        return cparams;
     }
 };
 
@@ -375,13 +377,12 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
     std::vector<cmd_params_instance> instances;
 
     for (const auto & m : params.model)
-    for (const auto & nb : params.n_batch)
-    for (const auto & fk : params.f32_kv)
     for (const auto & nl : params.n_gpu_layers)
     for (const auto & mg : params.main_gpu)
-    for (const auto & mmq : params.mul_mat_q)
-    for (const auto & lv : params.low_vram)
     for (const auto & ts : params.tensor_split)
+    for (const auto & nb : params.n_batch)
+    for (const auto & fk : params.f32_kv)
+    for (const auto & mmq : params.mul_mat_q)
     for (const auto & nt : params.n_threads) {
         cmd_params_instance instance = {
             /* .model        = */ m,
@@ -393,7 +394,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
             /* .n_gpu_layers = */ nl,
             /* .main_gpu     = */ mg,
             /* .mul_mat_q    = */ mmq,
-            /* .low_vram     = */ lv,
             /* .tensor_split = */ ts,
         };
         instances.push_back(instance);
@@ -404,6 +404,56 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
 static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_params & params) {
     std::vector<cmd_params_instance> instances;
 
+#if 1
+    // this ordering minimizes the number of times that each model needs to be reloaded
+    for (const auto & m : params.model)
+    for (const auto & nl : params.n_gpu_layers)
+    for (const auto & mg : params.main_gpu)
+    for (const auto & ts : params.tensor_split)
+    for (const auto & nb : params.n_batch)
+    for (const auto & fk : params.f32_kv)
+    for (const auto & mmq : params.mul_mat_q)
+    for (const auto & nt : params.n_threads) {
+        for (const auto & n_prompt : params.n_prompt) {
+            if (n_prompt == 0) {
+                continue;
+            }
+            cmd_params_instance instance = {
+                /* .model        = */ m,
+                /* .n_prompt     = */ n_prompt,
+                /* .n_gen        = */ 0,
+                /* .n_batch      = */ nb,
+                /* .f32_kv       = */ fk,
+                /* .n_threads    = */ nt,
+                /* .n_gpu_layers = */ nl,
+                /* .main_gpu     = */ mg,
+                /* .mul_mat_q    = */ mmq,
+                /* .tensor_split = */ ts,
+            };
+            instances.push_back(instance);
+        }
+
+        for (const auto & n_gen : params.n_gen) {
+            if (n_gen == 0) {
+                continue;
+            }
+            cmd_params_instance instance = {
+                /* .model        = */ m,
+                /* .n_prompt     = */ 0,
+                /* .n_gen        = */ n_gen,
+                /* .n_batch      = */ nb,
+                /* .f32_kv       = */ fk,
+                /* .n_threads    = */ nt,
+                /* .n_gpu_layers = */ nl,
+                /* .main_gpu     = */ mg,
+                /* .mul_mat_q    = */ mmq,
+                /* .tensor_split = */ ts,
+            };
+            instances.push_back(instance);
+        }
+    }
+#else
+    // this ordering separates the prompt and generation tests
     for (const auto & n_prompt : params.n_prompt) {
         if (n_prompt == 0) {
             continue;
@@ -419,6 +469,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
         auto instances_gen = get_cmd_params_instances_int(params, n_gen, 0);
         instances.insert(instances.end(), instances_gen.begin(), instances_gen.end());
     }
+#endif
 
     return instances;
 }
@@ -443,7 +494,6 @@ struct test {
     int n_gpu_layers;
     int main_gpu;
     bool mul_mat_q;
-    bool low_vram;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
     int n_prompt;
     int n_gen;
@@ -463,7 +513,6 @@ struct test {
         n_gpu_layers = inst.n_gpu_layers;
         main_gpu = inst.main_gpu;
         mul_mat_q = inst.mul_mat_q;
-        low_vram = inst.low_vram;
         tensor_split = inst.tensor_split;
         n_prompt = inst.n_prompt;
         n_gen = inst.n_gen;
@@ -524,7 +573,7 @@ struct test {
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_threads", "f16_kv",
-            "n_gpu_layers", "main_gpu", "mul_mat_q", "low_vram", "tensor_split",
+            "n_gpu_layers", "main_gpu", "mul_mat_q", "tensor_split",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -543,7 +592,7 @@ struct test {
             return INT;
         }
         if (field == "cuda" || field == "opencl" || field == "metal" || field == "gpu_blas" || field == "blas" ||
-            field == "f16_kv" || field == "mul_mat_q" || field == "low_vram") {
+            field == "f16_kv" || field == "mul_mat_q") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -574,7 +623,7 @@ struct test {
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
-            std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), std::to_string(low_vram), tensor_split_str,
+            std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), tensor_split_str,
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
             std::to_string(avg_ts()), std::to_string(stdev_ts())
@@ -606,9 +655,9 @@ struct printer {
     virtual ~printer() {}
 
     FILE * fout;
-    virtual void print_header(const cmd_params & params) { (void) params; };
+    virtual void print_header(const cmd_params & params) { (void) params; }
     virtual void print_test(const test & t) = 0;
-    virtual void print_footer() { };
+    virtual void print_footer() { }
 };
 
 struct csv_printer : public printer {
@@ -766,9 +815,6 @@ struct markdown_printer : public printer {
         if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
             fields.push_back("mul_mat_q");
         }
-        if (params.low_vram.size() > 1 || params.low_vram != cmd_params_defaults.low_vram) {
-            fields.push_back("low_vram");
-        }
         if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
             fields.push_back("tensor_split");
         }
@@ -889,21 +935,27 @@ 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(ctx));
     int n_processed = 0;
+
+    llama_set_n_threads(ctx, n_threads, n_threads);
+
     while (n_processed < n_prompt) {
         int n_tokens = std::min(n_prompt - n_processed, n_batch);
-        llama_eval(ctx, tokens.data(), n_tokens, n_past + n_processed, n_threads);
+        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
         n_processed += n_tokens;
     }
 }
 
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
     llama_token token = llama_token_bos(ctx);
+
+    llama_set_n_threads(ctx, n_threads, n_threads);
+
     for (int i = 0; i < n_gen; i++) {
-        llama_eval(ctx, &token, 1, n_past + i, n_threads);
+        llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
     }
 }
 
-static void llama_null_log_callback(enum llama_log_level level, const char * text, void * user_data) {
+static void llama_null_log_callback(enum ggml_log_level level, const char * text, void * user_data) {
     (void) level;
     (void) text;
     (void) user_data;
@@ -958,17 +1010,25 @@ int main(int argc, char ** argv) {
 
     std::vector<cmd_params_instance> params_instances = get_cmd_params_instances(params);
 
+    llama_model * lmodel = nullptr;
+    const cmd_params_instance * prev_inst = nullptr;
+
     for (const auto & inst : params_instances) {
-        // TODO: keep the model between tests when possible
-        llama_context_params lparams = inst.to_llama_params();
+        // keep the same model between tests when possible
+        if (!lmodel || !prev_inst || !inst.equal_mparams(*prev_inst)) {
+            if (lmodel) {
+                llama_free_model(lmodel);
+            }
 
-        llama_model * lmodel  = llama_load_model_from_file(inst.model.c_str(), lparams);
-        if (lmodel == NULL) {
-            fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, inst.model.c_str());
-            return 1;
+            lmodel = llama_load_model_from_file(inst.model.c_str(), inst.to_llama_mparams());
+            if (lmodel == NULL) {
+                fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, inst.model.c_str());
+                return 1;
+            }
+            prev_inst = &inst;
         }
 
-        llama_context * ctx = llama_new_context_with_model(lmodel, lparams);
+        llama_context * ctx = llama_new_context_with_model(lmodel, inst.to_llama_cparams());
         if (ctx == NULL) {
             fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, inst.model.c_str());
             llama_free_model(lmodel);
@@ -977,6 +1037,8 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
         // warmup run
         if (t.n_prompt > 0) {
             test_prompt(ctx, std::min(2, t.n_batch), 0, t.n_batch, t.n_threads);
@@ -986,6 +1048,8 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
+            llama_kv_cache_tokens_rm(ctx, -1, -1);
+
             uint64_t t_start = get_time_ns();
             if (t.n_prompt > 0) {
                 test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
@@ -1002,9 +1066,10 @@ int main(int argc, char ** argv) {
         llama_print_timings(ctx);
 
         llama_free(ctx);
-        llama_free_model(lmodel);
     }
 
+    llama_free_model(lmodel);
+
     p->print_footer();
 
     llama_backend_free();

examples/main/README.md

@@ -262,7 +262,8 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### Number of Threads
 
--   `-t N, --threads N`: Set the number of threads to use during computation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). Using the correct number of threads can greatly improve performance.
+-   `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). Using the correct number of threads can greatly improve performance.
+-   `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. In some systems, it is beneficial to use a higher number of threads during batch processing than during generation. If not specified, the number of threads used for batch processing will be the same as the number of threads used for generation.
 
 ### Mlock
 
@@ -305,6 +306,5 @@ These options provide extra functionality and customization when running the LLa
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
--   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.

examples/main/main.cpp

@@ -124,7 +124,7 @@ int main(int argc, char ** argv) {
     console::init(params.simple_io, params.use_color);
     atexit([]() { console::cleanup(); });
 
-    if (params.perplexity) {
+    if (params.logits_all) {
         printf("\n************\n");
         printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
         printf("************\n\n");
@@ -140,12 +140,17 @@ int main(int argc, char ** argv) {
         return 0;
     }
 
-    if (params.rope_freq_base != 10000.0) {
-        LOG_TEE("%s: warning: changing RoPE frequency base to %g (default 10000.0)\n", __func__, params.rope_freq_base);
+    if (params.n_ctx != 0 && params.n_ctx < 8) {
+        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
+    }
+
+    if (params.rope_freq_base != 0.0) {
+        LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
     }
 
-    if (params.rope_freq_scale != 1.0) {
-        LOG_TEE("%s: warning: scaling RoPE frequency by %g (default 1.0)\n", __func__, params.rope_freq_scale);
+    if (params.rope_freq_scale != 0.0) {
+        LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
     LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
@@ -184,29 +189,19 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    const int n_ctx_train = llama_n_ctx_train(ctx);
-    if (params.n_ctx > n_ctx_train) {
+    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx = llama_n_ctx(ctx);
+    LOG("n_ctx: %d\n", n_ctx);
+
+    if (n_ctx > n_ctx_train) {
         LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
-                __func__, n_ctx_train, params.n_ctx);
-    } else if (params.n_ctx < 8) {
-        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
-        params.n_ctx = 8;
+                __func__, n_ctx_train, n_ctx);
     }
 
     // print system information
     {
         LOG_TEE("\n");
-        LOG_TEE("system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
-    }
-
-    // export the cgraph and exit
-    if (params.export_cgraph) {
-        llama_eval_export(ctx, "llama.ggml");
-        llama_free(ctx);
-        llama_free_model(model);
-
-        return 0;
+        LOG_TEE("%s\n", get_system_info(params).c_str());
     }
 
     std::string path_session = params.path_prompt_cache;
@@ -220,7 +215,7 @@ int main(int argc, char ** argv) {
         if (fp != NULL) {
             std::fclose(fp);
 
-            session_tokens.resize(params.n_ctx);
+            session_tokens.resize(n_ctx);
             size_t n_token_count_out = 0;
             if (!llama_load_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
                 LOG_TEE("%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
@@ -235,7 +230,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const bool add_bos = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
     LOG("add_bos: %d\n", add_bos);
 
     std::vector<llama_token> embd_inp;
@@ -276,9 +271,6 @@ int main(int argc, char ** argv) {
         LOG("guidance_offset:     %s", log_tostr(guidance_offset));
     }
 
-    const int n_ctx = llama_n_ctx(ctx);
-    LOG("n_ctx: %d\n", n_ctx);
-
     if ((int) embd_inp.size() > n_ctx - 4) {
         LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
         return 1;
@@ -475,7 +467,7 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    const int n_vocab = llama_n_vocab(ctx);
+    const int n_vocab = llama_n_vocab(model);
 
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
@@ -508,17 +500,22 @@ int main(int argc, char ** argv) {
                     break;
                 }
 
-                const int n_left = n_past - params.n_keep;
-                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d\n", n_past, n_left, n_ctx, params.n_keep);
+                const int n_left    = n_past - params.n_keep - 1;
+                const int n_discard = n_left/2;
 
-                // always keep the first token - BOS
-                n_past          = std::max(1, params.n_keep);
-                n_past_guidance = std::max(1, params.n_keep + guidance_offset);
+                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                    n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                n_past -= n_discard;
 
-                // insert n_left/2 tokens at the start of embd from last_tokens
-                embd.insert(embd.begin(), last_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_tokens.end() - embd.size());
+                if (ctx_guidance) {
+                    n_past_guidance -= n_discard;
+                }
+
+                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));
 
@@ -580,7 +577,7 @@ int main(int argc, char ** argv) {
 
                 for (int i = 0; i < input_size; i += params.n_batch) {
                     int n_eval = std::min(input_size - i, params.n_batch);
-                    if (llama_eval(ctx_guidance, input_buf + i, n_eval, n_past_guidance, params.n_threads)) {
+                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
                         LOG_TEE("%s : failed to eval\n", __func__);
                         return 1;
                     }
@@ -597,7 +594,7 @@ int main(int argc, char ** argv) {
 
                 LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
 
-                if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) {
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
                     LOG_TEE("%s : failed to eval\n", __func__);
                     return 1;
                 }
@@ -855,7 +852,7 @@ int main(int argc, char ** argv) {
     llama_backend_free();
 
 #ifndef LOG_DISABLE_LOGS
-    LOG_TEE("Log end\n")
+    LOG_TEE("Log end\n");
 #endif // LOG_DISABLE_LOGS
 
     return 0;

examples/make-ggml.py

@@ -1,22 +1,25 @@
 #!/usr/bin/env python3
 """
-This script converts Hugging Face llama models to GGML and quantizes them.
+This script converts Hugging Face Llama, StarCoder, Falcon, Baichuan, and GPT-NeoX models to GGUF and quantizes them.
 
 Usage:
-python make-ggml.py --model {model_dir_or_hf_repo_name} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]
+python make-ggml.py {model_dir_or_hf_repo_name} --model_type {model_type} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]
 
 Arguments:
-- --model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
+- model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
+- --model_type: (Required) The type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.
 - --outname: (Optional) The name of the output model. If not specified, the last part of the model directory path or the Hugging Face model repo name will be used.
 - --outdir: (Optional) The directory where the output model(s) will be stored. If not specified, '../models/{outname}' will be used.
 - --quants: (Optional) The types of quantization to apply. This should be a space-separated list. The default is 'Q4_K_M Q5_K_S'.
 - --keep_fp16: (Optional) If specified, the FP16 model will not be deleted after the quantized models are created.
 
-Quant types:
+Old quant types (some base model types require these):
 - Q4_0: small, very high quality loss - legacy, prefer using Q3_K_M
 - Q4_1: small, substantial quality loss - legacy, prefer using Q3_K_L
 - Q5_0: medium, balanced quality - legacy, prefer using Q4_K_M
 - Q5_1: medium, low quality loss - legacy, prefer using Q5_K_M
+
+New quant types (recommended):
 - Q2_K: smallest, extreme quality loss - not recommended
 - Q3_K: alias for Q3_K_M
 - Q3_K_S: very small, very high quality loss
@@ -40,9 +43,7 @@ import argparse
 import os
 from huggingface_hub import snapshot_download
 
-def main(model, outname, outdir, quants, keep_fp16):
-    ggml_version = "v3"
-
+def main(model, model_type, outname, outdir, quants, keep_fp16):
     if not os.path.isdir(model):
         print(f"Model not found at {model}. Downloading...")
         try:
@@ -63,17 +64,20 @@ def main(model, outname, outdir, quants, keep_fp16):
     print("Building llama.cpp")
     subprocess.run(f"cd .. && make quantize", shell=True, check=True)
 
-    fp16 = f"{outdir}/{outname}.ggml{ggml_version}.fp16.bin"
+    fp16 = f"{outdir}/{outname}.gguf.fp16.bin"
 
-    print(f"Making unquantised GGML at {fp16}")
+    print(f"Making unquantised GGUF at {fp16}")
     if not os.path.isfile(fp16):
-        subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
+        if model_type != "llama":
+            subprocess.run(f"python3 ../convert-{model_type}-hf-to-gguf.py {model} 1 --outfile {fp16}", shell=True, check=True)
+        else:
+            subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
     else:
         print(f"Unquantised GGML already exists at: {fp16}")
 
     print("Making quants")
     for type in quants:
-        outfile = f"{outdir}/{outname}.ggml{ggml_version}.{type}.bin"
+        outfile = f"{outdir}/{outname}.gguf.{type}.bin"
         print(f"Making {type} : {outfile}")
         subprocess.run(f"../quantize {fp16} {outfile} {type}", shell=True, check=True)
 
@@ -81,8 +85,9 @@ def main(model, outname, outdir, quants, keep_fp16):
         os.remove(fp16)
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description='Convert/Quantize HF to GGML. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
-    parser.add_argument('--model', required=True, help='Downloaded model dir or Hugging Face model repo name')
+    parser = argparse.ArgumentParser(description='Convert/Quantize HF models to GGUF. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
+    parser.add_argument('model', help='Downloaded model dir or Hugging Face model repo name')
+    parser.add_argument('--model_type', required=True, choices=['llama', 'starcoder', 'falcon', 'baichuan', 'gptneox'], help='Type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.')
     parser.add_argument('--outname', default=None, help='Output model(s) name')
     parser.add_argument('--outdir', default=None, help='Output directory')
     parser.add_argument('--quants', nargs='*', default=["Q4_K_M", "Q5_K_S"], help='Quant types')
@@ -90,4 +95,4 @@ if __name__ == "__main__":
 
     args = parser.parse_args()
 
-    main(args.model, args.outname, args.outdir, args.quants, args.keep_fp16)
+    main(args.model, args.model_type, args.outname, args.outdir, args.quants, args.keep_fp16)

examples/parallel/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET parallel)
+add_executable(${TARGET} parallel.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/parallel/README.md

@@ -0,0 +1,3 @@
+# llama.cpp/example/parallel
+
+Simplified simluation for serving incoming requests in parallel

examples/parallel/parallel.cpp

@@ -0,0 +1,380 @@
+// A basic application simulating a server with multiple clients.
+// The clients submite requests to the server and they are processed in parallel.
+
+#include "build-info.h"
+
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+// trim whitespace from the beginning and end of a string
+static std::string trim(const std::string & str) {
+    size_t start = 0;
+    size_t end = str.size();
+
+    while (start < end && isspace(str[start])) {
+        start += 1;
+    }
+
+    while (end > start && isspace(str[end - 1])) {
+        end -= 1;
+    }
+
+    return str.substr(start, end - start);
+}
+
+static std::string k_system =
+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:)";
+
+static std::vector<std::string> k_prompts = {
+    "What is the meaning of life?",
+    "Tell me an interesting fact about llamas.",
+    "What is the best way to cook a steak?",
+    "Are you familiar with the Special Theory of Relativity and can you explain it to me?",
+    "Recommend some interesting books to read.",
+    "What is the best way to learn a new language?",
+    "How to get a job at Google?",
+    "If you could have any superpower, what would it be?",
+    "I want to learn how to play the piano.",
+};
+
+struct client {
+    int32_t id = 0;
+
+    llama_seq_id seq_id = -1;
+
+    llama_token sampled;
+
+    int64_t t_start_prompt;
+    int64_t t_start_gen;
+
+    int32_t n_prompt  = 0;
+    int32_t n_decoded = 0;
+    int32_t i_batch   = -1;
+
+    std::string input;
+    std::string prompt;
+    std::string response;
+
+    std::vector<llama_token> tokens_prev;
+};
+
+int main(int argc, char ** argv) {
+    srand(1234);
+
+    gpt_params params;
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    // number of simultaneous "clients" to simulate
+    const int32_t n_clients = params.n_parallel;
+
+    // requests to simulate
+    const int32_t n_seq = params.n_sequences;
+
+    // insert new requests as soon as the previous one is done
+    const bool cont_batching = params.cont_batching;
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("parallel", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // init llama.cpp
+    llama_backend_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the target model
+    params.logits_all = true;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+
+    fprintf(stderr, "\n\n");
+    fflush(stderr);
+
+    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.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();
+
+    llama_seq_id g_seq_id = 0;
+
+    // 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(params.n_ctx, 0);
+
+    int32_t n_total_prompt = 0;
+    int32_t n_total_gen    = 0;
+    int32_t n_cache_miss   = 0;
+
+    const auto t_main_start = ggml_time_us();
+
+    LOG_TEE("%s: Simulating parallel requests from clients:\n", __func__);
+    LOG_TEE("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
+    LOG_TEE("\n");
+
+    {
+        LOG_TEE("%s: Evaluating the system prompt ...\n", __func__);
+
+        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) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+
+        // assign the system KV cache to all parallel sequences
+        for (int32_t i = 1; i < n_clients; ++i) {
+            llama_kv_cache_seq_cp(ctx, 0, i, 0, n_tokens_system);
+        }
+
+        LOG_TEE("\n");
+    }
+
+    LOG_TEE("Processing requests ...\n\n");
+
+    while (true) {
+        batch.n_tokens = 0;
+
+        // decode any currently ongoing sequences
+        for (auto & client : clients) {
+            if (client.seq_id == -1) {
+                continue;
+            }
+
+            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) {
+            // all sequences have ended - clear the entire KV cache
+            for (int i = 0; i < n_clients; ++i) {
+                llama_kv_cache_seq_rm(ctx, i, n_tokens_system, -1);
+            }
+
+            LOG_TEE("%s: clearing the KV cache\n", __func__);
+        }
+
+        // insert new sequences for decoding
+        if (cont_batching || batch.n_tokens == 0) {
+            for (auto & client : clients) {
+                if (client.seq_id == -1 && g_seq_id < n_seq) {
+                    client.seq_id = g_seq_id;
+
+                    client.t_start_prompt = ggml_time_us();
+                    client.t_start_gen    = 0;
+
+                    client.input    = k_prompts[rand() % k_prompts.size()];
+                    client.prompt   = client.input + "\nAssistant:";
+                    client.response = "";
+
+                    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) {
+                        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
+                    if (batch.n_tokens > 0) {
+                        batch.logits[batch.n_tokens - 1] = true;
+                    }
+
+                    client.n_prompt  = tokens_prompt.size();
+                    client.n_decoded = 0;
+                    client.i_batch   = batch.n_tokens - 1;
+
+                    LOG_TEE("\033[1mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
+
+                    g_seq_id += 1;
+
+                    // insert new requests one-by-one
+                    //if (cont_batching) {
+                    //    break;
+                    //}
+                }
+            }
+        }
+
+        if (batch.n_tokens == 0) {
+            break;
+        }
+
+        // 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 1;
+                }
+
+                LOG("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
+
+                n_cache_miss += 1;
+
+                // retry with half the batch size to try to find a free slot in the KV cache
+                n_batch /= 2;
+                i -= n_batch;
+
+                continue;
+            }
+
+            LOG("%s : decoded batch of %d tokens\n", __func__, n_tokens);
+
+            for (auto & client : clients) {
+                if (client.i_batch < (int) i || client.i_batch >= (int) (i + n_tokens)) {
+                    continue;
+                }
+
+                //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_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
+                    // have their prompt already processed
+                    client.t_start_gen = ggml_time_us();
+                }
+
+                // 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;
+
+                //printf("client %d, seq %d, token %d, pos %d, batch %d: %s\n",
+                //        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(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)) {
+                    // basic reverse prompt
+                    const size_t pos = client.response.find("User:");
+                    if (pos != std::string::npos) {
+                        client.response = client.response.substr(0, pos);
+                    }
+
+                    // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
+                    llama_kv_cache_seq_rm(ctx, client.id, n_tokens_system, n_ctx);
+
+                    const auto t_main_end = ggml_time_us();
+
+                    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,
+                            ::trim(client.input).c_str(),
+                            ::trim(client.response).c_str());
+
+                    n_total_prompt += client.n_prompt;
+                    n_total_gen    += client.n_decoded;
+
+                    client.seq_id = -1;
+                }
+
+                client.i_batch = -1;
+            }
+        }
+    }
+
+    const auto t_main_end = ggml_time_us();
+
+    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\n");
+
+    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/perplexity/README.md

@@ -1,3 +1,21 @@
 # perplexity
 
 TODO
+
+## Llama 2 70B Scorechart
+Quantization | Model size (GiB) | Perplexity | Delta to fp16
+-- | -- | -- | --
+Q4_0 | 36.20 | 3.5550 | 3.61%
+Q4_1 | 40.20 | 3.5125 | 2.37%
+Q5_0 | 44.20 | 3.4744 | 1.26%
+Q2_K | 27.27 | 3.7339 | 8.82%
+Q3_K_S | 27.86 | 3.7019 | 7.89%
+Q3_K_M | 30.83 | 3.5932 | 4.72%
+Q3_K_L | 33.67 | 3.5617 | 3.80%
+Q4_K_S | 36.39 | 3.4852 | 1.57%
+Q4_K_M | 38.54 | 3.4725 | 1.20%
+Q5_K_S | 44.20 | 3.4483 | 0.50%
+Q5_K_M | 45.41 | 3.4451 | 0.40%
+Q6_K | 52.70 | 3.4367 | 0.16%
+fp16 | 128.5 | 3.4313 | -
+

examples/perplexity/perplexity.cpp

@@ -80,7 +80,9 @@ static void write_logfile(
 static std::vector<float> softmax(const std::vector<float>& logits) {
     std::vector<float> probs(logits.size());
     float max_logit = logits[0];
-    for (float v : logits) max_logit = std::max(max_logit, v);
+    for (float v : logits) {
+        max_logit = std::max(max_logit, v);
+    }
     double sum_exp = 0.0;
     for (size_t i = 0; i < logits.size(); i++) {
         // Subtract the maximum logit value from the current logit value for numerical stability
@@ -89,15 +91,21 @@ static std::vector<float> softmax(const std::vector<float>& logits) {
         sum_exp += exp_logit;
         probs[i] = exp_logit;
     }
-    for (size_t i = 0; i < probs.size(); i++) probs[i] /= sum_exp;
+    for (size_t i = 0; i < probs.size(); i++) {
+        probs[i] /= sum_exp;
+    }
     return probs;
 }
 
 static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
     float max_logit = logits[0];
-    for (int i = 1; i < n_vocab; ++i) max_logit = std::max(max_logit, logits[i]);
+    for (int i = 1; i < n_vocab; ++i) {
+        max_logit = std::max(max_logit, logits[i]);
+    }
     double sum_exp = 0.0;
-    for (int i = 0; i < n_vocab; ++i) sum_exp += expf(logits[i] - max_logit);
+    for (int i = 0; i < n_vocab; ++i) {
+        sum_exp += expf(logits[i] - max_logit);
+    }
     return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
 }
 
@@ -108,7 +116,8 @@ static void process_logits(
     std::mutex mutex;
     int counter = 0;
     auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
-        double local_nll = 0, local_nll2 = 0;
+        double local_nll  = 0;
+        double local_nll2 = 0;
         while (true) {
             std::unique_lock<std::mutex> lock(mutex);
             int i = counter++;
@@ -126,10 +135,13 @@ static void process_logits(
             prob_history[i]  = results.prob;
         }
     };
-    for (auto & w : workers) w = std::thread(compute);
+    for (auto & w : workers) {
+        w = std::thread(compute);
+    }
     compute();
-    for (auto & w : workers) w.join();
-
+    for (auto & w : workers) {
+        w.join();
+    }
 }
 
 static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
@@ -138,22 +150,24 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
-    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
     const bool add_bos = is_spm;
 
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
     std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
 
-    if (int(tokens.size()) < 2*params.n_ctx) {
-        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*params.n_ctx,
-                params.n_ctx);
+    const int n_ctx = llama_n_ctx(ctx);
+
+    if (int(tokens.size()) < 2*n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
+                n_ctx);
         fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
         return {std::move(tokens), 0., {}, {}};
     }
 
-    std::vector<float>       logit_history;
-    std::vector<float>       prob_history;
+    std::vector<float> logit_history;
+    std::vector<float> prob_history;
 
     logit_history.resize(tokens.size());
     prob_history.resize(tokens.size());
@@ -163,20 +177,20 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
         return {tokens, -1, logit_history, prob_history};
     }
 
-    const int calc_chunk = params.n_ctx;
+    const int calc_chunk = n_ctx;
 
     fprintf(stderr, "%s: have %zu tokens. Calculation chunk = %d\n", __func__, tokens.size(), calc_chunk);
 
     if (int(tokens.size()) <= calc_chunk) {
         fprintf(stderr, "%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__,
-                tokens.size(), params.n_ctx, params.ppl_stride);
+                tokens.size(), n_ctx, params.ppl_stride);
         return {tokens, -1, logit_history, prob_history};
     }
 
     const int n_chunk_max = (tokens.size() - calc_chunk + params.ppl_stride - 1)  / params.ppl_stride;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
-    const int n_vocab = llama_n_vocab(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_batch = params.n_batch;
 
     int count = 0;
@@ -195,12 +209,15 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
 
         const auto t_start = std::chrono::high_resolution_clock::now();
 
+        // clear the KV cache
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
 
             //fprintf(stderr, "    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
-            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 //fprintf(stderr, "%s : failed to eval\n", __func__);
                 return {tokens, -1, logit_history, prob_history};
             }
@@ -235,7 +252,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
         }
 
         //fprintf(stderr, "%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
-        for (int j = params.n_ctx - params.ppl_stride - 1; j < params.n_ctx - 1; ++j) {
+        for (int j = n_ctx - params.ppl_stride - 1; j < n_ctx - 1; ++j) {
 
             // Calculate probability of next token, given the previous ones.
             const std::vector<float> tok_logits(
@@ -272,8 +289,9 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
-    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
     const bool add_bos = is_spm;
+    const int n_ctx = llama_n_ctx(ctx);
 
     auto tim1 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
@@ -283,9 +301,9 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     auto tim2 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
 
-    if (int(tokens.size()) < 2*params.n_ctx) {
-        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*params.n_ctx,
-                params.n_ctx);
+    if (int(tokens.size()) < 2*n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
+                n_ctx);
         fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
         return {std::move(tokens), 0., {}, {}};
     }
@@ -296,10 +314,10 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     std::vector<float> prob_history;
     prob_history.resize(tokens.size());
 
-    const int n_chunk_max = tokens.size() / params.n_ctx;
+    const int n_chunk_max = tokens.size() / n_ctx;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
-    const int n_vocab = llama_n_vocab(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_batch = params.n_batch;
 
     int count = 0;
@@ -311,15 +329,18 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
     for (int i = 0; i < n_chunk; ++i) {
-        const int start =     i * params.n_ctx;
-        const int end   = start + params.n_ctx;
+        const int start =     i * n_ctx;
+        const int end   = start + n_ctx;
 
-        const int num_batches = (params.n_ctx + n_batch - 1) / n_batch;
+        const int num_batches = (n_ctx + n_batch - 1) / n_batch;
 
         std::vector<float> logits;
 
         const auto t_start = std::chrono::high_resolution_clock::now();
 
+        // clear the KV cache
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
@@ -332,7 +353,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
                 tokens[batch_start] = llama_token_bos(ctx);
             }
 
-            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return {tokens, -1, logit_history, prob_history};
             }
@@ -340,7 +361,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             // restore the original token in case it was set to BOS
             tokens[batch_start] = token_org;
 
-            const auto batch_logits = llama_get_logits(ctx);
+            const auto * batch_logits = llama_get_logits(ctx);
             logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
         }
 
@@ -369,10 +390,10 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         // Example, we have a context window of 512, we will compute perplexity for each of the
         // last 256 tokens.  Then, we split the input up into context window size chunks to
         // process the entire prompt.
-        const int first = params.n_ctx/2;
-        process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, params.n_ctx - 1 - first,
+        const int first = n_ctx/2;
+        process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
                        workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
-        count += params.n_ctx - first - 1;
+        count += n_ctx - first - 1;
 
         // perplexity is e^(average negative log-likelihood)
         if (params.ppl_output_type == 0) {
@@ -381,7 +402,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             double av = nll/count;
             double av2 = nll2/count - av*av;
             if (av2 > 0) av2 = sqrt(av2/(count-1));
-            printf("%8d  %.4lf  %4lf  %4lf\n", i*params.n_ctx, std::exp(nll / count), av, av2);
+            printf("%8d  %.4lf  %4lf  %4lf\n", i*n_ctx, std::exp(nll / count), av, av2);
         }
         fflush(stdout);
     }
@@ -402,7 +423,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 }
 
 static std::vector<float> hellaswag_evaluate_tokens(
-    llama_context * ctx, const std::vector<int>& tokens, int n_past, int n_batch, int n_vocab, int n_thread
+    llama_context * ctx, std::vector<int> & tokens, int n_past, int n_batch, int n_vocab
 ) {
     std::vector<float> result;
     result.reserve(tokens.size() * n_vocab);
@@ -410,7 +431,7 @@ static std::vector<float> hellaswag_evaluate_tokens(
     for (size_t i_chunk = 0; i_chunk < n_chunk; ++i_chunk) {
         size_t n_tokens = tokens.size() - i_chunk * n_batch;
         n_tokens = std::min(n_tokens, size_t(n_batch));
-        if (llama_eval(ctx, tokens.data() + i_chunk * n_batch, n_tokens, n_past, n_thread)) {
+        if (llama_decode(ctx, llama_batch_get_one(tokens.data() + i_chunk * n_batch, n_tokens, n_past, 0))) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return {};
         }
@@ -457,7 +478,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     size_t hs_task_count = prompt_lines.size()/6;
     fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
 
-    const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
     fprintf(stderr, "================================= is_spm = %d\n", is_spm);
 
     // This is needed as usual for LLaMA models
@@ -512,7 +533,8 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     printf("\ntask\tacc_norm\n");
 
     double acc = 0.0f;
-    const int n_vocab = llama_n_vocab(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_ctx = llama_n_ctx(ctx);
 
     std::vector<std::vector<int>> ending_tokens(4);
 
@@ -540,7 +562,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         auto query_size = query_embd.size();
 
         // Stop if query wont fit the ctx window
-        if (query_size > (size_t)params.n_ctx) {
+        if (query_size > (size_t)n_ctx) {
             fprintf(stderr, "%s : number of tokens in query %zu > n_ctxl\n", __func__, query_size);
             return;
         }
@@ -550,7 +572,10 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             query_embd.resize(32);
         }
 
-        auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab, params.n_threads);
+        // clear the KV cache
+        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()) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return;
@@ -587,7 +612,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             query_size = query_embd.size();
 
             // Stop if query wont fit the ctx window
-            if (context_size + query_size > (size_t)params.n_ctx) {
+            if (context_size + query_size > (size_t)n_ctx) {
                 fprintf(stderr, "%s : number of tokens in query %zu > n_ctxl\n", __func__, query_size);
                 return;
             }
@@ -599,7 +624,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             //}
 
             // Evaluate the query
-            logits = hellaswag_evaluate_tokens(ctx, query_embd, context_size, params.n_batch, n_vocab, params.n_threads);
+            logits = hellaswag_evaluate_tokens(ctx, query_embd, context_size, params.n_batch, n_vocab);
             if (logits.empty()) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return;
@@ -661,7 +686,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    params.perplexity = true;
+    params.logits_all = true;
     params.n_batch = std::min(params.n_batch, params.n_ctx);
 
     if (params.ppl_stride > 0) {
@@ -695,7 +720,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    const int n_ctx_train = llama_n_ctx_train(ctx);
+    const int n_ctx_train = llama_n_ctx_train(model);
     if (params.n_ctx > n_ctx_train) {
         fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
                 __func__, n_ctx_train, params.n_ctx);
@@ -704,8 +729,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
     struct results_perplexity results;

examples/quantize-stats/quantize-stats.cpp

@@ -309,21 +309,22 @@ int main(int argc, char ** argv) {
     llama_context * ctx;
 
     {
-        auto lparams = llama_context_default_params();
+        auto mparams = llama_model_default_params();
+        mparams.use_mlock  = false;
 
-        lparams.n_ctx      = 256;
-        lparams.seed       = 1;
-        lparams.f16_kv     = false;
-        lparams.use_mlock  = false;
-
-        model = llama_load_model_from_file(params.model.c_str(), lparams);
+        model = llama_load_model_from_file(params.model.c_str(), mparams);
 
         if (model == NULL) {
             fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
             return 1;
         }
 
-        ctx = llama_new_context_with_model(model, lparams);
+        auto cparams = llama_context_default_params();
+        cparams.n_ctx      = 256;
+        cparams.seed       = 1;
+        cparams.f16_kv     = false;
+
+        ctx = llama_new_context_with_model(model, cparams);
 
         if (ctx == NULL) {
             fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());

examples/quantize/README.md

@@ -1,3 +1,44 @@
 # quantize
 
 TODO
+
+## Llama 2 7B
+
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.35
+Q3_K_S | 3.50
+Q3_K_M | 3.91
+Q3_K_L | 4.27
+Q4_K_S | 4.58
+Q4_K_M | 4.84
+Q5_K_S | 5.52
+Q5_K_M | 5.68
+Q6_K | 6.56
+
+## Llama 2 13B
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.34
+Q3_K_S | 3.48
+Q3_K_M | 3.89
+Q3_K_L | 4.26
+Q4_K_S | 4.56
+Q4_K_M | 4.83
+Q5_K_S | 5.51
+Q5_K_M | 5.67
+Q6_K | 6.56
+
+# Llama 2 70B
+
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.40
+Q3_K_S | 3.47
+Q3_K_M | 3.85
+Q3_K_L | 4.19
+Q4_K_S | 4.53
+Q4_K_M | 4.80
+Q5_K_S | 5.50
+Q5_K_M | 5.65
+Q6_K | 6.56

examples/quantize/quantize.cpp

@@ -72,6 +72,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 // usage:
 //  ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
 //
+[[noreturn]]
 static void usage(const char * executable) {
     printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");

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

@@ -23,23 +23,17 @@ int main(int argc, char ** argv) {
         params.n_predict = 16;
     }
 
-    auto lparams = llama_context_default_params();
-
-    lparams.n_ctx     = params.n_ctx;
-    lparams.seed      = params.seed;
-    lparams.f16_kv    = params.memory_f16;
-    lparams.use_mmap  = params.use_mmap;
-    lparams.use_mlock = params.use_mlock;
-
     auto n_past = 0;
     auto last_n_tokens_data = std::vector<llama_token>(params.repeat_last_n, 0);
 
     // init
-    auto model = llama_load_model_from_file(params.model.c_str(), lparams);
+    llama_model * model;
+    llama_context * ctx;
+
+    std::tie(model, ctx) = llama_init_from_gpt_params( params );
     if (model == nullptr) {
         return 1;
     }
-    auto ctx = llama_new_context_with_model(model, lparams);
     if (ctx == nullptr) {
         llama_free_model(model);
         return 1;
@@ -54,7 +48,7 @@ int main(int argc, char ** argv) {
     }
 
     // evaluate prompt
-    llama_eval(ctx, tokens.data(), n_prompt_tokens, n_past, params.n_threads);
+    llama_decode(ctx, llama_batch_get_one(tokens.data(), n_prompt_tokens, n_past, 0));
 
     last_n_tokens_data.insert(last_n_tokens_data.end(), tokens.data(), tokens.data() + n_prompt_tokens);
     n_past += n_prompt_tokens;
@@ -78,8 +72,8 @@ int main(int argc, char ** argv) {
     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(ctx);
+        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++) {
@@ -91,7 +85,7 @@ int main(int argc, char ** argv) {
         last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());
-        if (llama_eval(ctx, &next_token, 1, n_past, params.n_threads)) {
+        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);
             llama_free_model(model);
@@ -106,7 +100,7 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
 
     // make new context
-    auto ctx2 = llama_new_context_with_model(model, lparams);
+    auto * ctx2 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
 
     // Load state (rng, logits, embedding and kv_cache) from file
     {
@@ -138,8 +132,8 @@ int main(int argc, char ** argv) {
 
     // second run
     for (auto i = 0; i < params.n_predict; i++) {
-        auto logits = llama_get_logits(ctx2);
-        auto n_vocab = llama_n_vocab(ctx2);
+        auto * logits = llama_get_logits(ctx2);
+        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++) {
@@ -151,7 +145,7 @@ int main(int argc, char ** argv) {
         last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());
-        if (llama_eval(ctx2, &next_token, 1, n_past, params.n_threads)) {
+        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);