fix: LLAMA_CUBLAS=1 undefined reference 'shm_open' (#1080)

* Q4_2 quantization with rmse-optimized scale and quants

For quantize-stats we get
q4_2: rmse 0.00159301, maxerr 0.17480469, 95pct<0.0030, median<0.0012

For 7B perplexity with BLAS enabled we get 6.2038 after 655 chunks.

Quantization is slow (~90 seconds on my Mac for 7B) as not
multi-threaded as in PR #896.

* ggml : satisfy the sanitizer builds

Not sure why this makes them fail

* Better follow ggml conventions for function names

* Fixed type as per reviewer comment

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.devops/full.Dockerfile

@@ -5,9 +5,10 @@ FROM ubuntu:$UBUNTU_VERSION as build
 RUN apt-get update && \
     apt-get install -y build-essential python3 python3-pip
 
+COPY requirements.txt requirements.txt
+
 RUN pip install --upgrade pip setuptools wheel \
-    && pip install numpy requests sentencepiece tqdm \
-    && pip install torch --index-url https://download.pytorch.org/whl/cpu
+    && pip install -r requirements.txt
 
 WORKDIR /app
 

.github/workflows/build.yml

@@ -8,6 +8,8 @@ on:
         required: true
         type: boolean
   push:
+    branches:
+      - master
     paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
   pull_request:
     types: [opened, synchronize, edited, reopened, review_requested, ready_for_review]
@@ -18,6 +20,8 @@ env:
 
 jobs:
   ubuntu-latest-make:
+    if: github.event.pull_request.draft == false
+
     runs-on: ubuntu-latest
 
     steps:
@@ -37,6 +41,8 @@ jobs:
           make
 
   ubuntu-latest-cmake:
+    if: github.event.pull_request.draft == false
+
     runs-on: ubuntu-latest
 
     steps:
@@ -65,6 +71,8 @@ jobs:
           ctest --verbose
 
   ubuntu-latest-cmake-sanitizer:
+    if: github.event.pull_request.draft == false
+
     runs-on: ubuntu-latest
 
     continue-on-error: true
@@ -101,6 +109,8 @@ jobs:
           ctest --verbose
 
   macOS-latest-make:
+    if: github.event.pull_request.draft == false
+
     runs-on: macos-latest
 
     steps:
@@ -119,6 +129,8 @@ jobs:
           make
 
   macOS-latest-cmake:
+    if: github.event.pull_request.draft == false
+
     runs-on: macOS-latest
 
     steps:
@@ -146,6 +158,8 @@ jobs:
           ctest --verbose
 
   windows-latest-cmake:
+    if: github.event.pull_request.draft == false
+
     runs-on: windows-latest
 
     strategy:

.github/workflows/docker.yml

@@ -18,6 +18,8 @@ on:
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
+    if: github.event.pull_request.draft == false
+
     runs-on: ubuntu-latest
     env:
       COMMIT_SHA: ${{ github.sha }}

.gitignore

@@ -1,11 +1,15 @@
 *.o
 *.a
+.DS_Store
+.build/
 .cache/
+.direnv/
+.envrc
+.swiftpm
+.venv
 .vs/
 .vscode/
-.DS_Store
 
-.build/
 build/
 build-em/
 build-debug/
@@ -24,17 +28,15 @@ models/*
 /perplexity
 /embedding
 /benchmark-q4_0-matmult
+/vdot
 /Pipfile
 
 arm_neon.h
 compile_commands.json
 
-.envrc
-.direnv/
-
-.venv
 __pycache__
-.swiftpm
 
 zig-out/
 zig-cache/
+
+ppl-*.txt

CMakeLists.txt

@@ -55,6 +55,8 @@ option(LLAMA_SANITIZE_UNDEFINED     "llama: enable undefined sanitizer"
 option(LLAMA_AVX                    "llama: enable AVX"                                     ON)
 option(LLAMA_AVX2                   "llama: enable AVX2"                                    ON)
 option(LLAMA_AVX512                 "llama: enable AVX512"                                  OFF)
+option(LLAMA_AVX512_VBMI            "llama: enable AVX512-VBMI"                             OFF)
+option(LLAMA_AVX512_VNNI            "llama: enable AVX512-VNNI"                             OFF)
 option(LLAMA_FMA                    "llama: enable FMA"                                     ON)
 # in MSVC F16C is implied with AVX2/AVX512
 if (NOT MSVC)
@@ -64,6 +66,7 @@ endif()
 # 3rd party libs
 option(LLAMA_ACCELERATE             "llama: enable Accelerate framework"                    ON)
 option(LLAMA_OPENBLAS               "llama: use OpenBLAS"                                   OFF)
+option(LLAMA_CUBLAS                 "llama: use cuBLAS"                                     OFF)
 
 option(LLAMA_BUILD_TESTS            "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES         "llama: build examples" ${LLAMA_STANDALONE})
@@ -107,6 +110,7 @@ if (APPLE AND LLAMA_ACCELERATE)
         message(WARNING "Accelerate framework not found")
     endif()
 endif()
+
 if (LLAMA_OPENBLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
@@ -120,11 +124,50 @@ if (LLAMA_OPENBLAS)
         add_compile_definitions(GGML_USE_OPENBLAS)
         add_link_options(${BLAS_LIBRARIES})
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} openblas)
+
+        # find header file
+        set(OPENBLAS_INCLUDE_SEARCH_PATHS
+            /usr/include
+            /usr/include/openblas
+            /usr/include/openblas-base
+            /usr/local/include
+            /usr/local/include/openblas
+            /usr/local/include/openblas-base
+            /opt/OpenBLAS/include
+            $ENV{OpenBLAS_HOME}
+            $ENV{OpenBLAS_HOME}/include
+            )
+        find_path(OPENBLAS_INC NAMES cblas.h PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS})
+        add_compile_options(-I${OPENBLAS_INC})
     else()
         message(WARNING "OpenBLAS not found")
     endif()
 endif()
 
+if (LLAMA_CUBLAS)
+    cmake_minimum_required(VERSION 3.17)
+
+    find_package(CUDAToolkit)
+    if (CUDAToolkit_FOUND)
+        message(STATUS "cuBLAS found")
+
+        enable_language(CUDA)
+
+        set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
+
+        add_compile_definitions(GGML_USE_CUBLAS)
+
+        if (LLAMA_STATIC)
+            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+        else()
+            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+        endif()
+
+    else()
+        message(WARNING "cuBLAS not found")
+    endif()
+endif()
+
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
         set(c_flags
@@ -136,7 +179,6 @@ if (LLAMA_ALL_WARNINGS)
             -Wshadow
             -Wstrict-prototypes
             -Wpointer-arith
-            -Wno-unused-function
         )
         set(cxx_flags
             -Wall
@@ -204,11 +246,26 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
     message(STATUS "x86 detected")
     if (MSVC)
         if (LLAMA_AVX512)
-            add_compile_options(/arch:AVX512)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX512>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX512>)
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (LLAMA_AVX512_VBMI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+            endif()
+            if (LLAMA_AVX512_VNNI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+            endif()
         elseif (LLAMA_AVX2)
-            add_compile_options(/arch:AVX2)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX2>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX2>)
         elseif (LLAMA_AVX)
-            add_compile_options(/arch:AVX)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX>)
         endif()
     else()
         if (LLAMA_F16C)
@@ -225,9 +282,13 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
         endif()
         if (LLAMA_AVX512)
             add_compile_options(-mavx512f)
-            # add_compile_options(-mavx512cd)
-            # add_compile_options(-mavx512dq)
-            # add_compile_options(-mavx512bw)
+            add_compile_options(-mavx512bw)
+        endif()
+        if (LLAMA_AVX512_VBMI)
+            add_compile_options(-mavx512vbmi)
+        endif()
+        if (LLAMA_AVX512_VNNI)
+            add_compile_options(-mavx512vnni)
         endif()
     endif()
 else()
@@ -241,7 +302,8 @@ endif()
 
 add_library(ggml OBJECT
             ggml.c
-            ggml.h)
+            ggml.h
+            ${GGML_CUDA_SOURCES})
 
 target_include_directories(ggml PUBLIC .)
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
@@ -263,6 +325,14 @@ if (BUILD_SHARED_LIBS)
     target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)
 endif()
 
+if (GGML_CUDA_SOURCES)
+    message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
+    set_property(TARGET ggml PROPERTY CUDA_ARCHITECTURES OFF)
+    set_property(TARGET ggml PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
+    set_property(TARGET llama PROPERTY CUDA_ARCHITECTURES OFF)
+endif()
+
+
 #
 # programs, examples and tests
 #
@@ -274,4 +344,5 @@ endif ()
 
 if (LLAMA_BUILD_EXAMPLES)
     add_subdirectory(examples)
+    add_subdirectory(pocs)
 endif()

Makefile

@@ -1,3 +1,6 @@
+# Define the default target now so that it is always the first target
+default: main quantize quantize-stats perplexity embedding vdot
+
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
 endif
@@ -36,7 +39,7 @@ CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
 # warnings
-CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith -Wno-unused-function
+CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith
 CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
 
 # OS specific
@@ -97,6 +100,13 @@ ifdef LLAMA_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
 	LDFLAGS += -lopenblas
 endif
+ifdef LLAMA_CUBLAS
+	CFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include
+	LDFLAGS += -lcublas_static -lculibos -lcudart_static -lcublasLt_static -lpthread -ldl -lrt -L/usr/local/cuda/lib64
+	OBJS	+= ggml-cuda.o
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	nvcc -arch=native -c -o $@ $<
+endif
 ifdef LLAMA_GPROF
 	CFLAGS   += -pg
 	CXXFLAGS += -pg
@@ -133,51 +143,52 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main quantize perplexity embedding
-
 #
 # Build library
 #
 
 ggml.o: ggml.c ggml.h
-	$(CC)  $(CFLAGS)   -c ggml.c -o ggml.o
+	$(CC)  $(CFLAGS)   -c $< -o $@
 
-llama.o: llama.cpp llama.h llama_util.h
-	$(CXX) $(CXXFLAGS) -c llama.cpp -o llama.o
+llama.o: llama.cpp ggml.h llama.h llama_util.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 common.o: examples/common.cpp examples/common.h
-	$(CXX) $(CXXFLAGS) -c examples/common.cpp -o common.o
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 clean:
 	rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-q4_0-matmult
 
-main: examples/main/main.cpp ggml.o llama.o common.o
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp ggml.o llama.o common.o -o main $(LDFLAGS)
+main: examples/main/main.cpp ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-quantize: examples/quantize/quantize.cpp ggml.o llama.o
-	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp ggml.o llama.o -o quantize $(LDFLAGS)
+quantize: examples/quantize/quantize.cpp ggml.o llama.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o
-	$(CXX) $(CXXFLAGS) examples/quantize-stats/quantize-stats.cpp ggml.o llama.o -o quantize-stats $(LDFLAGS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o
-	$(CXX) $(CXXFLAGS) examples/perplexity/perplexity.cpp ggml.o llama.o common.o -o perplexity $(LDFLAGS)
+perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o
-	$(CXX) $(CXXFLAGS) examples/embedding/embedding.cpp ggml.o llama.o common.o -o embedding $(LDFLAGS)
+embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-libllama.so: llama.o ggml.o
-	$(CXX) $(CXXFLAGS) -shared -fPIC -o libllama.so llama.o ggml.o $(LDFLAGS)
+vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+
+libllama.so: llama.o ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 #
 # Tests
 #
 
-benchmark: ggml.o
-	$(CXX) $(CXXFLAGS) examples/benchmark/benchmark-q4_0-matmult.c ggml.o -o benchmark-q4_0-matmult $(LDFLAGS)
+benchmark: examples/benchmark/benchmark-q4_0-matmult.c ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o benchmark-q4_0-matmult $(LDFLAGS)
 	./benchmark-q4_0-matmult
 
 .PHONY: tests

README.md

@@ -7,14 +7,19 @@
 
 Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
+**Warnings**
+
+- `Q4_2` and `Q4_3` are still in development. Do not expect any kind of backward compatibility until they are finalized
+
 **Hot topics:**
 
+- [Added LoRA support](https://github.com/ggerganov/llama.cpp/pull/820)
 - [Add GPU support to ggml](https://github.com/ggerganov/llama.cpp/discussions/915)
 - [Roadmap Apr 2023](https://github.com/ggerganov/llama.cpp/discussions/784)
 
 ## Description
 
-The main goal is to run the model using 4-bit quantization on a MacBook
+The main goal of llama.cpp is to run the llama model using 4-bit quantization on a MacBook.
 
 - Plain C/C++ implementation without dependencies
 - Apple silicon first-class citizen - optimized via ARM NEON and Accelerate framework
@@ -50,6 +55,7 @@ New features will probably be added mostly through community contributions.
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
+- Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 
 **UI:**
 
@@ -150,7 +156,7 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8
 
 ## Usage
 
-Here are the step for the LLaMA-7B model.
+Here are the steps for the LLaMA-7B model.
 
 ### Get the Code
 
@@ -192,10 +198,10 @@ ls ./models
 65B 30B 13B 7B tokenizer_checklist.chk tokenizer.model
 
 # install Python dependencies
-python3 -m pip install torch numpy sentencepiece
+python3 -m pip install -r requirements.txt
 
 # convert the 7B model to ggml FP16 format
-python3 convert-pth-to-ggml.py models/7B/ 1
+python3 convert.py models/7B/
 
 # quantize the model to 4-bits (using method 2 = q4_0)
 ./quantize ./models/7B/ggml-model-f16.bin ./models/7B/ggml-model-q4_0.bin 2
@@ -208,8 +214,7 @@ When running the larger models, make sure you have enough disk space to store al
 
 ### Memory/Disk Requirements
 
-As the models are currently fully loaded into memory, you will need adequate disk space to save them
-and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
+As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
 
 | model | original size | quantized size (4-bit) |
 |-------|---------------|------------------------|
@@ -221,18 +226,18 @@ and sufficient RAM to load them. At the moment, memory and disk requirements are
 ### Interactive mode
 
 If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
-In this mode, you can always interrupt generation by pressing Ctrl+C and enter one or more lines of text which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt which makes LLaMa emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
+In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMa emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
 
-Here is an example few-shot interaction, invoked with the command
+Here is an example of a few-shot interaction, invoked with the command
 
 ```bash
-# default arguments using 7B model
+# default arguments using a 7B model
 ./examples/chat.sh
 
-# advanced chat with 13B model
+# advanced chat with a 13B model
 ./examples/chat-13B.sh
 
-# custom arguments using 13B model
+# custom arguments using a 13B model
 ./main -m ./models/13B/ggml-model-q4_0.bin -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt
 ```
 
@@ -271,7 +276,7 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
 ### Using [GPT4All](https://github.com/nomic-ai/gpt4all)
 
 - Obtain the `gpt4all-lora-quantized.bin` model
-- It is distributed in the old `ggml` format which is now obsoleted
+- It is distributed in the old `ggml` format, which is now obsoleted
 - You have to convert it to the new format using [./convert-gpt4all-to-ggml.py](./convert-gpt4all-to-ggml.py). You may also need to
 convert the model from the old format to the new format with [./migrate-ggml-2023-03-30-pr613.py](./migrate-ggml-2023-03-30-pr613.py):
 
@@ -285,7 +290,7 @@ convert the model from the old format to the new format with [./migrate-ggml-202
 
 ### Obtaining and verifying the Facebook LLaMA original model and Stanford Alpaca model data
 
-- **Under no circumstances share IPFS, magnet links, or any other links to model downloads anywhere in this respository, including in issues, discussions or pull requests. They will be immediately deleted.**
+- **Under no circumstances should IPFS, magnet links, or any other links to model downloads be shared anywhere in this repository, including in issues, discussions, or pull requests. They will be immediately deleted.**
 - The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository.
 - Refer to [Facebook's LLaMA repository](https://github.com/facebookresearch/llama/pull/73/files) if you need to request access to the model data.
 - Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files.
@@ -297,29 +302,27 @@ convert the model from the old format to the new format with [./migrate-ggml-202
 
   `shasum -a 256 --ignore-missing -c SHA256SUMS` on macOS
 
-- If your issue is with model generation quality then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
-  - LLaMA:
-    - [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/)
-    - [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971)
-  - GPT-3
-    - [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165)
-  - GPT-3.5 / InstructGPT / ChatGPT:
-    - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
-    - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
+- If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
+- LLaMA:
+- [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/)
+- [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971)
+- GPT-3
+- [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165)
+- GPT-3.5 / InstructGPT / ChatGPT:
+- [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
+- [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
 
-### Perplexity (Measuring model quality)
+### Perplexity (measuring model quality)
 
-You can use the `perplexity` example to measure perplexity over the given prompt.  For more background,
-see https://huggingface.co/docs/transformers/perplexity.  However, in general, lower perplexity is better for LLMs.
+You can use the `perplexity` example to measure perplexity over the given prompt. For more background, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity). However, in general, lower perplexity is better for LLMs.
 
 #### Latest measurements
 
-The latest perplexity scores for the various model sizes and quantizations are being tracked in [discussion #406](https://github.com/ggerganov/llama.cpp/discussions/406).  `llama.cpp` is measuring very well
-compared to the baseline implementations.  Quantization has a small negative impact to quality, but, as you can see, running
+The latest perplexity scores for the various model sizes and quantizations are being tracked in [discussion #406](https://github.com/ggerganov/llama.cpp/discussions/406). `llama.cpp` is measuring very well compared to the baseline implementations. Quantization has a small negative impact on quality, but, as you can see, running
 13B at q4_0 beats the 7B f16 model by a significant amount.
 
-All measurements are done against wikitext2 test dataset (https://paperswithcode.com/dataset/wikitext-2), with default options (512 length context).
-Note that the changing the context length will have a significant impact on perplexity (longer context = better perplexity).
+All measurements are done against the wikitext2 test dataset (https://paperswithcode.com/dataset/wikitext-2), with default options (512 length context).
+Note that changing the context length will have a significant impact on perplexity (longer context = better perplexity).
 ```
 Perplexity - model options
 5.5985 - 13B, q4_0
@@ -361,7 +364,7 @@ https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b0
 
 #### Prerequisites
 * Docker must be installed and running on your system.
-* Create a folder to store big models & intermediate files (in ex. im using /llama/models)
+* Create a folder to store big models & intermediate files (ex. /llama/models)
 
 #### Images
 We have two Docker images available for this project:
@@ -375,17 +378,17 @@ The easiest way to download the models, convert them to ggml and optimize them i
 
 Replace `/path/to/models` below with the actual path where you downloaded the models.
 
- ```bash
+```bash
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B
 ```
 
-On complete, you are ready to play!
+On completion, you are ready to play!
 
 ```bash
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512
 ```
 
-or with light image:
+or with a light image:
 
 ```bash
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512
@@ -406,7 +409,7 @@ docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /mode
 - Always consider cross-compatibility with other operating systems and architectures
 - Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
 - There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
-- Clean-up any trailing whitespaces, use 4 spaces indentation, brackets on same line, `void * ptr`, `int & a`
+- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 
 ### Docs

convert-ggml-to-pth.py

@@ -1,299 +0,0 @@
-# Author: github.com/ductai199x
-import argparse
-import os
-import struct
-
-import numpy as np
-import torch
-from numba import njit
-from tqdm.auto import tqdm
-
-
-def read_header(fin):
-    values = struct.unpack("i" * 9, fin.read(4 * 9))
-    _, _, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype = values
-    return {
-        "vocab_size": vocab_size,
-        "dim": dim,
-        "multiple_of": multiple_of,
-        "n_heads": n_heads,
-        "n_layers": n_layers,
-    }, ftype
-
-
-def read_tokens(fin, vocab_size):
-    tokens = []
-    for _ in range(vocab_size):
-        text_len = struct.unpack("i", fin.read(4))[0]
-        text_bytes = fin.read(text_len)
-        try:
-            text = text_bytes.decode()
-        except UnicodeDecodeError:
-            text = text_bytes.decode(errors="replace")
-        score = struct.unpack("f", fin.read(4))[0]
-        tokens.append((text, score))
-    return tokens
-
-
-@njit
-def dequantize_weights_numba(fin_data, n_rows, n_cols):
-    qk = 32
-    nb = n_cols // qk
-    bs = 4 + (qk // 2)
-
-    weights = np.zeros((n_rows, n_cols), dtype=np.float32)
-    data_pos = 0
-
-    for row in range(n_rows):
-        for block in range(nb):
-            d = np.frombuffer(fin_data[data_pos : data_pos + 4], dtype=np.float32)[0]
-            data_pos += 4
-            packed_values = fin_data[data_pos : data_pos + (qk // 2)]
-            data_pos += qk // 2
-
-            for i in range(qk // 2):
-                packed_value = packed_values[i]
-                v0 = np.float32((packed_value & 0b00001111) - 8) * d
-                v1 = np.float32((packed_value >> 4) - 8) * d
-
-                weights[row, block * qk + 2 * i] = v0
-                weights[row, block * qk + 2 * i + 1] = v1
-
-    return weights
-
-
-def dequantize_weights(fin, n_rows, n_cols):
-    qk = 32
-    nb = n_cols // qk
-    data_size = n_rows * n_cols // 2 + n_rows * nb * 4
-    fin_data = fin.read(data_size)
-    return dequantize_weights_numba(fin_data, n_rows, n_cols)
-
-
-def read_variables(fin):
-    model = {}
-    pbar = tqdm(total=os.path.getsize(fin.name), unit="B", unit_scale=True, desc="Reading variables")
-    while True:
-        start_pos = fin.tell()
-        try:
-            n_dims, name_length, ftype_cur = struct.unpack("iii", fin.read(4 * 3))
-        except struct.error:
-            break
-
-        shape = tuple(struct.unpack("i" * n_dims, fin.read(4 * n_dims)))
-        shape = shape[::-1]
-        name = fin.read(name_length).decode()
-
-        # ensure tensor data is aligned
-        tensor_data_offset = fin.tell()
-        tensor_data_offset = (tensor_data_offset + 31) & -32
-        fin.seek(tensor_data_offset)
-
-        if ftype_cur == 2:
-            # 4-bit quantized weights
-            dtype = np.uint8
-            data = dequantize_weights(fin, shape[0], shape[1])
-            data = data.reshape(shape)
-        elif ftype_cur == 0:
-            dtype = np.float32
-            data_size = np.prod(shape)
-            data = np.fromfile(fin, dtype=dtype, count=data_size).reshape(shape)
-        elif ftype_cur == 1:
-            dtype = np.float16
-            data_size = np.prod(shape)
-            data = np.fromfile(fin, dtype=dtype, count=data_size).reshape(shape)
-
-        model[name] = torch.tensor(data, dtype=torch.float32 if dtype == np.float32 else torch.float16)
-
-        pbar.update(fin.tell() - start_pos)
-
-    return model
-
-
-def convert_to_hf_format(model, hparams):
-    # This works for llama 7B, need to test with other models
-    n_layers = hparams["n_layers"]
-    n_heads = hparams["n_heads"]
-    dim = hparams["dim"]
-    dims_per_head = dim // n_heads
-    base = 10000.0
-    inv_freq = 1.0 / (base ** (torch.arange(0, dims_per_head, 2).float() / dims_per_head))
-
-    # permute for sliced rotary
-    def permute(w):
-        return w.view(n_heads, dim // n_heads // 2, 2, dim).transpose(1, 2).reshape(dim, dim)
-
-    state_dict = {}
-    for layer_i in range(n_layers):
-        state_dict.update(
-            {
-                f"model.layers.{layer_i}.self_attn.q_proj.weight": permute(
-                    model[f"layers.{layer_i}.attention.wq.weight"]
-                ),
-                f"model.layers.{layer_i}.self_attn.k_proj.weight": permute(
-                    model[f"layers.{layer_i}.attention.wk.weight"]
-                ),
-                f"model.layers.{layer_i}.self_attn.v_proj.weight": model[
-                    f"layers.{layer_i}.attention.wv.weight"
-                ],
-                f"model.layers.{layer_i}.self_attn.o_proj.weight": model[
-                    f"layers.{layer_i}.attention.wo.weight"
-                ],
-                f"model.layers.{layer_i}.mlp.gate_proj.weight": model[
-                    f"layers.{layer_i}.feed_forward.w1.weight"
-                ],
-                f"model.layers.{layer_i}.mlp.down_proj.weight": model[
-                    f"layers.{layer_i}.feed_forward.w2.weight"
-                ],
-                f"model.layers.{layer_i}.mlp.up_proj.weight": model[
-                    f"layers.{layer_i}.feed_forward.w3.weight"
-                ],
-                f"model.layers.{layer_i}.input_layernorm.weight": model[
-                    f"layers.{layer_i}.attention_norm.weight"
-                ],
-                f"model.layers.{layer_i}.post_attention_layernorm.weight": model[
-                    f"layers.{layer_i}.ffn_norm.weight"
-                ],
-            }
-        )
-        state_dict[f"model.layers.{layer_i}.self_attn.rotary_emb.inv_freq"] = inv_freq
-    state_dict.update(
-        {
-            "model.embed_tokens.weight": model["tok_embeddings.weight"],
-            "model.norm.weight": model["norm.weight"],
-            "lm_head.weight": model["output.weight"],
-        }
-    )
-
-    return state_dict
-
-
-def chat(model, hparams, llama_dir):
-    from transformers import (GenerationConfig, LlamaForCausalLM,
-                              LlamaTokenizer, StoppingCriteria,
-                              StoppingCriteriaList)
-    from transformers.models.llama.configuration_llama import LlamaConfig
-
-    class StoppingCriteriaSub(StoppingCriteria):
-        def __init__(self):
-            super().__init__()
-
-        def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, stops=[]):
-            print(tokenizer.decode(input_ids[0]), end="", flush=True)
-            if input_ids[0][-1] == 13:
-                return True
-
-            return False
-
-    config = LlamaConfig(
-        vocab_size=hparams["vocab_size"],
-        dim=hparams["dim"],
-        num_hidden_layers=hparams["n_layers"],
-        num_attention_heads=hparams["n_heads"],
-    )
-
-    llama = LlamaForCausalLM(config=config)
-    llama.load_state_dict(state_dict=model, strict=True)
-    tokenizer = LlamaTokenizer.from_pretrained(llama_dir)
-
-    device = torch.device("cpu")
-    llama = llama.to(device)
-
-    ctx = """You are AI.
-This is a dialog, where User interacts with AI. AI is helpful, kind, obedient, honest, respectful, direct, concise, should try to protect User's privacy, and knows its own limits. Also, AI must answer User and AI cannot stop the conversation by itself.
-User: Hello, AI.
-AI: Hello! How can I assist you today?
-"""
-    print(ctx.rstrip("\n"))
-    while True:
-        print("-" * 60)
-        prompt = input("User: ")
-        if ctx != "":
-            ctx = f"{ctx}User: {prompt}\n"
-        else:
-            ctx = f"{prompt}\nAI:"
-
-        ctx = (ctx[-1920:]) if len(ctx) >= 2048 else ctx
-
-        print("-" * 60)
-        if len(ctx.strip()) > 0:
-            input_ids = tokenizer(ctx, return_tensors="pt")["input_ids"].to(device)
-            generation_config = GenerationConfig(
-                temperature=0.8,
-                top_p=0.95,
-                top_k=50,
-                repetition_penalty=1.1764,
-            )
-            with torch.no_grad():
-                generation_output = llama.generate(
-                    input_ids=input_ids,
-                    generation_config=generation_config,
-                    return_dict_in_generate=True,
-                    output_scores=True,
-                    max_length=2048,
-                    do_sample=True,
-                    stopping_criteria=StoppingCriteriaList([StoppingCriteriaSub()]),
-                )
-            s = generation_output.sequences[0]
-            decoded = tokenizer.decode(s)
-            ctx = f"{decoded}\n"
-
-
-def main():
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--input_dir", "-i", type=str, required=True, help="The input directory containing the ggml files."
-    )
-    parser.add_argument(
-        "--prefix",
-        "-p",
-        type=str,
-        required=True,
-        help="The prefix of the ggml files (ggml-model-f16 or ggml-model-q4_0).",
-    )
-    parser.add_argument(
-        "--hf",
-        action="store_true",
-        help="Whether to save the model in the Hugging Face format. (default: False)",
-    )
-    parser.add_argument(
-        "--chat", "-c", action="store_true", help="Whether to open a chat with the model. (default: False)"
-    )
-    args = parser.parse_args()
-
-    llama_dir = os.path.abspath(f"{args.input_dir}/../")
-
-    ggml_files = sorted(
-        [f"{args.input_dir}/{f}" for f in os.listdir(args.input_dir) if f.startswith(args.prefix)]
-    )
-
-    fin = open(ggml_files[0], "rb")
-    hparams, ftype = read_header(fin)
-    tokens = read_tokens(fin, hparams["vocab_size"])
-    model = read_variables(fin)
-
-    for f in tqdm(ggml_files[1:]):
-        fin = open(f, "rb")
-        read_header(fin)
-        read_tokens(fin, hparams["vocab_size"])
-        model.update(read_variables(fin))
-
-    if args.hf:
-        model = convert_to_hf_format(model, hparams)
-
-    pth_ckpt = {
-        "state_dict": model,
-        "hparams": hparams,
-        "tokens": tokens,
-    }
-
-    torch.save(pth_ckpt, f"{args.input_dir}/{args.prefix}-to-torch.pth")
-
-    if args.chat:
-        if not args.hf:
-            model = convert_to_hf_format(model, hparams)
-        chat(model, hparams, llama_dir)
-
-
-if __name__ == "__main__":
-    main()

convert-gpt4all-to-ggml.py

@@ -1,107 +0,0 @@
-#!/usr/bin/env python3
-
-#
-# TODO: deduplicate GPT4All with convert-unversioned-ggml-to-ggml.py
-#
-
-# Original by https://github.com/eiz
-# https://github.com/ggerganov/llama.cpp/issues/324#issuecomment-1476227818
-import argparse
-import glob
-import os
-import struct
-import sys
-from sentencepiece import SentencePieceProcessor
-
-HPARAMS = keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"]
-
-def parse_args():
-    parser = argparse.ArgumentParser(description='Upgrade a GPT4All model to the current format')
-    parser.add_argument('gpt4all_model', help='path to gpt4all-lora-quantized.bin')
-    parser.add_argument('tokenizer_model', help='path to LLaMA tokenizer.model file')
-    return parser.parse_args()
-
-def read_header(f_in):
-    struct_fmt = "i" * (3 + len(HPARAMS))
-    struct_size = struct.calcsize(struct_fmt)
-    buf = f_in.read(struct_size)
-    return struct.unpack(struct_fmt, buf)
-
-def write_header(f_out, header):
-    (magic, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype) = header
-
-    if magic != 0x67676d6c:
-        raise Exception('Invalid file magic. Must be an old style ggml file.')
-
-    values = [
-        0x67676d66, # magic: ggml in hex
-        1,          # file version
-        vocab_size,
-        dim,
-        multiple_of,
-        n_heads,
-        n_layers,
-        rot,
-        ftype
-    ]
-    f_out.write(struct.pack("i" * len(values), *values))
-
-def write_tokens(fout, tokenizer):
-    for i in range(tokenizer.vocab_size()):
-        if tokenizer.is_unknown(i):
-            text = " \u2047 ".encode()
-        elif tokenizer.is_control(i):
-            text = b""
-        elif tokenizer.is_byte(i):
-            piece = tokenizer.id_to_piece(i)
-            if len(piece) != 6:
-                print(f"Invalid token: {piece}")
-                sys.exit(1)
-            byte_value = int(piece[3:-1], 16)
-            text = struct.pack("B", byte_value)
-        else:
-            text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
-        fout.write(struct.pack("i", len(text)))
-        fout.write(text)
-        fout.write(struct.pack("f", tokenizer.get_score(i)))
-
-    # TODO: GPT4All - add extra <pad> token
-    text = "<pad>".encode()
-    fout.write(struct.pack("i", len(text)))
-    fout.write(text)
-    fout.write(struct.pack("f", 0.0))
-
-def read_tokens(f_in, tokenizer):
-    for i in range(tokenizer.vocab_size()):
-        len_b = f_in.read(4)
-        (length,) = struct.unpack("i", len_b)
-        f_in.read(length)
-
-def copy_all_data(f_out, f_in):
-    while True:
-        buf = f_in.read(1024 * 1024)
-        if not buf:
-            break
-        f_out.write(buf)
-
-def convert_one_file(path_in, tokenizer):
-    path_tmp = f"{path_in}.tmp"
-    path_orig= f"{path_in}.orig"
-    print(f"converting {path_in}")
-    with open(path_in, "rb") as f_in, open(path_tmp, "wb") as f_out:
-        write_header(f_out, read_header(f_in))
-        read_tokens(f_in, tokenizer)
-        write_tokens(f_out, tokenizer)
-        copy_all_data(f_out, f_in)
-    os.rename(path_in, path_orig)
-    os.rename(path_tmp, path_in)
-
-def main():
-    args = parse_args()
-
-    tokenizer = SentencePieceProcessor(args.tokenizer_model)
-
-    convert_one_file(args.gpt4all_model, tokenizer)
-
-if __name__ == "__main__":
-    main()

convert-gptq-to-ggml.py

@@ -1,172 +0,0 @@
-# Convert a GPTQ quantized LLaMA model to a ggml compatible file
-# Based on: https://github.com/qwopqwop200/GPTQ-for-LLaMa
-#
-import os
-import re
-import sys
-import json
-import struct
-import numpy as np
-import torch
-from sentencepiece import SentencePieceProcessor
-
-if len(sys.argv) != 4:
-    print("Usage: convert-gptq-to-ggml.py llamaXXb-4bit.pt tokenizer.model out.bin\n")
-    sys.exit(1)
-
-fname_model = sys.argv[1]
-fname_tokenizer = sys.argv[2]
-dir_out = sys.argv[3]
-
-model = torch.load(fname_model, map_location="cpu")
-
-n_vocab, n_embd = model['model.embed_tokens.weight'].shape
-n_layer = 1 + max(int(m.group(1)) for name in model
-                  if (m := re.match(r'model\.layers\.([0-9]+)', name)))
-
-# hardcoded:
-n_mult = 256
-n_head = {32: 32, 40: 40, 60: 52, 80: 64}[n_layer]
-
-tokenizer = SentencePieceProcessor(fname_tokenizer)
-
-assert tokenizer.vocab_size() == n_vocab
-
-fname_out = sys.argv[3]
-
-fout = open(fname_out, "wb")
-
-fout.write(struct.pack("i", 0x67676d66)) # magic: ggmf in hex
-fout.write(struct.pack("i", 1)) # file version
-fout.write(struct.pack("i", n_vocab))
-fout.write(struct.pack("i", n_embd))
-fout.write(struct.pack("i", n_mult))
-fout.write(struct.pack("i", n_head))
-fout.write(struct.pack("i", n_layer))
-fout.write(struct.pack("i", n_embd // n_head)) # rot (obsolete)
-fout.write(struct.pack("i", 4))
-
-
-# This loop unchanged from convert-pth-to-ggml.py:
-for i in range(tokenizer.vocab_size()):
-    if tokenizer.is_unknown(i):
-        text = " \u2047 ".encode()
-    elif tokenizer.is_control(i):
-        text = b""
-    elif tokenizer.is_byte(i):
-        piece = tokenizer.id_to_piece(i)
-        if len(piece) != 6:
-            print(f"Invalid token: {piece}")
-            sys.exit(1)
-        byte_value = int(piece[3:-1], 16)
-        text = struct.pack("B", byte_value)
-    else:
-        text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
-    fout.write(struct.pack("i", len(text)))
-    fout.write(text)
-    fout.write(struct.pack("f", tokenizer.get_score(i)))
-
-def write_header(shape, dst_name, ftype_cur):
-    sname = dst_name.encode()
-    fout.write(struct.pack("iii", len(shape), len(sname), ftype_cur))
-    fout.write(struct.pack("i" * len(shape), *shape[::-1]))
-    fout.write(sname)
-
-    # ensure tensor data is aligned
-    tensor_data_offset = fout.tell()
-    tensor_data_offset = (tensor_data_offset + 31) & -32
-    fout.seek(tensor_data_offset)
-
-def convert_non_q4(src_name, dst_name):
-    v = model[src_name]
-    shape = v.shape
-    print(f"Processing non-Q4 variable: {src_name} with shape: {shape} and type: {v.dtype}")
-    if len(shape) == 1:
-        print("  Converting to float32")
-        v = v.to(torch.float32)
-
-    ftype_cur = {torch.float16: 1, torch.float32: 0}[v.dtype]
-
-    # header
-    write_header(shape, dst_name, ftype_cur)
-
-    # data
-    v.numpy().tofile(fout)
-
-def convert_q4(src_name, dst_name, permute=False):
-    zeros = model[f"{src_name}.zeros"].numpy()
-    scales = model[f"{src_name}.scales"].numpy()
-    bias = model[f"{src_name}.bias"].numpy()
-    qweight = model[f"{src_name}.qweight"].numpy().T # transpose
-
-    # Q4_1 does not support bias; good thing the bias is always all zeros.
-    assert not np.any(bias)
-
-    # Each int32 item is actually 8 int4 items packed together, and it's transposed.
-    shape = (qweight.shape[0], qweight.shape[1] * 8)
-
-    print(f"Processing Q4 variable: {src_name} with shape: {shape}")
-
-    # The output format has the int4 weights in groups of 32 rather than 8.
-    # It looks like this:
-    # For each row:
-    #   For each group of 32 columns:
-    #     - addend (float32, 4 bytes)
-    #     - scale (float32, 4 bytes)
-    #     - weights (int4 * 32, 16 bytes)
-    # Note that in the input, the scales and addends are shared between all
-    # the columns in a row, so we end up wasting quite a bit of memory with
-    # repeated scales and addends.
-
-    addends = -zeros # flip sign
-
-    # Since the output format is mixed between integers and floats, we have
-    # to hackily view the floats as int32s just so numpy will let us
-    # concatenate them.
-    addends_view = addends.view(dtype=np.int32)
-    scales_view = scales.view(dtype=np.int32)
-
-    # Split into groups of 4 columns (i.e. 32 columns of quantized data):
-    grouped = qweight.reshape([qweight.shape[0], qweight.shape[1] // 4, 4])
-
-    # Repeat addends and scales:
-    addends_rep = np.atleast_3d(addends_view).repeat(grouped.shape[1], axis=1)
-    scales_rep = np.atleast_3d(scales_view).repeat(grouped.shape[1], axis=1)
-
-    blob = np.concatenate([scales_rep, addends_rep, grouped], axis=2, casting='no')
-
-    if permute:
-        # Permute some rows to undo the permutation done by convert_llama_weights_to_hf.py.
-        # This can be done after the above conversion because it doesn't affect column order/layout.
-        blob = (blob.reshape(n_head, 2, shape[0] // n_head // 2, *blob.shape[1:])
-                    .swapaxes(1, 2)
-                    .reshape(blob.shape))
-
-    # header
-    write_header(shape, dst_name, 3) # ftype = Q4_1
-
-    # data
-    blob.tofile(fout)
-
-convert_non_q4("model.embed_tokens.weight", "tok_embeddings.weight")
-convert_non_q4("model.norm.weight", "norm.weight")
-convert_non_q4("lm_head.weight", "output.weight")
-
-for i in range(n_layer):
-    convert_q4(f"model.layers.{i}.self_attn.q_proj", f"layers.{i}.attention.wq.weight", permute=True)
-    convert_q4(f"model.layers.{i}.self_attn.k_proj", f"layers.{i}.attention.wk.weight", permute=True)
-    convert_q4(f"model.layers.{i}.self_attn.v_proj", f"layers.{i}.attention.wv.weight")
-    convert_q4(f"model.layers.{i}.self_attn.o_proj", f"layers.{i}.attention.wo.weight")
-
-    convert_q4(f"model.layers.{i}.mlp.gate_proj", f"layers.{i}.feed_forward.w1.weight")
-    convert_q4(f"model.layers.{i}.mlp.down_proj", f"layers.{i}.feed_forward.w2.weight")
-    convert_q4(f"model.layers.{i}.mlp.up_proj",   f"layers.{i}.feed_forward.w3.weight")
-
-    convert_non_q4(f"model.layers.{i}.input_layernorm.weight", f"layers.{i}.attention_norm.weight")
-    convert_non_q4(f"model.layers.{i}.post_attention_layernorm.weight", f"layers.{i}.ffn_norm.weight")
-
-
-fout.close()
-
-print(f"Done. Output file: {fname_out}")
-print()

convert-lora-to-ggml.py

@@ -0,0 +1,124 @@
+import json
+import os
+import re
+import struct
+import sys
+from typing import Any, Dict, Sequence, TextIO
+
+import torch
+
+from convert import DATA_TYPE_TO_FTYPE, NUMPY_TYPE_TO_DATA_TYPE, DataType
+
+HF_SUBLAYER_TO_GGML = {
+    "self_attn.q_proj": "attention.wq",
+    "self_attn.k_proj": "attention.wk",
+    "self_attn.v_proj": "attention.wv",
+    "self_attn.o_proj": "attention.wo",
+    "mlp.gate_proj": "feed_forward.w1",
+    "mlp.down_proj": "feed_forward.w2",
+    "mlp.up_proj": "feed_forward.w3",
+    "input_layernorm": "attention_norm",
+    "post_attention_layernorm": "ffn_norm",
+    # "norm": "norm",
+    # "embed_tokens": "tok_embeddings",
+    # "lm_head": "output",
+}
+
+
+def translate_tensor_name(t: str) -> str:
+    match = re.match(r".*layers\.(\d+)\.(\w+\.\w+)\.lora_(A|B)\.weight", t)
+    if match:
+        nn = match.group(1)
+        sub_layer = match.group(2)
+        lora_type = match.group(3)
+
+        sub_layer_renamed = HF_SUBLAYER_TO_GGML.get(sub_layer)
+        if sub_layer_renamed is None:
+            print(f"Error: unrecognized sub-layer {sub_layer} in tensor {t}")
+            sys.exit(1)
+
+        output_string = (
+            f"layers.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}"
+        )
+        return output_string
+    else:
+        print(f"Error: unrecognized tensor {t}")
+        sys.exit(1)
+
+
+def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
+    fout.write(b"ggla"[::-1])  # magic (ggml lora)
+    fout.write(struct.pack("i", 1))  # file version
+    fout.write(struct.pack("ii", params["r"], params["lora_alpha"]))
+
+
+def write_tensor_header(
+    self, name: str, shape: Sequence[int], data_type: DataType
+) -> None:
+    sname = name.encode("utf-8")
+    fout.write(
+        struct.pack(
+            "iii",
+            len(shape),
+            len(sname),
+            DATA_TYPE_TO_FTYPE[NUMPY_TYPE_TO_DATA_TYPE[data_type]],
+        )
+    )
+    fout.write(struct.pack("i" * len(shape), *shape[::-1]))
+    fout.write(sname)
+    fout.seek((fout.tell() + 31) & -32)
+
+
+if len(sys.argv) != 2:
+    print(f"Usage: python {sys.argv[0]} <path>")
+    print(
+        "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
+    )
+    sys.exit(1)
+
+input_json = os.path.join(sys.argv[1], "adapter_config.json")
+input_model = os.path.join(sys.argv[1], "adapter_model.bin")
+output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
+
+model = torch.load(input_model, map_location="cpu")
+
+with open(input_json, "r") as f:
+    params = json.load(f)
+
+if params["peft_type"] != "LORA":
+    print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
+    sys.exit(1)
+
+if params["fan_in_fan_out"] == True:
+    print("Error: param fan_in_fan_out is not supported")
+    sys.exit(1)
+
+if params["bias"] is not None and params["bias"] != "none":
+    print("Error: param bias is not supported")
+    sys.exit(1)
+
+# TODO: these seem to be layers that have been trained but without lora.
+# doesn't seem widely used but eventually should be supported
+if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
+    print("Error: param modules_to_save is not supported")
+    sys.exit(1)
+
+with open(output_path, "wb") as fout:
+    fout.truncate()
+
+    write_file_header(fout, params)
+    for k, v in model.items():
+        if k.endswith("lora_A.weight"):
+            if v.dtype != torch.float16 and v.dtype != torch.float32:
+                v = v.float()
+            v = v.T
+        else:
+            v = v.float()
+
+        t = v.numpy()
+        tname = translate_tensor_name(k)
+        print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
+        write_tensor_header(fout, tname, t.shape, t.dtype)
+        t.tofile(fout)
+
+print(f"Converted {input_json} and {input_model} to {output_path}")

convert-pth-to-ggml.py

@@ -1,274 +1,11 @@
-# Convert a LLaMA model checkpoint to a ggjt compatible file
-#
-# Load the model using Torch
-# Iterate over all variables and write them to a binary file.
-#
-# For each variable, write the following:
-#   - Number of dimensions (int)
-#   - Name length (int)
-#   - Dimensions (int[n_dims])
-#   - Name (char[name_length])
-#   - Data (float[n_dims])
-#
-# At the start of the ggml file we write the model parameters
-# and vocabulary.
-#
+# Compatibility stub
 
 import argparse
-import os
-import sys
-import json
-import struct
-import numpy as np
-import torch
 
-from sentencepiece import SentencePieceProcessor
+import convert
 
-QK = 32
-
-GGML_TYPE_Q4_0  = 0
-GGML_TYPE_Q4_1  = 1
-GGML_TYPE_I8    = 2
-GGML_TYPE_I16   = 3
-GGML_TYPE_I32   = 4
-GGML_TYPE_F16   = 5
-GGML_TYPE_F32   = 6
-
-WTYPES = {
-    0: GGML_TYPE_F32,
-    1: GGML_TYPE_F16,
-    2: GGML_TYPE_Q4_0,
-    3: GGML_TYPE_Q4_1,
-}
-
-GGML_BLCK_SIZE = {
-    GGML_TYPE_Q4_0:  QK,
-    GGML_TYPE_Q4_1:  QK,
-    GGML_TYPE_I8:    1,
-    GGML_TYPE_I16:   1,
-    GGML_TYPE_I32:   1,
-    GGML_TYPE_F16:   1,
-    GGML_TYPE_F32:   1,
-}
-
-GGML_TYPE_SIZE = {
-    GGML_TYPE_Q4_0: 4   + QK//2,
-    GGML_TYPE_Q4_1: 4*2 + QK//2,
-    GGML_TYPE_I8:   1,
-    GGML_TYPE_I16:  2,
-    GGML_TYPE_I32:  4,
-    GGML_TYPE_F16:  2,
-    GGML_TYPE_F32:  4,
-}
-
-def ggml_nelements(shape):
-    r = 1
-    for i in shape:
-        r *= i
-    return r
-
-def ggml_nbytes(shape, ftype):
-    x = ggml_nelements(shape)
-    t = WTYPES[ftype]
-    x *= GGML_TYPE_SIZE[t]
-    x //= GGML_BLCK_SIZE[t]
-    return x
-
-def parse_args():
-    parser = argparse.ArgumentParser(description='Convert a LLaMA model checkpoint to a ggml compatible file')
-    parser.add_argument('dir_model',  help='directory containing the model checkpoint')
-    parser.add_argument('ftype',      help='file type (0: float32, 1: float16)', type=int, choices=[0, 1], default=1)
-    parser.add_argument('vocab_only', help='only write vocab to file', type=int, default=0, nargs='?')
-    return parser.parse_args()
-
-def get_n_parts(dim):
-    mappings = {4096: 1, 5120: 2, 6656: 4, 8192: 8}
-    n_parts = mappings.get(dim)
-    if n_parts is None:
-        print(f"Invalid dim: {dim}")
-        sys.exit(1)
-
-    print(f"n_parts = {n_parts}\n")
-    return n_parts
-
-def load_hparams_and_tokenizer(dir_model):
-    # `dir_model` is something like `models/7B` or `models/7B/`.
-    # "tokenizer.model" is expected under model's parent dir.
-    # When `dir_model` is a symlink, f"{dir_model}/../tokenizer.model" would not be found.
-    # Let's use the model's parent dir directly.
-    model_parent_dir = os.path.dirname(os.path.normpath(dir_model))
-    fname_hparams = f"{dir_model}/params.json"
-    fname_tokenizer = f"{model_parent_dir}/tokenizer.model"
-    with open(fname_hparams, "r") as f:
-        hparams = json.load(f)
-        print(hparams)
-    tokenizer = SentencePieceProcessor(fname_tokenizer)
-    hparams.update({"vocab_size": tokenizer.vocab_size()})
-    return hparams, tokenizer
-
-def write_header(fout, hparams, ftype):
-    keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"]
-    values = [
-        0x67676a74,  # magic: ggjt in hex
-        1, # file version
-        *[hparams[key] for key in keys],
-        hparams["dim"] // hparams["n_heads"],  # rot (obsolete)
-        ftype
-    ]
-    fout.write(struct.pack("i" * len(values), *values))
-
-def write_tokens(fout, tokenizer):
-    for i in range(tokenizer.vocab_size()):
-        if tokenizer.is_unknown(i):
-            text = " \u2047 ".encode()
-        elif tokenizer.is_control(i):
-            text = b""
-        elif tokenizer.is_byte(i):
-            piece = tokenizer.id_to_piece(i)
-            if len(piece) != 6:
-                print(f"Invalid token: {piece}")
-                sys.exit(1)
-            byte_value = int(piece[3:-1], 16)
-            text = struct.pack("B", byte_value)
-        else:
-            text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
-        fout.write(struct.pack("i", len(text)))
-        fout.write(text)
-        fout.write(struct.pack("f", tokenizer.get_score(i)))
-
-def process_and_write_variables(fout, model, ftype, part_id, n_parts):
-    for name, datao in model.items():
-        if name.endswith("freqs"):
-            continue
-
-        # remove dimensions with a single element
-        data = datao.numpy().squeeze()
-        partshape = data.shape
-        n_dims = len(data.shape)
-        assert n_dims in (1, 2)
-
-        print(f"Processing variable: {name} with shape: {partshape} and type: {datao.dtype}")
-
-        # coerce single-dimensional tensors from float16 to float32
-        ftype_cur = 1
-        if ftype == 0 or n_dims == 1:
-            print("  Converting to float32")
-            data = data.astype(np.float32)
-            ftype_cur = 0
-        blck_size = GGML_BLCK_SIZE[WTYPES[ftype_cur]]
-        type_size = GGML_TYPE_SIZE[WTYPES[ftype_cur]]
-
-        # determine dimension along which multipart tensor is sharded
-        #
-        # split_dim 0 regex:
-        #   - output.*
-        #   - layers.*.attention.wq.weight
-        #   - layers.*.attention.wk.weight
-        #   - layers.*.attention.wv.weight
-        #   - layers.*.feed_forward.w1.weight
-        #   - layers.*.feed_forward.w3.weight
-        #
-        # split_dim 1 regex:
-        #   - tok_embeddings.*
-        #   - layers.*.attention.wo.weight
-        #   - layers.*.feed_forward.w2.weight
-        #
-        if n_dims > 1:
-            split_dim = 1
-            if "tok_embeddings" in name:
-                split_dim = 1
-            elif "layers" in name:
-                if "attention.wo.weight" in name:
-                    split_dim = 1
-                elif "feed_forward.w2.weight" in name:
-                    split_dim = 1
-                else:
-                    split_dim = 0
-            elif "output" in name:
-                split_dim = 0
-
-        # output tensor header
-        fullshape = list(partshape)
-        if n_dims > 1:
-            fullshape[split_dim] *= n_parts
-        sname = name.encode()
-        fout.write(struct.pack("iii", n_dims, len(sname), ftype_cur))
-        for dim in reversed(fullshape):
-            fout.write(struct.pack("i", dim))
-        fout.write(sname)
-
-        # ensure tensor data is aligned
-        tensor_data_offset = fout.tell()
-        while tensor_data_offset % QK != 0:
-            fout.write(struct.pack("B", 0))
-            tensor_data_offset += 1
-
-        # output unified mappable tensor data
-        if n_dims == 1 or n_parts == 1:
-            # copy tensor which we thankfully received in one piece
-            if part_id == 0:
-                data.tofile(fout)
-        elif split_dim == 0:
-            # reassemble multifile tensor containing some of the rows
-            rows_per_chunk = partshape[0]
-            current_row = part_id * rows_per_chunk
-            bytes_per_row = fullshape[1] // blck_size * type_size
-            offset = current_row * bytes_per_row
-            fout.seek(tensor_data_offset + offset)
-            data.tofile(fout)
-        elif split_dim == 1:
-            # reassemble multifile tensor containing some of the cols
-            cols_per_chunk = partshape[1]
-            current_col = part_id * cols_per_chunk
-            bytes_per_row = fullshape[1] // blck_size * type_size
-            offset_current_col = current_col // blck_size * type_size
-            for row in range(partshape[0]):
-                offset_row = row * bytes_per_row
-                offset = offset_row + offset_current_col
-                fout.seek(tensor_data_offset + offset)
-                data[row].tofile(fout)
-
-        # advance file position to next tensor
-        fout.seek(tensor_data_offset + ggml_nbytes(fullshape, ftype_cur))
-
-def main():
-    args = parse_args()
-    dir_model = args.dir_model
-    ftype = args.ftype
-    ftype_str = ["f32", "f16"]
-    hparams, tokenizer = load_hparams_and_tokenizer(dir_model)
-
-    print(args)
-
-    # if only writing vocab to file
-    if args.vocab_only:
-        fname_model = f"{dir_model}/consolidated.00.pth"
-        fname_out = f"{dir_model}/ggml-vocab.bin"
-        print(f"Extracting only the vocab from '{fname_model}'\n")
-        with open(fname_out, "wb") as fout:
-            write_header(fout, hparams, ftype)
-            write_tokens(fout, tokenizer)
-        print(f"Done. Output file: {fname_out}\n")
-        return
-
-    n_parts = get_n_parts(hparams["dim"])
-    fname_out = f"{dir_model}/ggml-model-{ftype_str[ftype]}.bin"
-
-    # we output a single file for ggml
-    with open(fname_out, "wb") as fout:
-        write_header(fout, hparams, ftype)
-        write_tokens(fout, tokenizer)
-        offset_of_tensors = fout.tell()
-        # the tensors we load could be split across multiple files
-        for part_id in range(n_parts):
-            fout.seek(offset_of_tensors)
-            print(f"Processing part {part_id+1} of {n_parts}\n")
-            fname_model = f"{dir_model}/consolidated.0{part_id}.pth"
-            model = torch.load(fname_model, map_location="cpu")
-            process_and_write_variables(fout, model, ftype, part_id, n_parts)
-            del model
-
-    print(f"Done. Output file: {fname_out}\n")
-
-if __name__ == "__main__":
-    main()
+parser = argparse.ArgumentParser(description='Convert a LLaMA model checkpoint to a ggml compatible file')
+parser.add_argument('dir_model',  help='directory containing the model checkpoint')
+parser.add_argument('ftype',      help='file type (0: float32, 1: float16)', type=int, choices=[0, 1], default=1)
+args = parser.parse_args()
+convert.main(['--outtype', 'f16' if args.ftype == 1 else 'f32', '--', args.dir_model])

convert-unversioned-ggml-to-ggml.py

@@ -1,100 +0,0 @@
-#!/usr/bin/env python3
-# Original by https://github.com/eiz
-# https://github.com/ggerganov/llama.cpp/issues/324#issuecomment-1476227818
-import argparse
-import glob
-import os
-import struct
-import sys
-from sentencepiece import SentencePieceProcessor
-
-HPARAMS = keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"]
-
-def parse_args():
-    parser = argparse.ArgumentParser(description='Upgrade old ggml model files to the current format')
-    parser.add_argument('dir_model', help='directory containing ggml .bin files')
-    parser.add_argument('tokenizer_model', help='path to LLaMA tokenizer.model file')
-    return parser.parse_args()
-
-def read_header(f_in):
-    struct_fmt = "i" * (3 + len(HPARAMS))
-    struct_size = struct.calcsize(struct_fmt)
-    buf = f_in.read(struct_size)
-    return struct.unpack(struct_fmt, buf)
-
-def write_header(f_out, header):
-    (magic, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype) = header
-
-    if magic != 0x67676d6c:
-        raise Exception('Invalid file magic. Must be an old style ggml file.')
-
-    values = [
-        0x67676d66,  # magic: ggml in hex
-        1, # file version
-        vocab_size,
-        dim,
-        multiple_of,
-        n_heads,
-        n_layers,
-        rot,
-        ftype
-    ]
-    f_out.write(struct.pack("i" * len(values), *values))
-
-def write_tokens(fout, tokenizer):
-    for i in range(tokenizer.vocab_size()):
-        if tokenizer.is_unknown(i):
-            text = " \u2047 ".encode()
-        elif tokenizer.is_control(i):
-            text = b""
-        elif tokenizer.is_byte(i):
-            piece = tokenizer.id_to_piece(i)
-            if len(piece) != 6:
-                print(f"Invalid token: {piece}")
-                sys.exit(1)
-            byte_value = int(piece[3:-1], 16)
-            text = struct.pack("B", byte_value)
-        else:
-            text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
-        fout.write(struct.pack("i", len(text)))
-        fout.write(text)
-        fout.write(struct.pack("f", tokenizer.get_score(i)))
-
-def read_tokens(f_in, tokenizer):
-    for i in range(tokenizer.vocab_size()):
-        len_b = f_in.read(4)
-        (length,) = struct.unpack("i", len_b)
-        f_in.read(length)
-
-def copy_all_data(f_out, f_in):
-    while True:
-        buf = f_in.read(1024 * 1024)
-        if not buf:
-            break
-        f_out.write(buf)
-
-def convert_one_file(path_in, tokenizer):
-    path_tmp = f"{path_in}.tmp"
-    path_orig= f"{path_in}.orig"
-    print(f"converting {path_in}")
-    with open(path_in, "rb") as f_in, open(path_tmp, "wb") as f_out:
-        write_header(f_out, read_header(f_in))
-        read_tokens(f_in, tokenizer)
-        write_tokens(f_out, tokenizer)
-        copy_all_data(f_out, f_in)
-    os.rename(path_in, path_orig)
-    os.rename(path_tmp, path_in)
-
-def main():
-    args = parse_args()
-    files = []
-    files.extend(glob.glob(f"{args.dir_model}/*.bin"))
-    files.extend(glob.glob(f"{args.dir_model}/*.bin.*"))
-
-    tokenizer = SentencePieceProcessor(args.tokenizer_model)
-
-    for file in files:
-        convert_one_file(file, tokenizer)
-
-if __name__ == "__main__":
-    main()

examples/benchmark/benchmark-q4_0-matmult.c

@@ -24,7 +24,7 @@
 
 float tensor_sum_elements(struct ggml_tensor * tensor) {
     float sum = 0;
-    if (tensor->type==6) {
+    if (tensor->type==GGML_TYPE_F32) {
         for (int j = 0; j < tensor->ne[1]; j++) {
             for (int k = 0; k < tensor->ne[0]; k++) {
                 sum +=  ((float *) tensor->data)[j*tensor->ne[0]+k];

examples/common.cpp

@@ -7,12 +7,6 @@
 #include <iterator>
 #include <algorithm>
 
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <malloc.h> // using malloc.h with MSC/MINGW
-#elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
-#include <alloca.h>
-#endif
-
 #if defined (_WIN32)
 #include <fcntl.h>
 #include <io.h>
@@ -145,6 +139,19 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.model = argv[i];
+        } else if (arg == "--lora") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter = argv[i];
+            params.use_mmap = false;
+        } else if (arg == "--lora-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.lora_base = argv[i];
         } else if (arg == "-i" || arg == "--interactive") {
             params.interactive = true;
         } else if (arg == "--embedding") {
@@ -248,6 +255,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     }
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
     fprintf(stderr, "  --verbose-prompt      print prompt before generation\n");
+    fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
     fprintf(stderr, "  -m FNAME, --model FNAME\n");
     fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
     fprintf(stderr, "\n");

examples/common.h

@@ -31,11 +31,12 @@ struct gpt_params {
 
     std::string model  = "models/lamma-7B/ggml-model.bin"; // model path
     std::string prompt = "";
-    std::string input_prefix = ""; // string to prefix user inputs with
-
-
+    std::string input_prefix = "";       // string to prefix user inputs with
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
 
+    std::string lora_adapter = "";  // lora adapter path
+    std::string lora_base = "";     // base model path for the lora adapter
+
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs

examples/embedding/embedding.cpp

@@ -1,6 +1,8 @@
 #include "common.h"
 #include "llama.h"
 
+#include <ctime>
+
 int main(int argc, char ** argv) {
     gpt_params params;
     params.model = "models/llama-7B/ggml-model.bin";

examples/main/main.cpp

@@ -11,6 +11,7 @@
 #include <cmath>
 #include <cstdio>
 #include <cstring>
+#include <ctime>
 #include <fstream>
 #include <iostream>
 #include <string>
@@ -113,6 +114,17 @@ int main(int argc, char ** argv) {
         }
     }
 
+    if (!params.lora_adapter.empty()) {
+        int err = llama_apply_lora_from_file(ctx,
+                                             params.lora_adapter.c_str(),
+                                             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__);
+            return 1;
+        }
+    }
+
     // print system information
     {
         fprintf(stderr, "\n");

examples/perplexity/perplexity.cpp

@@ -2,6 +2,7 @@
 #include "llama.h"
 
 #include <cmath>
+#include <ctime>
 
 std::vector<float> softmax(const std::vector<float>& logits) {
     std::vector<float> probs(logits.size());
@@ -27,20 +28,27 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
 
     int count = 0;
     int seq_count = tokens.size() / params.n_ctx;
+    int n_vocab = llama_n_vocab(ctx);
 
     double nll = 0.0;
-
-    fprintf(stderr, "%s : calculating perplexity over %d chunks\n", __func__, seq_count);
+    fprintf(stderr, "%s : calculating perplexity over %d chunks, batch_size=%d\n", __func__, seq_count, params.n_batch);
 
     for (int i = 0; i < seq_count; ++i) {
         int start = i * params.n_ctx;
-        int end = start + params.n_ctx - 1; // TODO: this is not optimal, e.g. it makes the batch 511 instead of 512
-                                            //       it is better to always be power of 2 for better performance
-        std::vector<llama_token> embd(tokens.begin() + start, tokens.begin() + end);
+        int end = start + params.n_ctx;
+
+        std::vector<float> logits;
+        int num_batches = (params.n_ctx + params.n_batch - 1) / params.n_batch;
         auto start_t = std::chrono::high_resolution_clock::now();
-        if (llama_eval(ctx, embd.data(), embd.size(), 0, params.n_threads)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return;
+        for (int j = 0; j < num_batches; ++j) {
+            int batch_start = start + j * params.n_batch;
+            int batch_size = std::min(end - batch_start, params.n_batch);
+            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * params.n_batch, params.n_threads)) {
+                fprintf(stderr, "%s : failed to eval\n", __func__);
+                return;
+            }
+            auto batch_logits = llama_get_logits(ctx);
+            logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
         }
         auto end_t = std::chrono::high_resolution_clock::now();
         if (i == 0) {
@@ -59,15 +67,12 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
         // 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.
-
-        auto logits = llama_get_logits(ctx);
-        for (int j = params.n_ctx / 2; j < params.n_ctx - 1; ++j) {
+        for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) {
             // Calculate probability of next token, given the previous ones.
-            int n_vocab = llama_n_vocab(ctx);
             std::vector<float> tok_logits(
-                logits + j * n_vocab,
-                logits + (j + 1) * n_vocab);
-            const float prob = softmax(tok_logits)[tokens[start + j + 1]];
+                logits.begin() + j * n_vocab,
+                logits.begin() + (j + 1) * n_vocab);
+            float prob = softmax(tok_logits)[tokens[start + j + 1]];
             nll += -std::log(prob);
             ++count;
         }
@@ -82,11 +87,13 @@ int main(int argc, char ** argv) {
     gpt_params params;
     params.model = "models/llama-7B/ggml-model.bin";
 
+    params.n_batch = 512;
     if (gpt_params_parse(argc, argv, params) == false) {
         return 1;
     }
 
     params.perplexity = true;
+    params.n_batch = std::min(params.n_batch, params.n_ctx);
 
     if (params.n_ctx > 2048) {
         fprintf(stderr, "%s: warning: model does not support context sizes greater than 2048 tokens (%d specified);"
@@ -127,6 +134,17 @@ int main(int argc, char ** argv) {
         }
     }
 
+    if (!params.lora_adapter.empty()) {
+        int err = llama_apply_lora_from_file(ctx,
+                                             params.lora_adapter.c_str(),
+                                             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__);
+            return 1;
+        }
+    }
+
     // print system information
     {
         fprintf(stderr, "\n");

examples/quantize-stats/quantize-stats.cpp

@@ -16,9 +16,6 @@
 #include <unordered_map>
 #include <vector>
 
-static const char * type_strs[] = { "q4_0", "q4_1", "i8", "i16", "i32", "f16", "f32"  };
-static_assert(sizeof(type_strs) == GGML_TYPE_COUNT * sizeof(char *), "Incomplete type list");
-
 struct quantize_stats_params {
     std::string model = "models/7B/ggml-model-f16.bin";
     bool verbose = false;
@@ -224,7 +221,7 @@ int main(int argc, char ** argv) {
                 break;
             }
             int j;
-            for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], type_strs[j]) != 0; j++) {
+            for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], ggml_type_name((ggml_type) j)) != 0; j++) {
                 // find match
             }
             if (j < GGML_TYPE_COUNT) {
@@ -279,7 +276,7 @@ int main(int argc, char ** argv) {
             continue;
         }
         if (params.verbose) {
-            printf("%s: type %s, size %" PRId64 "\n", kv_tensor.first.c_str(), type_strs[kv_tensor.second->type], ggml_nelements(kv_tensor.second));
+            printf("%s: type %s, size %" PRId64 "\n", kv_tensor.first.c_str(), ggml_type_name(kv_tensor.second->type), ggml_nelements(kv_tensor.second));
         }
         if (kv_tensor.second->type == GGML_TYPE_F16) {
             is_f16 = true;
@@ -304,13 +301,14 @@ int main(int argc, char ** argv) {
 
     // loop throught quantization types
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
+        const ggml_type type = (ggml_type) i;
         if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
             continue;
         }
         quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
         if (qfns.quantize_row_q && qfns.dequantize_row_q) {
             if (params.verbose) {
-                printf("testing %s ...\n",  type_strs[i]);
+                printf("testing %s ...\n",  ggml_type_name(type));
             }
 
             error_stats global_stats {};
@@ -322,7 +320,7 @@ int main(int argc, char ** argv) {
                 if (params.verbose) {
                     printf("  %s ...\n",  kv_tensor.first.c_str());
                 }
-                std::string layer_name { type_strs[i] };
+                std::string layer_name { ggml_type_name(type) };
                 layer_name += "::" + kv_tensor.first;
                 test_roundtrip_on_layer(
                         layer_name,
@@ -337,7 +335,7 @@ int main(int argc, char ** argv) {
                 );
             }
 
-            print_error_stats(type_strs[i], global_stats, params.print_histogram);
+            print_error_stats(ggml_type_name(type), global_stats, params.print_histogram);
         }
     }
 

examples/quantize/quantize.cpp

@@ -14,6 +14,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
         fprintf(stderr, "  type = %d - q4_0\n", LLAMA_FTYPE_MOSTLY_Q4_0);
         fprintf(stderr, "  type = %d - q4_1\n", LLAMA_FTYPE_MOSTLY_Q4_1);
+        fprintf(stderr, "  type = %d - q4_2\n", LLAMA_FTYPE_MOSTLY_Q4_2);
         return 1;
     }
 

flake.nix

@@ -10,7 +10,6 @@
           inherit system;
         };
         llama-python = pkgs.python310.withPackages (ps: with ps; [
-          torch
           numpy
           sentencepiece
         ]);

ggml-cuda.cu

@@ -0,0 +1,116 @@
+#include <stdint.h>
+#include <cuda_fp16.h>
+#include "ggml-cuda.h"
+
+typedef uint16_t ggml_fp16_t;
+static_assert(sizeof(__half) == sizeof(ggml_fp16_t), "wrong fp16 size");
+
+#define QK4_0 32
+typedef struct {
+    float   d;              // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    float   d;              // delta
+    float   m;              // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK4_2 16
+typedef struct {
+    __half d;               // delta
+    uint8_t qs[QK4_2 / 2];  // nibbles / quants
+} block_q4_2;
+static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding");
+
+
+static __global__ void dequantize_block_q4_0(const void * vx, float * y) {
+    const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_0; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = (vi0 - 8)*d;
+        const float v1 = (vi1 - 8)*d;
+
+        y[i*QK4_0 + l + 0] = v0;
+        y[i*QK4_0 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q4_1(const void * vx, float * y) {
+    const block_q4_1 * x = (const block_q4_1 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+    const float m = x[i].m;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_1; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = vi0*d + m;
+        const float v1 = vi1*d + m;
+
+        y[i*QK4_1 + l + 0] = v0;
+        y[i*QK4_1 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q4_2(const void * vx, float * y) {
+    const block_q4_2 * x = (const block_q4_2 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_2; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = (vi0 - 8)*d;
+        const float v1 = (vi1 - 8)*d;
+
+        y[i*QK4_2 + l + 0] = v0;
+        y[i*QK4_2 + l + 1] = v1;
+    }
+}
+
+extern "C" {
+    __host__ void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+        const int nb = k / QK4_0;
+        dequantize_block_q4_0<<<nb, 1, 0, stream>>>(vx, y);
+    }
+
+    __host__ void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+        const int nb = k / QK4_1;
+        dequantize_block_q4_1<<<nb, 1, 0, stream>>>(vx, y);
+    }
+
+    __host__ void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+        const int nb = k / QK4_2;
+        dequantize_block_q4_2<<<nb, 1, 0, stream>>>(vx, y);
+    }
+}

ggml-cuda.h

@@ -0,0 +1,11 @@
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+
+#ifdef  __cplusplus
+}
+#endif

ggml.h

@@ -177,11 +177,12 @@ extern "C" {
 #include <stddef.h>
 #include <stdbool.h>
 
-#define GGML_MAX_DIMS     4
-#define GGML_MAX_NODES    4096
-#define GGML_MAX_PARAMS   16
-#define GGML_MAX_CONTEXTS 64
-#define GGML_MAX_OPT      4
+#define GGML_MAX_DIMS          4
+#define GGML_MAX_NODES         4096
+#define GGML_MAX_PARAMS        16
+#define GGML_MAX_CONTEXTS      64
+#define GGML_MAX_OPT           4
+#define GGML_DEFAULT_N_THREADS 4
 
 #ifdef __ARM_NEON
 // we use the built-in 16-bit float type
@@ -203,6 +204,8 @@ enum ggml_type {
     GGML_TYPE_F16  = 1,
     GGML_TYPE_Q4_0 = 2,
     GGML_TYPE_Q4_1 = 3,
+    GGML_TYPE_Q4_2 = 4,
+    GGML_TYPE_Q8_0 = 5,
     GGML_TYPE_I8,
     GGML_TYPE_I16,
     GGML_TYPE_I32,
@@ -252,6 +255,9 @@ enum ggml_op {
     GGML_OP_FLASH_ATTN,
     GGML_OP_FLASH_FF,
 
+    GGML_OP_MAP_UNARY,
+    GGML_OP_MAP_BINARY,
+
     GGML_OP_COUNT,
 };
 
@@ -350,6 +356,8 @@ int    ggml_blck_size (enum ggml_type type);
 size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
 float  ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
 
+const char * ggml_type_name(enum ggml_type type);
+
 size_t ggml_element_size(const struct ggml_tensor * tensor);
 
 struct ggml_context * ggml_init(struct ggml_init_params params);
@@ -423,6 +431,12 @@ struct ggml_tensor * ggml_add(
         struct ggml_tensor  * a,
         struct ggml_tensor  * b);
 
+
+struct ggml_tensor * ggml_add_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
 struct ggml_tensor * ggml_sub(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -651,6 +665,21 @@ struct ggml_tensor * ggml_flash_ff(
         struct ggml_tensor  * c0,
         struct ggml_tensor  * c1);
 
+// Mapping operations
+typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
+typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
+
+struct ggml_tensor * ggml_map_unary_f32(
+        struct ggml_context        * ctx,
+        struct ggml_tensor         * a,
+        const  ggml_unary_op_f32_t fun);
+
+struct ggml_tensor * ggml_map_binary_f32(
+        struct ggml_context         * ctx,
+        struct ggml_tensor          * a,
+        struct ggml_tensor          * b,
+        const  ggml_binary_op_f32_t fun);
+
 //
 // automatic differentiation
 //
@@ -778,6 +807,7 @@ enum ggml_opt_result ggml_opt(
 
 size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
 size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * hist);
 
 //
 // system info
@@ -786,6 +816,8 @@ size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t *
 int ggml_cpu_has_avx(void);
 int ggml_cpu_has_avx2(void);
 int ggml_cpu_has_avx512(void);
+int ggml_cpu_has_avx512_vbmi(void);
+int ggml_cpu_has_avx512_vnni(void);
 int ggml_cpu_has_fma(void);
 int ggml_cpu_has_neon(void);
 int ggml_cpu_has_arm_fma(void);
@@ -793,6 +825,7 @@ int ggml_cpu_has_f16c(void);
 int ggml_cpu_has_fp16_va(void);
 int ggml_cpu_has_wasm_simd(void);
 int ggml_cpu_has_blas(void);
+int ggml_cpu_has_cublas(void);
 int ggml_cpu_has_sse3(void);
 int ggml_cpu_has_vsx(void);
 
@@ -815,6 +848,7 @@ typedef struct {
     dequantize_row_q_t dequantize_row_q;
     quantize_row_q_t   quantize_row_q;
     quantize_row_q_t   quantize_row_q_reference;
+    quantize_row_q_t   quantize_row_q_dot;
     vec_dot_q_t        vec_dot_q;
 } quantize_fns_t;
 

llama.cpp

@@ -1,6 +1,8 @@
 // Defines fileno on msys:
 #ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#include <cstdint>
+#include <cstdio>
 #endif
 
 #include "llama_util.h"
@@ -9,6 +11,7 @@
 #include "ggml.h"
 
 #include <array>
+#include <ctime>
 #include <cinttypes>
 #include <fstream>
 #include <random>
@@ -41,35 +44,51 @@ static const size_t MB = 1024*1024;
 // TODO: dynamically determine these sizes
 //       needs modifications in ggml
 
-static const std::map<e_model, size_t> MEM_REQ_SCRATCH0 = {
-    { MODEL_7B,    512ull*MB },
-    { MODEL_13B,   512ull*MB },
-    { MODEL_30B,   512ull*MB },
-    { MODEL_65B,   512ull*MB },
-};
+static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0()
+{
+    static std::map<e_model, size_t> _MEM_REQ_SCRATCH0 = {
+        { MODEL_7B,    512ull * MB },
+        { MODEL_13B,   512ull * MB },
+        { MODEL_30B,   512ull * MB },
+        { MODEL_65B,   512ull * MB },
+    };
+    return _MEM_REQ_SCRATCH0;
+}
 
-static const std::map<e_model, size_t> MEM_REQ_SCRATCH1 = {
-    { MODEL_7B,    512ull*MB },
-    { MODEL_13B,   512ull*MB },
-    { MODEL_30B,   512ull*MB },
-    { MODEL_65B,   512ull*MB },
+static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
+{
+    static std::map<e_model, size_t> _MEM_REQ_SCRATCH1 = {
+        { MODEL_7B,    512ull * MB },
+        { MODEL_13B,   512ull * MB },
+        { MODEL_30B,   512ull * MB },
+        { MODEL_65B,   512ull * MB },
+    };
+    return _MEM_REQ_SCRATCH1;
 };
 
 // 2*n_embd*n_ctx*n_layer*sizeof(float16)
-static const std::map<e_model, size_t> MEM_REQ_KV_SELF = {
-    { MODEL_7B,   1026ull*MB },
-    { MODEL_13B,  1608ull*MB },
-    { MODEL_30B,  3124ull*MB },
-    { MODEL_65B,  5120ull*MB },
+static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
+{
+    static std::map<e_model, size_t> _MEM_REQ_KV_SELF = {
+        { MODEL_7B,   1026ull * MB },
+        { MODEL_13B,  1608ull * MB },
+        { MODEL_30B,  3124ull * MB },
+        { MODEL_65B,  5120ull * MB },
+    };
+    return _MEM_REQ_KV_SELF;
 };
 
 // this is mostly needed for temporary mul_mat buffers to dequantize the data
 // not actually needed if BLAS is disabled
-static const std::map<e_model, size_t> MEM_REQ_EVAL = {
-    { MODEL_7B,   768ull*MB },
-    { MODEL_13B, 1024ull*MB },
-    { MODEL_30B, 1280ull*MB },
-    { MODEL_65B, 1536ull*MB },
+static const std::map<e_model, size_t> & MEM_REQ_EVAL()
+{
+    static std::map<e_model, size_t> _MEM_REQ_EVAL = {
+        { MODEL_7B,   768ull * MB },
+        { MODEL_13B, 1024ull * MB },
+        { MODEL_30B, 1280ull * MB },
+        { MODEL_65B, 1536ull * MB },
+    };
+    return _MEM_REQ_EVAL;
 };
 
 // default hparams (LLaMA 7B)
@@ -261,22 +280,12 @@ static size_t checked_div(size_t a, size_t b) {
 }
 
 static std::string llama_format_tensor_shape(const std::vector<uint32_t> & ne) {
-    std::string ret = "[" + std::to_string(ne.at(0));
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%5u", ne.at(0));
     for (size_t i = 1; i < ne.size(); i++) {
-        ret += " x " + std::to_string(ne.at(i));
-    }
-    ret += "]";
-    return ret;
-}
-
-static const char * llama_format_type(enum ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_F32: return "f32";
-        case GGML_TYPE_F16: return "f16";
-        case GGML_TYPE_Q4_0: return "q4_0";
-        case GGML_TYPE_Q4_1: return "q4_1";
-        default: LLAMA_ASSERT(false);
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), " x %5u", ne.at(i));
     }
+    return buf;
 }
 
 static size_t llama_calc_tensor_size(const std::vector<uint32_t> & ne, enum ggml_type type) {
@@ -469,6 +478,7 @@ struct llama_file_loader {
                 case GGML_TYPE_F16:
                 case GGML_TYPE_Q4_0:
                 case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q4_2:
                     break;
                 default: {
                     throw format("unrecognized tensor type %u\n", shard.type);
@@ -541,6 +551,7 @@ struct llama_file_saver {
             case GGML_TYPE_F16:
             case GGML_TYPE_Q4_0:
             case GGML_TYPE_Q4_1:
+            case GGML_TYPE_Q4_2:
                 break;
             default: LLAMA_ASSERT(false);
         }
@@ -626,6 +637,7 @@ struct llama_model_loader {
             throw format("llama.cpp: tensor '%s' has wrong shape; expected %s, got %s",
                          name.c_str(), llama_format_tensor_shape(ne).c_str(), llama_format_tensor_shape(lt.ne).c_str());
         }
+
         return get_tensor_for(lt);
     }
 
@@ -828,6 +840,7 @@ static const char *llama_ftype_name(enum llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q4_1: return "mostly Q4_1";
         case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16:
                                       return "mostly Q4_1, some F16";
+        case LLAMA_FTYPE_MOSTLY_Q4_2: return "mostly Q4_2";
         default:                      return "unknown, may not work";
     }
 }
@@ -908,13 +921,13 @@ static void llama_model_load_internal(
         const size_t mem_required =
             ctx_size +
             mmapped_size +
-            MEM_REQ_SCRATCH0.at(model.type) +
-            MEM_REQ_SCRATCH1.at(model.type) +
-            MEM_REQ_EVAL.at    (model.type);
+            MEM_REQ_SCRATCH0().at(model.type) +
+            MEM_REQ_SCRATCH1().at(model.type) +
+            MEM_REQ_EVAL().at(model.type);
 
         // this is the memory required by one llama_state
         const size_t mem_required_state =
-            scale*MEM_REQ_KV_SELF.at(model.type);
+            scale*MEM_REQ_KV_SELF().at(model.type);
 
         fprintf(stderr, "%s: mem required  = %7.2f MB (+ %7.2f MB per state)\n", __func__,
                 mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
@@ -951,8 +964,8 @@ static void llama_model_load_internal(
         ml->ggml_ctx = ctx;
 
         model.tok_embeddings = ml->get_tensor("tok_embeddings.weight", {n_embd, n_vocab});
-        model.norm   = ml->get_tensor("norm.weight", {n_embd});
-        model.output = ml->get_tensor("output.weight", {n_embd, n_vocab});
+        model.norm           = ml->get_tensor("norm.weight",           {n_embd});
+        model.output         = ml->get_tensor("output.weight",         {n_embd, n_vocab});
 
         model.layers.resize(n_layer);
         for (uint32_t i = 0; i < n_layer; ++i) {
@@ -1056,7 +1069,7 @@ static bool llama_eval_internal(
     // for big prompts, if BLAS is enabled, it is better to use only one thread
     // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
     ggml_cgraph gf = {};
-    gf.n_threads = N >= 32 && ggml_cpu_has_blas() ? 1 : n_threads;
+    gf.n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_cublas() ? 1 : n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, tokens, N*ggml_element_size(embd));
@@ -1561,6 +1574,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     switch (ftype) {
         case LLAMA_FTYPE_MOSTLY_Q4_0: quantized_type = GGML_TYPE_Q4_0; break;
         case LLAMA_FTYPE_MOSTLY_Q4_1: quantized_type = GGML_TYPE_Q4_1; break;
+        case LLAMA_FTYPE_MOSTLY_Q4_2: quantized_type = GGML_TYPE_Q4_2; break;
         default: throw format("invalid output file type %d\n", ftype);
     };
 
@@ -1579,10 +1593,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         tensor.data = read_data.addr;
         model_loader->load_data_for(tensor);
 
-        printf("[%zu/%zu] %36s - %s, type = %6s, ",
+        printf("[%4zu/%4zu] %36s - %16s, type = %6s, ",
                ++idx, model_loader->tensors_map.tensors.size(),
                tensor.name.c_str(), llama_format_tensor_shape(tensor.ne).c_str(),
-               llama_format_type(tensor.type));
+               ggml_type_name(tensor.type));
 
         // This used to be a regex, but <regex> has an extreme cost to compile times.
         bool quantize = tensor.name.rfind("weight") == tensor.name.size() - 6; // ends with 'weight'?
@@ -1615,7 +1629,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                     f32_data[i] = ggml_fp16_to_fp32(f16_data[i]);
                 }
             } else {
-                throw format("type %s unsupported for integer quantization", llama_format_type(tensor.type));
+                throw format("type %s unsupported for integer quantization", ggml_type_name(tensor.type));
             }
 
             printf("quantizing .. ");
@@ -1634,6 +1648,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                     {
                         new_size = ggml_quantize_q4_1(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
                     } break;
+                case GGML_TYPE_Q4_2:
+                    {
+                        new_size = ggml_quantize_q4_2(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
+                    } break;
                 default:
                     LLAMA_ASSERT(false);
             }
@@ -1741,10 +1759,10 @@ struct llama_context * llama_init_from_file(
             ctx->embedding.resize(hparams.n_embd);
         }
 
-        ctx->buf_compute.resize(MEM_REQ_EVAL.at(ctx->model.type));
+        ctx->buf_compute.resize(MEM_REQ_EVAL().at(ctx->model.type));
 
-        ctx->buf_scratch[0].resize(MEM_REQ_SCRATCH0.at(ctx->model.type));
-        ctx->buf_scratch[1].resize(MEM_REQ_SCRATCH1.at(ctx->model.type));
+        ctx->buf_scratch[0].resize(MEM_REQ_SCRATCH0().at(ctx->model.type));
+        ctx->buf_scratch[1].resize(MEM_REQ_SCRATCH1().at(ctx->model.type));
     }
 
     return ctx;
@@ -1767,6 +1785,254 @@ int llama_model_quantize(
     }
 }
 
+int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char * path_lora, const char * path_base_model, int n_threads) {
+    fprintf(stderr, "%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
+
+    auto & model = ctx->model;
+
+    const int64_t t_start_lora_us = ggml_time_us();
+
+    auto fin = std::ifstream(path_lora, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, path_lora);
+        return 1;
+    }
+
+    // verify magic and version
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != 'ggla') {
+            fprintf(stderr, "%s: bad file magic\n", __func__);
+            return 1;
+        }
+        uint32_t format_version;
+        fin.read((char *) &format_version, sizeof(format_version));
+
+        if (format_version != 1) {
+            fprintf(stderr, "%s: unsupported file version\n", __func__ );
+            return 1;
+        }
+    }
+
+    int32_t lora_r;
+    int32_t lora_alpha;
+    fin.read((char *) &lora_r, sizeof(lora_r));
+    fin.read((char *) &lora_alpha, sizeof(lora_alpha));
+    float scaling = (float)lora_alpha / (float)lora_r;
+
+    fprintf(stderr, "%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
+
+
+    // create a temporary ggml context to store the lora tensors
+    // todo: calculate size from biggest possible tensor
+    std::vector<uint8_t> lora_buf(1024ull * 1024ull * 1024ull);
+    struct ggml_init_params params;
+    params.mem_size   = lora_buf.size();
+    params.mem_buffer = lora_buf.data();
+    params.no_alloc   = false;
+
+    ggml_context * lora_ctx = ggml_init(params);
+    std::unordered_map<std::string, struct ggml_tensor *> lora_tensors;
+
+    // create a name -> tensor map of the model to accelerate lookups
+    std::unordered_map<std::string, struct ggml_tensor*> model_tensors;
+    for (auto & kv: model.tensors_by_name) {
+        model_tensors.insert(kv);
+    }
+
+
+    // load base model
+    std::unique_ptr<llama_model_loader> model_loader;
+    ggml_context * base_ctx = NULL;
+    llama_buffer base_buf;
+    if (path_base_model) {
+        fprintf(stderr, "%s: loading base model from '%s'\n", __func__, path_base_model);
+        model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*vocab_only*/ false));
+
+        size_t ctx_size, mmapped_size;
+        model_loader->calc_sizes(&ctx_size, &mmapped_size);
+        base_buf.resize(ctx_size);
+
+        ggml_init_params base_params;
+        base_params.mem_size   = base_buf.size;
+        base_params.mem_buffer = base_buf.addr;
+        base_params.no_alloc   = model_loader->use_mmap;
+
+        base_ctx = ggml_init(base_params);
+
+        model_loader->ggml_ctx = base_ctx;
+
+        // maybe this should in llama_model_loader
+        if (model_loader->use_mmap) {
+            model_loader->mapping.reset(new llama_mmap(&model_loader->file_loaders.at(0)->file, /* prefetch */ false));
+        }
+    }
+
+    // read tensors and apply
+    bool warned = false;
+    int n_tensors = 0;
+    while (true) {
+        int32_t n_dims;
+        int32_t length;
+        int32_t ftype;
+
+        fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+        fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+        fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+        if (fin.eof()) {
+            break;
+        }
+
+        int32_t ne[2] = { 1, 1 };
+        for (int i = 0; i < n_dims; ++i) {
+            fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+        }
+
+        std::string name(length, 0);
+        fin.read(&name[0], length);
+
+        // check for lora suffix and get the type of tensor
+        const std::string lora_suffix = ".lora";
+        size_t pos = name.rfind(lora_suffix);
+        if (pos == std::string::npos) {
+            fprintf(stderr, "%s: error: '%s' is not a lora tensor\n", __func__, name.c_str());
+            return 1;
+        }
+
+        std::string lora_type = name.substr(pos + lora_suffix.length());
+        std::string base_name = name;
+        base_name.erase(pos);
+        // fprintf(stderr, "%s: %s => %s (lora type %s) ", __func__, name.c_str(),base_name.c_str(), lora_type.c_str());
+
+        if (model_tensors.find(base_name.data()) == model_tensors.end()) {
+            fprintf(stderr, "%s: unknown tensor '%s' in lora adapter\n", __func__, name.data());
+            return 1;
+        }
+
+        // create ggml tensor
+        ggml_type wtype;
+        switch (ftype) {
+            case 0: wtype = GGML_TYPE_F32;  break;
+            case 1: wtype = GGML_TYPE_F16;  break;
+            default:
+                    {
+                        fprintf(stderr, "%s: invalid tensor data type '%d'\n",
+                                __func__, ftype);
+                        return false;
+                    }
+        }
+        ggml_tensor* lora_tensor;
+        if (n_dims == 2) {
+            lora_tensor = ggml_new_tensor_2d(lora_ctx, wtype, ne[0], ne[1]);
+        }
+        else {
+            fprintf(stderr, "%s: unsupported tensor dimension %d\n", __func__, n_dims);
+            return 1;
+        }
+
+        // load tensor data
+        size_t offset = fin.tellg();
+        size_t tensor_data_size = ggml_nbytes(lora_tensor);
+        offset = (offset + 31) & -32;
+        fin.seekg(offset);
+        fin.read((char*)lora_tensor->data, tensor_data_size);
+
+        lora_tensors[name] = lora_tensor;
+
+        // check if we have both A and B tensors and apply
+        if (lora_tensors.find(base_name + ".loraA") != lora_tensors.end() &&
+            lora_tensors.find(base_name + ".loraB") != lora_tensors.end()) {
+
+            ggml_tensor * dest_t = model_tensors[base_name];
+            ggml_tensor * base_t;
+            if (model_loader) {
+                // load from base model
+                if (model_loader->tensors_map.name_to_idx.find(base_name) == model_loader->tensors_map.name_to_idx.end()) {
+                    fprintf(stderr, "%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
+                    return 1;
+                }
+                size_t idx = model_loader->tensors_map.name_to_idx[base_name];
+                llama_load_tensor & lt = model_loader->tensors_map.tensors[idx];
+                base_t = model_loader->get_tensor(base_name, { (uint32_t)dest_t->ne[0], (uint32_t)dest_t->ne[1] });
+                lt.data = (uint8_t *) lt.ggml_tensor->data;
+                model_loader->load_data_for(lt);
+                lt.ggml_tensor->data = lt.data;
+            }
+            else {
+                base_t = dest_t;
+            }
+
+            if (base_t->type == GGML_TYPE_Q4_0 || base_t->type == GGML_TYPE_Q4_1 || base_t->type == GGML_TYPE_Q4_2) {
+                if (!warned) {
+                    fprintf(stderr, "%s: warning: using a lora adapter with a quantized model may result in poor quality, "
+                                    "use a f16 or f32 base model with --lora-base\n", __func__);
+                    warned = true;
+                }
+            }
+
+            ggml_tensor * loraA = lora_tensors[base_name + ".loraA"];
+            ggml_tensor * loraB = lora_tensors[base_name + ".loraB"];
+
+            if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
+                fprintf(stderr, "%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
+                               " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]);
+                return 1;
+            }
+
+            // w = w + BA*s
+            ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB);
+
+            if (scaling != 1.0f) {
+                ggml_tensor * scale_tensor = ggml_new_f32(lora_ctx, scaling);
+                BA = ggml_scale(lora_ctx, BA, scale_tensor);
+            }
+
+            ggml_tensor * r;
+            if (base_t == dest_t) {
+                r = ggml_add_inplace(lora_ctx, dest_t, BA);
+            }
+            else {
+                r = ggml_add(lora_ctx, base_t, BA);
+                r = ggml_cpy(lora_ctx, r, dest_t);
+            }
+
+            struct ggml_cgraph gf = ggml_build_forward(r);
+            gf.n_threads = n_threads;
+            ggml_graph_compute(lora_ctx, &gf);
+
+            // we won't need these tensors again, reset the context to save memory
+            ggml_free(lora_ctx);
+            lora_ctx = ggml_init(params);
+            lora_tensors.clear();
+
+            n_tensors++;
+            if (n_tensors % 4 == 0)
+                fprintf(stderr, ".");
+        }
+    }
+
+    // TODO: this should be in a destructor, it will leak on failure
+    ggml_free(lora_ctx);
+    if (base_ctx) {
+        ggml_free(base_ctx);
+    }
+
+    const int64_t t_lora_us = ggml_time_us() - t_start_lora_us;
+    fprintf(stderr, " done (%.2f ms)\n", t_lora_us / 1000.0);
+
+    return 0;
+}
+
+int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lora, const char * path_base_model, int n_threads) {
+    try {
+        return llama_apply_lora_from_file_internal(ctx, path_lora, path_base_model, n_threads);
+    } catch (const std::string & err) {
+        fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.c_str());
+        return 1;
+    }
+}
+
 // Returns the KV cache that will contain the context for the
 // ongoing prediction with the model.
 const uint8_t * llama_get_kv_cache(struct llama_context * ctx) {
@@ -1924,18 +2190,20 @@ const char * llama_print_system_info(void) {
     static std::string s;
 
     s  = "";
-    s += "AVX = "       + std::to_string(ggml_cpu_has_avx())       + " | ";
-    s += "AVX2 = "      + std::to_string(ggml_cpu_has_avx2())      + " | ";
-    s += "AVX512 = "    + std::to_string(ggml_cpu_has_avx512())    + " | ";
-    s += "FMA = "       + std::to_string(ggml_cpu_has_fma())       + " | ";
-    s += "NEON = "      + std::to_string(ggml_cpu_has_neon())      + " | ";
-    s += "ARM_FMA = "   + std::to_string(ggml_cpu_has_arm_fma())   + " | ";
-    s += "F16C = "      + std::to_string(ggml_cpu_has_f16c())      + " | ";
-    s += "FP16_VA = "   + std::to_string(ggml_cpu_has_fp16_va())   + " | ";
-    s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | ";
-    s += "BLAS = "      + std::to_string(ggml_cpu_has_blas())      + " | ";
-    s += "SSE3 = "      + std::to_string(ggml_cpu_has_sse3())      + " | ";
-    s += "VSX = "       + std::to_string(ggml_cpu_has_vsx())       + " | ";
+    s += "AVX = "         + std::to_string(ggml_cpu_has_avx())         + " | ";
+    s += "AVX2 = "        + std::to_string(ggml_cpu_has_avx2())        + " | ";
+    s += "AVX512 = "      + std::to_string(ggml_cpu_has_avx512())      + " | ";
+    s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | ";
+    s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | ";
+    s += "FMA = "         + std::to_string(ggml_cpu_has_fma())         + " | ";
+    s += "NEON = "        + std::to_string(ggml_cpu_has_neon())        + " | ";
+    s += "ARM_FMA = "     + std::to_string(ggml_cpu_has_arm_fma())     + " | ";
+    s += "F16C = "        + std::to_string(ggml_cpu_has_f16c())        + " | ";
+    s += "FP16_VA = "     + std::to_string(ggml_cpu_has_fp16_va())     + " | ";
+    s += "WASM_SIMD = "   + std::to_string(ggml_cpu_has_wasm_simd())   + " | ";
+    s += "BLAS = "        + std::to_string(ggml_cpu_has_blas())        + " | ";
+    s += "SSE3 = "        + std::to_string(ggml_cpu_has_sse3())        + " | ";
+    s += "VSX = "         + std::to_string(ggml_cpu_has_vsx())         + " | ";
 
     return s.c_str();
 }

llama.h

@@ -72,6 +72,7 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
+        LLAMA_FTYPE_MOSTLY_Q4_2 = 5,  // except 1d tensors
     };
 
     LLAMA_API struct llama_context_params llama_context_default_params();
@@ -96,6 +97,18 @@ extern "C" {
             const char * fname_out,
       enum llama_ftype   ftype);
 
+    // Apply a LoRA adapter to a loaded model
+    // path_base_model is the path to a higher quality model to use as a base for
+    // the layers modified by the adapter. Can be NULL to use the current loaded model.
+    // The model needs to be reloaded before applying a new adapter, otherwise the adapter
+    // will be applied on top of the previous one
+    // Returns 0 on success
+    LLAMA_API int llama_apply_lora_from_file(
+            struct llama_context * ctx,
+                      const char * path_lora,
+                      const char * path_base_model,
+                             int   n_threads);
+
     // Returns the KV cache that will contain the context for the
     // ongoing prediction with the model.
     LLAMA_API const uint8_t * llama_get_kv_cache(struct llama_context * ctx);

llama_util.h

@@ -43,8 +43,12 @@
     } while (0)
 
 #ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((format(gnu_printf, 1, 2)))
+#else
 __attribute__((format(printf, 1, 2)))
 #endif
+#endif
 static std::string format(const char * fmt, ...) {
     va_list ap, ap2;
     va_start(ap, fmt);
@@ -57,7 +61,7 @@ static std::string format(const char * fmt, ...) {
     va_end(ap2);
     va_end(ap);
     return std::string(buf.data(), size);
-};
+}
 
 struct llama_file {
     // use FILE * so we don't have to re-open the file to mmap
@@ -164,7 +168,7 @@ struct llama_mmap {
 #ifdef _POSIX_MAPPED_FILES
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file) {
+    llama_mmap(struct llama_file * file, bool prefetch = true) {
         size = file->size;
         int fd = fileno(file->fp);
         int flags = MAP_SHARED;
@@ -172,15 +176,16 @@ struct llama_mmap {
         flags |= MAP_POPULATE;
 #endif
         addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
-        close(fd);
         if (addr == MAP_FAILED) {
             throw format("mmap failed: %s", strerror(errno));
         }
 
-        // Advise the kernel to preload the mapped memory
-        if (madvise(addr, file->size, MADV_WILLNEED)) {
-            fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
-                    strerror(errno));
+        if (prefetch) {
+            // Advise the kernel to preload the mapped memory
+            if (madvise(addr, file->size, MADV_WILLNEED)) {
+                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+                        strerror(errno));
+            }
         }
     }
 
@@ -190,14 +195,13 @@ struct llama_mmap {
 #elif defined(_WIN32)
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file) {
+    llama_mmap(struct llama_file * file, bool prefetch = true) {
         size = file->size;
 
         HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
 
         HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
         DWORD error = GetLastError();
-        CloseHandle(hFile);
 
         if (hMapping == NULL) {
             throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str());
@@ -212,13 +216,15 @@ struct llama_mmap {
         }
 
         #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
-        // Advise the kernel to preload the mapped memory
-        WIN32_MEMORY_RANGE_ENTRY range;
-        range.VirtualAddress = addr;
-        range.NumberOfBytes = (SIZE_T)size;
-        if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
-            fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
-                    llama_format_win_err(GetLastError()).c_str());
+        if (prefetch) {
+            // Advise the kernel to preload the mapped memory
+            WIN32_MEMORY_RANGE_ENTRY range;
+            range.VirtualAddress = addr;
+            range.NumberOfBytes = (SIZE_T)size;
+            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+            }
         }
         #else
         #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")

migrate-ggml-2023-03-30-pr613.py

@@ -1,311 +0,0 @@
-# Migrate ggml file(s) with ggmf magic to ggml file with ggjt magic
-#
-# We caused a breaking change to the file format on 2023-03-30 in:
-#     https://github.com/ggerganov/llama.cpp/pull/613
-#
-# (1) If you still have the Meta LLaMA .pth files, then close this
-#     file now; you can just run `convert-pth-to-ggml.py` again to
-#     migrate to the new format. The tool is easier to use too. It
-#     isn't necessary anymore to manage split output files because
-#     the new format always combines things into a single file.
-#
-# (2) If you deleted the Meta LLaMA .pth files due to save on disk
-#     space, then this tool is intended to help you.  Please check
-#     out the instructions below.
-#
-# USAGE
-#
-#     python migrate-ggml-2023-03-30-pr613.py INPUT OUTPUT
-#
-# PREREQUISITES
-#
-#     pip install numpy
-#     cd llama.cpp
-#     make -j4
-#
-# EXAMPLE (7B MODEL)
-#
-#     # you can replace all the 'f16' with 'q4_0' if you're using quantized weights
-#     python migrate-ggml-2023-03-30-pr613.py models/7B/ggml-model-f16.bin models/7B/ggml-model-f16-ggjt.bin
-#
-#     # check that it works
-#     ./main -m models/7B/ggml-model-f16-ggjt.bin -p 'Question: Do you love me?'
-#
-#     # you can delete the old files
-#     rm -f models/7B/ggml-model-f16.bin
-#     mv models/7B/ggml-model-f16-ggjt.bin models/7B/ggml-model-f16.bin
-#
-# EXAMPLE (13B MODEL)
-#
-#     # you can replace all the 'f16' with 'q4_0' if you're using quantized weights
-#     python migrate-ggml-2023-03-30-pr613.py models/13B/ggml-model-f16.bin models/13B/ggml-model-f16-ggjt.bin
-#
-#     # check that it works
-#     ./main -m models/13B/ggml-model-f16-ggjt.bin -p 'Question: Do you love me?'
-#
-#     # you can delete the old files
-#     rm -f models/13B/ggml-model-f16.bin*
-#     mv models/13B/ggml-model-f16-ggjt.bin models/13B/ggml-model-f16.bin
-#
-
-import argparse
-import os
-import sys
-import json
-import struct
-import numpy as np
-
-QK = 32
-
-GGML_TYPE_Q4_0  = 0
-GGML_TYPE_Q4_1  = 1
-GGML_TYPE_I8    = 2
-GGML_TYPE_I16   = 3
-GGML_TYPE_I32   = 4
-GGML_TYPE_F16   = 5
-GGML_TYPE_F32   = 6
-
-WTYPE_NAMES = {
-    0: "F32",
-    1: "F16",
-    2: "Q4_0",
-    3: "Q4_1",
-}
-
-WTYPES = {
-    0: GGML_TYPE_F32,
-    1: GGML_TYPE_F16,
-    2: GGML_TYPE_Q4_0,
-    3: GGML_TYPE_Q4_1,
-}
-
-GGML_BLCK_SIZE = {
-    GGML_TYPE_Q4_0:  QK,
-    GGML_TYPE_Q4_1:  QK,
-    GGML_TYPE_I8:    1,
-    GGML_TYPE_I16:   1,
-    GGML_TYPE_I32:   1,
-    GGML_TYPE_F16:   1,
-    GGML_TYPE_F32:   1,
-}
-
-GGML_TYPE_SIZE = {
-    GGML_TYPE_Q4_0: 4   + QK//2,
-    GGML_TYPE_Q4_1: 4*2 + QK//2,
-    GGML_TYPE_I8:   1,
-    GGML_TYPE_I16:  2,
-    GGML_TYPE_I32:  4,
-    GGML_TYPE_F16:  2,
-    GGML_TYPE_F32:  4,
-}
-
-HPARAMS = [
-    'magic',    # int32
-    'version',  # int32
-    'n_vocab',  # int32
-    'n_embd',   # int32
-    'n_mult',   # int32
-    'n_head',   # int32
-    'n_layer',  # int32
-    'n_rot',    # int32
-    'f16',      # int32
-]
-
-def read_hparams(fin):
-    struct_fmt = "i" * len(HPARAMS)
-    struct_size = struct.calcsize(struct_fmt)
-    buf = fin.read(struct_size)
-    ints = struct.unpack(struct_fmt, buf)
-    hparams = dict(zip(HPARAMS, ints))
-    return hparams
-
-def write_hparams(fout, hparams):
-    struct_fmt = "i" * len(HPARAMS)
-    struct_size = struct.calcsize(struct_fmt)
-    ints = [hparams[h] for h in HPARAMS]
-    fout.write(struct.pack(struct_fmt, *ints))
-
-def read_tokens(fin, hparams):
-    tokens = []
-    for i in range(hparams['n_vocab']):
-        len_b = fin.read(4)
-        (length,) = struct.unpack("i", len_b)
-        word = fin.read(length)
-        score_b = fin.read(4)
-        (score,) = struct.unpack("f", score_b)
-        tokens.append((word, score))
-    return tokens
-
-def write_tokens(fout, tokens):
-    for word, score in tokens:
-        fout.write(struct.pack("i", len(word)))
-        fout.write(word)
-        fout.write(struct.pack("f", score))
-
-def ggml_nelements(shape):
-    r = 1
-    for i in shape:
-        r *= i
-    return r
-
-def ggml_nbytes(shape, ftype):
-    x = ggml_nelements(shape)
-    t = WTYPES[ftype]
-    x *= GGML_TYPE_SIZE[t]
-    x //= GGML_BLCK_SIZE[t]
-    return x
-
-def copy_tensors(fin, fout, part_id, n_parts):
-    while True:
-
-        b = fin.read(4)
-        if not b: break
-        (n_dims,) = struct.unpack("i", b)
-        b = fin.read(4)
-        (length,) = struct.unpack("i", b)
-        b = fin.read(4)
-        (ftype,) = struct.unpack("i", b)
-
-        assert n_dims in (1, 2)
-
-        partshape = list(range(n_dims))
-        for i in range(n_dims):
-            b = fin.read(4)
-            partshape[i] = struct.unpack("i", b)[0]
-        partshape = list(reversed(partshape))
-
-        name = fin.read(length)
-        data = fin.read(ggml_nbytes(partshape, ftype))
-
-        blck_size = GGML_BLCK_SIZE[WTYPES[ftype]]
-        type_size = GGML_TYPE_SIZE[WTYPES[ftype]]
-
-        print(f"Processing tensor {name} with shape: {partshape} and type: {WTYPE_NAMES[ftype]}")
-
-        # determine dimension along which multipart tensor is sharded
-        #
-        # split_dim 0 regex:
-        #   - output.*
-        #   - layers.*.attention.wq.weight
-        #   - layers.*.attention.wk.weight
-        #   - layers.*.attention.wv.weight
-        #   - layers.*.feed_forward.w1.weight
-        #   - layers.*.feed_forward.w3.weight
-        #
-        # split_dim 1 regex:
-        #   - tok_embeddings.*
-        #   - layers.*.attention.wo.weight
-        #   - layers.*.feed_forward.w2.weight
-        #
-        if n_dims > 1:
-            split_dim = 1
-            if b"tok_embeddings" in name:
-                split_dim = 1
-            elif b"layers" in name:
-                if b"attention.wo.weight" in name:
-                    split_dim = 1
-                elif b"feed_forward.w2.weight" in name:
-                    split_dim = 1
-                else:
-                    split_dim = 0
-            elif b"output" in name:
-                split_dim = 0
-
-        # output tensor header
-        fullshape = list(partshape)
-        if n_dims > 1:
-            fullshape[split_dim] *= n_parts
-        fout.write(struct.pack("iii", n_dims, len(name), ftype))
-        for dim in reversed(fullshape):
-            fout.write(struct.pack("i", dim))
-        fout.write(name)
-
-        # ensure tensor data is aligned
-        tensor_data_offset = fout.tell()
-        while tensor_data_offset % QK != 0:
-            fout.write(struct.pack("B", 0))
-            tensor_data_offset += 1
-
-        # output unified mappable tensor data
-        if n_dims == 1 or n_parts == 1:
-            # copy tensor which we thankfully received in one piece
-            if part_id == 0:
-                fout.write(data)
-        elif split_dim == 0:
-            # reassemble multifile tensor containing some of the rows
-            rows_per_chunk = partshape[0]
-            current_row = part_id * rows_per_chunk
-            bytes_per_row = fullshape[1] // blck_size * type_size
-            offset = current_row * bytes_per_row
-            fout.seek(tensor_data_offset + offset)
-            fout.write(data)
-        elif split_dim == 1:
-            # reassemble multifile tensor containing some of the cols
-            cols_per_chunk = partshape[1]
-            current_col = part_id * cols_per_chunk
-            bpr = partshape[1] // blck_size * type_size
-            bytes_per_row = fullshape[1] // blck_size * type_size
-            offset_current_col = current_col // blck_size * type_size
-            for row in range(partshape[0]):
-                offset_row = row * bytes_per_row
-                offset = offset_row + offset_current_col
-                fout.seek(tensor_data_offset + offset)
-                fout.write(data[row * bpr:row * bpr + bpr])
-
-        # advance file position to next tensor
-        fout.seek(tensor_data_offset + ggml_nbytes(fullshape, ftype))
-
-def parse_args():
-    parser = argparse.ArgumentParser(description='Migrate from GGML to new GGJT file format')
-    parser.add_argument('fin_path', help='your old ggml file (leave out the .1 .2 etc.)')
-    parser.add_argument('fout_path', help='your new ggjt file name')
-    return parser.parse_args()
-
-def main():
-    args = parse_args()
-    assert args.fin_path
-    assert args.fout_path
-    assert args.fin_path != args.fout_path
-
-    with open(args.fin_path, "rb") as fin:
-        hparams = read_hparams(fin)
-        tokens = read_tokens(fin, hparams)
-
-    if hparams['magic'] == 0x67676a74:  # ggjt
-        print(f"{args.fin_path}: input ggml has already been converted to 'ggjt' magic\n")
-        sys.exit(1)
-
-    if hparams['magic'] != 0x67676d66:  # ggmf
-        print(f"{args.fin_path}: input ggml file doesn't have expected 'ggmf' magic: {hparams['magic']:#x}\n")
-        sys.exit(1)
-
-    hparams['magic'] = 0x67676a74  # ggjt
-
-    # count number of multipart files by convention
-    n_parts = 1
-    while True:
-        if os.path.exists(f"{args.fin_path}.{n_parts}"):
-            n_parts += 1
-        else:
-            break
-
-    # we output a single file for ggml
-    with open(args.fout_path, "wb") as fout:
-        write_hparams(fout, hparams)
-        write_tokens(fout, tokens)
-        offset_of_tensors = fout.tell()
-        # the tensors we load could be split across multiple files
-        for part_id in range(n_parts):
-            fout.seek(offset_of_tensors)
-            print(f"Processing part {part_id+1} of {n_parts}\n")
-            fin_path = args.fin_path
-            if part_id > 0:
-                fin_path += f".{part_id}"
-            with open(fin_path, "rb") as fin:
-                read_tokens(fin, read_hparams(fin))
-                copy_tensors(fin, fout, part_id, n_parts)
-
-    print(f"Done. Output file: {args.fout_path}\n")
-
-if __name__ == "__main__":
-    main()

pocs/CMakeLists.txt

@@ -0,0 +1,12 @@
+# dependencies
+
+find_package(Threads REQUIRED)
+
+# third-party
+
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+
+if (EMSCRIPTEN)
+else()
+    add_subdirectory(vdot)
+endif()

pocs/vdot/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET vdot)
+add_executable(${TARGET} vdot.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

pocs/vdot/vdot.cpp

@@ -0,0 +1,305 @@
+#include <cstdio>
+#include <vector>
+#include <random>
+#include <chrono>
+#include <cstdlib>
+#include <cmath>
+#include <cassert>
+#include <cstring>
+#include <array>
+
+#include <ggml.h>
+
+constexpr int kVecSize = 1 << 18;
+
+float drawFromGaussianPdf(std::mt19937& rndm) {
+    constexpr double kScale = 1./(1. + std::mt19937::max());
+    constexpr double kTwoPiTimesScale = 6.28318530717958647692*kScale;
+    static float lastX;
+    static bool haveX = false;
+    if (haveX) { haveX = false; return lastX; }
+    auto r = sqrt(-2*log(1 - kScale*rndm()));
+    auto phi = kTwoPiTimesScale * rndm();
+    lastX = r*sin(phi);
+    haveX = true;
+    return r*cos(phi);
+}
+void fillRandomGaussianFloats(std::vector<float>& values, std::mt19937& rndm, float mean = 0) {
+    for (auto& v : values) v = mean + drawFromGaussianPdf(rndm);
+}
+
+// Copy-pasted from ggml.c
+#define QK4_0 32
+typedef struct {
+    float   d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    float   d;          // delta
+    float   m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+// Copy-pasted from ggml.c
+#define QK8_0 32
+typedef struct {
+    float   d;          // delta
+    int8_t  qs[QK8_0];  // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
+
+// "Scalar" dot product between the quantized vector x and float vector y
+inline double dot(int n, const block_q4_0* x, const float* y) {
+    const static float kValues[16] = {-8.f, -7.f, -6.f, -5.f, -4.f, -3.f, -2.f, -1.f, 0.f, 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f};
+    constexpr uint32_t kMask1 = 0x0f0f0f0f;
+    uint32_t u1, u2;
+    auto q1 = (const uint8_t*)&u1;
+    auto q2 = (const uint8_t*)&u2;
+    double sum = 0;
+    for (int i=0; i<n; ++i) {
+        float d = x->d;
+        auto u = (const uint32_t*)x->qs;
+        float s = 0;
+        for (int k=0; k<4; ++k) {
+            u1 = u[k] & kMask1;
+            u2 = (u[k] >> 4) & kMask1;
+            s += y[0]*kValues[q1[0]] + y[1]*kValues[q2[0]] +
+                 y[2]*kValues[q1[1]] + y[3]*kValues[q2[1]] +
+                 y[4]*kValues[q1[2]] + y[5]*kValues[q2[2]] +
+                 y[6]*kValues[q1[3]] + y[7]*kValues[q2[3]];
+            y += 8;
+        }
+        sum += s*d;
+        ++x;
+    }
+    return sum;
+}
+// Alternative version of the above. Faster on my Mac (~45 us vs ~55 us per dot product),
+// but about the same on X86_64 (Ryzen 7950X CPU).
+inline double dot3(int n, const block_q4_0* x, const float* y) {
+    const static std::pair<float,float> kValues[256] = {
+        {-8.f, -8.f}, {-7.f, -8.f}, {-6.f, -8.f}, {-5.f, -8.f}, {-4.f, -8.f}, {-3.f, -8.f}, {-2.f, -8.f}, {-1.f, -8.f},
+        { 0.f, -8.f}, { 1.f, -8.f}, { 2.f, -8.f}, { 3.f, -8.f}, { 4.f, -8.f}, { 5.f, -8.f}, { 6.f, -8.f}, { 7.f, -8.f},
+        {-8.f, -7.f}, {-7.f, -7.f}, {-6.f, -7.f}, {-5.f, -7.f}, {-4.f, -7.f}, {-3.f, -7.f}, {-2.f, -7.f}, {-1.f, -7.f},
+        { 0.f, -7.f}, { 1.f, -7.f}, { 2.f, -7.f}, { 3.f, -7.f}, { 4.f, -7.f}, { 5.f, -7.f}, { 6.f, -7.f}, { 7.f, -7.f},
+        {-8.f, -6.f}, {-7.f, -6.f}, {-6.f, -6.f}, {-5.f, -6.f}, {-4.f, -6.f}, {-3.f, -6.f}, {-2.f, -6.f}, {-1.f, -6.f},
+        { 0.f, -6.f}, { 1.f, -6.f}, { 2.f, -6.f}, { 3.f, -6.f}, { 4.f, -6.f}, { 5.f, -6.f}, { 6.f, -6.f}, { 7.f, -6.f},
+        {-8.f, -5.f}, {-7.f, -5.f}, {-6.f, -5.f}, {-5.f, -5.f}, {-4.f, -5.f}, {-3.f, -5.f}, {-2.f, -5.f}, {-1.f, -5.f},
+        { 0.f, -5.f}, { 1.f, -5.f}, { 2.f, -5.f}, { 3.f, -5.f}, { 4.f, -5.f}, { 5.f, -5.f}, { 6.f, -5.f}, { 7.f, -5.f},
+        {-8.f, -4.f}, {-7.f, -4.f}, {-6.f, -4.f}, {-5.f, -4.f}, {-4.f, -4.f}, {-3.f, -4.f}, {-2.f, -4.f}, {-1.f, -4.f},
+        { 0.f, -4.f}, { 1.f, -4.f}, { 2.f, -4.f}, { 3.f, -4.f}, { 4.f, -4.f}, { 5.f, -4.f}, { 6.f, -4.f}, { 7.f, -4.f},
+        {-8.f, -3.f}, {-7.f, -3.f}, {-6.f, -3.f}, {-5.f, -3.f}, {-4.f, -3.f}, {-3.f, -3.f}, {-2.f, -3.f}, {-1.f, -3.f},
+        { 0.f, -3.f}, { 1.f, -3.f}, { 2.f, -3.f}, { 3.f, -3.f}, { 4.f, -3.f}, { 5.f, -3.f}, { 6.f, -3.f}, { 7.f, -3.f},
+        {-8.f, -2.f}, {-7.f, -2.f}, {-6.f, -2.f}, {-5.f, -2.f}, {-4.f, -2.f}, {-3.f, -2.f}, {-2.f, -2.f}, {-1.f, -2.f},
+        { 0.f, -2.f}, { 1.f, -2.f}, { 2.f, -2.f}, { 3.f, -2.f}, { 4.f, -2.f}, { 5.f, -2.f}, { 6.f, -2.f}, { 7.f, -2.f},
+        {-8.f, -1.f}, {-7.f, -1.f}, {-6.f, -1.f}, {-5.f, -1.f}, {-4.f, -1.f}, {-3.f, -1.f}, {-2.f, -1.f}, {-1.f, -1.f},
+        { 0.f, -1.f}, { 1.f, -1.f}, { 2.f, -1.f}, { 3.f, -1.f}, { 4.f, -1.f}, { 5.f, -1.f}, { 6.f, -1.f}, { 7.f, -1.f},
+        {-8.f,  0.f}, {-7.f,  0.f}, {-6.f,  0.f}, {-5.f,  0.f}, {-4.f,  0.f}, {-3.f,  0.f}, {-2.f,  0.f}, {-1.f,  0.f},
+        { 0.f,  0.f}, { 1.f,  0.f}, { 2.f,  0.f}, { 3.f,  0.f}, { 4.f,  0.f}, { 5.f,  0.f}, { 6.f,  0.f}, { 7.f,  0.f},
+        {-8.f,  1.f}, {-7.f,  1.f}, {-6.f,  1.f}, {-5.f,  1.f}, {-4.f,  1.f}, {-3.f,  1.f}, {-2.f,  1.f}, {-1.f,  1.f},
+        { 0.f,  1.f}, { 1.f,  1.f}, { 2.f,  1.f}, { 3.f,  1.f}, { 4.f,  1.f}, { 5.f,  1.f}, { 6.f,  1.f}, { 7.f,  1.f},
+        {-8.f,  2.f}, {-7.f,  2.f}, {-6.f,  2.f}, {-5.f,  2.f}, {-4.f,  2.f}, {-3.f,  2.f}, {-2.f,  2.f}, {-1.f,  2.f},
+        { 0.f,  2.f}, { 1.f,  2.f}, { 2.f,  2.f}, { 3.f,  2.f}, { 4.f,  2.f}, { 5.f,  2.f}, { 6.f,  2.f}, { 7.f,  2.f},
+        {-8.f,  3.f}, {-7.f,  3.f}, {-6.f,  3.f}, {-5.f,  3.f}, {-4.f,  3.f}, {-3.f,  3.f}, {-2.f,  3.f}, {-1.f,  3.f},
+        { 0.f,  3.f}, { 1.f,  3.f}, { 2.f,  3.f}, { 3.f,  3.f}, { 4.f,  3.f}, { 5.f,  3.f}, { 6.f,  3.f}, { 7.f,  3.f},
+        {-8.f,  4.f}, {-7.f,  4.f}, {-6.f,  4.f}, {-5.f,  4.f}, {-4.f,  4.f}, {-3.f,  4.f}, {-2.f,  4.f}, {-1.f,  4.f},
+        { 0.f,  4.f}, { 1.f,  4.f}, { 2.f,  4.f}, { 3.f,  4.f}, { 4.f,  4.f}, { 5.f,  4.f}, { 6.f,  4.f}, { 7.f,  4.f},
+        {-8.f,  5.f}, {-7.f,  5.f}, {-6.f,  5.f}, {-5.f,  5.f}, {-4.f,  5.f}, {-3.f,  5.f}, {-2.f,  5.f}, {-1.f,  5.f},
+        { 0.f,  5.f}, { 1.f,  5.f}, { 2.f,  5.f}, { 3.f,  5.f}, { 4.f,  5.f}, { 5.f,  5.f}, { 6.f,  5.f}, { 7.f,  5.f},
+        {-8.f,  6.f}, {-7.f,  6.f}, {-6.f,  6.f}, {-5.f,  6.f}, {-4.f,  6.f}, {-3.f,  6.f}, {-2.f,  6.f}, {-1.f,  6.f},
+        { 0.f,  6.f}, { 1.f,  6.f}, { 2.f,  6.f}, { 3.f,  6.f}, { 4.f,  6.f}, { 5.f,  6.f}, { 6.f,  6.f}, { 7.f,  6.f},
+        {-8.f,  7.f}, {-7.f,  7.f}, {-6.f,  7.f}, {-5.f,  7.f}, {-4.f,  7.f}, {-3.f,  7.f}, {-2.f,  7.f}, {-1.f,  7.f},
+        { 0.f,  7.f}, { 1.f,  7.f}, { 2.f,  7.f}, { 3.f,  7.f}, { 4.f,  7.f}, { 5.f,  7.f}, { 6.f,  7.f}, { 7.f,  7.f}
+    };
+    double sum = 0;
+    for (int i=0; i<n; ++i) {
+        float d = x->d;
+        auto q = x->qs;
+        float s = 0;
+        for (int k=0; k<4; ++k) {
+            s += y[0]*kValues[q[0]].first + y[1]*kValues[q[0]].second +
+                 y[2]*kValues[q[1]].first + y[3]*kValues[q[1]].second +
+                 y[4]*kValues[q[2]].first + y[5]*kValues[q[2]].second +
+                 y[6]*kValues[q[3]].first + y[7]*kValues[q[3]].second;
+            y += 8; q += 4;
+        }
+        sum += s*d;
+        ++x;
+    }
+    return sum;
+}
+
+inline double dot41(int n, const block_q4_1* x, const float* y) {
+    const static float kValues[16] = {0.f, 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f, 8.f, 9.f, 10.f, 11.f, 12.f, 13.f, 14.f, 15.f};
+    constexpr uint32_t kMask1 = 0x0f0f0f0f;
+    uint32_t u1, u2;
+    auto q1 = (const uint8_t*)&u1;
+    auto q2 = (const uint8_t*)&u2;
+    double sum = 0;
+    for (int i=0; i<n; ++i) {
+        auto u = (const uint32_t*)x->qs;
+        float s = 0, s1 = 0;
+        for (int k=0; k<4; ++k) {
+            u1 = u[k] & kMask1;
+            u2 = (u[k] >> 4) & kMask1;
+            s += y[0]*kValues[q1[0]] + y[1]*kValues[q2[0]] +
+                 y[2]*kValues[q1[1]] + y[3]*kValues[q2[1]] +
+                 y[4]*kValues[q1[2]] + y[5]*kValues[q2[2]] +
+                 y[6]*kValues[q1[3]] + y[7]*kValues[q2[3]];
+            s1 += y[0] + y[1] + y[2] + y[3] + y[4] + y[5] + y[6] + y[7];
+            y += 8;
+        }
+        sum += s*x->d + s1*x->m;
+        ++x;
+    }
+    return sum;
+}
+
+// Copy-pasted from ggml.c
+static void quantize_row_q8_0_reference(const float *x, block_q8_0 *y, int k) {
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    for (int i = 0; i < nb; i++) {
+        float amax = 0.0f; // absolute max
+
+        for (int l = 0; l < QK8_0; l++) {
+            const float v = x[i*QK8_0 + l];
+            amax = std::max(amax, fabsf(v));
+        }
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = d;
+
+        for (int l = 0; l < QK8_0; ++l) {
+            const float   v  = x[i*QK8_0 + l]*id;
+            y[i].qs[l] = roundf(v);
+        }
+    }
+}
+
+// Copy-pasted from ggml.c
+static void dot_q4_q8(const int n, float* s, const void* vx, const void* vy) {
+    const int nb = n / QK8_0;
+    const block_q4_0* x = (const block_q4_0*)vx;
+    const block_q8_0* y = (const block_q8_0*)vy;
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+        const float d0 = x[i].d;
+        const float d1 = y[i].d;
+
+        const uint8_t * p0 = x[i].qs;
+        const  int8_t * p1 = y[i].qs;
+
+        int sumi = 0;
+        for (int j = 0; j < QK8_0/2; j++) {
+            const uint8_t v0 = p0[j];
+
+            const int i0 = (int8_t) (v0 & 0xf) - 8;
+            const int i1 = (int8_t) (v0 >> 4)  - 8;
+
+            const int i2 = p1[2*j + 0];
+            const int i3 = p1[2*j + 1];
+
+            sumi += i0*i2 + i1*i3;
+        }
+        sumf += d0*d1*sumi;
+    }
+    *s = sumf;
+}
+
+int main(int argc, char** argv) {
+
+    int nloop = argc > 1 ? atoi(argv[1]) : 10;
+    bool scalar = argc > 2 ? atoi(argv[2]) : false;
+    bool useQ4_1 = argc > 3 ? atoi(argv[3]) : false;
+
+    if (scalar && useQ4_1) {
+        printf("It is not possible to use Q4_1 quantization and scalar implementations\n");
+        return 1;
+    }
+
+    std::mt19937 rndm(1234);
+
+    std::vector<float> x1(kVecSize), y1(kVecSize);
+    int n4 = useQ4_1 ? kVecSize / QK4_1 : kVecSize / QK4_0; n4 = 64*((n4 + 63)/64);
+    int n8 = kVecSize / QK8_0; n8 = 64*((n8 + 63)/64);
+
+    auto funcs = useQ4_1 ? ggml_internal_get_quantize_fn(GGML_TYPE_Q4_1) : ggml_internal_get_quantize_fn(GGML_TYPE_Q4_0);
+
+    std::vector<block_q4_0> q40;
+    std::vector<block_q4_1> q41;
+    if (useQ4_1) q41.resize(n4);
+    else q40.resize(n4);
+    std::vector<block_q8_0> q8(n8);
+    std::vector<int64_t> H(16, 0);
+    double sumt = 0, sumt2 = 0, maxt = 0;
+    double sumqt = 0, sumqt2 = 0, maxqt = 0;
+    double sum = 0, sumq = 0, exactSum = 0;
+    for (int iloop=0; iloop<nloop; ++iloop) {
+
+        // Fill vector x with random numbers
+        fillRandomGaussianFloats(x1, rndm);
+
+        // Fill vector y with random numbers
+        fillRandomGaussianFloats(y1, rndm);
+
+        // Compute the exact dot product
+        for (int k=0; k<kVecSize; ++k) exactSum += x1[k]*y1[k];
+
+        // quantize x.
+        // Note, we do not include this in the timing as in practical application
+        // we already have the quantized model weights.
+        if (useQ4_1) {
+            funcs.quantize_row_q(x1.data(), q41.data(), kVecSize);
+        } else {
+            funcs.quantize_row_q(x1.data(), q40.data(), kVecSize);
+        }
+
+        // Now measure time the dot product needs using the "scalar" version above
+        auto t1 = std::chrono::high_resolution_clock::now();
+        if (useQ4_1) sum += dot41(kVecSize / QK4_1, q41.data(), y1.data());
+        else sum += dot(kVecSize / QK4_0, q40.data(), y1.data());
+        auto t2 = std::chrono::high_resolution_clock::now();
+        auto t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
+        sumt += t; sumt2 += t*t; maxt = std::max(maxt, t);
+
+        // And now measure the time needed to quantize y and perform the dot product with the quantized y
+        t1 = std::chrono::high_resolution_clock::now();
+        float result;
+        if (scalar) {
+            quantize_row_q8_0_reference(y1.data(), q8.data(), kVecSize);
+            dot_q4_q8(kVecSize, &result, q40.data(), q8.data());
+        }
+        else {
+            funcs.quantize_row_q_dot(y1.data(), q8.data(), kVecSize);
+            if (useQ4_1) funcs.vec_dot_q(kVecSize, &result, q41.data(), q8.data());
+            else funcs.vec_dot_q(kVecSize, &result, q40.data(), q8.data());
+        }
+        sumq += result;
+        t2 = std::chrono::high_resolution_clock::now();
+        t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
+        sumqt += t; sumqt2 += t*t; maxqt = std::max(maxqt, t);
+
+    }
+
+    // Report the time (and the average of the dot products so the compiler does not come up with the idea
+    // of optimizing away the function calls after figuring that the result is not used).
+    sum /= nloop; sumq /= nloop;
+    exactSum /= nloop;
+    printf("Exact result: <dot> = %g\n",exactSum);
+    printf("<dot> = %g, %g\n",sum,sumq);
+    sumt /= nloop; sumt2 /= nloop; sumt2 -= sumt*sumt;
+    if (sumt2 > 0) sumt2 = sqrt(sumt2);
+    printf("time = %g +/- %g us. maxt = %g us\n",sumt,sumt2,maxt);
+    sumqt /= nloop; sumqt2 /= nloop; sumqt2 -= sumqt*sumqt;
+    if (sumqt2 > 0) sumqt2 = sqrt(sumqt2);
+    printf("timeq = %g +/- %g us. maxt = %g us\n",sumqt,sumqt2,maxqt);
+    return 0;
+}

requirements.txt

@@ -0,0 +1,2 @@
+numpy==1.24
+sentencepiece==0.1.98

tests/test-tokenizer-0.cpp

@@ -5,13 +5,17 @@
 #include <map>
 #include <vector>
 
-static const std::map<std::string, std::vector<llama_token>> k_tests = {
-    { "Hello World",        { 1,  10994,   2787, }, },
-    { " Hello World",       { 1,  15043,   2787, }, },
-    { " Hello World!",      { 1,  15043,   2787,  29991, }, },
-    { " this is 🦙.cpp",    { 1,    445,    338,  29871,    243,    162,    169,    156,  29889,   8223, }, },
-    { "w048 7tuijk dsdfhu", { 1,  29893,  29900,  29946,  29947,  29871,  29955,   9161,  13535,  18031,   2176,   6905, }, },
-    { "нещо на Български",  { 1,    821,   4851,    665,   1386,  29713,   1305, }, },
+static const std::map<std::string, std::vector<llama_token>> & k_tests()
+{
+    static std::map<std::string, std::vector<llama_token>> _k_tests = {
+        { "Hello World",        { 1,  10994,   2787, }, },
+        { " Hello World",       { 1,  15043,   2787, }, },
+        { " Hello World!",      { 1,  15043,   2787,  29991, }, },
+        { " this is 🦙.cpp",    { 1,    445,    338,  29871,    243,    162,    169,    156,  29889,   8223, }, },
+        { "w048 7tuijk dsdfhu", { 1,  29893,  29900,  29946,  29947,  29871,  29955,   9161,  13535,  18031,   2176,   6905, }, },
+        { "нещо на Български",  { 1,    821,   4851,    665,   1386,  29713,   1305, }, },
+    };
+    return _k_tests;
 };
 
 int main(int argc, char **argv) {
@@ -47,7 +51,7 @@ int main(int argc, char **argv) {
         return 2;
     }
 
-    for (const auto & test_kv : k_tests) {
+    for (const auto & test_kv : k_tests()) {
         std::vector<llama_token> res(test_kv.first.size());
         const int n = llama_tokenize(ctx, test_kv.first.c_str(), res.data(), res.size(), true);
         res.resize(n);