ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684)

* Starting to add k-quantization to ggml

I think it is better to have quantization separate from
ggml. For now just adding the k-quants there, but it would be
better to also factor out the existing ggml quantizations.

* Adding Q3_K and Q8_K (de)-quantization

* Q3_K now working on CUDA and AVX2/scalar

CUDA is not ideal - ~50% slower than Q4_0 for
single token prediction, about the same in batch
mode (perplexity). CPU single token is ~55 ms
(on Ryzen 7950X).

* Some improvement for Q3_K on CUDA

It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0.

* Some more CUDA optimizations for Q3_K

Single token is now 20.5 ms/token (~20% slower than Q4_0).
Perplexity is on par with Q4_0.

* Adding Q4_K - scalar, AVX2, CUDA

Performance is the same or perhaps very slightly better than Q4_0 on the CPU.
On the GPU, single token prediction is ~10% better than Q4_0,
batch mode (perplexity is about the same).

* Adding Q6_K - scalar, AVX2, CUDA

Performance is ~40% lower compared to Q4_K on the CPU.
This is to be expected, considering that we are memory bound
on the CPU and the 6-bit model is ~44% larger than the 4-bit.
On the GPU, single token prediction is ~6% lower than Q4_0,
batch mode (perplexity) is even closer (but still slower).

* Adding Q5_K - scalar, AVX2, CUDA

Performance is ~20% lower compared to Q4_K on the CPU.
This is to be expected, considering that we are memory bound
on the CPU and the 5-bit model is ~22% larger than the 4-bit.
On the GPU, single token prediction is about the same as Q4_0
for both, single token and batch prediction.

* Per convention, all QX_K quantizations use Q5_K for output.weight

* Adding quantization mixes

* Quantization mixes: didn't quite get what I wanted in the last commit

* Q4_K dot product for ARM_NEON

* Q6_K dot product for ARM_NEON

* Q5_K dot product for ARM_NEON

* Adding Q3_K dot for ARM_NEON

It is 22% slower than Q4_K, despite the smaller model size.
On x86_64, where we are memory bound, the Q3_K model is
quite a bit faster than Q4_K.

* A very slightly faster ARM_NEON Q3_K dot

* Adding Q2_K - just CUDA for now

Token prediction is pretty good - about 15.5 ms on a RTX 4080.
Perplexity is about the same as Q4_K.

* Adding scalar and AVX2 Q2_K dot

* Adding ARM_NEON Q2_K dot

About the same performance as Q4_K.

* A slightly faster ARM_NEON Q2_K dot

Single token prediction is now ~36 ms on M2 Max.
The code is much simpler too.

* Fixed bug in Q2_K CUDA dot product kernel

Stranegly enough, for the few prompts I tried with the 7B model
the responses looked perfectly reasonable. Only realized something
is not quite right when I tried the larger models and started getting
nonse back.

In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X
box iusing CUDA and model fully loaded on the GPU are
  ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B.
The max number of layers that fit in VRAM for The 65B is 32.
With that, we get ~330 ms per token, which is not that much faster
than just running on the CPU (~470 ms per token).

* Don't print zeros/NaNs when no count histogram has been collected

* A 10% faster CUDA vector dot kernel for Q3_K

Q3_K is now running at ~18.5 ms / token on CUDA,
so the gap to Q4_0 is only 10%.
It seems memory acccess pattern is more important for
performance than the amount of computation the kernel
does.

* A slightly daster Q4_K AVX2 dot product

For perplexity, where we are less memory bound, time per
pass drops by ~5%. Barely measurable difference for single
token prediction.

* A slightly faster ARM_NEON A4_K dot product

* Minor

* Fix quantization error test

We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit
quantization variants.

* Fix docker build

I have been sloppy with vector reinterpret casts on ARM_NEON.
It seems clang is very forgiving in that regard.

* Added forgotten ggml.o dependence on k_quants.h to the Makefile

* Had unintentionally committed the Makefile with -Ofast enabled

* ggml : rename k_quants -> ggml-quants-k, use lowercase in code

---------

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

.devops/full.Dockerfile

@@ -3,7 +3,7 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION as build
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip
+    apt-get install -y build-essential python3 python3-pip git
 
 COPY requirements.txt requirements.txt
 

.devops/main.Dockerfile

@@ -3,7 +3,7 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION as build
 
 RUN apt-get update && \
-    apt-get install -y build-essential
+    apt-get install -y build-essential git
 
 WORKDIR /app
 

.devops/tools.sh

@@ -11,7 +11,7 @@ shift
 arg2="$@"
 
 if [[ $arg1 == '--convert' || $arg1 == '-c' ]]; then
-    python3 ./convert-pth-to-ggml.py $arg2
+    python3 ./convert.py $arg2
 elif [[ $arg1 == '--quantize' || $arg1 == '-q' ]]; then
     ./quantize $arg2
 elif [[ $arg1 == '--run' || $arg1 == '-r' ]]; then
@@ -32,7 +32,7 @@ else
     echo "  --run (-r): Run a model previously converted into ggml"
     echo "              ex: -m /models/7B/ggml-model-q4_0.bin -p \"Building a website can be done in 10 simple steps:\" -n 512"
     echo "  --convert (-c): Convert a llama model into ggml"
-    echo "              ex: \"/models/7B/\" 1"
+    echo "              ex: --outtype f16 \"/models/7B/\" "
     echo "  --quantize (-q): Optimize with quantization process ggml"
     echo "              ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2"
     echo "  --all-in-one (-a): Execute --convert & --quantize"

.github/workflows/build.yml

@@ -10,10 +10,10 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp']
 
 env:
  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
@@ -151,21 +151,21 @@ jobs:
     env:
       OPENBLAS_VERSION: 0.3.23
       OPENCL_VERSION: 2023.04.17
-      CLBLAST_VERSION: 1.5.3
+      CLBLAST_VERSION: 1.6.0
 
     strategy:
       matrix:
         include:
           - build: 'avx2'
-            defines: ''
+            defines: '-DLLAMA_BUILD_SERVER=ON'
           - build: 'avx'
-            defines: '-DLLAMA_AVX2=OFF'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF'
           - build: 'avx512'
-            defines: '-DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'clblast'
-            defines: '-DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
           - build: 'openblas'
-            defines: '-DLLAMA_OPENBLAS=ON -DBLAS_LIBRARIES="/LIBPATH:$env:RUNNER_TEMP/openblas/lib" -DOPENBLAS_INC="$env:RUNNER_TEMP/openblas/include"'
+            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
 
     steps:
       - name: Clone
@@ -184,13 +184,13 @@ jobs:
         id: get_clblast
         if: ${{ matrix.build == 'clblast' }}
         run: |
-          curl.exe -o $env:RUNNER_TEMP/clblast.zip -L "https://github.com/CNugteren/CLBlast/releases/download/${env:CLBLAST_VERSION}/CLBlast-${env:CLBLAST_VERSION}-Windows-x64.zip"
+          curl.exe -o $env:RUNNER_TEMP/clblast.7z -L "https://github.com/CNugteren/CLBlast/releases/download/${env:CLBLAST_VERSION}/CLBlast-${env:CLBLAST_VERSION}-windows-x64.7z"
           curl.exe -o $env:RUNNER_TEMP/CLBlast.LICENSE.txt -L "https://github.com/CNugteren/CLBlast/raw/${env:CLBLAST_VERSION}/LICENSE"
-          mkdir $env:RUNNER_TEMP/clblast
-          tar.exe -xvf $env:RUNNER_TEMP/clblast.zip -C $env:RUNNER_TEMP/clblast
+          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/clblast.7z
+          rename-item $env:RUNNER_TEMP/CLBlast-${env:CLBLAST_VERSION}-windows-x64 clblast
           foreach ($f in (gci -Recurse -Path "$env:RUNNER_TEMP/clblast" -Filter '*.cmake')) {
             $txt = Get-Content -Path $f -Raw
-            $txt.Replace('C:/dependencies/opencl/', "$($env:RUNNER_TEMP.Replace('\','/'))/opencl/") | Set-Content -Path $f -Encoding UTF8
+            $txt.Replace('C:/vcpkg/packages/opencl_x64-windows/', "$($env:RUNNER_TEMP.Replace('\','/'))/opencl/") | Set-Content -Path $f -Encoding UTF8
           }
 
       - name: Download OpenBLAS
@@ -213,7 +213,6 @@ jobs:
           cd build
           cmake .. ${{ matrix.defines }}
           cmake --build . --config Release
-          cp ../LICENSE ./bin/Release/llama.cpp.txt
 
       - name: Add clblast.dll
         id: add_clblast_dll
@@ -258,6 +257,7 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
+          Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
           7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -292,7 +292,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_CUBLAS=ON
+          cmake .. -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON
           cmake --build . --config Release
 
       - name: Get commit hash

.gitignore

@@ -17,6 +17,7 @@ build-release/
 build-static/
 build-cublas/
 build-opencl/
+build-metal/
 build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/

BLIS.md

@@ -0,0 +1,67 @@
+BLIS Installation Manual
+------------------------
+
+BLIS is a portable software framework for high-performance BLAS-like dense linear algebra libraries. It has received awards and recognition, including the 2023 James H. Wilkinson Prize for Numerical Software and the 2020 SIAM Activity Group on Supercomputing Best Paper Prize. BLIS provides a new BLAS-like API and a compatibility layer for traditional BLAS routine calls. It offers features such as object-based API, typed API, BLAS and CBLAS compatibility layers.
+
+Project URL: https://github.com/flame/blis
+
+### Prepare:
+
+Compile BLIS:
+
+```bash
+git clone https://github.com/flame/blis
+cd blis
+./configure --enable-cblas -t openmp,pthreads auto
+# will install to /usr/local/ by default.
+make -j
+```
+
+Install BLIS:
+
+```bash
+sudo make install
+```
+
+We recommend using openmp since it's easier to modify the cores been used.
+
+### llama.cpp compilation
+
+Makefile:
+
+```bash
+make LLAMA_BLIS=1 -j
+# make LLAMA_BLIS=1 benchmark-matmult
+```
+
+CMake:
+
+```bash
+mkdir build
+cd build
+cmake -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=FLAME ..
+make -j
+```
+
+### llama.cpp execution
+
+According to the BLIS documentation, we could set the following
+environment variables to modify the behavior of openmp:
+
+```
+export GOMP_GPU_AFFINITY="0-19"
+export BLIS_NUM_THREADS=14
+```
+
+And then run the binaries as normal.
+
+
+### Intel specific issue
+
+Some might get the error message saying that `libimf.so` cannot be found.
+Please follow this [stackoverflow page](https://stackoverflow.com/questions/70687930/intel-oneapi-2022-libimf-so-no-such-file-or-directory-during-openmpi-compila).
+
+### Reference:
+
+1. https://github.com/flame/blis#getting-started
+2. https://github.com/flame/blis/blob/master/docs/Multithreading.md

CMakeLists.txt

@@ -37,40 +37,45 @@ endif()
 #
 
 # general
-option(LLAMA_STATIC                 "llama: static link libraries"                          OFF)
-option(LLAMA_NATIVE                 "llama: enable -march=native flag"                      OFF)
-option(LLAMA_LTO                    "llama: enable link time optimization"                  OFF)
+option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
+option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      OFF)
+option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
 
 # debug
-option(LLAMA_ALL_WARNINGS           "llama: enable all compiler warnings"                   ON)
-option(LLAMA_ALL_WARNINGS_3RD_PARTY "llama: enable all compiler warnings in 3rd party libs" OFF)
-option(LLAMA_GPROF                  "llama: enable gprof"                                   OFF)
+option(LLAMA_ALL_WARNINGS               "llama: enable all compiler warnings"                   ON)
+option(LLAMA_ALL_WARNINGS_3RD_PARTY     "llama: enable all compiler warnings in 3rd party libs" OFF)
+option(LLAMA_GPROF                      "llama: enable gprof"                                   OFF)
 
 # sanitizers
-option(LLAMA_SANITIZE_THREAD        "llama: enable thread sanitizer"                        OFF)
-option(LLAMA_SANITIZE_ADDRESS       "llama: enable address sanitizer"                       OFF)
-option(LLAMA_SANITIZE_UNDEFINED     "llama: enable undefined sanitizer"                     OFF)
+option(LLAMA_SANITIZE_THREAD            "llama: enable thread sanitizer"                        OFF)
+option(LLAMA_SANITIZE_ADDRESS           "llama: enable address sanitizer"                       OFF)
+option(LLAMA_SANITIZE_UNDEFINED         "llama: enable undefined sanitizer"                     OFF)
 
 # instruction set specific
-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)
+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)
-    option(LLAMA_F16C               "llama: enable F16C"                                    ON)
+    option(LLAMA_F16C                   "llama: enable F16C"                                    ON)
 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_CLBLAST                "llama: use CLBlast"                                    OFF)
+option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
+option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
+set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
+option(LLAMA_CUBLAS                          "llama: use cuBLAS"                                OFF)
+set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
+set(LLAMA_CUDA_DMMV_Y       "1" CACHE STRING "llama: y block size for dmmv CUDA kernels")
+option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
+option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
 
-option(LLAMA_BUILD_TESTS            "llama: build tests"    ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_EXAMPLES         "llama: build examples" ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_EXAMPLES             "llama: build examples" ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_SERVER               "llama: build server example"                           OFF)
 
 #
 # Build info header
@@ -145,36 +150,28 @@ if (APPLE AND LLAMA_ACCELERATE)
     endif()
 endif()
 
-if (LLAMA_OPENBLAS)
+if (LLAMA_BLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
     endif()
-
-    set(BLA_VENDOR OpenBLAS)
+    if ($(CMAKE_VERSION) VERSION_GREATER_EQUAL 3.22)
+        set(BLA_SIZEOF_INTEGER 8)
+    endif()
+    set(BLA_VENDOR ${LLAMA_BLAS_VENDOR})
     find_package(BLAS)
     if (BLAS_FOUND)
-        message(STATUS "OpenBLAS found")
+        message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
 
+        add_compile_options(${BLAS_LINKER_FLAGS})
         add_compile_definitions(GGML_USE_OPENBLAS)
-        add_link_options(${BLAS_LIBRARIES})
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} openblas)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
 
-        # 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})
+        message("${BLAS_LIBRARIES} ${BLAS_INCLUDE_DIRS}")
+        include_directories(${BLAS_INCLUDE_DIRS})
     else()
-        message(WARNING "OpenBLAS not found")
+        message(WARNING "BLAS not found, please refer to "
+        "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
+        " to set correct LLAMA_BLAS_VENDOR")
     endif()
 endif()
 
@@ -187,9 +184,11 @@ if (LLAMA_CUBLAS)
 
         enable_language(CUDA)
 
-        set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+        add_compile_definitions(GGML_CUDA_DMMV_Y=${LLAMA_CUDA_DMMV_Y})
 
         if (LLAMA_STATIC)
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -202,12 +201,37 @@ if (LLAMA_CUBLAS)
     endif()
 endif()
 
+if (LLAMA_METAL)
+    find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+    find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+    find_library(METALPERFORMANCE_FRAMEWORK MetalPerformanceShaders REQUIRED)
+
+    set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
+
+    add_compile_definitions(GGML_USE_METAL)
+    add_compile_definitions(GGML_METAL_NDEBUG)
+
+    # get full path to the file
+    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
+
+    # copy ggml-metal.metal to bin directory
+    configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
+
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
+        ${FOUNDATION_LIBRARY}
+        ${METAL_FRAMEWORK}
+        ${METALKIT_FRAMEWORK}
+        ${METALPERFORMANCE_FRAMEWORK}
+        )
+endif()
+
 if (LLAMA_CLBLAST)
     find_package(CLBlast)
     if (CLBlast_FOUND)
         message(STATUS "CLBlast found")
 
-        set(GGML_OPENCL_SOURCES ggml-opencl.c ggml-opencl.h)
+        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
 
         add_compile_definitions(GGML_USE_CLBLAST)
 
@@ -372,8 +396,12 @@ endif()
 add_library(ggml OBJECT
             ggml.c
             ggml.h
-            ${GGML_CUDA_SOURCES}
-            ${GGML_OPENCL_SOURCES})
+            ggml-quants-k.h
+            ggml-quants-k.c
+            ${GGML_SOURCES_CUDA}
+            ${GGML_SOURCES_OPENCL}
+            ${GGML_SOURCES_METAL}
+            )
 
 target_include_directories(ggml PUBLIC .)
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
@@ -386,21 +414,25 @@ endif()
 add_library(llama
             llama.cpp
             llama.h
-            llama-util.h)
+            llama-util.h
+            )
 
 target_include_directories(llama PUBLIC .)
 target_compile_features(llama PUBLIC cxx_std_11) # don't bump
-target_link_libraries(llama PRIVATE ggml ${LLAMA_EXTRA_LIBS})
+target_link_libraries(llama PRIVATE
+    ggml
+    ${LLAMA_EXTRA_LIBS}
+    )
 
 if (BUILD_SHARED_LIBS)
     set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)
 endif()
 
-if (GGML_CUDA_SOURCES)
+if (GGML_SOURCES_CUDA)
     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 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()
 

Makefile

@@ -1,5 +1,11 @@
 # Define the default target now so that it is always the first target
-default: main quantize quantize-stats perplexity embedding vdot
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot
+
+ifdef LLAMA_BUILD_SERVER
+	BUILD_TARGETS += server
+endif
+
+default: $(BUILD_TARGETS)
 
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
@@ -34,11 +40,18 @@ endif
 #
 
 # keep standard at C11 and C++11
-CFLAGS   = -I.              -O3 -std=c11   -fPIC
-CXXFLAGS = -I. -I./examples -O3 -std=c++11 -fPIC
+# -Ofast tends to produce faster code, but may not be available for some compilers.
+#OPT = -Ofast
+OPT = -O3
+CFLAGS   = -I.              $(OPT) -std=c11   -fPIC
+CXXFLAGS = -I. -I./examples $(OPT) -std=c++11 -fPIC
 LDFLAGS  =
 
-ifndef LLAMA_DEBUG
+ifdef LLAMA_DEBUG
+	CFLAGS   += -O0 -g
+	CXXFLAGS += -O0 -g
+	LDFLAGS  += -g
+else
 	CFLAGS   += -DNDEBUG
 	CXXFLAGS += -DNDEBUG
 endif
@@ -95,6 +108,7 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 	#CFLAGS   += -mfma -mf16c -mavx
 	#CXXFLAGS += -mfma -mf16c -mavx
 endif
+
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
 	ifneq (,$(findstring POWER9,$(POWER9_M)))
@@ -106,6 +120,7 @@ ifneq ($(filter ppc64%,$(UNAME_M)),)
 		CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
 	endif
 endif
+
 ifndef LLAMA_NO_ACCELERATE
 	# Mac M1 - include Accelerate framework.
 	# `-framework Accelerate` works on Mac Intel as well, with negliable performance boost (as of the predict time).
@@ -113,7 +128,8 @@ ifndef LLAMA_NO_ACCELERATE
 		CFLAGS  += -DGGML_USE_ACCELERATE
 		LDFLAGS += -framework Accelerate
 	endif
-endif
+endif # LLAMA_NO_ACCELERATE
+
 ifdef LLAMA_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
 	ifneq ($(shell grep -e "Arch Linux" -e "ID_LIKE=arch" /etc/os-release 2>/dev/null),)
@@ -121,7 +137,13 @@ ifdef LLAMA_OPENBLAS
 	else
 		LDFLAGS += -lopenblas
 	endif
-endif
+endif # LLAMA_OPENBLAS
+
+ifdef LLAMA_BLIS
+	CFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
+	LDFLAGS += -lblis -L/usr/local/lib
+endif # LLAMA_BLIS
+
 ifdef LLAMA_CUBLAS
 	CFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	CXXFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
@@ -129,11 +151,23 @@ ifdef LLAMA_CUBLAS
 	OBJS      += ggml-cuda.o
 	NVCC      = nvcc
 	NVCCFLAGS = --forward-unknown-to-host-compiler -arch=native
+ifdef LLAMA_CUDA_DMMV_X
+	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+else
+	NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
+endif # LLAMA_CUDA_DMMV_X
+ifdef LLAMA_CUDA_DMMV_Y
+	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=$(LLAMA_CUDA_DMMV_Y)
+else
+	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=1
+endif # LLAMA_CUDA_DMMV_Y
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
-endif
+endif # LLAMA_CUBLAS
+
 ifdef LLAMA_CLBLAST
-	CFLAGS  += -DGGML_USE_CLBLAST
+	CFLAGS   += -DGGML_USE_CLBLAST
+	CXXFLAGS += -DGGML_USE_CLBLAST
 	# Mac provides OpenCL as a framework
 	ifeq ($(UNAME_S),Darwin)
 		LDFLAGS += -lclblast -framework OpenCL
@@ -141,23 +175,38 @@ ifdef LLAMA_CLBLAST
 		LDFLAGS += -lclblast -lOpenCL
 	endif
 	OBJS    += ggml-opencl.o
-ggml-opencl.o: ggml-opencl.c ggml-opencl.h
+
+ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # LLAMA_CLBLAST
+
+ifdef LLAMA_METAL
+	CFLAGS   += -DGGML_USE_METAL -DGGML_METAL_NDEBUG
+	CXXFLAGS += -DGGML_USE_METAL
+	LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
+	OBJS     += ggml-metal.o
+
+ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
-endif
+endif # LLAMA_METAL
+
 ifneq ($(filter aarch64%,$(UNAME_M)),)
 	# Apple M1, M2, etc.
 	# Raspberry Pi 3, 4, Zero 2 (64-bit)
 	CFLAGS   += -mcpu=native
 	CXXFLAGS += -mcpu=native
 endif
+
 ifneq ($(filter armv6%,$(UNAME_M)),)
 	# Raspberry Pi 1, Zero
 	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
 endif
+
 ifneq ($(filter armv7%,$(UNAME_M)),)
 	# Raspberry Pi 2
 	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
 endif
+
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 3, 4, Zero 2 (32-bit)
 	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
@@ -182,7 +231,10 @@ $(info )
 # Build library
 #
 
-ggml.o: ggml.c ggml.h ggml-cuda.h
+ggml.o: ggml.c ggml.h ggml-cuda.h ggml-quants-k.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+ggml-quants-k.o: ggml-quants-k.c ggml-quants-k.h ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
 llama.o: llama.cpp ggml.h ggml-cuda.h llama.h llama-util.h
@@ -195,33 +247,36 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state build-info.h
+	rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server vdot build-info.h
 
 #
 # Examples
 #
 
-main: examples/main/main.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+main: examples/main/main.cpp build-info.h ggml.o ggml-quants-k.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-quantize: examples/quantize/quantize.cpp build-info.h ggml.o llama.o $(OBJS)
+quantize: examples/quantize/quantize.cpp build-info.h ggml.o llama.o ggml-quants-k.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.cpp build-info.h ggml.o llama.o $(OBJS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp build-info.h ggml.o llama.o ggml-quants-k.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+perplexity: examples/perplexity/perplexity.cpp build-info.h ggml.o llama.o common.o ggml-quants-k.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+embedding: examples/embedding/embedding.cpp build-info.h ggml.o llama.o common.o ggml-quants-k.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o common.o ggml-quants-k.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
+
 build-info.h: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
@@ -238,9 +293,9 @@ benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	./$@
 
-vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
+vdot: pocs/vdot/vdot.cpp ggml.o ggml-quants-k.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-.PHONY: tests
+.PHONY: tests clean
 tests:
 	bash ./tests/run-tests.sh

README.md

@@ -9,9 +9,11 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 **Hot topics:**
 
-- Quantization formats `Q4` and `Q8` have changed again (19 May) - [(info)](https://github.com/ggerganov/llama.cpp/pull/1508)
-- Quantization formats `Q4` and `Q5` have changed - requantize any old models [(info)](https://github.com/ggerganov/llama.cpp/pull/1405)
-- [Roadmap May 2023](https://github.com/ggerganov/llama.cpp/discussions/1220)
+- GPU support with Metal (Apple Silicon): https://github.com/ggerganov/llama.cpp/pull/1642
+- High-quality 2,3,4,5,6-bit quantization: https://github.com/ggerganov/llama.cpp/pull/1684
+- Multi-GPU support: https://github.com/ggerganov/llama.cpp/pull/1607
+- Training LLaMA models from scratch: https://github.com/ggerganov/llama.cpp/pull/1652
+- CPU threading improvements: https://github.com/ggerganov/llama.cpp/pull/1632
 
 <details>
   <summary>Table of Contents</summary>
@@ -51,12 +53,11 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quantization on a MacBook
 
 - Plain C/C++ implementation without dependencies
-- Apple silicon first-class citizen - optimized via ARM NEON and Accelerate framework
+- Apple silicon first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
 - Mixed F16 / F32 precision
 - 4-bit, 5-bit and 8-bit integer quantization support
-- Runs on the CPU
-- OpenBLAS support
+- Supports OpenBLAS/Apple BLAS/ARM Performance Lib/ATLAS/BLIS/Intel MKL/NVHPC/ACML/SCSL/SGIMATH and [more](https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors) in BLAS
 - cuBLAS and CLBlast support
 
 The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022).
@@ -236,15 +237,41 @@ In order to build llama.cpp you have three different options.
     zig build -Drelease-fast
     ```
 
+### Metal Build
+
+Using Metal allows the computation to be executed on the GPU for Apple devices:
+
+- Using `make`:
+
+  ```bash
+  LLAMA_METAL=1 make
+  ```
+
+- Using `CMake`:
+
+    ```bash
+    mkdir build-metal
+    cd build-metal
+    cmake -DLLAMA_METAL=ON ..
+    cmake --build . --config Release
+    ```
+
+When built with Metal support, you can enable GPU inference with the `--gpu-layers|-ngl` command-line argument.
+Any value larger than 0 will offload the computation to the GPU. For example:
+
+```bash
+./main -m ./models/7B/ggml-model-q4_0.bin -n 128 -ngl 1
+```
+
 ### BLAS Build
 
 Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). BLAS doesn't affect the normal generation performance. There are currently three different implementations of it:
 
-- Accelerate Framework:
+- **Accelerate Framework**:
 
   This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
 
-- OpenBLAS:
+- **OpenBLAS**:
 
   This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine.
 
@@ -274,11 +301,26 @@ Building the program with BLAS support may lead to some performance improvements
       ```bash
       mkdir build
       cd build
-      cmake .. -DLLAMA_OPENBLAS=ON
+      cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS
       cmake --build . --config Release
       ```
 
-- cuBLAS
+- **BLIS**
+
+  Check [BLIS.md](BLIS.md) for more information.
+
+- **Intel MKL**
+
+  By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. You may also specify it by:
+
+  ```bash
+  mkdir build
+  cd build
+  cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+  cmake --build . -config Release
+  ```
+
+- **cuBLAS**
 
   This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
   - Using `make`:
@@ -293,8 +335,83 @@ Building the program with BLAS support may lead to some performance improvements
     cmake .. -DLLAMA_CUBLAS=ON
     cmake --build . --config Release
     ```
+  Note: Because llama.cpp uses multiple CUDA streams for matrix multiplication results [are not guaranteed to be reproducible](https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility). If you need reproducibility, set `GGML_CUDA_MAX_STREAMS` in the file `ggml-cuda.cu` to 1.
 
-Note: Because llama.cpp uses multiple CUDA streams for matrix multiplication results [are not guaranteed to be reproducible](https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility). If you need reproducibility, set `GGML_CUDA_MAX_STREAMS` in the file `ggml-cuda.cu` to 1.
+  The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used.
+
+- **CLBlast**
+
+  OpenCL acceleration is provided by the matrix multiplication kernels from the [CLBlast](https://github.com/CNugteren/CLBlast) project and custom kernels for ggml that can generate tokens on the GPU.
+
+  You will need the [OpenCL SDK](https://github.com/KhronosGroup/OpenCL-SDK).
+    - For Ubuntu or Debian, the packages `opencl-headers`, `ocl-icd` may be needed.
+
+    - <details>
+        <summary>Installing the OpenCL SDK from source</summary>
+
+        ```sh
+        git clone --recurse-submodules https://github.com/KhronosGroup/OpenCL-SDK.git
+        mkdir OpenCL-SDK/build
+        cd OpenCL-SDK/build
+        cmake .. -DBUILD_DOCS=OFF \
+          -DBUILD_EXAMPLES=OFF \
+          -DBUILD_TESTING=OFF \
+          -DOPENCL_SDK_BUILD_SAMPLES=OFF \
+          -DOPENCL_SDK_TEST_SAMPLES=OFF
+        cmake --build . --config Release
+        cmake --install . --prefix /some/path
+        ```
+      </details>
+
+  Installing CLBlast: it may be found in your operating system's packages.
+
+  - <details>
+    <summary>If not, then installing from source:</summary>
+
+      ```sh
+      git clone https://github.com/CNugteren/CLBlast.git
+      mkdir CLBlast/build
+      cd CLBLast/build
+      cmake .. -DBUILD_SHARED_LIBS=OFF -DTUNERS=OFF
+      cmake --build . --config Release
+      cmake --install . --prefix /some/path
+      ```
+
+      Where `/some/path` is where the built library will be installed (default is `/usr/local`).
+    </details>
+
+  Building:
+
+  - Build with make:
+    ```sh
+    make LLAMA_CLBLAST=1
+    ```
+  - CMake:
+    ```sh
+    mkdir build
+    cd build
+    cmake .. -DLLAMA_CLBLAST=ON -DCLBlast_dir=/some/path
+    cmake --build . --config Release
+    ```
+
+  Running:
+
+  The CLBlast build supports `--gpu-layers|-ngl` like the CUDA version does.
+
+  To select the correct platform (driver) and device (GPU), you can use the environment variables `GGML_OPENCL_PLATFORM` and `GGML_OPENCL_DEVICE`.
+  The selection can be a number (starting from 0) or a text string to search:
+
+  ```sh
+  GGML_OPENCL_PLATFORM=1 ./main ...
+  GGML_OPENCL_DEVICE=2 ./main ...
+  GGML_OPENCL_PLATFORM=Intel ./main ...
+  GGML_OPENCL_PLATFORM=AMD GGML_OPENCL_DEVICE=1 ./main ...
+  ```
+
+  The default behavior is to find the first GPU device, but when it is an integrated GPU on a laptop, for instance, the selectors are useful.
+  Using the variables it is possible to select a CPU-based driver as well, if so desired.
+
+  You can get a list of platforms and devices from the `clinfo -l` command, etc.
 
 ### Prepare Data & Run
 
@@ -376,6 +493,25 @@ Note the use of `--color` to distinguish between user input and generated text.
 
 ![image](https://user-images.githubusercontent.com/1991296/224575029-2af3c7dc-5a65-4f64-a6bb-517a532aea38.png)
 
+### Persistent Interaction
+
+The prompt, user inputs, and model generations can be saved and resumed across calls to `./main` by leveraging `--prompt-cache` and `--prompt-cache-all`. The `./examples/chat-persistent.sh` script demonstrates this with support for long-running, resumable chat sessions. To use this example, you must provide a file to cache the initial chat prompt and a directory to save the chat session, and may optionally provide the same variables as `chat-13B.sh`. The same prompt cache can be reused for new chat sessions. Note that both prompt cache and chat directory are tied to the initial prompt (`PROMPT_TEMPLATE`) and the model file.
+
+```bash
+# Start a new chat
+PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/default ./examples/chat-persistent.sh
+
+# Resume that chat
+PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/default ./examples/chat-persistent.sh
+
+# Start a different chat with the same prompt/model
+PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/another ./examples/chat-persistent.sh
+
+# Different prompt cache for different prompt/model
+PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
+    CHAT_SAVE_DIR=./chat/bob ./examples/chat-persistent.sh
+```
+
 ### Instruction mode with Alpaca
 
 1. First, download the `ggml` Alpaca model into the `./models` folder

convert-pth-to-ggml.py

@@ -4,7 +4,9 @@ import argparse
 
 import convert
 
-parser = argparse.ArgumentParser(description='Convert a LLaMA model checkpoint to a ggml compatible file')
+parser = argparse.ArgumentParser(
+    description="""[DEPRECATED - use `convert.py` instead]
+    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()

examples/CMakeLists.txt

@@ -37,4 +37,10 @@ else()
     add_subdirectory(save-load-state)
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
+    if (LLAMA_METAL)
+        add_subdirectory(metal)
+    endif()
+    if (LLAMA_BUILD_SERVER)
+        add_subdirectory(server)
+    endif()
 endif()

examples/chat-persistent.sh

@@ -23,8 +23,8 @@ CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin"
 NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt"
 NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin"
 
-SESSION_SIZE_MSG_PATTERN='main: session file matches \d+ / \d+'
-SAMPLE_TIME_MSG_PATTERN='sample time =\s+\d+.\d+ ms /\s+\d+'
+SESSION_SIZE_MSG_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'
+SAMPLE_TIME_MSG_PATTERN='sample time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
 SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
 
 CTX_SIZE=2048

examples/common.cpp

@@ -251,6 +251,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.model = argv[i];
+        } else if (arg == "-a" || arg == "--alias") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model_alias = argv[i];
         } else if (arg == "--lora") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -283,11 +289,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
             params.n_gpu_layers = std::stoi(argv[i]);
+#else
+            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
+            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+#endif
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--mtest") {
             params.mem_test = true;
+        } else if (arg == "--export") {
+            params.export_cgraph = true;
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
         } else if (arg == "-r" || arg == "--reverse-prompt") {
@@ -410,7 +423,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
     fprintf(stderr, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     fprintf(stderr, "  --no-penalize-nl      do not penalize newline token\n");
-    fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value\n");
+    fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    fprintf(stderr, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
     fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
     fprintf(stderr, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     fprintf(stderr, "  --perplexity          compute perplexity over the prompt\n");
@@ -421,9 +435,12 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     if (llama_mmap_supported()) {
         fprintf(stderr, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
     fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
     fprintf(stderr, "                        number of layers to store in VRAM\n");
+#endif
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
+    fprintf(stderr, "  --export              export the computation graph to 'llama.ggml'\n");
     fprintf(stderr, "  --verbose-prompt      print prompt before generation\n");
     fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");

examples/common.h

@@ -45,6 +45,7 @@ struct gpt_params {
     float   mirostat_eta      = 0.10f; // learning rate
 
     std::string model             = "models/7B/ggml-model.bin"; // model path
+    std::string model_alias       = "unknown"; // model alias
     std::string prompt            = "";
     std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
     std::string input_prefix      = "";  // string to prefix user inputs with
@@ -70,6 +71,7 @@ struct gpt_params {
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool mem_test          = false; // compute maximum memory usage
+    bool export_cgraph     = false; // export the computation graph
     bool verbose_prompt    = false; // print prompt tokens before generation
 };
 

examples/main/README.md

@@ -69,8 +69,8 @@ In this section, we cover the most commonly used options for running the `main`
 -   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
 -   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
 -   `-ins, --instruct`: Run the program in instruction mode, which is particularly useful when working with Alpaca models.
--   `-n N, --n_predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
--   `-c N, --ctx_size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+-   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
+-   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
 
 ## Input Prompts
 
@@ -136,9 +136,9 @@ During text generation, LLaMA models have a limited context size, which means th
 
 ### Context Size
 
-The `--ctx_size` option allows you to set the size of the prompt context used by the LLaMA models during text generation. A larger context size helps the model to better comprehend and generate responses for longer input or conversations.
+The `--ctx-size` option allows you to set the size of the prompt context used by the LLaMA models during text generation. A larger context size helps the model to better comprehend and generate responses for longer input or conversations.
 
--   `-c N, --ctx_size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
+-   `-c N, --ctx-size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
 
 ### Keep Prompt
 
@@ -146,7 +146,7 @@ The `--keep` option allows users to retain the original prompt when the model ru
 
 -   `--keep N`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context. By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
 
-By utilizing context management options like `--ctx_size` and `--keep`, you can maintain a more coherent and consistent interaction with the LLaMA models, ensuring that the generated text remains relevant to the original prompt or conversation.
+By utilizing context management options like `--ctx-size` and `--keep`, you can maintain a more coherent and consistent interaction with the LLaMA models, ensuring that the generated text remains relevant to the original prompt or conversation.
 
 ## Generation Flags
 
@@ -154,11 +154,11 @@ The following options allow you to control the text generation process and fine-
 
 ### Number of Tokens to Predict
 
--   `-n N, --n_predict N`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity).
+-   `-n N, --n-predict N`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity).
 
-The `--n_predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text. A value of -1 will cause text to be generated without limit.
+The `--n-predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text. A value of -1 will cause text to be generated without limit.
 
-It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `n_predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter.
+It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `n-predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter.
 
 ### Temperature
 
@@ -170,33 +170,33 @@ Example usage: `--temp 0.5`
 
 ### Repeat Penalty
 
--   `--repeat_penalty N`: Control the repetition of token sequences in the generated text (default: 1.1).
--   `--repeat_last_n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx_size).
+-   `--repeat-penalty N`: Control the repetition of token sequences in the generated text (default: 1.1).
+-   `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
 -   `--no-penalize-nl`: Disable penalization for newline tokens when applying the repeat penalty.
 
-The `repeat_penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.1.
+The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.1.
 
-The `repeat_last_n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx_size`).
+The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx-size`).
 
 Use the `--no-penalize-nl` option to disable newline penalization when applying the repeat penalty. This option is particularly useful for generating chat conversations, dialogues, code, poetry, or any text where newline tokens play a significant role in structure and formatting. Disabling newline penalization helps maintain the natural flow and intended formatting in these specific use cases.
 
-Example usage: `--repeat_penalty 1.15 --repeat_last_n 128 --no-penalize-nl`
+Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl`
 
 ### Top-K Sampling
 
--   `--top_k N`: Limit the next token selection to the K most probable tokens (default: 40).
+-   `--top-k N`: Limit the next token selection to the K most probable tokens (default: 40).
 
-Top-k sampling is a text generation method that selects the next token only from the top k most likely tokens predicted by the model. It helps reduce the risk of generating low-probability or nonsensical tokens, but it may also limit the diversity of the output. A higher value for top_k (e.g., 100) will consider more tokens and lead to more diverse text, while a lower value (e.g., 10) will focus on the most probable tokens and generate more conservative text. The default value is 40.
+Top-k sampling is a text generation method that selects the next token only from the top k most likely tokens predicted by the model. It helps reduce the risk of generating low-probability or nonsensical tokens, but it may also limit the diversity of the output. A higher value for top-k (e.g., 100) will consider more tokens and lead to more diverse text, while a lower value (e.g., 10) will focus on the most probable tokens and generate more conservative text. The default value is 40.
 
-Example usage: `--top_k 30`
+Example usage: `--top-k 30`
 
 ### Top-P Sampling
 
--   `--top_p N`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
+-   `--top-p N`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
 
-Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top_p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. The default value is 0.9.
+Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top-p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. The default value is 0.9.
 
-Example usage: `--top_p 0.95`
+Example usage: `--top-p 0.95`
 
 ### Tail Free Sampling (TFS)
 
@@ -217,16 +217,16 @@ Example usage: `--typical 0.9`
 ### Mirostat Sampling
 
 -   `--mirostat N`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0).
--   `--mirostat_lr N`: Set the Mirostat learning rate, parameter eta (default: 0.1).
--   `--mirostat_ent N`: Set the Mirostat target entropy, parameter tau (default: 5.0).
+-   `--mirostat-lr N`: Set the Mirostat learning rate, parameter eta (default: 0.1).
+-   `--mirostat-ent N`: Set the Mirostat target entropy, parameter tau (default: 5.0).
 
 Mirostat is an algorithm that actively maintains the quality of generated text within a desired range during text generation. It aims to strike a balance between coherence and diversity, avoiding low-quality output caused by excessive repetition (boredom traps) or incoherence (confusion traps).
 
-The `--mirostat_lr` option sets the Mirostat learning rate (eta). The learning rate influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. The default value is `0.1`.
+The `--mirostat-lr` option sets the Mirostat learning rate (eta). The learning rate influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. The default value is `0.1`.
 
-The `--mirostat_ent` option sets the Mirostat target entropy (tau), which represents the desired perplexity value for the generated text. Adjusting the target entropy allows you to control the balance between coherence and diversity in the generated text. A lower value will result in more focused and coherent text, while a higher value will lead to more diverse and potentially less coherent text. The default value is `5.0`.
+The `--mirostat-ent` option sets the Mirostat target entropy (tau), which represents the desired perplexity value for the generated text. Adjusting the target entropy allows you to control the balance between coherence and diversity in the generated text. A lower value will result in more focused and coherent text, while a higher value will lead to more diverse and potentially less coherent text. The default value is `5.0`.
 
-Example usage: `--mirostat 2 --mirostat_lr 0.05 --mirostat_ent 3.0`
+Example usage: `--mirostat 2 --mirostat-lr 0.05 --mirostat-ent 3.0`
 
 ### Logit Bias
 
@@ -264,15 +264,15 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### Memory Float 32
 
--   `--memory_f32`: Use 32-bit floats instead of 16-bit floats for memory key+value, allowing higher quality inference at the cost of higher memory usage.
+-   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement and cached prompt file size but does not appear to increase generation quality in a measurable way. Not recommended.
 
 ### Batch Size
 
--   `-b N, --batch_size N`: Set the batch size for prompt processing (default: 512). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
+-   `-b N, --batch-size N`: Set the batch size for prompt processing (default: 512). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
 
 ### Prompt Caching
 
--   `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs.
+-   `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs. **Note**: Restoring a cached prompt does not imply restoring the exact state of the session at the point it was saved. So even when specifying a specific seed, you are not guaranteed to get the same sequence of tokens as the original generation.
 
 ### Quantization
 
@@ -285,5 +285,6 @@ These options provide extra functionality and customization when running the LLa
 -   `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated.
 -   `--verbose-prompt`: Print the prompt before generating text.
 -   `--mtest`: Test the model's functionality by running a series of tests to ensure it's working properly.
+-   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.

examples/main/main.cpp

@@ -134,8 +134,13 @@ int main(int argc, char ** argv) {
         return 0;
     }
 
-    // Add a space in front of the first character to match OG llama tokenizer behavior
-    params.prompt.insert(0, 1, ' ');
+    // export the cgraph and exit
+    if (params.export_cgraph) {
+        llama_eval_export(ctx, "llama.ggml");
+        llama_free(ctx);
+
+        return 0;
+    }
 
     std::string path_session = params.path_prompt_cache;
     std::vector<llama_token> session_tokens;
@@ -155,6 +160,7 @@ int main(int argc, char ** argv) {
                 return 1;
             }
             session_tokens.resize(n_token_count_out);
+            llama_set_rng_seed(ctx, params.seed);
 
             fprintf(stderr, "%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
         } else {
@@ -163,7 +169,16 @@ int main(int argc, char ** argv) {
     }
 
     // tokenize the prompt
-    auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+    std::vector<llama_token> embd_inp;
+
+    if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
+        // Add a space in front of the first character to match OG llama tokenizer behavior
+        params.prompt.insert(0, 1, ' ');
+
+        embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+    } else {
+        embd_inp = session_tokens;
+    }
 
     const int n_ctx = llama_n_ctx(ctx);
 
@@ -181,7 +196,9 @@ int main(int argc, char ** argv) {
             }
             n_matching_session_tokens++;
         }
-        if (n_matching_session_tokens >= embd_inp.size()) {
+        if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {
+            fprintf(stderr, "%s: using full prompt from session file\n", __func__);
+        } else if (n_matching_session_tokens >= embd_inp.size()) {
             fprintf(stderr, "%s: session file has exact match for prompt!\n", __func__);
         } else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
             fprintf(stderr, "%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
@@ -192,6 +209,13 @@ int main(int argc, char ** argv) {
         }
     }
 
+    // if we will use the cache for the full prompt without reaching the end of the cache, force
+    // reevaluation of the last token token to recalculate the cached logits
+    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() &&
+            session_tokens.size() > embd_inp.size()) {
+        session_tokens.resize(embd_inp.size() - 1);
+    }
+
     // number of tokens to keep when resetting context
     if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct) {
         params.n_keep = (int)embd_inp.size();

examples/metal/CMakeLists.txt

@@ -0,0 +1,3 @@
+set(TEST_TARGET metal)
+add_executable(${TEST_TARGET} metal.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml)

examples/metal/metal.cpp

@@ -0,0 +1,102 @@
+// Evaluate a statically exported ggml computation graph with Metal
+//
+// - First, export a LLaMA graph:
+//
+//  $ ./bin/main -m ../models/7B/ggml-model-q4_0.bin --export
+//
+// - Run this tool to evaluate the exported graph:
+//
+//  $ ./bin/metal llama.ggml
+//
+// The purpose of this tool is mostly for debugging and demonstration purposes.
+// The main limitation of exporting computation graphs is that their sizes are static which often
+// can be a problem for real-world applications.
+//
+
+#include "ggml.h"
+#include "ggml-metal.h"
+
+#include <cstdio>
+#include <cstring>
+#include <cstdlib>
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    if (argc != 2) {
+        fprintf(stderr, "Usage: %s llama.ggml\n", argv[0]);
+        return -1;
+    }
+
+    const char * fname_cgraph = argv[1];
+
+    // load the compute graph
+    struct ggml_context * ctx_data = NULL;
+    struct ggml_context * ctx_eval = NULL;
+
+    struct ggml_cgraph gf = ggml_graph_import(fname_cgraph, &ctx_data, &ctx_eval);
+    gf.n_threads = 1;
+
+    // this allocates all Metal resources and memory buffers
+    auto * ctx_metal = ggml_metal_init();
+
+    ggml_metal_add_buffer(ctx_metal, "data", ggml_get_mem_buffer(ctx_data), ggml_get_mem_size(ctx_data));
+    ggml_metal_add_buffer(ctx_metal, "eval", ggml_get_mem_buffer(ctx_eval), ggml_get_mem_size(ctx_eval));
+
+    // main
+    {
+        struct ggml_tensor * input = ggml_graph_get_tensor(&gf, "embd");
+        *(int32_t *) input->data = 1; // BOS
+
+        ggml_metal_set_tensor(ctx_metal, input);
+
+        // warmup
+        ggml_metal_graph_compute(ctx_metal, &gf);
+
+        const int n_iter = 16;
+
+        const int64_t t0 = ggml_time_us();
+
+        // the actual inference happens here
+        for (int i = 0; i < n_iter; ++i) {
+            ggml_metal_graph_compute(ctx_metal, &gf);
+        }
+
+        const int64_t t1 = ggml_time_us();
+
+        printf("time: %.2f ms, %.2f ms/tok\n", (t1 - t0) / 1000.0, (t1 - t0) / 1000.0 / n_iter);
+    }
+
+    // debug output
+    {
+        struct ggml_tensor * logits = gf.nodes[gf.n_nodes - 1];
+        ggml_metal_get_tensor(ctx_metal, logits);
+
+        float * ptr = (float *) ggml_get_data(logits);
+
+        printf("logits: ");
+        for (int i = 0; i < 10; i++) {
+            printf("%8.4f ", ptr[i]);
+        }
+        printf("\n");
+        int imax = 0;
+        double sum = 0.0;
+        double vmax = -1e9;
+        for (int i = 0; i < 32000; i++) {
+            sum += (double) ptr[i];
+            if (ptr[i] > vmax) {
+                vmax = ptr[i];
+                imax = i;
+            }
+        }
+        printf("sum: %f, imax = %d, vmax = %f\n", sum, imax, vmax);
+    }
+
+    ggml_metal_free(ctx_metal);
+
+    ggml_free(ctx_data);
+    ggml_free(ctx_eval);
+
+    return 0;
+}
+

examples/quantize-stats/quantize-stats.cpp

@@ -282,8 +282,9 @@ int main(int argc, char ** argv) {
                 break;
             }
             int j;
-            for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], ggml_type_name((ggml_type) j)) != 0; j++) {
-                // find match
+            for (j = 0; j < GGML_TYPE_COUNT; ++j) {
+               const auto * name = ggml_type_name((ggml_type) j);
+               if (name && strcmp(argv[i], name) == 0) break;
             }
             if (j < GGML_TYPE_COUNT) {
                 params.include_types.push_back((ggml_type) j);

examples/quantize/quantize.cpp

@@ -7,11 +7,23 @@
 #include <string>
 
 static const std::map<std::string, llama_ftype> LLAMA_FTYPE_MAP = {
-  {"q4_0", LLAMA_FTYPE_MOSTLY_Q4_0},
-  {"q4_1", LLAMA_FTYPE_MOSTLY_Q4_1},
-  {"q5_0", LLAMA_FTYPE_MOSTLY_Q5_0},
-  {"q5_1", LLAMA_FTYPE_MOSTLY_Q5_1},
-  {"q8_0", LLAMA_FTYPE_MOSTLY_Q8_0},
+  {"q4_0",   LLAMA_FTYPE_MOSTLY_Q4_0},
+  {"q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1},
+  {"q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0},
+  {"q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1},
+  {"q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0},
+  {"q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K},
+  {"q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M},
+  {"q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S},
+  {"q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M},
+  {"q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L},
+  {"q4_K",   LLAMA_FTYPE_MOSTLY_Q4_K_M},
+  {"q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S},
+  {"q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M},
+  {"q5_K",   LLAMA_FTYPE_MOSTLY_Q5_K_M},
+  {"q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S},
+  {"q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M},
+  {"q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K},
 };
 
 bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::string & ftype_str_out) {

examples/server/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(TARGET server)
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+add_executable(${TARGET} server.cpp json.hpp httplib.h)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/server/README.md

@@ -0,0 +1,312 @@
+# llama.cpp/example/server
+
+This example allow you to have a llama.cpp http server to interact from a web page or consume the API.
+
+## Table of Contents
+
+1. [Quick Start](#quick-start)
+2. [Node JS Test](#node-js-test)
+3. [API Endpoints](#api-endpoints)
+4. [More examples](#more-examples)
+5. [Common Options](#common-options)
+6. [Performance Tuning and Memory Options](#performance-tuning-and-memory-options)
+
+## Quick Start
+
+To get started right away, run the following command, making sure to use the correct path for the model you have:
+
+#### Unix-based systems (Linux, macOS, etc.):
+
+```bash
+./server -m models/7B/ggml-model.bin --ctx_size 2048
+```
+
+#### Windows:
+
+```powershell
+server.exe -m models\7B\ggml-model.bin --ctx_size 2048
+```
+
+That will start a server that by default listens on `127.0.0.1:8080`. You can consume the endpoints with Postman or NodeJS with axios library.
+
+## Node JS Test
+
+You need to have [Node.js](https://nodejs.org/en) installed.
+
+```bash
+mkdir llama-client
+cd llama-client
+npm init
+npm install axios
+```
+
+Create a index.js file and put inside this:
+
+```javascript
+const axios = require("axios");
+
+const prompt = `Building a website can be done in 10 simple steps:`;
+
+async function Test() {
+    let result = await axios.post("http://127.0.0.1:8080/completion", {
+        prompt,
+        batch_size: 128,
+        n_predict: 512,
+    });
+
+    // the response is received until completion finish
+    console.log(result.data.content);
+}
+
+Test();
+```
+
+And run it:
+
+```bash
+node .
+```
+
+## API Endpoints
+
+You can interact with this API Endpoints. This implementations just support chat style interaction.
+
+-   **POST** `hostname:port/completion`: Setting up the Llama Context to begin the completions tasks.
+
+*Options:*
+
+`batch_size`: Set the batch size for prompt processing (default: 512).
+
+`temperature`: Adjust the randomness of the generated text (default: 0.8).
+
+`top_k`: Limit the next token selection to the K most probable tokens (default: 40).
+
+`top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
+
+`n_predict`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity).
+
+`threads`: Set the number of threads to use during computation.
+
+`n_keep`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context. By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
+
+`as_loop`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
+
+`interactive`: It allows interacting with the completion, and the completion stops as soon as it encounters a `stop word`. To enable this, set to `true`.
+
+`prompt`: Provide a prompt. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate.
+
+`stop`: Specify the words or characters that indicate a stop. These words will not be included in the completion, so make sure to add them to the prompt for the next iteration.
+
+`exclude`: Specify the words or characters you do not want to appear in the completion. These words will not be included in the completion, so make sure to add them to the prompt for the next iteration.
+
+-   **POST** `hostname:port/embedding`: Generate embedding of a given text
+
+*Options:*
+
+`content`: Set the text to get generate the embedding.
+
+`threads`: Set the number of threads to use during computation.
+
+To use this endpoint, you need to start the server with the `--embedding` option added.
+
+-   **POST** `hostname:port/tokenize`: Tokenize a given text
+
+*Options:*
+
+`content`: Set the text to tokenize.
+
+-   **GET** `hostname:port/next-token`: Receive the next token predicted, execute this request in a loop. Make sure set `as_loop` as `true` in the completion request.
+
+*Options:*
+
+`stop`: Set `hostname:port/next-token?stop=true` to stop the token generation.
+
+## More examples
+
+### Interactive mode
+
+This mode allows interacting in a chat-like manner. It is recommended for models designed as assistants such as `Vicuna`, `WizardLM`, `Koala`, among others. Make sure to add the correct stop word for the corresponding model.
+
+The prompt should be generated by you, according to the model's guidelines. You should keep adding the model's completions to the context as well.
+
+This example works well for `Vicuna - version 1`.
+
+```javascript
+const axios = require("axios");
+
+let prompt = `A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.
+### Human: Hello, Assistant.
+### Assistant: Hello. How may I help you today?
+### Human: Please tell me the largest city in Europe.
+### Assistant: Sure. The largest city in Europe is Moscow, the capital of Russia.`;
+
+async function ChatCompletion(answer) {
+    // the user's next question to the prompt
+    prompt += `\n### Human: ${answer}\n`
+
+    result = await axios.post("http://127.0.0.1:8080/completion", {
+        prompt,
+        batch_size: 128,
+        temperature: 0.2,
+        top_k: 40,
+        top_p: 0.9,
+        n_keep: -1,
+        n_predict: 2048,
+        stop: ["\n### Human:"], // when detect this, stop completion
+        exclude: ["### Assistant:"], // no show in the completion
+        threads: 8,
+        as_loop: true, // use this to request the completion token by token
+        interactive: true, // enable the detection of a stop word
+    });
+
+    // create a loop to receive every token predicted
+    // note: this operation is blocking, avoid use this in a ui thread
+
+    let message = "";
+    while (true) {
+        // you can stop the inference adding '?stop=true' like this http://127.0.0.1:8080/next-token?stop=true
+        result = await axios.get("http://127.0.0.1:8080/next-token");
+        process.stdout.write(result.data.content);
+        message += result.data.content;
+
+        // to avoid an infinite loop
+        if (result.data.stop) {
+            console.log("Completed");
+            // make sure to add the completion to the prompt.
+            prompt += `### Assistant: ${message}`;
+            break;
+        }
+    }
+}
+
+// This function should be called every time a question to the model is needed.
+async function Test() {
+    // the server can't inference in paralell
+    await ChatCompletion("Write a long story about a time magician in a fantasy world");
+    await ChatCompletion("Summary the story");
+}
+
+Test();
+```
+
+### Alpaca example
+
+**Temporaly note:** no tested, if you have the model, please test it and report me some issue
+
+```javascript
+const axios = require("axios");
+
+let prompt = `Below is an instruction that describes a task. Write a response that appropriately completes the request.
+`;
+
+async function DoInstruction(instruction) {
+    prompt += `\n\n### Instruction:\n\n${instruction}\n\n### Response:\n\n`;
+    result = await axios.post("http://127.0.0.1:8080/completion", {
+        prompt,
+        batch_size: 128,
+        temperature: 0.2,
+        top_k: 40,
+        top_p: 0.9,
+        n_keep: -1,
+        n_predict: 2048,
+        stop: ["### Instruction:\n\n"], // when detect this, stop completion
+        exclude: [], // no show in the completion
+        threads: 8,
+        as_loop: true, // use this to request the completion token by token
+        interactive: true, // enable the detection of a stop word
+    });
+
+    // create a loop to receive every token predicted
+    // note: this operation is blocking, avoid use this in a ui thread
+
+    let message = "";
+    while (true) {
+        result = await axios.get("http://127.0.0.1:8080/next-token");
+        process.stdout.write(result.data.content);
+        message += result.data.content;
+
+        // to avoid an infinite loop
+        if (result.data.stop) {
+            console.log("Completed");
+            // make sure to add the completion and the user's next question to the prompt.
+            prompt += message;
+            break;
+        }
+    }
+}
+
+// This function should be called every time a instruction to the model is needed.
+DoInstruction("Destroy the world"); // as joke
+```
+
+### Embeddings
+
+First, run the server with `--embedding` option:
+
+```bash
+server -m models/7B/ggml-model.bin --ctx_size 2048 --embedding
+```
+
+Run this code in NodeJS:
+
+```javascript
+const axios = require('axios');
+
+async function Test() {
+    let result = await axios.post("http://127.0.0.1:8080/embedding", {
+        content: `Hello`,
+        threads: 5
+    });
+    // print the embedding array
+    console.log(result.data.embedding);
+}
+
+Test();
+```
+
+### Tokenize
+
+Run this code in NodeJS:
+
+```javascript
+const axios = require('axios');
+
+async function Test() {
+    let result = await axios.post("http://127.0.0.1:8080/tokenize", {
+        content: `Hello`
+    });
+    // print the embedding array
+    console.log(result.data.tokens);
+}
+
+Test();
+```
+
+## Common Options
+
+-   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
+-   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+-   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+-   `--embedding`: Enable the embedding mode. **Completion function doesn't work in this mode**.
+-   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`;
+-   `--port`: Set the port to listen. Default: `8080`.
+
+### RNG Seed
+
+-   `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed).
+
+The RNG seed is used to initialize the random number generator that influences the text generation process. By setting a specific seed value, you can obtain consistent and reproducible results across multiple runs with the same input and settings. This can be helpful for testing, debugging, or comparing the effects of different options on the generated text to see when they diverge. If the seed is set to a value less than 0, a random seed will be used, which will result in different outputs on each run.
+
+## Performance Tuning and Memory Options
+
+### No Memory Mapping
+
+-   `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed. However, if the model is larger than your total amount of RAM or if your system is low on available memory, using mmap might increase the risk of pageouts, negatively impacting performance.
+
+### Memory Float 32
+
+-   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement but does not appear to increase generation quality in a measurable way. Not recommended.
+
+## Limitations:
+
+-   The actual implementation of llama.cpp need a `llama-state` for handle multiple contexts and clients, but this could require more powerful hardware.

examples/server/server.cpp

@@ -0,0 +1,742 @@
+#include <httplib.h>
+#include <json.hpp>
+#include "common.h"
+#include "llama.h"
+
+struct server_params
+{
+  std::string hostname = "127.0.0.1";
+  int32_t port = 8080;
+};
+
+struct llama_server_context
+{
+  bool as_loop = false;
+  bool has_next_token = false;
+  std::string generated_text = "";
+
+  int32_t num_tokens_predicted = 0;
+  int32_t n_past = 0;
+  int32_t n_consumed = 0;
+  int32_t n_session_consumed = 0;
+  int32_t n_remain = 0;
+
+  std::vector<llama_token> embd;
+  std::vector<llama_token> last_n_tokens;
+  std::vector<llama_token> processed_tokens;
+  std::vector<llama_token> llama_token_newline;
+  std::vector<llama_token> embd_inp;
+  std::vector<std::vector<llama_token>> no_show_words;
+  std::vector<llama_token> tokens_predicted;
+
+  llama_context *ctx;
+  gpt_params params;
+
+  void rewind() {
+    as_loop = false;
+    params.antiprompt.clear();
+    no_show_words.clear();
+    num_tokens_predicted = 0;
+    generated_text = "";
+  }
+
+  bool loadModel(gpt_params params_)
+  {
+    params = params_;
+    ctx = llama_init_from_gpt_params(params);
+    if (ctx == NULL)
+    {
+      fprintf(stderr, "%s: error: unable to load model\n", __func__);
+      return false;
+    }
+    // determine newline token
+    llama_token_newline = ::llama_tokenize(ctx, "\n", false);
+    last_n_tokens.resize(params.n_ctx);
+    std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
+    return true;
+  }
+
+  bool loadPrompt() {
+    params.prompt.insert(0, 1, ' '); // always add a first space
+    std::vector<llama_token> prompt_tokens = ::llama_tokenize(ctx, params.prompt, true);
+    // compare the evaluated prompt with the new prompt
+    int new_prompt_len = 0;
+    for (size_t i = 0; i < prompt_tokens.size(); i++) {
+      if (i < processed_tokens.size() &&
+        processed_tokens[i] == prompt_tokens[i])
+      {
+        continue;
+      }
+      else
+      {
+        embd_inp.push_back(prompt_tokens[i]);
+        if(new_prompt_len == 0) {
+          if(int32_t(i) - 1 < n_past) {
+            processed_tokens.erase(processed_tokens.begin() + i, processed_tokens.end());
+          }
+          // Evaluate the new fragment prompt from the last token processed.
+          n_past = processed_tokens.size();
+        }
+        new_prompt_len ++;
+      }
+    }
+    if(n_past > 0 && params.interactive) {
+      n_remain -= new_prompt_len;
+    }
+    if ((int)embd_inp.size() > params.n_ctx - 4)
+    {
+      return false;
+    }
+    has_next_token = true;
+    return true;
+  }
+
+  void beginCompletion()
+  {
+    if(n_remain == 0) {
+      // number of tokens to keep when resetting context
+      if (params.n_keep < 0 || params.n_keep > (int)embd_inp.size())
+      {
+        params.n_keep = (int)embd_inp.size();
+      }
+    }
+    n_remain = params.n_predict;
+  }
+
+  llama_token nextToken() {
+    llama_token result = -1;
+    if (embd.size() > 0)
+    {
+      if (n_past + (int)embd.size() > params.n_ctx)
+      {
+        // Reset context
+        const int n_left = n_past - params.n_keep;
+        n_past = std::max(1, params.n_keep);
+        processed_tokens.erase(processed_tokens.begin() + n_past, processed_tokens.end());
+        embd.insert(embd.begin(), last_n_tokens.begin() + params.n_ctx - n_left / 2 - embd.size(), last_n_tokens.end() - embd.size());
+      }
+      for (int i = 0; i < (int)embd.size(); i += params.n_batch)
+      {
+        int n_eval = (int)embd.size() - i;
+        if (n_eval > params.n_batch)
+        {
+          n_eval = params.n_batch;
+        }
+        if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads))
+        {
+          fprintf(stderr, "%s : failed to eval\n", __func__);
+          has_next_token = false;
+          return result;
+        }
+        n_past += n_eval;
+      }
+    }
+    embd.clear();
+    if ((int)embd_inp.size() <= n_consumed && has_next_token)
+    {
+      // out of user input, sample next token
+      const float temp = params.temp;
+      // const int32_t top_k = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k;
+      const float top_p = params.top_p;
+      const float tfs_z = params.tfs_z;
+      const float typical_p = params.typical_p;
+      const int32_t repeat_last_n = params.repeat_last_n < 0 ? params.n_ctx : params.repeat_last_n;
+      const float repeat_penalty = params.repeat_penalty;
+      const float alpha_presence = params.presence_penalty;
+      const float alpha_frequency = params.frequency_penalty;
+      const int mirostat = params.mirostat;
+      const float mirostat_tau = params.mirostat_tau;
+      const float mirostat_eta = params.mirostat_eta;
+      const bool penalize_nl = params.penalize_nl;
+      llama_token id = 0;
+      {
+        auto logits = llama_get_logits(ctx);
+        auto n_vocab = llama_n_vocab(ctx);
+
+        // Apply params.logit_bias map
+        for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++)
+        {
+          logits[it->first] += it->second;
+        }
+
+        std::vector<llama_token_data> candidates;
+        candidates.reserve(n_vocab);
+        for (llama_token token_id = 0; token_id < n_vocab; token_id++)
+        {
+          candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+        }
+
+        llama_token_data_array candidates_p = {candidates.data(), candidates.size(), false};
+
+        // Apply penalties
+        float nl_logit = logits[llama_token_nl()];
+        auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), params.n_ctx);
+        llama_sample_repetition_penalty(ctx, &candidates_p,
+                                        last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
+                                        last_n_repeat, repeat_penalty);
+        llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
+                                                      last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
+                                                      last_n_repeat, alpha_frequency, alpha_presence);
+        if (!penalize_nl)
+        {
+          logits[llama_token_nl()] = nl_logit;
+        }
+
+        if (temp <= 0)
+        {
+          // Greedy sampling
+          id = llama_sample_token_greedy(ctx, &candidates_p);
+        }
+        else
+        {
+          if (mirostat == 1)
+          {
+            static float mirostat_mu = 2.0f * mirostat_tau;
+            const int mirostat_m = 100;
+            llama_sample_temperature(ctx, &candidates_p, temp);
+            id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
+          }
+          else if (mirostat == 2)
+          {
+            static float mirostat_mu = 2.0f * mirostat_tau;
+            llama_sample_temperature(ctx, &candidates_p, temp);
+            id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
+          }
+          else
+          {
+            // Temperature sampling
+            llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
+            llama_sample_typical(ctx, &candidates_p, typical_p, 1);
+            llama_sample_top_p(ctx, &candidates_p, top_p, 1);
+            llama_sample_temperature(ctx, &candidates_p, temp);
+            id = llama_sample_token(ctx, &candidates_p);
+          }
+        }
+        last_n_tokens.erase(last_n_tokens.begin());
+        last_n_tokens.push_back(id);
+        processed_tokens.push_back(id);
+        num_tokens_predicted++;
+      }
+
+      // replace end of text token with newline token when in interactive mode
+      if (id == llama_token_eos() && params.interactive)
+      {
+        id = llama_token_newline.front();
+        if (params.antiprompt.size() != 0)
+        {
+          // tokenize and inject first reverse prompt
+          const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false);
+          embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
+        }
+      }
+
+      // add it to the context
+      embd.push_back(id);
+      for (auto id : embd)
+      {
+        result = id;
+      }
+      // decrement remaining sampling budget
+      --n_remain;
+    }
+    else
+    {
+      // some user input remains from prompt or interaction, forward it to processing
+      while ((int)embd_inp.size() > n_consumed)
+      {
+        embd.push_back(embd_inp[n_consumed]);
+        last_n_tokens.erase(last_n_tokens.begin());
+        last_n_tokens.push_back(embd_inp[n_consumed]);
+        processed_tokens.push_back(embd_inp[n_consumed]);
+        ++n_consumed;
+        if ((int)embd.size() >= params.n_batch)
+        {
+          break;
+        }
+      }
+    }
+    if (params.interactive && (int)embd_inp.size() <= n_consumed)
+    {
+      // check for reverse prompt
+      if (params.antiprompt.size())
+      {
+        std::string last_output;
+        for (auto id : last_n_tokens)
+        {
+          last_output += llama_token_to_str(ctx, id);
+        }
+        has_next_token = true;
+        // Check if each of the reverse prompts appears at the end of the output.
+        for (std::string &antiprompt : params.antiprompt)
+        {
+          if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos)
+          {
+            has_next_token = false;
+            return result;
+          }
+        }
+      }
+      if (n_past > 0)
+      {
+        has_next_token = true;
+      }
+    }
+
+    if (!embd.empty() && embd.back() == llama_token_eos()) {
+        has_next_token = false;
+    }
+
+    if (params.interactive && n_remain <= 0 && params.n_predict != -1)
+    {
+      n_remain = params.n_predict;
+    }
+    has_next_token = n_remain != 0;
+    return result;
+  }
+
+  std::string doCompletion()
+  {
+    llama_token token = nextToken();
+    if (token == -1) {
+      return "";
+    }
+    tokens_predicted.clear();
+    tokens_predicted.push_back(token);
+
+    // Avoid add the no show words to the response
+    for (std::vector<llama_token> word_tokens : no_show_words)
+    {
+      size_t match_token = 1;
+      if (tokens_predicted.front() == word_tokens.front())
+      {
+        bool execute_matching = true;
+        if (tokens_predicted.size() > 1) { // if previus tokens had been tested
+          for (size_t i = 1; i < word_tokens.size(); i++)
+          {
+            if (i >= tokens_predicted.size()) {
+              match_token = i;
+              break;
+            }
+            if (tokens_predicted[i] == word_tokens[i])
+            {
+              continue;
+            }
+            else
+            {
+              execute_matching = false;
+              break;
+            }
+          }
+        }
+        while (execute_matching) {
+          if (match_token == word_tokens.size()) {
+            return "";
+          }
+          token = nextToken();
+          tokens_predicted.push_back(token);
+          if (token == word_tokens[match_token])
+          { // the token follow the sequence
+            match_token++;
+          }
+          else if (match_token < word_tokens.size())
+          { // no complete all word sequence
+            break;
+          }
+        }
+      }
+    }
+    if(as_loop) {
+      generated_text = "";
+    }
+    for (llama_token tkn : tokens_predicted)
+    {
+      generated_text += llama_token_to_str(ctx, tkn);
+    }
+    return generated_text;
+  }
+
+  std::vector<float> embedding(std::string content, int threads) {
+    content.insert(0, 1, ' ');
+    std::vector<llama_token> tokens = ::llama_tokenize(ctx, content, true);
+    if (tokens.size() > 0)
+    {
+      if (llama_eval(ctx, tokens.data(), tokens.size(), 0, threads))
+      {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        std::vector<float> embeddings_;
+        return embeddings_;
+      }
+    }
+    const int n_embd = llama_n_embd(ctx);
+    const auto embeddings = llama_get_embeddings(ctx);
+    std::vector<float> embeddings_(embeddings, embeddings + n_embd);
+    return embeddings_;
+  }
+};
+
+using namespace httplib;
+
+using json = nlohmann::json;
+
+void server_print_usage(int /*argc*/, char **argv, const gpt_params &params)
+{
+  fprintf(stderr, "usage: %s [options]\n", argv[0]);
+  fprintf(stderr, "\n");
+  fprintf(stderr, "options:\n");
+  fprintf(stderr, "  -h, --help            show this help message and exit\n");
+  fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
+  fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+  fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+  fprintf(stderr, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
+  fprintf(stderr, "  --embedding           enable embedding mode\n");
+  fprintf(stderr, "  --keep                number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
+  if (llama_mlock_supported())
+  {
+    fprintf(stderr, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+  }
+  if (llama_mmap_supported())
+  {
+    fprintf(stderr, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+  }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+  fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
+  fprintf(stderr, "                        number of layers to store in VRAM\n");
+#endif
+  fprintf(stderr, "  -m FNAME, --model FNAME\n");
+  fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+  fprintf(stderr, "  -a ALIAS, --alias ALIAS\n");
+  fprintf(stderr, "                        set an alias for the model, will be added as `model` field in completion response\n");
+  fprintf(stderr, "  --host                ip address to listen (default 127.0.0.1)\n");
+  fprintf(stderr, "  --port PORT           port to listen (default 8080)\n");
+  fprintf(stderr, "\n");
+}
+
+bool server_params_parse(int argc, char **argv, server_params &sparams, gpt_params &params)
+{
+  gpt_params default_params;
+  std::string arg;
+  bool invalid_param = false;
+
+  for (int i = 1; i < argc; i++)
+  {
+    arg = argv[i];
+    if (arg == "--port")
+    {
+      if (++i >= argc)
+      {
+        invalid_param = true;
+        break;
+      }
+      sparams.port = std::stoi(argv[i]);
+    }
+    else if (arg == "--host")
+    {
+      if (++i >= argc)
+      {
+        invalid_param = true;
+        break;
+      }
+      sparams.hostname = argv[i];
+    }
+    else if (arg == "-s" || arg == "--seed")
+    {
+#if defined(GGML_USE_CUBLAS)
+      fprintf(stderr, "WARNING: when using cuBLAS generation results are NOT guaranteed to be reproducible.\n");
+#endif
+      if (++i >= argc)
+      {
+        invalid_param = true;
+        break;
+      }
+      params.seed = std::stoi(argv[i]);
+    }
+    else if (arg == "-m" || arg == "--model")
+    {
+      if (++i >= argc)
+      {
+        invalid_param = true;
+        break;
+      }
+      params.model = argv[i];
+    }
+    else if (arg == "-a" || arg == "--alias")
+    {
+      if (++i >= argc)
+      {
+        invalid_param = true;
+        break;
+      }
+      params.model_alias = argv[i];
+    }
+    else if (arg == "--embedding")
+    {
+      params.embedding = true;
+    }
+    else if (arg == "-h" || arg == "--help")
+    {
+      server_print_usage(argc, argv, default_params);
+      exit(0);
+    }
+    else if (arg == "-c" || arg == "--ctx-size" || arg == "--ctx_size")
+    {
+      if (++i >= argc)
+      {
+        invalid_param = true;
+        break;
+      }
+      params.n_ctx = std::stoi(argv[i]);
+    }
+    else if (arg == "--memory-f32" || arg == "--memory_f32")
+    {
+      params.memory_f16 = false;
+    }
+    else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers")
+    {
+      if (++i >= argc)
+      {
+        invalid_param = true;
+        break;
+      }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+      params.n_gpu_layers = std::stoi(argv[i]);
+#else
+      fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
+      fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+#endif
+    }
+    else
+    {
+      fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+      server_print_usage(argc, argv, default_params);
+      exit(1);
+    }
+  }
+
+  if (invalid_param)
+  {
+    fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+    server_print_usage(argc, argv, default_params);
+    exit(1);
+  }
+  return true;
+}
+
+bool parse_options_completion(json body, llama_server_context& llama, Response &res) {
+  if (!body["threads"].is_null())
+  {
+    llama.params.n_threads = body["threads"].get<int>();
+  }
+  if (!body["n_predict"].is_null())
+  {
+    llama.params.n_predict = body["n_predict"].get<int>();
+  }
+  if (!body["top_k"].is_null())
+  {
+    llama.params.top_k = body["top_k"].get<int>();
+  }
+  if (!body["top_p"].is_null())
+  {
+    llama.params.top_p = body["top_p"].get<float>();
+  }
+  if (!body["temperature"].is_null())
+  {
+    llama.params.temp = body["temperature"].get<float>();
+  }
+  if (!body["batch_size"].is_null())
+  {
+    llama.params.n_batch = body["batch_size"].get<int>();
+  }
+  if (!body["n_keep"].is_null())
+  {
+    llama.params.n_keep = body["n_keep"].get<int>();
+  }
+  if (!body["as_loop"].is_null())
+  {
+    llama.as_loop = body["as_loop"].get<bool>();
+  }
+  if (!body["interactive"].is_null())
+  {
+    llama.params.interactive = body["interactive"].get<bool>();
+  }
+  if (!body["prompt"].is_null())
+  {
+    llama.params.prompt = body["prompt"].get<std::string>();
+  }
+  else
+  {
+    json data = {
+        {"status", "error"},
+        {"reason", "You need to pass the prompt"}};
+    res.set_content(data.dump(), "application/json");
+    res.status = 400;
+    return false;
+  }
+  if (!body["stop"].is_null())
+  {
+    std::vector<std::string> stop_words = body["stop"].get<std::vector<std::string>>();
+    for (std::string stop_word : stop_words)
+    {
+      llama.params.antiprompt.push_back(stop_word);
+      llama.no_show_words.push_back(::llama_tokenize(llama.ctx, stop_word, false));
+    }
+  }
+  if (!body["exclude"].is_null())
+  {
+    std::vector<std::string> no_show_words = body["exclude"].get<std::vector<std::string>>();
+    for (std::string no_show : no_show_words)
+    {
+      llama.no_show_words.push_back(::llama_tokenize(llama.ctx, no_show, false));
+    }
+  }
+  return true;
+}
+
+int main(int argc, char **argv)
+{
+  // own arguments required by this example
+  gpt_params params;
+  server_params sparams;
+
+  // struct that contains llama context and inference
+  llama_server_context llama;
+  params.model = "ggml-model.bin";
+
+  if (server_params_parse(argc, argv, sparams, params) == false)
+  {
+    return 1;
+  }
+
+  if (params.seed <= 0)
+  {
+    params.seed = time(NULL);
+  }
+
+  fprintf(stderr, "%s: seed = %d\n", __func__, params.seed);
+
+  // load the model
+  if (!llama.loadModel(params))
+  {
+    return 1;
+  }
+
+  Server svr;
+
+  svr.Get("/", [](const Request &, Response &res)
+          { res.set_content("<h1>llama.cpp server works</h1>", "text/html"); });
+
+  svr.Post("/completion", [&llama](const Request &req, Response &res)
+            {
+              if(llama.params.embedding) {
+                json data = {
+                    {"status", "error"},
+                    {"reason", "To use completion function disable embedding mode"}};
+                res.set_content(data.dump(), "application/json");
+                res.status = 400;
+                return;
+              }
+
+              llama.rewind();
+
+              if(parse_options_completion(json::parse(req.body), llama, res) == false){
+                return;
+              }
+
+              if (!llama.loadPrompt())
+              {
+                json data = {
+                    {"status", "error"},
+                    {"reason", "Context too long, please be more specific"}};
+                res.set_content(data.dump(), "application/json");
+                res.status = 400;
+                return;
+              }
+
+              llama.beginCompletion();
+              if(llama.as_loop) {
+                json data = {
+                    {"status", "done" } };
+                return res.set_content(data.dump(), "application/json");
+              } else {
+                // loop inference until finish completion
+                while (llama.has_next_token)
+                {
+                  llama.doCompletion();
+                }
+                try
+                {
+                  json data = {
+                      {"model", llama.params.model_alias },
+                      {"content", llama.generated_text },
+                      {"tokens_predicted", llama.num_tokens_predicted}};
+                  return res.set_content(data.dump(), "application/json");
+                }
+                catch (const json::exception &e)
+                {
+                  // Some tokens have bad UTF-8 strings, the json parser is very sensitive
+                  json data = {
+                      {"content", "Bad encoding token"},
+                      {"tokens_predicted", 0}};
+                  return res.set_content(data.dump(), "application/json");
+                }
+              } });
+
+  svr.Post("/tokenize", [&llama](const Request &req, Response &res)
+            {
+              json body = json::parse(req.body);
+              json data = {
+                    {"tokens", ::llama_tokenize(llama.ctx, body["content"].get<std::string>(), false) } };
+                return res.set_content(data.dump(), "application/json");
+            });
+
+  svr.Post("/embedding", [&llama](const Request &req, Response &res)
+            {
+              if(!llama.params.embedding) {
+                std::vector<float> empty;
+                json data = {
+                    {"embedding", empty}};
+                fprintf(stderr, "[llama-server] : You need enable embedding mode adding: --embedding option\n");
+                return res.set_content(data.dump(), "application/json");
+              }
+              json body = json::parse(req.body);
+              std::string content = body["content"].get<std::string>();
+              int threads = body["threads"].get<int>();
+              json data = {
+                    {"embedding", llama.embedding(content, threads) } };
+              return res.set_content(data.dump(), "application/json");
+            });
+
+  svr.Get("/next-token", [&llama](const Request &req, Response &res)
+          {
+            if(llama.params.embedding) {
+                res.set_content("{}", "application/json");
+                return;
+            }
+            std::string result = "";
+            if (req.has_param("stop")) {
+                llama.has_next_token = false;
+            } else {
+              result = llama.doCompletion(); // inference next token
+            }
+            try {
+              json data = {
+                        {"content", result },
+                        {"stop", !llama.has_next_token }};
+              return res.set_content(data.dump(), "application/json");
+            } catch (const json::exception &e) {
+              // Some tokens have bad UTF-8 strings, the json parser is very sensitive
+              json data = {
+                        {"content", "" },
+                        {"stop", !llama.has_next_token }};
+              return res.set_content(data.dump(), "application/json");
+            }
+          });
+
+  fprintf(stderr, "%s: http server Listening at http://%s:%i\n", __func__, sparams.hostname.c_str(), sparams.port);
+
+  if(params.embedding) {
+    fprintf(stderr, "NOTE: Mode embedding enabled. Completion function doesn't work in this mode.\n");
+  }
+
+  // change hostname and port
+  svr.listen(sparams.hostname, sparams.port);
+}

ggml-cuda.cu

@@ -3,6 +3,7 @@
 #include <stdint.h>
 #include <stdio.h>
 #include <atomic>
+#include <assert.h>
 
 #include <cuda_runtime.h>
 #include <cublas_v2.h>
@@ -35,6 +36,7 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
 typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, float & v0, float & v1);
 typedef void (*to_fp32_cuda_t)(const void * x, float * y, int k, cudaStream_t stream);
 typedef void (*dequantize_mul_mat_vec_cuda_t)(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream);
+typedef void (*dot_kernel_k_t)(const void * vx, const int ib, const int iqs, const float * y, float & v);
 
 // QK = number of values after dequantization
 // QR = QK / number of values before dequantization
@@ -83,9 +85,64 @@ typedef struct {
 } block_q8_0;
 static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
 
+//================================= k-quants
+
+#define QK_K 256
+
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    half d;                  // super-block scale for quantized scales
+    half dmin;               // super-block scale for quantized mins
+} block_q2_k;
+static_assert(sizeof(block_q2_k) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_k block size/padding");
+
+typedef struct {
+    uint8_t hmask[QK_K/8];
+    uint8_t qs[QK_K/4]; // nibbles / quants
+    uint8_t scales[3*QK_K/64];
+    half d;
+} block_q3_k;
+static_assert(sizeof(block_q3_k) == sizeof(ggml_fp16_t) + QK_K / 4 + 11 * QK_K / 64, "wrong q3_k block size/padding");
+
+typedef struct {
+    half d;                    // super-block scale for quantized scales
+    half dmin;                 // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_k;
+static_assert(sizeof(block_q4_k) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_k block size/padding");
+
+typedef struct {
+    half    d;                   // super-block scale for quantized scales
+    half    dmin;                // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_k;
+static_assert(sizeof(block_q5_k) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2 + QK_K/8, "wrong q5_k block size/padding");
+
+typedef struct {
+    uint8_t ql[QK_K/2];   // quants, lower 4 bits
+    uint8_t qh[QK_K/4];   // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales
+    half    d;         // delta
+} block_q6_k;
+static_assert(sizeof(block_q6_k) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_k block size/padding");
+
+#define WARP_SIZE 32
+
 #define CUDA_MUL_BLOCK_SIZE 256
+
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
-#define CUDA_DMMV_BLOCK_SIZE 32 // dmmv = dequantize_mul_mat_vec
+
+// dmmv = dequantize_mul_mat_vec
+#ifndef GGML_CUDA_DMMV_X
+#define GGML_CUDA_DMMV_X 32
+#endif
+#ifndef GGML_CUDA_DMMV_Y
+#define GGML_CUDA_DMMV_Y 1
+#endif
 
 static __global__ void mul_f32(const float * x, const float * y, float * dst, const int kx, const int ky) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
@@ -174,6 +231,337 @@ static __device__ void dequantize_q8_0(const void * vx, const int ib, const int
     v1 = vi1*d;
 }
 
+//================================== k-quants
+
+static __global__ void dequantize_block_q2_k(const void * vx, float * yy) {
+
+    const int i   = blockIdx.x;
+    const int tid = threadIdx.x;
+    const int n   = tid/32;
+    const int l   = tid - 32*n;
+    const int is  = 8*n + l/16;
+
+    const block_q2_k * x = (const block_q2_k *) vx;
+
+    const uint8_t q = x[i].qs[32*n + l];
+    float * y = yy + i*QK_K + 128*n;
+
+    float dall = x[i].d;
+    float dmin = x[i].dmin;
+    y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
+    y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
+    y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
+    y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
+
+}
+
+static __device__ void vec_dot_q2_k(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
+
+    const block_q2_k * x = (const block_q2_k *) vx;
+
+    // if n is 0, we want to do the lower 128, else the upper 128,
+    // covering y[l+0],  y[l+32], y[l+64], y[l+96] and
+    //          y[l+16], y[l+48], y[l+80], y[l+112]
+    int n = iqs/128;                // 0 or 1
+    int r = iqs - 128*n;            // 0...120 in steps of 8
+    int l = r/8;                    // 0...15 in steps of 1
+
+    const float   * y = yy + 128*n + l;
+    const uint8_t * q = x[ib].qs + 32*n + l;
+    const uint8_t * s = x[ib].scales + 8*n;
+
+    const float dall = x[ib].d;
+    const float dmin = x[ib].dmin;
+
+    float sum = y[  0] * (dall * ((s[0] & 0xF) * ((q[ 0] >> 0) & 3)) - dmin * (s[0] >> 4))
+              + y[ 32] * (dall * ((s[2] & 0xF) * ((q[ 0] >> 2) & 3)) - dmin * (s[2] >> 4))
+              + y[ 64] * (dall * ((s[4] & 0xF) * ((q[ 0] >> 4) & 3)) - dmin * (s[4] >> 4))
+              + y[ 96] * (dall * ((s[6] & 0xF) * ((q[ 0] >> 6) & 3)) - dmin * (s[6] >> 4))
+              + y[ 16] * (dall * ((s[1] & 0xF) * ((q[16] >> 0) & 3)) - dmin * (s[1] >> 4))
+              + y[ 48] * (dall * ((s[3] & 0xF) * ((q[16] >> 2) & 3)) - dmin * (s[3] >> 4))
+              + y[ 80] * (dall * ((s[5] & 0xF) * ((q[16] >> 4) & 3)) - dmin * (s[5] >> 4))
+              + y[112] * (dall * ((s[7] & 0xF) * ((q[16] >> 6) & 3)) - dmin * (s[7] >> 4));
+
+    result = sum;
+
+}
+
+static __global__ void dequantize_block_q3_k(const void * vx, float * yy) {
+
+    int r = threadIdx.x/4;
+    int i = blockIdx.x;
+    int tid = r/2;
+    int is0 = r%2;
+    int l0 = 16*is0 + 4*(threadIdx.x%4);
+    int n = tid / 4;
+    int j = tid - 4*n;
+
+    const block_q3_k * x = (const block_q3_k *) vx;
+
+    uint8_t m = 1 << (4*n + j);
+    int is = 8*n + 2*j + is0;
+    int shift = 2*j;
+
+    int8_t us = is <  4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
+                is <  8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
+                is < 12 ? (x[i].scales[is-8] >>  4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
+                          (x[i].scales[is-8] >>  4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
+    float d_all = x[i].d;
+    float dl = d_all * (us - 32);
+
+    float * y = yy + i*QK_K + 128*n + 32*j;
+    const uint8_t * q = x[i].qs + 32*n;
+    const uint8_t * hm = x[i].hmask;
+
+    for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
+
+}
+
+static __device__ void vec_dot_q3_k(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
+
+    const block_q3_k * x = (const block_q3_k *) vx;
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    // if n is 0, we want to do the lower 128, else the upper 128,
+    // covering y[l+0],  y[l+32], y[l+64], y[l+96] and
+    //          y[l+16], y[l+48], y[l+80], y[l+112]
+    int n = iqs/128;                // 0 or 1
+    int r = iqs - 128*n;            // 0...120 in steps of 8
+    int l = r/8;                    // 0...15 in steps of 1
+
+    const float   * y = yy + 128*n + l;
+    const uint8_t * q = x[ib].qs + 32*n + l;
+    const uint8_t * hm = x[ib].hmask + l;
+    const int8_t  * s = (const int8_t *)utmp + 8*n;
+
+    memcpy(aux, x[ib].scales, 12);
+    utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+    utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+    utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+    utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+    const float dall = x[ib].d;
+
+    const uint8_t m = 1 << (4*n);
+
+    float sum = y[  0] * (s[0] - 32) * (((q[ 0] >> 0) & 3) - (hm[ 0] & (m << 0) ? 0 : 4))
+              + y[ 32] * (s[2] - 32) * (((q[ 0] >> 2) & 3) - (hm[ 0] & (m << 1) ? 0 : 4))
+              + y[ 64] * (s[4] - 32) * (((q[ 0] >> 4) & 3) - (hm[ 0] & (m << 2) ? 0 : 4))
+              + y[ 96] * (s[6] - 32) * (((q[ 0] >> 6) & 3) - (hm[ 0] & (m << 3) ? 0 : 4))
+              + y[ 16] * (s[1] - 32) * (((q[16] >> 0) & 3) - (hm[16] & (m << 0) ? 0 : 4))
+              + y[ 48] * (s[3] - 32) * (((q[16] >> 2) & 3) - (hm[16] & (m << 1) ? 0 : 4))
+              + y[ 80] * (s[5] - 32) * (((q[16] >> 4) & 3) - (hm[16] & (m << 2) ? 0 : 4))
+              + y[112] * (s[7] - 32) * (((q[16] >> 6) & 3) - (hm[16] & (m << 3) ? 0 : 4));
+
+    result = sum * dall;
+
+}
+
+static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
+    if (j < 4) {
+        d = q[j] & 63; m = q[j + 4] & 63;
+    } else {
+        d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
+        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
+    }
+}
+
+static __global__ void dequantize_block_q4_k(const void * vx, float * yy) {
+    const block_q4_k * x = (const block_q4_k *) vx;
+
+    const int i = blockIdx.x;
+
+    //// assume 64 threads - this is very slightly better than the one below
+    //const int tid = threadIdx.x;
+    //const int il  = tid/16;
+    //const int ir  = tid%16;
+    //const int is  = 2*il;
+    //const int n   = 2;
+
+    // assume 32 threads
+    const int tid = threadIdx.x;
+    const int il  = tid/8;
+    const int ir  = tid%8;
+    const int is  = 2*il;
+    const int n   = 4;
+
+    float * y = yy + i*QK_K + 64*il + n*ir;
+
+    const float dall = x[i].d;
+    const float dmin = x[i].dmin;
+
+    const uint8_t * q = x[i].qs + 32*il + n*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, sc, m);
+    const float d1 = dall * sc; const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, sc, m);
+    const float d2 = dall * sc; const float m2 = dmin * m;
+    for (int l = 0; l < n; ++l) {
+        y[l + 0] = d1 * (q[l] & 0xF) - m1;
+        y[l +32] = d2 * (q[l] >>  4) - m2;
+    }
+}
+
+static __device__ void vec_dot_q4_k(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
+
+    const block_q4_k * x = (const block_q4_k *) vx;
+
+                                    // iqs is in 0...248 in steps of 8 =>
+    const int j  = iqs / 64;        // j  is in 0...3
+    const int ir = (iqs - 64*j)/2;  // ir is in 0...28 in steps of 4
+    const int is = 2*j;             // is is in 0...6 in steps of 2
+
+    const float   * y = yy + 64*j + ir;
+    const uint8_t * q = x[ib].qs + 32*j + ir;
+
+    const float dall = x[ib].d;
+    const float dmin = x[ib].dmin;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[ib].scales, sc, m);
+    const float d1 = dall * sc;
+    const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[ib].scales, sc, m);
+    const float d2 = dall * sc;
+    const float m2 = dmin * m;
+
+    float sum = 0;
+    for (int k = 0; k < 4; ++k) {
+        sum += y[k +  0] * (d1 * (q[k] & 0xF) - m1);
+        sum += y[k + 32] * (d2 * (q[k] >>  4) - m2);
+    }
+    result = sum;
+
+}
+
+static __global__ void dequantize_block_q5_k(const void * vx, float * yy) {
+    const block_q5_k * x = (const block_q5_k *) vx;
+
+    const int i = blockIdx.x;
+
+    // assume 64 threads - this is very slightly better than the one below
+    const int tid = threadIdx.x;
+    const int il  = tid/16;   // il is in 0...3
+    const int ir  = tid%16;   // ir is in 0...15
+    const int is  = 2*il;     // is is in 0...6
+
+    float * y = yy + i*QK_K + 64*il + 2*ir;
+
+    const float dall = x[i].d;
+    const float dmin = x[i].dmin;
+
+    const uint8_t * ql = x[i].qs + 32*il + 2*ir;
+    const uint8_t * qh = x[i].qh + 2*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, sc, m);
+    const float d1 = dall * sc; const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, sc, m);
+    const float d2 = dall * sc; const float m2 = dmin * m;
+
+    uint8_t   hm  = 1 << (2*il);
+    y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1;
+    y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1;
+    hm <<= 1;
+    y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
+    y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
+}
+
+static __device__ void vec_dot_q5_k(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
+
+    const block_q5_k * x = (const block_q5_k *) vx;
+
+                                    // iqs is in 0...248 in steps of 8 =>
+    const int j  = iqs / 64;        // j  is in 0...3
+    const int ir = (iqs - 64*j)/2;  // ir is in 0...28 in steps of 4
+    const int is = 2*j;             // is is in 0...6 in steps of 2
+
+    const float   * y  = yy + 64*j + ir;
+    const uint8_t * ql = x[ib].qs + 32*j + ir;
+    const uint8_t * qh = x[ib].qh + ir;
+
+    const float dall = x[ib].d;
+    const float dmin = x[ib].dmin;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[ib].scales, sc, m);
+    const float d1 = dall * sc;
+    const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[ib].scales, sc, m);
+    const float d2 = dall * sc;
+    const float m2 = dmin * m;
+
+    uint8_t   hm  = 1 << is;
+    float sum = 0;
+    for (int k = 0; k < 4; ++k) {
+        sum += y[k +  0] * (d1 * ((ql[k] & 0xF) + (qh[k] & hm ? 16 : 0)) - m1);
+    }
+    hm <<= 1;
+    for (int k = 0; k < 4; ++k) {
+        sum += y[k + 32] * (d2 * ((ql[k] >>  4) + (qh[k] & hm ? 16 : 0)) - m2);
+    }
+    result = sum;
+
+}
+
+static __global__ void dequantize_block_q6_k(const void * vx, float * yy) {
+    const block_q6_k * x = (const block_q6_k *) vx;
+
+    const int i = blockIdx.x;
+
+    // assume 64 threads - this is very slightly better than the one below
+    const int tid = threadIdx.x;
+    const int ip  = tid/32;   // ip is 0 or 1
+    const int il  = tid - 32*ip; // 0...32
+    const int is  = 8*ip + il/16;
+
+    float * y = yy + i*QK_K + 128*ip + il;
+
+    const float d = x[i].d;
+
+    const uint8_t * ql = x[i].ql + 64*ip + il;
+    const uint8_t   qh = x[i].qh[32*ip + il];
+    const int8_t  * sc = x[i].scales + is;
+
+    y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
+    y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+    y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
+}
+
+static __device__ void vec_dot_q6_k(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
+
+    const block_q6_k * x = (const block_q6_k *) vx;
+
+    const int ip = iqs / 128;        // 0 or 1
+    const int il = (iqs - 128*ip)/8; // 0...15
+    const int is = 8*ip;
+
+    const float * y = yy + 128*ip + il;
+
+    const float d = x[ib].d;
+
+    const uint8_t * ql = x[ib].ql + 64*ip + il;
+    const uint8_t * qh = x[ib].qh + 32*ip + il;
+    const int8_t  * sc = x[ib].scales + is;
+
+    result = y[  0] * d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh[ 0] >> 0) & 3) << 4)) - 32)
+           + y[ 32] * d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh[ 0] >> 2) & 3) << 4)) - 32)
+           + y[ 64] * d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh[ 0] >> 4) & 3) << 4)) - 32)
+           + y[ 96] * d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh[ 0] >> 6) & 3) << 4)) - 32)
+           + y[ 16] * d * sc[1] * ((int8_t)((ql[16] & 0xF) | (((qh[16] >> 0) & 3) << 4)) - 32)
+           + y[ 48] * d * sc[3] * ((int8_t)((ql[48] & 0xF) | (((qh[16] >> 2) & 3) << 4)) - 32)
+           + y[ 80] * d * sc[5] * ((int8_t)((ql[16]  >> 4) | (((qh[16] >> 4) & 3) << 4)) - 32)
+           + y[112] * d * sc[7] * ((int8_t)((ql[48]  >> 4) | (((qh[16] >> 6) & 3) << 4)) - 32);
+
+}
+
 static __device__ void convert_f16(const void * vx, const int ib, const int iqs, float & v0, float & v1){
     const half * x = (const half *) vx;
 
@@ -200,41 +588,86 @@ static __global__ void dequantize_block(const void * vx, float * y, const int k)
     dequantize_kernel(vx, ib, iqs, v0, v1);
 }
 
-template <int block_size, int qk, int qr, dequantize_kernel_t dequantize_kernel>
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
 static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y, float * dst, const int ncols) {
-    const int row = blockIdx.x;
+    // qk = quantized weights per x block
+    // qr = number of quantized weights per data value in x block
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
     const int tid = threadIdx.x;
 
+    const int iter_stride = 2*GGML_CUDA_DMMV_X;
+    const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
     const int y_offset = qr == 1 ? 1 : qk/2;
 
-    __shared__ float tmp[block_size]; // separate sum for each thread
-    tmp[tid] = 0;
+    float tmp = 0; // partial sum for thread in warp
 
-    for (int i = 0; i < ncols/block_size; i += 2) {
-        const int col = i*block_size + 2*tid;
-        const int ib = (row*ncols + col)/qk; // block index
-        const int iqs = (col%qk)/qr; // quant index
+    for (int i = 0; i < ncols; i += iter_stride) {
+        const int col = i + vals_per_iter*tid;
+        const int ib = (row*ncols + col)/qk; // x block index
+        const int iqs = (col%qk)/qr; // x quant index
         const int iybs = col - col%qk; // y block start index
 
-        // dequantize
-        float v0, v1;
-        dequantize_kernel(vx, ib, iqs, v0, v1);
+// processing >2 values per i iter is faster for fast GPUs
+#pragma unroll
+        for (int j = 0; j < vals_per_iter; j += 2) {
+            // process 2 vals per j iter
 
-        // matrix multiplication
-        tmp[tid] += v0 * y[iybs + iqs + 0];
-        tmp[tid] += v1 * y[iybs + iqs + y_offset];
+            // dequantize
+            float v0, v1;
+            dequantize_kernel(vx, ib, iqs + j/qr, v0, v1);
+            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
+
+            // matrix multiplication
+            tmp += v0 * y[iybs + iqs + j/qr + 0];
+            tmp += v1 * y[iybs + iqs + j/qr + y_offset];
+            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
+        }
     }
 
     // sum up partial sums and write back result
     __syncthreads();
-    for (int s=block_size/2; s>0; s>>=1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        __syncthreads();
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
     }
+
     if (tid == 0) {
-        dst[row] = tmp[0];
+        dst[row] = tmp;
+    }
+}
+
+template <int n_thread, dot_kernel_k_t dot_kernel>
+static __global__ void dequantize_mul_mat_vec_k(const void * vx, const float * y, float * dst, const int ncols) {
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    const int tid = threadIdx.x;
+
+    const int iter_stride = QK_K;
+    const int vals_per_iter = iter_stride / n_thread;
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = 0; i < ncols; i += iter_stride) {
+        const int col = i + vals_per_iter*tid;
+        const int ib = ib0 + col/QK_K; // x block index
+        const int iqs = col%QK_K; // x quant index
+        const int iybs = col - col%QK_K; // y block start index
+
+        float v;
+        dot_kernel(vx, ib, iqs, y + iybs, v);
+        tmp += v;
+    }
+
+    // sum up partial sums and write back result
+    __syncthreads();
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
+    }
+
+    if (tid == 0) {
+        dst[row] = tmp;
     }
 }
 
@@ -268,34 +701,100 @@ static void dequantize_row_q8_0_cuda(const void * vx, float * y, const int k, cu
     dequantize_block<QK8_0, QR8_0, dequantize_q8_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
+static void dequantize_row_q2_k_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q2_k<<<nb, 64, 0, stream>>>(vx, y);
+}
+
+static void dequantize_row_q3_k_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q3_k<<<nb, 64, 0, stream>>>(vx, y);
+}
+
+static void dequantize_row_q4_k_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q4_k<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+static void dequantize_row_q5_k_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q5_k<<<nb, 64, 0, stream>>>(vx, y);
+}
+
+static void dequantize_row_q6_k_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q6_k<<<nb, 64, 0, stream>>>(vx, y);
+}
+
 static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK4_0, QR4_0, dequantize_q4_0>
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK4_1, QR4_1, dequantize_q4_1>
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK5_0, QR5_0, dequantize_q5_0>
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK5_1, QR5_1, dequantize_q5_1>
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK8_0, QR8_0, dequantize_q8_0>
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void dequantize_mul_mat_vec_q2_k_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2;
+    const dim3 block_dims(32, ny, 1);
+    dequantize_mul_mat_vec_k<32, vec_dot_q2_k><<<(nrows + ny - 1)/ny, block_dims, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void dequantize_mul_mat_vec_q3_k_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const dim3 block_dims(32, 2, 1);
+    dequantize_mul_mat_vec_k<32, vec_dot_q3_k><<<nrows/2, block_dims, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void dequantize_mul_mat_vec_q4_k_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const dim3 block_dims(32, 2, 1);
+    dequantize_mul_mat_vec_k<32, vec_dot_q4_k><<<nrows/2, block_dims, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void dequantize_mul_mat_vec_q5_k_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const dim3 block_dims(32, 2, 1);
+    dequantize_mul_mat_vec_k<32, vec_dot_q5_k><<<nrows/2, block_dims, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void dequantize_mul_mat_vec_q6_k_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const dim3 block_dims(32, 2, 1);
+    dequantize_mul_mat_vec_k<32, vec_dot_q6_k><<<nrows/2, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
@@ -304,9 +803,11 @@ static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, c
 }
 
 static void convert_mul_mat_vec_f16_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
-    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, 32, 1, convert_f16>
-        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
+    dequantize_mul_mat_vec<1, 1, convert_f16>
+        <<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
@@ -321,6 +822,16 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_q5_1_cuda;
         case GGML_TYPE_Q8_0:
             return dequantize_row_q8_0_cuda;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_k_cuda;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_k_cuda;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_k_cuda;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_k_cuda;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_k_cuda;
         case GGML_TYPE_F16:
             return convert_fp16_to_fp32_cuda;
         default:
@@ -340,6 +851,16 @@ static dequantize_mul_mat_vec_cuda_t ggml_get_dequantize_mul_mat_vec_cuda(ggml_t
             return dequantize_mul_mat_vec_q5_1_cuda;
         case GGML_TYPE_Q8_0:
             return dequantize_mul_mat_vec_q8_0_cuda;
+        case GGML_TYPE_Q2_K:
+            return dequantize_mul_mat_vec_q2_k_cuda;
+        case GGML_TYPE_Q3_K:
+            return dequantize_mul_mat_vec_q3_k_cuda;
+        case GGML_TYPE_Q4_K:
+            return dequantize_mul_mat_vec_q4_k_cuda;
+        case GGML_TYPE_Q5_K:
+            return dequantize_mul_mat_vec_q5_k_cuda;
+        case GGML_TYPE_Q6_K:
+            return dequantize_mul_mat_vec_q6_k_cuda;
         case GGML_TYPE_F16:
             return convert_mul_mat_vec_f16_cuda;
         default:
@@ -758,12 +1279,14 @@ static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor
                 CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0));
 
                 // compute
+                //printf("Calling dmmv\n");
                 dmmv(c_Q, c_Y, c_D, ne00, ne01, cudaStream);
                 CUDA_CHECK(cudaGetLastError());
 
             } else { // general dequantization kernel + cuBLAS matrix matrix multiplication
                 float * c_X = d_X + i * x_ne;
 
+//typedef void (*to_fp32_cuda_t)(const void * x, float * y, int k, cudaStream_t stream);
                 // convert src0 to fp32 on device
                 to_fp32_cuda(c_Q, c_X, x_ne, cudaStream2);
                 CUDA_CHECK(cudaGetLastError());

ggml-metal.h

@@ -0,0 +1,63 @@
+// An interface allowing to compute ggml_cgraph with Metal
+//
+// This is a fully functional interface that extends ggml with GPU support for Apple devices.
+// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
+//
+// How it works?
+//
+// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
+// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
+// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
+//
+// You only need to make sure that all memory buffers that you used during the graph creation
+// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
+// used during the graph evaluation to determine the arguments of the compute kernels.
+//
+// Synchronization between device and host memory (for example for input and output tensors)
+// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
+//
+
+#pragma once
+
+#include <stddef.h>
+#include <stdbool.h>
+
+// max memory buffers that can be mapped to the device
+#define GGML_METAL_MAX_BUFFERS 16
+
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_metal_context;
+
+struct ggml_metal_context * ggml_metal_init(void);
+void ggml_metal_free(struct ggml_metal_context * ctx);
+
+// creates a mapping between a host memory buffer and a device memory buffer
+// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
+// - the mapping is used during computation to determine the arguments of the compute kernels
+// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
+//
+bool ggml_metal_add_buffer(
+        struct ggml_metal_context * ctx,
+                       const char * name,
+                             void * data,
+                           size_t   size);
+
+// set data from host memory into the device
+void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+
+// get data from the device into host memory
+void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+
+// same as ggml_graph_compute but uses Metal
+void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+
+#ifdef __cplusplus
+}
+#endif
+

ggml-metal.m

@@ -0,0 +1,672 @@
+#import "ggml-metal.h"
+
+#import "ggml.h"
+
+#import <Foundation/Foundation.h>
+
+#import <Metal/Metal.h>
+#import <MetalPerformanceShaders/MetalPerformanceShaders.h>
+
+#ifdef GGML_METAL_NDEBUG
+#define metal_printf(...)
+#else
+#define metal_printf(...) fprintf(stderr, __VA_ARGS__)
+#endif
+
+#define UNUSED(x) (void)(x)
+
+struct ggml_metal_buffer {
+    const char * name;
+
+    void   * data;
+    size_t   size;
+
+    id<MTLBuffer> metal;
+};
+
+struct ggml_metal_context {
+    float * logits;
+
+    id<MTLDevice>       device;
+    id<MTLCommandQueue> queue;
+    id<MTLLibrary>      library;
+
+    int n_buffers;
+    struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
+
+    // custom kernels
+#define GGML_METAL_DECL_KERNEL(name) \
+    id<MTLFunction>             function_##name; \
+    id<MTLComputePipelineState> pipeline_##name
+
+    GGML_METAL_DECL_KERNEL(add);
+    GGML_METAL_DECL_KERNEL(mul);
+    GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
+    GGML_METAL_DECL_KERNEL(scale);
+    GGML_METAL_DECL_KERNEL(silu);
+    GGML_METAL_DECL_KERNEL(relu);
+    GGML_METAL_DECL_KERNEL(soft_max);
+    GGML_METAL_DECL_KERNEL(diag_mask_inf);
+    GGML_METAL_DECL_KERNEL(get_rows_q4_0);
+    GGML_METAL_DECL_KERNEL(rms_norm);
+    GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
+    GGML_METAL_DECL_KERNEL(rope);
+    GGML_METAL_DECL_KERNEL(cpy_f32_f16);
+    GGML_METAL_DECL_KERNEL(cpy_f32_f32);
+
+#undef GGML_METAL_DECL_KERNEL
+};
+
+// MSL code
+// TODO: move the contents here when ready
+//       for now it is easier to work in a separate file
+static NSString * const msl_library_source = @"see metal.metal";
+
+struct ggml_metal_context * ggml_metal_init(void) {
+    fprintf(stderr, "%s: allocating\n", __func__);
+
+    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+
+    ctx->device = MTLCreateSystemDefaultDevice();
+    ctx->queue  = [ctx->device newCommandQueue];
+
+    // determine if we can use MPS
+    if (MPSSupportsMTLDevice(ctx->device)) {
+        fprintf(stderr, "%s: using MPS\n", __func__);
+    } else {
+        fprintf(stderr, "%s: not using MPS\n", __func__);
+        GGML_ASSERT(false && "MPS not supported");
+    }
+
+#if 0
+    // compile from source string and show compile log
+    {
+        NSError * error = nil;
+
+        ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            exit(1);
+        }
+    }
+#else
+    UNUSED(msl_library_source);
+
+    // read the source from "ggml-metal.metal" into a string and use newLibraryWithSource
+    {
+        NSError * error = nil;
+
+        //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
+        NSString * path = [[NSBundle mainBundle] pathForResource:@"ggml-metal" ofType:@"metal"];
+        fprintf(stderr, "%s: loading '%s'\n", __func__, [path UTF8String]);
+
+        NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            exit(1);
+        }
+
+        ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            exit(1);
+        }
+    }
+#endif
+
+    // load kernels
+    {
+#define GGML_METAL_ADD_KERNEL(name) \
+        ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
+        ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:nil]; \
+        fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
+
+        GGML_METAL_ADD_KERNEL(add);
+        GGML_METAL_ADD_KERNEL(mul);
+        GGML_METAL_ADD_KERNEL(mul_row);
+        GGML_METAL_ADD_KERNEL(scale);
+        GGML_METAL_ADD_KERNEL(silu);
+        GGML_METAL_ADD_KERNEL(relu);
+        GGML_METAL_ADD_KERNEL(soft_max);
+        GGML_METAL_ADD_KERNEL(diag_mask_inf);
+        GGML_METAL_ADD_KERNEL(get_rows_q4_0);
+        GGML_METAL_ADD_KERNEL(rms_norm);
+        GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
+        GGML_METAL_ADD_KERNEL(rope);
+        GGML_METAL_ADD_KERNEL(cpy_f32_f16);
+        GGML_METAL_ADD_KERNEL(cpy_f32_f32);
+
+#undef GGML_METAL_ADD_KERNEL
+    }
+
+    return ctx;
+}
+
+void ggml_metal_free(struct ggml_metal_context * ctx) {
+    fprintf(stderr, "%s: deallocating\n", __func__);
+
+    free(ctx);
+}
+
+// finds the Metal buffer that contains the tensor data on the GPU device
+// the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
+// Metal buffer based on the host memory pointer
+//
+static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
+    //fprintf(stderr, "%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
+
+    for (int i = 0; i < ctx->n_buffers; ++i) {
+        const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
+
+        if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
+            *offs = (size_t) ioffs;
+
+            //fprintf(stderr, "%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
+
+            return ctx->buffers[i].metal;
+        }
+    }
+
+    fprintf(stderr, "%s: error: buffer is nil\n", __func__);
+
+    return nil;
+}
+
+bool ggml_metal_add_buffer(
+        struct ggml_metal_context * ctx,
+                     const char * name,
+                           void * data,
+                         size_t   size) {
+    if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
+        fprintf(stderr, "%s: too many buffers\n", __func__);
+        return false;
+    }
+
+    if (data) {
+        // verify that the buffer does not overlap with any of the existing buffers
+        for (int i = 0; i < ctx->n_buffers; ++i) {
+            const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
+
+            if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
+                fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
+                return false;
+            }
+        }
+
+        ctx->buffers[ctx->n_buffers].name = name;
+        ctx->buffers[ctx->n_buffers].data = data;
+        ctx->buffers[ctx->n_buffers].size = size;
+        ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytes:data length:size options:MTLResourceStorageModeShared];
+
+        ++ctx->n_buffers;
+
+        fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB\n", __func__, name, size / 1024.0 / 1024.0);
+    }
+
+    return true;
+}
+
+void ggml_metal_set_tensor(
+        struct ggml_metal_context * ctx,
+        struct ggml_tensor * t) {
+    metal_printf("%s: set input for tensor '%s'\n", __func__, t->name);
+
+    size_t offs;
+    id<MTLBuffer> id_dst = ggml_metal_get_buffer(ctx, t, &offs);
+
+    memcpy((void *) ((uint8_t *) id_dst.contents + offs), t->data, ggml_nbytes(t));
+}
+
+void ggml_metal_get_tensor(
+        struct ggml_metal_context * ctx,
+        struct ggml_tensor * t) {
+    metal_printf("%s: extract results for tensor '%s'\n", __func__, t->name);
+
+    size_t offs;
+    id<MTLBuffer> id_src = ggml_metal_get_buffer(ctx, t, &offs);
+
+    memcpy(t->data, (void *) ((uint8_t *) id_src.contents + offs), ggml_nbytes(t));
+}
+
+void ggml_metal_graph_compute(
+        struct ggml_metal_context * ctx,
+             struct ggml_cgraph * gf) {
+    metal_printf("%s: evaluating graph\n", __func__);
+
+    size_t offs_src0 = 0;
+    size_t offs_src1 = 0;
+    size_t offs_dst  = 0;
+
+    id<MTLCommandBuffer> command_buffer  = [ctx->queue commandBuffer];
+    id<MTLComputeCommandEncoder> encoder = nil;
+
+    for (int i = 0; i < gf->n_nodes; ++i) {
+        //metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+
+        struct ggml_tensor * src0 = gf->nodes[i]->src0;
+        struct ggml_tensor * src1 = gf->nodes[i]->src1;
+        struct ggml_tensor * dst  = gf->nodes[i];
+
+        const int64_t  ne00 = src0 ? src0->ne[0] : 0;
+        const int64_t  ne01 = src0 ? src0->ne[1] : 0;
+        const int64_t  ne02 = src0 ? src0->ne[2] : 0;
+        const int64_t  ne03 = src0 ? src0->ne[3] : 0;
+
+        const uint64_t nb00 = src0 ? src0->nb[0] : 0;
+        const uint64_t nb01 = src0 ? src0->nb[1] : 0;
+        const uint64_t nb02 = src0 ? src0->nb[2] : 0;
+        const uint64_t nb03 = src0 ? src0->nb[3] : 0;
+
+        const int64_t  ne10 = src1 ? src1->ne[0] : 0;
+        const int64_t  ne11 = src1 ? src1->ne[1] : 0;
+        const int64_t  ne12 = src1 ? src1->ne[2] : 0;
+        const int64_t  ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
+
+        const uint64_t nb10 = src1 ? src1->nb[0] : 0;
+        const uint64_t nb11 = src1 ? src1->nb[1] : 0;
+        const uint64_t nb12 = src1 ? src1->nb[2] : 0;
+        const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
+
+        const int64_t  ne0  = dst ? dst->ne[0] : 0;
+        const int64_t  ne1  = dst ? dst->ne[1] : 0;
+        const int64_t  ne2  = dst ? dst->ne[2] : 0;
+        const int64_t  ne3  = dst ? dst->ne[3] : 0;
+
+        const uint64_t nb0  = dst ? dst->nb[0] : 0;
+        const uint64_t nb1  = dst ? dst->nb[1] : 0;
+        const uint64_t nb2  = dst ? dst->nb[2] : 0;
+        const uint64_t nb3  = dst ? dst->nb[3] : 0;
+
+        const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
+        const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
+        const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
+
+        id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
+        id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
+        id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
+
+        //metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
+        //if (src0) {
+        //    metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
+        //            ggml_is_contiguous(src0), src0->name);
+        //}
+        //if (src1) {
+        //    metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
+        //            ggml_is_contiguous(src1), src1->name);
+        //}
+        //if (dst) {
+        //    metal_printf("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
+        //            dst->name);
+        //}
+
+        switch (dst->op) {
+            case GGML_OP_RESHAPE:
+            case GGML_OP_VIEW:
+            case GGML_OP_TRANSPOSE:
+            case GGML_OP_PERMUTE:
+                {
+                    // noop
+                } break;
+            case GGML_OP_ADD:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    [encoder setComputePipelineState:ctx->pipeline_add];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+            case GGML_OP_MUL:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    if (ggml_nelements(src1) == ne10) {
+                        // src1 is a row
+                        [encoder setComputePipelineState:ctx->pipeline_mul_row];
+                    } else {
+                        [encoder setComputePipelineState:ctx->pipeline_mul];
+                    }
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+            case GGML_OP_SCALE:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    const float scale = *(const float *) src1->data;
+
+                    [encoder setComputePipelineState:ctx->pipeline_scale];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                    [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+            case GGML_OP_SILU:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    [encoder setComputePipelineState:ctx->pipeline_silu];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+            case GGML_OP_RELU:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    [encoder setComputePipelineState:ctx->pipeline_relu];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+            case GGML_OP_SOFT_MAX:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    const int nth = 32;
+
+                    [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                    [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                    [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                    [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                } break;
+            case GGML_OP_DIAG_MASK_INF:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    const int n_past = ((int32_t *)(src1->data))[0];
+
+                    [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                    [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
+                    [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
+                    [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+            case GGML_OP_MUL_MAT:
+                {
+                    // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
+
+                    GGML_ASSERT(ne00 == ne10);
+                    GGML_ASSERT(ne02 == ne12);
+
+                    if (ggml_is_contiguous(src0) &&
+                        ggml_is_contiguous(src1) &&
+                        (src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
+
+                        if (encoder != nil) {
+                            [encoder endEncoding];
+                            encoder = nil;
+                        }
+
+                        MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
+                        MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
+
+                        // for F32 x F32 we use MPS
+                        MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
+                            matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
+
+                        MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
+                            matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
+
+                        MPSMatrixDescriptor * desc  = [MPSMatrixDescriptor
+                            matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
+
+                        MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
+                            initWithDevice:ctx->device transposeLeft:false transposeRight:true
+                                resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
+
+                        // we need to do ne02 multiplications
+                        // TODO: is there a way to do this in parallel - currently very slow ..
+                        // TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
+                        for (int64_t i02 = 0; i02 < ne02; ++i02) {
+                            size_t offs_src0_cur = offs_src0 + i02*nb02;
+                            size_t offs_src1_cur = offs_src1 + i02*nb12;
+                            size_t offs_dst_cur  = offs_dst  + i02*nb2;
+
+                            MPSMatrix * mat_src0 = [[MPSMatrix alloc] initWithBuffer:id_src0 offset:offs_src0_cur descriptor:desc0];
+                            MPSMatrix * mat_src1 = [[MPSMatrix alloc] initWithBuffer:id_src1 offset:offs_src1_cur descriptor:desc1];
+                            MPSMatrix * mat_dst  = [[MPSMatrix alloc] initWithBuffer:id_dst  offset:offs_dst_cur  descriptor:desc ];
+
+                            [mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
+                        }
+                    } else {
+                        if (encoder == nil) {
+                            encoder = [command_buffer computeCommandEncoder];
+                        }
+
+                        int nth0 = 32;
+                        int nth1 = 1;
+
+                        // use custom matrix x vector kernel
+                        switch (src0t) {
+                            case GGML_TYPE_Q4_0:
+                                {
+                                    GGML_ASSERT(ne02 == 1);
+                                    GGML_ASSERT(ne12 == 1);
+
+                                    nth0 = 8;
+                                    nth1 = 4;
+                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
+                                } break;
+                            case GGML_TYPE_F16:
+                                {
+                                    GGML_ASSERT(ne02 == ne12);
+
+                                    nth0 = 32;
+                                    nth1 = 1;
+                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                } break;
+                            default: GGML_ASSERT(false && "not implemented");
+                        };
+
+
+                        [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                        [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                        [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                        [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                        [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                        [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
+                        [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
+                        [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
+                        [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
+                        [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
+                        [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
+                        [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
+                        [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
+                        [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
+                        [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
+
+                        if (src0t == GGML_TYPE_Q4_0) {
+                            [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                        } else {
+                            [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                        }
+                    }
+                } break;
+            case GGML_OP_GET_ROWS:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    switch (src0->type) {
+                        case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
+                        default: GGML_ASSERT(false && "not implemented");
+                    }
+
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                    [encoder setBytes:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
+                    [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
+                    [encoder setBytes:&(dst->nb[1])  length:sizeof(uint64_t) atIndex:5];
+
+                    const int64_t n = ggml_nelements(src1);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+            case GGML_OP_RMS_NORM:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    const float eps = 1e-6f;
+
+                    const int nth = 256;
+
+                    [encoder setComputePipelineState:ctx->pipeline_rms_norm];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                    [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
+                    [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
+                    [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
+                    [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+
+                    const int64_t nrows = ggml_nrows(src0);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                } break;
+            case GGML_OP_ROPE:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    const int n_dims = ((int32_t *) src1->data)[1];
+                    const int mode   = ((int32_t *) src1->data)[2];
+
+                    const int n_past = ((int32_t *)(src1->data))[0];
+
+                    [encoder setComputePipelineState:ctx->pipeline_rope];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                    [encoder setBytes:&ne00   length:sizeof( int64_t) atIndex:2];
+                    [encoder setBytes:&ne01   length:sizeof( int64_t) atIndex:3];
+                    [encoder setBytes:&ne02   length:sizeof( int64_t) atIndex:4];
+                    [encoder setBytes:&ne03   length:sizeof( int64_t) atIndex:5];
+                    [encoder setBytes:&nb00   length:sizeof(uint64_t) atIndex:6];
+                    [encoder setBytes:&nb01   length:sizeof(uint64_t) atIndex:7];
+                    [encoder setBytes:&nb02   length:sizeof(uint64_t) atIndex:8];
+                    [encoder setBytes:&nb03   length:sizeof(uint64_t) atIndex:9];
+                    [encoder setBytes:&ne0    length:sizeof( int64_t) atIndex:10];
+                    [encoder setBytes:&ne1    length:sizeof( int64_t) atIndex:11];
+                    [encoder setBytes:&ne2    length:sizeof( int64_t) atIndex:12];
+                    [encoder setBytes:&ne3    length:sizeof( int64_t) atIndex:13];
+                    [encoder setBytes:&nb0    length:sizeof(uint64_t) atIndex:14];
+                    [encoder setBytes:&nb1    length:sizeof(uint64_t) atIndex:15];
+                    [encoder setBytes:&nb2    length:sizeof(uint64_t) atIndex:16];
+                    [encoder setBytes:&nb3    length:sizeof(uint64_t) atIndex:17];
+                    [encoder setBytes:&n_past length:sizeof(     int) atIndex:18];
+                    [encoder setBytes:&n_dims length:sizeof(     int) atIndex:19];
+                    [encoder setBytes:&mode   length:sizeof(     int) atIndex:20];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+            case GGML_OP_CPY:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    const int nth = 32;
+
+                    switch (src0t) {
+                        case GGML_TYPE_F32:
+                            {
+                                switch (dstt) {
+                                    case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16]; break;
+                                    case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32]; break;
+                                    default: GGML_ASSERT(false && "not implemented");
+                                };
+                            } break;
+                        default: GGML_ASSERT(false && "not implemented");
+                    }
+
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                    [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
+                    [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
+                    [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
+                    [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
+                    [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
+                    [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
+                    [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
+                    [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
+                    [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
+                    [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
+                    [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
+                    [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
+                    [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
+                    [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
+                    [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
+                    [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                } break;
+            default:
+                fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                GGML_ASSERT(false);
+        }
+    }
+
+    if (encoder != nil) {
+        [encoder endEncoding];
+        encoder = nil;
+    }
+
+    [command_buffer commit];
+    [command_buffer waitUntilCompleted];
+
+    {
+        const double time_elapsed = [command_buffer GPUEndTime] - [command_buffer GPUStartTime];
+        UNUSED(time_elapsed);
+
+        metal_printf("%s: time elapsed = %f ms\n", __func__, time_elapsed * 1000.0);
+    }
+}

ggml-metal.metal

@@ -0,0 +1,489 @@
+#include <metal_stdlib>
+
+using namespace metal;
+
+#define MAX(x, y) ((x) > (y) ? (x) : (y))
+
+#define QK4_0 32
+#define QR4_0 2
+typedef struct {
+    half    d;             // delta
+    uint8_t qs[QK4_0 / 2]; // nibbles / quants
+} block_q4_0;
+
+static void dequantize_row_q4_0(device const block_q4_0 * x, device float * y, int k) {
+    const int qk = QK4_0;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        const half d = x[i].d;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int x0 = (x[i].qs[j] & 0x0F) - 8;
+            const int x1 = (x[i].qs[j] >>   4) - 8;
+
+            y[i*qk + j + 0   ] = x0*d;
+            y[i*qk + j + qk/2] = x1*d;
+        }
+    }
+}
+
+kernel void kernel_add(
+        device const float * src0,
+        device const float * src1,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] + src1[tpig];
+}
+
+kernel void kernel_mul(
+        device const float * src0,
+        device const float * src1,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * src1[tpig];
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_mul_row(
+        device const float * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * src1[tpig % ne00];
+}
+
+kernel void kernel_scale(
+        device const float * src0,
+        device       float * dst,
+        constant     float & scale,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * scale;
+}
+
+kernel void kernel_silu(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    float x = src0[tpig];
+    dst[tpig] = x / (1.0f + exp(-x));
+}
+
+kernel void kernel_relu(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = max(0.0f, src0[tpig]);
+}
+
+kernel void kernel_soft_max(
+        device const float * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    device const float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    device       float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    // parallel max
+    buf[tpitg[0]] = -INFINITY;
+    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+        buf[tpitg[0]] = MAX(buf[tpitg[0]], psrc0[i00]);
+    }
+
+    // reduce
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    for (uint i = ntg[0]/2; i > 0; i /= 2) {
+        if (tpitg[0] < i) {
+            buf[tpitg[0]] = MAX(buf[tpitg[0]], buf[tpitg[0] + i]);
+        }
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    // broadcast
+    if (tpitg[0] == 0) {
+        buf[0] = buf[0];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    const float max = buf[0];
+
+    // parallel sum
+    buf[tpitg[0]] = 0.0f;
+    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+        buf[tpitg[0]] += exp(psrc0[i00] - max);
+    }
+
+    // reduce
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    for (uint i = ntg[0]/2; i > 0; i /= 2) {
+        if (tpitg[0] < i) {
+            buf[tpitg[0]] += buf[tpitg[0] + i];
+        }
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    // broadcast
+    if (tpitg[0] == 0) {
+        buf[0] = buf[0];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    const float sum = buf[0];
+
+    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+        pdst[i00] = exp(psrc0[i00] - max) / sum;
+    }
+}
+
+kernel void kernel_diag_mask_inf(
+        device const float * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant       int & n_past,
+        uint3 tpig[[thread_position_in_grid]]) {
+    const int64_t i02 = tpig[2];
+    const int64_t i01 = tpig[1];
+    const int64_t i00 = tpig[0];
+
+    if (i00 > n_past + i01) {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
+    } else {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
+    }
+}
+
+kernel void kernel_get_rows_q4_0(
+        device const  void * src0,
+        device const   int * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb1,
+        uint tpig[[thread_position_in_grid]]) {
+    const int i = tpig;
+    const int r = ((device int32_t *) src1)[i];
+
+    dequantize_row_q4_0(
+            (device const block_q4_0 *) ((device char *) src0 + r*nb01),
+                       (device float *) ((device char *)  dst + i*nb1), ne00);
+}
+
+kernel void kernel_rms_norm(
+        device const  void * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant     float & eps,
+        threadgroup float  * sum [[threadgroup(0)]],
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * x = (device const float *) ((device const char *) src0 + tgpig*nb01);
+
+    // parallel sum
+    sum[tpitg] = 0.0f;
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        sum[tpitg] += x[i00] * x[i00];
+    }
+
+    // reduce
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    for (uint i = ntg/2; i > 0; i /= 2) {
+        if (tpitg < i) {
+            sum[tpitg] += sum[tpitg + i];
+        }
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    // broadcast
+    if (tpitg == 0) {
+        sum[0] /= ne00;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    const float mean  = sum[0];
+    const float scale = 1.0f/sqrt(mean + eps);
+
+    device float * y = dst + tgpig*ne00;
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        y[i00] = x[i00] * scale;
+    }
+}
+
+kernel void kernel_mul_mat_q4_0_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        threadgroup float  * sum [[threadgroup(0)]],
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2  tpig[[thread_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2  tptg[[threads_per_threadgroup]]) {
+    const int nb = ne00/QK4_0;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+
+    device const block_q4_0 * x = (device const block_q4_0 *) src0 + r0*nb;
+    device const float      * y = (device const float      *) src1 + r1*ne10;
+
+    const uint nth = tptg.x*tptg.y;
+    const uint ith = tptg.y*tpitg.x + tpitg.y;
+
+    sum[ith] = 0.0f;
+
+    for (int i = tpitg.x; i < nb; i += tptg.x) {
+        device const uchar4 * x0p = (device const uchar4 *) (x + i)->qs;
+        device const float4 * y0p = (device const float4 *) (y + i*QK4_0);
+
+        const float d = (float)((x + i)->d);
+
+        const uchar4 x0v = *(x0p + tpitg.y);
+        const float4 y0v = *(y0p + tpitg.y + 0);
+        const float4 y1v = *(y0p + tpitg.y + 4);
+
+        float acc = 0.0f;
+
+        for (int j = 0; j < 4; ++j) {
+            const int x0 = x0v[j] & 0x0F;
+            const int x1 = x0v[j] >>   4;
+
+            const float y0 = y0v[j];
+            const float y1 = y1v[j];
+
+            acc += (x0 - 8)*y0 + (x1 - 8)*y1;
+        }
+
+        sum[ith] += acc*d;
+    }
+
+    // accumulate the sum from all threads in the threadgroup
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    for (uint i = nth/2; i > 0; i /= 2) {
+        if (ith < i) {
+            sum[ith] += sum[ith + i];
+        }
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    if (ith == 0) {
+        dst[r1*ne0 + r0] = sum[0];
+    }
+}
+
+kernel void kernel_mul_mat_f16_f32(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        threadgroup float  * sum [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tpig[[thread_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+    const int64_t im = tgpig.z;
+
+    device const half  * x = (device const half  *) (src0 + r0*nb01 + im*nb02);
+    device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
+
+    sum[tpitg.x] = 0.0f;
+
+    for (int i = tpitg.x; i < ne00; i += tptg.x) {
+        sum[tpitg.x] += (float) x[i] * (float) y[i];
+    }
+
+    // accumulate the sum from all threads in the threadgroup
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    for (uint i = tptg.x/2; i > 0; i /= 2) {
+        if (tpitg.x < i) {
+            sum[tpitg.x] += sum[tpitg.x + i];
+        }
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    if (tpitg.x == 0) {
+        dst[im*ne1*ne0 + r1*ne0 + r0] = sum[0];
+    }
+}
+
+kernel void kernel_rope(
+        device const  void * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        constant       int & n_past,
+        constant       int & n_dims,
+        constant       int & mode,
+        uint3 tpig[[thread_position_in_grid]]) {
+    const int64_t i3 = tpig[2];
+    const int64_t i2 = tpig[1];
+    const int64_t i1 = tpig[0];
+
+    const bool is_neox = mode & 2;
+    const float theta_scale = pow(10000.0, -2.0f/n_dims);
+
+    const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2);
+
+    float theta = (float)p;
+
+    if (!is_neox) {
+        for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+            const float cos_theta = cos(theta);
+            const float sin_theta = sin(theta);
+
+            theta *= theta_scale;
+
+            device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            device       float * dst_data  = (device float *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[1];
+
+            dst_data[0] = x0*cos_theta - x1*sin_theta;
+            dst_data[1] = x0*sin_theta + x1*cos_theta;
+        }
+    } else {
+        // TODO: implement
+    }
+}
+
+kernel void kernel_cpy_f32_f16(
+        device const float * src0,
+        device        half * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    device half * dst_data = (device half *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f32_f32(
+        device const float * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}

ggml-opencl.c

@@ -1,474 +0,0 @@
-#include "ggml-opencl.h"
-
-#define CL_TARGET_OPENCL_VERSION 110
-#include <clblast_c.h>
-
-#include <stdlib.h>
-#include <stdio.h>
-#include <string.h>
-
-#include "ggml.h"
-
-#define MULTILINE_QUOTE(...) #__VA_ARGS__
-static const char * program_source = MULTILINE_QUOTE(
-
-typedef char int8_t;
-typedef uchar uint8_t;
-typedef int int32_t;
-typedef uint uint32_t;
-
-struct __attribute__ ((packed)) block_q4_0
-{
-    half d;
-    uint8_t qs[16]; /* QK4_0 / 2 */
-};
-
-struct __attribute__ ((packed)) block_q4_1
-{
-    half d;
-    half m;
-    uint8_t qs[16]; /* QK4_1 / 2 */
-};
-
-struct __attribute__ ((packed)) block_q5_0
-{
-    half d;
-    uint32_t qh;
-    uint8_t qs[16]; /* QK5_0 / 2 */
-};
-
-struct __attribute__ ((packed)) block_q5_1
-{
-    half d;
-    half m;
-    uint32_t qh;
-    uint8_t qs[16]; /* QK5_1 / 2 */
-};
-
-struct __attribute__ ((packed)) block_q8_0
-{
-    half d;
-    int8_t qs[32]; /* QK8_0 */
-};
-
-
-__kernel void dequantize_row_q4_0(__global struct block_q4_0* x, __global float* y) {
-    const uint i = get_global_id(0) / 32; /* QK4_0 */
-    const uint j = get_local_id(0);
-
-    const float d = vload_half(0, (__global half*) &x[i].d);
-
-    const int x0 = (x[i].qs[j] & 0xf) - 8;
-    const int x1 = (x[i].qs[j] >>  4) - 8;
-
-    y[i*32 + j + 0 ] = x0*d;
-    y[i*32 + j + 16] = x1*d;
-}
-
-__kernel void dequantize_row_q4_1(__global struct block_q4_1* x, __global float* y) {
-    const uint i = get_global_id(0) / 32; /* QK4_1 */
-    const uint j = get_local_id(0);
-
-    const float d = vload_half(0, (__global half*) &x[i].d);
-    const float m = vload_half(0, (__global half*) &x[i].m);
-
-    const int x0 = (x[i].qs[j] & 0xf);
-    const int x1 = (x[i].qs[j] >>  4);
-
-    y[i*32 + j + 0 ] = x0*d + m;
-    y[i*32 + j + 16] = x1*d + m;
-}
-
-__kernel void dequantize_row_q5_0(__global struct block_q5_0* x, __global float* y) {
-    const uint i = get_global_id(0) / 32; /* QK5_0 */
-    const uint j = get_local_id(0);
-
-    const float d = vload_half(0, (__global half*) &x[i].d);
-
-    uint32_t qh = x[i].qh;
-
-    const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-    const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-    const int32_t x0 = ((x[i].qs[j] & 0xf) | xh_0) - 16;
-    const int32_t x1 = ((x[i].qs[j] >>  4) | xh_1) - 16;
-
-    y[i*32 + j + 0 ] = x0*d;
-    y[i*32 + j + 16] = x1*d;
-}
-
-__kernel void dequantize_row_q5_1(__global struct block_q5_1* x, __global float* y) {
-    const uint i = get_global_id(0) / 32; /* QK5_1 */
-    const uint j = get_local_id(0);
-
-    const float d = vload_half(0, (__global half*) &x[i].d);
-    const float m = vload_half(0, (__global half*) &x[i].m);
-
-    uint32_t qh = x[i].qh;
-
-    const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-    const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-    const int x0 = (x[i].qs[j] & 0xf) | xh_0;
-    const int x1 = (x[i].qs[j] >>  4) | xh_1;
-
-    y[i*32 + j + 0 ] = x0*d + m;
-    y[i*32 + j + 16] = x1*d + m;
-}
-
-__kernel void dequantize_row_q8_0(__global struct block_q8_0* x, __global float* y) {
-    const uint i = get_global_id(0) / 32; /* QK8_0 */
-    const uint j = get_local_id(0);
-
-    const float d = vload_half(0, (__global half*) &x[i].d);
-    y[i*32 + j] = x[i].qs[j]*d;
-}
-
-);
-
-#define CL_CHECK(err)                                               \
-    do {                                                            \
-        cl_int err_ = (err);                                        \
-        if (err_ != CL_SUCCESS) {                                   \
-            fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n",  \
-                #err, err_, __FILE__, __LINE__);                    \
-            exit(1);                                                \
-        }                                                           \
-    } while (0)
-
-#define CLBLAST_CHECK(err)                                          \
-    do {                                                            \
-        CLBlastStatusCode err_ = (err);                             \
-        if (err_ != CLBlastSuccess) {                               \
-            fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n",  \
-                #err, err_, __FILE__, __LINE__);                    \
-            exit(1);                                                \
-        }                                                           \
-    } while (0)
-
-static cl_platform_id platform;
-static cl_device_id device;
-static cl_context context;
-static cl_command_queue queue;
-static cl_program program;
-static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q5_0, kernel_q5_1, kernel_q8_0;
-static cl_mem cl_buffer_a, cl_buffer_qb, cl_buffer_b, cl_buffer_c;
-static size_t cl_size_a = 0, cl_size_qb = 0, cl_size_b = 0, cl_size_c = 0;
-
-static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
-    cl_program p;
-    char *program_log;
-    size_t program_size, log_size;
-    int err;
-
-    program_size = strlen(program_buffer);
-
-    p = clCreateProgramWithSource(ctx, 1, (const char**)&program_buffer, &program_size, &err);
-    if(err < 0) {
-        fprintf(stderr, "OpenCL error creating program");
-        exit(1);
-    }
-
-    err = clBuildProgram(p, 0, NULL, NULL, NULL, NULL);
-    if(err < 0) {
-
-        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
-        program_log = (char*) malloc(log_size + 1);
-        program_log[log_size] = '\0';
-        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, log_size + 1, program_log, NULL);
-        printf("%s\n", program_log);
-        free(program_log);
-        exit(1);
-    }
-
-    return p;
-}
-
-void ggml_cl_init(void) {
-    cl_int err = 0;
-
-    struct cl_device;
-    struct cl_platform {
-        cl_platform_id id;
-        unsigned number;
-        char name[128];
-        char vendor[128];
-        struct cl_device * devices;
-        unsigned n_devices;
-        struct cl_device * default_device;
-    };
-
-    struct cl_device {
-        struct cl_platform * platform;
-        cl_device_id id;
-        unsigned number;
-        cl_device_type type;
-        char name[128];
-    };
-
-    enum { NPLAT = 16, NDEV = 16 };
-
-    struct cl_platform platforms[NPLAT];
-    unsigned n_platforms = 0;
-    struct cl_device devices[NDEV];
-    unsigned n_devices = 0;
-    struct cl_device * default_device = NULL;
-
-    platform = NULL;
-    device = NULL;
-
-    cl_platform_id platform_ids[NPLAT];
-    CL_CHECK(clGetPlatformIDs(NPLAT, platform_ids, &n_platforms));
-
-    for (unsigned i = 0; i < n_platforms; i++) {
-        struct cl_platform * p = &platforms[i];
-        p->number = i;
-        p->id = platform_ids[i];
-        CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_NAME, sizeof(p->name), &p->name, NULL));
-        CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_VENDOR, sizeof(p->vendor), &p->vendor, NULL));
-
-        cl_device_id device_ids[NDEV];
-        cl_int clGetDeviceIDsError = clGetDeviceIDs(p->id, CL_DEVICE_TYPE_ALL, NDEV, device_ids, &p->n_devices);
-        if (clGetDeviceIDsError == CL_DEVICE_NOT_FOUND) {
-            p->n_devices = 0;
-        } else {
-            CL_CHECK(clGetDeviceIDsError);
-        }
-        p->devices = p->n_devices > 0 ? &devices[n_devices] : NULL;
-        p->default_device = NULL;
-
-        for (unsigned j = 0; j < p->n_devices; j++) {
-            struct cl_device * d = &devices[n_devices];
-            d->number = n_devices++;
-            d->id = device_ids[j];
-            d->platform = p;
-            CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_NAME, sizeof(d->name), &d->name, NULL));
-            CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_TYPE, sizeof(d->type), &d->type, NULL));
-
-            if (p->default_device == NULL && d->type == CL_DEVICE_TYPE_GPU) {
-                p->default_device = d;
-            }
-        }
-
-        if (default_device == NULL && p->default_device != NULL) {
-            default_device = p->default_device;
-        }
-    }
-
-    if (n_devices == 0) {
-        fprintf(stderr, "ggml_opencl: could find any OpenCL devices.\n");
-        exit(1);
-    }
-
-    char * user_platform_string = getenv("GGML_OPENCL_PLATFORM");
-    char * user_device_string = getenv("GGML_OPENCL_DEVICE");
-    int user_platform_number = -1;
-    int user_device_number = -1;
-
-    unsigned n;
-    if (user_platform_string != NULL && sscanf(user_platform_string, " %u", &n) == 1 && n < n_platforms) {
-        user_platform_number = (int)n;
-    }
-    if (user_device_string != NULL && sscanf(user_device_string, " %u", &n) == 1 && n < n_devices) {
-        user_device_number = (int)n;
-    }
-
-    struct cl_device * selected_devices = devices;
-    unsigned n_selected_devices = n_devices;
-
-    if (user_platform_number == -1 && user_platform_string != NULL && user_platform_string[0] != 0) {
-        for (unsigned i = 0; i < n_platforms; i++) {
-            struct cl_platform * p = &platforms[i];
-            if (strstr(p->name, user_platform_string) != NULL ||
-                strstr(p->vendor, user_platform_string) != NULL) {
-                user_platform_number = (int)i;
-                break;
-            }
-        }
-        if (user_platform_number == -1) {
-            fprintf(stderr, "ggml_opencl: no platform matching '%s' was found.\n", user_platform_string);
-            exit(1);
-        }
-    }
-    if (user_platform_number != -1) {
-        struct cl_platform * p = &platforms[user_platform_number];
-        selected_devices = p->devices;
-        n_selected_devices = p->n_devices;
-        default_device = p->default_device;
-        if (n_selected_devices == 0) {
-            fprintf(stderr, "ggml_opencl: selected platform '%s' does not have any devices.\n", p->name);
-            exit(1);
-        }
-    }
-
-    if (user_device_number == -1 && user_device_string != NULL && user_device_string[0] != 0) {
-        for (unsigned i = 0; i < n_selected_devices; i++) {
-            struct cl_device * d = &selected_devices[i];
-            if (strstr(d->name, user_device_string) != NULL) {
-                user_device_number = d->number;
-                break;
-            }
-        }
-        if (user_device_number == -1) {
-            fprintf(stderr, "ggml_opencl: no device matching '%s' was found.\n", user_device_string);
-            exit(1);
-        }
-    }
-    if (user_device_number != -1) {
-        selected_devices = &devices[user_device_number];
-        n_selected_devices = 1;
-        default_device = &selected_devices[0];
-    }
-
-    GGML_ASSERT(n_selected_devices > 0);
-
-    if (default_device == NULL) {
-        default_device = &selected_devices[0];
-    }
-
-    fprintf(stderr, "ggml_opencl: selecting platform: '%s'\n", default_device->platform->name);
-    fprintf(stderr, "ggml_opencl: selecting device: '%s'\n", default_device->name);
-    if (default_device->type != CL_DEVICE_TYPE_GPU) {
-        fprintf(stderr, "ggml_opencl: warning, not a GPU: '%s'.\n", default_device->name);
-    }
-
-    platform = default_device->platform->id;
-    device = default_device->id;
-
-    cl_context_properties properties[] = {
-        (intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0
-    };
-
-    CL_CHECK((context = clCreateContext(properties, 1, &device, NULL, NULL, &err), err));
-
-    CL_CHECK((queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err),
-        (err != CL_INVALID_PROPERTY && err != CL_INVALID_VALUE ? err :
-        (queue = clCreateCommandQueue(context, device, 0, &err), err)
-    )));
-
-    program = build_program_from_source(context, device, program_source);
-
-    // Prepare dequantize kernels
-    CL_CHECK((kernel_q4_0 = clCreateKernel(program, "dequantize_row_q4_0", &err), err));
-    CL_CHECK((kernel_q4_1 = clCreateKernel(program, "dequantize_row_q4_1", &err), err));
-    CL_CHECK((kernel_q5_0 = clCreateKernel(program, "dequantize_row_q5_0", &err), err));
-    CL_CHECK((kernel_q5_1 = clCreateKernel(program, "dequantize_row_q5_1", &err), err));
-    CL_CHECK((kernel_q8_0 = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
-}
-
-static void ggml_cl_malloc(size_t req_size, size_t* cur_size, cl_mem_flags flags, cl_mem* buf) {
-    if (req_size <= *cur_size) {
-        return;
-    }
-
-    // Reallocate buffer with enough space
-    if (*cur_size > 0) {
-        clReleaseMemObject(*buf);
-    }
-    cl_int err;
-    CL_CHECK((*buf = clCreateBuffer(context, flags, req_size, NULL, &err), err));
-    *cur_size = req_size;
-}
-
-void ggml_cl_sgemm_wrapper(
-        const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b,
-        const int m, const int n, const int k,
-        const float alpha, const void *host_a, const int lda,
-        const float *host_b, const int ldb, const float beta,
-        float *host_c, const int ldc, const int btype) {
-
-    cl_kernel kernel;
-    size_t global = n * k, local, size_qb;
-    bool dequant;
-
-    switch (btype) {
-    case GGML_TYPE_F32:
-        dequant = false;
-        break;
-    case GGML_TYPE_Q4_0:
-        dequant = true;
-        kernel = kernel_q4_0;
-        local = 16;
-        size_qb = global * (sizeof(ggml_fp16_t) + local) / 32;
-        break;
-    case GGML_TYPE_Q4_1:
-        dequant = true;
-        kernel = kernel_q4_1;
-        local = 16;
-        size_qb = global * (sizeof(ggml_fp16_t) * 2 + local) / 32;
-        break;
-    case GGML_TYPE_Q5_0:
-        dequant = true;
-        kernel = kernel_q5_0;
-        local = 16;
-        size_qb = global * (sizeof(ggml_fp16_t) + sizeof(uint32_t) + local) / 32;
-        break;
-    case GGML_TYPE_Q5_1:
-        dequant = true;
-        kernel = kernel_q5_1;
-        local = 16;
-        size_qb = global * (sizeof(ggml_fp16_t) * 2 + sizeof(uint32_t) + local) / 32;
-        break;
-    case GGML_TYPE_Q8_0:
-        dequant = true;
-        kernel = kernel_q8_0;
-        local = 32;
-        size_qb = global * (sizeof(ggml_fp16_t) + local) / 32;
-        break;
-    default:
-        fprintf(stderr, "Error: Unsupported OpenCL btype %d\n", btype);
-        abort();
-    }
-
-    const size_t size_a =  m * k * sizeof(float);
-    const size_t size_b =  n * k * sizeof(float);
-    const size_t size_c =  m * n * sizeof(float);
-
-    // Prepare buffers
-    ggml_cl_malloc(size_a, &cl_size_a, CL_MEM_READ_ONLY, &cl_buffer_a);
-    if (dequant) {
-        ggml_cl_malloc(size_qb, &cl_size_qb, CL_MEM_READ_ONLY, &cl_buffer_qb);
-    }
-    ggml_cl_malloc(size_b, &cl_size_b, CL_MEM_READ_WRITE, &cl_buffer_b);
-    ggml_cl_malloc(size_c, &cl_size_c, CL_MEM_WRITE_ONLY, &cl_buffer_c);
-
-    cl_event ev_a, ev_qb, ev_b;
-
-    if (dequant) {
-        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &cl_buffer_qb));
-        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &cl_buffer_b));
-        CL_CHECK(clEnqueueWriteBuffer(queue, cl_buffer_qb, CL_FALSE, 0, size_qb, host_b, 0, NULL, &ev_qb));
-    } else {
-        CL_CHECK(clEnqueueWriteBuffer(queue, cl_buffer_b, CL_FALSE, 0, size_b, host_b, 0, NULL, &ev_b));
-    }
-
-    CL_CHECK(clEnqueueWriteBuffer(queue, cl_buffer_a, CL_FALSE, 0, size_a, host_a, 0, NULL, &ev_a));
-    if (dequant) {
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 1, &ev_qb, &ev_b));
-        CL_CHECK(clReleaseEvent(ev_qb));
-    }
-    CL_CHECK(clWaitForEvents(1, &ev_a));
-    CL_CHECK(clWaitForEvents(1, &ev_b));
-    CL_CHECK(clReleaseEvent(ev_a));
-    CL_CHECK(clReleaseEvent(ev_b));
-
-    cl_event ev_sgemm;
-    CLBLAST_CHECK(CLBlastSgemm(
-        (CLBlastLayout)order,
-        (CLBlastTranspose)trans_a, (CLBlastTranspose)trans_b,
-        m, n, k,
-        alpha,
-        cl_buffer_a, 0, lda,
-        cl_buffer_b, 0, ldb,
-        beta,
-        cl_buffer_c, 0, ldc,
-        &queue, &ev_sgemm));
-
-    cl_event ev_c;
-    CL_CHECK(clEnqueueReadBuffer(queue, cl_buffer_c, CL_TRUE, 0, size_c, host_c, 1, &ev_sgemm, &ev_c));
-
-    // Wait for completion
-    CL_CHECK(clWaitForEvents(1, &ev_c));
-    CL_CHECK(clReleaseEvent(ev_sgemm));
-    CL_CHECK(clReleaseEvent(ev_c));
-}

ggml-opencl.h

@@ -1,23 +1,23 @@
 #pragma once
 
+#include "ggml.h"
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
 void ggml_cl_init(void);
 
-enum ggml_blas_order {
-    GGML_BLAS_ORDER_ROW_MAJOR = 101,
-    GGML_BLAS_ORDER_COLUMN_MAJOR = 102,
-};
+void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
 
-enum ggml_blas_op {
-    GGML_BLAS_OP_N = 111,
-    GGML_BLAS_OP_T = 112,
-    GGML_BLAS_OP_C = 113,
-};
+void * ggml_cl_host_malloc(size_t size);
+void   ggml_cl_host_free(void * ptr);
 
-void ggml_cl_sgemm_wrapper(const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b, const int m, const int n, const int k, const float alpha, const void *host_a, const int lda, const float *host_b, const int ldb, const float beta, float *host_c, const int ldc, const int btype);
+void ggml_cl_transform_tensor(struct ggml_tensor * tensor);
+void ggml_cl_load_data(const char * fname, struct ggml_tensor * tensor, size_t offset);
 
 #ifdef  __cplusplus
 }

ggml-quants-k.h

@@ -0,0 +1,122 @@
+#pragma once
+
+#include "ggml.h"
+
+#include <stdint.h>
+#include <assert.h>
+#include <stddef.h>
+
+// Super-block size
+#define QK_K 256
+
+//
+// Super-block quantization structures
+//
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elemenets each
+// Effectively 2.5625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    ggml_fp16_t d;           // super-block scale for quantized scales
+    ggml_fp16_t dmin;        // super-block scale for quantized mins
+} block_q2_k;
+static_assert(sizeof(block_q2_k) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_k block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 3.4375 bits per weight
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_k;
+static_assert(sizeof(block_q3_k) == sizeof(ggml_fp16_t) + QK_K / 4 + 11 * QK_K / 64, "wrong q3_k block size/padding");
+
+// 4-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+typedef struct {
+    ggml_fp16_t d;             // super-block scale for quantized scales
+    ggml_fp16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_k;
+static_assert(sizeof(block_q4_k) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_k block size/padding");
+
+// 5-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+typedef struct {
+    ggml_fp16_t d;               // super-block scale for quantized scales
+    ggml_fp16_t dmin;            // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_k;
+static_assert(sizeof(block_q5_k) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2 + QK_K/8, "wrong q5_k block size/padding");
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_fp16_t d;           // super-block scale
+} block_q6_k;
+static_assert(sizeof(block_q6_k) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_k block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_k;
+static_assert(sizeof(block_q8_k) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_k block size/padding");
+
+
+// Quantization
+void quantize_row_q2_k_reference(const float * restrict x, block_q2_k * restrict y, int k);
+void quantize_row_q3_k_reference(const float * restrict x, block_q3_k * restrict y, int k);
+void quantize_row_q4_k_reference(const float * restrict x, block_q4_k * restrict y, int k);
+void quantize_row_q5_k_reference(const float * restrict x, block_q5_k * restrict y, int k);
+void quantize_row_q6_k_reference(const float * restrict x, block_q6_k * restrict y, int k);
+void quantize_row_q8_k_reference(const float * restrict x, block_q8_k * restrict y, int k);
+
+void quantize_row_q2_k(const float * restrict x, void * restrict y, int k);
+void quantize_row_q3_k(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_k(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_k(const float * restrict x, void * restrict y, int k);
+void quantize_row_q6_k(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_k(const float * restrict x, void * restrict y, int k);
+
+// Dequantization
+void dequantize_row_q2_k(const block_q2_k * restrict x, float * restrict y, int k);
+void dequantize_row_q3_k(const block_q3_k * restrict x, float * restrict y, int k);
+void dequantize_row_q4_k(const block_q4_k * restrict x, float * restrict y, int k);
+void dequantize_row_q5_k(const block_q5_k * restrict x, float * restrict y, int k);
+void dequantize_row_q6_k(const block_q6_k * restrict x, float * restrict y, int k);
+void dequantize_row_q8_k(const block_q8_k * restrict x, float * restrict y, int k);
+
+// Dot product
+void ggml_vec_dot_q2_k_q8_k(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q3_k_q8_k(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_k_q8_k(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_k_q8_k(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q6_k_q8_k(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+// Quantization with histogram collection
+size_t ggml_quantize_q2_k(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q3_k(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_k(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q5_k(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q6_k(const float * src, void * dst, int n, int k, int64_t * hist);
+

ggml.h

@@ -198,6 +198,7 @@
 #define GGML_MAX_PARAMS        256
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_OPT           4
+#define GGML_MAX_NAME          32
 #define GGML_DEFAULT_N_THREADS 4
 
 #define GGML_ASSERT(x) \
@@ -240,6 +241,13 @@ extern "C" {
         GGML_TYPE_Q5_1 = 7,
         GGML_TYPE_Q8_0 = 8,
         GGML_TYPE_Q8_1 = 9,
+        // k-quantizations
+        GGML_TYPE_Q2_K = 10,
+        GGML_TYPE_Q3_K = 11,
+        GGML_TYPE_Q4_K = 12,
+        GGML_TYPE_Q5_K = 13,
+        GGML_TYPE_Q6_K = 14,
+        GGML_TYPE_Q8_K = 15,
         GGML_TYPE_I8,
         GGML_TYPE_I16,
         GGML_TYPE_I32,
@@ -249,6 +257,7 @@ extern "C" {
     enum ggml_backend {
         GGML_BACKEND_CPU = 0,
         GGML_BACKEND_CUDA = 1,
+        GGML_BACKEND_CL = 2,
     };
 
     // model file types
@@ -262,6 +271,11 @@ extern "C" {
         GGML_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
     };
 
     // available tensor operations:
@@ -371,11 +385,13 @@ extern "C" {
 
         void * data;
 
-        char name[32];
+        char name[GGML_MAX_NAME];
 
         char padding[16];
     };
 
+    static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
+
     // computation graph
     struct ggml_cgraph {
         int n_nodes;
@@ -421,6 +437,7 @@ extern "C" {
     GGML_API void    ggml_print_objects(const struct ggml_context * ctx);
 
     GGML_API int64_t ggml_nelements(const struct ggml_tensor * tensor);
+    GGML_API int64_t ggml_nrows    (const struct ggml_tensor * tensor);
     GGML_API size_t  ggml_nbytes   (const struct ggml_tensor * tensor);
 
     GGML_API int     ggml_blck_size (enum ggml_type type);
@@ -428,6 +445,7 @@ extern "C" {
     GGML_API float   ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
 
     GGML_API const char * ggml_type_name(enum ggml_type type);
+    GGML_API const char * ggml_op_name  (enum ggml_op   op);
 
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
 
@@ -436,14 +454,24 @@ extern "C" {
     // TODO: temporary until model loading of ggml examples is refactored
     GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
 
+    GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
+
+    // use this to compute the memory overhead of a tensor
+    GGML_API size_t ggml_tensor_overhead(void);
+
     // main
 
     GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
-    GGML_API void    ggml_free(struct ggml_context * ctx);
+    GGML_API void                  ggml_free(struct ggml_context * ctx);
 
     GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
 
-    GGML_API size_t  ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
+    GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
+    GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
+
+    GGML_API void *  ggml_get_mem_buffer(struct ggml_context * ctx);
+    GGML_API size_t  ggml_get_mem_size  (struct ggml_context * ctx);
 
     GGML_API struct ggml_tensor * ggml_new_tensor(
             struct ggml_context * ctx,
@@ -483,6 +511,8 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
     GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src);
 
+    GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
+
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
     GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
@@ -969,6 +999,11 @@ extern "C" {
     GGML_API void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph);
     GGML_API void ggml_graph_reset  (struct ggml_cgraph * cgraph);
 
+    GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
+
+    GGML_API void               ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
+    GGML_API struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
+
     // print info and performance information for the graph
     GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
 

llama.cpp

@@ -12,6 +12,12 @@
 #include "ggml.h"
 #ifdef GGML_USE_CUBLAS
 #include "ggml-cuda.h"
+#elif defined(GGML_USE_CLBLAST)
+#include "ggml-opencl.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
 #endif
 
 #include <array>
@@ -40,6 +46,7 @@
 // available llama models
 enum e_model {
     MODEL_UNKNOWN,
+    MODEL_3B,
     MODEL_7B,
     MODEL_13B,
     MODEL_30B,
@@ -56,6 +63,7 @@ static const size_t MB = 1024*1024;
 static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0()
 {
     static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,    256ull * MB },
         { MODEL_7B,    512ull * MB },
         { MODEL_13B,   512ull * MB },
         { MODEL_30B,   512ull * MB },
@@ -67,6 +75,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0()
 static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
 {
     static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,    256ull * MB },
         { MODEL_7B,    512ull * MB },
         { MODEL_13B,   512ull * MB },
         { MODEL_30B,   512ull * MB },
@@ -79,6 +88,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
 static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
 {
     static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,    682ull * MB },
         { MODEL_7B,   1026ull * MB },
         { MODEL_13B,  1608ull * MB },
         { MODEL_30B,  3124ull * MB },
@@ -92,6 +102,7 @@ static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
 static const std::map<e_model, size_t> & MEM_REQ_EVAL()
 {
     static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,   512ull * MB },
         { MODEL_7B,   768ull * MB },
         { MODEL_13B, 1024ull * MB },
         { MODEL_30B, 1280ull * MB },
@@ -236,6 +247,10 @@ struct llama_context {
     llama_ctx_buffer buf_compute;
     llama_ctx_buffer buf_scratch[LLAMA_MAX_SCRATCH_BUFFERS];
 
+#ifdef GGML_USE_METAL
+    ggml_metal_context * ctx_metal = NULL;
+#endif
+
     int    buf_last = 0;
     size_t buf_max_size[LLAMA_MAX_SCRATCH_BUFFERS] = { 0 };
 
@@ -500,6 +515,11 @@ struct llama_file_loader {
                 case GGML_TYPE_Q5_0:
                 case GGML_TYPE_Q5_1:
                 case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q2_K:
+                case GGML_TYPE_Q3_K:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
                     break;
                 default: {
                     throw format("unrecognized tensor type %u\n", shard.type);
@@ -575,6 +595,11 @@ struct llama_file_saver {
             case GGML_TYPE_Q5_0:
             case GGML_TYPE_Q5_1:
             case GGML_TYPE_Q8_0:
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+            case GGML_TYPE_Q6_K:
                 break;
             default: LLAMA_ASSERT(false);
         }
@@ -891,12 +916,23 @@ static const char *llama_ftype_name(enum llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q5_0: return "mostly Q5_0";
         case LLAMA_FTYPE_MOSTLY_Q5_1: return "mostly Q5_1";
         case LLAMA_FTYPE_MOSTLY_Q8_0: return "mostly Q8_0";
+        // K-quants
+        case LLAMA_FTYPE_MOSTLY_Q2_K: return "mostly Q2_K";
+        case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "mostly Q3_K - Small";
+        case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "mostly Q3_K - Medium";
+        case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "mostly Q3_K - Large";
+        case LLAMA_FTYPE_MOSTLY_Q4_K_S: return "mostly Q4_K - Small";
+        case LLAMA_FTYPE_MOSTLY_Q4_K_M: return "mostly Q4_K - Medium";
+        case LLAMA_FTYPE_MOSTLY_Q5_K_S: return "mostly Q5_K - Small";
+        case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "mostly Q5_K - Medium";
+        case LLAMA_FTYPE_MOSTLY_Q6_K: return "mostly Q6_K";
         default:                      return "unknown, may not work";
     }
 }
 
 static const char *llama_model_type_name(e_model type) {
     switch (type) {
+        case MODEL_3B: return "3B";
         case MODEL_7B: return "7B";
         case MODEL_13B: return "13B";
         case MODEL_30B: return "30B";
@@ -930,6 +966,7 @@ static void llama_model_load_internal(
 
     {
         switch (hparams.n_layer) {
+            case 26: model.type = e_model::MODEL_3B; break;
             case 32: model.type = e_model::MODEL_7B; break;
             case 40: model.type = e_model::MODEL_13B; break;
             case 60: model.type = e_model::MODEL_30B; break;
@@ -1001,8 +1038,12 @@ static void llama_model_load_internal(
         }
     }
 
-#ifdef GGML_USE_CUBLAS
+#if defined(GGML_USE_CUBLAS)
 #define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_CUDA
+    fprintf(stderr, "%s: using CUDA for GPU acceleration\n", __func__);
+#elif defined(GGML_USE_CLBLAST)
+#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_CL
+    fprintf(stderr, "%s: using OpenCL for GPU acceleration\n", __func__);
 #else
 #define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_CPU
 #endif
@@ -1054,7 +1095,7 @@ static void llama_model_load_internal(
             layer.w2 = ml->get_tensor(layers_i + ".feed_forward.w2.weight", {  n_ff,   n_embd}, backend);
             layer.w3 = ml->get_tensor(layers_i + ".feed_forward.w3.weight", {n_embd,   n_ff},   backend);
 
-            if (backend == GGML_BACKEND_CUDA) {
+            if (backend == LLAMA_BACKEND_OFFLOAD) {
                 vram_total +=
                     ggml_nbytes(layer.attention_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk)             +
                     ggml_nbytes(layer.wv)             + ggml_nbytes(layer.wo) + ggml_nbytes(layer.attention_norm) +
@@ -1075,7 +1116,7 @@ static void llama_model_load_internal(
             mmapped_size - vram_total + // weights in VRAM not in memory
             MEM_REQ_SCRATCH0().at(model.type) +
             MEM_REQ_SCRATCH1().at(model.type) +
-            MEM_REQ_EVAL().at(model.type);
+            MEM_REQ_EVAL().at    (model.type);
 
         // this is the memory required by one llama_state
         const size_t mem_required_state =
@@ -1084,14 +1125,14 @@ static void llama_model_load_internal(
         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);
 
-#ifdef GGML_USE_CUBLAS
         const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
 
-        fprintf(stderr, "%s: [cublas] offloading %d layers to GPU\n", __func__, n_gpu);
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+        fprintf(stderr, "%s: offloading %d layers to GPU\n", __func__, n_gpu);
         if (n_gpu_layers > (int) hparams.n_layer) {
-            fprintf(stderr, "%s: [cublas] offloading output layer to GPU\n", __func__);
+            fprintf(stderr, "%s: offloading output layer to GPU\n", __func__);
         }
-        fprintf(stderr, "%s: [cublas] total VRAM used: %zu MB\n", __func__, vram_total / 1024 / 1024);
+        fprintf(stderr, "%s: total VRAM used: %zu MB\n", __func__, vram_total / 1024 / 1024);
 #else
         (void) n_gpu_layers;
 #endif
@@ -1104,7 +1145,7 @@ static void llama_model_load_internal(
 
     ml->load_all_data(progress_callback, progress_callback_user_data, use_mlock ? &lctx.model.mlock_mmap : NULL);
 
-#ifdef GGML_USE_CUBLAS
+#if defined(GGML_USE_CUBLAS)
     {
         size_t done_size = 0;
         size_t data_size = 0;
@@ -1125,7 +1166,28 @@ static void llama_model_load_internal(
             done_size += lt.size;
         }
     }
-#endif // GGML_USE_CUBLAS
+#elif defined(GGML_USE_CLBLAST)
+    {
+        size_t done_size = 0;
+        size_t data_size = 0;
+        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
+            data_size += lt.size;
+            if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
+                done_size += lt.size;
+            }
+        }
+        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
+            if (lt.ggml_tensor->backend != GGML_BACKEND_CL) {
+                continue;
+            }
+            if (progress_callback) {
+                progress_callback((float) done_size / data_size, progress_callback_user_data);
+            }
+            ggml_cl_load_data(fname.c_str(), lt.ggml_tensor, lt.shards.at(0).file_off);
+            done_size += lt.size;
+        }
+    }
+#endif
 
     if (progress_callback) {
         progress_callback(1.0f, progress_callback_user_data);
@@ -1161,17 +1223,19 @@ static bool llama_model_load(
 
 // evaluate the transformer
 //
-//   - lctx:      llama context
-//   - tokens:    new batch of tokens to process
-//   - n_past:    the context size so far
-//   - n_threads: number of threads to use
+//   - lctx:         llama context
+//   - tokens:       new batch of tokens to process
+//   - n_past:       the context size so far
+//   - n_threads:    number of threads to use
+//   - cgraph_fname: filename of the exported computation graph
 //
 static bool llama_eval_internal(
-        llama_context & lctx,
-    const llama_token * tokens,
-            const int   n_tokens,
-            const int   n_past,
-            const int   n_threads) {
+        llama_context &  lctx,
+    const llama_token *  tokens,
+            const int    n_tokens,
+            const int    n_past,
+            const int    n_threads,
+            const char * cgraph_fname) {
 
     // enforce that the first token is BOS
     if (n_past == 0 && tokens[0] != llama_token_bos()) {
@@ -1217,13 +1281,18 @@ static bool llama_eval_internal(
     ggml_set_name(embd, "embd");
     memcpy(embd->data, tokens, N*ggml_element_size(embd));
 
+#ifdef GGML_USE_METAL
+    if (lctx.ctx_metal && N == 1) {
+        ggml_metal_set_tensor(lctx.ctx_metal, embd);
+    }
+#endif
+
+    struct ggml_tensor * cur;
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
 
     for (int il = 0; il < n_layer; ++il) {
         struct ggml_tensor * inpSA = inpL;
 
-        struct ggml_tensor * cur;
-
         lctx.use_buf(ctx0, 0);
 
         // norm
@@ -1237,6 +1306,7 @@ static bool llama_eval_internal(
         // self-attention
         {
             // compute Q and K and RoPE them
+
             struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
             struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
             ggml_set_name(Qcur, "Qcur");
@@ -1246,6 +1316,7 @@ static bool llama_eval_internal(
             {
                 // compute the transposed [N, n_embd] V matrix
                 struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), n_embd, N));
+                ggml_set_name(Vcur, "Vcur");
 
                 struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
                 struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
@@ -1291,7 +1362,6 @@ static bool llama_eval_internal(
             struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
             ggml_set_name(KQ_soft_max, "KQ_soft_max");
 
-
             // split cached V into n_head heads
             struct ggml_tensor * V =
                 ggml_view_3d(ctx0, kv_self.v,
@@ -1373,26 +1443,61 @@ static bool llama_eval_internal(
 
     // norm
     {
+        cur = ggml_rms_norm(ctx0, inpL);
 
-        inpL = ggml_rms_norm(ctx0, inpL);
+        // cur = cur*norm(broadcasted)
+        cur = ggml_mul(ctx0, cur, model.norm);
 
-        // inpL = inpL*norm(broadcasted)
-        inpL = ggml_mul(ctx0, inpL, model.norm);
-
-        embeddings = inpL;
+        embeddings = cur;
     }
 
     // lm_head
-    inpL = ggml_mul_mat(ctx0, model.output, inpL);
+    cur = ggml_mul_mat(ctx0, model.output, cur);
 
     lctx.use_buf(ctx0, -1);
 
     // logits -> probs
-    //inpL = ggml_soft_max_inplace(ctx0, inpL);
+    //cur = ggml_soft_max_inplace(ctx0, cur);
 
     // run the computation
-    ggml_build_forward_expand(&gf, inpL);
-    ggml_graph_compute       (ctx0, &gf);
+    ggml_build_forward_expand(&gf, cur);
+
+#ifdef GGML_USE_METAL
+    if (lctx.ctx_metal && N == 1) {
+        ggml_metal_graph_compute(lctx.ctx_metal, &gf);
+        ggml_metal_get_tensor   (lctx.ctx_metal, cur);
+    } else {
+        // IMPORTANT:
+        // Since we don't have efficient Matrix x Matrix Metal multiplication yet, we fallback to vanilla
+        // ggml_graph_compute(). It uses Apple's Accelerate CBLAS API which takes advantage of the ANE or the AMX
+        // coprocessor.
+        //
+        // When we implement Matrix x Matrix Metal multiplication, we can avoid this branch.
+        // But for now, we have focused only on Matrix x Vector Metal multiplication.
+        //
+        // TODO: avoid these syncs via shared memory (ref #1696)
+        //
+        if (lctx.ctx_metal) {
+            // We need to sync the GPU KV cache with the CPU KV cache
+            ggml_metal_get_tensor(lctx.ctx_metal, kv_self.k);
+            ggml_metal_get_tensor(lctx.ctx_metal, kv_self.v);
+        }
+
+        ggml_graph_compute(ctx0, &gf);
+
+        if (lctx.ctx_metal) {
+            // We need to sync the CPU KV cache with the GPU KV cache
+            ggml_metal_set_tensor(lctx.ctx_metal, kv_self.k);
+            ggml_metal_set_tensor(lctx.ctx_metal, kv_self.v);
+        }
+    }
+#else
+    ggml_graph_compute(ctx0, &gf);
+#endif
+
+    if (cgraph_fname) {
+        ggml_graph_export(&gf, cgraph_fname);
+    }
 
 #ifdef GGML_PERF
     // print timing information per ggml operation (for debugging purposes)
@@ -1406,7 +1511,7 @@ static bool llama_eval_internal(
     //}
 
     //embd_w.resize(n_vocab*N);
-    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(cur), sizeof(float)*n_vocab*N);
 
     // update kv token count
     lctx.model.kv_self.n = n_past + N;
@@ -1417,11 +1522,11 @@ static bool llama_eval_internal(
 
         if (lctx.logits_all) {
             logits_out.resize(n_vocab * N);
-            memcpy(logits_out.data(), (float *) ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+            memcpy(logits_out.data(), (float *) ggml_get_data(cur), sizeof(float)*n_vocab*N);
         } else {
             // return result for just the last token
             logits_out.resize(n_vocab);
-            memcpy(logits_out.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+            memcpy(logits_out.data(), (float *) ggml_get_data(cur) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
         }
     }
 
@@ -2028,8 +2133,18 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_0: quantized_type = GGML_TYPE_Q5_0; break;
         case LLAMA_FTYPE_MOSTLY_Q5_1: quantized_type = GGML_TYPE_Q5_1; break;
         case LLAMA_FTYPE_MOSTLY_Q8_0: quantized_type = GGML_TYPE_Q8_0; break;
+        // K-quants
+        case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break;
+        case LLAMA_FTYPE_MOSTLY_Q3_K_S:
+        case LLAMA_FTYPE_MOSTLY_Q3_K_M:
+        case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break;
+        case LLAMA_FTYPE_MOSTLY_Q4_K_S:
+        case LLAMA_FTYPE_MOSTLY_Q4_K_M: quantized_type = GGML_TYPE_Q4_K; break;
+        case LLAMA_FTYPE_MOSTLY_Q5_K_S:
+        case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break;
+        case LLAMA_FTYPE_MOSTLY_Q6_K: quantized_type = GGML_TYPE_Q6_K; break;
         default: throw format("invalid output file type %d\n", ftype);
-    };
+    }
 
     if (nthread <= 0) {
         nthread = std::thread::hardware_concurrency();
@@ -2039,6 +2154,20 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                                                                             /*vocab_only*/ false));
     llama_file_saver file_saver(fname_out.c_str(), model_loader->file_loaders.at(0).get(), ftype);
 
+    int n_attention_wv    = 0;
+    int n_feed_forward_w2 = 0;
+    for (auto& tensor : model_loader->tensors_map.tensors) {
+        if (tensor.name.find("attention.wv.weight") != std::string::npos) {
+            ++n_attention_wv;
+        }
+        else if (tensor.name.find("feed_forward.w2.weight") != std::string::npos) {
+            ++n_feed_forward_w2;
+        }
+    }
+
+    int i_attention_wv = 0;
+    int i_feed_forward_w2 = 0;
+
     size_t total_size_org = 0;
     size_t total_size_new = 0;
     std::vector<int64_t> hist_all(1 << 4, 0);
@@ -2081,6 +2210,27 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             printf("size = %8.3f MB\n", tensor.size/1024.0/1024.0);
         } else {
             new_type = quantized_type;
+            if (tensor.name == "output.weight") new_type = GGML_TYPE_Q6_K;
+            else if (tensor.name.find("attention.wv.weight") != std::string::npos) {
+                if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+                else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
+                         (i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8 ||
+                         (i_attention_wv - n_attention_wv/8)%3 == 2)) new_type = GGML_TYPE_Q6_K;
+                ++i_attention_wv;
+            }
+            else if (tensor.name.find("feed_forward.w2.weight") != std::string::npos) {
+                if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+                else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
+                         (i_feed_forward_w2 < n_feed_forward_w2/8 || i_feed_forward_w2 >= 7*n_feed_forward_w2/8 ||
+                         (i_feed_forward_w2 - n_feed_forward_w2/8)%3 == 2)) new_type = GGML_TYPE_Q6_K;
+                ++i_feed_forward_w2;
+            }
+            else if (tensor.name.find("attention.wo.weight") != std::string::npos) {
+                if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+            }
             float * f32_data;
             size_t nelements = tensor.ne.at(0) * tensor.ne.at(1);
             llama_buffer f32_conv_buf;
@@ -2148,12 +2298,16 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             }
 
             printf("size = %8.2f MB -> %8.2f MB | hist: ", tensor.size/1024.0/1024.0, new_size/1024.0/1024.0);
+            int64_t tot_count = 0;
             for (size_t i = 0; i < hist_cur.size(); i++) {
                 hist_all[i] += hist_cur[i];
+                tot_count += hist_cur[i];
             }
 
-            for (size_t i = 0; i < hist_cur.size(); i++) {
-                printf("%5.3f ", hist_cur[i] / float(nelements));
+            if (tot_count > 0) {
+                for (size_t i = 0; i < hist_cur.size(); i++) {
+                    printf("%5.3f ", hist_cur[i] / float(nelements));
+                }
             }
             printf("\n");
         }
@@ -2171,11 +2325,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             sum_all += hist_all[i];
         }
 
-        printf("%s: hist: ", __func__);
-        for (size_t i = 0; i < hist_all.size(); i++) {
-            printf("%5.3f ", hist_all[i] / float(sum_all));
+        if (sum_all > 0) {
+            printf("%s: hist: ", __func__);
+            for (size_t i = 0; i < hist_all.size(); i++) {
+                printf("%5.3f ", hist_all[i] / float(sum_all));
+            }
+            printf("\n");
         }
-        printf("\n");
     }
 }
 
@@ -2217,8 +2373,8 @@ struct llama_context * llama_init_from_file(
     ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
     if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_gpu_layers, memory_type,
-                          params.use_mmap, params.use_mlock, params.vocab_only,
-                          params.progress_callback, params.progress_callback_user_data)) {
+                params.use_mmap, params.use_mlock, params.vocab_only,
+                params.progress_callback, params.progress_callback_user_data)) {
         fprintf(stderr, "%s: failed to load model\n", __func__);
         llama_free(ctx);
         return nullptr;
@@ -2256,6 +2412,25 @@ struct llama_context * llama_init_from_file(
         ctx->buf_scratch[1].resize(MEM_REQ_SCRATCH1().at(ctx->model.type));
     }
 
+#ifdef GGML_USE_METAL
+    if (params.n_gpu_layers > 0) {
+        // this allocates all Metal resources and memory buffers
+        ctx->ctx_metal = ggml_metal_init();
+
+        if (params.use_mmap) {
+            ggml_metal_add_buffer(ctx->ctx_metal, "data", ctx->model.mapping->addr, ctx->model.mapping->size);
+            ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr,    ctx->buf_compute.size);
+        } else {
+            ggml_metal_add_buffer(ctx->ctx_metal, "data", ggml_get_mem_buffer(ctx->model.ctx), ggml_get_mem_size(ctx->model.ctx));
+            ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr,               ctx->buf_compute.size);
+        }
+
+        ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->model.kv_self.buf.addr, ctx->model.kv_self.buf.size);
+        ggml_metal_add_buffer(ctx->ctx_metal, "scr0", ctx->buf_scratch[0].addr,    ctx->buf_scratch[0].size);
+        ggml_metal_add_buffer(ctx->ctx_metal, "scr1", ctx->buf_scratch[1].addr,    ctx->buf_scratch[1].size);
+    }
+#endif
+
     return ctx;
 }
 
@@ -2871,7 +3046,7 @@ int llama_eval(
                          int   n_tokens,
                          int   n_past,
                          int   n_threads) {
-    if (!llama_eval_internal(*ctx, tokens, n_tokens, n_past, n_threads)) {
+    if (!llama_eval_internal(*ctx, tokens, n_tokens, n_past, n_threads, nullptr)) {
         fprintf(stderr, "%s: failed to eval\n", __func__);
         return 1;
     }
@@ -2886,6 +3061,20 @@ int llama_eval(
     return 0;
 }
 
+int llama_eval_export(struct llama_context * ctx, const char * fname) {
+    const int n_batch = 1;
+    const int n_ctx   = 512 - n_batch;
+
+    const std::vector<llama_token> tmp(n_batch, llama_token_bos());
+
+    if (!llama_eval_internal(*ctx, tmp.data(), tmp.size(), n_ctx, 1, fname)) {
+        fprintf(stderr, "%s: failed to eval\n", __func__);
+        return 1;
+    }
+
+    return 0;
+}
+
 int llama_tokenize(
         struct llama_context * ctx,
                   const char * text,

llama.h

@@ -31,6 +31,11 @@
 #define LLAMA_SESSION_MAGIC          LLAMA_FILE_MAGIC_GGSN
 #define LLAMA_SESSION_VERSION        1
 
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
+// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
+#define LLAMA_SUPPORTS_GPU_OFFLOAD
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -89,6 +94,15 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K          = 10,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
     };
 
     LLAMA_API struct llama_context_params llama_context_default_params();
@@ -168,6 +182,12 @@ extern "C" {
                              int   n_past,
                              int   n_threads);
 
+    // Export a static computation graph for context of 511 and batch size of 1
+    // NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
+    //       parameters here to keep things simple
+    // IMPORTANT: do not use for anything else other than debugging and testing!
+    LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
+
     // Convert the provided text into tokens.
     // The tokens pointer must be large enough to hold the resulting tokens.
     // Returns the number of tokens on success, no more than n_max_tokens

tests/test-quantize-fns.cpp

@@ -12,6 +12,8 @@
 
 const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001;
 const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002;
+const float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075;
+const float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040;
 const float MAX_DOT_PRODUCT_ERROR = 0.02;
 
 const char* RESULT_STR[] = {"ok", "FAILED"};
@@ -122,7 +124,10 @@ int main(int argc, char * argv[]) {
 
         if (qfns.quantize_row_q && qfns.dequantize_row_q) {
             const float total_error = total_quantization_error(qfns, test_size, test_data.data());
-            failed = !(total_error < MAX_QUANTIZATION_TOTAL_ERROR);
+            const float max_quantization_error =
+                type == GGML_TYPE_Q2_K ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
+                type == GGML_TYPE_Q3_K ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS : MAX_QUANTIZATION_TOTAL_ERROR;
+            failed = !(total_error < max_quantization_error);
             num_failed += failed;
             if (failed || verbose) {
                 printf("%5s absolute quantization error:    %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], total_error);