convert : update phi-2 to latest HF repo (#4903)

* convert : update phi-2 to latest HF repo

ggml-ci

* py : try to fix flake stuff

.devops/full-cuda.Dockerfile

@@ -14,7 +14,8 @@ ARG CUDA_DOCKER_ARCH=all
 RUN apt-get update && \
     apt-get install -y build-essential python3 python3-pip git
 
-COPY requirements.txt requirements.txt
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
 
 RUN pip install --upgrade pip setuptools wheel \
     && pip install -r requirements.txt

.devops/full-rocm.Dockerfile

@@ -23,7 +23,8 @@ ARG ROCM_DOCKER_ARCH=\
     gfx1101 \
     gfx1102
 
-COPY requirements.txt requirements.txt
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
 
 RUN pip install --upgrade pip setuptools wheel \
     && pip install -r requirements.txt

.devops/full.Dockerfile

@@ -5,7 +5,8 @@ FROM ubuntu:$UBUNTU_VERSION as build
 RUN apt-get update && \
     apt-get install -y build-essential python3 python3-pip git
 
-COPY requirements.txt requirements.txt
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
 
 RUN pip install --upgrade pip setuptools wheel \
     && pip install -r requirements.txt

.devops/main-rocm.Dockerfile

@@ -23,7 +23,8 @@ ARG ROCM_DOCKER_ARCH=\
     gfx1101 \
     gfx1102
 
-COPY requirements.txt requirements.txt
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
 
 RUN pip install --upgrade pip setuptools wheel \
     && pip install -r requirements.txt

.devops/nix/apps.nix

@@ -0,0 +1,22 @@
+{
+  perSystem =
+    { config, lib, ... }:
+    {
+      apps =
+        let
+          inherit (config.packages) default;
+          binaries = [
+            "llama"
+            "llama-embedding"
+            "llama-server"
+            "quantize"
+            "train-text-from-scratch"
+          ];
+          mkApp = name: {
+            type = "app";
+            program = "${default}/bin/${name}";
+          };
+        in
+        lib.genAttrs binaries mkApp;
+    };
+}

.devops/nix/devshells.nix

@@ -0,0 +1,13 @@
+{
+  perSystem =
+    { config, lib, ... }:
+    {
+      devShells =
+        lib.concatMapAttrs
+          (name: package: {
+            ${name} = package.passthru.shell;
+            ${name + "-extra"} = package.passthru.shell-extra;
+          })
+          config.packages;
+    };
+}

.devops/nix/jetson-support.nix

@@ -0,0 +1,39 @@
+{ inputs, ... }:
+{
+  perSystem =
+    {
+      config,
+      system,
+      lib,
+      pkgsCuda,
+      ...
+    }:
+    {
+      legacyPackages =
+        let
+          caps.llamaPackagesXavier = "7.2";
+          caps.llamaPackagesOrin = "8.7";
+          caps.llamaPackagesTX2 = "6.2";
+          caps.llamaPackagesNano = "5.3";
+
+          pkgsFor =
+            cap:
+            import inputs.nixpkgs {
+              inherit system;
+              config = {
+                cudaSupport = true;
+                cudaCapabilities = [ cap ];
+                cudaEnableForwardCompat = false;
+                inherit (pkgsCuda.config) allowUnfreePredicate;
+              };
+            };
+        in
+        builtins.mapAttrs (name: cap: (pkgsFor cap).callPackage ./scope.nix { }) caps;
+
+      packages = lib.optionalAttrs (system == "aarch64-linux") {
+        jetson-xavier = config.legacyPackages.llamaPackagesXavier.llama-cpp;
+        jetson-orin = config.legacyPackages.llamaPackagesOrin.llama-cpp;
+        jetson-nano = config.legacyPackages.llamaPackagesNano.llama-cpp;
+      };
+    };
+}

.devops/nix/nixpkgs-instances.nix

@@ -0,0 +1,35 @@
+{ inputs, ... }:
+{
+  # The _module.args definitions are passed on to modules as arguments. E.g.
+  # the module `{ pkgs ... }: { /* config */ }` implicitly uses
+  # `_module.args.pkgs` (defined in this case by flake-parts).
+  perSystem =
+    { system, ... }:
+    {
+      _module.args = {
+        pkgsCuda = import inputs.nixpkgs {
+          inherit system;
+          # Ensure dependencies use CUDA consistently (e.g. that openmpi, ucc,
+          # and ucx are built with CUDA support)
+          config.cudaSupport = true;
+          config.allowUnfreePredicate =
+            p:
+            builtins.all
+              (
+                license:
+                license.free
+                || builtins.elem license.shortName [
+                  "CUDA EULA"
+                  "cuDNN EULA"
+                ]
+              )
+              (p.meta.licenses or [ p.meta.license ]);
+        };
+        # Ensure dependencies use ROCm consistently
+        pkgsRocm = import inputs.nixpkgs {
+          inherit system;
+          config.rocmSupport = true;
+        };
+      };
+    };
+}

.devops/nix/package.nix

@@ -0,0 +1,265 @@
+{
+  lib,
+  config,
+  stdenv,
+  mkShell,
+  cmake,
+  ninja,
+  pkg-config,
+  git,
+  python3,
+  mpi,
+  openblas, # TODO: Use the generic `blas` so users could switch between alternative implementations
+  cudaPackages,
+  darwin,
+  rocmPackages,
+  clblast,
+  useBlas ? builtins.all (x: !x) [
+    useCuda
+    useMetalKit
+    useOpenCL
+    useRocm
+  ],
+  useCuda ? config.cudaSupport,
+  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
+  useMpi ? false, # Increases the runtime closure size by ~700M
+  useOpenCL ? false,
+  useRocm ? config.rocmSupport,
+  llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
+}@inputs:
+
+let
+  inherit (lib)
+    cmakeBool
+    cmakeFeature
+    optionals
+    strings
+    versionOlder
+    ;
+
+  # It's necessary to consistently use backendStdenv when building with CUDA support,
+  # otherwise we get libstdc++ errors downstream.
+  stdenv = throw "Use effectiveStdenv instead";
+  effectiveStdenv = if useCuda then cudaPackages.backendStdenv else inputs.stdenv;
+
+  suffices =
+    lib.optionals useBlas [ "BLAS" ]
+    ++ lib.optionals useCuda [ "CUDA" ]
+    ++ lib.optionals useMetalKit [ "MetalKit" ]
+    ++ lib.optionals useMpi [ "MPI" ]
+    ++ lib.optionals useOpenCL [ "OpenCL" ]
+    ++ lib.optionals useRocm [ "ROCm" ];
+
+  pnameSuffix =
+    strings.optionalString (suffices != [ ])
+      "-${strings.concatMapStringsSep "-" strings.toLower suffices}";
+  descriptionSuffix =
+    strings.optionalString (suffices != [ ])
+      ", accelerated with ${strings.concatStringsSep ", " suffices}";
+
+  # TODO: package the Python in this repository in a Nix-like way.
+  # It'd be nice to migrate to buildPythonPackage, as well as ensure this repo
+  # is PEP 517-compatible, and ensure the correct .dist-info is generated.
+  # https://peps.python.org/pep-0517/
+  llama-python = python3.withPackages (
+    ps: [
+      ps.numpy
+      ps.sentencepiece
+    ]
+  );
+
+  # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
+  llama-python-extra = python3.withPackages (
+    ps: [
+      ps.numpy
+      ps.sentencepiece
+      ps.torchWithoutCuda
+      ps.transformers
+    ]
+  );
+
+  # apple_sdk is supposed to choose sane defaults, no need to handle isAarch64
+  # separately
+  darwinBuildInputs =
+    with darwin.apple_sdk.frameworks;
+    [
+      Accelerate
+      CoreVideo
+      CoreGraphics
+    ]
+    ++ optionals useMetalKit [ MetalKit ];
+
+  cudaBuildInputs = with cudaPackages; [
+    cuda_cccl.dev # <nv/target>
+
+    # A temporary hack for reducing the closure size, remove once cudaPackages
+    # have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
+    cuda_cudart.dev
+    cuda_cudart.lib
+    cuda_cudart.static
+    libcublas.dev
+    libcublas.lib
+    libcublas.static
+  ];
+
+  rocmBuildInputs = with rocmPackages; [
+    clr
+    hipblas
+    rocblas
+  ];
+in
+
+effectiveStdenv.mkDerivation (
+  finalAttrs: {
+    pname = "llama-cpp${pnameSuffix}";
+    version = llamaVersion;
+
+    src = lib.cleanSourceWith {
+      filter =
+        name: type:
+        !(builtins.any (_: _) [
+          (lib.hasSuffix ".nix" name) # Ignore *.nix files when computing outPaths
+          (name == "README.md") # Ignore *.md changes whe computing outPaths
+          (lib.hasPrefix "." name) # Skip hidden files and directories
+        ]);
+      src = lib.cleanSource ../../.;
+    };
+
+    postPatch = ''
+      substituteInPlace ./ggml-metal.m \
+        --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
+
+      # TODO: Package up each Python script or service appropriately.
+      # If we were to migrate to buildPythonPackage and prepare the `pyproject.toml`,
+      # we could make those *.py into setuptools' entrypoints
+      substituteInPlace ./*.py --replace "/usr/bin/env python" "${llama-python}/bin/python"
+    '';
+
+    nativeBuildInputs =
+      [
+        cmake
+        ninja
+        pkg-config
+        git
+      ]
+      ++ optionals useCuda [
+        cudaPackages.cuda_nvcc
+
+        # TODO: Replace with autoAddDriverRunpath
+        # once https://github.com/NixOS/nixpkgs/pull/275241 has been merged
+        cudaPackages.autoAddOpenGLRunpathHook
+      ];
+
+    buildInputs =
+      optionals effectiveStdenv.isDarwin darwinBuildInputs
+      ++ optionals useCuda cudaBuildInputs
+      ++ optionals useMpi [ mpi ]
+      ++ optionals useOpenCL [ clblast ]
+      ++ optionals useRocm rocmBuildInputs;
+
+    cmakeFlags =
+      [
+        (cmakeBool "LLAMA_NATIVE" true)
+        (cmakeBool "LLAMA_BUILD_SERVER" true)
+        (cmakeBool "BUILD_SHARED_LIBS" true)
+        (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
+        (cmakeBool "LLAMA_BLAS" useBlas)
+        (cmakeBool "LLAMA_CLBLAST" useOpenCL)
+        (cmakeBool "LLAMA_CUBLAS" useCuda)
+        (cmakeBool "LLAMA_HIPBLAS" useRocm)
+        (cmakeBool "LLAMA_METAL" useMetalKit)
+        (cmakeBool "LLAMA_MPI" useMpi)
+      ]
+      ++ optionals useCuda [
+        (
+          with cudaPackages.flags;
+          cmakeFeature "CMAKE_CUDA_ARCHITECTURES" (
+            builtins.concatStringsSep ";" (map dropDot cudaCapabilities)
+          )
+        )
+      ]
+      ++ optionals useRocm [
+        (cmakeFeature "CMAKE_C_COMPILER" "hipcc")
+        (cmakeFeature "CMAKE_CXX_COMPILER" "hipcc")
+
+        # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
+        # in https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
+        # and select the line that matches the current nixpkgs version of rocBLAS.
+        # Should likely use `rocmPackages.clr.gpuTargets`.
+        "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
+      ]
+      ++ optionals useMetalKit [ (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1") ]
+      ++ optionals useBlas [ (lib.cmakeFeature "LLAMA_BLAS_VENDOR" "OpenBLAS") ];
+
+    # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
+    # if they haven't been added yet.
+    postInstall = ''
+      mv $out/bin/main $out/bin/llama
+      mv $out/bin/server $out/bin/llama-server
+      mkdir -p $out/include
+      cp $src/llama.h $out/include/
+    '';
+
+    # Define the shells here, but don't add in the inputsFrom to avoid recursion.
+    passthru = {
+      inherit
+        useBlas
+        useCuda
+        useMetalKit
+        useMpi
+        useOpenCL
+        useRocm
+        ;
+
+      shell = mkShell {
+        name = "shell-${finalAttrs.finalPackage.name}";
+        description = "contains numpy and sentencepiece";
+        buildInputs = [ llama-python ];
+        inputsFrom = [ finalAttrs.finalPackage ];
+      };
+
+      shell-extra = mkShell {
+        name = "shell-extra-${finalAttrs.finalPackage.name}";
+        description = "contains numpy, sentencepiece, torchWithoutCuda, and transformers";
+        buildInputs = [ llama-python-extra ];
+        inputsFrom = [ finalAttrs.finalPackage ];
+      };
+    };
+
+    meta = {
+      # Configurations we don't want even the CI to evaluate. Results in the
+      # "unsupported platform" messages. This is mostly a no-op, because
+      # cudaPackages would've refused to evaluate anyway.
+      badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin;
+
+      # Configurations that are known to result in build failures. Can be
+      # overridden by importing Nixpkgs with `allowBroken = true`.
+      broken = (useMetalKit && !effectiveStdenv.isDarwin);
+
+      description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
+      homepage = "https://github.com/ggerganov/llama.cpp/";
+      license = lib.licenses.mit;
+
+      # Accommodates `nix run` and `lib.getExe`
+      mainProgram = "llama";
+
+      # These people might respond, on the best effort basis, if you ping them
+      # in case of Nix-specific regressions or for reviewing Nix-specific PRs.
+      # Consider adding yourself to this list if you want to ensure this flake
+      # stays maintained and you're willing to invest your time. Do not add
+      # other people without their consent. Consider removing people after
+      # they've been unreachable for long periods of time.
+
+      # Note that lib.maintainers is defined in Nixpkgs, but you may just add
+      # an attrset following the same format as in
+      # https://github.com/NixOS/nixpkgs/blob/f36a80e54da29775c78d7eff0e628c2b4e34d1d7/maintainers/maintainer-list.nix
+      maintainers = with lib.maintainers; [
+        philiptaron
+        SomeoneSerge
+      ];
+
+      # Extend `badPlatforms` instead
+      platforms = lib.platforms.all;
+    };
+  }
+)

.devops/nix/scope.nix

@@ -0,0 +1,12 @@
+{
+  lib,
+  newScope,
+  llamaVersion ? "0.0.0",
+}:
+
+lib.makeScope newScope (
+  self: {
+    inherit llamaVersion;
+    llama-cpp = self.callPackage ./package.nix { };
+  }
+)

.github/ISSUE_TEMPLATE/bug.md

@@ -6,179 +6,4 @@ assignees: ''
 
 ---
 
-# Prerequisites
-
-Please answer the following questions for yourself before submitting an issue.
-
-- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
-- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
-- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed).
-- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share.
-
-# Expected Behavior
-
-Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do.
-
-# Current Behavior
-
-Please provide a detailed written description of what `llama.cpp` did, instead.
-
-# Environment and Context
-
-Please provide detailed information about your computer setup. This is important in case the issue is not reproducible except for under certain specific conditions.
-
-* Physical (or virtual) hardware you are using, e.g. for Linux:
-
-`$ lscpu`
-
-* Operating System, e.g. for Linux:
-
-`$ uname -a`
-
-* SDK version, e.g. for Linux:
-
-```
-$ python3 --version
-$ make --version
-$ g++ --version
-```
-
-# Failure Information (for bugs)
-
-Please help provide information about the failure / bug.
-
-# Steps to Reproduce
-
-Please provide detailed steps for reproducing the issue. We are not sitting in front of your screen, so the more detail the better.
-
-1. step 1
-2. step 2
-3. step 3
-4. etc.
-
-# Failure Logs
-
-Please include any relevant log snippets or files. If it works under one configuration but not under another, please provide logs for both configurations and their corresponding outputs so it is easy to see where behavior changes.
-
-Also, please try to **avoid using screenshots** if at all possible. Instead, copy/paste the console output and use [Github's markdown](https://docs.github.com/en/get-started/writing-on-github/getting-started-with-writing-and-formatting-on-github/basic-writing-and-formatting-syntax) to cleanly format your logs for easy readability.
-
-Example environment info:
-```
-llama.cpp$ git log | head -1
-commit 2af23d30434a677c6416812eea52ccc0af65119c
-
-llama.cpp$ lscpu | egrep "AMD|Flags"
-Vendor ID:                       AuthenticAMD
-Model name:                      AMD Ryzen Threadripper 1950X 16-Core Processor
-Flags:                           fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid amd_dcm aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb hw_pstate ssbd ibpb vmmcall fsgsbase bmi1 avx2 smep bmi2 rdseed adx smap clflushopt sha_ni xsaveopt xsavec xgetbv1 xsaves clzero irperf xsaveerptr arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif overflow_recov succor smca sme sev
-Virtualization:                  AMD-V
-
-llama.cpp$ python3 --version
-Python 3.10.9
-
-llama.cpp$ pip list | egrep "torch|numpy|sentencepiece"
-numpy                         1.24.2
-numpydoc                      1.5.0
-sentencepiece                 0.1.97
-torch                         1.13.1
-torchvision                   0.14.1
-
-llama.cpp$ make --version | head -1
-GNU Make 4.3
-
-$ md5sum ./models/65B/ggml-model-q4_0.bin
-dbdd682cce80e2d6e93cefc7449df487  ./models/65B/ggml-model-q4_0.bin
-```
-
-Example run with the Linux command [perf](https://www.brendangregg.com/perf.html)
-```
-llama.cpp$ perf stat ./main -m ./models/65B/ggml-model-q4_0.bin -t 16 -n 1024 -p "Please close your issue when it has been answered."
-main: seed = 1679149377
-llama_model_load: loading model from './models/65B/ggml-model-q4_0.bin' - please wait ...
-llama_model_load: n_vocab = 32000
-llama_model_load: n_ctx   = 512
-llama_model_load: n_embd  = 8192
-llama_model_load: n_mult  = 256
-llama_model_load: n_head  = 64
-llama_model_load: n_layer = 80
-llama_model_load: n_rot   = 128
-llama_model_load: f16     = 2
-llama_model_load: n_ff    = 22016
-llama_model_load: n_parts = 8
-llama_model_load: ggml ctx size = 41477.73 MB
-llama_model_load: memory_size =  2560.00 MB, n_mem = 40960
-llama_model_load: loading model part 1/8 from './models/65B/ggml-model-q4_0.bin'
-llama_model_load: .......................................................................................... done
-llama_model_load: model size =  4869.09 MB / num tensors = 723
-llama_model_load: loading model part 2/8 from './models/65B/ggml-model-q4_0.bin.1'
-llama_model_load: .......................................................................................... done
-llama_model_load: model size =  4869.09 MB / num tensors = 723
-llama_model_load: loading model part 3/8 from './models/65B/ggml-model-q4_0.bin.2'
-llama_model_load: .......................................................................................... done
-llama_model_load: model size =  4869.09 MB / num tensors = 723
-llama_model_load: loading model part 4/8 from './models/65B/ggml-model-q4_0.bin.3'
-llama_model_load: .......................................................................................... done
-llama_model_load: model size =  4869.09 MB / num tensors = 723
-llama_model_load: loading model part 5/8 from './models/65B/ggml-model-q4_0.bin.4'
-llama_model_load: .......................................................................................... done
-llama_model_load: model size =  4869.09 MB / num tensors = 723
-llama_model_load: loading model part 6/8 from './models/65B/ggml-model-q4_0.bin.5'
-llama_model_load: .......................................................................................... done
-llama_model_load: model size =  4869.09 MB / num tensors = 723
-llama_model_load: loading model part 7/8 from './models/65B/ggml-model-q4_0.bin.6'
-llama_model_load: .......................................................................................... done
-llama_model_load: model size =  4869.09 MB / num tensors = 723
-llama_model_load: loading model part 8/8 from './models/65B/ggml-model-q4_0.bin.7'
-llama_model_load: .......................................................................................... done
-llama_model_load: model size =  4869.09 MB / num tensors = 723
-
-system_info: n_threads = 16 / 32 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 |
-
-main: prompt: 'Please close your issue when it has been answered.'
-main: number of tokens in prompt = 11
-     1 -> ''
- 12148 -> 'Please'
-  3802 -> ' close'
-   596 -> ' your'
-  2228 -> ' issue'
-   746 -> ' when'
-   372 -> ' it'
-   756 -> ' has'
-  1063 -> ' been'
-  7699 -> ' answered'
- 29889 -> '.'
-
-sampling parameters: temp = 0.800000, top_k = 40, top_p = 0.950000, repeat_last_n = 64, repeat_penalty = 1.300000
-
-
-Please close your issue when it has been answered.
-@duncan-donut: I'm trying to figure out what kind of "support" you need for this script and why, exactly? Is there a question about how the code works that hasn't already been addressed in one or more comments below this ticket, or are we talking something else entirely like some sorta bugfixing job because your server setup is different from mine??
-I can understand if your site needs to be running smoothly and you need help with a fix of sorts but there should really be nothing wrong here that the code itself could not handle. And given that I'm getting reports about how it works perfectly well on some other servers, what exactly are we talking? A detailed report will do wonders in helping us get this resolved for ya quickly so please take your time and describe the issue(s) you see as clearly & concisely as possible!!
-@duncan-donut: I'm not sure if you have access to cPanel but you could try these instructions. It is worth a shot! Let me know how it goes (or what error message, exactly!) when/if ya give that code a go? [end of text]
-
-
-main: mem per token = 71159620 bytes
-main:     load time = 19309.95 ms
-main:   sample time =   168.62 ms
-main:  predict time = 223895.61 ms / 888.47 ms per token
-main:    total time = 246406.42 ms
-
- Performance counter stats for './main -m ./models/65B/ggml-model-q4_0.bin -t 16 -n 1024 -p Please close your issue when it has been answered.':
-
-        3636882.89 msec task-clock                #   14.677 CPUs utilized
-             13509      context-switches          #    3.714 /sec
-              2436      cpu-migrations            #    0.670 /sec
-          10476679      page-faults               #    2.881 K/sec
-    13133115082869      cycles                    #    3.611 GHz                      (16.77%)
-       29314462753      stalled-cycles-frontend   #    0.22% frontend cycles idle     (16.76%)
-    10294402631459      stalled-cycles-backend    #   78.39% backend cycles idle      (16.74%)
-    23479217109614      instructions              #    1.79  insn per cycle
-                                                  #    0.44  stalled cycles per insn  (16.76%)
-     2353072268027      branches                  #  647.002 M/sec                    (16.77%)
-        1998682780      branch-misses             #    0.08% of all branches          (16.76%)
-
-     247.802177522 seconds time elapsed
-
-    3618.573072000 seconds user
-      18.491698000 seconds sys
-```
+Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug.

.github/workflows/build.yml

@@ -515,7 +515,6 @@ jobs:
       - name: Build Xcode project
         run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
 
-
 #  freeBSD-latest:
 #    runs-on: macos-12
 #    steps:

.github/workflows/docker.yml

@@ -52,6 +52,36 @@ jobs:
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
+      # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
+      - name: Free Disk Space (Ubuntu)
+        uses: jlumbroso/free-disk-space@main
+        with:
+          # this might remove tools that are actually needed,
+          # if set to "true" but frees about 6 GB
+          tool-cache: false
+
+          # all of these default to true, but feel free to set to
+          # "false" if necessary for your workflow
+          android: true
+          dotnet: true
+          haskell: true
+          large-packages: true
+          docker-images: true
+          swap-storage: true
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
       - name: Build and push Docker image (versioned)
         if: github.event_name == 'push'
         uses: docker/build-push-action@v4
@@ -59,7 +89,7 @@ jobs:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
       - name: Build and push Docker image (tagged)
@@ -68,5 +98,5 @@ jobs:
           context: .
           push: ${{ github.event_name == 'push' }}
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}"
+          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
           file: ${{ matrix.config.dockerfile }}

.github/workflows/nix-ci.yml

@@ -0,0 +1,112 @@
+name: Nix CI
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix']
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix']
+
+jobs:
+  nix-eval:
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ ubuntu-latest, macos-latest ]
+    runs-on: ${{ matrix.os }}
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+    - name: Install Nix
+      uses: DeterminateSystems/nix-installer-action@v9
+      with:
+        github-token: ${{ secrets.GITHUB_TOKEN }}
+        extra-conf: |
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+    - uses: DeterminateSystems/magic-nix-cache-action@v2
+      with:
+        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
+    - name: List all flake outputs
+      run: nix flake show --all-systems
+    - name: Show all output paths
+      run: >
+          nix run github:nix-community/nix-eval-jobs
+          -- --gc-roots-dir gcroot
+          --flake
+          ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
+  nix-build:
+    if: ${{ vars.CACHIX_NAME != '' }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ ubuntu-latest, macos-latest ]
+    runs-on: ${{ matrix.os }}
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+    - name: Install Nix
+      uses: DeterminateSystems/nix-installer-action@v9
+      with:
+        github-token: ${{ secrets.GITHUB_TOKEN }}
+        extra-conf: |
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+    - uses: DeterminateSystems/magic-nix-cache-action@v2
+      with:
+        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
+    - name: Set-up cachix to push the results to
+      uses: cachix/cachix-action@v13
+      with:
+        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
+        name: ${{ vars.CACHIX_NAME }}
+    - name: Build
+      run: >
+          nix run github:Mic92/nix-fast-build
+          -- --skip-cached --no-nom
+          --flake
+          ".#checks.$(nix eval --raw --impure --expr builtins.currentSystem)"
+  nix-build-aarch64:
+    if: ${{ vars.CACHIX_NAME != '' }}
+    runs-on: ubuntu-latest
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+    - name: Install QEMU
+      # Copy-paste from https://github.com/orgs/community/discussions/8305#discussioncomment-5888654
+      run: |
+        sudo apt-get install -y qemu-user-static qemu-system-aarch64
+        sudo usermod -a -G kvm $USER
+    - name: Install Nix
+      uses: DeterminateSystems/nix-installer-action@v9
+      with:
+        github-token: ${{ secrets.GITHUB_TOKEN }}
+        extra-conf: |
+          extra-platforms = aarch64-linux
+          extra-system-features = nixos-test kvm
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+    - uses: DeterminateSystems/magic-nix-cache-action@v2
+      with:
+        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
+    - name: Set-up cachix to push the results to
+      uses: cachix/cachix-action@v13
+      with:
+        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
+        name: ${{ vars.CACHIX_NAME }}
+    - name: Show all output paths
+      run: >
+          nix run github:nix-community/nix-eval-jobs
+          -- --gc-roots-dir gcroot
+          --flake
+          ".#packages.aarch64-linux"
+    - name: Build
+      run: >
+          nix run github:Mic92/nix-fast-build
+          -- --skip-cached --no-nom
+          --systems aarch64-linux
+          --flake
+          ".#checks.aarch64-linux"

.github/workflows/nix-flake-update.yml

@@ -0,0 +1,22 @@
+name: update-flake-lock
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: '0 0 * * 0' # runs weekly on Sunday at 00:00
+
+jobs:
+  lockfile:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+      - name: Install Nix
+        uses: DeterminateSystems/nix-installer-action@main
+      - name: Update flake.lock
+        uses: DeterminateSystems/update-flake-lock@main
+        with:
+          pr-title: "nix: update flake.lock"
+          pr-labels: |
+            nix
+          pr-reviewers: philiptaron,SomeoneSerge
+          token: ${{ secrets.FLAKE_TOKEN }}

.github/workflows/nix-publish-flake.yml

@@ -0,0 +1,36 @@
+# Make the flake discoverable on https://flakestry.dev and https://flakehub.com/flakes
+name: "Publish a flake to flakestry & flakehub"
+on:
+    push:
+        tags:
+        - "*"
+    workflow_dispatch:
+        inputs:
+            tag:
+                description: "The existing tag to publish"
+                type: "string"
+                required: true
+jobs:
+    flakestry-publish:
+        runs-on: ubuntu-latest
+        permissions:
+            id-token: "write"
+            contents: "read"
+        steps:
+            - uses: flakestry/flakestry-publish@main
+              with:
+                version: "${{ inputs.tag || github.ref_name }}"
+    flakehub-publish:
+      runs-on: "ubuntu-latest"
+      permissions:
+        id-token: "write"
+        contents: "read"
+      steps:
+        - uses: "actions/checkout@v4"
+          with:
+            ref: "${{ (inputs.tag != null) && format('refs/tags/{0}', inputs.tag) || '' }}"
+        - uses: "DeterminateSystems/nix-installer-action@main"
+        - uses: "DeterminateSystems/flakehub-push@main"
+          with:
+            visibility: "public"
+            tag: "${{ inputs.tag }}"

.github/workflows/python-check-requirements.yml

@@ -0,0 +1,29 @@
+name: Python check requirements.txt
+
+on:
+  push:
+    paths:
+      - 'scripts/check-requirements.sh'
+      - 'convert*.py'
+      - 'requirements.txt'
+      - 'requirements/*.txt'
+  pull_request:
+    paths:
+      - 'scripts/check-requirements.sh'
+      - 'convert*.py'
+      - 'requirements.txt'
+      - 'requirements/*.txt'
+
+jobs:
+  python-check-requirements:
+    runs-on: ubuntu-latest
+    name: check-requirements
+    steps:
+      - name: Check out source repository
+        uses: actions/checkout@v3
+      - name: Set up Python environment
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.11"
+      - name: Run check-requirements.sh script
+        run:  bash scripts/check-requirements.sh nocleanup

.gitignore

@@ -43,13 +43,16 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
+/imatrix
 /infill
 /libllama.so
 /llama-bench
 /llava-cli
 /lookahead
+/lookup
 /main
 /metal
+/passkey
 /perplexity
 /q8dot
 /quantize

CMakeLists.txt

@@ -91,9 +91,11 @@ set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for
 set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                              "llama: max. batch size for using peer access")
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
+option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
+option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 
@@ -153,9 +155,9 @@ if (APPLE AND LLAMA_ACCELERATE)
 endif()
 
 if (LLAMA_METAL)
-    find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
-    find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
-    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+    find_library(METAL_FRAMEWORK    Metal      REQUIRED)
+    find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
 
     message(STATUS "Metal framework found")
     set(GGML_HEADERS_METAL ggml-metal.h)
@@ -172,6 +174,35 @@ if (LLAMA_METAL)
     # copy ggml-metal.metal to bin directory
     configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
 
+    if (LLAMA_METAL_SHADER_DEBUG)
+        # custom command to do the following:
+        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
+        #
+        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
+        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
+        # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
+        #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
+        set(XC_FLAGS -fno-fast-math -fno-inline -g)
+        if (LLAMA_QKK_64)
+            set(XC_FLAGS ${XC_FLAGS} -DQK_K=64)
+        endif()
+
+        add_custom_command(
+            OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
+            COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            DEPENDS ggml-metal.metal
+            COMMENT "Compiling Metal kernels"
+        )
+
+        add_custom_target(
+            ggml-metal ALL
+            DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        )
+    endif()
+
     set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
         ${FOUNDATION_LIBRARY}
         ${METAL_FRAMEWORK}
@@ -199,7 +230,11 @@ if (LLAMA_BLAS)
             if (${LLAMA_BLAS_VENDOR} MATCHES "Generic")
                 pkg_check_modules(DepBLAS REQUIRED blas)
             elseif (${LLAMA_BLAS_VENDOR} MATCHES "OpenBLAS")
-                pkg_check_modules(DepBLAS REQUIRED openblas)
+                # As of openblas v0.3.22, the 64-bit is named openblas64.pc
+                pkg_check_modules(DepBLAS openblas64)
+                if (NOT DepBLAS_FOUND)
+                    pkg_check_modules(DepBLAS REQUIRED openblas)
+                endif()
             elseif (${LLAMA_BLAS_VENDOR} MATCHES "FLAME")
                 pkg_check_modules(DepBLAS REQUIRED blis)
             elseif (${LLAMA_BLAS_VENDOR} MATCHES "ATLAS")
@@ -301,6 +336,8 @@ if (LLAMA_CUBLAS)
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
         endif()
 
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cuda_driver)
+
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
         # 52 == lowest CUDA 12 standard
         # 60 == f16 CUDA intrinsics
@@ -377,6 +414,9 @@ if (LLAMA_HIPBLAS)
     if (${hipblas_FOUND} AND ${hip_FOUND})
         message(STATUS "HIP and hipBLAS found")
         add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        if (LLAMA_HIP_UMA)
+            add_compile_definitions(GGML_HIP_UMA)
+        endif()
         add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
         if (BUILD_SHARED_LIBS)
             set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)

Makefile

@@ -1,8 +1,8 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
-	main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
+	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
 	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
-	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead tests/test-c.o
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
@@ -65,7 +65,7 @@ test: $(TEST_TARGETS)
 			./$$test_target; \
 		fi; \
 		if [ $$? -ne 0 ]; then \
-			printf 'Test $$test_target FAILED!\n\n' $$test_target; \
+			printf 'Test %s FAILED!\n\n' $$test_target; \
 			failures=$$(( failures + 1 )); \
 		else \
 			printf 'Test %s passed.\n\n' $$test_target; \
@@ -282,8 +282,17 @@ endif
 ifneq ($(filter aarch64%,$(UNAME_M)),)
 	# Apple M1, M2, etc.
 	# Raspberry Pi 3, 4, Zero 2 (64-bit)
+	# Nvidia Jetson
 	MK_CFLAGS   += -mcpu=native
 	MK_CXXFLAGS += -mcpu=native
+	JETSON_RELEASE_INFO = $(shell jetson_release)
+	ifdef JETSON_RELEASE_INFO
+		ifneq ($(filter TX2%,$(JETSON_RELEASE_INFO)),)
+			JETSON_EOL_MODULE_DETECT = 1
+			CC = aarch64-unknown-linux-gnu-gcc
+			cxx = aarch64-unknown-linux-gnu-g++
+		endif
+	endif
 endif
 
 ifneq ($(filter armv6%,$(UNAME_M)),)
@@ -357,15 +366,16 @@ ifdef LLAMA_BLIS
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
-	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
-	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
+	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include -I/usr/local/cuda/targets/aarch64-linux/include
+	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
-	MK_NVCCFLAGS  = --forward-unknown-to-host-compiler -use_fast_math
-
+	MK_NVCCFLAGS  = -use_fast_math
+ifndef JETSON_EOL_MODULE_DETECT
+	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
+endif # JETSON_EOL_MODULE_DETECT
 ifdef LLAMA_DEBUG
 	MK_NVCCFLAGS += -lineinfo
-endif
-
+endif # LLAMA_DEBUG
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(LLAMA_CUDA_NVCC)
 else
@@ -417,7 +427,11 @@ ifdef LLAMA_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+ifdef JETSON_EOL_MODULE_DETECT
+	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+else
 	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+endif # JETSON_EOL_MODULE_DETECT
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
@@ -452,6 +466,9 @@ ifdef LLAMA_HIPBLAS
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
 	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+ifdef LLAMA_HIP_UMA
+	MK_CPPFLAGS += -DGGML_HIP_UMA
+endif # LLAMA_HIP_UMA
 	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
 	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
@@ -597,6 +614,9 @@ quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.
 perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+imatrix: examples/imatrix/imatrix.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
@@ -606,7 +626,7 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
 
-gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
+gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
@@ -645,6 +665,12 @@ parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 ifdef LLAMA_METAL
 metal: examples/metal/metal.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)

Package.swift

@@ -13,21 +13,17 @@ let package = Package(
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
+    dependencies: [
+        .package(url: "https://github.com/ggerganov/ggml.git", .branch("release"))
+    ],
     targets: [
         .target(
             name: "llama",
+            dependencies: ["ggml"],
             path: ".",
-            exclude: [],
+            exclude: ["ggml-metal.metal"],
             sources: [
-                "ggml.c",
                 "llama.cpp",
-                "ggml-alloc.c",
-                "ggml-backend.c",
-                "ggml-quants.c",
-                "ggml-metal.m",
-            ],
-            resources: [
-                .process("ggml-metal.metal")
             ],
             publicHeadersPath: "spm-headers",
             cSettings: [

README.md

@@ -10,6 +10,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
+- New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow
 - Collecting Apple Silicon performance stats:
   - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
   - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
@@ -102,6 +103,8 @@ as the main playground for developing new features for the [ggml](https://github
 - [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek)
 - [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
 - [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
+- [x] [PLaMo-13B](https://github.com/ggerganov/llama.cpp/pull/3557)
+- [x] [GPT-2](https://huggingface.co/gpt2)
 
 **Multimodal models:**
 
@@ -116,6 +119,7 @@ as the main playground for developing new features for the [ggml](https://github
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
+- JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
@@ -123,6 +127,7 @@ as the main playground for developing new features for the [ggml](https://github
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
 - React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
 - Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
+- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
 
 **UI:**
 
@@ -131,6 +136,8 @@ as the main playground for developing new features for the [ggml](https://github
 - [withcatai/catai](https://github.com/withcatai/catai)
 - [semperai/amica](https://github.com/semperai/amica)
 - [psugihara/FreeChat](https://github.com/psugihara/FreeChat)
+- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
+- [iohub/collama](https://github.com/iohub/coLLaMA)
 
 ---
 
@@ -381,20 +388,37 @@ Building the program with BLAS support may lead to some performance improvements
 
   Check [BLIS.md](docs/BLIS.md) for more information.
 
-- #### Intel MKL
+- #### Intel oneMKL
+  - Using manual oneAPI installation:
+    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. Otherwise please install oneAPI and follow the below steps:
+      ```bash
+      mkdir build
+      cd build
+      source /opt/intel/oneapi/setvars.sh # You can skip this step if  in oneapi-runtime docker image, only required for manual installation
+      cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON
+      cmake --build . --config Release
+      ```
 
-  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:
+  - Using oneAPI docker image:
+    If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-runtime](https://hub.docker.com/r/intel/oneapi-runtime)
 
-  ```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
-  ```
+      ```bash
+      mkdir build
+      cd build
+      cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON
+      cmake --build . --config Release
+      ```
+
+  Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni.
+
+  Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
 
 - #### cuBLAS
 
   This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
+
+  For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
+
   - Using `make`:
     ```bash
     make LLAMA_CUBLAS=1
@@ -432,14 +456,21 @@ Building the program with BLAS support may lead to some performance improvements
     ```bash
     make LLAMA_HIPBLAS=1
     ```
-  - Using `CMake` for Linux:
+  - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
     ```bash
-    mkdir build
-    cd build
-    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON
-    cmake --build .
+    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ \
+        cmake -H. -Bbuild -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+        && cmake --build build -- -j 16
     ```
-  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS):
+    On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON"`.
+    However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
+
+  - Using `make` (example for target gfx1030, build with 16 CPU threads):
+    ```bash
+    make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gxf1030
+    ```
+
+  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
     ```bash
     set PATH=%HIP_PATH%\bin;%PATH%
     mkdir build
@@ -448,10 +479,11 @@ Building the program with BLAS support may lead to some performance improvements
     cmake --build .
     ```
     Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
+    Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
 
 
   The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
-  If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3.
+  If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
   The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
 
   | Option                  | Legal values           | Default | Description |
@@ -982,6 +1014,8 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /
 - There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
 - Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
+- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
+- Matrix multiplication is unconventional: [`z = ggml_mul_mat(ctx, x, y)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means `zT = x @ yT`
 
 ### Docs
 

awq-py/README.md

@@ -0,0 +1,116 @@
+# AWQ: Activation-aware Weight Quantization for LLM - version apply to llamacpp
+[[Paper](https://arxiv.org/abs/2306.00978)][[Original Repo](https://github.com/mit-han-lab/llm-awq)][[Easy-to-use Repo](https://github.com/casper-hansen/AutoAWQ)]
+
+**Supported models:**
+
+- [X] LLaMA
+- [x] LLaMA 2
+- [X] MPT
+- [X] Mistral AI v0.1
+- [ ] Bloom
+- [ ] Mixtral MoE
+
+**TODO:**
+- [x] Update version work with both MPT and MPT-AWQ model
+- [ ] Add OPT model
+- [ ] Add Bloom model
+- [ ] Add Mixtral MoE
+- [ ] Support w3, w2
+
+
+## Contents
+
+- [Install](##Install)
+- [Convert](##Convert)
+- [Quantize](##Quantize)
+- [Test](##Test)
+- [Benchmark](##Benchmark)
+- [Results](##Results)
+
+## Install
+Install requirements
+```bash
+pip install -r requirements.txt
+```
+Get the pre-computed AWQ search results for multiple model families, including LLaMA, LLaMA2, MPT, OPT
+```bash
+git clone https://huggingface.co/datasets/mit-han-lab/awq-model-zoo awq_cache
+```
+
+## Convert
+Example for llama model
+```bash
+# For llama7b and llama2 models
+python convert.py models/llama-7b/ --awq-path awq_cache/llama-7b-w4-g128.pt --outfile models/llama_7b_fp16.gguf
+# For mistral and mpt models
+python convert-hf-to-gguf.py models/mpt-7b/ --awq-path awq_cache/llama-7b-w4-g128.pt --outfile models/mpt_7b_fp16.gguf
+```
+
+## Quantize
+```bash
+# We only benchmark and confirm the results on q4_0, q4_1, and q2_k types.
+./quantize models/llama_7b_fp16.gguf models/llama_7b_q4_0.gguf q4_0
+```
+
+## Test
+```bash
+# For all models.
+./build/bin/main -m models/llama_7b_q4_0.gguf -n 128 --prompt "Once upon a time"
+```
+
+## Benchmark
+The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
+```bash
+# For llama and llama2, and mistral models.
+./perplexity -m models/llama_7b_q4_0.gguf -f datasets/wikitext-2-raw/wiki.test.raw
+```
+
+## Results
+Results are run on OpenBLAS (CPU) and CuBLAS (GPU) for fair comparison
+We use three types of llamacpp quantization methods to work with our version, including q4_0, q4_1, and q2_k
+
+### Llama 7B (Build with OpenBLAS)
+
+| Model      | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|-----------:|--------------|-------:|-------:|-------:|-------:|
+|Llama 7B    | perplexity   | 5.9066 | 6.1214 | 6.0643 | 6.5808 |
+|Llama 7B    | file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|Llama 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-LLama 7B| perplexity   | 5.9175 | 6.0252 | 5.9987 | 6.3692 |
+|AWQ-LLama 7B| file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|AWQ-LLama 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+
+### Llama2 7B (Build with CuBLAS)
+
+| Model       | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|------------:|--------------|-------:|-------:|-------:|-------:|
+|Llama2 7B    | perplexity   | 5.8664 | 6.0260 | 6.0656 | 6.4496 |
+|Llama2 7B    | file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|Llama2 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-LLama2 7B| perplexity   | 5.8801 | 6.0054 | 5.9849 | 6.3650 |
+|AWQ-LLama2 7B| file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|AWQ-LLama2 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+
+### Mistral 7B v0.1 (Build with CuBLAS)
+
+| Model        | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|-------------:|--------------|-------:|-------:|-------:|-------:|
+|Mistral 7B    | perplexity   | 5.6931 | 5.8202 | 5.8268 | 6.1645 |
+|Mistral 7B    | file size     |  14.5G |   4.1G |   4.5G |   3.1G |
+|Mistral 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-Mistral 7B| perplexity   | 5.6934 | 5.8020 | 5.7691 | 6.0426 |
+|AWQ-Mistral 7B| file size     |  14.5G |   4.1G |   4.5G |   3.1G |
+|AWQ-Mistral 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+### MPT 7B (Build with OpenBLAS)
+
+| Model    | Measure      | F16    | Q4_0   | Q4_1   | Q2_K    |
+|---------:|--------------|-------:|-------:|-------:|--------:|
+|MPT 7B    | perplexity   | 8.4369 | 8.7956 | 8.6265 | 11.4913 |
+|MPT 7B    | file size    |  13.7G  |   3.9G |   4.3G |   2.8G  |
+|MPT 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6  |
+|AWQ-MPT 7B| perplexity   | 8.4944 | 8.7053 |  8.6750 | 10.2873|
+|AWQ-MPT 7B| file size    |  13.7G  |   3.9G |   4.3G |   2.8G  |
+|AWQ-MPT 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6  |

awq-py/awq/apply_awq.py

@@ -0,0 +1,254 @@
+"""
+Implements the AWQ for llama.cpp use cases.
+Original paper: https://arxiv.org/abs/2306.00978
+
+This code is based on versions of the AWQ implementation found in the following repositories:
+* https://github.com/mit-han-lab/llm-awq
+* https://github.com/casper-hansen/AutoAWQ
+"""
+
+import os
+import torch
+import torch.nn as nn
+
+from transformers import AutoModelForCausalLM, AutoConfig
+from transformers.models.bloom.modeling_bloom import BloomGelu
+from transformers.models.llama.modeling_llama import LlamaRMSNorm
+from transformers.activations import GELUActivation
+
+
+class ScaledActivation(nn.Module):
+    """
+    ScaledActivation module wraps an existing activation function and applies a
+    scale factor to its output.
+
+    Args:
+        module (nn.Module): The activation function to be scaled.
+        scales (torch.Tensor): A tensor of size (num_features,) containing the initial
+            scale factors for each feature.
+
+    Returns:
+        torch.Tensor: The scaled output of the activation function.
+    """
+
+    def __init__(self, module, scales):
+        super().__init__()
+        self.act = module
+        self.scales = nn.Parameter(scales.data)
+
+    def forward(self, x):
+        return self.act(x) / self.scales.view(1, 1, -1).to(x.device)
+
+
+def set_op_by_name(layer, name, new_module):
+    """
+    Set the new module for given module's name.
+
+    Args:
+        layer (nn.Module): The layer in which to replace the submodule.
+        name (str): The path to the submodule to be replaced, using dot notation
+            to access nested modules.
+        new_module (nn.Module): The new module to replace the existing one.
+    """
+    levels = name.split(".")
+    if len(levels) > 1:
+        mod_ = layer
+        for l_idx in range(len(levels) - 1):
+            if levels[l_idx].isdigit():
+                mod_ = mod_[int(levels[l_idx])]
+            else:
+                mod_ = getattr(mod_, levels[l_idx])
+        setattr(mod_, levels[-1], new_module)
+    else:
+        setattr(layer, name, new_module)
+
+
+def get_op_by_name(module, op_name):
+    """
+    Retrieves a submodule within a given layer based on its name.
+
+    Args:
+        module (nn.Module): The layer containing the submodule to find.
+        op_name (str): The name of the submodule.
+
+    Returns:
+        nn.Module: The requested submodule found within the given layer.
+
+    Raises:
+        ValueError: If the specified submodule cannot be found within the layer.
+    """
+    for name, m in module.named_modules():
+        if name == op_name:
+            return m
+    raise ValueError(f"Cannot find op {op_name} in module {module}")
+
+
+@torch.no_grad()
+def scale_ln_fcs(ln, fcs, scales):
+    """
+    Scales the weights of a LayerNorm and a list of fully-connected layers proportionally.
+
+    Args:
+        ln (nn.LayerNorm): The LayerNorm module to be scaled.
+        fcs (List[nn.Linear]): A list of fully-connected layers to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+    """
+
+    if not isinstance(fcs, list):
+        fcs = [fcs]
+
+    scales = scales.to(ln.weight.device)
+
+    ln.weight.div_(scales)
+    if hasattr(ln, "bias") and ln.bias is not None:
+        ln.bias.div_(scales)
+
+    for fc in fcs:
+        fc.weight.mul_(scales.view(1, -1))
+
+    for p in ln.parameters():
+        assert torch.isnan(p).sum() == 0
+    for fc in fcs:
+        for p in fc.parameters():
+            assert torch.isnan(p).sum() == 0
+
+
+@torch.no_grad()
+def scale_fc_fc(fc1, fc2, scales):
+    """
+    Scales the weights of two fully-connected layers in a specific pattern.
+
+    Args:
+        fc1 (nn.Linear): The first fully-connected layer to be scaled.
+        fc2 (nn.Linear): The second fully-connected layer to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+    """
+    assert isinstance(fc1, nn.Linear)
+    assert isinstance(fc2, nn.Linear)
+
+    scales = scales.to(fc1.weight.device)
+
+    fc1.weight[-scales.size(0):].div_(scales.view(-1, 1))
+    if fc1.bias is not None:
+        fc1.bias.div_(scales.view(-1))
+
+    fc2.weight.mul_(scales.view(1, -1))
+
+    for p in fc1.parameters():
+        assert torch.isnan(p).sum() == 0
+    for p in fc2.parameters():
+        assert torch.isnan(p).sum() == 0
+
+
+@torch.no_grad()
+def scale_gelu_fc(gelu, fc, scales):
+    """
+    Scales the weight of a GELU activation and a fully-connected layer proportionally.
+
+    Args:
+        gelu (Union[nn.GELU, BloomGelu, GELUActivation]): The GELU activation module to be scaled.
+        fc (nn.Linear): The fully-connected layer to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+
+    Raises:
+        TypeError: If the `gelu` module is not of type `nn.GELU`, `BloomGelu`, or `GELUActivation`.
+        TypeError: If the `fc` module is not of type `nn.Linear`.
+    """
+    assert isinstance(gelu, (nn.GELU, BloomGelu, GELUActivation))
+    assert isinstance(fc, nn.Linear)
+
+    fc.weight.mul_(scales.view(1, -1).to(fc.weight.device))
+
+    for p in fc.parameters():
+        assert torch.isnan(p).sum() == 0
+
+
+def apply_scale(module, scales_list, input_feat_dict=None):
+    """
+    Applies different scaling strategies to layers based on their type and hierarchy within a given module.
+
+    Args:
+        module (nn.Module): The module containing the layers to be scaled.
+        scales_list (List[Tuple[str, List[str], torch.Tensor]]): A list of tuples containing:
+            * prev_op_name (str): The name of the preceding operation or module,
+                relative to which the layers to be scaled are located.
+            * layer_names (List[str]): A list of names of the layers to be scaled, relative to the preceding operation.
+            * scales (torch.Tensor): A 1D tensor of size (num_features,) containing the scaling factors for each feature.
+        input_feat_dict (Optional[Dict[str, torch.Tensor]]): A dictionary mapping layer names to their corresponding
+            input features (optional).
+    """
+    for prev_op_name, layer_names, scales in scales_list:
+        prev_op = get_op_by_name(module, prev_op_name)
+        layers = [get_op_by_name(module, name) for name in layer_names]
+
+        prev_op.cuda()
+        for layer in layers:
+            layer.cuda()
+        scales.cuda()
+
+        if isinstance(prev_op, nn.Linear):
+            assert len(layers) == 1
+            scale_fc_fc(prev_op, layers[0], scales)
+        elif isinstance(prev_op, (nn.LayerNorm, LlamaRMSNorm)) or "rmsnorm" in str(prev_op.__class__).lower():
+            scale_ln_fcs(prev_op, layers, scales)
+        elif isinstance(prev_op, (nn.GELU, BloomGelu, GELUActivation)):
+            new_module = ScaledActivation(prev_op, scales)
+            set_op_by_name(module, prev_op_name, new_module)
+            scale_gelu_fc(prev_op, layers[0], scales)
+        else:
+            raise NotImplementedError(f"prev_op {type(prev_op)} not supported yet!")
+
+        # apply the scaling to input feat if given; prepare it for clipping
+        if input_feat_dict is not None:
+            for layer_name in layer_names:
+                inp = input_feat_dict[layer_name]
+                inp.div_(scales.view(1, -1).to(inp.device))
+
+        prev_op.cpu()
+        for layer in layers:
+            layer.cpu()
+        scales.cpu()
+
+
+@torch.no_grad()
+def apply_clip(module, clip_list):
+    """
+    Applies element-wise clipping to the weight of a specific layer within a given module.
+
+    Args:
+        module (nn.Module): The module containing the layer to be clipped.
+        clip_list (List[Tuple[str, torch.Tensor]]): A list of tuples containing:
+            * name (str): The name of the layer to be clipped, relative to the root of the module.
+            * max_val (torch.Tensor): A 1D or 2D tensor defining the upper bound for each element of the layer's weight.
+    """
+    for name, max_val in clip_list:
+        layer = get_op_by_name(module, name)
+        layer.cuda()
+        max_val = max_val.to(layer.weight.device)
+        org_shape = layer.weight.shape
+        layer.weight.data = layer.weight.data.reshape(*max_val.shape[:2], -1)
+        layer.weight.data = torch.clamp(layer.weight.data, -max_val, max_val)
+        layer.weight.data = layer.weight.data.reshape(org_shape)
+        layer.cpu()
+
+
+def add_scale_weights(model_path, scale_path, tmp_path):
+    """
+    Adds pre-computed Activation Weight Quantization (AWQ) results to a model,
+    including scaling factors and clipping bounds.
+
+    Args:
+        model_path (str): Path to the pre-trained model to be equipped with AWQ.
+        scale_path (str): Path to the AWQ scale factors (.pt file).
+        tmp_path (str): Path to the temporary directory where the equipped model will be saved.
+    """
+    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(
+        model_path, config=config, trust_remote_code=True
+    )
+    model.eval()
+    awq_results = torch.load(str(scale_path), map_location="cpu")
+    apply_scale(model, awq_results["scale"])
+    apply_clip(model, awq_results["clip"])
+    model.save_pretrained(str(tmp_path))
+    os.system(f"cp {str(model_path)}/tokenizer* {str(tmp_path)}")

awq-py/requirements.txt

@@ -0,0 +1,2 @@
+torch>=2.1.1
+transformers>=4.32.0

ci/run.sh

@@ -30,6 +30,12 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
+CMAKE_EXTRA=""
+
+if [ ! -z ${GG_BUILD_METAL} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
+fi
+
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -81,8 +87,8 @@ function gg_run_ctest_debug {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Debug ..     ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                          ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
@@ -109,8 +115,8 @@ function gg_run_ctest_release {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ..   ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
         (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log

common/CMakeLists.txt

@@ -65,4 +65,4 @@ endif()
 
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE llama build_info)
+target_link_libraries(${TARGET} PRIVATE build_info PUBLIC llama)

common/common.cpp

@@ -220,6 +220,20 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_ctx = std::stoi(argv[i]);
+        } else if (arg == "--grp-attn-n" || arg == "-gan") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+
+            params.grp_attn_n = std::stoi(argv[i]);
+        } else if (arg == "--grp-attn-w" || arg == "-gaw") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+
+            params.grp_attn_w = std::stoi(argv[i]);
         } else if (arg == "--rope-freq-base") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -529,9 +543,8 @@ bool gpt_params_parse_ex(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
+#ifndef LLAMA_SUPPORTS_GPU_OFFLOAD
             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
@@ -540,9 +553,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
             params.n_gpu_layers_draft = std::stoi(argv[i]);
-#else
+#ifndef LLAMA_SUPPORTS_GPU_OFFLOAD
             fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers-draft option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
 #endif
@@ -551,25 +563,44 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifdef GGML_USE_CUBLAS
             params.main_gpu = std::stoi(argv[i]);
-#else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
-#endif
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the main GPU has no effect.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--split-mode" || arg == "-sm") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string arg_next = argv[i];
+            if (arg_next == "none") {
+                params.split_mode = LLAMA_SPLIT_NONE;
+            } else if (arg_next == "layer") {
+                params.split_mode = LLAMA_SPLIT_LAYER;
+            } else if (arg_next == "row") {
+                params.split_mode = LLAMA_SPLIT_ROW;
+            } else {
+                invalid_param = true;
+                break;
+            }
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n");
+#endif // GGML_USE_CUBLAS
         } else if (arg == "--tensor-split" || arg == "-ts") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-#ifdef GGML_USE_CUBLAS
             std::string arg_next = argv[i];
 
             // split string by , and /
             const std::regex regex{R"([,/]+)"};
             std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
             std::vector<std::string> split_arg{it, {}};
-            GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
-
+            if (split_arg.size() >= LLAMA_MAX_DEVICES) {
+                invalid_param = true;
+                break;
+            }
             for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
                 if (i < split_arg.size()) {
                     params.tensor_split[i] = std::stof(split_arg[i]);
@@ -577,14 +608,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                     params.tensor_split[i] = 0.0f;
                 }
             }
-#else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
-#endif // GGML_USE_CUBLAS
-        } else if (arg == "--no-mul-mat-q" || arg == "-nommq") {
-#ifdef GGML_USE_CUBLAS
-            params.mul_mat_q = false;
-#else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting a tensor split has no effect.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
@@ -616,6 +641,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.ppl_stride = std::stoi(argv[i]);
+        } else if (arg == "-ptc" || arg == "--print-token-count") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_print = std::stoi(argv[i]);
         } else if (arg == "--ppl-output-type") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -798,7 +829,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("\n");
     printf("options:\n");
     printf("  -h, --help            show this help message and exit\n");
-    printf("      --version         show version and build info\n");
+    printf("  --version             show version and build info\n");
     printf("  -i, --interactive     run in interactive mode\n");
     printf("  --interactive-first   run in interactive mode and wait for input right away\n");
     printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
@@ -895,15 +926,20 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        number of layers to store in VRAM\n");
     printf("  -ngld N, --n-gpu-layers-draft N\n");
     printf("                        number of layers to store in VRAM for the draft model\n");
-    printf("  -ts SPLIT --tensor-split SPLIT\n");
-    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-#ifdef GGML_USE_CUBLAS
-    printf("  -nommq, --no-mul-mat-q\n");
-    printf("                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
-    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
-#endif // GGML_USE_CUBLAS
+    printf("  -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
+    printf("                        how to split the model across multiple GPUs, one of:\n");
+    printf("                          - none: use one GPU only\n");
+    printf("                          - layer (default): split layers and KV across GPUs\n");
+    printf("                          - row: split rows across GPUs\n");
+    printf("  -ts SPLIT, --tensor-split SPLIT\n");
+    printf("                        fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for the model (with split-mode = none),\n");
+    printf("                        or for intermediate results and KV (with split-mode = row) (default: %d)\n", params.main_gpu);
 #endif
+    printf("  -gan N, --grp-attn-n N\n");
+    printf("                        group-attention factor (default: %d)\n", params.grp_attn_n);
+    printf("  -gaw N, --grp-attn-w N\n");
+    printf("                        group-attention width (default: %.1f)\n", (double)params.grp_attn_w);
     printf("  --verbose-prompt      print prompt before generation\n");
     printf("  -dkvc, --dump-kv-cache\n");
     printf("                        verbose print of the KV cache\n");
@@ -920,12 +956,14 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -m FNAME, --model FNAME\n");
     printf("                        model path (default: %s)\n", params.model.c_str());
     printf("  -md FNAME, --model-draft FNAME\n");
-    printf("                        draft model for speculative decoding (default: %s)\n", params.model.c_str());
+    printf("                        draft model for speculative decoding\n");
     printf("  -ld LOGDIR, --logdir LOGDIR\n");
     printf("                        path under which to save YAML logs (no logging if unset)\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("                        advanced option to override model metadata by key. may be specified multiple times.\n");
     printf("                        types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
+    printf("  -ptc N, --print-token-count N\n");
+    printf("                        print token count every N tokens (default: %d)\n", params.n_print);
     printf("\n");
 #ifndef LOG_DISABLE_LOGS
     log_print_usage();
@@ -1015,6 +1053,7 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
     mparams.main_gpu        = params.main_gpu;
+    mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
@@ -1029,6 +1068,9 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
 }
 
 static ggml_type kv_cache_type_from_str(const std::string & s) {
+    if (s == "f32") {
+        return GGML_TYPE_F32;
+    }
     if (s == "f16") {
         return GGML_TYPE_F16;
     }
@@ -1394,6 +1436,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "build_number: %d\n",        LLAMA_BUILD_NUMBER);
     fprintf(stream, "cpu_has_arm_fma: %s\n",     ggml_cpu_has_arm_fma()     ? "true" : "false");
     fprintf(stream, "cpu_has_avx: %s\n",         ggml_cpu_has_avx()         ? "true" : "false");
+    fprintf(stream, "cpu_has_avx_vnni: %s\n",    ggml_cpu_has_avx_vnni()    ? "true" : "false");
     fprintf(stream, "cpu_has_avx2: %s\n",        ggml_cpu_has_avx2()        ? "true" : "false");
     fprintf(stream, "cpu_has_avx512: %s\n",      ggml_cpu_has_avx512()      ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");

common/common.h

@@ -51,7 +51,7 @@ struct gpt_params {
     int32_t n_ctx                           = 512;   // context size
     int32_t n_batch                         = 512;   // batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                          = 0;     // number of tokens to keep from initial prompt
-    int32_t n_draft                         = 16;    // number of tokens to draft during speculative decoding
+    int32_t n_draft                         = 8;     // number of tokens to draft during speculative decoding
     int32_t n_chunks                        = -1;    // max number of chunks to process (-1 = unlimited)
     int32_t n_parallel                      = 1;     // number of parallel sequences to decode
     int32_t n_sequences                     = 1;     // number of sequences to decode
@@ -59,9 +59,13 @@ struct gpt_params {
     float   p_split                         = 0.1f;  // speculative decoding split probability
     int32_t n_gpu_layers                    = -1;    // number of layers to store in VRAM (-1 - use default)
     int32_t n_gpu_layers_draft              = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
+    llama_split_mode split_mode             = LLAMA_SPLIT_LAYER; // how to split the model across GPUs
     int32_t main_gpu                        = 0;     // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0};   // how split tensors should be distributed across GPUs
     int32_t n_beams                         = 0;     // if non-zero then use beam search of given width.
+    int32_t grp_attn_n                      = 1;     // group-attention factor
+    int32_t grp_attn_w                      = 512;   // group-attention width
+    int32_t n_print                         = -1;    // print token count every n tokens (-1 = disabled)
     float   rope_freq_base                  = 0.0f;  // RoPE base frequency
     float   rope_freq_scale                 = 0.0f;  // RoPE frequency scaling factor
     float   yarn_ext_factor                 = -1.0f; // YaRN extrapolation mix factor

common/sampling.cpp

@@ -149,11 +149,12 @@ static void sampler_queue(
     }
 }
 
-llama_token llama_sampling_sample(
+static llama_token llama_sampling_sample_impl(
                   struct llama_sampling_context * ctx_sampling,
                   struct llama_context * ctx_main,
                   struct llama_context * ctx_cfg,
-                  const int idx) {
+                  const int idx,
+                  bool is_resampling) {  // Add a parameter to indicate if we are resampling
     const llama_sampling_params & params = ctx_sampling->params;
 
     const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
@@ -173,8 +174,17 @@ llama_token llama_sampling_sample(
 
     llama_token id = 0;
 
+    // Get a pointer to the logits
     float * logits = llama_get_logits_ith(ctx_main, idx);
 
+    // Declare original_logits at the beginning of the function scope
+    std::vector<float> original_logits;
+
+    if (!is_resampling) {
+        // Only make a copy of the original logits if we are not in the resampling phase, not sure if I actually have to do this.
+        original_logits = std::vector<float>(logits, logits + llama_n_vocab(llama_get_model(ctx_main)));
+    }
+
     // apply params.logit_bias map
     for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
         logits[it->first] += it->second;
@@ -193,12 +203,14 @@ llama_token llama_sampling_sample(
     }
 
     // apply penalties
-    if (!prev.empty()) {
+    const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
+    const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);
+    if (penalty_tokens_used_size) {
         const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
 
         llama_sample_repetition_penalties(ctx_main, &cur_p,
-                prev.data() + prev.size() - penalty_last_n,
-                penalty_last_n, penalty_repeat, penalty_freq, penalty_present);
+                penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size,
+                penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present);
 
         if (!penalize_nl) {
             for (size_t idx = 0; idx < cur_p.size; idx++) {
@@ -210,7 +222,8 @@ llama_token llama_sampling_sample(
         }
     }
 
-    if (ctx_sampling->grammar != NULL) {
+    // If we are in the resampling phase, apply grammar checks before sampling logic
+    if (is_resampling && ctx_sampling->grammar != NULL) {
         llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
     }
 
@@ -252,9 +265,40 @@ llama_token llama_sampling_sample(
         }
     }
 
+    if (ctx_sampling->grammar != NULL && !is_resampling) {
+        // Create an array with a single token data element for the sampled id
+        llama_token_data single_token_data = {id, logits[id], 0.0f};
+        llama_token_data_array single_token_data_array = { &single_token_data, 1, false };
+
+        // Apply grammar constraints to the single token
+        llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar);
+
+        // Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY
+        bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
+
+        // If the token is not valid according to the grammar, perform resampling
+        if (!is_valid) {
+            LOG("Resampling because token %d: '%s' does not meet grammar rules\n", id, llama_token_to_piece(ctx_main, id).c_str());
+
+            // Restore logits from the copy
+            std::copy(original_logits.begin(), original_logits.end(), logits);
+
+            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, true);  // Pass true for is_resampling
+        }
+    }
+
     return id;
 }
 
+llama_token llama_sampling_sample(
+                  struct llama_sampling_context * ctx_sampling,
+                  struct llama_context * ctx_main,
+                  struct llama_context * ctx_cfg,
+                  const int idx) {
+    // Call the implementation function with is_resampling set to false by default
+    return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
+}
+
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,

common/sampling.h

@@ -36,6 +36,9 @@ typedef struct llama_sampling_params {
     float       cfg_scale     = 1.f; // how strong is guidance
 
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
+
+    std::vector<llama_token> penalty_prompt_tokens;
+    bool                     use_penalty_prompt_tokens = false;
 } llama_sampling_params;
 
 // general sampler context

common/train.cpp

@@ -1107,7 +1107,7 @@ void print_common_train_usage(int /*argc*/, char ** /*argv*/, const struct train
     fprintf(stderr, "  --sample-start STR         Sets the starting point for samples after the specified pattern. If empty use every token position as sample start. (default '%s')\n", params->sample_start.c_str());
     fprintf(stderr, "  --include-sample-start     Include the sample start in the samples. (default off)\n");
     fprintf(stderr, "  --escape                   process sample start escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
-    fprintf(stderr, "  --overlapping-samples      Samples my overlap, will include sample-start of second and following samples. When off, samples will end at begin of next sample. (default off)\n");
+    fprintf(stderr, "  --overlapping-samples      Samples may overlap, will include sample-start of second and following samples. When off, samples will end at begin of next sample. (default off)\n");
     fprintf(stderr, "  --fill-with-next-samples   Samples shorter than context length will be followed by the next (shuffled) samples. (default off)\n");
     fprintf(stderr, "  --separate-with-eos        When fill-with-next-samples, insert end-of-sequence token between samples.%s\n", params->separate_with_eos ? " (default)" : "");
     fprintf(stderr, "  --separate-with-bos        When fill-with-next-samples, insert begin-of-sequence token between samples.%s\n", params->separate_with_bos ? " (default)" : "");

convert-hf-to-gguf.py

@@ -23,6 +23,15 @@ if 'NO_LOCAL_GGUF' not in os.environ:
 import gguf
 
 
+# check for any of the given keys in the dictionary and return the value of the first key found
+def get_key_opts(d, keys):
+    for k in keys:
+        if k in d:
+            return d[k]
+    print(f"Could not find any of {keys}")
+    sys.exit()
+
+
 ###### MODEL DEFINITIONS ######
 
 class SentencePieceTokenTypes(IntEnum):
@@ -46,7 +55,7 @@ class Model:
         self.part_names = self._get_part_names()
         self.hparams = Model.load_hparams(self.dir_model)
         self.model_arch = self._get_model_architecture()
-        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess)
+        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
 
     def set_vocab(self):
         self._set_vocab_gpt2()
@@ -182,8 +191,12 @@ class Model:
             return QwenModel
         if model_architecture == "MixtralForCausalLM":
             return MixtralModel
+        if model_architecture == "GPT2LMHeadModel":
+            return GPT2Model
         if model_architecture == "PhiForCausalLM":
             return Phi2Model
+        if model_architecture == "PlamoForCausalLM":
+            return PlamoModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -223,8 +236,12 @@ class Model:
             return gguf.MODEL_ARCH.QWEN
         if arch == "MixtralForCausalLM":
             return gguf.MODEL_ARCH.LLAMA
+        if arch == "GPT2LMHeadModel":
+            return gguf.MODEL_ARCH.GPT2
         if arch == "PhiForCausalLM":
             return gguf.MODEL_ARCH.PHI2
+        if arch == "PlamoForCausalLM":
+            return gguf.MODEL_ARCH.PLAMO
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -234,7 +251,7 @@ class Model:
         tokens: list[bytearray] = []
         toktypes: list[int] = []
 
-        from transformers import AutoTokenizer  # type: ignore[attr-defined]
+        from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(dir_model)
         vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
         assert max(tokenizer.vocab.values()) < vocab_size
@@ -249,10 +266,11 @@ class Model:
                 toktypes.append(gguf.TokenType.USER_DEFINED)
             elif reverse_vocab[i] in added_vocab:
                 tokens.append(reverse_vocab[i])
-                if tokenizer.added_tokens_decoder[i].special:
-                    toktypes.append(gguf.TokenType.CONTROL)
-                else:
-                    toktypes.append(gguf.TokenType.USER_DEFINED)
+                if hasattr(tokenizer, "added_tokens_decoder"):
+                    if tokenizer.added_tokens_decoder[i].special:
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
             else:
                 tokens.append(reverse_vocab[i])
                 toktypes.append(gguf.TokenType.NORMAL)
@@ -460,7 +478,11 @@ class MPTModel(Model):
             data = data_torch.squeeze().numpy()
 
             # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if "scales" in name:
+                new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias", ".scales"))
+                new_name = new_name.replace("scales", "act.scales")
+            else:
+                new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
             if new_name is None:
                 print(f"Can not map tensor {name!r}")
                 sys.exit()
@@ -805,10 +827,17 @@ class PersimmonModel(Model):
         hidden_size = self.hparams["hidden_size"]
 
         self.gguf_writer.add_name('persimmon-8b-chat')
+        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hidden_size)
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
-        self.gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+
+        # NOTE: not sure about this change - why does the model not have a rope dimension count when it is smaller
+        #       than the head size?
+        #       ref: https://github.com/ggerganov/llama.cpp/pull/4889
+        # self.gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+        self.gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
+
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
@@ -844,7 +873,7 @@ class StableLMModel(Model):
         hparams = self.hparams
         block_count = hparams["num_hidden_layers"]
 
-        self.gguf_writer.add_name(dir_model.name)
+        self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_block_count(block_count)
@@ -890,7 +919,7 @@ class QwenModel(Model):
         tokens: list[bytearray] = []
         toktypes: list[int] = []
 
-        from transformers import AutoTokenizer  # type: ignore[attr-defined]
+        from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
         vocab_size = hparams["vocab_size"]
         assert max(tokenizer.get_vocab().values()) < vocab_size
@@ -985,32 +1014,182 @@ class QwenModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-class Phi2Model(Model):
+class GPT2Model(Model):
     def set_gguf_parameters(self):
-        block_count = self.hparams["n_layer"]
-
-        self.gguf_writer.add_name("Phi2")
-        self.gguf_writer.add_context_length(self.hparams["n_positions"])
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_context_length(self.hparams["n_ctx"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
-        self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(self.hparams["n_head"])
-        self.gguf_writer.add_head_count_kv(self.hparams["n_head"])
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
-        self.gguf_writer.add_rope_dimension_count(self.hparams["rotary_dim"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def write_tensors(self):
+        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+        for name, data_torch in self.get_tensors():
+            # we don't need these
+            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq", ".attn.bias")):
+                continue
+
+            if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_proj.weight")):
+                data_torch = data_torch.transpose(1, 0)
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+            # note: GPT2 output is tied to (same as) wte in original model
+            if new_name == "token_embd.weight":
+                print(f"output.weight, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+                self.gguf_writer.add_tensor("output.weight", data)
+
+
+class Phi2Model(Model):
+    def set_gguf_parameters(self):
+        block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"])
+
+        rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
+        n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"])
+        n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"])
+
+        self.gguf_writer.add_name("Phi2")
+        self.gguf_writer.add_context_length(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"]))
+
+        self.gguf_writer.add_embedding_length(n_embd)
+        self.gguf_writer.add_feed_forward_length(4 * n_embd)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head)
+        self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"]))
+        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_add_bos_token(False)
 
 
+class PlamoModel(Model):
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
+
+        self.gguf_writer.add_name("PLaMo")
+        self.gguf_writer.add_context_length(4096)  # not in config.json
+        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
+        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(5)  # hparams["num_key_value_heads"]) is wrong
+        self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
+
+    def shuffle_attn_q_weight(self, data_torch):
+        assert data_torch.size() == (5120, 5120)
+        data_torch = data_torch.reshape(8, 5, 128, 5120)
+        data_torch = torch.permute(data_torch, (1, 0, 2, 3))
+        data_torch = torch.reshape(data_torch, (5120, 5120))
+        return data_torch
+
+    def shuffle_attn_output_weight(self, data_torch):
+        assert data_torch.size() == (5120, 5120)
+        data_torch = data_torch.reshape(5120, 8, 5, 128)
+        data_torch = torch.permute(data_torch, (0, 2, 1, 3))
+        data_torch = torch.reshape(data_torch, (5120, 5120))
+        return data_torch
+
+    def write_tensors(self):
+        block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+        for name, data_torch in self.get_tensors():
+            if "self_attn.rotary_emb.inv_freq" in name:
+                continue
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            # shuffle for broadcasting of gqa in ggml_mul_mat
+            if new_name.endswith("attn_q.weight"):
+                data_torch = self.shuffle_attn_q_weight(data_torch)
+            elif new_name.endswith("attn_output.weight"):
+                data_torch = self.shuffle_attn_output_weight(data_torch)
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+
 ###### CONVERSION LOGIC ######
 
 
 def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Convert a huggingface model to a GGML compatible file")
+    parser = argparse.ArgumentParser(
+        description="Convert a huggingface model to a GGML compatible file")
     parser.add_argument(
         "--vocab-only", action="store_true",
         help="extract only the vocab",
     )
+    parser.add_argument(
+        "--awq-path", type=Path, default=None,
+        help="Path to scale awq cache file")
     parser.add_argument(
         "--outfile", type=Path,
         help="path to write to; default: based on input",
@@ -1028,43 +1207,62 @@ def parse_args() -> argparse.Namespace:
     return parser.parse_args()
 
 
-args = parse_args()
+def main() -> None:
+    args = parse_args()
 
-dir_model = args.model
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file=sys.stderr)
-    sys.exit(1)
+    dir_model = args.model
 
-ftype_map = {
-    "f32": gguf.GGMLQuantizationType.F32,
-    "f16": gguf.GGMLQuantizationType.F16,
-}
+    if args.awq_path:
+        sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
+        from awq.apply_awq import add_scale_weights
+        tmp_model_path = args.model / "weighted_model"
+        dir_model = tmp_model_path
+        if tmp_model_path.is_dir():
+            print(f"{tmp_model_path} exists as a weighted model.")
+        else:
+            tmp_model_path.mkdir(parents=True, exist_ok=True)
+            print("Saving new weighted model ...")
+            add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
+            print(f"Saved weighted model at {tmp_model_path}.")
 
-if args.outfile is not None:
-    fname_out = args.outfile
-else:
-    # output in the same directory as the model by default
-    fname_out = dir_model / f'ggml-model-{args.outtype}.gguf'
+    if not dir_model.is_dir():
+        print(f'Error: {args.model} is not a directory', file=sys.stderr)
+        sys.exit(1)
 
-print(f"Loading model: {dir_model.name}")
+    ftype_map = {
+        "f32": gguf.GGMLQuantizationType.F32,
+        "f16": gguf.GGMLQuantizationType.F16,
+    }
 
-hparams = Model.load_hparams(dir_model)
-
-with torch.inference_mode():
-    model_class = Model.from_model_architecture(hparams["architectures"][0])
-    model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
-
-    print("Set model parameters")
-    model_instance.set_gguf_parameters()
-
-    print("Set model tokenizer")
-    model_instance.set_vocab()
-
-    if args.vocab_only:
-        print(f"Exporting model vocab to '{fname_out}'")
-        model_instance.write_vocab()
+    if args.outfile is not None:
+        fname_out = args.outfile
     else:
-        print(f"Exporting model to '{fname_out}'")
-        model_instance.write()
+        # output in the same directory as the model by default
+        fname_out = dir_model / f'ggml-model-{args.outtype}.gguf'
 
-    print(f"Model successfully exported to '{fname_out}'")
+    print(f"Loading model: {dir_model.name}")
+
+    hparams = Model.load_hparams(dir_model)
+
+    with torch.inference_mode():
+        model_class = Model.from_model_architecture(hparams["architectures"][0])
+        model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
+
+        print("Set model parameters")
+        model_instance.set_gguf_parameters()
+
+        print("Set model tokenizer")
+        model_instance.set_vocab()
+
+        if args.vocab_only:
+            print(f"Exporting model vocab to '{fname_out}'")
+            model_instance.write_vocab()
+        else:
+            print(f"Exporting model to '{fname_out}'")
+            model_instance.write()
+
+        print(f"Model successfully exported to '{fname_out}'")
+
+
+if __name__ == '__main__':
+    main()

convert-lora-to-ggml.py

@@ -47,95 +47,96 @@ def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_ty
     fout.seek((fout.tell() + 31) & -32)
 
 
-if len(sys.argv) < 2:
-    print(f"Usage: python {sys.argv[0]} <path> [arch]")
-    print(
-        "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
-    )
-    print(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
-    sys.exit(1)
+if __name__ == '__main__':
+    if len(sys.argv) < 2:
+        print(f"Usage: python {sys.argv[0]} <path> [arch]")
+        print(
+            "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
+        )
+        print(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
+        sys.exit(1)
 
-input_json = os.path.join(sys.argv[1], "adapter_config.json")
-input_model = os.path.join(sys.argv[1], "adapter_model.bin")
-output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
+    input_json = os.path.join(sys.argv[1], "adapter_config.json")
+    input_model = os.path.join(sys.argv[1], "adapter_model.bin")
+    output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
 
-model = torch.load(input_model, map_location="cpu")
-arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
+    model = torch.load(input_model, map_location="cpu")
+    arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
 
-if arch_name not in gguf.MODEL_ARCH_NAMES.values():
-    print(f"Error: unsupported architecture {arch_name}")
-    sys.exit(1)
+    if arch_name not in gguf.MODEL_ARCH_NAMES.values():
+        print(f"Error: unsupported architecture {arch_name}")
+        sys.exit(1)
 
-arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
-name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
+    arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
+    name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
 
-with open(input_json, "r") as f:
-    params = json.load(f)
+    with open(input_json, "r") as f:
+        params = json.load(f)
 
-if params["peft_type"] != "LORA":
-    print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
-    sys.exit(1)
+    if params["peft_type"] != "LORA":
+        print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
+        sys.exit(1)
 
-if params["fan_in_fan_out"] is True:
-    print("Error: param fan_in_fan_out is not supported")
-    sys.exit(1)
+    if params["fan_in_fan_out"] is True:
+        print("Error: param fan_in_fan_out is not supported")
+        sys.exit(1)
 
-if params["bias"] is not None and params["bias"] != "none":
-    print("Error: param bias is not supported")
-    sys.exit(1)
+    if params["bias"] is not None and params["bias"] != "none":
+        print("Error: param bias is not supported")
+        sys.exit(1)
 
-# TODO: these seem to be layers that have been trained but without lora.
-# doesn't seem widely used but eventually should be supported
-if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
-    print("Error: param modules_to_save is not supported")
-    sys.exit(1)
+    # TODO: these seem to be layers that have been trained but without lora.
+    # doesn't seem widely used but eventually should be supported
+    if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
+        print("Error: param modules_to_save is not supported")
+        sys.exit(1)
 
-with open(output_path, "wb") as fout:
-    fout.truncate()
+    with open(output_path, "wb") as fout:
+        fout.truncate()
 
-    write_file_header(fout, params)
-    for k, v in model.items():
-        orig_k = k
-        if k.endswith(".default.weight"):
-            k = k.replace(".default.weight", ".weight")
-        if k in ["llama_proj.weight", "llama_proj.bias"]:
-            continue
-        if k.endswith("lora_A.weight"):
-            if v.dtype != torch.float16 and v.dtype != torch.float32:
+        write_file_header(fout, params)
+        for k, v in model.items():
+            orig_k = k
+            if k.endswith(".default.weight"):
+                k = k.replace(".default.weight", ".weight")
+            if k in ["llama_proj.weight", "llama_proj.bias"]:
+                continue
+            if k.endswith("lora_A.weight"):
+                if v.dtype != torch.float16 and v.dtype != torch.float32:
+                    v = v.float()
+                v = v.T
+            else:
                 v = v.float()
-            v = v.T
-        else:
-            v = v.float()
 
-        t = v.detach().numpy()
+            t = v.detach().numpy()
 
-        prefix = "base_model.model."
-        if k.startswith(prefix):
-            k = k[len(prefix) :]
+            prefix = "base_model.model."
+            if k.startswith(prefix):
+                k = k[len(prefix) :]
 
-        lora_suffixes = (".lora_A.weight", ".lora_B.weight")
-        if k.endswith(lora_suffixes):
-            suffix = k[-len(lora_suffixes[0]):]
-            k = k[: -len(lora_suffixes[0])]
-        else:
-            print(f"Error: unrecognized tensor name {orig_k}")
-            sys.exit(1)
+            lora_suffixes = (".lora_A.weight", ".lora_B.weight")
+            if k.endswith(lora_suffixes):
+                suffix = k[-len(lora_suffixes[0]):]
+                k = k[: -len(lora_suffixes[0])]
+            else:
+                print(f"Error: unrecognized tensor name {orig_k}")
+                sys.exit(1)
 
-        tname = name_map.get_name(k)
-        if tname is None:
-            print(f"Error: could not map tensor name {orig_k}")
-            print(" Note: the arch parameter must be specified if the model is not llama")
-            sys.exit(1)
+            tname = name_map.get_name(k)
+            if tname is None:
+                print(f"Error: could not map tensor name {orig_k}")
+                print(" Note: the arch parameter must be specified if the model is not llama")
+                sys.exit(1)
 
-        if suffix == ".lora_A.weight":
-            tname += ".weight.loraA"
-        elif suffix == ".lora_B.weight":
-            tname += ".weight.loraB"
-        else:
-            assert False
+            if suffix == ".lora_A.weight":
+                tname += ".weight.loraA"
+            elif suffix == ".lora_B.weight":
+                tname += ".weight.loraB"
+            else:
+                assert False
 
-        print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
-        write_tensor_header(fout, tname, t.shape, t.dtype)
-        t.tofile(fout)
+            print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
+            write_tensor_header(fout, tname, t.shape, t.dtype)
+            t.tofile(fout)
 
-print(f"Converted {input_json} and {input_model} to {output_path}")
+    print(f"Converted {input_json} and {input_model} to {output_path}")

convert-persimmon-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import torch
 import os
 from pprint import pprint

examples/CMakeLists.txt

@@ -31,9 +31,12 @@ else()
     add_subdirectory(quantize-stats)
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
+    add_subdirectory(passkey)
     add_subdirectory(speculative)
     add_subdirectory(lookahead)
+    add_subdirectory(lookup)
     add_subdirectory(train-text-from-scratch)
+    add_subdirectory(imatrix)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/baby-llama/baby-llama.cpp

@@ -575,10 +575,7 @@ static struct ggml_tensor * forward(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]
@@ -844,10 +841,7 @@ static struct ggml_tensor * forward_batch(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, n_batch]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
             assert_shape_4d(KQ_scaled, n_past + N, N, n_head, n_batch);
 
             // KQ_masked = mask_past(KQ_scaled)
@@ -1131,10 +1125,7 @@ static struct ggml_tensor * forward_lora(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]

examples/base-translate.sh

@@ -0,0 +1,61 @@
+#!/bin/bash
+#
+# Few-shot translation example.
+# Requires a base model (i.e. no fine-tuned or instruct models).
+#
+# Usage:
+#
+#   cd llama.cpp
+#   make -j
+#
+#   ./examples/base-translate.sh <model-base> "<text>" [extra-main-args]
+#
+
+if [ $# -lt 2 ]; then
+  echo "Usage: ./base-translate.sh <model-base> \"<text>\" [extra-main-args]"
+  exit 1
+fi
+
+eargs=""
+if [ $# -gt 2 ]; then
+  eargs="${@:3}"
+fi
+
+ftmp="__llama.cpp_example_tmp__.txt"
+trap "rm -f $ftmp" EXIT
+
+echo "Translate from English to French:
+
+===
+
+sea otter, peppermint, plush girafe:
+
+sea otter => loutre de mer
+peppermint => menthe poivrée
+plush girafe => girafe peluche
+
+===
+
+violin
+
+violin => violon
+
+===
+
+phone, computer, mouse, keyboard:
+
+phone => téléphone
+computer => ordinateur
+mouse => souris
+keyboard => clavier
+
+===
+" > $ftmp
+
+echo "$2
+" >> $ftmp
+
+model=$1
+
+# generate the most likely continuation until the string "===" is found
+./main -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs

examples/batched-bench/batched-bench.cpp

@@ -88,7 +88,10 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = llama_model_default_params();
 
+    const std::vector<float> t_split (LLAMA_MAX_DEVICES, 0.0f);
+
     model_params.n_gpu_layers = n_gpu_layers;
+    model_params.tensor_split = t_split.data();
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 

examples/batched/batched.cpp

@@ -69,6 +69,7 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> tokens_list;
     tokens_list = ::llama_tokenize(model, params.prompt, true);
+
     const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size())*n_parallel;
 
     // initialize the context

examples/export-lora/export-lora.cpp

@@ -245,9 +245,8 @@ static struct lora_data * load_lora(struct lora_info * info) {
     params_ggml.no_alloc   = true;
     result->ctx = ggml_init(params_ggml);
 
-    uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
     uint32_t magic   = file.read_u32();
-    if (magic != LLAMA_FILE_MAGIC_LORA) {
+    if (magic != LLAMA_FILE_MAGIC_GGLA) {
         die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
     }
     uint32_t version = file.read_u32();
@@ -309,7 +308,7 @@ static struct ggml_cgraph * build_graph_lora(
 ) {
     struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
     if (scaling != 1.0f) {
-        ab = ggml_scale(ctx, ab, ggml_new_f32(ctx, scaling));
+        ab = ggml_scale(ctx, ab, scaling);
     }
     struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
 

examples/finetune/README.md

@@ -61,7 +61,7 @@ For example to apply 40% of the 'shakespeare' LORA adapter, 80% of the 'bible' L
   --lora lora-open-llama-3b-v2-q8_0-yet-another-one-LATEST.bin
 ```
 
-The scale numbers don't need to add up to one, and you can also use numbers greater than 1 to further increase the influence of an adapter. But making the values to big will sometimes result in worse output. Play around to find good values.
+The scale numbers don't need to add up to one, and you can also use numbers greater than 1 to further increase the influence of an adapter. But making the values too big will sometimes result in worse output. Play around to find good values.
 
 Gradient checkpointing reduces the memory requirements by ~50% but increases the runtime.
 If you have enough RAM, you can make finetuning a bit faster by disabling checkpointing with `--no-checkpointing`.

examples/finetune/finetune.cpp

@@ -3,15 +3,9 @@
 #include "llama.h"
 #include "common.h"
 #include "train.h"
-#include <unordered_map>
 #include <vector>
-#include <cassert>
-#include <climits>
 #include <cstring>
-#include <cstdarg>
 #include <ctime>
-#include <random>
-#include <stdexcept>
 #include <algorithm>
 #include <string>
 
@@ -196,13 +190,13 @@ static const char * LLM_TENSOR_FFN_DOWN      = "blk.%d.ffn_down";
 static const char * LLM_TENSOR_FFN_UP        = "blk.%d.ffn_up";
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab:   %u\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:     %u\n", __func__, params->n_ctx);
-    printf("%s: n_embd:    %u\n", __func__, params->n_embd);
-    printf("%s: n_ff:      %u\n", __func__, params->n_ff);
-    printf("%s: n_head:    %u\n", __func__, params->n_head);
-    printf("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
-    printf("%s: n_layer:   %u\n", __func__, params->n_layer);
+    printf("%s: n_vocab               : %u\n", __func__, params->n_vocab);
+    printf("%s: n_ctx                 : %u\n", __func__, params->n_ctx);
+    printf("%s: n_embd                : %u\n", __func__, params->n_embd);
+    printf("%s: n_ff                  : %u\n", __func__, params->n_ff);
+    printf("%s: n_head                : %u\n", __func__, params->n_head);
+    printf("%s: n_head_kv             : %u\n", __func__, params->n_head_kv);
+    printf("%s: n_layer               : %u\n", __func__, params->n_layer);
     printf("%s: norm_rms_eps          : %f\n", __func__, params->f_norm_rms_eps);
     printf("%s: rope_freq_base        : %f\n", __func__, params->rope_freq_base);
     printf("%s: rope_freq_scale       : %f\n", __func__, params->rope_freq_scale);
@@ -269,7 +263,7 @@ static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_h
     float rope_freq_scale = 1.0f;
     GGUF_GET_KEY(ctx, hparams->f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
     GGUF_GET_KEY(ctx, hparams->rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
-    GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
+    GGUF_GET_KEY(ctx, rope_freq_scale,         gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
     if (rope_freq_scale != 1.0f) {
         hparams->rope_freq_scale = 1.0f / rope_freq_scale;
     }
@@ -612,6 +606,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     const int n_rot       = hparams.n_embd_head();
     const int n_embd_head = hparams.n_embd_head();
     const int n_embd_gqa  = hparams.n_embd_gqa();
+
     const float rms_norm_eps    = hparams.f_norm_rms_eps;
     const float rope_freq_base  = hparams.rope_freq_base;
     const float rope_freq_scale = hparams.rope_freq_scale;
@@ -680,10 +675,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         checkpoints.push_back(t01);
     }
 
-    struct ggml_tensor * kv_scale = NULL;
-    if (!enable_flash_attn) {
-        kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
-    }
+    const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
 
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
@@ -781,32 +773,32 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     // make sure some tensors are not reallocated by inserting new temporary nodes depending on them
     int n_leafs_before = gb->n_leafs;
     int n_nodes_before = gb->n_nodes;
-    struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
+
     // output tensors
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
     // input gradient
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
     GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
     ggml_allocr_alloc(alloc, t36->grad);
     // KQ_pos
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
 
     // make sure base model tensors data cannot be used in viewable operations
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, one));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, one));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, 1.0f));
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, 1.0f));
     }
 
     // allocating checkpoints in one block to reduce memory fragmentation

examples/gguf/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET gguf)
 add_executable(${TARGET} gguf.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/gguf/gguf.cpp

@@ -1,5 +1,4 @@
 #include "ggml.h"
-#include "llama.h"
 
 #include <cstdio>
 #include <cinttypes>

examples/imatrix/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET imatrix)
+add_executable(${TARGET} imatrix.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/imatrix/imatrix.cpp

@@ -0,0 +1,380 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <sstream>
+#include <thread>
+#include <mutex>
+#include <vector>
+#include <fstream>
+#include <unordered_map>
+#include <algorithm>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+struct Stats {
+    std::vector<float> values;
+    int ncall = 0;
+};
+
+struct StatParams {
+    std::string ofile = "imatrix.dat";
+    int         n_output_frequency = 10;
+    int         verbosity = 1;
+    bool        collect_output_weight = false;
+};
+
+class IMatrixCollector {
+public:
+    IMatrixCollector() = default;
+    void set_parameters(StatParams&& params) { m_params = std::move(params); }
+    void collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1);
+    void save_imatrix() const;
+private:
+    std::unordered_map<std::string, Stats> m_stats;
+    StatParams                             m_params;
+    std::mutex                             m_mutex;
+    int                                    m_last_call = 0;
+};
+
+void IMatrixCollector::collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) {
+    if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return;
+    if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return;
+    std::lock_guard<std::mutex> lock(m_mutex);
+    auto& e = m_stats[src0->name];
+    if (e.values.empty()) {
+        e.values.resize(src1->ne[0], 0);
+    }
+    else if (e.values.size() != (size_t)src1->ne[0]) {
+        fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
+        exit(1); //GGML_ASSERT(false);
+    }
+    ++e.ncall;
+    if (m_params.verbosity > 1) {
+        printf("%s[%d]: %s, %d x %d, %d\n",__func__,m_last_call,src0->name,(int)src1->ne[0],(int)src1->ne[1],(int)src1->type);
+    }
+    for (int row = 0; row < (int)src1->ne[1]; ++row) {
+        const float * x = (const float *)src1->data + row * src1->ne[0];
+        for (int j = 0; j < (int)src1->ne[0]; ++j) {
+            e.values[j] += x[j]*x[j];
+        }
+    }
+    if (e.ncall > m_last_call) {
+        m_last_call = e.ncall;
+        if (m_last_call % m_params.n_output_frequency == 0) {
+            save_imatrix();
+        }
+    }
+}
+
+void IMatrixCollector::save_imatrix() const {
+    const char * fname = m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str();
+    std::ofstream out(fname, std::ios::binary);
+    int n_entries = m_stats.size();
+    out.write((const char*)&n_entries, sizeof(n_entries));
+    for (auto& p : m_stats) {
+        int len = p.first.size();
+        out.write((const char*)&len, sizeof(len));
+        out.write(p.first.c_str(), len);
+        out.write((const char*)&p.second.ncall, sizeof(p.second.ncall));
+        int nval = p.second.values.size();
+        out.write((const char*)&nval, sizeof(nval));
+        if (nval > 0) out.write((const char*)p.second.values.data(), nval*sizeof(float));
+    }
+    if (m_params.verbosity > 0) {
+        fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n",__func__,m_last_call,fname);
+    }
+}
+
+static IMatrixCollector g_collector;
+
+static void ik_collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) {
+    g_collector.collect_imatrix(src0, src1);
+}
+
+
+struct results_log_softmax {
+    double log_softmax;
+    float  logit;
+    float  prob;
+};
+
+static std::vector<float> softmax(const std::vector<float>& logits) {
+    std::vector<float> probs(logits.size());
+    float max_logit = logits[0];
+    for (float v : logits) {
+        max_logit = std::max(max_logit, v);
+    }
+    double sum_exp = 0.0;
+    for (size_t i = 0; i < logits.size(); i++) {
+        // Subtract the maximum logit value from the current logit value for numerical stability
+        const float logit = logits[i] - max_logit;
+        const float exp_logit = expf(logit);
+        sum_exp += exp_logit;
+        probs[i] = exp_logit;
+    }
+    for (size_t i = 0; i < probs.size(); i++) {
+        probs[i] /= sum_exp;
+    }
+    return probs;
+}
+
+static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
+    float max_logit = logits[0];
+    for (int i = 1; i < n_vocab; ++i) {
+        max_logit = std::max(max_logit, logits[i]);
+    }
+    double sum_exp = 0.0;
+    for (int i = 0; i < n_vocab; ++i) {
+        sum_exp += expf(logits[i] - max_logit);
+    }
+    return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
+}
+
+static void process_logits(
+    int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
+    double & nll, double & nll2, float * logit_history, float * prob_history
+) {
+    std::mutex mutex;
+    int counter = 0;
+    auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
+        double local_nll  = 0;
+        double local_nll2 = 0;
+        while (true) {
+            std::unique_lock<std::mutex> lock(mutex);
+            int i = counter++;
+            if (i >= n_token) {
+                nll += local_nll; nll2 += local_nll2;
+                break;
+            }
+            lock.unlock();
+            const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
+            const double v = -results.log_softmax;
+            local_nll += v;
+            local_nll2 += v*v;
+
+            logit_history[i] = results.logit;
+            prob_history[i]  = results.prob;
+        }
+    };
+    for (auto & w : workers) {
+        w = std::thread(compute);
+    }
+    compute();
+    for (auto & w : workers) {
+        w.join();
+    }
+}
+
+static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
+
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    const int n_ctx = llama_n_ctx(ctx);
+
+    auto tim1 = std::chrono::high_resolution_clock::now();
+    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
+
+    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    auto tim2 = std::chrono::high_resolution_clock::now();
+    fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
+
+    if (int(tokens.size()) < 2*n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n",__func__,2*n_ctx,
+                n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return false;
+    }
+
+    std::vector<float> logit_history;
+    logit_history.resize(tokens.size());
+
+    std::vector<float> prob_history;
+    prob_history.resize(tokens.size());
+
+    const int n_chunk_max = tokens.size() / n_ctx;
+
+    const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_batch = params.n_batch;
+
+    int count = 0;
+    double nll = 0.0;
+    double nll2 = 0.0;
+
+    fprintf(stderr, "%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch);
+
+    std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
+
+    for (int i = 0; i < n_chunk; ++i) {
+        const int start =     i * n_ctx;
+        const int end   = start + n_ctx;
+
+        const int num_batches = (n_ctx + n_batch - 1) / n_batch;
+
+        std::vector<float> logits;
+
+        const auto t_start = std::chrono::high_resolution_clock::now();
+
+        // clear the KV cache
+        llama_kv_cache_clear(ctx);
+
+        for (int j = 0; j < num_batches; ++j) {
+            const int batch_start = start + j * n_batch;
+            const int batch_size  = std::min(end - batch_start, n_batch);
+
+            // save original token and restore it after eval
+            const auto token_org = tokens[batch_start];
+
+            // add BOS token for the first batch of each chunk
+            if (add_bos && j == 0) {
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
+            }
+
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
+                fprintf(stderr, "%s : failed to eval\n", __func__);
+                return false;
+            }
+
+            // restore the original token in case it was set to BOS
+            tokens[batch_start] = token_org;
+
+            const auto * batch_logits = llama_get_logits(ctx);
+            logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+        }
+
+        const auto t_end = std::chrono::high_resolution_clock::now();
+
+        if (i == 0) {
+            const float t_total = std::chrono::duration<float>(t_end - t_start).count();
+            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            int total_seconds = (int)(t_total * n_chunk);
+            if (total_seconds >= 60*60) {
+                fprintf(stderr, "%d hours ", total_seconds / (60*60));
+                total_seconds = total_seconds % (60*60);
+            }
+            fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
+        }
+
+        const int first = n_ctx/2;
+        process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+                       workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
+        count += n_ctx - first - 1;
+
+        printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        fflush(stdout);
+    }
+    printf("\n");
+
+    nll2 /= count;
+    nll /= count;
+    const double ppl = exp(nll);
+    nll2 -= nll * nll;
+    if (nll2 > 0) {
+        nll2 = sqrt(nll2/(count-1));
+        printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
+    } else {
+        printf("Unexpected negative standard deviation of log(prob)\n");
+    }
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+
+    StatParams sparams;
+    std::vector<char*> args;
+    args.push_back(argv[0]);
+    int iarg = 1;
+    for (; iarg < argc-1; ++iarg) {
+        std::string arg{argv[iarg]};
+        if (arg == "-o" || arg == "--output-file") {
+            sparams.ofile = argv[++iarg];
+        }
+        else if (arg == "-ofreq" || arg == "--output-frequency") {
+            sparams.n_output_frequency = std::stoi(argv[++iarg]);
+        }
+        else if (arg == "-ow" || arg == "--output-weight") {
+            sparams.collect_output_weight = std::stoi(argv[++iarg]);
+        }
+        else if (arg == "--verbosity") {
+            sparams.verbosity = std::stoi(argv[++iarg]);
+        } else {
+            args.push_back(argv[iarg]);
+        }
+    }
+    if (iarg < argc) {
+        args.push_back(argv[iarg]);
+    }
+
+    gpt_params params;
+    params.n_batch = 512;
+    if (!gpt_params_parse(args.size(), args.data(), params)) {
+        return 1;
+    }
+
+    g_collector.set_parameters(std::move(sparams));
+
+    ggml_set_imatrix_collection(ik_collect_imatrix);
+
+    params.logits_all = true;
+    params.n_batch = std::min(params.n_batch, params.n_ctx);
+
+    print_build_info();
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+
+    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.random_prompt) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    llama_backend_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+
+    // load the model and apply lora adapter, if any
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        return 1;
+    }
+
+    const int n_ctx_train = llama_n_ctx_train(model);
+    if (params.n_ctx > n_ctx_train) {
+        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, params.n_ctx);
+    }
+
+    // print system information
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+    }
+
+    bool OK = compute_imatrix(ctx, params);
+    if (!OK) {
+        return 1;
+    }
+
+    g_collector.save_imatrix();
+
+    llama_print_timings(ctx);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/llama-bench/llama-bench.cpp

@@ -128,6 +128,25 @@ static std::string get_gpu_info() {
 // command line params
 enum output_formats {CSV, JSON, MARKDOWN, SQL};
 
+static const char * output_format_str(output_formats format) {
+    switch (format) {
+        case CSV:      return "csv";
+        case JSON:     return "json";
+        case MARKDOWN: return "md";
+        case SQL:      return "sql";
+        default: GGML_ASSERT(!"invalid output format");
+    }
+}
+
+static const char * split_mode_str(llama_split_mode mode) {
+    switch (mode) {
+        case LLAMA_SPLIT_NONE:  return "none";
+        case LLAMA_SPLIT_LAYER: return "layer";
+        case LLAMA_SPLIT_ROW:   return "row";
+        default: GGML_ASSERT(!"invalid split mode");
+    }
+}
+
 struct cmd_params {
     std::vector<std::string> model;
     std::vector<int> n_prompt;
@@ -137,7 +156,9 @@ struct cmd_params {
     std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
     std::vector<int> n_gpu_layers;
+    std::vector<llama_split_mode> split_mode;
     std::vector<int> main_gpu;
+    std::vector<bool> no_kv_offload;
     std::vector<bool> mul_mat_q;
     std::vector<std::array<float, LLAMA_MAX_DEVICES>> tensor_split;
     int reps;
@@ -154,7 +175,9 @@ static const cmd_params cmd_params_defaults = {
     /* type_v        */ {GGML_TYPE_F16},
     /* n_threads     */ {get_num_physical_cores()},
     /* n_gpu_layers  */ {99},
+    /* split_mode    */ {LLAMA_SPLIT_LAYER},
     /* main_gpu      */ {0},
+    /* no_kv_offload */ {false},
     /* mul_mat_q     */ {true},
     /* tensor_split  */ {{}},
     /* reps          */ 5,
@@ -167,20 +190,22 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("\n");
     printf("options:\n");
     printf("  -h, --help\n");
-    printf("  -m, --model <filename>            (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
-    printf("  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
-    printf("  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
-    printf("  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    printf("  -ctk <t>, --cache-type-k <t>      (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
-    printf("  -ctv <t>, --cache-type-v <t>      (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
-    printf("  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
-    printf("  -ngl, --n-gpu-layers <n>          (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
-    printf("  -mg, --main-gpu <i>               (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
-    printf("  -mmq, --mul-mat-q <0|1>           (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
-    printf("  -ts, --tensor_split <ts0/ts1/..>               \n");
-    printf("  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
-    printf("  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : cmd_params_defaults.output_format == MARKDOWN ? "md" : "sql");
-    printf("  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
+    printf("  -m, --model <filename>              (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
+    printf("  -p, --n-prompt <n>                  (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
+    printf("  -n, --n-gen <n>                     (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
+    printf("  -b, --batch-size <n>                (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
+    printf("  -ctk <t>, --cache-type-k <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
+    printf("  -ctv <t>, --cache-type-v <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
+    printf("  -t, --threads <n>                   (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
+    printf("  -ngl, --n-gpu-layers <n>            (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    printf("  -sm, --split-mode <none|layer|row>  (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
+    printf("  -mg, --main-gpu <i>                 (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
+    printf("  -nkvo, --no-kv-offload <0|1>        (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
+    printf("  -mmq, --mul-mat-q <0|1>             (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
+    printf("  -ts, --tensor_split <ts0/ts1/..>    (default: 0)\n");
+    printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
+    printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
+    printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     printf("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
 }
@@ -303,12 +328,41 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<int>(argv[i], split_delim);
             params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
+        } else if (arg == "-sm" || arg == "--split-mode") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<std::string>(argv[i], split_delim);
+            std::vector<llama_split_mode> modes;
+            for (const auto & m : p) {
+                llama_split_mode mode;
+                if (m == "none") {
+                    mode = LLAMA_SPLIT_NONE;
+                } else if (m == "layer") {
+                    mode = LLAMA_SPLIT_LAYER;
+                } else if (m == "row") {
+                    mode = LLAMA_SPLIT_ROW;
+                } else {
+                    invalid_param = true;
+                    break;
+                }
+                modes.push_back(mode);
+            }
+            params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
         } else if (arg == "-mg" || arg == "--main-gpu") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.main_gpu = split<int>(argv[i], split_delim);
+        } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<bool>(argv[i], split_delim);
+            params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
         } else if (arg == "-mmq" || arg == "--mul-mat-q") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -382,7 +436,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
     if (params.type_v.empty())       { params.type_v = cmd_params_defaults.type_v; }
     if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
+    if (params.split_mode.empty())   { params.split_mode = cmd_params_defaults.split_mode; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
+    if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; }
     if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
     if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
     if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
@@ -399,7 +455,9 @@ struct cmd_params_instance {
     ggml_type type_v;
     int n_threads;
     int n_gpu_layers;
+    llama_split_mode split_mode;
     int main_gpu;
+    bool no_kv_offload;
     bool mul_mat_q;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
 
@@ -407,6 +465,7 @@ struct cmd_params_instance {
         llama_model_params mparams = llama_model_default_params();
 
         mparams.n_gpu_layers = n_gpu_layers;
+        mparams.split_mode = split_mode;
         mparams.main_gpu = main_gpu;
         mparams.tensor_split = tensor_split.data();
 
@@ -416,6 +475,7 @@ struct cmd_params_instance {
     bool equal_mparams(const cmd_params_instance & other) const {
         return model == other.model &&
                n_gpu_layers == other.n_gpu_layers &&
+               split_mode == other.split_mode &&
                main_gpu == other.main_gpu &&
                tensor_split == other.tensor_split;
     }
@@ -428,54 +488,26 @@ struct cmd_params_instance {
         cparams.type_k = type_k;
         cparams.type_v = type_v;
         cparams.mul_mat_q = mul_mat_q;
+        cparams.offload_kqv = !no_kv_offload;
 
         return cparams;
     }
 };
 
-static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_params & params, int n_gen, int n_prompt) {
-    std::vector<cmd_params_instance> instances;
-
-    for (const auto & m : params.model)
-    for (const auto & nl : params.n_gpu_layers)
-    for (const auto & mg : params.main_gpu)
-    for (const auto & ts : params.tensor_split)
-    for (const auto & nb : params.n_batch)
-    for (const auto & tk : params.type_k)
-    for (const auto & tv : params.type_v)
-    for (const auto & mmq : params.mul_mat_q)
-    for (const auto & nt : params.n_threads) {
-        cmd_params_instance instance = {
-            /* .model        = */ m,
-            /* .n_prompt     = */ n_prompt,
-            /* .n_gen        = */ n_gen,
-            /* .n_batch      = */ nb,
-            /* .type_k       = */ tk,
-            /* .type_v       = */ tv,
-            /* .n_threads    = */ nt,
-            /* .n_gpu_layers = */ nl,
-            /* .main_gpu     = */ mg,
-            /* .mul_mat_q    = */ mmq,
-            /* .tensor_split = */ ts,
-        };
-        instances.push_back(instance);
-    }
-    return instances;
-}
-
 static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_params & params) {
     std::vector<cmd_params_instance> instances;
 
-#if 1
     // this ordering minimizes the number of times that each model needs to be reloaded
     for (const auto & m : params.model)
     for (const auto & nl : params.n_gpu_layers)
+    for (const auto & sm : params.split_mode)
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & nb : params.n_batch)
     for (const auto & tk : params.type_k)
     for (const auto & tv : params.type_v)
     for (const auto & mmq : params.mul_mat_q)
+    for (const auto & nkvo : params.no_kv_offload)
     for (const auto & nt : params.n_threads) {
         for (const auto & n_prompt : params.n_prompt) {
             if (n_prompt == 0) {
@@ -490,7 +522,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
+                /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
+                /* .no_kv_offload= */ nkvo,
                 /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
             };
@@ -510,31 +544,15 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
+                /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
+                /* .no_kv_offload= */ nkvo,
                 /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
             };
             instances.push_back(instance);
         }
     }
-#else
-    // this ordering separates the prompt and generation tests
-    for (const auto & n_prompt : params.n_prompt) {
-        if (n_prompt == 0) {
-            continue;
-        }
-        auto instances_prompt = get_cmd_params_instances_int(params, 0, n_prompt);
-        instances.insert(instances.end(), instances_prompt.begin(), instances_prompt.end());
-    }
-
-    for (const auto & n_gen : params.n_gen) {
-        if (n_gen == 0) {
-            continue;
-        }
-        auto instances_gen = get_cmd_params_instances_int(params, n_gen, 0);
-        instances.insert(instances.end(), instances_gen.begin(), instances_gen.end());
-    }
-#endif
 
     return instances;
 }
@@ -558,7 +576,9 @@ struct test {
     ggml_type type_k;
     ggml_type type_v;
     int n_gpu_layers;
+    llama_split_mode split_mode;
     int main_gpu;
+    bool no_kv_offload;
     bool mul_mat_q;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
     int n_prompt;
@@ -578,7 +598,9 @@ struct test {
         type_k = inst.type_k;
         type_v = inst.type_v;
         n_gpu_layers = inst.n_gpu_layers;
+        split_mode = inst.split_mode;
         main_gpu = inst.main_gpu;
+        no_kv_offload = inst.no_kv_offload;
         mul_mat_q = inst.mul_mat_q;
         tensor_split = inst.tensor_split;
         n_prompt = inst.n_prompt;
@@ -640,7 +662,9 @@ struct test {
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_threads", "type_k", "type_v",
-            "n_gpu_layers", "main_gpu", "mul_mat_q", "tensor_split",
+            "n_gpu_layers", "split_mode",
+            "main_gpu", "no_kv_offload",
+            "mul_mat_q", "tensor_split",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -659,7 +683,7 @@ struct test {
             return INT;
         }
         if (field == "cuda" || field == "opencl" || field == "metal" || field == "gpu_blas" || field == "blas" ||
-            field == "f16_kv" || field == "mul_mat_q") {
+            field == "f16_kv" || field == "no_kv_offload" || field == "mul_mat_q") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -690,7 +714,9 @@ struct test {
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
-            std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), tensor_split_str,
+            std::to_string(n_gpu_layers), split_mode_str(split_mode),
+            std::to_string(main_gpu), std::to_string(no_kv_offload),
+            std::to_string(mul_mat_q), tensor_split_str,
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
             std::to_string(avg_ts()), std::to_string(stdev_ts())
@@ -845,12 +871,18 @@ struct markdown_printer : public printer {
         if (field == "n_gpu_layers") {
             return "ngl";
         }
+        if (field == "split_mode") {
+            return "sm";
+        }
         if (field == "n_threads") {
             return "threads";
         }
         if (field == "mul_mat_q") {
             return "mmq";
         }
+        if (field == "no_kv_offload") {
+            return "nkvo";
+        }
         if (field == "tensor_split") {
             return "ts";
         }
@@ -882,9 +914,15 @@ struct markdown_printer : public printer {
         if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
             fields.push_back("main_gpu");
         }
+        if (params.split_mode.size() > 1 || params.split_mode != cmd_params_defaults.split_mode) {
+            fields.push_back("split_mode");
+        }
         if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
             fields.push_back("mul_mat_q");
         }
+        if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) {
+            fields.push_back("no_kv_offload");
+        }
         if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
             fields.push_back("tensor_split");
         }

examples/llama.swiftui/README.md

@@ -1,7 +1,12 @@
-# llama.swiftui
+# llama.cpp/examples/llama.swiftui
 
-Local inference of llama.cpp on an iPhone.
-So far I only tested with starcoder 1B model, but it can most likely handle 7B models as well.
+Local inference of llama.cpp on an iPhone. This is a sample app that can be used as a starting
+point for more advanced projects.
+
+For usage instructions and performance stats, check the following discussion: https://github.com/ggerganov/llama.cpp/discussions/4508
+
+![image](https://github.com/ggerganov/llama.cpp/assets/1991296/2b40284f-8421-47a2-b634-74eece09a299)
+
+Video demonstration:
 
 https://github.com/bachittle/llama.cpp/assets/39804642/e290827a-4edb-4093-9642-2a5e399ec545
-

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -1,6 +1,5 @@
 import Foundation
-
-// import llama
+import llama
 
 enum LlamaError: Error {
     case couldNotInitializeContext
@@ -159,7 +158,7 @@ actor LlamaContext {
             new_token_id = llama_sample_token_greedy(context, &candidates_p)
         }
 
-        if new_token_id == llama_token_eos(context) || n_cur == n_len {
+        if new_token_id == llama_token_eos(model) || n_cur == n_len {
             print("\n")
             let new_token_str = String(cString: temporary_invalid_cchars + [0])
             temporary_invalid_cchars.removeAll()

examples/llama.swiftui/llama.cpp.swift/bridging-header.h

@@ -1,5 +0,0 @@
-//
-//  Use this file to import your target's public headers that you would like to expose to Swift.
-//
-
-#import "llama.h"

examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj

@@ -7,51 +7,34 @@
 	objects = {
 
 /* Begin PBXBuildFile section */
-		542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 542376072B0D9BFB008E6A1C /* ggml-quants.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */ = {isa = PBXBuildFile; fileRef = 5423760A2B0D9C4B008E6A1C /* ggml-backend.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		542378792ACE3F3500834A7B /* ggml-metal.metal in Resources */ = {isa = PBXBuildFile; fileRef = 549479C82AC9E10B00E0F78B /* ggml-metal.metal */; };
-		542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09B2AC8723900A8AEE9 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -DGGML_USE_METAL -DGGML_USE_K_QUANTS -O3"; }; };
-		542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 542EA0A12AC8729100A8AEE9 /* llama.cpp */; settings = {COMPILER_FLAGS = "-DGGML_USE_K_QUANTS -DGGML_USE_METAL -O3"; }; };
 		549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; };
-		549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */ = {isa = PBXBuildFile; fileRef = 549479C52AC9E0F200E0F78B /* ggml-metal.m */; settings = {COMPILER_FLAGS = "-fno-objc-arc -DGGML_SWIFT -DGGML_USE_METAL -O3"; }; };
+		79E1D9CD2B4CD16E005F8E46 /* InputButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */; };
 		7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; };
 		8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */; };
 		8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83782AC328BD0096AF73 /* ContentView.swift */; };
 		8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837A2AC328BE0096AF73 /* Assets.xcassets */; };
-		8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */; };
 		8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 8A39BE092AC7601000BFEB40 /* Accelerate.framework */; };
 		8A3F84242AC4C891005E2EE8 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 8A3F84232AC4C891005E2EE8 /* models */; };
 		8A907F332AC7138A006146EA /* LibLlama.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A907F322AC7134E006146EA /* LibLlama.swift */; };
 		8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */; };
+		DF810E132B4A5BA200301144 /* llama in Frameworks */ = {isa = PBXBuildFile; productRef = DF810E122B4A5BA200301144 /* llama */; };
+		F1FE20E22B465ECA00B45541 /* LoadCustomButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */; };
 /* End PBXBuildFile section */
 
 /* Begin PBXFileReference section */
-		542376062B0D9BEA008E6A1C /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../ggml-quants.h"; sourceTree = "<group>"; };
-		542376072B0D9BFB008E6A1C /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../ggml-quants.c"; sourceTree = "<group>"; };
-		542376092B0D9C40008E6A1C /* ggml-backend.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../ggml-backend.h"; sourceTree = "<group>"; };
-		5423760A2B0D9C4B008E6A1C /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../ggml-backend.c"; sourceTree = "<group>"; };
-		542EA09B2AC8723900A8AEE9 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../ggml.c; sourceTree = "<group>"; };
-		542EA09C2AC8723900A8AEE9 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../ggml.h; sourceTree = "<group>"; };
-		542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../ggml-alloc.h"; sourceTree = "<group>"; };
-		542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../ggml-alloc.c"; sourceTree = "<group>"; };
-		542EA0A12AC8729100A8AEE9 /* llama.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = llama.cpp; path = ../../llama.cpp; sourceTree = "<group>"; };
-		542EA0A22AC8729100A8AEE9 /* llama.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = llama.h; path = ../../llama.h; sourceTree = "<group>"; };
-		549479C52AC9E0F200E0F78B /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../ggml-metal.m"; sourceTree = "<group>"; };
-		549479C62AC9E0F200E0F78B /* ggml-metal.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-metal.h"; path = "../../ggml-metal.h"; sourceTree = "<group>"; };
-		549479C82AC9E10B00E0F78B /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../ggml-metal.metal"; sourceTree = "<group>"; };
 		549479CA2AC9E16000E0F78B /* Metal.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Metal.framework; path = System/Library/Frameworks/Metal.framework; sourceTree = SDKROOT; };
+		79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = InputButton.swift; sourceTree = "<group>"; };
 		7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = DownloadButton.swift; sourceTree = "<group>"; };
-		8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = "bridging-header.h"; sourceTree = "<group>"; };
 		8A1C83732AC328BD0096AF73 /* llama.swiftui.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = llama.swiftui.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = llama_swiftuiApp.swift; sourceTree = "<group>"; };
 		8A1C83782AC328BD0096AF73 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = "<group>"; };
 		8A1C837A2AC328BE0096AF73 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
-		8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = "Preview Assets.xcassets"; sourceTree = "<group>"; };
 		8A39BE092AC7601000BFEB40 /* Accelerate.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Accelerate.framework; path = System/Library/Frameworks/Accelerate.framework; sourceTree = SDKROOT; };
 		8A3F84232AC4C891005E2EE8 /* models */ = {isa = PBXFileReference; lastKnownFileType = folder; name = models; path = llama.swiftui/Resources/models; sourceTree = "<group>"; };
 		8A907F322AC7134E006146EA /* LibLlama.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibLlama.swift; sourceTree = "<group>"; };
 		8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LlamaState.swift; sourceTree = "<group>"; };
+		DF2D2FE72B4A59BE00FCB72D /* llama.cpp */ = {isa = PBXFileReference; lastKnownFileType = wrapper; name = llama.cpp; path = ../..; sourceTree = "<group>"; };
+		F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LoadCustomButton.swift; sourceTree = "<group>"; };
 /* End PBXFileReference section */
 
 /* Begin PBXFrameworksBuildPhase section */
@@ -59,6 +42,7 @@
 			isa = PBXFrameworksBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
+				DF810E132B4A5BA200301144 /* llama in Frameworks */,
 				549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */,
 				8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */,
 			);
@@ -67,30 +51,10 @@
 /* End PBXFrameworksBuildPhase section */
 
 /* Begin PBXGroup section */
-		8A08D1F62AC7383900FE6CD4 /* llama.cpp */ = {
-			isa = PBXGroup;
-			children = (
-				5423760A2B0D9C4B008E6A1C /* ggml-backend.c */,
-				542376092B0D9C40008E6A1C /* ggml-backend.h */,
-				542376062B0D9BEA008E6A1C /* ggml-quants.h */,
-				542376072B0D9BFB008E6A1C /* ggml-quants.c */,
-				549479C82AC9E10B00E0F78B /* ggml-metal.metal */,
-				549479C62AC9E0F200E0F78B /* ggml-metal.h */,
-				549479C52AC9E0F200E0F78B /* ggml-metal.m */,
-				542EA09B2AC8723900A8AEE9 /* ggml.c */,
-				542EA09C2AC8723900A8AEE9 /* ggml.h */,
-				542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */,
-				542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */,
-				542EA0A12AC8729100A8AEE9 /* llama.cpp */,
-				542EA0A22AC8729100A8AEE9 /* llama.h */,
-			);
-			name = llama.cpp;
-			sourceTree = "<group>";
-		};
 		8A1C836A2AC328BD0096AF73 = {
 			isa = PBXGroup;
 			children = (
-				8A08D1F62AC7383900FE6CD4 /* llama.cpp */,
+				DF2D2FE72B4A59BE00FCB72D /* llama.cpp */,
 				8A907F312AC7134E006146EA /* llama.cpp.swift */,
 				8A3F84232AC4C891005E2EE8 /* models */,
 				8A1C83752AC328BD0096AF73 /* llama.swiftui */,
@@ -115,19 +79,10 @@
 				8A9F7C4A2AC332BF008AE1EA /* UI */,
 				8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */,
 				8A1C837A2AC328BE0096AF73 /* Assets.xcassets */,
-				8A1C837C2AC328BE0096AF73 /* Preview Content */,
 			);
 			path = llama.swiftui;
 			sourceTree = "<group>";
 		};
-		8A1C837C2AC328BE0096AF73 /* Preview Content */ = {
-			isa = PBXGroup;
-			children = (
-				8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */,
-			);
-			path = "Preview Content";
-			sourceTree = "<group>";
-		};
 		8A39BE082AC7601000BFEB40 /* Frameworks */ = {
 			isa = PBXGroup;
 			children = (
@@ -155,7 +110,6 @@
 		8A907F312AC7134E006146EA /* llama.cpp.swift */ = {
 			isa = PBXGroup;
 			children = (
-				8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */,
 				8A907F322AC7134E006146EA /* LibLlama.swift */,
 			);
 			path = llama.cpp.swift;
@@ -166,6 +120,8 @@
 			children = (
 				7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */,
 				8A1C83782AC328BD0096AF73 /* ContentView.swift */,
+				F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */,
+				79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */,
 			);
 			path = UI;
 			sourceTree = "<group>";
@@ -195,6 +151,7 @@
 			);
 			name = llama.swiftui;
 			packageProductDependencies = (
+				DF810E122B4A5BA200301144 /* llama */,
 			);
 			productName = llama.swiftui;
 			productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */;
@@ -241,9 +198,7 @@
 			isa = PBXResourcesBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				542378792ACE3F3500834A7B /* ggml-metal.metal in Resources */,
 				8A3F84242AC4C891005E2EE8 /* models in Resources */,
-				8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */,
 				8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
@@ -255,17 +210,13 @@
 			isa = PBXSourcesBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */,
-				549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */,
-				542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */,
+				F1FE20E22B465ECA00B45541 /* LoadCustomButton.swift in Sources */,
 				8A907F332AC7138A006146EA /* LibLlama.swift in Sources */,
-				542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */,
 				8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */,
 				8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */,
 				8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */,
 				7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */,
-				542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */,
-				5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */,
+				79E1D9CD2B4CD16E005F8E46 /* InputButton.swift in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -395,12 +346,10 @@
 			isa = XCBuildConfiguration;
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
-				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
-				DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
-				DEVELOPMENT_TEAM = STLSG3FG8Q;
+				DEVELOPMENT_TEAM = K5UQJPP73A;
 				ENABLE_PREVIEWS = YES;
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
@@ -416,11 +365,12 @@
 				MARKETING_VERSION = 1.0;
 				PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
 				PRODUCT_NAME = "$(TARGET_NAME)";
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator xros xrsimulator";
+				SUPPORTS_XR_DESIGNED_FOR_IPHONE_IPAD = NO;
 				SWIFT_EMIT_LOC_STRINGS = YES;
-				SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
 				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
 				SWIFT_VERSION = 5.0;
-				TARGETED_DEVICE_FAMILY = "1,2";
+				TARGETED_DEVICE_FAMILY = "1,2,7";
 			};
 			name = Debug;
 		};
@@ -428,12 +378,10 @@
 			isa = XCBuildConfiguration;
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
-				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
-				DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
-				DEVELOPMENT_TEAM = STLSG3FG8Q;
+				DEVELOPMENT_TEAM = K5UQJPP73A;
 				ENABLE_PREVIEWS = YES;
 				GENERATE_INFOPLIST_FILE = YES;
 				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
@@ -449,10 +397,11 @@
 				MARKETING_VERSION = 1.0;
 				PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
 				PRODUCT_NAME = "$(TARGET_NAME)";
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator xros xrsimulator";
+				SUPPORTS_XR_DESIGNED_FOR_IPHONE_IPAD = NO;
 				SWIFT_EMIT_LOC_STRINGS = YES;
-				SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
 				SWIFT_VERSION = 5.0;
-				TARGETED_DEVICE_FAMILY = "1,2";
+				TARGETED_DEVICE_FAMILY = "1,2,7";
 			};
 			name = Release;
 		};
@@ -478,6 +427,13 @@
 			defaultConfigurationName = Release;
 		};
 /* End XCConfigurationList section */
+
+/* Begin XCSwiftPackageProductDependency section */
+		DF810E122B4A5BA200301144 /* llama */ = {
+			isa = XCSwiftPackageProductDependency;
+			productName = llama;
+		};
+/* End XCSwiftPackageProductDependency section */
 	};
 	rootObject = 8A1C836B2AC328BD0096AF73 /* Project object */;
 }

examples/llama.swiftui/llama.swiftui/Assets.xcassets/AccentColor.colorset/Contents.json

@@ -1,11 +0,0 @@
-{
-  "colors" : [
-    {
-      "idiom" : "universal"
-    }
-  ],
-  "info" : {
-    "author" : "xcode",
-    "version" : 1
-  }
-}

examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift

@@ -1,9 +1,20 @@
 import Foundation
 
+struct Model: Identifiable {
+    var id = UUID()
+    var name: String
+    var url: String
+    var filename: String
+    var status: String?
+}
+
 @MainActor
 class LlamaState: ObservableObject {
     @Published var messageLog = ""
     @Published var cacheCleared = false
+    @Published var downloadedModels: [Model] = []
+    @Published var undownloadedModels: [Model] = []
+    let NS_PER_S = 1_000_000_000.0
 
     private var llamaContext: LlamaContext?
     private var defaultModelUrl: URL? {
@@ -12,37 +23,130 @@ class LlamaState: ObservableObject {
     }
 
     init() {
+        loadModelsFromDisk()
+        loadDefaultModels()
+    }
+
+    private func loadModelsFromDisk() {
+        do {
+            let documentsURL = getDocumentsDirectory()
+            let modelURLs = try FileManager.default.contentsOfDirectory(at: documentsURL, includingPropertiesForKeys: nil, options: [.skipsHiddenFiles, .skipsSubdirectoryDescendants])
+            for modelURL in modelURLs {
+                let modelName = modelURL.deletingPathExtension().lastPathComponent
+                downloadedModels.append(Model(name: modelName, url: "", filename: modelURL.lastPathComponent, status: "downloaded"))
+            }
+        } catch {
+            print("Error loading models from disk: \(error)")
+        }
+    }
+
+    private func loadDefaultModels() {
         do {
             try loadModel(modelUrl: defaultModelUrl)
         } catch {
             messageLog += "Error!\n"
         }
-    }
 
-    func loadModel(modelUrl: URL?) throws {
-        messageLog += "Loading model...\n"
-        if let modelUrl {
-            llamaContext = try LlamaContext.create_context(path: modelUrl.path())
-            messageLog += "Loaded model \(modelUrl.lastPathComponent)\n"
-        } else {
-            messageLog += "Could not locate model\n"
+        for model in defaultModels {
+            let fileURL = getDocumentsDirectory().appendingPathComponent(model.filename)
+            if FileManager.default.fileExists(atPath: fileURL.path) {
+
+            } else {
+                var undownloadedModel = model
+                undownloadedModel.status = "download"
+                undownloadedModels.append(undownloadedModel)
+            }
         }
     }
 
+    func getDocumentsDirectory() -> URL {
+        let paths = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)
+        return paths[0]
+    }
+    private let defaultModels: [Model] = [
+        Model(name: "TinyLlama-1.1B (Q4_0, 0.6 GiB)",url: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true",filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf", status: "download"),
+        Model(
+            name: "TinyLlama-1.1B Chat (Q8_0, 1.1 GiB)",
+            url: "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/main/tinyllama-1.1b-chat-v1.0.Q8_0.gguf?download=true",
+            filename: "tinyllama-1.1b-chat-v1.0.Q8_0.gguf", status: "download"
+        ),
+
+        Model(
+            name: "TinyLlama-1.1B (F16, 2.2 GiB)",
+            url: "https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf?download=true",
+            filename: "tinyllama-1.1b-f16.gguf", status: "download"
+        ),
+
+        Model(
+            name: "Phi-2.7B (Q4_0, 1.6 GiB)",
+            url: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf?download=true",
+            filename: "phi-2-q4_0.gguf", status: "download"
+        ),
+
+        Model(
+            name: "Phi-2.7B (Q8_0, 2.8 GiB)",
+            url: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q8_0.gguf?download=true",
+            filename: "phi-2-q8_0.gguf", status: "download"
+        ),
+
+        Model(
+            name: "Mistral-7B-v0.1 (Q4_0, 3.8 GiB)",
+            url: "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF/resolve/main/mistral-7b-v0.1.Q4_0.gguf?download=true",
+            filename: "mistral-7b-v0.1.Q4_0.gguf", status: "download"
+        ),
+        Model(
+            name: "OpenHermes-2.5-Mistral-7B (Q3_K_M, 3.52 GiB)",
+            url: "https://huggingface.co/TheBloke/OpenHermes-2.5-Mistral-7B-GGUF/resolve/main/openhermes-2.5-mistral-7b.Q3_K_M.gguf?download=true",
+            filename: "openhermes-2.5-mistral-7b.Q3_K_M.gguf", status: "download"
+        )
+    ]
+    func loadModel(modelUrl: URL?) throws {
+        if let modelUrl {
+            messageLog += "Loading model...\n"
+            llamaContext = try LlamaContext.create_context(path: modelUrl.path())
+            messageLog += "Loaded model \(modelUrl.lastPathComponent)\n"
+
+            // Assuming that the model is successfully loaded, update the downloaded models
+            updateDownloadedModels(modelName: modelUrl.lastPathComponent, status: "downloaded")
+        } else {
+            messageLog += "Load a model from the list below\n"
+        }
+    }
+
+
+    private func updateDownloadedModels(modelName: String, status: String) {
+        undownloadedModels.removeAll { $0.name == modelName }
+    }
+
+
     func complete(text: String) async {
         guard let llamaContext else {
             return
         }
 
+        let t_start = DispatchTime.now().uptimeNanoseconds
         await llamaContext.completion_init(text: text)
+        let t_heat_end = DispatchTime.now().uptimeNanoseconds
+        let t_heat = Double(t_heat_end - t_start) / NS_PER_S
+
         messageLog += "\(text)"
 
-        while await llamaContext.n_cur <= llamaContext.n_len {
+        while await llamaContext.n_cur < llamaContext.n_len {
             let result = await llamaContext.completion_loop()
             messageLog += "\(result)"
         }
+
+        let t_end = DispatchTime.now().uptimeNanoseconds
+        let t_generation = Double(t_end - t_heat_end) / NS_PER_S
+        let tokens_per_second = Double(await llamaContext.n_len) / t_generation
+
         await llamaContext.clear()
-        messageLog += "\n\ndone\n"
+        messageLog += """
+            \n
+            Done
+            Heat up took \(t_heat)s
+            Generated \(tokens_per_second) t/s\n
+            """
     }
 
     func bench() async {
@@ -56,10 +160,10 @@ class LlamaState: ObservableObject {
         messageLog += await llamaContext.model_info() + "\n"
 
         let t_start = DispatchTime.now().uptimeNanoseconds
-        await llamaContext.bench(pp: 8, tg: 4, pl: 1) // heat up
+        let _ = await llamaContext.bench(pp: 8, tg: 4, pl: 1) // heat up
         let t_end = DispatchTime.now().uptimeNanoseconds
 
-        let t_heat = Double(t_end - t_start) / 1_000_000_000.0
+        let t_heat = Double(t_end - t_start) / NS_PER_S
         messageLog += "Heat up time: \(t_heat) seconds, please wait...\n"
 
         // if more than 5 seconds, then we're probably running on a slow device

examples/llama.swiftui/llama.swiftui/Preview Content/Preview Assets.xcassets/Contents.json

@@ -1,6 +0,0 @@
-{
-  "info" : {
-    "author" : "xcode",
-    "version" : 1
-  }
-}

examples/llama.swiftui/llama.swiftui/UI/ContentView.swift

@@ -2,138 +2,57 @@ import SwiftUI
 
 struct ContentView: View {
     @StateObject var llamaState = LlamaState()
-
     @State private var multiLineText = ""
-
-    private static func cleanupModelCaches() {
-        // Delete all models (*.gguf)
-        let fileManager = FileManager.default
-        let documentsUrl =  FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0]
-        do {
-            let fileURLs = try fileManager.contentsOfDirectory(at: documentsUrl, includingPropertiesForKeys: nil)
-            for fileURL in fileURLs {
-                if fileURL.pathExtension == "gguf" {
-                    try fileManager.removeItem(at: fileURL)
-                }
-            }
-        } catch {
-            print("Error while enumerating files \(documentsUrl.path): \(error.localizedDescription)")
-        }
-    }
+    @State private var showingHelp = false    // To track if Help Sheet should be shown
 
     var body: some View {
-        VStack {
-            ScrollView(.vertical, showsIndicators: true) {
-                Text(llamaState.messageLog)
-                .font(.system(size: 12))
-                .frame(maxWidth: .infinity, alignment: .leading)
-                .padding()
-                .onTapGesture {
-                    UIApplication.shared.sendAction(#selector(UIResponder.resignFirstResponder), to: nil, from: nil, for: nil)
-                }
-            }
-
-            TextEditor(text: $multiLineText)
-                .frame(height: 80)
-                .padding()
-                .border(Color.gray, width: 0.5)
-
-            HStack {
-                Button("Send") {
-                    sendText()
-                }
-                .padding(8)
-                .background(Color.blue)
-                .foregroundColor(.white)
-                .cornerRadius(8)
-
-                Button("Bench") {
-                    bench()
-                }
-                .padding(8)
-                .background(Color.blue)
-                .foregroundColor(.white)
-                .cornerRadius(8)
-
-                Button("Clear") {
-                    clear()
-                }
-                .padding(8)
-                .background(Color.blue)
-                .foregroundColor(.white)
-                .cornerRadius(8)
-
-                Button("Copy") {
-                    UIPasteboard.general.string = llamaState.messageLog
-                }
-                .padding(8)
-                .background(Color.blue)
-                .foregroundColor(.white)
-                .cornerRadius(8)
-            }
-
+        NavigationView {
             VStack {
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "TinyLlama-1.1B (Q4_0, 0.6 GiB)",
-                    modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true",
-                    filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf"
-                )
-                .font(.system(size: 12))
-                .padding(.top, 4)
-                .frame(maxWidth: .infinity, alignment: .leading)
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "TinyLlama-1.1B (Q8_0, 1.1 GiB)",
-                    modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q8_0.gguf?download=true",
-                    filename: "tinyllama-1.1b-1t-openorca.Q8_0.gguf"
-                )
-                .font(.system(size: 12))
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "TinyLlama-1.1B (F16, 2.2 GiB)",
-                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf?download=true",
-                    filename: "tinyllama-1.1b-f16.gguf"
-                )
-                .font(.system(size: 12))
-                .frame(maxWidth: .infinity, alignment: .leading)
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "Phi-2.7B (Q4_0, 1.6 GiB)",
-                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf?download=true",
-                    filename: "phi-2-q4_0.gguf"
-                )
-                .font(.system(size: 12))
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "Phi-2.7B (Q8_0, 2.8 GiB)",
-                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q8_0.gguf?download=true",
-                    filename: "phi-2-q8_0.gguf"
-                )
-                .font(.system(size: 12))
-                .frame(maxWidth: .infinity, alignment: .leading)
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "Mistral-7B-v0.1 (Q4_0, 3.8 GiB)",
-                    modelUrl: "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF/resolve/main/mistral-7b-v0.1.Q4_0.gguf?download=true",
-                    filename: "mistral-7b-v0.1.Q4_0.gguf"
-                )
-                .font(.system(size: 12))
-
-                Button("Clear downloaded models") {
-                    ContentView.cleanupModelCaches()
-                    llamaState.cacheCleared = true
+                ScrollView(.vertical, showsIndicators: true) {
+                    Text(llamaState.messageLog)
+                        .font(.system(size: 12))
+                        .frame(maxWidth: .infinity, alignment: .leading)
+                        .padding()
+                        .onTapGesture {
+                            UIApplication.shared.sendAction(#selector(UIResponder.resignFirstResponder), to: nil, from: nil, for: nil)
+                        }
                 }
-                .padding(8)
-                .font(.system(size: 12))
+
+                TextEditor(text: $multiLineText)
+                    .frame(height: 80)
+                    .padding()
+                    .border(Color.gray, width: 0.5)
+
+                HStack {
+                    Button("Send") {
+                        sendText()
+                    }
+
+                    Button("Bench") {
+                        bench()
+                    }
+
+                    Button("Clear") {
+                        clear()
+                    }
+
+                    Button("Copy") {
+                        UIPasteboard.general.string = llamaState.messageLog
+                    }
+                }
+                .buttonStyle(.bordered)
+                .padding()
+
+                NavigationLink(destination: DrawerView(llamaState: llamaState)) {
+                    Text("View Models")
+                }
+                .padding()
+
             }
+            .padding()
+            .navigationBarTitle("Model Settings", displayMode: .inline)
+
         }
-        .padding()
     }
 
     func sendText() {
@@ -154,8 +73,73 @@ struct ContentView: View {
             await llamaState.clear()
         }
     }
+    struct DrawerView: View {
+
+        @ObservedObject var llamaState: LlamaState
+        @State private var showingHelp = false
+        func delete(at offsets: IndexSet) {
+            offsets.forEach { offset in
+                let model = llamaState.downloadedModels[offset]
+                let fileURL = getDocumentsDirectory().appendingPathComponent(model.filename)
+                do {
+                    try FileManager.default.removeItem(at: fileURL)
+                } catch {
+                    print("Error deleting file: \(error)")
+                }
+            }
+
+            // Remove models from downloadedModels array
+            llamaState.downloadedModels.remove(atOffsets: offsets)
+        }
+
+        func getDocumentsDirectory() -> URL {
+            let paths = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)
+            return paths[0]
+        }
+        var body: some View {
+            List {
+                Section(header: Text("Download Models From Hugging Face")) {
+                    HStack {
+                        InputButton(llamaState: llamaState)
+                    }
+                }
+                Section(header: Text("Downloaded Models")) {
+                    ForEach(llamaState.downloadedModels) { model in
+                        DownloadButton(llamaState: llamaState, modelName: model.name, modelUrl: model.url, filename: model.filename)
+                    }
+                    .onDelete(perform: delete)
+                }
+                Section(header: Text("Default Models")) {
+                    ForEach(llamaState.undownloadedModels) { model in
+                        DownloadButton(llamaState: llamaState, modelName: model.name, modelUrl: model.url, filename: model.filename)
+                    }
+                }
+
+            }
+            .listStyle(GroupedListStyle())
+            .navigationBarTitle("Model Settings", displayMode: .inline).toolbar {
+                ToolbarItem(placement: .navigationBarTrailing) {
+                    Button("Help") {
+                        showingHelp = true
+                    }
+                }
+            }.sheet(isPresented: $showingHelp) {    // Sheet for help modal
+                VStack(alignment: .leading) {
+                    VStack(alignment: .leading) {
+                        Text("1. Make sure the model is in GGUF Format")
+                               .padding()
+                        Text("2. Copy the download link of the quantized model")
+                               .padding()
+                    }
+                    Spacer()
+                   }
+            }
+        }
+    }
 }
 
-//#Preview {
-//    ContentView()
-//}
+struct ContentView_Previews: PreviewProvider {
+    static var previews: some View {
+        ContentView()
+    }
+}

examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift

@@ -53,6 +53,8 @@ struct DownloadButton: View {
 
                     llamaState.cacheCleared = false
 
+                    let model = Model(name: modelName, url: modelUrl, filename: filename, status: "downloaded")
+                    llamaState.downloadedModels.append(model)
                     status = "downloaded"
                 }
             } catch let err {
@@ -93,7 +95,7 @@ struct DownloadButton: View {
                         print("Error: \(err.localizedDescription)")
                     }
                 }) {
-                    Text("\(modelName) (Downloaded)")
+                    Text("Load \(modelName)")
                 }
             } else {
                 Text("Unknown status")

examples/llama.swiftui/llama.swiftui/UI/InputButton.swift

@@ -0,0 +1,131 @@
+import SwiftUI
+
+struct InputButton: View {
+    @ObservedObject var llamaState: LlamaState
+    @State private var inputLink: String = ""
+    @State private var status: String = "download"
+    @State private var filename: String = ""
+
+    @State private var downloadTask: URLSessionDownloadTask?
+    @State private var progress = 0.0
+    @State private var observation: NSKeyValueObservation?
+
+    private static func extractModelInfo(from link: String) -> (modelName: String, filename: String)? {
+        guard let url = URL(string: link),
+              let lastPathComponent = url.lastPathComponent.components(separatedBy: ".").first,
+              let modelName = lastPathComponent.components(separatedBy: "-").dropLast().joined(separator: "-").removingPercentEncoding,
+              let filename = lastPathComponent.removingPercentEncoding else {
+            return nil
+        }
+
+        return (modelName, filename)
+    }
+
+    private static func getFileURL(filename: String) -> URL {
+        FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0].appendingPathComponent(filename)
+    }
+
+    private func download() {
+        guard let extractedInfo = InputButton.extractModelInfo(from: inputLink) else {
+            // Handle invalid link or extraction failure
+            return
+        }
+
+        let (modelName, filename) = extractedInfo
+        self.filename = filename  // Set the state variable
+
+        status = "downloading"
+        print("Downloading model \(modelName) from \(inputLink)")
+        guard let url = URL(string: inputLink) else { return }
+        let fileURL = InputButton.getFileURL(filename: filename)
+
+        downloadTask = URLSession.shared.downloadTask(with: url) { temporaryURL, response, error in
+            if let error = error {
+                print("Error: \(error.localizedDescription)")
+                return
+            }
+
+            guard let response = response as? HTTPURLResponse, (200...299).contains(response.statusCode) else {
+                print("Server error!")
+                return
+            }
+
+            do {
+                if let temporaryURL = temporaryURL {
+                    try FileManager.default.copyItem(at: temporaryURL, to: fileURL)
+                    print("Writing to \(filename) completed")
+
+                    llamaState.cacheCleared = false
+
+                    let model = Model(name: modelName, url: self.inputLink, filename: filename, status: "downloaded")
+                    llamaState.downloadedModels.append(model)
+                    status = "downloaded"
+                }
+            } catch let err {
+                print("Error: \(err.localizedDescription)")
+            }
+        }
+
+        observation = downloadTask?.progress.observe(\.fractionCompleted) { progress, _ in
+            self.progress = progress.fractionCompleted
+        }
+
+        downloadTask?.resume()
+    }
+
+    var body: some View {
+        VStack {
+            HStack {
+                TextField("Paste Quantized Download Link", text: $inputLink)
+                    .textFieldStyle(RoundedBorderTextFieldStyle())
+
+                Button(action: {
+                    downloadTask?.cancel()
+                    status = "download"
+                }) {
+                    Text("Cancel")
+                }
+            }
+
+            if status == "download" {
+                Button(action: download) {
+                    Text("Download Custom Model")
+                }
+            } else if status == "downloading" {
+                Button(action: {
+                    downloadTask?.cancel()
+                    status = "download"
+                }) {
+                    Text("Downloading \(Int(progress * 100))%")
+                }
+            } else if status == "downloaded" {
+                Button(action: {
+                    let fileURL = InputButton.getFileURL(filename: self.filename)
+                    if !FileManager.default.fileExists(atPath: fileURL.path) {
+                        download()
+                        return
+                    }
+                    do {
+                        try llamaState.loadModel(modelUrl: fileURL)
+                    } catch let err {
+                        print("Error: \(err.localizedDescription)")
+                    }
+                }) {
+                    Text("Load Custom Model")
+                }
+            } else {
+                Text("Unknown status")
+            }
+        }
+        .onDisappear() {
+            downloadTask?.cancel()
+        }
+        .onChange(of: llamaState.cacheCleared) { newValue in
+            if newValue {
+                downloadTask?.cancel()
+                let fileURL = InputButton.getFileURL(filename: self.filename)
+                status = FileManager.default.fileExists(atPath: fileURL.path) ? "downloaded" : "download"
+            }
+        }
+    }
+}

examples/llama.swiftui/llama.swiftui/UI/LoadCustomButton.swift

@@ -0,0 +1,44 @@
+import SwiftUI
+import UniformTypeIdentifiers
+
+struct LoadCustomButton: View {
+    @ObservedObject private var llamaState: LlamaState
+    @State private var showFileImporter = false
+
+    init(llamaState: LlamaState) {
+        self.llamaState = llamaState
+    }
+
+    var body: some View {
+        VStack {
+            Button(action: {
+                showFileImporter = true
+            }) {
+                Text("Load Custom Model")
+            }
+        }
+        .fileImporter(
+            isPresented: $showFileImporter,
+            allowedContentTypes: [UTType(filenameExtension: "gguf", conformingTo: .data)!],
+            allowsMultipleSelection: false
+        ) { result in
+            switch result {
+            case .success(let files):
+                files.forEach { file in
+                    let gotAccess = file.startAccessingSecurityScopedResource()
+                    if !gotAccess { return }
+
+                    do {
+                        try llamaState.loadModel(modelUrl: file.absoluteURL)
+                    } catch let err {
+                        print("Error: \(err.localizedDescription)")
+                    }
+
+                    file.stopAccessingSecurityScopedResource()
+                }
+            case .failure(let error):
+                print(error)
+            }
+        }
+    }
+}

examples/llava/CMakeLists.txt

@@ -24,7 +24,8 @@ endif()
 
 if (NOT MSVC)
     target_compile_options(llava PRIVATE -Wno-cast-qual) # stb_image.h
-    endif()
+endif()
+
 if(TARGET BUILD_INFO)
     add_dependencies(llava BUILD_INFO)
 endif()
@@ -32,5 +33,5 @@ endif()
 set(TARGET llava-cli)
 add_executable(llava-cli llava-cli.cpp)
 install(TARGETS llava-cli RUNTIME)
-target_link_libraries(llava-cli PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(llava-cli PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(llava PRIVATE cxx_std_11)

examples/llava/clip.cpp

@@ -16,12 +16,19 @@
 #include "clip.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
+#include "ggml-backend.h"
+
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
 
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 
-#define CLIP_DEBUG
-
 static std::string format(const char * fmt, ...) {
     va_list ap;
     va_list ap2;
@@ -119,26 +126,30 @@ static struct ggml_tensor * get_tensor(struct ggml_context * ctx, const std::str
 }
 
 static std::string get_ftype(int ftype) {
-    switch (ftype) {
-    case 0:
-        return "f32";
-    case 1:
-        return "f16";
-    case 2:
-        return "q4_0";
-    case 3:
-        return "q4_1";
-    case 6:
-        return "q5_0";
-    case 7:
-        return "q5_1";
-    case 8:
-        return "q8_0";
-    default:
-        throw std::runtime_error(format("%s: Unrecognized file type: %d\n", __func__, ftype));
-    }
+    return ggml_type_name(static_cast<ggml_type>(ftype));
 }
 
+//
+// image data
+//
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
 //
 // clip layers
 //
@@ -196,39 +207,31 @@ struct clip_vision_model {
     struct ggml_tensor * mm_2_b;
 };
 
-// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
-struct clip_buffer {
-    uint8_t * data = NULL;
-    size_t size = 0;
-
-    void resize(size_t size) {
-        delete[] data;
-        data = new uint8_t[size];
-        this->size = size;
-    }
-
-    ~clip_buffer() { delete[] data; }
-};
-
 struct clip_ctx {
-    bool has_text_encoder = false;
-    bool has_vision_encoder = false;
+    bool has_text_encoder    = false;
+    bool has_vision_encoder  = false;
     bool has_llava_projector = false;
+
     struct clip_vision_model vision_model;
+
     float image_mean[3];
     float image_std[3];
     bool use_gelu = false;
     int32_t ftype = 1;
-    struct ggml_context * ctx;
+
     struct gguf_context * ctx_gguf;
+    struct ggml_context * ctx_data;
+
+    std::vector<uint8_t> buf_compute_meta;
 
     // memory buffers to evaluate the model
-    clip_buffer buf_compute;
-    clip_buffer buf_alloc;
-    ggml_allocr * alloc = NULL;
+    ggml_backend_buffer_t params_buffer = NULL;
+    ggml_backend_buffer_t compute_buffer = NULL;
+    ggml_backend_t backend = NULL;
+    ggml_allocr * compute_alloc = NULL;
 };
 
-static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_image_f32_batch * imgs) {
+static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
     if (!ctx->has_vision_encoder) {
         printf("This gguf file seems to have no vision encoder\n");
         return nullptr;
@@ -249,28 +252,24 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
     //const int projection_dim = hparams.projection_dim;
     const float eps = hparams.eps;
     int batch_size = imgs->size;
-    if(ctx->has_llava_projector) {
+    if (ctx->has_llava_projector) {
         GGML_ASSERT(batch_size == 1);
     }
 
-    const auto & buf_compute = ctx->buf_compute;
-
     struct ggml_init_params params = {
-        /*.mem_size =*/ buf_compute.size,
-        /*.mem_buffer =*/ buf_compute.data,
-        /*.no_alloc =*/ false,
+        /*.mem_size   =*/ ctx->buf_compute_meta.size(),
+        /*.mem_buffer =*/ ctx->buf_compute_meta.data(),
+        /*.no_alloc   =*/ true,
     };
 
-    params.no_alloc = true;
-
     struct ggml_context * ctx0 = ggml_init(params);
     struct ggml_cgraph * gf = ggml_new_graph(ctx0);
 
     struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size, image_size, 3, batch_size);
-    ggml_allocr_alloc(ctx->alloc, inp_raw);
+    ggml_allocr_alloc(ctx->compute_alloc, inp_raw);
 
-    if (!ggml_allocr_is_measure(ctx->alloc)) {
-        float * data = (float *)ggml_get_data(inp_raw);
+    if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
+        float * data = (float *)malloc(ggml_nbytes(inp_raw));
 
         for (size_t i = 0; i < imgs->size; i++) {
             const int nx = imgs->data[i].nx;
@@ -283,12 +282,14 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
                 for (int k = 0; k < 3; k++) {
                     for (int y = 0; y < ny; y++) {
                         for (int x = 0; x < nx; x++) {
-                            data[(b * 3 * n) + k * n + y * nx + x] = imgs->data[b].data[3 * (y * nx + x) + k];
+                            data[(b * 3 * n) + k * n + y * nx + x] = imgs->data[b].buf[3 * (y * nx + x) + k];
                         }
                     }
                 }
             }
         }
+        ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw));
+        free(data);
     }
 
     struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
@@ -298,42 +299,39 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
 
     // concat class_embeddings and patch_embeddings
     struct ggml_tensor * embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
-    ggml_allocr_alloc(ctx->alloc, embeddings);
-    if (!ggml_allocr_is_measure(ctx->alloc)) {
-        ggml_set_zero(embeddings);
+    ggml_allocr_alloc(ctx->compute_alloc, embeddings);
+    if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
+        void* zero_mem = malloc(ggml_nbytes(embeddings));
+        memset(zero_mem, 0, ggml_nbytes(embeddings));
+        ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
+        free(zero_mem);
     }
 
-    struct ggml_tensor * temp = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, 1, batch_size);
-    ggml_allocr_alloc(ctx->alloc, temp);
+    embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
+            embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
 
-    embeddings = ggml_acc(ctx0, embeddings, ggml_repeat(ctx0, model.class_embedding, temp), embeddings->nb[1],
-                          embeddings->nb[2], embeddings->nb[3], 0);
-    embeddings =
-        ggml_acc(ctx0, embeddings, inp, embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+    embeddings = ggml_acc(ctx0, embeddings, inp,
+            embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
 
     struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
-    ggml_allocr_alloc(ctx->alloc, positions);
-    if (!ggml_allocr_is_measure(ctx->alloc)) {
+    ggml_allocr_alloc(ctx->compute_alloc, positions);
+    if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
+        int* positions_data = (int*)malloc(ggml_nbytes(positions));
         for (int i = 0; i < num_positions; i++) {
-            ggml_set_i32_1d(positions, i, i);
+            positions_data[i] = i;
         }
+        ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+        free(positions_data);
     }
 
     embeddings =
-        ggml_add(ctx0, embeddings, ggml_repeat(ctx0, ggml_get_rows(ctx0, model.position_embeddings, positions), embeddings));
+        ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
 
     // pre-layernorm
     {
         embeddings = ggml_norm(ctx0, embeddings, eps);
 
-        embeddings = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, model.pre_ln_w, embeddings), embeddings),
-                              ggml_repeat(ctx0, model.pre_ln_b, embeddings));
-    }
-
-    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
-    ggml_allocr_alloc(ctx->alloc, KQ_scale);
-    if (!ggml_allocr_is_measure(ctx->alloc)) {
-        ggml_set_f32(KQ_scale, 1.0f / sqrt((float)d_head));
+        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b);
     }
 
     // loop over layers
@@ -346,30 +344,30 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
         {
             cur = ggml_norm(ctx0, cur, eps);
 
-            cur = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, model.layers[il].ln_1_w, cur), cur),
-                           ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_1_w),
+                           model.layers[il].ln_1_b);
         }
 
         // self-attention
         {
 
             struct ggml_tensor * Q =
-                ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].q_b, cur), ggml_mul_mat(ctx0, model.layers[il].q_w, cur));
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b);
 
-            Q = ggml_scale_inplace(ctx0, Q, KQ_scale);
+            Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
             Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size);
             Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
             Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size);
 
             struct ggml_tensor * K =
-                ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].k_b, cur), ggml_mul_mat(ctx0, model.layers[il].k_w, cur));
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b);
 
             K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
             K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
             K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
 
             struct ggml_tensor * V =
-                ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].v_b, cur), ggml_mul_mat(ctx0, model.layers[il].v_w, cur));
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].v_w, cur), model.layers[il].v_b);
 
             V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size);
             V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
@@ -385,7 +383,7 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
         }
 
         // attention output
-        cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].o_b, cur), ggml_mul_mat(ctx0, model.layers[il].o_w, cur));
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].o_w, cur), model.layers[il].o_b);
 
         // re-add the layer input, e.g., residual
         cur = ggml_add(ctx0, cur, embeddings);
@@ -396,12 +394,11 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
         {
             cur = ggml_norm(ctx0, cur, eps);
 
-            cur = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, model.layers[il].ln_2_w, cur), cur),
-                           ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_2_w), model.layers[il].ln_2_b);
         }
 
         cur = ggml_mul_mat(ctx0, model.layers[il].ff_i_w, cur);
-        cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].ff_i_b, cur), cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_i_b);
 
         if (ctx->use_gelu) {
             cur = ggml_gelu_inplace(ctx0, cur);
@@ -410,7 +407,7 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
         }
 
         cur = ggml_mul_mat(ctx0, model.layers[il].ff_o_w, cur);
-        cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].ff_o_b, cur), cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_o_b);
 
         // residual 2
         cur = ggml_add(ctx0, embeddings, cur);
@@ -423,23 +420,26 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
         embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
 
         struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches);
-        ggml_allocr_alloc(ctx->alloc, patches);
-        if (!ggml_allocr_is_measure(ctx->alloc)) {
-            for (int i = 0; i < num_patches; ++i) {
-                ggml_set_i32_1d(patches, i, i+1);
+        ggml_allocr_alloc(ctx->compute_alloc, patches);
+        if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
+            int* patches_data = (int*)malloc(ggml_nbytes(patches));
+            for (int i = 0; i < num_patches; i++) {
+                patches_data[i] = i + 1;
             }
+            ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+            free(patches_data);
         }
 
         embeddings = ggml_get_rows(ctx0, embeddings, patches);
 
         // mm projection 0
         embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
-        embeddings = ggml_add(ctx0, ggml_repeat(ctx0, model.mm_0_b, embeddings), embeddings);
+        embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
 
         embeddings = ggml_gelu(ctx0, embeddings);
 
         embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
-        embeddings = ggml_add(ctx0, ggml_repeat(ctx0, model.mm_2_b, embeddings), embeddings);
+        embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
     }
 
     // build the graph
@@ -452,7 +452,6 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
 
 // read and create ggml_context containing the tensors and their data
 struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
-
     struct ggml_context * meta = NULL;
 
     struct gguf_init_params params = {
@@ -485,7 +484,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         printf("%s: ftype:        %s\n", __func__, ftype_str.c_str());
         printf("\n");
     }
-
+    const int n_tensors = gguf_get_n_tensors(ctx);
     // kv
     if (verbosity >= 3) {
         const int n_kv = gguf_get_n_kv(ctx);
@@ -499,28 +498,41 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     }
 
     // data
-    size_t ctx_size = 0;
+    size_t buffer_size = 0;
     {
-        const int n_tensors = gguf_get_n_tensors(ctx);
-
         for (int i = 0; i < n_tensors; ++i) {
             const char * name = gguf_get_tensor_name(ctx, i);
             const size_t offset = gguf_get_tensor_offset(ctx, i);
-
             struct ggml_tensor * cur = ggml_get_tensor(meta, name);
-            ctx_size += sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE;
             size_t tensor_size = ggml_nbytes(cur);
-            size_t padded_size = ggml_nbytes_pad(cur);
-            ctx_size += padded_size;
+            buffer_size += tensor_size;
             if (verbosity >= 3) {
-                printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, padded_size=%zu, offset=%zu\n", __func__, i,
-                       ggml_n_dims(cur), cur->name, tensor_size, padded_size, offset);
+                printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu\n", __func__, i,
+                       ggml_n_dims(cur), cur->name, tensor_size, offset);
             }
         }
     }
 
+    buffer_size += n_tensors * 128 /* CLIP PADDING */;
+
     clip_ctx * new_clip = new clip_ctx;
 
+#ifdef GGML_USE_CUBLAS
+    new_clip->backend = ggml_backend_cuda_init(0);
+    printf("%s: CLIP using CUDA backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_METAL
+    new_clip->backend = ggml_backend_metal_init();
+    printf("%s: CLIP using Metal backend\n", __func__);
+#endif
+
+
+    if (!new_clip->backend) {
+        new_clip->backend = ggml_backend_cpu_init();
+        printf("%s: CLIP using CPU backend\n", __func__);
+    }
+
     // model size and capabilities
     {
         int idx = get_key_idx(ctx, KEY_HAS_TEXT_ENC);
@@ -545,21 +557,24 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             printf("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
             printf("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
             printf("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
-            printf("%s: model size:     %.2f MB\n", __func__, (ctx_size / 1024.0 / 1024.0));
+            printf("%s: model size:     %.2f MB\n", __func__, buffer_size / 1024.0 / 1024.0);
             printf("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
         }
     }
 
+    printf("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, buffer_size / (1024.0 * 1024.0), n_tensors);
+
     // load tensors
     {
+        std::vector<uint8_t> read_buf;
         struct ggml_init_params params = {
-            /*.mem_size =*/ ctx_size,
+            /*.mem_size =*/ (n_tensors + 1) * ggml_tensor_overhead(),
             /*.mem_buffer =*/ NULL,
-            /*.no_alloc =*/ false,
+            /*.no_alloc =*/ true,
         };
 
-        new_clip->ctx = ggml_init(params);
-        if (!new_clip->ctx) {
+        new_clip->ctx_data = ggml_init(params);
+        if (!new_clip->ctx_data) {
             fprintf(stderr, "%s: ggml_init() failed\n", __func__);
             clip_free(new_clip);
             return nullptr;
@@ -572,13 +587,21 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             return nullptr;
         }
 
-        const int n_tensors = gguf_get_n_tensors(ctx);
+        // add tensors to context
         for (int i = 0; i < n_tensors; ++i) {
             const char * name = gguf_get_tensor_name(ctx, i);
             struct ggml_tensor * t = ggml_get_tensor(meta, name);
-            struct ggml_tensor * cur = ggml_dup_tensor(new_clip->ctx, t);
+            struct ggml_tensor * cur = ggml_dup_tensor(new_clip->ctx_data, t);
             ggml_set_name(cur, name);
+        }
 
+        // alloc memory and offload data
+        new_clip->params_buffer = ggml_backend_alloc_buffer(new_clip->backend, buffer_size);
+        ggml_allocr* alloc = ggml_allocr_new_from_buffer(new_clip->params_buffer);
+        for (int i = 0; i < n_tensors; ++i) {
+            const char * name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor * cur = ggml_get_tensor(new_clip->ctx_data, name);
+            ggml_allocr_alloc(alloc, cur);
             const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
             fin.seekg(offset, std::ios::beg);
             if (!fin) {
@@ -586,10 +609,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
                 clip_free(new_clip);
                 return nullptr;
             }
-
-            fin.read(reinterpret_cast<char *>(cur->data), ggml_nbytes(t));
+            int num_bytes = ggml_nbytes(cur);
+            if (ggml_backend_buffer_is_host(new_clip->params_buffer)) {
+                // for the CPU and Metal backend, we can read directly into the tensor
+                fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
+            } else {
+                // read into a temporary buffer first, then copy to device memory
+                read_buf.resize(num_bytes);
+                fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
+                ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
+            }
         }
-
+        ggml_allocr_free(alloc);
         fin.close();
     }
 
@@ -598,20 +629,20 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         // load vision model
         auto & vision_model = new_clip->vision_model;
         auto & hparams = vision_model.hparams;
-        hparams.hidden_size = get_u32(ctx, format(KEY_N_EMBD, "vision"));
-        hparams.n_head = get_u32(ctx, format(KEY_N_HEAD, "vision"));
+        hparams.hidden_size    = get_u32(ctx, format(KEY_N_EMBD, "vision"));
+        hparams.n_head         = get_u32(ctx, format(KEY_N_HEAD, "vision"));
         hparams.n_intermediate = get_u32(ctx, format(KEY_N_FF, "vision"));
-        hparams.n_layer = get_u32(ctx, format(KEY_N_BLOCK, "vision"));
-        hparams.image_size = get_u32(ctx, KEY_IMAGE_SIZE);
-        hparams.patch_size = get_u32(ctx, KEY_PATCH_SIZE);
+        hparams.n_layer        = get_u32(ctx, format(KEY_N_BLOCK, "vision"));
+        hparams.image_size     = get_u32(ctx, KEY_IMAGE_SIZE);
+        hparams.patch_size     = get_u32(ctx, KEY_PATCH_SIZE);
         hparams.projection_dim = get_u32(ctx, format(KEY_PROJ_DIM, "vision"));
-        hparams.eps = get_f32(ctx, format(KEY_LAYER_NORM_EPS, "vision"));
+        hparams.eps            = get_f32(ctx, format(KEY_LAYER_NORM_EPS, "vision"));
 
         int idx_mean = get_key_idx(ctx, KEY_IMAGE_MEAN);
-        int idx_std = get_key_idx(ctx, KEY_IMAGE_STD);
+        int idx_std  = get_key_idx(ctx, KEY_IMAGE_STD);
         for (int i = 0; i < 3; ++i) {
             new_clip->image_mean[i] = *((const float *)gguf_get_arr_data(ctx, idx_mean));
-            new_clip->image_std[i] = *((const float *)gguf_get_arr_data(ctx, idx_std));
+            new_clip->image_std[i]  = *((const float *)gguf_get_arr_data(ctx, idx_std));
         }
 
         if (verbosity >= 2) {
@@ -625,35 +656,35 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             printf("v_n_layer          %d\n", hparams.n_layer);
         }
 
-        vision_model.patch_embeddings = get_tensor(new_clip->ctx, TN_PATCH_EMBD);
-        vision_model.class_embedding = get_tensor(new_clip->ctx, TN_CLASS_EMBD);
-        vision_model.position_embeddings = get_tensor(new_clip->ctx, format(TN_POS_EMBD, "v"));
-        vision_model.pre_ln_w = get_tensor(new_clip->ctx, format(TN_LN_PRE, "v", "weight"));
-        vision_model.pre_ln_b = get_tensor(new_clip->ctx, format(TN_LN_PRE, "v", "bias"));
-        vision_model.mm_0_w = get_tensor(new_clip->ctx, format(TN_LLAVA_PROJ, 0, "weight"));
-        vision_model.mm_0_b = get_tensor(new_clip->ctx, format(TN_LLAVA_PROJ, 0, "bias"));
-        vision_model.mm_2_w = get_tensor(new_clip->ctx, format(TN_LLAVA_PROJ, 2, "weight"));
-        vision_model.mm_2_b = get_tensor(new_clip->ctx, format(TN_LLAVA_PROJ, 2, "bias"));
+        vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
+        vision_model.class_embedding     = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
+        vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
+        vision_model.pre_ln_w            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
+        vision_model.pre_ln_b            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
+        vision_model.mm_0_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
+        vision_model.mm_0_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
+        vision_model.mm_2_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
+        vision_model.mm_2_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
 
         vision_model.layers.resize(hparams.n_layer);
         for (int il = 0; il < hparams.n_layer; ++il) {
             auto & layer = vision_model.layers[il];
-            layer.k_w = get_tensor(new_clip->ctx, format(TN_ATTN_K, "v", il, "weight"));
-            layer.q_w = get_tensor(new_clip->ctx, format(TN_ATTN_Q, "v", il, "weight"));
-            layer.v_w = get_tensor(new_clip->ctx, format(TN_ATTN_V, "v", il, "weight"));
-            layer.o_w = get_tensor(new_clip->ctx, format(TN_ATTN_OUTPUT, "v", il, "weight"));
-            layer.ln_1_w = get_tensor(new_clip->ctx, format(TN_LN_1, "v", il, "weight"));
-            layer.ln_2_w = get_tensor(new_clip->ctx, format(TN_LN_2, "v", il, "weight"));
-            layer.ff_i_w = get_tensor(new_clip->ctx, format(TN_FFN_DOWN, "v", il, "weight"));
-            layer.ff_o_w = get_tensor(new_clip->ctx, format(TN_FFN_UP, "v", il, "weight"));
-            layer.k_b = get_tensor(new_clip->ctx, format(TN_ATTN_K, "v", il, "bias"));
-            layer.q_b = get_tensor(new_clip->ctx, format(TN_ATTN_Q, "v", il, "bias"));
-            layer.v_b = get_tensor(new_clip->ctx, format(TN_ATTN_V, "v", il, "bias"));
-            layer.o_b = get_tensor(new_clip->ctx, format(TN_ATTN_OUTPUT, "v", il, "bias"));
-            layer.ln_1_b = get_tensor(new_clip->ctx, format(TN_LN_1, "v", il, "bias"));
-            layer.ln_2_b = get_tensor(new_clip->ctx, format(TN_LN_2, "v", il, "bias"));
-            layer.ff_i_b = get_tensor(new_clip->ctx, format(TN_FFN_DOWN, "v", il, "bias"));
-            layer.ff_o_b = get_tensor(new_clip->ctx, format(TN_FFN_UP, "v", il, "bias"));
+            layer.k_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_K,      "v", il, "weight"));
+            layer.q_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_Q,      "v", il, "weight"));
+            layer.v_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_V,      "v", il, "weight"));
+            layer.o_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_OUTPUT, "v", il, "weight"));
+            layer.ln_1_w = get_tensor(new_clip->ctx_data, format(TN_LN_1,        "v", il, "weight"));
+            layer.ln_2_w = get_tensor(new_clip->ctx_data, format(TN_LN_2,        "v", il, "weight"));
+            layer.ff_i_w = get_tensor(new_clip->ctx_data, format(TN_FFN_DOWN,    "v", il, "weight"));
+            layer.ff_o_w = get_tensor(new_clip->ctx_data, format(TN_FFN_UP,      "v", il, "weight"));
+            layer.k_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_K,      "v", il, "bias"));
+            layer.q_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_Q,      "v", il, "bias"));
+            layer.v_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_V,      "v", il, "bias"));
+            layer.o_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_OUTPUT, "v", il, "bias"));
+            layer.ln_1_b = get_tensor(new_clip->ctx_data, format(TN_LN_1,        "v", il, "bias"));
+            layer.ln_2_b = get_tensor(new_clip->ctx_data, format(TN_LN_2,        "v", il, "bias"));
+            layer.ff_i_b = get_tensor(new_clip->ctx_data, format(TN_FFN_DOWN,    "v", il, "bias"));
+            layer.ff_o_b = get_tensor(new_clip->ctx_data, format(TN_FFN_UP,      "v", il, "bias"));
         }
     }
 
@@ -661,45 +692,45 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 
     new_clip->ctx_gguf = ctx;
 
-// measure mem requirement and allocate
+    // measure mem requirement and allocate
     {
-        static const size_t tensor_alignment = 32;
-        new_clip->buf_compute.resize(ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead());
-        new_clip->alloc = ggml_allocr_new_measure(tensor_alignment);
+        new_clip->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead());
+        new_clip->compute_alloc = ggml_allocr_new_measure_from_backend(new_clip->backend);
         clip_image_f32_batch batch;
         batch.size = 1;
         ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
-        size_t alloc_size = ggml_allocr_alloc_graph(new_clip->alloc, gf) + tensor_alignment;
-        ggml_allocr_free(new_clip->alloc);
-        new_clip->buf_alloc.resize(alloc_size);
-        new_clip->alloc = ggml_allocr_new(new_clip->buf_alloc.data, new_clip->buf_alloc.size, tensor_alignment);
+        size_t compute_memory_buffer_size = ggml_allocr_alloc_graph(new_clip->compute_alloc, gf);
+        ggml_allocr_free(new_clip->compute_alloc);
+        new_clip->compute_buffer = ggml_backend_alloc_buffer(new_clip->backend, compute_memory_buffer_size);
+        new_clip->compute_alloc = ggml_allocr_new_from_buffer(new_clip->compute_buffer);
 
-        printf("%s: total allocated memory: %.2f MB\n", __func__, (new_clip->buf_compute.size + alloc_size)/1024.0/1024.0);
+        printf("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
     }
 
     return new_clip;
 }
 
-clip_image_u8 * make_clip_image_u8() {
-    auto img = new clip_image_u8();
-    return img;
+struct clip_image_u8 * clip_image_u8_init() {
+    return new clip_image_u8();
 }
-clip_image_f32 * make_clip_image_f32() { return new clip_image_f32(); }
 
-void clip_image_u8_free(clip_image_u8 * img) { if (img->data) { delete[] img->data; } delete img; }
-void clip_image_f32_free(clip_image_f32 * img) { if (img->data) { delete[] img->data; } delete img; }
+struct clip_image_f32 * clip_image_f32_init() {
+    return new clip_image_f32();
+}
+
+void clip_image_u8_free (struct clip_image_u8  * img) { delete img; }
+void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
 
 static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) {
     img->nx = nx;
     img->ny = ny;
-    img->size = nx * ny * 3;
-    img->data = new uint8_t[img->size]();
-    memcpy(img->data, data, img->size);
+    img->buf.resize(3 * nx * ny);
+    memcpy(img->buf.data(), data, img->buf.size());
 }
 
 bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
     int nx, ny, nc;
-    auto data = stbi_load(fname, &nx, &ny, &nc, 3);
+    auto * data = stbi_load(fname, &nx, &ny, &nc, 3);
     if (!data) {
         fprintf(stderr, "%s: failed to load image '%s'\n", __func__, fname);
         return false;
@@ -711,7 +742,7 @@ bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
 
 bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img) {
     int nx, ny, nc;
-    auto data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
+    auto * data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
     if (!data) {
         fprintf(stderr, "%s: failed to decode image bytes\n", __func__);
         return false;
@@ -723,7 +754,7 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
 
 // normalize: x = (x - mean) / std
 // TODO: implement bicubic interpolation instead of linear.
-bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32 * res, const bool pad2square) {
+bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32 * res, const bool pad2square) {
     if (!ctx->has_vision_encoder) {
         printf("This gguf file seems to have no vision encoder\n");
         return false;
@@ -732,18 +763,17 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
     // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
     // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
 
-    clip_image_u8 * temp = make_clip_image_u8(); // we will keep the input image data here temporarily
+    clip_image_u8 * temp = clip_image_u8_init(); // we will keep the input image data here temporarily
     if (pad2square && img->nx != img->ny) {
         int longer_side = std::max(img->nx, img->ny);
         temp->nx = longer_side;
         temp->ny = longer_side;
-        temp->size = 3 * longer_side * longer_side;
-        temp->data = new uint8_t[temp->size]();
-        uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA
+        temp->buf.resize(3 * longer_side * longer_side);
+        const uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA
 
         // fill with background color
-        for (size_t i = 0; i < temp->size; i++) {
-            temp->data[i] = bc[i % 3];
+        for (size_t i = 0; i < temp->buf.size(); i++) {
+            temp->buf[i] = bc[i % 3];
         }
 
         // copy from the input image
@@ -751,17 +781,16 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
             for (int x = 0; x < img->nx; x++) {
                 const int i = 3 * (y * img->nx + x);
                 const int j = 3 * (y * temp->nx + x);
-                temp->data[j] = img->data[i];
-                temp->data[j+1] = img->data[i+1];
-                temp->data[j+2] = img->data[i+2];
+                temp->buf[j]   = img->buf[i];
+                temp->buf[j+1] = img->buf[i+1];
+                temp->buf[j+2] = img->buf[i+2];
             }
         }
     } else {
-        temp->nx   = img->nx;
-        temp->ny   = img->ny;
-        temp->size = img->size;
-        temp->data = new uint8_t[temp->size]();
-        memcpy(&temp->data[0], &img->data[0], temp->size); // copy
+        temp->nx = img->nx;
+        temp->ny = img->ny;
+        temp->buf.resize(img->buf.size());
+        memcpy(temp->buf.data(), img->buf.data(), temp->buf.size());
     }
 
     const int nx = temp->nx;
@@ -772,8 +801,7 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
 
     res->nx = nx2;
     res->ny = ny2;
-    res->size = 3 * nx2 * ny2;
-    res->data = new float[res->size]();
+    res->buf.resize(3 * nx2 * ny2);
 
     const float scale = std::max(nx, ny) / (float)ctx->vision_model.hparams.image_size;
 
@@ -804,10 +832,10 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
                 const int j10 = 3 * (y1 * nx + x0) + c;
                 const int j11 = 3 * (y1 * nx + x1) + c;
 
-                const float v00 = temp->data[j00];
-                const float v01 = temp->data[j01];
-                const float v10 = temp->data[j10];
-                const float v11 = temp->data[j11];
+                const float v00 = temp->buf[j00];
+                const float v01 = temp->buf[j01];
+                const float v10 = temp->buf[j10];
+                const float v11 = temp->buf[j11];
 
                 const float v0 = v00 * (1.0f - dx) + v01 * dx;
                 const float v1 = v10 * (1.0f - dx) + v11 * dx;
@@ -818,7 +846,7 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
 
                 const int i = 3 * (y * nx3 + x) + c;
 
-                res->data[i] = ((float(v2) / 255.0f) - m3[c]) / s3[c];
+                res->buf[i] = ((float(v2) / 255.0f) - m3[c]) / s3[c];
             }
         }
     }
@@ -828,12 +856,13 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
 }
 
 void clip_free(clip_ctx * ctx) {
-    ggml_free(ctx->ctx);
+    ggml_free(ctx->ctx_data);
     gguf_free(ctx->ctx_gguf);
+
     delete ctx;
 }
 
-bool clip_image_encode(const clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
+bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
     if (!ctx->has_vision_encoder) {
         printf("This gguf file seems to have no vision encoder\n");
         return false;
@@ -845,8 +874,7 @@ bool clip_image_encode(const clip_ctx * ctx, const int n_threads, clip_image_f32
     return clip_image_batch_encode(ctx, n_threads, &imgs, vec);
 }
 
-bool clip_image_batch_encode(const clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {
-
+bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {
     if (!ctx->has_vision_encoder) {
         printf("This gguf file seems to have no vision encoder\n");
         return false;
@@ -858,29 +886,29 @@ bool clip_image_batch_encode(const clip_ctx * ctx, const int n_threads, const cl
     }
 
     // reset alloc buffer to clean the memory from previous invocations
-    ggml_allocr_reset(ctx->alloc);
+    ggml_allocr_reset(ctx->compute_alloc);
 
     // build the inference graph
     ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
-    ggml_allocr_alloc_graph(ctx->alloc, gf);
+    ggml_allocr_alloc_graph(ctx->compute_alloc, gf);
 
-    struct ggml_cplan plan = ggml_graph_plan(gf, n_threads);
-    if (plan.work_size > 0) {
-        plan.work_data = (uint8_t *)malloc(plan.work_size);
+    if (ggml_backend_is_cpu(ctx->backend)) {
+        ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
     }
 
-    ggml_graph_compute(gf, &plan);
+#ifdef GGML_USE_METAL
+    if (ggml_backend_is_metal(ctx->backend)) {
+        ggml_backend_metal_set_n_cb(ctx->backend, n_threads);
+    }
+#endif
+
+    ggml_backend_graph_compute(ctx->backend, gf);
 
     // the last node is the embedding tensor
-struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 1];
 
     // copy the embeddings to the location passed by the user
-    memcpy(vec, ggml_get_data_f32(embeddings), ggml_nbytes(embeddings));
-
-    if (plan.work_size > 0) {
-        free(plan.work_data);
-    }
-
+    ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
     return true;
 }
 
@@ -888,32 +916,15 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
 
     ggml_type type = GGML_TYPE_Q4_1;
 
-    switch (itype) {
-    case 2:
-        type = GGML_TYPE_Q4_0;
-        break;
-    case 3:
-        type = GGML_TYPE_Q4_1;
-        break;
-    case 6:
-        type = GGML_TYPE_Q5_0;
-        break;
-    case 7:
-        type = GGML_TYPE_Q5_1;
-        break;
-    case 8:
-        type = GGML_TYPE_Q8_0;
-        break;
-    default:
-        fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype);
-        return false;
-    };
+    assert(itype < GGML_TYPE_COUNT);
+    type = static_cast<ggml_type>(itype);
+
+    auto * ctx_clip = clip_model_load(fname_inp, 2);
 
-    auto ctx_clip = clip_model_load(fname_inp, 2);
     const auto & ctx_src = ctx_clip->ctx_gguf;
-    const auto & ctx_data = ctx_clip->ctx;
+    const auto & ctx_data = ctx_clip->ctx_data;
 
-    auto ctx_out = gguf_init_empty();
+    auto * ctx_out = gguf_init_empty();
     gguf_set_kv(ctx_out, ctx_src);
     gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION);
     gguf_set_val_u32(ctx_out, "general.file_type", itype);
@@ -966,6 +977,10 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
 
         if (quantize) {
             new_type = type;
+            if (new_type >= GGML_TYPE_Q2_K && name.find("embd") != std::string::npos) {
+                new_type = GGML_TYPE_Q8_0; // ggml_get_rows needs non K type
+                // fprintf(stderr, "%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
+            }
             const size_t n_elms = ggml_nelements(cur);
             float * f32_data;
 
@@ -1010,6 +1025,21 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
                 case GGML_TYPE_Q8_0: {
                     new_size = ggml_quantize_q8_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
                 } break;
+                case GGML_TYPE_Q2_K: {
+                    new_size = ggml_quantize_q2_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q3_K: {
+                    new_size = ggml_quantize_q3_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q4_K: {
+                    new_size = ggml_quantize_q4_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q5_K: {
+                    new_size = ggml_quantize_q5_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q6_K: {
+                    new_size = ggml_quantize_q6_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
                 default: {
                     fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, new_type);
                     return false;
@@ -1051,8 +1081,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     gguf_free(ctx_out);
 
     {
-        printf("%s: original size  = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
-        printf("%s: quantized size  = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
+        printf("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
+        printf("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
 
         int64_t sum_all = 0;
         for (size_t i = 0; i < hist_all.size(); ++i) {

examples/llava/clip.h

@@ -35,31 +35,14 @@ struct clip_vision_hparams {
     float eps;
 };
 
-/** load mmproj model */
-CLIP_API struct clip_ctx * clip_model_load(const char * fname, const int verbosity);
-/** free mmproj model */
+CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
+
 CLIP_API void clip_free(struct clip_ctx * ctx);
 
-size_t clip_embd_nbytes(const struct clip_ctx * ctx);
-int clip_n_patches(const struct clip_ctx * ctx);
-int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
 
-// RGB uint8 image
-struct clip_image_u8 {
-    int nx;
-    int ny;
-    uint8_t * data = NULL;
-    size_t size;
-};
-
-// RGB float32 image (NHWC)
-// Memory layout: RGBRGBRGB...
-struct clip_image_f32 {
-    int nx;
-    int ny;
-    float * data = NULL;
-    size_t size;
-};
+CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
 struct clip_image_u8_batch {
     struct clip_image_u8 * data;
@@ -71,21 +54,22 @@ struct clip_image_f32_batch {
     size_t size;
 };
 
-struct clip_image_u8 * make_clip_image_u8();
-struct clip_image_f32 * make_clip_image_f32();
-CLIP_API void clip_image_u8_free(clip_image_u8 * img);
-CLIP_API void clip_image_f32_free(clip_image_f32 * img);
+CLIP_API struct clip_image_u8  * clip_image_u8_init ();
+CLIP_API struct clip_image_f32 * clip_image_f32_init();
+
+CLIP_API void clip_image_u8_free (struct clip_image_u8 * img);
+CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
+
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
+
 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */
 CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
 
-bool clip_image_preprocess(const struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, const bool pad2square);
-bool clip_image_encode(const struct clip_ctx * ctx, const int n_threads, struct clip_image_f32 * img, float * vec);
+CLIP_API bool clip_image_preprocess  (struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, bool pad2square);
+CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
+CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
 
-bool clip_image_batch_encode(const struct clip_ctx * ctx, const int n_threads, const struct clip_image_f32_batch * imgs,
-                             float * vec);
-
-bool clip_model_quantize(const char * fname_inp, const char * fname_out, const int itype);
+CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
 
 #ifdef __cplusplus
 }

examples/llava/llava-cli.cpp

@@ -39,73 +39,11 @@ static bool eval_string(struct llama_context * ctx_llama, const char* str, int n
     return true;
 }
 
-// TODO: use common/sampling.h
-static llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
-    auto & sparams = params.sparams;
-
-    // out of user input, sample next token
-    const float   temp      = sparams.temp;
-    const int32_t top_k     = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : sparams.top_k;
-    const float   top_p     = sparams.top_p;
-    const float   tfs_z     = sparams.tfs_z;
-    const float   typical_p = sparams.typical_p;
-    // const int32_t repeat_last_n   = sparams.repeat_last_n < 0 ? n_ctx : sparams.repeat_last_n;
-    // const float   repeat_penalty  = sparams.repeat_penalty;
-    // const float   alpha_presence  = sparams.presence_penalty;
-    // const float   alpha_frequency = sparams.frequency_penalty;
-    const int     mirostat     = sparams.mirostat;
-    const float   mirostat_tau = sparams.mirostat_tau;
-    const float   mirostat_eta = sparams.mirostat_eta;
-    // const bool    penalize_nl     = sparams.penalize_nl;
-
-    llama_token id = 0;
-    {
-        auto logits  = llama_get_logits(ctx_llama);
-        auto n_vocab = llama_n_vocab(llama_get_model(ctx_llama));
-
-        // Apply params.logit_bias map
-        for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) {
-            logits[it->first] += it->second;
-        }
-
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-        }
-
-        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-        if (temp <= 0) {
-              // Greedy sampling
-            id = llama_sample_token_greedy(ctx_llama, &candidates_p);
-        } else {
-            if (mirostat == 1) {
-                static float mirostat_mu = 2.0f * mirostat_tau;
-                const  int mirostat_m    = 100;
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token_mirostat(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
-            } else if (mirostat == 2) {
-                static float mirostat_mu = 2.0f * mirostat_tau;
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token_mirostat_v2(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-            } else {
-                  // Temperature sampling
-                llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1);
-                llama_sample_tail_free(ctx_llama, &candidates_p, tfs_z, 1);
-                llama_sample_typical(ctx_llama, &candidates_p, typical_p, 1);
-                llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1);
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token(ctx_llama, &candidates_p);
-            }
-        }
-    }
-
-    return id;
-}
-
-static const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) {
-    int id = sample_id(ctx_llama, params);
+static const char * sample(struct llama_sampling_context * ctx_sampling,
+                           struct llama_context * ctx_llama,
+                           int * n_past) {
+    const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
+    llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
     static std::string ret;
     if (id == llama_token_eos(llama_get_model(ctx_llama))) {
         ret = "</s>";
@@ -174,8 +112,8 @@ struct llava_context {
 };
 
 static void show_additional_info(int /*argc*/, char ** argv) {
-    printf("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    printf("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+    fprintf(stderr, "\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    fprintf(stderr, "  note: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
 static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params) {
@@ -185,7 +123,7 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
     auto prompt = params->prompt;
     if (prompt_contains_image(prompt)) {
         if (!params->image.empty()) {
-            printf("using base64 encoded image instead of command line image path\n");
+            fprintf(stderr, "using base64 encoded image instead of command line image path\n");
         }
         embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt);
         if (!embed) {
@@ -217,16 +155,19 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
 
     // generate the response
 
-    printf("\n");
+    fprintf(stderr, "\n");
+
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
 
     for (int i = 0; i < max_tgt_len; i++) {
-        const char * tmp = sample(ctx_llava->ctx_llama, *params, &n_past);
+        const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
         if (strcmp(tmp, "</s>") == 0) break;
 
         printf("%s", tmp);
         fflush(stdout);
     }
 
+    llama_sampling_free(ctx_sampling);
     printf("\n");
 }
 
@@ -302,6 +243,9 @@ int main(int argc, char ** argv) {
     }
 
     auto image_embed = load_image(ctx_llava, &params);
+    if (!image_embed) {
+        return 1;
+    }
 
     // process the prompt
     process_prompt(ctx_llava, image_embed, &params, params.prompt);

examples/llava/llava.cpp

@@ -10,7 +10,7 @@
 #include "base64.hpp"
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
-    clip_image_f32 * img_res = make_clip_image_f32();
+    clip_image_f32 * img_res = clip_image_f32_init();
     if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) {
         fprintf(stderr, "%s: unable to preprocess image\n", __func__);
         clip_image_f32_free(img_res);
@@ -86,7 +86,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
 }
 
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
-    clip_image_u8 * img = make_clip_image_u8();
+    clip_image_u8 * img = clip_image_u8_init();
     if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
         clip_image_u8_free(img);
         fprintf(stderr, "%s: can't load image from bytes, is it a valid image?", __func__);

examples/lookup/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET lookup)
+add_executable(${TARGET} lookup.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/lookup/README.md

@@ -0,0 +1,13 @@
+# llama.cpp/examples/lookup
+
+Demonstration of Prompt Lookup Decoding
+
+https://github.com/apoorvumang/prompt-lookup-decoding
+
+The key parameters for lookup decoding are `ngram_min`, `ngram_max` and `n_draft`. The first two determine the size of the ngrams to search for in the prompt for a match. The latter specifies how many subsequent tokens to draft if a match is found.
+
+More info:
+
+https://github.com/ggerganov/llama.cpp/pull/4484
+https://github.com/ggerganov/llama.cpp/issues/4226
+

examples/lookup/lookup.cpp

@@ -0,0 +1,230 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv){
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+
+    // max/min n-grams size to search for in prompt
+    const int ngram_max = 4;
+    const int ngram_min = 1;
+
+    // length of the candidate / draft sequence, if match is found
+    const int n_draft = params.n_draft;
+
+    const bool dump_kv_cache = params.dump_kv_cache;
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("lookup", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // init llama.cpp
+    llama_backend_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+
+    // tokenize the prompt
+    const bool add_bos = llama_should_add_bos_token(model);
+    LOG("add_bos tgt: %d\n", add_bos);
+
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+
+    const int max_context_size     = llama_n_ctx(ctx);
+    const int max_tokens_list_size = max_context_size - 4;
+
+    if ((int) inp.size() > max_tokens_list_size) {
+        fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
+        return 1;
+    }
+
+    fprintf(stderr, "\n\n");
+
+    for (auto id : inp) {
+        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
+    }
+
+    fflush(stderr);
+
+    const int n_input = inp.size();
+
+    const auto t_enc_start = ggml_time_us();
+
+    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
+    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
+
+    const auto t_enc_end = ggml_time_us();
+
+    int n_predict = 0;
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    int n_past = inp.size();
+
+    bool has_eos = false;
+
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
+
+    std::vector<llama_token> draft;
+
+    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);
+
+    // debug
+    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, 1);
+
+    const auto t_dec_start = ggml_time_us();
+
+    while (true) {
+        // debug
+        if (dump_kv_cache) {
+            llama_kv_cache_view_update(ctx, &kvc_view);
+            dump_kv_cache_view_seqs(kvc_view, 40);
+        }
+
+        // print current draft sequence
+        LOG("drafted %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, draft).c_str());
+
+        int i_dft = 0;
+        while (true) {
+            // sample from the target model
+            llama_token id = llama_sampling_sample(ctx_sampling, ctx, NULL, i_dft);
+
+            llama_sampling_accept(ctx_sampling, ctx, id, true);
+
+            const std::string token_str = llama_token_to_piece(ctx, id);
+
+            if (!params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+
+            if (id == llama_token_eos(model)) {
+                has_eos = true;
+            }
+
+            ++n_predict;
+
+            // check if the target token matches the draft
+            if (i_dft < (int) draft.size() && id == draft[i_dft]) {
+                LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
+                ++n_accept;
+                ++n_past;
+                ++i_dft;
+                inp.push_back(id);
+
+                if (params.use_color) {
+                    // color accepted draft token
+                    printf("\033[34m%s\033[0m", token_str.c_str());
+                    fflush(stdout);
+                }
+                continue;
+            }
+
+            if (params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+            fflush(stdout);
+
+
+            LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
+
+            draft.clear();
+            draft.push_back(id);
+            inp.push_back(id);
+            break;
+        }
+
+        if ((params.n_predict > 0 && n_predict > params.n_predict) || has_eos) {
+            break;
+        }
+
+        // KV cache management
+        // clean the cache of draft tokens that weren't accepted
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
+        llama_batch_clear(batch_tgt);
+        llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
+
+        // generate n_pred tokens through prompt lookup
+        auto prompt_lookup = [&]() -> void {
+            int inp_size = inp.size();
+            for (int ngram_size = ngram_max ; ngram_size > ngram_min; --ngram_size){
+                const llama_token * ngram = &inp[inp_size - ngram_size];
+
+                for (int i = 0; i <= (int) inp_size - (ngram_size * 2); ++i) {
+                    bool match = true;
+                    for (int j = 0; j < ngram_size; ++j) {
+                        if (inp[i + j] != ngram[j]) {
+                            match = false;
+                            break;
+                        }
+                    }
+
+                    if (match) {
+                        const int startIdx = i + ngram_size;
+                        const int endIdx = startIdx + n_draft;
+                        if (endIdx < inp_size) {
+                            for (int j = startIdx; j < endIdx; ++j) {
+                                LOG(" - draft candidate %d: %d\n", j, inp[j]);
+                                draft.push_back(inp[j]);
+                                llama_batch_add(batch_tgt, inp[j], n_past + (j - startIdx) + 1, { 0 }, true);
+                                ++n_drafted;
+                            }
+                            return;
+                        }
+                    }
+                }
+            }
+            return;
+        };
+
+        prompt_lookup();
+
+        llama_decode(ctx, batch_tgt);
+        ++n_past;
+
+        draft.erase(draft.begin());
+    }
+
+    auto t_dec_end = ggml_time_us();
+
+    LOG_TEE("\n\n");
+
+    LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
+    LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
+
+    LOG_TEE("\n");
+    LOG_TEE("n_draft   = %d\n", n_draft);
+    LOG_TEE("n_predict = %d\n", n_predict);
+    LOG_TEE("n_drafted = %d\n", n_drafted);
+    LOG_TEE("n_accept  = %d\n", n_accept);
+    LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
+
+    LOG_TEE("\ntarget:\n");
+    llama_print_timings(ctx);
+
+    llama_sampling_free(ctx_sampling);
+    llama_batch_free(batch_tgt);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}

examples/main-cmake-pkg/CMakeLists.txt

@@ -7,28 +7,13 @@ find_package(Llama 0.0.1 REQUIRED)
 # Bake common functionality in with target. Because applications
 # using the relocatable Llama package should be outside of the
 # source tree, main-cmake-pkg pretends the dependencies are built-in.
-
 set(_common_path "${CMAKE_CURRENT_LIST_DIR}/../../common")
-add_library(common OBJECT
-    ${_common_path}/common.h
-    ${_common_path}/common.cpp
-    ${_common_path}/console.h
-    ${_common_path}/console.cpp
-    ${_common_path}/grammar-parser.h
-    ${_common_path}/grammar-parser.cpp
-    ${_common_path}/sampling.h
-    ${_common_path}/sampling.cpp
-    )
-
-# WARNING: because build-info.h is auto-generated, it will only
-# be available after the user has built the llama.cpp sources.
-#
-configure_file(${_common_path}/../build-info.h
-    ${CMAKE_CURRENT_BINARY_DIR}/build-info.h
-    COPYONLY)
-
-target_include_directories(common PUBLIC ${LLAMA_INCLUDE_DIR}
-    ${CMAKE_CURRENT_BINARY_DIR})
+add_library(common OBJECT)
+file(GLOB _common_files
+    "${_common_path}/*.h"
+    "${_common_path}/*.cpp"
+)
+target_sources(common PRIVATE ${_common_files})
 
 # If the common project was part of "main-cmake-pkg" the transient
 # defines would automatically be attached. Because the common func-

examples/main/main.cpp

@@ -439,6 +439,21 @@ int main(int argc, char ** argv) {
     LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
     LOG_TEE("sampling order: \n%s\n", llama_sampling_order_print(sparams).c_str());
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+
+    // group-attention state
+    // number of grouped KV tokens so far (used only if params.grp_attn_n > 1)
+    int ga_i = 0;
+
+    const int ga_n = params.grp_attn_n;
+    const int ga_w = params.grp_attn_w;
+
+    if (ga_n != 1) {
+        GGML_ASSERT(ga_n > 0                    && "grp_attn_n must be positive");                     // NOLINT
+        GGML_ASSERT(ga_w % ga_n == 0            && "grp_attn_w must be a multiple of grp_attn_n");     // NOLINT
+      //GGML_ASSERT(n_ctx_train % ga_w == 0     && "n_ctx_train must be a multiple of grp_attn_w");    // NOLINT
+      //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
+        LOG_TEE("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
+    }
     LOG_TEE("\n\n");
 
     if (params.interactive) {
@@ -485,7 +500,7 @@ int main(int argc, char ** argv) {
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
         if (!embd.empty()) {
-            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // Note: (n_ctx - 4) here is to match the logic for commandline prompt handling via
             // --prompt or --file which uses the same value.
             int max_embd_size = n_ctx - 4;
 
@@ -500,37 +515,61 @@ int main(int argc, char ** argv) {
                 fflush(stdout);
             }
 
-            // infinite text generation via context swapping
-            // if we run out of context:
-            // - take the n_keep first tokens from the original prompt (via n_past)
-            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
-            if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
-                if (params.n_predict == -2) {
-                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
-                    break;
+            if (ga_n == 1) {
+                // infinite text generation via context shifting
+                // if we run out of context:
+                // - take the n_keep first tokens from the original prompt (via n_past)
+                // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
+                if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
+                    if (params.n_predict == -2) {
+                        LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                        break;
+                    }
+
+                    const int n_left    = n_past - params.n_keep - 1;
+                    const int n_discard = n_left/2;
+
+                    LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                            n_past, n_left, n_ctx, params.n_keep, n_discard);
+
+                    llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                    n_past -= n_discard;
+
+                    if (ctx_guidance) {
+                        n_past_guidance -= n_discard;
+                    }
+
+                    LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+
+                    LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
+
+                    LOG("clear session path\n");
+                    path_session.clear();
                 }
+            } else {
+                // context extension via Self-Extend
+                while (n_past >= ga_i + ga_w) {
+                    const int ib = (ga_n*ga_i)/ga_w;
+                    const int bd = (ga_w/ga_n)*(ga_n - 1);
+                    const int dd = (ga_w/ga_n) - ib*bd - ga_w;
 
-                const int n_left    = n_past - params.n_keep - 1;
-                const int n_discard = n_left/2;
+                    LOG("\n");
+                    LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i, n_past, ib*bd, ga_i + ib*bd, n_past + ib*bd);
+                    LOG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
+                    LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
 
-                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
-                    n_past, n_left, n_ctx, params.n_keep, n_discard);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i,                n_past,              ib*bd);
+                    llama_kv_cache_seq_div  (ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
-                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                    n_past -= bd;
 
-                n_past -= n_discard;
+                    ga_i += ga_w/ga_n;
 
-                if (ctx_guidance) {
-                    n_past_guidance -= n_discard;
+                    LOG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", n_past + bd, n_past, ga_i);
                 }
-
-                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
-
-                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
-
-                LOG("clear session path\n");
-                path_session.clear();
             }
 
             // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
@@ -611,6 +650,10 @@ int main(int argc, char ** argv) {
                 n_past += n_eval;
 
                 LOG("n_past = %d\n", n_past);
+                // Display total tokens alongside total time
+                if (params.n_print > 0 && n_past % params.n_print == 0) {
+                    LOG_TEE("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);
+                }
             }
 
             if (!embd.empty() && !path_session.empty()) {

examples/passkey/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET passkey)
+add_executable(${TARGET} passkey.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/passkey/README.md

@@ -0,0 +1,12 @@
+# llama.cpp/example/passkey
+
+See the following PRs for more info:
+
+- https://github.com/ggerganov/llama.cpp/pull/3856
+- https://github.com/ggerganov/llama.cpp/pull/4810
+
+### Usage
+
+```bash
+make -j && ./passkey ./models/llama-7b-v2/ggml-model-f16.gguf 250
+```

examples/passkey/passkey.cpp

@@ -0,0 +1,296 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (argc == 1 || argv[1][0] == '-') {
+        printf("usage: %s MODEL_PATH N_JUNK N_GRP I_POS SEED\n" , argv[0]);
+        return 1 ;
+    }
+
+    int seed = -1;
+
+    int n_junk = 250; // number of times to repeat the junk text
+    int n_keep = 32;  // number of tokens in the prompt prefix
+    int n_grp  = 1;   // if more than 1 - perform LongLM SelfExtend
+    int i_pos  = -1;  // position of the passkey in the junk text
+
+    if (argc >= 2) {
+        params.model = argv[1];
+    }
+
+    if (argc >= 3) {
+        n_junk = std::stoi(argv[2]);
+    }
+
+    if (argc >= 4) {
+        n_grp = std::stoi(argv[3]);
+    }
+
+    if (argc >= 5) {
+        i_pos = std::stoi(argv[4]);
+    }
+
+    if (argc >= 6) {
+        seed = std::stoi(argv[5]);
+    }
+
+    if (seed == -1) {
+        seed = time(NULL);
+    }
+
+    srand(seed);
+
+    if (i_pos == -1) {
+        i_pos = rand() % n_junk;
+    }
+
+    const std::string prompt_prefix = "There is an important info hidden inside a lot of irrelevant text. Find it and memorize them. I will quiz you about the important information there.";
+    const std::string prompt_suffix = " What is the pass key? The pass key is";
+
+    // generate junk text
+    params.prompt = prompt_prefix;
+
+    const int passkey = rand() % 50000 + 1;
+
+    for (int i = 0; i < n_junk; i++) {
+        if (i % n_junk == i_pos) {
+            params.prompt += " The pass key is " + std::to_string(passkey) + ". Remember it. " + std::to_string(passkey) + " is the pass key.";
+        }
+
+        params.prompt += " The grass is green. The sky is blue. The sun is yellow. Here we go. There and back again.";
+    }
+
+    params.prompt += prompt_suffix;
+
+    // init LLM
+
+    llama_backend_init(params.numa);
+
+    // initialize the model
+
+    llama_model_params model_params = llama_model_default_params();
+
+    model_params.n_gpu_layers = 99; // offload all layers to the GPU
+
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    // initialize the context
+
+    llama_context_params ctx_params = llama_context_default_params();
+
+    ctx_params.seed    = seed;
+    ctx_params.n_ctx   = llama_n_ctx_train(model)*n_grp + n_keep;
+    ctx_params.n_batch = 512;
+    ctx_params.n_threads       = params.n_threads;
+    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+
+    GGML_ASSERT(ctx_params.n_batch % n_grp == 0 && "n_batch must be divisible by n_grp");
+
+    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    // tokenize the prompt
+    std::vector<llama_token> tokens_list;
+    tokens_list = ::llama_tokenize(ctx, params.prompt, true);
+
+    // tokenize the prefix and use it as a sink
+    const int n_tokens_prefix = ::llama_tokenize(ctx, prompt_prefix, true).size();
+
+    const int n_tokens_all = tokens_list.size();
+
+    // we leave a margin of 16 tokens for the generated text - it should contain just the passkey
+    const int n_predict = 16;
+
+    // total length of the sequences including the prompt
+    const int n_len = n_tokens_all + n_predict;
+
+    const int n_ctx       = llama_n_ctx(ctx) - n_keep;
+    const int n_kv_req    = llama_n_ctx(ctx);
+    const int n_batch     = ctx_params.n_batch;
+    const int n_batch_grp = ctx_params.n_batch/n_grp;
+
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch);
+
+    // print the prompt token-by-token
+
+    LOG_TEE("\n");
+    LOG_TEE("prefix tokens: %d\n", n_tokens_prefix);
+    LOG_TEE("prompt tokens: %d\n", n_tokens_all);
+    //LOG_TEE("prompt: %s\n", params.prompt.c_str());
+
+    llama_batch batch = llama_batch_init(512, 0, 1);
+
+    int n_past = 0;
+
+    // fill the KV cache
+    for (int i = 0; i < n_ctx; i += n_batch) {
+        if (i > 0 && n_grp > 1) {
+            // if SelfExtend is enabled, we compress the position from the last batch by a factor of n_grp
+            const int ib = i/n_batch - 1;
+            const int bd = n_batch_grp*(n_grp - 1);
+
+            llama_kv_cache_seq_shift(ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_cache_seq_div  (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+
+            n_past -= bd;
+        }
+
+        llama_batch_clear(batch);
+
+        for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
+            llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
+        }
+
+        if (i + n_batch >= n_tokens_all) {
+            batch.logits[batch.n_tokens - 1] = true;
+        }
+
+        if (llama_decode(ctx, batch) != 0) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+
+        LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
+
+        if (i + n_batch >= n_tokens_all) {
+            break;
+        }
+    }
+
+    for (int i = n_ctx; i < n_tokens_all; i += n_batch) {
+        const int n_discard = n_batch;
+
+        LOG_TEE("%s: shifting KV cache with %d\n", __func__, n_discard);
+
+        llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+
+        n_past -= n_discard;
+
+        llama_batch_clear(batch);
+
+        for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
+            llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
+        }
+
+        if (i + n_batch >= n_tokens_all) {
+            batch.logits[batch.n_tokens - 1] = true;
+        }
+
+        if (llama_decode(ctx, batch) != 0) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+
+        LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
+    }
+
+    {
+        const int n_discard = n_past - n_ctx + n_predict;
+
+        if (n_discard > 0) {
+            LOG_TEE("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
+
+            llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+
+            n_past -= n_discard;
+        }
+    }
+
+    LOG_TEE("\n");
+    LOG_TEE("%s: passkey = %d, inserted at position %d / %d (token pos: ~%d)\n", __func__, passkey, i_pos, n_junk, (i_pos * n_tokens_all) / n_junk);
+    LOG_TEE("\n");
+
+    // main loop
+
+    int n_cur    = n_tokens_all;
+    int n_decode = 0;
+
+    LOG_TEE("%s", prompt_suffix.c_str());
+    fflush(stdout);
+
+    const auto t_main_start = ggml_time_us();
+
+    while (n_cur <= n_len) {
+        // sample the next token
+        {
+            auto   n_vocab = llama_n_vocab(model);
+            auto * logits  = llama_get_logits_ith(ctx, batch.n_tokens - 1);
+
+            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 };
+
+            // sample the most likely token
+            const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+
+            // is it an end of stream?
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
+                LOG_TEE("\n");
+
+                break;
+            }
+
+            LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+            fflush(stdout);
+
+            n_decode += 1;
+
+            // prepare the next batch
+            llama_batch_clear(batch);
+
+            // push this new token for next evaluation
+            llama_batch_add(batch, new_token_id, n_past++, { 0 }, true);
+        }
+
+        n_cur += 1;
+
+        // evaluate the current batch with the transformer model
+        if (llama_decode(ctx, batch)) {
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
+            return 1;
+        }
+    }
+
+    LOG_TEE("\n");
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+            __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
+
+    llama_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/pydantic-models-to-grammar-examples.py

@@ -0,0 +1,136 @@
+# Function calling example using pydantic models.
+
+import json
+from enum import Enum
+from typing import Union, Optional
+
+import requests
+from pydantic import BaseModel, Field
+
+import importlib
+from pydantic_models_to_grammar import generate_gbnf_grammar_and_documentation
+
+# Function to get completion on the llama.cpp server with grammar.
+def create_completion(prompt, grammar):
+    headers = {"Content-Type": "application/json"}
+    data = {"prompt": prompt, "grammar": grammar}
+
+    response = requests.post("http://127.0.0.1:8080/completion", headers=headers, json=data)
+    data = response.json()
+
+    print(data["content"])
+    return data["content"]
+
+
+# A function for the agent to send a message to the user.
+class SendMessageToUser(BaseModel):
+    """
+    Send a message to the User.
+    """
+    chain_of_thought: str = Field(..., description="Your chain of thought while sending the message.")
+    message: str = Field(..., description="Message you want to send to the user.")
+
+    def run(self):
+        print(self.message)
+
+
+# Enum for the calculator function.
+class MathOperation(Enum):
+    ADD = "add"
+    SUBTRACT = "subtract"
+    MULTIPLY = "multiply"
+    DIVIDE = "divide"
+
+
+# Very simple calculator tool for the agent.
+class Calculator(BaseModel):
+    """
+    Perform a math operation on two numbers.
+    """
+    number_one: Union[int, float] = Field(..., description="First number.")
+    operation: MathOperation = Field(..., description="Math operation to perform.")
+    number_two: Union[int, float] = Field(..., description="Second number.")
+
+    def run(self):
+        if self.operation == MathOperation.ADD:
+            return self.number_one + self.number_two
+        elif self.operation == MathOperation.SUBTRACT:
+            return self.number_one - self.number_two
+        elif self.operation == MathOperation.MULTIPLY:
+            return self.number_one * self.number_two
+        elif self.operation == MathOperation.DIVIDE:
+            return self.number_one / self.number_two
+        else:
+            raise ValueError("Unknown operation.")
+
+
+# Here the grammar gets generated by passing the available function models to generate_gbnf_grammar_and_documentation function. This also generates a documentation usable by the LLM.
+# pydantic_model_list is the list of pydanitc models
+# outer_object_name is an optional name for an outer object around the actual model object. Like a "function" object with "function_parameters" which contains the actual model object. If None, no outer object will be generated
+# outer_object_content is the name of outer object content.
+# model_prefix is the optional prefix for models in the documentation. (Default="Output Model")
+# fields_prefix is the prefix for the model fields in the documentation. (Default="Output Fields")
+gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation(
+    pydantic_model_list=[SendMessageToUser, Calculator], outer_object_name="function",
+    outer_object_content="function_parameters", model_prefix="Function", fields_prefix="Parameters")
+
+print(gbnf_grammar)
+print(documentation)
+
+system_message = "You are an advanced AI, tasked to assist the user by calling functions in JSON format. The following are the available functions and their parameters and types:\n\n" + documentation
+
+user_message = "What is 42 * 42?"
+prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
+
+text = create_completion(prompt=prompt, grammar=gbnf_grammar)
+# This should output something like this:
+# {
+#     "function": "calculator",
+#     "function_parameters": {
+#         "number_one": 42,
+#         "operation": "multiply",
+#         "number_two": 42
+#     }
+# }
+function_dictionary = json.loads(text)
+if function_dictionary["function"] == "calculator":
+    function_parameters = {**function_dictionary["function_parameters"]}
+
+    print(Calculator(**function_parameters).run())
+    # This should output: 1764
+
+
+# A example structured output based on pydantic models. The LLM will create an entry for a Book database out of an unstructured text.
+class Category(Enum):
+    """
+    The category of the book.
+    """
+    Fiction = "Fiction"
+    NonFiction = "Non-Fiction"
+
+
+class Book(BaseModel):
+    """
+    Represents an entry about a book.
+    """
+    title: str = Field(..., description="Title of the book.")
+    author: str = Field(..., description="Author of the book.")
+    published_year: Optional[int] = Field(..., description="Publishing year of the book.")
+    keywords: list[str] = Field(..., description="A list of keywords.")
+    category: Category = Field(..., description="Category of the book.")
+    summary: str = Field(..., description="Summary of the book.")
+
+
+# We need no additional parameters other than our list of pydantic models.
+gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation([Book])
+
+system_message = "You are an advanced AI, tasked to create a dataset entry in JSON for a Book. The following is the expected output model:\n\n" + documentation
+
+text = """The Feynman Lectures on Physics is a physics textbook based on some lectures by Richard Feynman, a Nobel laureate who has sometimes been called "The Great Explainer". The lectures were presented before undergraduate students at the California Institute of Technology (Caltech), during 1961–1963. The book's co-authors are Feynman, Robert B. Leighton, and Matthew Sands."""
+prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant"
+
+text = create_completion(prompt=prompt, grammar=gbnf_grammar)
+
+json_data = json.loads(text)
+
+print(Book(**json_data))

examples/quantize/quantize.cpp

@@ -18,6 +18,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0,   " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", },
     { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
+    { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
     { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
     { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },

examples/server/CMakeLists.txt

@@ -6,7 +6,7 @@ install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()

examples/server/README.md

@@ -23,6 +23,8 @@ Command line options:
 -   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
 -   `--port`: Set the port to listen. Default: `8080`.
 -   `--path`: path from which to serve static files (default examples/server/public)
+-   `--api-key`: Set an api key for request authorization. By default the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
+-   `--api-key-file`: path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`'s.
 -   `--embedding`: Enable embedding extraction, Default: disabled.
 -   `-np N`, `--parallel N`: Set the number of slots for process requests (default: 1)
 -   `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled)
@@ -109,6 +111,10 @@ node index.js
 ```
 
 ## API Endpoints
+- **GET** `/health`: Returns the current state of the server:
+    - `{"status": "loading model"}` if the model is still being loaded.
+    - `{"status": "error"}` if the model failed to load.
+    - `{"status": "ok"}` if the model is successfully loaded and the server is ready for further requests mentioned below.
 
 -   **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
@@ -148,6 +154,8 @@ node index.js
 
     `frequency_penalty`: Repeat alpha frequency penalty (default: 0.0, 0.0 = disabled);
 
+    `penalty_prompt`: This will replace the `prompt` for the purpose of the penalty evaluation. Can be either `null`, a string or an array of numbers representing tokens (default: `null` = use the original `prompt`).
+
     `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0).
 
     `mirostat_tau`: Set the Mirostat target entropy, parameter tau (default: 5.0).
@@ -164,37 +172,7 @@ node index.js
 
     `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
 
-    `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:` In this case, `[img-12]` will be replaced by the embeddings of the image id 12 in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
-
-    *Result JSON:*
-
-    Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
-
-    `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
-
-    `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
-
-    `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
-
-    `model`: The path to the model loaded with `-m`
-
-    `prompt`: The provided `prompt`
-
-    `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
-
-    `stopped_limit`: Indicating whether the completion stopped because `n_predict` tokens were generated before stop words or EOS was encountered
-
-    `stopped_word`: Indicating whether the completion stopped due to encountering a stopping word from `stop` JSON array provided
-
-    `stopping_word`: The stopping word encountered which stopped the generation (or "" if not stopped due to a stopping word)
-
-    `timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second`
-
-    `tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`)
-
-    `tokens_evaluated`: Number of tokens evaluated in total from the prompt
-
-    `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
+    `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
     `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
 
@@ -202,6 +180,45 @@ node index.js
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
+### Result JSON:
+
+* Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
+
+
+- `completion_probabilities`: An array of token probabilities for each completion. The array's length is `n_predict`. Each item in the array has the following structure:
+
+```
+{
+  "content": "<the token selected by the model>",
+  "probs": [
+    {
+      "prob": float,
+      "tok_str": "<most likely token>"
+    },
+    {
+      "prob": float,
+      "tok_str": "<second most likely tonen>"
+    },
+    ...
+  ]
+},
+```
+Notice that each `probs` is an array of length `n_probs`.
+
+- `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
+- `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
+- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
+- `model`: The path to the model loaded with `-m`
+- `prompt`: The provided `prompt`
+- `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
+- `stopped_limit`: Indicating whether the completion stopped because `n_predict` tokens were generated before stop words or EOS was encountered
+- `stopped_word`: Indicating whether the completion stopped due to encountering a stopping word from `stop` JSON array provided
+- `stopping_word`: The stopping word encountered which stopped the generation (or "" if not stopped due to a stopping word)
+- `timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second`
+- `tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`)
+- `tokens_evaluated`: Number of tokens evaluated in total from the prompt
+- `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
+
 -   **POST** `/tokenize`: Tokenize a given text.
 
     *Options:*
@@ -222,6 +239,8 @@ node index.js
 
     `content`: Set the text to process.
 
+    `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
+
 -   **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
 
     *Options:*

examples/server/completion.js.hpp

@@ -74,355 +74,376 @@ unsigned char completion_js[] = {
   0x6f, 0x6e, 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x27, 0x2c, 0x0a, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x27, 0x41, 0x63, 0x63, 0x65, 0x70, 0x74, 0x27,
   0x3a, 0x20, 0x27, 0x74, 0x65, 0x78, 0x74, 0x2f, 0x65, 0x76, 0x65, 0x6e,
-  0x74, 0x2d, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x27, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x73, 0x69, 0x67,
-  0x6e, 0x61, 0x6c, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c,
-  0x6c, 0x65, 0x72, 0x2e, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x0a,
-  0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20,
-  0x72, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73, 0x65, 0x2e, 0x62, 0x6f, 0x64,
-  0x79, 0x2e, 0x67, 0x65, 0x74, 0x52, 0x65, 0x61, 0x64, 0x65, 0x72, 0x28,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x64,
-  0x65, 0x63, 0x6f, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77,
-  0x20, 0x54, 0x65, 0x78, 0x74, 0x44, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72,
-  0x28, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63,
-  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b,
-  0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f,
-  0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f,
-  0x20, 0x42, 0x75, 0x66, 0x66, 0x65, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20,
-  0x70, 0x61, 0x72, 0x74, 0x69, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x72, 0x65,
-  0x61, 0x64, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20,
-  0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65,
-  0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75,
-  0x65, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c,
-  0x65, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72,
-  0x65, 0x73, 0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69,
-  0x74, 0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x61,
-  0x64, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69,
-  0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x6f,
-  0x6e, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x2f, 0x2f, 0x20, 0x41, 0x64, 0x64, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x6c,
-  0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x64, 0x61, 0x74, 0x61,
-  0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x75, 0x72, 0x72,
-  0x65, 0x6e, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66,
-  0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3d,
-  0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x2b, 0x20,
-  0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72, 0x2e, 0x64, 0x65, 0x63, 0x6f,
-  0x64, 0x65, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x2f, 0x2f, 0x20, 0x43, 0x68, 0x65, 0x63, 0x6b, 0x20, 0x69, 0x66,
-  0x20, 0x74, 0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x63, 0x68,
-  0x61, 0x72, 0x61, 0x63, 0x74, 0x65, 0x72, 0x20, 0x69, 0x73, 0x20, 0x61,
-  0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x0a,
+  0x74, 0x2d, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x27, 0x2c, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x28, 0x70, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x2e, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x20,
+  0x3f, 0x20, 0x7b, 0x27, 0x41, 0x75, 0x74, 0x68, 0x6f, 0x72, 0x69, 0x7a,
+  0x61, 0x74, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x60, 0x42, 0x65, 0x61,
+  0x72, 0x65, 0x72, 0x20, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
+  0x2e, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x7d, 0x60, 0x7d, 0x20,
+  0x3a, 0x20, 0x7b, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x3a,
+  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e,
+  0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x0a, 0x20, 0x20, 0x7d, 0x29,
+  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72,
+  0x65, 0x61, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x73, 0x70,
+  0x6f, 0x6e, 0x73, 0x65, 0x2e, 0x62, 0x6f, 0x64, 0x79, 0x2e, 0x67, 0x65,
+  0x74, 0x52, 0x65, 0x61, 0x64, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64,
+  0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x54, 0x65, 0x78,
+  0x74, 0x44, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a,
+  0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x6c,
+  0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20,
+  0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f, 0x20, 0x42, 0x75, 0x66,
+  0x66, 0x65, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x70, 0x61, 0x72, 0x74,
+  0x69, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x72, 0x65, 0x61, 0x64, 0x20, 0x6c,
+  0x69, 0x6e, 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f,
+  0x6e, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, 0x0a, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x63,
+  0x6f, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c,
+  0x74, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x72, 0x65,
+  0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x61, 0x64, 0x28, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x6f, 0x6e, 0x65, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72,
+  0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x41,
+  0x64, 0x64, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f,
+  0x76, 0x65, 0x72, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x74, 0x6f, 0x20,
+  0x74, 0x68, 0x65, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x20,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x61, 0x74,
+  0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3d, 0x20, 0x6c, 0x65, 0x66,
+  0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x2b, 0x20, 0x64, 0x65, 0x63, 0x6f,
+  0x64, 0x65, 0x72, 0x2e, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x28, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29,
+  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x43, 0x68, 0x65, 0x63, 0x6b, 0x20, 0x69, 0x66, 0x20, 0x74, 0x68, 0x65,
+  0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x72, 0x61, 0x63,
+  0x74, 0x65, 0x72, 0x20, 0x69, 0x73, 0x20, 0x61, 0x20, 0x6c, 0x69, 0x6e,
+  0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x6e, 0x64, 0x73,
+  0x57, 0x69, 0x74, 0x68, 0x4c, 0x69, 0x6e, 0x65, 0x42, 0x72, 0x65, 0x61,
+  0x6b, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2e, 0x65, 0x6e, 0x64,
+  0x73, 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, 0x5c, 0x6e, 0x27, 0x29, 0x3b,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x53,
+  0x70, 0x6c, 0x69, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x74, 0x65, 0x78,
+  0x74, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c,
+  0x69, 0x6e, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2e,
+  0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x5c, 0x6e, 0x27, 0x29, 0x3b,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x49,
+  0x66, 0x20, 0x74, 0x68, 0x65, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x64,
+  0x6f, 0x65, 0x73, 0x6e, 0x27, 0x74, 0x20, 0x65, 0x6e, 0x64, 0x20, 0x77,
+  0x69, 0x74, 0x68, 0x20, 0x61, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62,
+  0x72, 0x65, 0x61, 0x6b, 0x2c, 0x20, 0x74, 0x68, 0x65, 0x6e, 0x20, 0x74,
+  0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x6c, 0x69, 0x6e, 0x65,
+  0x20, 0x69, 0x73, 0x20, 0x69, 0x6e, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65,
+  0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x53, 0x74, 0x6f, 0x72, 0x65, 0x20, 0x69, 0x74, 0x20, 0x69, 0x6e, 0x20,
+  0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x74, 0x6f, 0x20,
+  0x62, 0x65, 0x20, 0x61, 0x64, 0x64, 0x65, 0x64, 0x20, 0x74, 0x6f, 0x20,
+  0x74, 0x68, 0x65, 0x20, 0x6e, 0x65, 0x78, 0x74, 0x20, 0x63, 0x68, 0x75,
+  0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x65, 0x6e,
+  0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x4c, 0x69, 0x6e, 0x65, 0x42, 0x72,
+  0x65, 0x61, 0x6b, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20,
+  0x3d, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x2e, 0x70, 0x6f, 0x70, 0x28,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65,
+  0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x3d,
+  0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f, 0x20, 0x52, 0x65, 0x73, 0x65,
+  0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x69,
+  0x66, 0x20, 0x77, 0x65, 0x20, 0x68, 0x61, 0x76, 0x65, 0x20, 0x61, 0x20,
+  0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x20, 0x61,
+  0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x65, 0x6e, 0x64, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2f, 0x20, 0x50, 0x61, 0x72, 0x73, 0x65, 0x20, 0x61, 0x6c,
+  0x6c, 0x20, 0x73, 0x73, 0x65, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x73,
+  0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x64, 0x64, 0x20, 0x74, 0x68, 0x65,
+  0x6d, 0x20, 0x74, 0x6f, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x4c, 0x69, 0x6e, 0x65,
-  0x42, 0x72, 0x65, 0x61, 0x6b, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x78, 0x74,
-  0x2e, 0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, 0x5c,
-  0x6e, 0x27, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x2f, 0x2f, 0x20, 0x53, 0x70, 0x6c, 0x69, 0x74, 0x20, 0x74, 0x68, 0x65,
-  0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x6c,
-  0x69, 0x6e, 0x65, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c,
-  0x65, 0x74, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x74,
-  0x65, 0x78, 0x74, 0x2e, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x5c,
-  0x6e, 0x27, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x2f, 0x2f, 0x20, 0x49, 0x66, 0x20, 0x74, 0x68, 0x65, 0x20, 0x74, 0x65,
-  0x78, 0x74, 0x20, 0x64, 0x6f, 0x65, 0x73, 0x6e, 0x27, 0x74, 0x20, 0x65,
-  0x6e, 0x64, 0x20, 0x77, 0x69, 0x74, 0x68, 0x20, 0x61, 0x20, 0x6c, 0x69,
-  0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x2c, 0x20, 0x74, 0x68,
-  0x65, 0x6e, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20,
-  0x6c, 0x69, 0x6e, 0x65, 0x20, 0x69, 0x73, 0x20, 0x69, 0x6e, 0x63, 0x6f,
-  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x2f, 0x2f, 0x20, 0x53, 0x74, 0x6f, 0x72, 0x65, 0x20, 0x69, 0x74,
-  0x20, 0x69, 0x6e, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72,
-  0x20, 0x74, 0x6f, 0x20, 0x62, 0x65, 0x20, 0x61, 0x64, 0x64, 0x65, 0x64,
-  0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6e, 0x65, 0x78, 0x74,
-  0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x61,
-  0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
-  0x28, 0x21, 0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x4c, 0x69,
-  0x6e, 0x65, 0x42, 0x72, 0x65, 0x61, 0x6b, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f,
-  0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x2e,
-  0x70, 0x6f, 0x70, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76,
-  0x65, 0x72, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f, 0x20,
-  0x52, 0x65, 0x73, 0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76,
-  0x65, 0x72, 0x20, 0x69, 0x66, 0x20, 0x77, 0x65, 0x20, 0x68, 0x61, 0x76,
-  0x65, 0x20, 0x61, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65,
-  0x61, 0x6b, 0x20, 0x61, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x65, 0x6e,
-  0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x50, 0x61, 0x72, 0x73,
-  0x65, 0x20, 0x61, 0x6c, 0x6c, 0x20, 0x73, 0x73, 0x65, 0x20, 0x65, 0x76,
-  0x65, 0x6e, 0x74, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x64, 0x64,
-  0x20, 0x74, 0x68, 0x65, 0x6d, 0x20, 0x74, 0x6f, 0x20, 0x72, 0x65, 0x73,
-  0x75, 0x6c, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x20, 0x3d, 0x20,
-  0x2f, 0x5e, 0x28, 0x5c, 0x53, 0x2b, 0x29, 0x3a, 0x5c, 0x73, 0x28, 0x2e,
-  0x2a, 0x29, 0x24, 0x2f, 0x67, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74,
-  0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x69, 0x6e,
-  0x65, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x61, 0x74, 0x63,
-  0x68, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x2e, 0x65, 0x78,
-  0x65, 0x63, 0x28, 0x6c, 0x69, 0x6e, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6d, 0x61,
-  0x74, 0x63, 0x68, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x5b,
-  0x6d, 0x61, 0x74, 0x63, 0x68, 0x5b, 0x31, 0x5d, 0x5d, 0x20, 0x3d, 0x20,
-  0x6d, 0x61, 0x74, 0x63, 0x68, 0x5b, 0x32, 0x5d, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69,
-  0x6e, 0x63, 0x65, 0x20, 0x77, 0x65, 0x20, 0x6b, 0x6e, 0x6f, 0x77, 0x20,
-  0x74, 0x68, 0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
-  0x61, 0x2e, 0x63, 0x70, 0x70, 0x2c, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73,
-  0x20, 0x6a, 0x75, 0x73, 0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65,
-  0x20, 0x74, 0x68, 0x65, 0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x20, 0x69, 0x6e,
-  0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75,
-  0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72,
-  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d,
-  0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28,
-  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29,
+  0x72, 0x65, 0x67, 0x65, 0x78, 0x20, 0x3d, 0x20, 0x2f, 0x5e, 0x28, 0x5c,
+  0x53, 0x2b, 0x29, 0x3a, 0x5c, 0x73, 0x28, 0x2e, 0x2a, 0x29, 0x24, 0x2f,
+  0x67, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f,
+  0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x69, 0x6e,
+  0x65, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x20, 0x3d, 0x20,
+  0x72, 0x65, 0x67, 0x65, 0x78, 0x2e, 0x65, 0x78, 0x65, 0x63, 0x28, 0x6c,
+  0x69, 0x6e, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x29,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x5b, 0x6d, 0x61, 0x74, 0x63,
+  0x68, 0x5b, 0x31, 0x5d, 0x5d, 0x20, 0x3d, 0x20, 0x6d, 0x61, 0x74, 0x63,
+  0x68, 0x5b, 0x32, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6e, 0x63, 0x65, 0x20,
+  0x77, 0x65, 0x20, 0x6b, 0x6e, 0x6f, 0x77, 0x20, 0x74, 0x68, 0x69, 0x73,
+  0x20, 0x69, 0x73, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70,
+  0x70, 0x2c, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x6a, 0x75, 0x73,
+  0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x20, 0x74, 0x68, 0x65,
+  0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x64, 0x61, 0x74,
+  0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64,
+  0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c,
+  0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x4a, 0x53, 0x4f,
+  0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x72, 0x65, 0x73, 0x75,
+  0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x72, 0x65, 0x73,
+  0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e,
+  0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x79, 0x69,
+  0x65, 0x6c, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x79, 0x69, 0x65, 0x6c, 0x64, 0x20, 0x72, 0x65,
+  0x73, 0x75, 0x6c, 0x74, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x69, 0x66,
+  0x20, 0x77, 0x65, 0x20, 0x67, 0x6f, 0x74, 0x20, 0x61, 0x20, 0x73, 0x74,
+  0x6f, 0x70, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x20, 0x66, 0x72, 0x6f,
+  0x6d, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2c, 0x20, 0x77, 0x65,
+  0x20, 0x77, 0x69, 0x6c, 0x6c, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x20,
+  0x68, 0x65, 0x72, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73,
+  0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, 0x6f,
+  0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67,
+  0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65,
+  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e,
+  0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20,
+  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e,
+  0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73,
+  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x66, 0x61,
+  0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b,
   0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d,
-  0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61,
-  0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
-  0x2f, 0x20, 0x79, 0x69, 0x65, 0x6c, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x79, 0x69, 0x65, 0x6c,
-  0x64, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x3b, 0x0a, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
-  0x2f, 0x20, 0x69, 0x66, 0x20, 0x77, 0x65, 0x20, 0x67, 0x6f, 0x74, 0x20,
-  0x61, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e,
-  0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72,
-  0x2c, 0x20, 0x77, 0x65, 0x20, 0x77, 0x69, 0x6c, 0x6c, 0x20, 0x62, 0x72,
-  0x65, 0x61, 0x6b, 0x20, 0x68, 0x65, 0x72, 0x65, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
-  0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61,
-  0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69,
-  0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74,
+  0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x3d,
+  0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28,
+  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x65,
+  0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e,
+  0x63, 0x70, 0x70, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x24,
+  0x7b, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f,
+  0x72, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x7d, 0x60, 0x29,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x20,
+  0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66,
+  0x20, 0x28, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x21, 0x3d, 0x3d,
+  0x20, 0x27, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x45, 0x72, 0x72, 0x6f, 0x72,
+  0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72,
+  0x28, 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x20, 0x65, 0x72, 0x72, 0x6f,
+  0x72, 0x3a, 0x20, 0x22, 0x2c, 0x20, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f,
+  0x77, 0x20, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x66,
+  0x69, 0x6e, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e,
+  0x61, 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d,
+  0x0a, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x63,
+  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f,
+  0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x2c, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x61, 0x6e, 0x20,
+  0x65, 0x76, 0x65, 0x6e, 0x74, 0x20, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74,
+  0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x79, 0x6f, 0x75, 0x20, 0x63, 0x61,
+  0x6e, 0x20, 0x73, 0x75, 0x62, 0x73, 0x63, 0x72, 0x69, 0x62, 0x65, 0x20,
+  0x74, 0x6f, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61,
+  0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20,
+  0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20,
+  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61,
+  0x72, 0x67, 0x65, 0x74, 0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20,
+  0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e,
+  0x2e, 0x6a, 0x73, 0x27, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x6e,
+  0x20, 0x3d, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e,
+  0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x28, 0x70, 0x72, 0x6f, 0x6d,
+  0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x6e, 0x2e, 0x61, 0x64, 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c,
+  0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x22, 0x6d, 0x65, 0x73,
+  0x73, 0x61, 0x67, 0x65, 0x22, 0x2c, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e,
+  0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74,
+  0x2e, 0x77, 0x72, 0x69, 0x74, 0x65, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b,
+  0x2e, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x2e, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45,
+  0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d,
+  0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61,
+  0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63,
+  0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20,
+  0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
+  0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x76, 0x65, 0x6e, 0x74,
+  0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x28, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63,
+  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61,
+  0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68,
+  0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29,
+  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66,
+  0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61,
+  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63,
+  0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f,
+  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
+  0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45,
+  0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73,
+  0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x6d, 0x65,
+  0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65,
+  0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e,
+  0x64, 0x61, 0x74, 0x61, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61,
   0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f,
   0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61,
-  0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67,
-  0x73, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64,
-  0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69,
-  0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20,
-  0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62,
-  0x72, 0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x20,
-  0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x2e, 0x6e, 0x61,
-  0x6d, 0x65, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x41, 0x62, 0x6f, 0x72,
-  0x74, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65,
-  0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x6c, 0x6c, 0x61, 0x6d,
-  0x61, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x22, 0x2c, 0x20,
-  0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x65, 0x3b, 0x0a, 0x20,
-  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x61, 0x6c, 0x6c, 0x79,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72,
-  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x72, 0x65,
-  0x74, 0x75, 0x72, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c,
-  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75,
-  0x72, 0x6e, 0x20, 0x61, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x20,
-  0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20,
-  0x79, 0x6f, 0x75, 0x20, 0x63, 0x61, 0x6e, 0x20, 0x73, 0x75, 0x62, 0x63,
-  0x72, 0x69, 0x62, 0x65, 0x20, 0x74, 0x6f, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f,
-  0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a, 0x2f,
-  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f,
-  0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76,
-  0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x7d, 0x20,
-  0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c,
-  0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x0a, 0x2f, 0x2f,
-  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
-  0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x20, 0x3d, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
-  0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
-  0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20,
-  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x2e, 0x61, 0x64, 0x64, 0x45,
-  0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72,
-  0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c, 0x20,
-  0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
-  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63,
-  0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74, 0x65, 0x28,
-  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c,
-  0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f, 0x2f,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76,
+  0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69,
+  0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28,
+  0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76,
+  0x65, 0x6e, 0x74, 0x28, 0x22, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74,
+  0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73,
+  0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a,
+  0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e,
+  0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73,
+  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b,
+  0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67,
+  0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
+  0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65,
+  0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74, 0x6f,
+  0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x74, 0x69, 0x6d, 0x69,
+  0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61,
+  0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61,
+  0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d,
+  0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76,
+  0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69,
+  0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28,
+  0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76,
+  0x65, 0x6e, 0x74, 0x28, 0x22, 0x64, 0x6f, 0x6e, 0x65, 0x22, 0x2c, 0x20,
+  0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x7b, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x7d, 0x20, 0x7d, 0x29,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x28, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e,
+  0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x3b, 0x0a, 0x7d, 0x0a, 0x0a,
+  0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x61, 0x20,
+  0x70, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x20, 0x74, 0x68, 0x61, 0x74,
+  0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x73, 0x20, 0x74, 0x6f,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74,
+  0x65, 0x64, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2e, 0x20, 0x54, 0x68, 0x69,
+  0x73, 0x20, 0x64, 0x6f, 0x65, 0x73, 0x20, 0x6e, 0x6f, 0x74, 0x20, 0x73,
+  0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x73, 0x74, 0x72, 0x65, 0x61,
+  0x6d, 0x69, 0x6e, 0x67, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x45,
+  0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f,
+  0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50,
+  0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70,
+  0x74, 0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28, 0x28, 0x63, 0x6f, 0x6e,
+  0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x2f,
+  0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75,
+  0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74, 0x65, 0x28, 0x63,
+  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x6f, 0x72, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63,
+  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61,
+  0x69, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d,
+  0x69, 0x73, 0x65, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a,
+  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d,
+  0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74, 0x65, 0x28, 0x63, 0x6f,
+  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78,
   0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c,
-  0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72,
-  0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70,
-  0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20,
-  0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d,
-  0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54,
-  0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20,
-  0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x28,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20,
-  0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x20,
+  0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x2c, 0x20,
+  0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x29,
+  0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x50, 0x72, 0x6f, 0x6d, 0x69,
+  0x73, 0x65, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x72, 0x65,
+  0x73, 0x6f, 0x6c, 0x76, 0x65, 0x2c, 0x20, 0x72, 0x65, 0x6a, 0x65, 0x63,
+  0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
   0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
-  0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f,
+  0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72,
+  0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f,
   0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e,
   0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20,
   0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74,
   0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f,
   0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b,
-  0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
   0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61,
   0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e,
-  0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70,
-  0x61, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65,
-  0x77, 0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e,
-  0x74, 0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c,
-  0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63,
-  0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x7d, 0x29,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75,
-  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65,
-  0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69,
-  0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
-  0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45,
-  0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73,
-  0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x67, 0x65,
-  0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74,
-  0x74, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65,
-  0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e,
-  0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74,
-  0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73,
-  0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
-  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74,
-  0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54,
-  0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74,
-  0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20,
-  0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28,
-  0x22, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b,
-  0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75,
-  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69,
-  0x6e, 0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
-  0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45,
-  0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73,
-  0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x64, 0x6f,
-  0x6e, 0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69,
-  0x6c, 0x3a, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x20, 0x7d, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x29,
-  0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e,
-  0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
-  0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c,
-  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75,
-  0x72, 0x6e, 0x20, 0x61, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65,
-  0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76,
-  0x65, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f,
-  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x64, 0x20, 0x74, 0x65, 0x78, 0x74,
-  0x2e, 0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x64, 0x6f, 0x65, 0x73, 0x20,
-  0x6e, 0x6f, 0x74, 0x20, 0x73, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20,
-  0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x69, 0x6e, 0x67, 0x0a, 0x2f, 0x2f,
-  0x0a, 0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a,
-  0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c,
-  0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28,
-  0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e,
-  0x28, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d,
-  0x3e, 0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72,
-  0x69, 0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29,
-  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f,
-  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x72, 0x0a,
-  0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
-  0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
-  0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70, 0x72, 0x6f,
-  0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69,
-  0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a,
-  0x2f, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f,
-  0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f,
-  0x6d, 0x69, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d,
-  0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d,
-  0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20,
-  0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x65, 0x77, 0x20,
-  0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x61, 0x73, 0x79, 0x6e,
-  0x63, 0x20, 0x28, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x2c, 0x20,
-  0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e,
-  0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74,
-  0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e,
-  0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70,
-  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d,
-  0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75,
-  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74,
-  0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c,
-  0x76, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x3b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68,
-  0x20, 0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x28,
-  0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x7d, 0x3b, 0x0a, 0x0a,
-  0x2f, 0x2a, 0x2a, 0x0a, 0x20, 0x2a, 0x20, 0x28, 0x64, 0x65, 0x70, 0x72,
-  0x65, 0x63, 0x61, 0x74, 0x65, 0x64, 0x29, 0x0a, 0x20, 0x2a, 0x2f, 0x0a,
-  0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
-  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65,
-  0x74, 0x65, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28,
-  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74,
-  0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2c, 0x20, 0x63, 0x61, 0x6c, 0x6c,
-  0x62, 0x61, 0x63, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28,
-  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20,
-  0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x61, 0x72,
-  0x61, 0x6d, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20,
-  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63, 0x6f,
-  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x7d, 0x29, 0x29,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62,
-  0x61, 0x63, 0x6b, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x3b, 0x0a,
-  0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x47, 0x65,
-  0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x20,
-  0x69, 0x6e, 0x66, 0x6f, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x68,
-  0x65, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2e, 0x20, 0x54, 0x68,
-  0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x75, 0x73, 0x65, 0x66, 0x75, 0x6c,
-  0x20, 0x66, 0x6f, 0x72, 0x20, 0x67, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67,
-  0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74,
-  0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x20, 0x61, 0x6e, 0x64, 0x20,
-  0x73, 0x6f, 0x20, 0x6f, 0x6e, 0x2e, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x28, 0x63, 0x6f,
+  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x72, 0x72,
+  0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x65, 0x72, 0x72, 0x6f, 0x72,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d,
+  0x29, 0x3b, 0x0a, 0x7d, 0x3b, 0x0a, 0x0a, 0x2f, 0x2a, 0x2a, 0x0a, 0x20,
+  0x2a, 0x20, 0x28, 0x64, 0x65, 0x70, 0x72, 0x65, 0x63, 0x61, 0x74, 0x65,
+  0x64, 0x29, 0x0a, 0x20, 0x2a, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72,
   0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
-  0x61, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d,
-  0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x67, 0x65,
-  0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74,
-  0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f,
-  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x61,
-  0x77, 0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x22,
-  0x2f, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x22,
-  0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28, 0x72, 0x20, 0x3d, 0x3e, 0x20,
-  0x72, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x28, 0x29, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
-  0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73,
-  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x7d, 0x0a
+  0x61, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x20, 0x3d, 0x20,
+  0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65,
+  0x72, 0x2c, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x29,
+  0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20,
+  0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c,
+  0x61, 0x6d, 0x61, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x70,
+  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c,
+  0x6c, 0x65, 0x72, 0x20, 0x7d, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x28, 0x63,
+  0x68, 0x75, 0x6e, 0x6b, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x7d,
+  0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x47, 0x65, 0x74, 0x20, 0x74, 0x68, 0x65,
+  0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x20, 0x69, 0x6e, 0x66, 0x6f, 0x20,
+  0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x68, 0x65, 0x20, 0x73, 0x65, 0x72,
+  0x76, 0x65, 0x72, 0x2e, 0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x69, 0x73,
+  0x20, 0x75, 0x73, 0x65, 0x66, 0x75, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x20,
+  0x67, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x74, 0x68, 0x65, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x20, 0x77, 0x69, 0x6e, 0x64,
+  0x6f, 0x77, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x6f, 0x20, 0x6f, 0x6e,
+  0x2e, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x4d, 0x6f, 0x64, 0x65,
+  0x6c, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e,
+  0x63, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x69, 0x66, 0x20, 0x28, 0x21, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74,
+  0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73,
+  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x6e, 0x65,
+  0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20,
+  0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x22, 0x2f, 0x6d, 0x6f, 0x64, 0x65,
+  0x6c, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x22, 0x29, 0x2e, 0x74, 0x68, 0x65,
+  0x6e, 0x28, 0x72, 0x20, 0x3d, 0x3e, 0x20, 0x72, 0x2e, 0x6a, 0x73, 0x6f,
+  0x6e, 0x28, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72,
+  0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e,
+  0x67, 0x73, 0x3b, 0x0a, 0x7d, 0x0a
 };
-unsigned int completion_js_len = 5099;
+unsigned int completion_js_len = 5346;

examples/server/public/completion.js

@@ -95,6 +95,15 @@ export async function* llama(prompt, params = {}, config = {}) {
               break;
             }
           }
+          if (result.error) {
+            result.error = JSON.parse(result.error);
+            if (result.error.content.includes('slot unavailable')) {
+              // Throw an error to be caught by upstream callers
+              throw new Error('slot unavailable');
+            } else {
+              console.error(`llama.cpp error: ${result.error.content}`);
+            }
+          }
           if (result.error) {
             result.error = JSON.parse(result.error);
             console.error(`llama.cpp error: ${result.error.content}`);

examples/server/public/index.html

@@ -427,7 +427,7 @@
         }
 
         if (data.timings) {
-          llamaStats.value = data.timings;
+          llamaStats.value = data;
         }
       }
 
@@ -880,7 +880,7 @@
       }
       return html`
         <span>
-          ${llamaStats.value.predicted_per_token_ms.toFixed()}ms per token, ${llamaStats.value.predicted_per_second.toFixed(2)} tokens per second
+          ${llamaStats.value.tokens_predicted} predicted, ${llamaStats.value.tokens_cached} cached, ${llamaStats.value.timings.predicted_per_token_ms.toFixed()}ms per token, ${llamaStats.value.timings.predicted_per_second.toFixed(2)} tokens per second
         </span>
       `
     }

examples/server/server.cpp

@@ -25,6 +25,8 @@
 #include <thread>
 #include <mutex>
 #include <chrono>
+#include <condition_variable>
+#include <atomic>
 
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
@@ -37,7 +39,7 @@ using json = nlohmann::json;
 struct server_params
 {
     std::string hostname = "127.0.0.1";
-    std::string api_key;
+    std::vector<std::string> api_keys;
     std::string public_path = "examples/server/public";
     int32_t port = 8080;
     int32_t read_timeout = 600;
@@ -81,7 +83,7 @@ static inline bool is_base64(uint8_t c)
     return (isalnum(c) || (c == '+') || (c == '/'));
 }
 
-static std::vector<uint8_t> base64_decode(std::string const &encoded_string)
+static std::vector<uint8_t> base64_decode(const std::string & encoded_string)
 {
     int i = 0;
     int j = 0;
@@ -145,9 +147,15 @@ static std::vector<uint8_t> base64_decode(std::string const &encoded_string)
 // parallel
 //
 
+enum server_state {
+    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
+    SERVER_STATE_READY,          // Server is ready and model is loaded
+    SERVER_STATE_ERROR           // An error occurred, load_model failed
+};
+
 enum task_type {
-    COMPLETION_TASK,
-    CANCEL_TASK
+    TASK_TYPE_COMPLETION,
+    TASK_TYPE_CANCEL,
 };
 
 struct task_server {
@@ -208,10 +216,10 @@ struct slot_image
     int32_t id;
 
     bool request_encode_image = false;
-    float* image_embedding = nullptr;
+    float * image_embedding = nullptr;
     int32_t image_tokens = 0;
 
-    clip_image_u8 img_data;
+    clip_image_u8 * img_data;
 
     std::string prefix_prompt; // before of this image
 };
@@ -433,20 +441,27 @@ struct llama_client_slot
 
         generated_token_probs.clear();
 
-        for (slot_image &img : images)
+        for (slot_image & img : images)
         {
             free(img.image_embedding);
-            delete[] img.img_data.data;
+            if (img.img_data) {
+                clip_image_u8_free(img.img_data);
+            }
             img.prefix_prompt = "";
         }
 
         images.clear();
-        // llama_set_rng_seed(ctx, params.seed); in batched the seed matter???????
     }
 
     bool has_budget(gpt_params &global_params) {
+        if (params.n_predict == -1 && global_params.n_predict == -1)
+        {
+            return true; // limitless
+        }
+
         n_remaining = -1;
-        if(params.n_predict != -1)
+
+        if (params.n_predict != -1)
         {
             n_remaining = params.n_predict - n_decoded;
         }
@@ -454,7 +469,8 @@ struct llama_client_slot
         {
             n_remaining = global_params.n_predict - n_decoded;
         }
-        return n_remaining > 0 || n_remaining == -1; // no budget || limitless
+
+        return n_remaining > 0; // no budget
     }
 
     bool available() const {
@@ -542,7 +558,9 @@ struct llama_server_context
     std::vector<task_result> queue_results;
     std::vector<task_multi>  queue_multitasks;
     std::mutex mutex_tasks; // also guards id_gen, and queue_multitasks
+    std::condition_variable condition_tasks;
     std::mutex mutex_results;
+    std::condition_variable condition_results;
 
     ~llama_server_context()
     {
@@ -761,6 +779,42 @@ struct llama_server_context
             slot->prompt = "";
         }
 
+        slot->sparams.penalty_prompt_tokens.clear();
+        slot->sparams.use_penalty_prompt_tokens = false;
+        const auto &penalty_prompt = data.find("penalty_prompt");
+        if (penalty_prompt != data.end())
+        {
+            if (penalty_prompt->is_string())
+            {
+                const auto penalty_prompt_string = penalty_prompt->get<std::string>();
+                auto penalty_tokens = llama_tokenize(model, penalty_prompt_string, false);
+                slot->sparams.penalty_prompt_tokens.swap(penalty_tokens);
+                if (slot->params.n_predict > 0)
+                {
+                    slot->sparams.penalty_prompt_tokens.reserve(slot->sparams.penalty_prompt_tokens.size() + slot->params.n_predict);
+                }
+                slot->sparams.use_penalty_prompt_tokens = true;
+            }
+            else if (penalty_prompt->is_array())
+            {
+                const auto n_tokens = penalty_prompt->size();
+                slot->sparams.penalty_prompt_tokens.reserve(n_tokens + std::max(0, slot->params.n_predict));
+                const int n_vocab = llama_n_vocab(model);
+                for (const auto &penalty_token : *penalty_prompt)
+                {
+                    if (penalty_token.is_number_integer())
+                    {
+                        const auto tok = penalty_token.get<llama_token>();
+                        if (tok >= 0 && tok < n_vocab)
+                        {
+                            slot->sparams.penalty_prompt_tokens.push_back(tok);
+                        }
+                    }
+                }
+                slot->sparams.use_penalty_prompt_tokens = true;
+            }
+        }
+
         slot->sparams.logit_bias.clear();
 
         if (json_value(data, "ignore_eos", false))
@@ -813,24 +867,17 @@ struct llama_server_context
             {
                 for (const auto &img : *images_data)
                 {
-                    std::string data_b64 = img["data"].get<std::string>();
+                    const std::vector<uint8_t> image_buffer = base64_decode(img["data"].get<std::string>());
+
                     slot_image img_sl;
                     img_sl.id = img.count("id") != 0 ? img["id"].get<int>() : slot->images.size();
-                    int width, height, channels;
-                    std::vector<uint8_t> image_buffer = base64_decode(data_b64);
-                    data_b64.clear();
-                    auto data = stbi_load_from_memory(image_buffer.data(), image_buffer.size(), &width, &height, &channels, 3);
-                    if (!data) {
+                    img_sl.img_data = clip_image_u8_init();
+                    if (!clip_image_load_from_bytes(image_buffer.data(), image_buffer.size(), img_sl.img_data))
+                    {
                         LOG_TEE("slot %i - failed to load image [id: %i]\n", slot->id, img_sl.id);
                         return false;
                     }
-                    LOG_TEE("slot %i - image loaded [id: %i] resolution (%i x %i)\n", slot->id, img_sl.id, width, height);
-                    img_sl.img_data.nx = width;
-                    img_sl.img_data.ny = height;
-                    img_sl.img_data.size = width * height * 3;
-                    img_sl.img_data.data = new uint8_t[width * height * 3]();
-                    memcpy(img_sl.img_data.data, data, width * height * 3);
-                    stbi_image_free(data);
+                    LOG_TEE("slot %i - loaded image\n", slot->id);
                     img_sl.request_encode_image = true;
                     slot->images.push_back(img_sl);
                 }
@@ -885,6 +932,7 @@ struct llama_server_context
             llama_sampling_free(slot->ctx_sampling);
         }
         slot->ctx_sampling = llama_sampling_init(slot->sparams);
+        llama_set_rng_seed(ctx, slot->params.seed);
         slot->command = LOAD_PROMPT;
 
         all_slots_are_idle = false;
@@ -992,6 +1040,12 @@ struct llama_server_context
         slot.generated_text += token_str;
         slot.has_next_token = true;
 
+        if (slot.ctx_sampling->params.use_penalty_prompt_tokens && result.tok != -1)
+        {
+            // we can change penalty_prompt_tokens because it is always created from scratch each request
+            slot.ctx_sampling->params.penalty_prompt_tokens.push_back(result.tok);
+        }
+
         // check if there is incomplete UTF-8 character at the end
         bool incomplete = false;
         for (unsigned i = 1; i < 5 && i <= slot.generated_text.size(); ++i)
@@ -1062,7 +1116,7 @@ struct llama_server_context
         }
 
         // check the limits
-        if (slot.n_decoded > 2 && slot.has_next_token && !slot.has_budget(params))
+        if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params))
         {
             slot.stopped_limit = true;
             slot.has_next_token = false;
@@ -1098,8 +1152,8 @@ struct llama_server_context
             {
                 continue;
             }
-            clip_image_f32 img_res;
-            if (!clip_image_preprocess(clp_ctx, &img.img_data, &img_res, /*pad2square =*/ true))
+            clip_image_f32 * img_res = clip_image_f32_init();
+            if (!clip_image_preprocess(clp_ctx, img.img_data, img_res, /*pad2square =*/ true))
             {
                 LOG_TEE("Error processing the given image");
                 clip_free(clp_ctx);
@@ -1114,11 +1168,12 @@ struct llama_server_context
                 return false;
             }
             LOG_TEE("slot %i - encoding image [id: %i]\n", slot.id, img.id);
-            if (!clip_image_encode(clp_ctx, params.n_threads, &img_res, img.image_embedding))
+            if (!clip_image_encode(clp_ctx, params.n_threads, img_res, img.image_embedding))
             {
                 LOG_TEE("Unable to encode image\n");
                 return false;
             }
+            clip_image_f32_free(img_res);
             img.request_encode_image = false;
         }
 
@@ -1127,7 +1182,7 @@ struct llama_server_context
 
     void send_error(task_server& task, std::string error)
     {
-        std::lock_guard<std::mutex> lock(mutex_results);
+        std::unique_lock<std::mutex> lock(mutex_results);
         task_result res;
         res.id = task.id;
         res.multitask_id = task.multitask_id;
@@ -1135,6 +1190,7 @@ struct llama_server_context
         res.error = true;
         res.result_json = { { "content", error } };
         queue_results.push_back(res);
+        condition_results.notify_all();
     }
 
     void add_multi_task(int id, std::vector<int>& sub_ids)
@@ -1144,6 +1200,7 @@ struct llama_server_context
         multi.id = id;
         std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end()));
         queue_multitasks.push_back(multi);
+        condition_tasks.notify_one();
     }
 
     void update_multi_task(int multitask_id, int subtask_id, task_result& result)
@@ -1155,6 +1212,7 @@ struct llama_server_context
             {
                 multitask.subtasks_remaining.erase(subtask_id);
                 multitask.results.push_back(result);
+                condition_tasks.notify_one();
             }
         }
     }
@@ -1173,7 +1231,7 @@ struct llama_server_context
             {"n_ctx",             slot.n_ctx},
             {"model",             params.model_alias},
             {"seed",              slot.params.seed},
-            {"temp",              slot.sparams.temp},
+            {"temperature",       slot.sparams.temp},
             {"top_k",             slot.sparams.top_k},
             {"top_p",             slot.sparams.top_p},
             {"min_p",             slot.sparams.min_p},
@@ -1183,6 +1241,8 @@ struct llama_server_context
             {"repeat_penalty",    slot.sparams.penalty_repeat},
             {"presence_penalty",  slot.sparams.penalty_present},
             {"frequency_penalty", slot.sparams.penalty_freq},
+            {"penalty_prompt_tokens", slot.sparams.penalty_prompt_tokens},
+            {"use_penalty_prompt_tokens", slot.sparams.use_penalty_prompt_tokens},
             {"mirostat",          slot.sparams.mirostat},
             {"mirostat_tau",      slot.sparams.mirostat_tau},
             {"mirostat_eta",      slot.sparams.mirostat_eta},
@@ -1200,7 +1260,7 @@ struct llama_server_context
 
     void send_partial_response(llama_client_slot &slot, completion_token_output tkn)
     {
-        std::lock_guard<std::mutex> lock(mutex_results);
+        std::unique_lock<std::mutex> lock(mutex_results);
         task_result res;
         res.id = slot.task_id;
         res.multitask_id = slot.multitask_id;
@@ -1219,7 +1279,7 @@ struct llama_server_context
         {
             std::vector<completion_token_output> probs_output = {};
             const std::vector<llama_token> to_send_toks = llama_tokenize(ctx, tkn.text_to_send, false);
-            size_t probs_pos = std::min(slot.sent_token_probs_index, slot.generated_token_probs.size());
+            size_t probs_pos      = std::min(slot.sent_token_probs_index,                       slot.generated_token_probs.size());
             size_t probs_stop_pos = std::min(slot.sent_token_probs_index + to_send_toks.size(), slot.generated_token_probs.size());
             if (probs_pos < probs_stop_pos)
             {
@@ -1236,11 +1296,12 @@ struct llama_server_context
         }
 
         queue_results.push_back(res);
+        condition_results.notify_all();
     }
 
     void send_final_response(llama_client_slot &slot)
     {
-        std::lock_guard<std::mutex> lock(mutex_results);
+        std::unique_lock<std::mutex> lock(mutex_results);
         task_result res;
         res.id = slot.task_id;
         res.multitask_id = slot.multitask_id;
@@ -1278,7 +1339,7 @@ struct llama_server_context
             {
                 probs = std::vector<completion_token_output>(
                                     slot.generated_token_probs.begin(),
-                                    slot.generated_token_probs.begin() + slot.sent_token_probs_index);
+                                    slot.generated_token_probs.end());
             }
             res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs);
         }
@@ -1296,11 +1357,12 @@ struct llama_server_context
         }
 
         queue_results.push_back(res);
+        condition_results.notify_all();
     }
 
     void send_embedding(llama_client_slot &slot)
     {
-        std::lock_guard<std::mutex> lock(mutex_results);
+        std::unique_lock<std::mutex> lock(mutex_results);
         task_result res;
         res.id = slot.task_id;
         res.multitask_id = slot.multitask_id;
@@ -1328,6 +1390,7 @@ struct llama_server_context
             };
         }
         queue_results.push_back(res);
+        condition_results.notify_all();
     }
 
     int request_completion(json data, bool infill, bool embedding, int multitask_id)
@@ -1339,11 +1402,11 @@ struct llama_server_context
         task.data = std::move(data);
         task.infill_mode = infill;
         task.embedding_mode = embedding;
-        task.type = COMPLETION_TASK;
+        task.type = TASK_TYPE_COMPLETION;
         task.multitask_id = multitask_id;
 
         // when a completion task's prompt array is not a singleton, we split it into multiple requests
-        if (task.data.at("prompt").size() > 1)
+        if (task.data.count("prompt") && task.data.at("prompt").size() > 1)
         {
             lock.unlock(); // entering new func scope
             return split_multiprompt_task(task);
@@ -1351,6 +1414,7 @@ struct llama_server_context
 
         // otherwise, it's a single-prompt task, we actually queue it
         queue_tasks.push_back(task);
+        condition_tasks.notify_one();
         return task.id;
     }
 
@@ -1358,13 +1422,10 @@ struct llama_server_context
     {
         while (true)
         {
-            std::this_thread::sleep_for(std::chrono::microseconds(5));
-            std::lock_guard<std::mutex> lock(mutex_results);
-
-            if (queue_results.empty())
-            {
-                continue;
-            }
+            std::unique_lock<std::mutex> lock(mutex_results);
+            condition_results.wait(lock, [&]{
+                return !queue_results.empty();
+            });
 
             for (int i = 0; i < (int) queue_results.size(); i++)
             {
@@ -1460,12 +1521,13 @@ struct llama_server_context
 
     void request_cancel(int task_id)
     {
-        std::lock_guard<std::mutex> lock(mutex_tasks);
+        std::unique_lock<std::mutex> lock(mutex_tasks);
         task_server task;
         task.id = id_gen++;
-        task.type = CANCEL_TASK;
+        task.type = TASK_TYPE_CANCEL;
         task.target_id = task_id;
         queue_tasks.push_back(task);
+        condition_tasks.notify_one();
     }
 
     int split_multiprompt_task(task_server& multiprompt_task)
@@ -1491,14 +1553,14 @@ struct llama_server_context
 
     void process_tasks()
     {
-        std::lock_guard<std::mutex> lock(mutex_tasks);
+        std::unique_lock<std::mutex> lock(mutex_tasks);
         while (!queue_tasks.empty())
         {
             task_server task = queue_tasks.front();
             queue_tasks.erase(queue_tasks.begin());
             switch (task.type)
             {
-                case COMPLETION_TASK: {
+                case TASK_TYPE_COMPLETION: {
                     llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
                     if (slot == nullptr)
                     {
@@ -1515,9 +1577,9 @@ struct llama_server_context
 
                     slot->reset();
 
-                    slot->infill = task.infill_mode;
-                    slot->embedding = task.embedding_mode;
-                    slot->task_id = task.id;
+                    slot->infill       = task.infill_mode;
+                    slot->embedding    = task.embedding_mode;
+                    slot->task_id      = task.id;
                     slot->multitask_id = task.multitask_id;
 
                     if (!launch_slot_with_data(slot, task.data))
@@ -1527,7 +1589,7 @@ struct llama_server_context
                         break;
                     }
                 } break;
-                case CANCEL_TASK: { // release slot linked with the task id
+                case TASK_TYPE_CANCEL: { // release slot linked with the task id
                     for (auto & slot : slots)
                     {
                         if (slot.task_id == task.target_id)
@@ -1563,6 +1625,7 @@ struct llama_server_context
 
                 std::lock_guard<std::mutex> lock(mutex_results);
                 queue_results.push_back(aggregate_result);
+                condition_results.notify_all();
 
                 queue_iterator = queue_multitasks.erase(queue_iterator);
             }
@@ -1593,8 +1656,10 @@ struct llama_server_context
                 LOG_TEE("all slots are idle and system prompt is empty, clear the KV cache\n");
                 kv_cache_clear();
             }
-            // avoid 100% usage of cpu all time
-            std::this_thread::sleep_for(std::chrono::milliseconds(5));
+            std::unique_lock<std::mutex> lock(mutex_tasks);
+            condition_tasks.wait(lock, [&]{
+                return !queue_tasks.empty();
+            });
         }
 
         for (llama_client_slot &slot : slots)
@@ -1652,7 +1717,6 @@ struct llama_server_context
 
             llama_batch_add(batch, slot.sampled, system_tokens.size() + slot.n_past, { slot.id }, true);
 
-            slot.n_decoded += 1;
             slot.n_past += 1;
         }
 
@@ -1667,7 +1731,8 @@ struct llama_server_context
                 const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty()) || !slot.images.empty();
 
                 // empty prompt passed -> release the slot and send empty response
-                if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt)
+                // note: infill mode allows empty prompt
+                if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt && !slot.infill)
                 {
                     slot.release();
                     slot.print_timings();
@@ -1870,6 +1935,7 @@ struct llama_server_context
 
                 llama_sampling_accept(slot.ctx_sampling, ctx, id, true);
 
+                slot.n_decoded += 1;
                 if (slot.n_decoded == 1)
                 {
                     slot.t_start_genereration = ggml_time_us();
@@ -1939,12 +2005,15 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
     printf("  -ngl N, --n-gpu-layers N\n");
     printf("                        number of layers to store in VRAM\n");
+    printf("  -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
+    printf("                        how to split the model across multiple GPUs, one of:\n");
+    printf("                          - none: use one GPU only\n");
+    printf("                          - layer (default): split layers and KV across GPUs\n");
+    printf("                          - row: split rows across GPUs\n");
     printf("  -ts SPLIT --tensor-split SPLIT\n");
-    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-    printf("  -nommq, --no-mul-mat-q\n");
-    printf("                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
-    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
+    printf("                        fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for the model (with split-mode = none),\n");
+    printf("                        or for intermediate results and KV (with split-mode = row)\n");
 #endif
     printf("  -m FNAME, --model FNAME\n");
     printf("                        model path (default: %s)\n", params.model.c_str());
@@ -1956,6 +2025,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
     printf("  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
     printf("  --api-key API_KEY     optional api key to enhance server security. If set, requests must include this key for access.\n");
+    printf("  --api-key-file FNAME  path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access.\n");
     printf("  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
     printf("  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
     printf("  -np N, --parallel N   number of slots for process requests (default: %d)\n", params.n_parallel);
@@ -1965,6 +2035,10 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  --mmproj MMPROJ_FILE  path to a multimodal projector file for LLaVA.\n");
     printf("  --log-disable         disables logging to a file.\n");
     printf("\n");
+    printf("  --override-kv KEY=TYPE:VALUE\n");
+    printf("                        advanced option to override model metadata by key. may be specified multiple times.\n");
+    printf("                        types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
+    printf("\n");
 }
 
 static void server_params_parse(int argc, char **argv, server_params &sparams,
@@ -2012,7 +2086,28 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 invalid_param = true;
                 break;
             }
-            sparams.api_key = argv[i];
+            sparams.api_keys.push_back(argv[i]);
+        }
+        else if (arg == "--api-key-file")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            std::ifstream key_file(argv[i]);
+            if (!key_file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            std::string key;
+            while (std::getline(key_file, key)) {
+               if (key.size() > 0) {
+                   sparams.api_keys.push_back(key);
+               }
+            }
+            key_file.close();
         }
         else if (arg == "--timeout" || arg == "-to")
         {
@@ -2161,6 +2256,33 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                         "See main README.md for information on enabling GPU BLAS support",
                         {{"n_gpu_layers", params.n_gpu_layers}});
 #endif
+        }
+        else if (arg == "--split-mode" || arg == "-sm")
+        {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string arg_next = argv[i];
+            if (arg_next == "none")
+            {
+                params.split_mode = LLAMA_SPLIT_NONE;
+            }
+            else if (arg_next == "layer")
+            {
+                params.split_mode = LLAMA_SPLIT_LAYER;
+            }
+            else if (arg_next == "row")
+            {
+                params.split_mode = LLAMA_SPLIT_ROW;
+            }
+            else {
+                invalid_param = true;
+                break;
+            }
+#ifndef GGML_USE_CUBLAS
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n");
+#endif // GGML_USE_CUBLAS
         }
         else if (arg == "--tensor-split" || arg == "-ts")
         {
@@ -2328,6 +2450,49 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             log_set_target(stdout);
             LOG_INFO("logging to file is disabled.", {});
         }
+        else if (arg == "--override-kv")
+        {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            char * sep = strchr(argv[i], '=');
+            if (sep == nullptr || sep - argv[i] >= 128) {
+                fprintf(stderr, "error: Malformed KV override: %s\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            struct llama_model_kv_override kvo;
+            std::strncpy(kvo.key, argv[i], sep - argv[i]);
+            kvo.key[sep - argv[i]] = 0;
+            sep++;
+            if (strncmp(sep, "int:", 4) == 0) {
+                sep += 4;
+                kvo.tag = LLAMA_KV_OVERRIDE_INT;
+                kvo.int_value = std::atol(sep);
+            } else if (strncmp(sep, "float:", 6) == 0) {
+                sep += 6;
+                kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
+                kvo.float_value = std::atof(sep);
+            } else if (strncmp(sep, "bool:", 5) == 0) {
+                sep += 5;
+                kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
+                if (std::strcmp(sep, "true") == 0) {
+                    kvo.bool_value = true;
+                } else if (std::strcmp(sep, "false") == 0) {
+                    kvo.bool_value = false;
+                } else {
+                    fprintf(stderr, "error: Invalid boolean value for KV override: %s\n", argv[i]);
+                    invalid_param = true;
+                    break;
+                }
+            } else {
+                fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            params.kv_overrides.push_back(kvo);
+        }
         else
         {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@@ -2335,6 +2500,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             exit(1);
         }
     }
+    if (!params.kv_overrides.empty()) {
+        params.kv_overrides.emplace_back(llama_model_kv_override());
+        params.kv_overrides.back().key[0] = 0;
+    }
 
     if (invalid_param)
     {
@@ -2344,7 +2513,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
     }
 }
 
-
 static std::string random_string()
 {
     static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
@@ -2393,26 +2561,33 @@ json oaicompat_completion_params_parse(
     llama_params["__oaicompat"] = true;
 
     // Map OpenAI parameters to llama.cpp parameters
-    llama_params["model"]             = json_value(body, "model", std::string("uknown"));
+    //
+    // For parameters that are defined by the OpenAI documentation (e.g.
+    // temperature), we explicitly specify OpenAI's intended default; we
+    // need to do that because sometimes OpenAI disagrees with llama.cpp
+    //
+    // https://platform.openai.com/docs/api-reference/chat/create
+    llama_sampling_params default_sparams;
+    llama_params["model"]             = json_value(body, "model", std::string("unknown"));
     llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
     llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
-    llama_params["temperature"]       = json_value(body, "temperature", 0.8);
-    llama_params["top_k"]             = json_value(body, "top_k", 40);
-    llama_params["top_p"]             = json_value(body, "top_p", 0.95);
+    llama_params["temperature"]       = json_value(body, "temperature", 0.0);
+    llama_params["top_k"]             = json_value(body, "top_k", default_sparams.top_k);
+    llama_params["top_p"]             = json_value(body, "top_p", 1.0);
     llama_params["n_predict"]         = json_value(body, "max_tokens", -1);
     llama_params["logit_bias"]        = json_value(body, "logit_bias",json::object());
     llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0);
     llama_params["presence_penalty"]  = json_value(body, "presence_penalty", 0.0);
-    llama_params["seed"]              = json_value(body, "seed", 0);
+    llama_params["seed"]              = json_value(body, "seed", LLAMA_DEFAULT_SEED);
     llama_params["stream"]            = json_value(body, "stream", false);
-    llama_params["mirostat"]          = json_value(body, "mirostat", false);
-    llama_params["mirostat_tau"]      = json_value(body, "mirostat_tau", 0.0);
-    llama_params["mirostat_eta"]      = json_value(body, "mirostat_eta", 0.0);
-    llama_params["penalize_nl"]       = json_value(body, "penalize_nl", false);
-    llama_params["typical_p"]         = json_value(body, "typical_p", 0.0);
-    llama_params["repeat_last_n"]     = json_value(body, "repeat_last_n", 0);
+    llama_params["mirostat"]          = json_value(body, "mirostat", default_sparams.mirostat);
+    llama_params["mirostat_tau"]      = json_value(body, "mirostat_tau", default_sparams.mirostat_tau);
+    llama_params["mirostat_eta"]      = json_value(body, "mirostat_eta", default_sparams.mirostat_eta);
+    llama_params["penalize_nl"]       = json_value(body, "penalize_nl", default_sparams.penalize_nl);
+    llama_params["typical_p"]         = json_value(body, "typical_p", default_sparams.typical_p);
+    llama_params["repeat_last_n"]     = json_value(body, "repeat_last_n", default_sparams.penalty_last_n);
     llama_params["ignore_eos"]        = json_value(body, "ignore_eos", false);
-    llama_params["tfs_z"]             = json_value(body, "tfs_z", 0.0);
+    llama_params["tfs_z"]             = json_value(body, "tfs_z", default_sparams.tfs_z);
 
     if (body.count("grammar") != 0) {
         llama_params["grammar"] = json_value(body, "grammar", json::object());
@@ -2465,8 +2640,8 @@ static json format_final_response_oaicompat(const json &request, const task_resu
             {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
             {"usage",
                 json{{"completion_tokens", num_tokens_predicted},
-                    {"prompt_tokens", num_prompt_tokens},
-                    {"total_tokens", num_tokens_predicted + num_prompt_tokens}}},
+                     {"prompt_tokens",     num_prompt_tokens},
+                     {"total_tokens",      num_tokens_predicted + num_prompt_tokens}}},
             {"id", gen_chatcmplid()}};
 
     if (server_verbose) {
@@ -2674,20 +2849,131 @@ int main(int argc, char **argv)
                                 {"system_info", llama_print_system_info()},
                             });
 
-    // load the model
-    if (!llama.load_model(params))
+    httplib::Server svr;
+
+    std::atomic<server_state> state{SERVER_STATE_LOADING_MODEL};
+
+    svr.set_default_headers({{"Server", "llama.cpp"}});
+
+    // CORS preflight
+    svr.Options(R"(.*)", [](const httplib::Request &req, httplib::Response &res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+        res.set_header("Access-Control-Allow-Credentials", "true");
+        res.set_header("Access-Control-Allow-Methods", "POST");
+        res.set_header("Access-Control-Allow-Headers", "*");
+    });
+
+    svr.Get("/health", [&](const httplib::Request&, httplib::Response& res) {
+        server_state current_state = state.load();
+        switch(current_state) {
+            case SERVER_STATE_READY:
+                res.set_content(R"({"status": "ok"})", "application/json");
+                res.status = 200; // HTTP OK
+                break;
+            case SERVER_STATE_LOADING_MODEL:
+                res.set_content(R"({"status": "loading model"})", "application/json");
+                res.status = 503; // HTTP Service Unavailable
+                break;
+            case SERVER_STATE_ERROR:
+                res.set_content(R"({"status": "error", "error": "Model failed to load"})", "application/json");
+                res.status = 500; // HTTP Internal Server Error
+                break;
+        }
+    });
+
+    svr.set_logger(log_server_request);
+
+    svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
+            {
+                const char fmt[] = "500 Internal Server Error\n%s";
+                char buf[BUFSIZ];
+                try
+                {
+                    std::rethrow_exception(std::move(ep));
+                }
+                catch (std::exception &e)
+                {
+                    snprintf(buf, sizeof(buf), fmt, e.what());
+                }
+                catch (...)
+                {
+                    snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
+                }
+                res.set_content(buf, "text/plain; charset=utf-8");
+                res.status = 500;
+            });
+
+    svr.set_error_handler([](const httplib::Request &, httplib::Response &res)
+            {
+                if (res.status == 401)
+                {
+                    res.set_content("Unauthorized", "text/plain; charset=utf-8");
+                }
+                if (res.status == 400)
+                {
+                    res.set_content("Invalid request", "text/plain; charset=utf-8");
+                }
+                else if (res.status == 404)
+                {
+                    res.set_content("File Not Found", "text/plain; charset=utf-8");
+                    res.status = 404;
+                }
+            });
+
+    // set timeouts and change hostname and port
+    svr.set_read_timeout (sparams.read_timeout);
+    svr.set_write_timeout(sparams.write_timeout);
+
+    if (!svr.bind_to_port(sparams.hostname, sparams.port))
     {
+        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port);
         return 1;
     }
 
-    llama.initialize();
+    // Set the base directory for serving static files
+    svr.set_base_dir(sparams.public_path);
 
-    httplib::Server svr;
+    // to make it ctrl+clickable:
+    LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
+
+    std::unordered_map<std::string, std::string> log_data;
+    log_data["hostname"] = sparams.hostname;
+    log_data["port"] = std::to_string(sparams.port);
+
+    if (sparams.api_keys.size() == 1) {
+        log_data["api_key"] = "api_key: ****" + sparams.api_keys[0].substr(sparams.api_keys[0].length() - 4);
+    } else if (sparams.api_keys.size() > 1) {
+        log_data["api_key"] = "api_key: " + std::to_string(sparams.api_keys.size()) + " keys loaded";
+    }
+
+    LOG_INFO("HTTP server listening", log_data);
+    // run the HTTP server in a thread - see comment below
+    std::thread t([&]()
+            {
+                if (!svr.listen_after_bind())
+                {
+                    state.store(SERVER_STATE_ERROR);
+                    return 1;
+                }
+
+                return 0;
+            });
+
+    // load the model
+    if (!llama.load_model(params))
+    {
+        state.store(SERVER_STATE_ERROR);
+        return 1;
+    } else {
+        llama.initialize();
+        state.store(SERVER_STATE_READY);
+        LOG_INFO("model loaded", {});
+    }
 
     // Middleware for API key validation
     auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool {
         // If API key is not set, skip validation
-        if (sparams.api_key.empty()) {
+        if (sparams.api_keys.empty()) {
             return true;
         }
 
@@ -2696,7 +2982,7 @@ int main(int argc, char **argv)
         std::string prefix = "Bearer ";
         if (auth_header.substr(0, prefix.size()) == prefix) {
             std::string received_api_key = auth_header.substr(prefix.size());
-            if (received_api_key == sparams.api_key) {
+            if (std::find(sparams.api_keys.begin(), sparams.api_keys.end(), received_api_key) != sparams.api_keys.end()) {
                 return true; // API key is valid
             }
         }
@@ -2710,10 +2996,6 @@ int main(int argc, char **argv)
         return false;
     };
 
-    svr.set_default_headers({{"Server", "llama.cpp"},
-                             {"Access-Control-Allow-Origin", "*"},
-                             {"Access-Control-Allow-Headers", "content-type"}});
-
     // this is only called if no index.html is found in the public --path
     svr.Get("/", [](const httplib::Request &, httplib::Response &res)
             {
@@ -2742,9 +3024,9 @@ int main(int argc, char **argv)
                 return false;
             });
 
-    svr.Get("/props", [&llama](const httplib::Request & /*req*/, httplib::Response &res)
+    svr.Get("/props", [&llama](const httplib::Request & req, httplib::Response &res)
             {
-                res.set_header("Access-Control-Allow-Origin", "*");
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 json data = {
                     { "user_name",      llama.name_user.c_str() },
                     { "assistant_name", llama.name_assistant.c_str() }
@@ -2754,6 +3036,7 @@ int main(int argc, char **argv)
 
     svr.Post("/completion", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 if (!validate_api_key(req, res)) {
                     return;
                 }
@@ -2821,10 +3104,9 @@ int main(int argc, char **argv)
                 }
             });
 
-
-
-    svr.Get("/v1/models", [&params](const httplib::Request&, httplib::Response& res)
+    svr.Get("/v1/models", [&params](const httplib::Request& req, httplib::Response& res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 std::time_t t = std::time(0);
 
                 json models = {
@@ -2842,9 +3124,11 @@ int main(int argc, char **argv)
                 res.set_content(models.dump(), "application/json; charset=utf-8");
             });
 
+
     // TODO: add mount point without "/v1" prefix -- how?
     svr.Post("/v1/chat/completions", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 if (!validate_api_key(req, res)) {
                     return;
                 }
@@ -2918,6 +3202,7 @@ int main(int argc, char **argv)
 
     svr.Post("/infill", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 if (!validate_api_key(req, res)) {
                     return;
                 }
@@ -2990,6 +3275,7 @@ int main(int argc, char **argv)
 
     svr.Post("/tokenize", [&llama](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 const json body = json::parse(req.body);
                 std::vector<llama_token> tokens;
                 if (body.count("content") != 0)
@@ -3002,6 +3288,7 @@ int main(int argc, char **argv)
 
     svr.Post("/detokenize", [&llama](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 const json body = json::parse(req.body);
                 std::string content;
                 if (body.count("tokens") != 0)
@@ -3016,6 +3303,7 @@ int main(int argc, char **argv)
 
     svr.Post("/embedding", [&llama](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 const json body = json::parse(req.body);
                 json prompt;
                 if (body.count("content") != 0)
@@ -3026,86 +3314,21 @@ int main(int argc, char **argv)
                 {
                     prompt = "";
                 }
-                const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false, true, -1);
+
+                json image_data;
+                if (body.count("image_data") != 0) {
+                    image_data = body["image_data"];
+                }
+                else
+                {
+                    image_data = "";
+                }
+
+                const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, false, true, -1);
                 task_result result = llama.next_result(task_id);
                 return res.set_content(result.result_json.dump(), "application/json; charset=utf-8");
             });
 
-    svr.set_logger(log_server_request);
-
-    svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
-            {
-                const char fmt[] = "500 Internal Server Error\n%s";
-                char buf[BUFSIZ];
-                try
-                {
-                    std::rethrow_exception(std::move(ep));
-                }
-                catch (std::exception &e)
-                {
-                    snprintf(buf, sizeof(buf), fmt, e.what());
-                }
-                catch (...)
-                {
-                    snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
-                }
-                res.set_content(buf, "text/plain; charset=utf-8");
-                res.status = 500;
-            });
-
-    svr.set_error_handler([](const httplib::Request &, httplib::Response &res)
-            {
-                if (res.status == 401)
-                {
-                    res.set_content("Unauthorized", "text/plain; charset=utf-8");
-                }
-                if (res.status == 400)
-                {
-                    res.set_content("Invalid request", "text/plain; charset=utf-8");
-                }
-                else if (res.status == 404)
-                {
-                    res.set_content("File Not Found", "text/plain; charset=utf-8");
-                    res.status = 404;
-                }
-            });
-
-    // set timeouts and change hostname and port
-    svr.set_read_timeout (sparams.read_timeout);
-    svr.set_write_timeout(sparams.write_timeout);
-
-    if (!svr.bind_to_port(sparams.hostname, sparams.port))
-    {
-        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port);
-        return 1;
-    }
-
-    // Set the base directory for serving static files
-    svr.set_base_dir(sparams.public_path);
-
-    // to make it ctrl+clickable:
-    LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
-
-    std::unordered_map<std::string, std::string> log_data;
-    log_data["hostname"] = sparams.hostname;
-    log_data["port"] = std::to_string(sparams.port);
-
-    if (!sparams.api_key.empty()) {
-        log_data["api_key"] = "api_key: ****" + sparams.api_key.substr(sparams.api_key.length() - 4);
-    }
-
-    LOG_INFO("HTTP server listening", log_data);
-    // run the HTTP server in a thread - see comment below
-    std::thread t([&]()
-            {
-                if (!svr.listen_after_bind())
-                {
-                    return 1;
-                }
-
-                return 0;
-            });
-
     // GG: if I put the main loop inside a thread, it crashes on the first request when build in Debug!?
     //     "Bus error: 10" - this is on macOS, it does not crash on Linux
     //std::thread t2([&]()

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

@@ -369,10 +369,7 @@ static struct ggml_tensor * llama_build_train_graphs(
     checkpoints.push_back(t00);
     checkpoints.push_back(t01);
 
-    struct ggml_tensor * kv_scale = NULL;
-    if (!enable_flash_attn) {
-        kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
-    }
+    const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
 
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
@@ -444,14 +441,13 @@ static struct ggml_tensor * llama_build_train_graphs(
         // make sure some tensors are not reallocated by inserting new temporary nodes depending on them
         int n_leafs_before = gb->n_leafs;
         int n_nodes_before = gb->n_nodes;
-        struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
         // output tensors
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
         // input gradient
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
         // KQ_pos
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
         GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
 
         ggml_allocr_alloc(alloc, t36->grad);

flake.lock

@@ -1,30 +1,30 @@
 {
   "nodes": {
-    "flake-utils": {
+    "flake-parts": {
       "inputs": {
-        "systems": "systems"
+        "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1694529238,
-        "narHash": "sha256-zsNZZGTGnMOf9YpHKJqMSsa0dXbfmxeoJ7xHlrt+xmY=",
-        "owner": "numtide",
-        "repo": "flake-utils",
-        "rev": "ff7b65b44d01cf9ba6a71320833626af21126384",
+        "lastModified": 1701473968,
+        "narHash": "sha256-YcVE5emp1qQ8ieHUnxt1wCZCC3ZfAS+SRRWZ2TMda7E=",
+        "owner": "hercules-ci",
+        "repo": "flake-parts",
+        "rev": "34fed993f1674c8d06d58b37ce1e0fe5eebcb9f5",
         "type": "github"
       },
       "original": {
-        "owner": "numtide",
-        "repo": "flake-utils",
+        "owner": "hercules-ci",
+        "repo": "flake-parts",
         "type": "github"
       }
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1698318101,
-        "narHash": "sha256-gUihHt3yPD7bVqg+k/UVHgngyaJ3DMEBchbymBMvK1E=",
+        "lastModified": 1703637592,
+        "narHash": "sha256-8MXjxU0RfFfzl57Zy3OfXCITS0qWDNLzlBAdwxGZwfY=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "63678e9f3d3afecfeafa0acead6239cdb447574c",
+        "rev": "cfc3698c31b1fb9cdcf10f36c9643460264d0ca8",
         "type": "github"
       },
       "original": {
@@ -34,26 +34,29 @@
         "type": "github"
       }
     },
-    "root": {
-      "inputs": {
-        "flake-utils": "flake-utils",
-        "nixpkgs": "nixpkgs"
-      }
-    },
-    "systems": {
+    "nixpkgs-lib": {
       "locked": {
-        "lastModified": 1681028828,
-        "narHash": "sha256-Vy1rq5AaRuLzOxct8nz4T6wlgyUR7zLU309k9mBC768=",
-        "owner": "nix-systems",
-        "repo": "default",
-        "rev": "da67096a3b9bf56a91d16901293e51ba5b49a27e",
+        "dir": "lib",
+        "lastModified": 1701253981,
+        "narHash": "sha256-ztaDIyZ7HrTAfEEUt9AtTDNoCYxUdSd6NrRHaYOIxtk=",
+        "owner": "NixOS",
+        "repo": "nixpkgs",
+        "rev": "e92039b55bcd58469325ded85d4f58dd5a4eaf58",
         "type": "github"
       },
       "original": {
-        "owner": "nix-systems",
-        "repo": "default",
+        "dir": "lib",
+        "owner": "NixOS",
+        "ref": "nixos-unstable",
+        "repo": "nixpkgs",
         "type": "github"
       }
+    },
+    "root": {
+      "inputs": {
+        "flake-parts": "flake-parts",
+        "nixpkgs": "nixpkgs"
+      }
     }
   },
   "root": "root",

flake.nix

@@ -1,139 +1,144 @@
 {
+  description = "Port of Facebook's LLaMA model in C/C++";
+
   inputs = {
     nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";
-    flake-utils.url = "github:numtide/flake-utils";
+    flake-parts.url = "github:hercules-ci/flake-parts";
   };
-  outputs = { self, nixpkgs, flake-utils }:
-    flake-utils.lib.eachDefaultSystem (system:
-      let
-        name = "llama.cpp";
-        src = ./.;
-        meta.mainProgram = "llama";
-        inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin;
-        buildInputs = with pkgs; [ openmpi ];
-        osSpecific = with pkgs; buildInputs ++ (
-          if isAarch64 && isDarwin then
-            with pkgs.darwin.apple_sdk_11_0.frameworks; [
-              Accelerate
-              MetalKit
-            ]
-          else if isAarch32 && isDarwin then
-            with pkgs.darwin.apple_sdk.frameworks; [
-              Accelerate
-              CoreGraphics
-              CoreVideo
-            ]
-          else if isDarwin then
-            with pkgs.darwin.apple_sdk.frameworks; [
-              Accelerate
-              CoreGraphics
-              CoreVideo
-            ]
-          else
-            with pkgs; [ openblas ]
-        );
-        pkgs = import nixpkgs { inherit system; };
-        nativeBuildInputs = with pkgs; [ cmake ninja pkg-config ];
-        cudatoolkit_joined = with pkgs; symlinkJoin {
-          # HACK(Green-Sky): nix currently has issues with cmake findcudatoolkit
-          # see https://github.com/NixOS/nixpkgs/issues/224291
-          # copied from jaxlib
-          name = "${cudaPackages.cudatoolkit.name}-merged";
-          paths = [
-            cudaPackages.cudatoolkit.lib
-            cudaPackages.cudatoolkit.out
-          ] ++ lib.optionals (lib.versionOlder cudaPackages.cudatoolkit.version "11") [
-            # for some reason some of the required libs are in the targets/x86_64-linux
-            # directory; not sure why but this works around it
-            "${cudaPackages.cudatoolkit}/targets/${system}"
-          ];
-        };
-        llama-python =
-          pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
-        # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
-        llama-python-extra =
-          pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece torchWithoutCuda transformers ]);
-        postPatch = ''
-          substituteInPlace ./ggml-metal.m \
-            --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
-          substituteInPlace ./*.py --replace '/usr/bin/env python' '${llama-python}/bin/python'
-        '';
-        postInstall = ''
-          mv $out/bin/main $out/bin/llama
-          mv $out/bin/server $out/bin/llama-server
-          mkdir -p $out/include
-          cp ${src}/llama.h $out/include/
-        '';
-        cmakeFlags = [ "-DLLAMA_NATIVE=OFF" "-DLLAMA_BUILD_SERVER=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ];
-      in
+
+  # Optional binary cache
+  nixConfig = {
+    extra-substituters = [
+      # Populated by the CI in ggerganov/llama.cpp
+      "https://llama-cpp.cachix.org"
+
+      # A development cache for nixpkgs imported with `config.cudaSupport = true`.
+      # Populated by https://hercules-ci.com/github/SomeoneSerge/nixpkgs-cuda-ci.
+      # This lets one skip building e.g. the CUDA-enabled openmpi.
+      # TODO: Replace once nix-community obtains an official one.
+      "https://cuda-maintainers.cachix.org"
+    ];
+
+    # Verify these are the same keys as published on
+    # - https://app.cachix.org/cache/llama-cpp
+    # - https://app.cachix.org/cache/cuda-maintainers
+    extra-trusted-public-keys = [
+      "llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc="
+      "cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E="
+    ];
+  };
+
+
+  # For inspection, use `nix flake show github:ggerganov/llama.cpp` or the nix repl:
+  #
+  # ```bash
+  # ❯ nix repl
+  # nix-repl> :lf github:ggerganov/llama.cpp
+  # Added 13 variables.
+  # nix-repl> outputs.apps.x86_64-linux.quantize
+  # { program = "/nix/store/00000000000000000000000000000000-llama.cpp/bin/quantize"; type = "app"; }
+  # ```
+  outputs =
+    { self, flake-parts, ... }@inputs:
+    let
+      # We could include the git revisions in the package names but those would
+      # needlessly trigger rebuilds:
+      # llamaVersion = self.dirtyShortRev or self.shortRev;
+
+      # Nix already uses cryptographic hashes for versioning, so we'll just fix
+      # the fake semver for now:
+      llamaVersion = "0.0.0";
+    in
+    flake-parts.lib.mkFlake { inherit inputs; }
+
       {
-        packages.default = pkgs.stdenv.mkDerivation {
-          inherit name src meta postPatch nativeBuildInputs postInstall;
-          buildInputs = osSpecific;
-          cmakeFlags = cmakeFlags
-            ++ (if isAarch64 && isDarwin then [
-            "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
-            "-DLLAMA_METAL=ON"
-          ] else [
-            "-DLLAMA_BLAS=ON"
-            "-DLLAMA_BLAS_VENDOR=OpenBLAS"
-          ]);
-        };
-        packages.opencl = pkgs.stdenv.mkDerivation {
-          inherit name src meta postPatch nativeBuildInputs postInstall;
-          buildInputs = with pkgs; buildInputs ++ [ clblast ];
-          cmakeFlags = cmakeFlags ++ [
-            "-DLLAMA_CLBLAST=ON"
-          ];
-        };
-        packages.cuda = pkgs.stdenv.mkDerivation {
-          inherit name src meta postPatch nativeBuildInputs postInstall;
-          buildInputs = with pkgs; buildInputs ++ [ cudatoolkit_joined ];
-          cmakeFlags = cmakeFlags ++ [
-            "-DLLAMA_CUBLAS=ON"
-          ];
-        };
-        packages.rocm = pkgs.stdenv.mkDerivation {
-          inherit name src meta postPatch nativeBuildInputs postInstall;
-          buildInputs = with pkgs.rocmPackages; buildInputs ++ [ clr hipblas rocblas ];
-          cmakeFlags = cmakeFlags ++ [
-            "-DLLAMA_HIPBLAS=1"
-            "-DCMAKE_C_COMPILER=hipcc"
-            "-DCMAKE_CXX_COMPILER=hipcc"
-            # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
-            # in github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
-            # and select the line that matches the current nixpkgs version of rocBLAS.
-            "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
-          ];
-        };
-        apps.llama-server = {
-          type = "app";
-          program = "${self.packages.${system}.default}/bin/llama-server";
-        };
-        apps.llama-embedding = {
-          type = "app";
-          program = "${self.packages.${system}.default}/bin/embedding";
-        };
-        apps.llama = {
-          type = "app";
-          program = "${self.packages.${system}.default}/bin/llama";
-        };
-        apps.quantize = {
-          type = "app";
-          program = "${self.packages.${system}.default}/bin/quantize";
-        };
-        apps.train-text-from-scratch = {
-          type = "app";
-          program = "${self.packages.${system}.default}/bin/train-text-from-scratch";
-        };
-        apps.default = self.apps.${system}.llama;
-        devShells.default = pkgs.mkShell {
-          buildInputs = [ llama-python ];
-          packages = nativeBuildInputs ++ osSpecific;
-        };
-        devShells.extra = pkgs.mkShell {
-          buildInputs = [ llama-python-extra ];
-          packages = nativeBuildInputs ++ osSpecific;
-        };
-      });
+
+        imports = [
+          .devops/nix/nixpkgs-instances.nix
+          .devops/nix/apps.nix
+          .devops/nix/devshells.nix
+          .devops/nix/jetson-support.nix
+        ];
+
+        # An overlay can be used to have a more granular control over llama-cpp's
+        # dependencies and configuration, than that offered by the `.override`
+        # mechanism. Cf. https://nixos.org/manual/nixpkgs/stable/#chap-overlays.
+        #
+        # E.g. in a flake:
+        # ```
+        # { nixpkgs, llama-cpp, ... }:
+        # let pkgs = import nixpkgs {
+        #     overlays = [ (llama-cpp.overlays.default) ];
+        #     system = "aarch64-linux";
+        #     config.allowUnfree = true;
+        #     config.cudaSupport = true;
+        #     config.cudaCapabilities = [ "7.2" ];
+        #     config.cudaEnableForwardCompat = false;
+        # }; in {
+        #     packages.aarch64-linux.llamaJetsonXavier = pkgs.llamaPackages.llama-cpp;
+        # }
+        # ```
+        #
+        # Cf. https://nixos.org/manual/nix/unstable/command-ref/new-cli/nix3-flake.html?highlight=flake#flake-format
+        flake.overlays.default =
+          (final: prev: {
+            llamaPackages = final.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
+            inherit (final.llamaPackages) llama-cpp;
+          });
+
+        systems = [
+          "aarch64-darwin"
+          "aarch64-linux"
+          "x86_64-darwin" # x86_64-darwin isn't tested (and likely isn't relevant)
+          "x86_64-linux"
+        ];
+
+        perSystem =
+          {
+            config,
+            lib,
+            system,
+            pkgs,
+            pkgsCuda,
+            pkgsRocm,
+            ...
+          }:
+          {
+            # Unlike `.#packages`, legacyPackages may contain values of
+            # arbitrary types (including nested attrsets) and may even throw
+            # exceptions. This attribute isn't recursed into by `nix flake
+            # show` either.
+            #
+            # You can add arbitrary scripts to `.devops/nix/scope.nix` and
+            # access them as `nix build .#llamaPackages.${scriptName}` using
+            # the same path you would with an overlay.
+            legacyPackages = {
+              llamaPackages = pkgs.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
+              llamaPackagesCuda = pkgsCuda.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
+              llamaPackagesRocm = pkgsRocm.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
+            };
+
+            # We don't use the overlay here so as to avoid making too many instances of nixpkgs,
+            # cf. https://zimbatm.com/notes/1000-instances-of-nixpkgs
+            packages =
+              {
+                default = config.legacyPackages.llamaPackages.llama-cpp;
+              }
+              // lib.optionalAttrs pkgs.stdenv.isLinux {
+                opencl = config.packages.default.override { useOpenCL = true; };
+                cuda = config.legacyPackages.llamaPackagesCuda.llama-cpp;
+
+                mpi-cpu = config.packages.default.override { useMpi = true; };
+                mpi-cuda = config.packages.default.override { useMpi = true; };
+              }
+              // lib.optionalAttrs (system == "x86_64-linux") {
+                rocm = config.legacyPackages.llamaPackagesRocm.llama-cpp;
+              };
+
+            # Packages exposed in `.#checks` will be built by the CI and by
+            # `nix flake check`. Currently we expose all packages, but we could
+            # make more granular choices
+            checks = config.packages;
+          };
+      };
 }

ggml-alloc.c

@@ -72,7 +72,7 @@ static void remove_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * t
 
 // check if a tensor is allocated by this buffer
 static bool ggml_tallocr_is_own(ggml_tallocr_t alloc, const struct ggml_tensor * tensor) {
-    return tensor->buffer == alloc->buffer;
+    return tensor->buffer == alloc->buffer && (!tensor->view_src || tensor->view_src->buffer == alloc->buffer);
 }
 
 static bool ggml_is_view(struct ggml_tensor * t) {
@@ -102,8 +102,6 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
         }
     }
 
-    AT_PRINTF("block %d\n", best_fit_block);
-
     if (best_fit_block == -1) {
         // the last block is our last resort
         struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
@@ -117,6 +115,7 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
             return;
         }
     }
+
     struct free_block * block = &alloc->free_blocks[best_fit_block];
     void * addr = block->addr;
     block->addr = (char*)block->addr + size;
@@ -129,6 +128,8 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
         }
     }
 
+    AT_PRINTF("block %d, addr %p\n", best_fit_block, addr);
+
     tensor->data = addr;
     tensor->buffer = alloc->buffer;
     if (!alloc->measure) {
@@ -229,6 +230,7 @@ void ggml_tallocr_reset(ggml_tallocr_t alloc) {
         alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
     } else {
         alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset;
+        ggml_backend_buffer_reset(alloc->buffer);
     }
 }
 
@@ -263,9 +265,9 @@ ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment) {
     return alloc;
 }
 
-ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend) {
+ggml_tallocr_t ggml_tallocr_new_measure_from_buft(struct ggml_backend_buffer_type * buft) {
     // create a backend buffer to get the correct tensor allocation sizes
-    ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, 1);
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, 1);
 
     // TODO: move alloc initialization to a common ggml_tallocr_new_impl function
     ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer);
@@ -275,13 +277,22 @@ ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backe
     return alloc;
 }
 
-ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size) {
-    ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, size);
+ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend) {
+    return ggml_tallocr_new_measure_from_buft(ggml_backend_get_default_buffer_type(backend));
+}
+
+ggml_tallocr_t ggml_tallocr_new_from_buft(struct ggml_backend_buffer_type * buft, size_t size) {
+    // create a backend buffer to get the correct tensor allocation sizes
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
     ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer);
     alloc->buffer_owned = true;
     return alloc;
 }
 
+ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size) {
+    return ggml_tallocr_new_from_buft(ggml_backend_get_default_buffer_type(backend), size);
+}
+
 ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer) {
     ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
 
@@ -449,11 +460,10 @@ static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool upd
     if (update_backend) {
         view->backend = view->view_src->backend;
     }
-    view->buffer  = view->view_src->buffer;
+    // views are initialized in the alloc buffer rather than the view_src buffer
+    view->buffer  = alloc->buffer;
     view->data    = (char *)view->view_src->data + view->view_offs;
 
-    // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
-    // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
     assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->buft == alloc->buffer->buft);
 
     if (!alloc->measure) {
@@ -736,6 +746,10 @@ void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n) {
 }
 
 void ggml_allocr_free(ggml_allocr_t alloc) {
+    if (alloc == NULL) {
+        return;
+    }
+
     ggml_gallocr_free(alloc->galloc);
     ggml_tallocr_free(alloc->talloc);
     free(alloc);
@@ -775,11 +789,22 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
     }
 
     if (nbytes == 0) {
-        fprintf(stderr, "%s: no tensors to allocate\n", __func__);
+        // all the tensors in the context are already allocated
+#ifndef NDEBUG
+        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
+#endif
         return NULL;
     }
 
     ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
+    if (buffer == NULL) {
+        // failed to allocate buffer
+#ifndef NDEBUG
+        fprintf(stderr, "%s: failed to allocate buffer\n", __func__);
+#endif
+        return NULL;
+    }
+
     ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
 
     for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
@@ -789,6 +814,11 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
             } else {
                 ggml_backend_view_init(buffer, t);
             }
+        } else {
+            if (t->view_src != NULL) {
+                // view of a pre-allocated tensor
+                ggml_backend_view_init(buffer, t);
+            }
         }
     }
 

ggml-alloc.h

@@ -52,8 +52,10 @@ typedef struct ggml_tallocr * ggml_tallocr_t;
 
 GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment);
 GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment);
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
+GGML_API ggml_tallocr_t ggml_tallocr_new_from_buft(struct ggml_backend_buffer_type * buft, size_t size);
 GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
+GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
+GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_buft(struct ggml_backend_buffer_type * buft);
 GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend);
 
 GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc);

ggml-backend-impl.h

@@ -16,10 +16,14 @@ extern "C" {
     typedef void * ggml_backend_buffer_type_context_t;
 
     struct ggml_backend_buffer_type_i {
+        const char *          (*get_name)        (ggml_backend_buffer_type_t buft);
         ggml_backend_buffer_t (*alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
         size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
-        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
+        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
         bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+        // check if tensor data is in host memory
+        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
+        bool                  (*is_host)         (ggml_backend_buffer_type_t buft);
     };
 
     struct ggml_backend_buffer_type {
@@ -31,15 +35,15 @@ extern "C" {
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        void     (*free_buffer)(ggml_backend_buffer_t buffer);
-        //void     (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
-        void *   (*get_base)   (ggml_backend_buffer_t buffer);
-        void     (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void     (*set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void     (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
-        void (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        const char * (*get_name)   (ggml_backend_buffer_t buffer);
+        void         (*free_buffer)(ggml_backend_buffer_t buffer);
+        void *       (*get_base)   (ggml_backend_buffer_t buffer);
+        void         (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        void         (*set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void         (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool         (*cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
+        void         (*clear)      (ggml_backend_buffer_t buffer, uint8_t value);
+        void         (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
     };
 
     struct ggml_backend_buffer {
@@ -47,6 +51,7 @@ extern "C" {
         ggml_backend_buffer_type_t    buft;
         ggml_backend_buffer_context_t context;
         size_t size;
+        enum ggml_backend_buffer_usage usage;
     };
 
     ggml_backend_buffer_t ggml_backend_buffer_init(
@@ -55,6 +60,8 @@ extern "C" {
                    ggml_backend_buffer_context_t   context,
                    size_t                          size);
 
+    // do not use directly, use ggml_backend_tensor_copy instead
+    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
 
     //
     // Backend
@@ -70,23 +77,21 @@ extern "C" {
         // buffer allocation
         ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
 
-        // (optional) asynchroneous tensor data access
+        // (optional) asynchronous tensor data access
         void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool (*cpy_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
-        // (optional) asynchroneous tensor copy
-        void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to_async)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-
-        void (*synchronize)     (ggml_backend_t backend);
+        // (optional) complete all pending operations
+        void (*synchronize)(ggml_backend_t backend);
 
         // compute graph with a plan
-        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
         void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
         void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
-        // compute graph without a plan
-        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        // compute graph without a plan (async)
+        bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
         // check if the backend supports an operation
         bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
@@ -98,7 +103,6 @@ extern "C" {
         ggml_backend_context_t context;
     };
 
-
     //
     // Backend registry
     //

ggml-backend.h

@@ -17,19 +17,31 @@ extern "C" {
     //
 
     // buffer type
-    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
-    GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
-    GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
-    GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
+    GGML_API const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
+    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
+    GGML_API size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
+    GGML_API size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
+    GGML_API bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
 
     // buffer
-    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
+    enum ggml_backend_buffer_usage {
+        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
+        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
+    };
+
+    GGML_API const char *               ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
+    GGML_API void                       ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API void *                     ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
+    GGML_API bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
+    GGML_API void                       ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
+    GGML_API void                       ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
 
     //
     // Backend
@@ -55,7 +67,7 @@ extern "C" {
 
     GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
     GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
     GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
 
     // tensor copy between different backends
@@ -76,6 +88,10 @@ extern "C" {
 
     GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
 
+#ifdef GGML_USE_CPU_HBM
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
+
     //
     // Backend registry
     //
@@ -133,23 +149,24 @@ extern "C" {
     typedef struct ggml_backend_sched * ggml_backend_sched_t;
 
     // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
-
-    GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
-
+    GGML_API ggml_backend_sched_t  ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size);
+    GGML_API void                  ggml_backend_sched_free(ggml_backend_sched_t sched);
     // Initialize backend buffers from a measure graph
-    GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+    GGML_API void                  ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+    // Get the number of splits of the last graph
+    GGML_API int                   ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
 
     GGML_API ggml_tallocr_t        ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
     GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);
 
-    GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API void                  ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API ggml_backend_t        ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 
-    // Allocate a graph on the backend scheduler
-    GGML_API void ggml_backend_sched_graph_compute(
-            ggml_backend_sched_t sched,
-            struct ggml_cgraph * graph);
+    // Allocate and compute graph on the backend scheduler
+    GGML_API void                  ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
 
+    // Reset all assignments and allocators - must be called before using the sched allocators to allocate inputs
+    GGML_API void                  ggml_backend_sched_reset(ggml_backend_sched_t sched);
 
     //
     // Utils
@@ -169,7 +186,7 @@ extern "C" {
     typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
 
     // Compare the output of two backends
-    GGML_API void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
 
     // Tensor initialization
     GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);

ggml-cuda.h

@@ -27,22 +27,6 @@ GGML_API void * ggml_cuda_host_malloc(size_t size);
 GGML_API void   ggml_cuda_host_free(void * ptr);
 
 GGML_API bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cuda_set_tensor_split(const float * tensor_split);
-GGML_API void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_free_data(struct ggml_tensor * tensor);
-
-GGML_API void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
-
-GGML_API void   ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
-GGML_API void   ggml_cuda_copy_to_device(struct ggml_tensor * tensor);
-
-GGML_API void   ggml_cuda_set_main_device(int main_device);
-GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
-GGML_API void   ggml_cuda_set_scratch_size(size_t scratch_size);
-GGML_API void   ggml_cuda_free_scratch(void);
 GGML_API bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
 
 GGML_API int    ggml_cuda_get_device_count(void);
@@ -52,13 +36,17 @@ GGML_API void   ggml_cuda_get_device_description(int device, char * description,
 GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
 
 GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
-GGML_API int  ggml_backend_cuda_get_device(ggml_backend_t backend);
 
 GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
-
-// pinned host buffer for use with CPU backend for faster copies between CPU and GPU
+// split tensor buffer that splits matrices by rows across multiple devices
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
+GGML_API int  ggml_backend_cuda_get_device_count(void);
+GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+
 #ifdef  __cplusplus
 }
 #endif

ggml-impl.h

@@ -5,6 +5,7 @@
 // GGML internal header
 
 #include <assert.h>
+#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
 #include <stddef.h>
 #include <stdbool.h>
 #include <string.h> // memcpy
@@ -227,6 +228,8 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_HASHTABLE_FULL ((size_t)-1)
 #define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
 
+struct ggml_hash_set ggml_hash_set_new(size_t size);
+
 bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
 
 // returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted

ggml-metal.h

@@ -87,7 +87,7 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
 
 // same as ggml_graph_compute but uses Metal
 // creates gf->n_threads command buffers in parallel
-void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+bool ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
 
 //
 // backend API
@@ -98,7 +98,10 @@ GGML_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
+GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+
 GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
+
 GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family

ggml-metal.m

@@ -87,6 +87,9 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(get_rows_q4_K);
     GGML_METAL_DECL_KERNEL(get_rows_q5_K);
     GGML_METAL_DECL_KERNEL(get_rows_q6_K);
+    GGML_METAL_DECL_KERNEL(get_rows_i32);
+    GGML_METAL_DECL_KERNEL(get_rows_iq2_xxs);
+    GGML_METAL_DECL_KERNEL(get_rows_iq2_xs);
     GGML_METAL_DECL_KERNEL(rms_norm);
     GGML_METAL_DECL_KERNEL(group_norm);
     GGML_METAL_DECL_KERNEL(norm);
@@ -105,6 +108,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mv_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_iq2_xxs_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_iq2_xs_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_id_f32_f32);
     //GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f16);
     GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f32);
@@ -120,6 +125,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mv_id_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_id_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_id_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_iq2_xxs_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_iq2_xs_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_f32_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
@@ -132,6 +139,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_iq2_xxs_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_iq2_xs_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_id_f32_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_id_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_id_q4_0_f32);
@@ -144,6 +153,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mm_id_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_id_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_id_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_iq2_xxs_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_iq2_xs_f32);
     GGML_METAL_DECL_KERNEL(rope_f32);
     GGML_METAL_DECL_KERNEL(rope_f16);
     GGML_METAL_DECL_KERNEL(alibi_f32);
@@ -180,7 +191,15 @@ struct ggml_metal_context {
 @implementation GGMLMetalClass
 @end
 
-ggml_log_callback ggml_metal_log_callback = NULL;
+
+static void ggml_metal_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
+    fprintf(stderr, "%s", msg);
+
+    UNUSED(level);
+    UNUSED(user_data);
+}
+
+ggml_log_callback ggml_metal_log_callback = ggml_metal_default_log_callback;
 void * ggml_metal_log_user_data = NULL;
 
 void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
@@ -251,6 +270,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         NSError * error = nil;
         NSString * libPath = [bundle pathForResource:@"default" ofType:@"metallib"];
         if (libPath != nil) {
+            // pre-compiled library found
             NSURL * libURL = [NSURL fileURLWithPath:libPath];
             GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]);
             ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
@@ -283,6 +303,13 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
             options = [MTLCompileOptions new];
             options.preprocessorMacros = @{ @"QK_K" : @(64) };
 #endif
+            // try to disable fast-math
+            // NOTE: this seems to have no effect whatsoever
+            //       instead, in order to disable fast-math, we have to build default.metallib from the command line
+            //       using xcrun -sdk macosx metal -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+            //       and go through the "pre-compiled library found" path above
+            //[options setFastMathEnabled:false];
+
             ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
         }
 
@@ -361,6 +388,9 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(get_rows_q4_K);
         GGML_METAL_ADD_KERNEL(get_rows_q5_K);
         GGML_METAL_ADD_KERNEL(get_rows_q6_K);
+        GGML_METAL_ADD_KERNEL(get_rows_i32);
+        GGML_METAL_ADD_KERNEL(get_rows_iq2_xxs);
+        GGML_METAL_ADD_KERNEL(get_rows_iq2_xs);
         GGML_METAL_ADD_KERNEL(rms_norm);
         GGML_METAL_ADD_KERNEL(group_norm);
         GGML_METAL_ADD_KERNEL(norm);
@@ -379,6 +409,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul_mv_q4_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_q5_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_q6_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_iq2_xxs_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_iq2_xs_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_id_f32_f32);
         //GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f16);
         GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f32);
@@ -394,6 +426,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul_mv_id_q4_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_id_q5_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_id_q6_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_iq2_xxs_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_iq2_xs_f32);
         if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
             GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
@@ -407,6 +441,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
             GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_iq2_xxs_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_iq2_xs_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_id_f32_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_id_f16_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_id_q4_0_f32);
@@ -419,6 +455,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
             GGML_METAL_ADD_KERNEL(mul_mm_id_q4_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_id_q5_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_id_q6_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_iq2_xxs_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_iq2_xs_f32);
         }
         GGML_METAL_ADD_KERNEL(rope_f32);
         GGML_METAL_ADD_KERNEL(rope_f16);
@@ -483,6 +521,9 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(get_rows_q4_K);
     GGML_METAL_DEL_KERNEL(get_rows_q5_K);
     GGML_METAL_DEL_KERNEL(get_rows_q6_K);
+    GGML_METAL_DEL_KERNEL(get_rows_i32);
+    GGML_METAL_DEL_KERNEL(get_rows_iq2_xxs);
+    GGML_METAL_DEL_KERNEL(get_rows_iq2_xs);
     GGML_METAL_DEL_KERNEL(rms_norm);
     GGML_METAL_DEL_KERNEL(group_norm);
     GGML_METAL_DEL_KERNEL(norm);
@@ -501,6 +542,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul_mv_q4_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_q5_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_q6_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_iq2_xxs_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_iq2_xs_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_id_f32_f32);
     //GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f16);
     GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f32);
@@ -516,6 +559,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul_mv_id_q4_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_id_q5_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_id_q6_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_iq2_xxs_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_iq2_xs_f32);
     if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
         GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
@@ -529,6 +574,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
         GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_iq2_xxs_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_iq2_xs_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_id_f32_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_id_f16_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_id_q4_0_f32);
@@ -541,6 +588,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
         GGML_METAL_DEL_KERNEL(mul_mm_id_q4_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_id_q5_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_id_q6_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_iq2_xxs_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_iq2_xs_f32);
     }
     GGML_METAL_DEL_KERNEL(rope_f32);
     GGML_METAL_DEL_KERNEL(rope_f16);
@@ -607,12 +656,24 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
 }
 
 // temporarily defined here for compatibility between ggml-backend and the old API
-struct ggml_backend_metal_buffer_context {
-    void * data;
+
+struct ggml_backend_metal_buffer {
+    void   * data;
+    size_t   size;
 
     id<MTLBuffer> metal;
 };
 
+struct ggml_backend_metal_buffer_context {
+    void * all_data;
+    size_t all_size;
+    bool owned;
+
+    // multiple buffers are used only to avoid the maximum buffer size limitation when using mmap
+    int n_buffers;
+    struct ggml_backend_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
+};
+
 // 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
@@ -622,17 +683,29 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
 
     const int64_t tsize = ggml_nbytes(t);
 
+    ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
+
     // compatibility with ggml-backend
-    if (t->buffer && t->buffer->buft == ggml_backend_metal_buffer_type()) {
-        struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) t->buffer->context;
+    if (buffer && buffer->buft == ggml_backend_metal_buffer_type()) {
+        struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
 
-        const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->data;
+        // find the view that contains the tensor fully
+        for (int i = 0; i < buf_ctx->n_buffers; ++i) {
+            const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
 
-        GGML_ASSERT(ioffs >= 0 && ioffs + tsize <= (int64_t) t->buffer->size);
+            //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
+            if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
+                *offs = (size_t) ioffs;
 
-        *offs = (size_t) ioffs;
+                //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
 
-        return buf_ctx->metal;
+                return buf_ctx->buffers[i].metal;
+            }
+        }
+
+        GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
+
+        return nil;
     }
 
     // find the view that contains the tensor fully
@@ -934,7 +1007,7 @@ static bool ggml_metal_supports_op(const struct ggml_tensor * op) {
             return false;
     }
 }
-void ggml_metal_graph_compute(
+bool ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
     @autoreleasepool {
@@ -1009,6 +1082,10 @@ void ggml_metal_graph_compute(
                     GGML_ASSERT(!"unsupported op");
                 }
 
+#ifndef GGML_METAL_NDEBUG
+                [encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(dst) encoding:NSUTF8StringEncoding]];
+#endif
+
                 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;
@@ -1198,7 +1275,7 @@ void ggml_metal_graph_compute(
                                 // not sure how to avoid this
                                 // TODO: make a simpler cpy_bytes kernel
 
-                                const int nth = MIN(1024, ne00);
+                                const int nth = MIN((int) ctx->pipeline_cpy_f32_f32.maxTotalThreadsPerThreadgroup, ne00);
 
                                 [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32];
                                 [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
@@ -1253,7 +1330,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26];
                             [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
 
-                            const int nth = MIN(1024, ne0);
+                            const int nth = MIN((int) ctx->pipeline_add.maxTotalThreadsPerThreadgroup, ne00);
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
@@ -1261,7 +1338,7 @@ void ggml_metal_graph_compute(
                         {
                             GGML_ASSERT(ggml_is_contiguous(src0));
 
-                            const float scale = *(const float *) src1->data;
+                            const float scale = *(const float *) dst->op_params;
 
                             int64_t n = ggml_nelements(dst);
 
@@ -1272,8 +1349,8 @@ void ggml_metal_graph_compute(
                                 [encoder setComputePipelineState:ctx->pipeline_scale];
                             }
 
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [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];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
@@ -1498,6 +1575,8 @@ void ggml_metal_graph_compute(
                                     case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
                                     case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
                                     case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
+                                    case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_mul_mm_iq2_xxs_f32]; break;
+                                    case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_mul_mm_iq2_xs_f32]; break;
                                     default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
                                 }
                                 [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
@@ -1610,6 +1689,18 @@ void ggml_metal_graph_compute(
                                             nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q6_K_f32];
                                         } break;
+                                    case GGML_TYPE_IQ2_XXS:
+                                        {
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_iq2_xxs_f32];
+                                        } break;
+                                    case GGML_TYPE_IQ2_XS:
+                                        {
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_iq2_xs_f32];
+                                        } break;
                                     default:
                                         {
                                             GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
@@ -1617,6 +1708,10 @@ void ggml_metal_graph_compute(
                                         }
                                 };
 
+                                if (ggml_is_quantized(src0t)) {
+                                    GGML_ASSERT(ne00 >= nth0*nth1);
+                                }
+
                                 [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];
@@ -1642,6 +1737,11 @@ void ggml_metal_graph_compute(
                                     src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
+                                else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
+                                    const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
+                                    [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
                                 else if (src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
@@ -1675,6 +1775,9 @@ void ggml_metal_graph_compute(
                             // TODO: make this more general
                             GGML_ASSERT(n_as <= 8);
 
+                            // max size of the src1ids array in the kernel stack
+                            GGML_ASSERT(ne11 <= 512);
+
                             struct ggml_tensor * src2 = gf->nodes[i]->src[2];
 
                             const int64_t  ne20 = src2 ? src2->ne[0] : 0;
@@ -1692,9 +1795,6 @@ void ggml_metal_graph_compute(
                             GGML_ASSERT(!ggml_is_transposed(src2));
                             GGML_ASSERT(!ggml_is_transposed(src1));
 
-                            GGML_ASSERT(ne20 % 32 == 0);
-                            // !!!!!!!!! TODO: this assert is probably required but not sure!
-                            //GGML_ASSERT(ne20 >= 64);
                             GGML_ASSERT(src1t == GGML_TYPE_F32);
 
                             const uint r2 = ne12/ne22;
@@ -1702,22 +1802,22 @@ void ggml_metal_graph_compute(
 
                             // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                             // to the matrix-vector kernel
-                            int ne11_mm_min = 1;
+                            int ne11_mm_min = n_as;
 
                             const int idx = ((int32_t *) dst->op_params)[0];
 
                             // batch size
                             GGML_ASSERT(ne01 == ne11);
 
-                            const int64_t _ne1 = 1; // kernel_mul_mm_impl needs a reference in constant memory
-
                             // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                             // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
                             // !!!
                             // TODO: for now, always use mat-vec kernels until we figure out how to improve the
                             //       indirect matrix multiplication
                             // !!!
-                            if ([ctx->device supportsFamily:MTLGPUFamilyApple7] && _ne1 > ne11_mm_min) {
+                            if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
+                                ne20 % 32 == 0 && ne20 >= 64 &&
+                                ne11 > ne11_mm_min) {
                                 switch (src2->type) {
                                     case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_f32_f32];  break;
                                     case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_f16_f32];  break;
@@ -1731,6 +1831,8 @@ void ggml_metal_graph_compute(
                                     case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q4_K_f32]; break;
                                     case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q5_K_f32]; break;
                                     case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q6_K_f32]; break;
+                                    case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_iq2_xxs_f32]; break;
+                                    case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_iq2_xs_f32]; break;
                                     default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
                                 }
                                 [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
@@ -1747,14 +1849,15 @@ void ggml_metal_graph_compute(
                                 [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];
+                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
                                 [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:15];
                                 [encoder setBytes:&r2      length:sizeof(r2)   atIndex:16];
                                 [encoder setBytes:&r3      length:sizeof(r3)   atIndex:17];
                                 [encoder setBytes:&idx     length:sizeof(idx)  atIndex:18];
                                 // TODO: how to make this an array? read Metal docs
-                                for (int j = 0; j < n_as; ++j) {
-                                    struct ggml_tensor * src_cur = dst->src[2 + j];
+                                for (int j = 0; j < 8; ++j) {
+                                    // NOTE: this is done like this to avoid uninitialized kernel arguments when n_as < 8
+                                    struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
 
                                     size_t offs_src_cur = 0;
                                     id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
@@ -1764,8 +1867,7 @@ void ggml_metal_graph_compute(
 
                                 [encoder setThreadgroupMemoryLength:8192 atIndex:0];
 
-                                // TODO: processing one row at a time (ne11 -> 1) is not efficient
-                                [encoder dispatchThreadgroups:MTLSizeMake( (_ne1 + 31)/32, (ne21 + 63)/64, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne11 + 31)/32, (ne21 + 63)/64, n_as*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                             } else {
                                 int nth0 = 32;
                                 int nth1 = 1;
@@ -1846,13 +1948,31 @@ void ggml_metal_graph_compute(
                                             nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q6_K_f32];
                                         } break;
+                                    case GGML_TYPE_IQ2_XXS:
+                                        {
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_iq2_xxs_f32];
+                                        } break;
+                                    case GGML_TYPE_IQ2_XS:
+                                        {
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_iq2_xs_f32];
+                                        } break;
                                     default:
                                         {
-                                            GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
+                                            GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
                                             GGML_ASSERT(false && "not implemented");
                                         }
                                 };
 
+                                if (ggml_is_quantized(src2t)) {
+                                    GGML_ASSERT(ne20 >= nth0*nth1);
+                                }
+
+                                const int64_t _ne1 = 1; // kernels needs a reference in constant memory
+
                                 [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];
@@ -1877,8 +1997,9 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&r3   length:sizeof(r3)   atIndex:21];
                                 [encoder setBytes:&idx  length:sizeof(idx)  atIndex:22];
                                 // TODO: how to make this an array? read Metal docs
-                                for (int j = 0; j < n_as; ++j) {
-                                    struct ggml_tensor * src_cur = dst->src[2 + j];
+                                for (int j = 0; j < 8; ++j) {
+                                    // NOTE: this is done like this to avoid uninitialized kernel arguments when n_as < 8
+                                    struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
 
                                     size_t offs_src_cur = 0;
                                     id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
@@ -1891,6 +2012,11 @@ void ggml_metal_graph_compute(
                                     src2t == GGML_TYPE_Q2_K) { // || src2t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
+                                else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) {
+                                    const int mem_size = src2t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
+                                    [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
                                 else if (src2t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
@@ -1927,6 +2053,9 @@ void ggml_metal_graph_compute(
                                 case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_K]; break;
                                 case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_K]; break;
                                 case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_K]; break;
+                                case GGML_TYPE_I32:  [encoder setComputePipelineState:ctx->pipeline_get_rows_i32];  break;
+                                case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_get_rows_iq2_xxs]; break;
+                                case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_get_rows_iq2_xs]; break;
                                 default: GGML_ASSERT(false && "not implemented");
                             }
 
@@ -2197,7 +2326,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
                             [encoder setBytes:&sf   length:sizeof(sf)   atIndex:18];
 
-                            const int nth = MIN(1024, ne0);
+                            const int nth = MIN((int) ctx->pipeline_upscale_f32.maxTotalThreadsPerThreadgroup, ne0);
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
@@ -2328,6 +2457,10 @@ void ggml_metal_graph_compute(
                             GGML_ASSERT(false);
                         }
                 }
+
+#ifndef GGML_METAL_NDEBUG
+                [encoder popDebugGroup];
+#endif
             }
 
             if (encoder != nil) {
@@ -2350,10 +2483,11 @@ void ggml_metal_graph_compute(
         MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status];
         if (status != MTLCommandBufferStatusCompleted) {
             GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
-            GGML_ASSERT(false);
+            return false;
         }
     }
 
+    return true;
     }
 }
 
@@ -2361,6 +2495,7 @@ void ggml_metal_graph_compute(
 
 // backend interface
 
+// default buffer
 static id<MTLDevice> g_backend_device = nil;
 static int g_backend_device_ref_count = 0;
 
@@ -2385,64 +2520,81 @@ static void ggml_backend_metal_free_device(void) {
     }
 }
 
-static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
-    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
+static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return "Metal";
 
-    return ctx->data;
+    UNUSED(buffer);
 }
 
 static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
 
-    [ctx->metal release];
+    for (int i = 0; i < ctx->n_buffers; i++) {
+        [ctx->buffers[i].metal release];
+    }
     ggml_backend_metal_free_device();
 
-    free(ctx->data);
-    free(ctx);
+    if (ctx->owned) {
+        free(ctx->all_data);
+    }
 
-    UNUSED(buffer);
+    free(ctx);
+}
+
+static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
+    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
+
+    return ctx->all_data;
 }
 
 static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
     memcpy((char *)tensor->data + offset, data, size);
 
     UNUSED(buffer);
 }
 
 static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
     memcpy(data, (const char *)tensor->data + offset, size);
 
     UNUSED(buffer);
 }
 
-static void ggml_backend_metal_buffer_cpy_tensor_from(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        memcpy(dst->data, src->data, ggml_nbytes(src));
+        return true;
+    }
+    return false;
 
     UNUSED(buffer);
 }
 
-static void ggml_backend_metal_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
+static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
 
-    UNUSED(buffer);
+    memset(ctx->all_data, value, ctx->all_size);
 }
 
-static struct ggml_backend_buffer_i metal_backend_buffer_i = {
+static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
+    /* .get_name        = */ ggml_backend_metal_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_metal_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_metal_buffer_get_base,
     /* .init_tensor     = */ NULL,
     /* .set_tensor      = */ ggml_backend_metal_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_metal_buffer_get_tensor,
-    /* .cpy_tensor_from = */ ggml_backend_metal_buffer_cpy_tensor_from,
-    /* .cpy_tensor_to   = */ ggml_backend_metal_buffer_cpy_tensor_to,
+    /* .cpy_tensor      = */ ggml_backend_metal_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_metal_buffer_clear,
+    /* .reset           = */ NULL,
 };
 
+// default buffer type
+
+static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "Metal";
+
+    UNUSED(buft);
+}
+
 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
 
@@ -2453,13 +2605,46 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
         size_aligned += (size_page - (size_aligned % size_page));
     }
 
-    ctx->data  = ggml_metal_host_malloc(size);
-    ctx->metal = [ggml_backend_metal_get_device() newBufferWithBytesNoCopy:ctx->data
+    id<MTLDevice> device = ggml_backend_metal_get_device();
+
+    ctx->all_data = ggml_metal_host_malloc(size_aligned);
+    ctx->all_size = size_aligned;
+    ctx->owned = true;
+    ctx->n_buffers = 1;
+
+    ctx->buffers[0].data = ctx->all_data;
+    ctx->buffers[0].size = size;
+    ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
                     length:size_aligned
                     options:MTLResourceStorageModeShared
                     deallocator:nil];
 
-    return ggml_backend_buffer_init(buft, metal_backend_buffer_i, ctx, size);
+    if (ctx->buffers[0].metal == nil) {
+        GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
+        free(ctx);
+        ggml_backend_metal_free_device();
+        return NULL;
+    }
+
+    GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
+
+
+#if TARGET_OS_OSX
+    GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
+            device.currentAllocatedSize / 1024.0 / 1024.0,
+            device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+
+    if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
+        GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
+    } else {
+        GGML_METAL_LOG_INFO("\n");
+    }
+#else
+    GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
+#endif
+
+
+    return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
 }
 
 static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
@@ -2470,16 +2655,24 @@ static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_t
 static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
     return ggml_backend_is_metal(backend) || ggml_backend_is_cpu(backend);
 
-    GGML_UNUSED(buft);
+    UNUSED(buft);
+}
+
+static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    UNUSED(buft);
 }
 
 ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
     static struct ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
         /* .iface = */ {
+            /* .get_name         = */ ggml_backend_metal_buffer_type_get_name,
             /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_metal_buffer_type_get_alignment,
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .supports_backend = */ ggml_backend_metal_buffer_type_supports_backend,
+            /* .is_host          = */ ggml_backend_metal_buffer_type_is_host,
         },
         /* .context = */ NULL,
     };
@@ -2487,6 +2680,95 @@ ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
     return &ggml_backend_buffer_type_metal;
 }
 
+// buffer from ptr
+
+ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
+    struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
+
+    ctx->all_data = data;
+    ctx->all_size = size;
+    ctx->owned = false;
+    ctx->n_buffers = 0;
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+
+    // page-align the data ptr
+    {
+        const uintptr_t offs = (uintptr_t) data % size_page;
+        data  = (void *) ((char *) data - offs);
+        size += offs;
+    }
+
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += (size_page - (size_aligned % size_page));
+    }
+
+    id<MTLDevice> device = ggml_backend_metal_get_device();
+
+    // the buffer fits into the max buffer size allowed by the device
+    if (size_aligned <= device.maxBufferLength) {
+        ctx->buffers[ctx->n_buffers].data = data;
+        ctx->buffers[ctx->n_buffers].size = size;
+
+        ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+        if (ctx->buffers[ctx->n_buffers].metal == nil) {
+            GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
+            return false;
+        }
+
+        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
+
+        ++ctx->n_buffers;
+    } else {
+        // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
+        // one of the views
+        const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
+        const size_t size_step = device.maxBufferLength - size_ovlp;
+        const size_t size_view = device.maxBufferLength;
+
+        for (size_t i = 0; i < size; i += size_step) {
+            const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
+
+            ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
+            ctx->buffers[ctx->n_buffers].size = size_step_aligned;
+
+            ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+            if (ctx->buffers[ctx->n_buffers].metal == nil) {
+                GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0);
+                return false;
+            }
+
+            GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, offs = %12ld", __func__, size_step_aligned / 1024.0 / 1024.0, i);
+            if (i + size_step < size) {
+                GGML_METAL_LOG_INFO("\n");
+            }
+
+            ++ctx->n_buffers;
+        }
+    }
+
+#if TARGET_OS_OSX
+    GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
+            device.currentAllocatedSize / 1024.0 / 1024.0,
+            device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+
+    if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
+        GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
+    } else {
+        GGML_METAL_LOG_INFO("\n");
+    }
+#else
+    GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
+#endif
+
+    return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
+}
+
+// backend
+
 static const char * ggml_backend_metal_name(ggml_backend_t backend) {
     return "Metal";
 
@@ -2499,20 +2781,16 @@ static void ggml_backend_metal_free(ggml_backend_t backend) {
     free(backend);
 }
 
-static void ggml_backend_metal_synchronize(ggml_backend_t backend) {
-    UNUSED(backend);
-}
-
 static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
     return ggml_backend_metal_buffer_type();
 
     UNUSED(backend);
 }
 
-static void ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+static bool ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
 
-    ggml_metal_graph_compute(metal_ctx, cgraph);
+    return ggml_metal_graph_compute(metal_ctx, cgraph);
 }
 
 static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
@@ -2521,33 +2799,22 @@ static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct
     UNUSED(backend);
 }
 
-static struct ggml_backend_i metal_backend_i = {
+static struct ggml_backend_i ggml_backend_metal_i = {
     /* .get_name                = */ ggml_backend_metal_name,
     /* .free                    = */ ggml_backend_metal_free,
     /* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_from_async   = */ NULL,
-    /* .cpy_tensor_to_async     = */ NULL,
-    /* .synchronize             = */ ggml_backend_metal_synchronize,
-    /* .graph_plan_create       = */ NULL, // the metal implementation does not require creating graph plans atm
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_metal_graph_compute,
     /* .supports_op             = */ ggml_backend_metal_supports_op,
 };
 
-// TODO: make a common log callback for all backends in ggml-backend
-static void ggml_backend_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
-    fprintf(stderr, "%s", msg);
-
-    UNUSED(level);
-    UNUSED(user_data);
-}
-
 ggml_backend_t ggml_backend_metal_init(void) {
-    ggml_metal_log_set_callback(ggml_backend_log_callback, NULL);
-
     struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
 
     if (ctx == NULL) {
@@ -2557,7 +2824,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
     ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
 
     *metal_backend = (struct ggml_backend) {
-        /* .interface = */ metal_backend_i,
+        /* .interface = */ ggml_backend_metal_i,
         /* .context   = */ ctx,
     };
 
@@ -2565,7 +2832,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
 }
 
 bool ggml_backend_is_metal(ggml_backend_t backend) {
-    return backend->iface.get_name == ggml_backend_metal_name;
+    return backend && backend->iface.get_name == ggml_backend_metal_name;
 }
 
 void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {

ggml-opencl.cpp

@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "ggml-opencl.h"
+#include "ggml-backend-impl.h"
 
 #include <array>
 #include <atomic>
@@ -10,7 +11,7 @@
 #include <sstream>
 #include <vector>
 
-#define CL_TARGET_OPENCL_VERSION 110
+#define CL_TARGET_OPENCL_VERSION 120
 #include <clblast.h>
 
 #if defined(_MSC_VER)
@@ -929,6 +930,12 @@ static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, co
 }
 
 void ggml_cl_init(void) {
+    static bool initialized = false;
+    if (initialized) {
+        return;
+    }
+    initialized = true;
+
     cl_int err;
 
     struct cl_device;
@@ -1483,8 +1490,8 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
     }
-    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
-    cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
+    cl_mem d_Y = src1->backend == GGML_BACKEND_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
+    cl_mem d_D =  dst->backend == GGML_BACKEND_GPU ? (cl_mem)  dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
 
     size_t x_offset = 0;
 
@@ -1501,7 +1508,9 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
 
                 for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
                     // copy src1 to device
-                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
+                    if (src1->backend == GGML_BACKEND_CPU) {
+                        CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
+                    }
 
                     CL_CHECK(clFinish(queue));
 
@@ -1522,8 +1531,10 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
                     }
 
                     // copy dst to host
-                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
+                    if (dst->backend == GGML_BACKEND_CPU) {
+                        float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                        CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
+                    }
                 }
             }
         }
@@ -1532,8 +1543,12 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     if (src0->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_X, x_size);
     }
-    ggml_cl_pool_free(d_Y, y_size);
-    ggml_cl_pool_free(d_D, d_size);
+    if (src1->backend != GGML_BACKEND_GPU) {
+        ggml_cl_pool_free(d_Y, y_size);
+    }
+    if (dst->backend != GGML_BACKEND_GPU) {
+        ggml_cl_pool_free(d_D, d_size);
+    }
 }
 
 static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
@@ -1598,6 +1613,8 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
                     CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
                 }
 
+                // FIXME: convert on device
+
                 for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
                     // convert src1 to fp16
                     // TODO: use multiple threads
@@ -1643,11 +1660,13 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
                     }
 
                     // copy dst to host, then convert to float
-                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
-
-                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-
-                    ggml_fp16_to_fp32_row(tmp, d, d_ne);
+                    if (dst->backend == GGML_BACKEND_CPU) {
+                        CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
+                        float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                        ggml_fp16_to_fp32_row(tmp, d, d_ne);
+                    } else {
+                        // FIXME: convert dst to fp32 on device
+                    }
                 }
             }
         }
@@ -1801,7 +1820,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
 }
 
 
-bool ggml_cl_can_mul_mat(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, const struct ggml_tensor * dst) {
     const int64_t ne10 = src1->ne[0];
 
     const int64_t ne0 = dst->ne[0];
@@ -1895,3 +1914,291 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
     tensor->extra = dst;
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 }
+
+// ggml-backend
+
+// buffer
+
+struct ggml_backend_opencl_buffer_context {
+    ~ggml_backend_opencl_buffer_context() {
+        if (buffer) {
+            clReleaseMemObject(buffer);
+        }
+        for (auto * sub_buffer : sub_buffers) {
+            clReleaseMemObject(sub_buffer);
+        }
+    }
+
+    cl_mem buffer;
+    std::vector<cl_mem> sub_buffers;
+};
+
+static void * const cl_ptr_base = (void *)(uintptr_t) 0x1000;
+
+static const char * ggml_backend_opencl_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return "OpenCL";
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    delete ctx;
+}
+
+static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return cl_ptr_base;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    if (tensor->view_src != NULL && tensor->view_offs == 0) {
+        tensor->extra = tensor->view_src->extra;
+    } else {
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        cl_buffer_region region = {(size_t)((char *)tensor->data - (char *)cl_ptr_base), ggml_nbytes(tensor)};
+        cl_int err;
+        cl_mem sub_buffer = clCreateSubBuffer(ctx->buffer, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        ctx->sub_buffers.push_back(sub_buffer);
+        tensor->extra = sub_buffer;
+    }
+    tensor->backend = GGML_BACKEND_GPU;
+}
+
+static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    cl_mem tensor_buffer = (cl_mem) tensor->extra;
+    CL_CHECK(clEnqueueWriteBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
+    CL_CHECK(clFinish(queue));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    cl_mem tensor_buffer = (cl_mem) tensor->extra;
+    CL_CHECK(clEnqueueReadBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
+    CL_CHECK(clFinish(queue));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    CL_CHECK(clEnqueueFillBuffer(queue, ctx->buffer, &value, sizeof(value), 0, buffer->size, 0, NULL, NULL));
+    CL_CHECK(clFinish(queue));
+}
+
+static void ggml_backend_opencl_buffer_reset(ggml_backend_buffer_t buffer) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    for (auto * sub_buffer : ctx->sub_buffers) {
+        clReleaseMemObject(sub_buffer);
+    }
+    ctx->sub_buffers.clear();
+}
+
+static ggml_backend_buffer_i ggml_backend_opencl_buffer_interface = {
+    /* .get_name        = */ ggml_backend_opencl_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_opencl_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_opencl_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_opencl_buffer_init_tensor,
+    /* .set_tensor      = */ ggml_backend_opencl_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_opencl_buffer_get_tensor,
+    /* .cpy_tensor      = */ NULL,
+    /* .clear           = */ ggml_backend_opencl_buffer_clear,
+    /* .reset           = */ ggml_backend_opencl_buffer_reset,
+};
+
+// buffer type
+
+static const char * ggml_backend_opencl_buffer_type_name(ggml_backend_buffer_type_t buffer_type) {
+    return "OpenCL";
+
+    GGML_UNUSED(buffer_type);
+}
+
+static ggml_backend_buffer_t ggml_backend_opencl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buffer_type, size_t size) {
+    ggml_cl_init();
+
+    cl_int err;
+    cl_mem mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err);
+    if (err != CL_SUCCESS) {
+        fprintf(stderr, "%s: failed to allocate %.2f MiB\n", __func__, size / 1024.0 / 1024.0);
+        return nullptr;
+    }
+
+    ggml_backend_opencl_buffer_context * ctx = new ggml_backend_opencl_buffer_context{mem, {}};
+
+    return ggml_backend_buffer_init(buffer_type, ggml_backend_opencl_buffer_interface, ctx, size);
+}
+
+static size_t ggml_backend_opencl_buffer_type_get_alignment(ggml_backend_buffer_type_t buffer_type) {
+    // FIXME: not thread safe, device may not be initialized yet
+    static cl_uint alignment = -1;
+    if (alignment == (cl_uint)-1) {
+        ggml_cl_init();
+        clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &alignment, NULL);
+    }
+    return alignment;
+
+    GGML_UNUSED(buffer_type);
+}
+
+static bool ggml_backend_opencl_buffer_type_supports_backend(ggml_backend_buffer_type_t buffer_type, ggml_backend_t backend) {
+    //return ggml_backend_is_opencl(backend); // opencl must be used through the cpu backend
+    return ggml_backend_is_cpu(backend);
+
+    GGML_UNUSED(buffer_type);
+}
+
+static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_opencl_buffer_type_name,
+    /* .alloc_buffer     = */ ggml_backend_opencl_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_opencl_buffer_type_get_alignment,
+    /* .get_alloc_size   = */ NULL,
+    /* .supports_backend = */ ggml_backend_opencl_buffer_type_supports_backend,
+    /* .is_host          = */ NULL,
+};
+
+
+ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type() {
+    static ggml_backend_buffer_type buffer_type = {
+        /* .iface   = */ ggml_backend_opencl_buffer_type_interface,
+        /* .context = */ nullptr,
+    };
+
+    return &buffer_type;
+}
+
+#if 0
+// host buffer type
+
+static const char * ggml_backend_opencl_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    return "CL_Host";
+
+    GGML_UNUSED(buft);
+}
+
+static const char * ggml_backend_opencl_host_buffer_name(ggml_backend_buffer_t buffer) {
+    return "CL_Host";
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_cl_host_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_opencl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * ptr = ggml_cl_host_malloc(size);
+
+    if (ptr == nullptr) {
+        // fallback to cpu buffer
+        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_opencl_host_buffer_name;
+    buffer->iface.free_buffer = ggml_backend_opencl_host_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_opencl_buffer_type_host = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_opencl_host_buffer_type_name,
+            /* .alloc_buffer     = */ ggml_backend_opencl_host_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
+            /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
+        },
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_opencl_buffer_type_host;
+}
+
+// backend
+
+static const char * ggml_backend_opencl_name(ggml_backend_t backend) {
+    return "OpenCL";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_opencl_free(ggml_backend_t backend) {
+    GGML_UNUSED(backend);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_opencl_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_opencl_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
+static bool ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
+    for (int i = 0; i < graph->n_nodes; ++i) {
+        ggml_tensor * node = graph->nodes[i];
+        switch (node->op) {
+            case GGML_OP_MUL_MAT:
+                ggml_cl_mul_mat(node->src[0], node->src[1], node, nullptr, 0);
+                break;
+            case GGML_OP_MUL:
+                ggml_cl_mul(node->src[0], node->src[1], node);
+                break;
+            default:
+                GGML_ASSERT(false);
+        }
+    }
+
+    return true;
+
+    GGML_UNUSED(backend);
+}
+
+static bool ggml_backend_opencl_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_MUL_MAT:
+            return ggml_cl_can_mul_mat(op->src[0], op->src[1], op);
+        case GGML_OP_MUL:
+            // return ggml_can_repeat_rows(op->src[1], op->src[0]);
+            return true;
+        default:
+            return false;
+    }
+
+    GGML_UNUSED(backend);
+}
+
+static ggml_backend_i opencl_backend_i = {
+    /* .get_name                = */ ggml_backend_opencl_name,
+    /* .free                    = */ ggml_backend_opencl_free,
+    /* .get_default_buffer_type = */ ggml_backend_opencl_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_from_async   = */ NULL,
+    /* .cpy_tensor_to_async     = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_opencl_graph_compute,
+    /* .supports_op             = */ ggml_backend_opencl_supports_op,
+};
+
+ggml_backend_t ggml_backend_opencl_init() {
+    ggml_backend_t backend = new ggml_backend {
+        /* .interface = */ opencl_backend_i,
+        /* .context   = */ nullptr
+    };
+
+    return backend;
+}
+
+bool ggml_backend_is_opencl(ggml_backend_t backend) {
+    return backend && backend->iface.get_name == ggml_backend_opencl_name;
+}
+#endif

ggml-opencl.h

@@ -1,24 +1,34 @@
 #pragma once
 
 #include "ggml.h"
+#include "ggml-backend.h"
 
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
-void ggml_cl_init(void);
+GGML_API void ggml_cl_init(void);
 
-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);
+GGML_API void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst);
+GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
 
-void * ggml_cl_host_malloc(size_t size);
-void   ggml_cl_host_free(void * ptr);
+// GGML_API void * ggml_cl_host_malloc(size_t size);
+// GGML_API void   ggml_cl_host_free(void * ptr);
 
-void ggml_cl_free_data(const struct ggml_tensor* tensor);
+GGML_API void ggml_cl_free_data(const struct ggml_tensor* tensor);
 
-void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
+GGML_API void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
+
+// backend API
+
+// GGML_API ggml_backend_t ggml_backend_opencl_init(void);
+
+// GGML_API bool ggml_backend_is_opencl(ggml_backend_t backend);
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
+// GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
 
 #ifdef  __cplusplus
 }