CUDA: fix Pascal FA, deq. KV to FP16 for batch > 8 (#7681)

* CUDA: quantized KV support for FA vec

* try CI fix

* fix commented-out kernel variants

* add q8_0 q4_0 tests

* fix nwarps > batch size

* split fattn compile via extern templates

* fix flake8

* fix metal tests

* fix cmake

* make generate_cu_files.py executable

* add autogenerated .cu files

* fix AMD

* error if type_v != FP16 and not flash_attn

* remove obsolete code

.devops/full-cuda.Dockerfile

@@ -31,6 +31,6 @@ ENV LLAMA_CUDA=1
 # Enable cURL
 ENV LLAMA_CURL=1
 
-RUN make
+RUN make -j$(nproc)
 
 ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/full-rocm.Dockerfile

@@ -45,6 +45,6 @@ ENV LLAMA_CURL=1
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev
 
-RUN make
+RUN make -j$(nproc)
 
 ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/full.Dockerfile

@@ -18,7 +18,7 @@ COPY . .
 ENV LLAMA_CURL=1
 
 
-RUN make
+RUN make -j$(nproc)
 
 ENV LC_ALL=C.utf8
 

.devops/main-cuda.Dockerfile

@@ -23,7 +23,7 @@ ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable CUDA
 ENV LLAMA_CUDA=1
 
-RUN make
+RUN make -j$(nproc)
 
 FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
 

.devops/main-intel.Dockerfile

@@ -2,6 +2,14 @@ ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
 
 FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
 
+RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
+    echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
+    chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
+    rm /etc/apt/sources.list.d/intel-graphics.list && \
+    wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
+    echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
+    chmod 644 /usr/share/keyrings/intel-graphics.gpg
+
 ARG LLAMA_SYCL_F16=OFF
 RUN apt-get update && \
     apt-get install -y git

.devops/main-rocm.Dockerfile

@@ -40,6 +40,6 @@ ENV LLAMA_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang
 ENV CXX=/opt/rocm/llvm/bin/clang++
 
-RUN make
+RUN make -j$(nproc)
 
 ENTRYPOINT [ "/app/main" ]

.devops/main.Dockerfile

@@ -9,7 +9,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN make
+RUN make -j$(nproc)
 
 FROM ubuntu:$UBUNTU_VERSION as runtime
 

.devops/nix/package.nix

@@ -214,7 +214,6 @@ effectiveStdenv.mkDerivation (
         (cmakeBool "LLAMA_CUDA" useCuda)
         (cmakeBool "LLAMA_HIPBLAS" useRocm)
         (cmakeBool "LLAMA_METAL" useMetalKit)
-        (cmakeBool "LLAMA_MPI" useMpi)
         (cmakeBool "LLAMA_VULKAN" useVulkan)
         (cmakeBool "LLAMA_STATIC" enableStatic)
       ]
@@ -227,20 +226,20 @@ effectiveStdenv.mkDerivation (
         )
       ]
       ++ 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"
+        (cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.llvm.clang}/bin/clang")
+        (cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets))
       ]
       ++ optionals useMetalKit [
         (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
         (cmakeBool "LLAMA_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
       ];
 
+    # Environment variables needed for ROCm
+    env = optionals useRocm {
+      ROCM_PATH = "${rocmPackages.clr}";
+      HIP_DEVICE_LIB_PATH = "${rocmPackages.rocm-device-libs}/amdgcn/bitcode";
+    };
+
     # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
     # if they haven't been added yet.
     postInstall = ''

.devops/server-cuda.Dockerfile

@@ -25,7 +25,7 @@ ENV LLAMA_CUDA=1
 # Enable cURL
 ENV LLAMA_CURL=1
 
-RUN make
+RUN make -j$(nproc)
 
 FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
 

.devops/server-intel.Dockerfile

@@ -2,6 +2,14 @@ ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
 
 FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
 
+RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
+    echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
+    chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
+    rm /etc/apt/sources.list.d/intel-graphics.list && \
+    wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
+    echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
+    chmod 644 /usr/share/keyrings/intel-graphics.gpg
+
 ARG LLAMA_SYCL_F16=OFF
 RUN apt-get update && \
     apt-get install -y git libcurl4-openssl-dev
@@ -19,6 +27,14 @@ RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
 
 FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
 
+RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
+    echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
+    chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
+    rm /etc/apt/sources.list.d/intel-graphics.list && \
+    wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
+    echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
+    chmod 644 /usr/share/keyrings/intel-graphics.gpg
+
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev
 

.devops/server-rocm.Dockerfile

@@ -45,6 +45,6 @@ ENV LLAMA_CURL=1
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev
 
-RUN make
+RUN make -j$(nproc)
 
 ENTRYPOINT [ "/app/server" ]

.devops/server.Dockerfile

@@ -11,7 +11,7 @@ COPY . .
 
 ENV LLAMA_CURL=1
 
-RUN make
+RUN make -j$(nproc)
 
 FROM ubuntu:$UBUNTU_VERSION as runtime
 

.devops/tools.sh

@@ -8,7 +8,7 @@ arg1="$1"
 shift
 
 if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
-    python3 ./convert.py "$@"
+    python3 ./convert-hf-to-gguf.py "$@"
 elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
     ./quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then

.github/ISSUE_TEMPLATE/01-bug-low.yml

@@ -0,0 +1,50 @@
+name: Low Severity Bugs
+description: Used to report low severity bugs in llama.cpp (e.g. cosmetic issues, non critical UI glitches)
+title: "Bug: "
+labels: ["bug-unconfirmed", "low severity"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to fill out this bug report!
+        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.
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened?
+      description: Also tell us, what did you expect to happen?
+      placeholder: Tell us what you see!
+    validations:
+      required: true
+  - type: textarea
+    id: version
+    attributes:
+      label: Name and Version
+      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
+      placeholder: |
+        $./main --version
+        version: 2999 (42b4109e)
+        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
+    validations:
+      required: true
+  - type: dropdown
+    id: operating-system
+    attributes:
+      label: What operating system are you seeing the problem on?
+      multiple: true
+      options:
+        - Linux
+        - Mac
+        - Windows
+        - BSD
+        - Other? (Please let us know in description)
+    validations:
+      required: false
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
+      render: shell

.github/ISSUE_TEMPLATE/02-bug-medium.yml

@@ -0,0 +1,50 @@
+name: Medium Severity Bug
+description: Used to report medium severity bugs in llama.cpp (e.g. Malfunctioning Features but generally still useable)
+title: "Bug: "
+labels: ["bug-unconfirmed", "medium severity"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to fill out this bug report!
+        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.
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened?
+      description: Also tell us, what did you expect to happen?
+      placeholder: Tell us what you see!
+    validations:
+      required: true
+  - type: textarea
+    id: version
+    attributes:
+      label: Name and Version
+      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
+      placeholder: |
+        $./main --version
+        version: 2999 (42b4109e)
+        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
+    validations:
+      required: true
+  - type: dropdown
+    id: operating-system
+    attributes:
+      label: What operating system are you seeing the problem on?
+      multiple: true
+      options:
+        - Linux
+        - Mac
+        - Windows
+        - BSD
+        - Other? (Please let us know in description)
+    validations:
+      required: false
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
+      render: shell

.github/ISSUE_TEMPLATE/03-bug-high.yml

@@ -0,0 +1,50 @@
+name: High Severity Bug
+description: Used to report high severity bugs in llama.cpp (e.g. Malfunctioning features hindering important common workflow)
+title: "Bug: "
+labels: ["bug-unconfirmed", "high severity"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to fill out this bug report!
+        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.
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened?
+      description: Also tell us, what did you expect to happen?
+      placeholder: Tell us what you see!
+    validations:
+      required: true
+  - type: textarea
+    id: version
+    attributes:
+      label: Name and Version
+      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
+      placeholder: |
+        $./main --version
+        version: 2999 (42b4109e)
+        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
+    validations:
+      required: true
+  - type: dropdown
+    id: operating-system
+    attributes:
+      label: What operating system are you seeing the problem on?
+      multiple: true
+      options:
+        - Linux
+        - Mac
+        - Windows
+        - BSD
+        - Other? (Please let us know in description)
+    validations:
+      required: false
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
+      render: shell

.github/ISSUE_TEMPLATE/04-bug-critical.yml

@@ -0,0 +1,50 @@
+name: Critical Severity Bug
+description: Used to report critical severity bugs in llama.cpp (e.g. Crashing, Corrupted, Dataloss)
+title: "Bug: "
+labels: ["bug-unconfirmed", "critical severity"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to fill out this bug report!
+        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.
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened?
+      description: Also tell us, what did you expect to happen?
+      placeholder: Tell us what you see!
+    validations:
+      required: true
+  - type: textarea
+    id: version
+    attributes:
+      label: Name and Version
+      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
+      placeholder: |
+        $./main --version
+        version: 2999 (42b4109e)
+        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
+    validations:
+      required: true
+  - type: dropdown
+    id: operating-system
+    attributes:
+      label: What operating system are you seeing the problem on?
+      multiple: true
+      options:
+        - Linux
+        - Mac
+        - Windows
+        - BSD
+        - Other? (Please let us know in description)
+    validations:
+      required: false
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
+      render: shell

.github/ISSUE_TEMPLATE/05-enhancement.yml

@@ -0,0 +1,51 @@
+name: Enhancement
+description: Used to request enhancements for llama.cpp
+title: "Feature Request: "
+labels: ["enhancement"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        [Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggerganov/llama.cpp/discussions/categories/ideas)
+
+  - type: checkboxes
+    id: prerequisites
+    attributes:
+      label: Prerequisites
+      description: Please confirm the following before submitting your enhancement request.
+      options:
+        - label: I am running the latest code. Mention the version if possible as well.
+          required: true
+        - label: I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
+          required: true
+        - label: I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
+          required: true
+        - label: I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new and useful enhancement to share.
+          required: true
+
+  - type: textarea
+    id: feature-description
+    attributes:
+      label: Feature Description
+      description: Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
+      placeholder: Detailed description of the enhancement
+    validations:
+      required: true
+
+  - type: textarea
+    id: motivation
+    attributes:
+      label: Motivation
+      description: Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
+      placeholder: Explanation of why this feature is needed and its benefits
+    validations:
+      required: true
+
+  - type: textarea
+    id: possible-implementation
+    attributes:
+      label: Possible Implementation
+      description: If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.
+      placeholder: Detailed description of potential implementation
+    validations:
+      required: false

.github/ISSUE_TEMPLATE/06-research.yml

@@ -0,0 +1,52 @@
+name: Research
+description: Track new technical research area
+title: "Research: "
+labels: ["research 🔬"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Don't forget to check for any [duplicate research issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3A%22research+%F0%9F%94%AC%22)
+
+  - type: checkboxes
+    id: research-stage
+    attributes:
+      label: Research Stage
+      description: Track general state of this research ticket
+      options:
+        - label: Background Research (Let's try to avoid reinventing the wheel)
+        - label: Hypothesis Formed (How do you think this will work and it's effect?)
+        - label: Strategy / Implementation Forming
+        - label: Analysis of results
+        - label: Debrief / Documentation (So people in the future can learn from us)
+
+  - type: textarea
+    id: background
+    attributes:
+      label: Previous existing literature and research
+      description: Whats the current state of the art and whats the motivation for this research?
+
+  - type: textarea
+    id: hypothesis
+    attributes:
+      label: Hypothesis
+      description: How do you think this will work and it's effect?
+
+  - type: textarea
+    id: implementation
+    attributes:
+      label: Implementation
+      description: Got an approach? e.g. a PR ready to go?
+
+  - type: textarea
+    id: analysis
+    attributes:
+      label: Analysis
+      description: How does the proposed implementation behave?
+
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
+      render: shell

.github/ISSUE_TEMPLATE/07-refactor.yml

@@ -0,0 +1,28 @@
+name: Refactor (Maintainers)
+description: Used to track refactoring opportunities
+title: "Refactor: "
+labels: ["refactor"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Don't forget to [check for existing refactor issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3Arefactoring) in case it's already covered.
+        Also you may want to check [Pull request refactor label as well](https://github.com/ggerganov/llama.cpp/pulls?q=is%3Aopen+is%3Apr+label%3Arefactoring) for duplicates too.
+
+  - type: textarea
+    id: background-description
+    attributes:
+      label: Background Description
+      description: Please provide a detailed written description of the pain points you are trying to solve.
+      placeholder: Detailed description behind your motivation to request refactor
+    validations:
+      required: true
+
+  - type: textarea
+    id: possible-approaches
+    attributes:
+      label: Possible Refactor Approaches
+      description: If you have some idea of possible approaches to solve this problem. You may want to make it a todo list.
+      placeholder: Your idea of possible refactoring opportunity/approaches
+    validations:
+      required: false

.github/ISSUE_TEMPLATE/bug.md

@@ -1,11 +0,0 @@
----
-name: Bug template
-about: Used to report bugs in llama.cpp
-labels: ["bug-unconfirmed"]
-assignees: ''
-
----
-
-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.
-
-If the bug concerns the server, please try to reproduce it first using the [server test scenario framework](https://github.com/ggerganov/llama.cpp/tree/master/examples/server/tests).

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,13 @@
+blank_issues_enabled: true
+contact_links:
+  - name: Got an idea?
+    url: https://github.com/ggerganov/llama.cpp/discussions/categories/ideas
+    about: Pop it there. It may then become an enhancement ticket.
+  - name: Got a question?
+    url: https://github.com/ggerganov/llama.cpp/discussions/categories/q-a
+    about: Ask a question there!
+  - name: Want to contribute?
+    url: https://github.com/ggerganov/llama.cpp/wiki/contribute
+    about: Head to the contribution guide page of the wiki for areas you can help with
+
+

.github/ISSUE_TEMPLATE/enhancement.md

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

.github/labeler.yml

@@ -0,0 +1,90 @@
+# https://github.com/actions/labeler
+Kompute:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-kompute.h
+            - ggml-kompute.cpp
+            - README-kompute.md
+Apple Metal:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-metal.h
+            - ggml-metal.cpp
+            - README-metal.md
+SYCL:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-sycl.h
+            - ggml-sycl.cpp
+            - README-sycl.md
+Nvidia GPU:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-cuda.h
+            - ggml-cuda/**
+Vulkan:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml_vk_generate_shaders.py
+            - ggml-vulkan*
+documentation:
+    - changed-files:
+        - any-glob-to-any-file:
+            - docs/**
+            - media/**
+testing:
+    - changed-files:
+        - any-glob-to-any-file:
+            - tests/**
+build:
+    - changed-files:
+        - any-glob-to-any-file:
+            - cmake/**
+            - CMakeLists.txt
+            - CMakePresets.json
+            - codecov.yml
+examples:
+    - changed-files:
+        - any-glob-to-any-file: examples/**
+devops:
+    - changed-files:
+        - any-glob-to-any-file:
+            - .devops/**
+            - .github/**
+            - ci/**
+python:
+    - changed-files:
+        - any-glob-to-any-file:
+            - "**/*.py"
+            - requirements/**
+            - gguf-py/**
+            - .flake8
+script:
+    - changed-files:
+        - any-glob-to-any-file:
+            - scripts/**
+android:
+    - changed-files:
+        - any-glob-to-any-file:
+            - examples/llama.android/**
+server:
+    - changed-files:
+        - any-glob-to-any-file:
+            - examples/server/**
+ggml:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml.c
+            - ggml.h
+            - ggml-*.c
+            - ggml-*.h
+            - ggml-cuda/**
+nix:
+    - changed-files:
+        - any-glob-to-any-file:
+            - "**/*.nix"
+            - .github/workflows/nix-*.yml
+            - .devops/nix/nixpkgs-instances.nix
+embedding:
+    - changed-files:
+        - any-glob-to-any-file: examples/embedding/

.github/workflows/build.yml

@@ -271,49 +271,15 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip
           name: llama-bin-ubuntu-x64.zip
 
-#  ubuntu-latest-cmake-sanitizer:
-#    runs-on: ubuntu-latest
-#
-#    continue-on-error: true
-#
-#    strategy:
-#      matrix:
-#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-#        build_type: [Debug, Release]
-#
-#    steps:
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v4
-#
-#      - name: Dependencies
-#        id: depends
-#        run: |
-#          sudo apt-get update
-#          sudo apt-get install build-essential
-#
-#      - name: Build
-#        id: cmake_build
-#        run: |
-#          mkdir build
-#          cd build
-#          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
-#          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
-#
-#      - name: Test
-#        id: cmake_test
-#        run: |
-#          cd build
-#          ctest -L main --verbose --timeout 900
-
-  ubuntu-latest-cmake-mpi:
+  ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
 
     continue-on-error: true
 
     strategy:
       matrix:
-        mpi_library: [mpich, libopenmpi-dev]
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug, Release]
 
     steps:
       - name: Clone
@@ -324,21 +290,21 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential ${{ matrix.mpi_library }}
+          sudo apt-get install build-essential
 
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake -DLLAMA_MPI=ON ..
-          cmake --build . --config Release -j $(nproc)
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Test
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose
+          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-rpc:
     runs-on: ubuntu-latest
@@ -392,6 +358,33 @@ jobs:
           cmake -DLLAMA_VULKAN=ON ..
           cmake --build . --config Release -j $(nproc)
 
+  ubuntu-22-cmake-hip:
+    runs-on: ubuntu-22.04
+    container: rocm/dev-ubuntu-22.04:6.0.2
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev
+
+      - name: Build with native CMake HIP support
+        id: cmake_build
+        run: |
+          cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DLLAMA_HIPBLAS=ON
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Build with legacy HIP support
+        id: cmake_build_legacy_hip
+        run: |
+          cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DLLAMA_HIPBLAS=ON
+          cmake --build build2 --config Release -j $(nproc)
+
   ubuntu-22-cmake-sycl:
     runs-on: ubuntu-22.04
 
@@ -989,6 +982,37 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
+  windows-latest-cmake-hip:
+    runs-on: windows-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Install
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+
+      - name: Build
+        id: cmake_build
+        run: |
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DLLAMA_HIPBLAS=ON
+          cmake --build build --config Release
+
   ios-xcode-build:
     runs-on: macos-latest
 

.github/workflows/labeler.yml

@@ -0,0 +1,17 @@
+name: "Pull Request Labeler"
+on:
+- pull_request_target
+
+jobs:
+  labeler:
+    permissions:
+      contents: read
+      pull-requests: write
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        repository: "ggerganov/llama.cpp"
+    - uses: actions/labeler@v5
+      with:
+        configuration-path: '.github/labeler.yml'

.github/workflows/server.yml

@@ -32,10 +32,8 @@ jobs:
 
     strategy:
       matrix:
-        # TODO: temporary disabled due to linux kernel issues
-        #sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        sanitizer: [UNDEFINED]
-        build_type: [Debug]
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [RelWithDebInfo]
         include:
           - build_type: Release
             sanitizer: ""
@@ -102,10 +100,8 @@ jobs:
               -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target server
 
-
       - name: Tests
         id: server_integration_tests
-        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
         run: |
           cd examples/server/tests
           PORT=8888 ./tests.sh

.github/workflows/zig-build.yml

@@ -1,29 +0,0 @@
-name: Zig CI
-
-on:
-  pull_request:
-  push:
-    branches:
-      - master
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  build:
-    strategy:
-      fail-fast: false
-      matrix:
-        runs-on: [ubuntu-latest, macos-latest, windows-latest]
-    runs-on: ${{ matrix.runs-on }}
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          submodules: recursive
-          fetch-depth: 0
-      - uses: goto-bus-stop/setup-zig@v2
-        with:
-          version: 0.11.0
-      - name: Build Summary
-        run: zig build --summary all -freference-trace

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.14)  # for add_link_options and implicit target directories.
+cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
 project("llama.cpp" C CXX)
 include(CheckIncludeFileCXX)
 
@@ -72,11 +72,13 @@ else()
     set(INS_ENB ON)
 endif()
 
+option(LLAMA_SVE                             "llama: enable SVE"                                OFF)
 option(LLAMA_AVX                             "llama: enable AVX"                                ${INS_ENB})
 option(LLAMA_AVX2                            "llama: enable AVX2"                               ${INS_ENB})
 option(LLAMA_AVX512                          "llama: enable AVX512"                             OFF)
 option(LLAMA_AVX512_VBMI                     "llama: enable AVX512-VBMI"                        OFF)
 option(LLAMA_AVX512_VNNI                     "llama: enable AVX512-VNNI"                        OFF)
+option(LLAMA_AVX512_BF16                     "llama: enable AVX512-BF16"                        OFF)
 option(LLAMA_FMA                             "llama: enable FMA"                                ${INS_ENB})
 # in MSVC F16C is implied with AVX2/AVX512
 if (NOT MSVC)
@@ -104,6 +106,7 @@ set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                              "llama: max. batch size for using peer access")
 option(LLAMA_CUDA_NO_PEER_COPY               "llama: do not use peer to peer copies"            OFF)
 option(LLAMA_CUDA_NO_VMM                     "llama: do not try to use CUDA VMM"                OFF)
+option(LLAMA_CUDA_FA_ALL_QUANTS              "llama: compile all quants for FlashAttention"     OFF)
 
 option(LLAMA_CURL                            "llama: use libcurl to download model from an URL" OFF)
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
@@ -122,9 +125,7 @@ set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
                                              "llama: metal minimum macOS version")
 set(LLAMA_METAL_STD "" CACHE STRING          "llama: metal standard version (-std flag)")
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
-option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_RPC                             "llama: use RPC"                                   OFF)
-option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
 set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
@@ -134,6 +135,8 @@ set(LLAMA_SCHED_MAX_COPIES  "4" CACHE STRING "llama: max input copies for pipeli
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
+option(LLAMA_LASX                            "llama: enable lasx"                               ON)
+option(LLAMA_LSX                             "llama: enable lsx"                                ON)
 
 # add perf arguments
 option(LLAMA_PERF                            "llama: enable perf"                               OFF)
@@ -382,10 +385,6 @@ if (LLAMA_LLAMAFILE)
     set(GGML_SOURCES_LLAMAFILE sgemm.cpp)
 endif()
 
-if (LLAMA_QKK_64)
-    add_compile_definitions(GGML_QKK_64)
-endif()
-
 if (LLAMA_CUBLAS)
     message(WARNING "LLAMA_CUBLAS is deprecated and will be removed in the future.\nUse LLAMA_CUDA instead")
     set(LLAMA_CUDA ON)
@@ -404,6 +403,8 @@ if (LLAMA_CUDA)
 
         file(GLOB GGML_SOURCES_CUDA "ggml-cuda/*.cu")
         list(APPEND GGML_SOURCES_CUDA "ggml-cuda.cu")
+        file(GLOB SRCS "ggml-cuda/template-instances/fattn-wmma*.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
 
         add_compile_definitions(GGML_USE_CUDA)
         add_compile_definitions(GGML_CUDA_USE_GRAPHS)
@@ -429,6 +430,18 @@ if (LLAMA_CUDA)
         if (LLAMA_CUDA_NO_PEER_COPY)
             add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
         endif()
+        if (LLAMA_CUDA_FA_ALL_QUANTS)
+            file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+        else()
+            file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        endif()
 
         if (LLAMA_STATIC)
             if (WIN32)
@@ -466,35 +479,6 @@ if (LLAMA_CUDA)
     endif()
 endif()
 
-if (LLAMA_MPI)
-    cmake_minimum_required(VERSION 3.10)
-    find_package(MPI)
-    if (MPI_C_FOUND)
-        message(STATUS "MPI found")
-
-        set(GGML_HEADERS_MPI ggml-mpi.h)
-        set(GGML_SOURCES_MPI ggml-mpi.c)
-
-        add_compile_definitions(GGML_USE_MPI)
-        add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
-
-        if (NOT MSVC)
-            add_compile_options(-Wno-cast-qual)
-        endif()
-
-        set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
-        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
-
-        # Even if you're only using the C header, C++ programs may bring in MPI
-        # C++ functions, so more linkage is needed
-        if (MPI_CXX_FOUND)
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}     ${MPI_CXX_LIBRARIES})
-        endif()
-    else()
-        message(WARNING "MPI not found")
-    endif()
-endif()
-
 if (LLAMA_RPC)
     add_compile_definitions(GGML_USE_RPC)
 
@@ -532,6 +516,12 @@ if (LLAMA_VULKAN)
 
         add_compile_definitions(GGML_USE_VULKAN)
 
+        # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
+        # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
+        if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+            add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
+        endif()
+
         if (LLAMA_VULKAN_CHECK_RESULTS)
             add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
         endif()
@@ -555,16 +545,37 @@ if (LLAMA_VULKAN)
 endif()
 
 if (LLAMA_HIPBLAS)
-    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
+    if ($ENV{ROCM_PATH})
+        set(ROCM_PATH $ENV{ROCM_PATH})
+    else()
+        set(ROCM_PATH /opt/rocm)
+    endif()
+    list(APPEND CMAKE_PREFIX_PATH ${ROCM_PATH})
 
-    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
-        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
+    # CMake on Windows doesn't support the HIP language yet
+    if(WIN32)
+        set(CXX_IS_HIPCC TRUE)
+    else()
+        string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
     endif()
 
-    if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
-        message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
-    endif()
+    if(CXX_IS_HIPCC)
+        if(LINUX)
+            if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+                message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+            endif()
 
+            message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
+                    " Prefer setting the HIP compiler directly. See README for details.")
+        endif()
+    else()
+        # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
+        if(AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
+            set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
+        endif()
+        cmake_minimum_required(VERSION 3.21)
+        enable_language(HIP)
+    endif()
     find_package(hip     REQUIRED)
     find_package(hipblas REQUIRED)
     find_package(rocblas REQUIRED)
@@ -575,6 +586,8 @@ if (LLAMA_HIPBLAS)
 
     file(GLOB GGML_SOURCES_ROCM "ggml-cuda/*.cu")
     list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
+    file(GLOB SRCS "ggml-cuda/template-instances/fattn-wmma*.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
 
     add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUDA)
 
@@ -594,17 +607,35 @@ if (LLAMA_HIPBLAS)
         add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
     endif()
 
+    if (LLAMA_CUDA_FA_ALL_QUANTS)
+        file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+    else()
+        file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+        file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+        file(GLOB SRCS "ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    endif()
+
     add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
     add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
     add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
 
-    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
+    if (CXX_IS_HIPCC)
+        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device)
+    else()
+        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
+    endif()
 
     if (LLAMA_STATIC)
         message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
     endif()
 
-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} PUBLIC hip::host roc::rocblas roc::hipblas)
 endif()
 
 if (LLAMA_SYCL)
@@ -627,6 +658,10 @@ if (LLAMA_SYCL)
         add_compile_definitions(GGML_SYCL_F16)
     endif()
 
+    if (LLAMA_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_SYCL_FORCE_MMQ)
+    endif()
+
     add_compile_options(-I./) #include DPCT
     add_compile_options(-I/${SYCL_INCLUDE_DIR})
 
@@ -1040,6 +1075,9 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
             list(APPEND ARCH_FLAGS -mno-unaligned-access)
         endif()
+        if (LLAMA_SVE)
+            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
+        endif()
     endif()
 elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
         (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
@@ -1064,6 +1102,10 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
                 add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
                 add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
             endif()
+            if (LLAMA_AVX512_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
+            endif()
         elseif (LLAMA_AVX2)
             list(APPEND ARCH_FLAGS /arch:AVX2)
         elseif (LLAMA_AVX)
@@ -1095,6 +1137,9 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
         if (LLAMA_AVX512_VNNI)
             list(APPEND ARCH_FLAGS -mavx512vnni)
         endif()
+        if (LLAMA_AVX512_BF16)
+            list(APPEND ARCH_FLAGS -mavx512bf16)
+        endif()
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
     message(STATUS "PowerPC detected")
@@ -1104,6 +1149,17 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
         list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
         #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
     endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+    message(STATUS "loongarch64 detected")
+
+    list(APPEND ARCH_FLAGS -march=loongarch64)
+    if (LLAMA_LASX)
+        list(APPEND ARCH_FLAGS -mlasx)
+    endif()
+    if (LLAMA_LSX)
+        list(APPEND ARCH_FLAGS -mlsx)
+    endif()
+
 else()
     message(STATUS "Unknown architecture")
 endif()
@@ -1192,7 +1248,6 @@ add_library(ggml OBJECT
             ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}
             ${GGML_SOURCES_OPENCL}    ${GGML_HEADERS_OPENCL}
             ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_MPI}       ${GGML_HEADERS_MPI}
             ${GGML_SOURCES_RPC}       ${GGML_HEADERS_RPC}
             ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
             ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
@@ -1280,7 +1335,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
 
 set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
         "${GGML_HEADERS_CUDA}"  "${GGML_HEADERS_OPENCL}"
-        "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
+        "${GGML_HEADERS_METAL}" "${GGML_HEADERS_EXTRA}")
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 install(TARGETS ggml PUBLIC_HEADER)
@@ -1289,7 +1344,7 @@ set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
 
 install(
-    FILES convert.py
+    FILES convert-hf-to-gguf.py
     PERMISSIONS
         OWNER_READ
         OWNER_WRITE

CMakePresets.json

@@ -1,4 +1,4 @@
-{
+{
   "version": 4,
   "configurePresets": [
     {
@@ -40,6 +40,10 @@
 
     { "name": "arm64-windows-msvc-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
     { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "release" ] },
-    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "release", "static" ] }
+    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "release", "static" ] },
+
+    { "name": "x64-windows-msvc-debug"  , "inherits": [ "base", "debug"   ] },
+    { "name": "x64-windows-msvc-release", "inherits": [ "base", "release" ] },
+    { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "release", "static" ] }
   ]
 }

Makefile

@@ -379,15 +379,16 @@ ifneq ($(filter ppc64le%,$(UNAME_M)),)
 	CUDA_POWER_ARCH = 1
 endif
 
+ifneq ($(filter loongarch64%,$(UNAME_M)),)
+	MK_CFLAGS   += -mlasx
+	MK_CXXFLAGS += -mlasx
+endif
+
 else
 	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
 
-ifdef LLAMA_QKK_64
-	MK_CPPFLAGS += -DGGML_QKK_64
-endif
-
 ifndef LLAMA_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
 	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
@@ -399,13 +400,6 @@ ifndef LLAMA_NO_ACCELERATE
 	endif
 endif # LLAMA_NO_ACCELERATE
 
-ifdef LLAMA_MPI
-	MK_CPPFLAGS += -DGGML_USE_MPI
-	MK_CFLAGS   += -Wno-cast-qual
-	MK_CXXFLAGS += -Wno-cast-qual
-	OBJS        += ggml-mpi.o
-endif # LLAMA_MPI
-
 ifdef LLAMA_OPENBLAS
 	MK_CPPFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags-only-I openblas)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
@@ -427,6 +421,15 @@ ifdef LLAMA_CUBLAS
 	LLAMA_CUDA := 1
 endif
 
+OBJS_CUDA_TEMP_INST      = $(patsubst %.cu,%.o,$(wildcard ggml-cuda/template-instances/fattn-wmma*.cu))
+ifdef LLAMA_CUDA_FA_ALL_QUANTS
+	OBJS_CUDA_TEMP_INST += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/template-instances/fattn-vec*.cu))
+else
+	OBJS_CUDA_TEMP_INST += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu))
+	OBJS_CUDA_TEMP_INST += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu))
+	OBJS_CUDA_TEMP_INST += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/template-instances/fattn-vec*f16-f16.cu))
+endif # LLAMA_CUDA_FA_ALL_QUANTS
+
 ifdef LLAMA_CUDA
 	ifneq ('', '$(wildcard /opt/cuda)')
 		CUDA_PATH ?= /opt/cuda
@@ -437,6 +440,7 @@ ifdef LLAMA_CUDA
 	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
 	OBJS         += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
+	OBJS         += $(OBJS_CUDA_TEMP_INST)
 	MK_NVCCFLAGS += -use_fast_math
 ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
@@ -447,6 +451,9 @@ endif # JETSON_EOL_MODULE_DETECT
 ifdef LLAMA_DEBUG
 	MK_NVCCFLAGS += -lineinfo
 endif # LLAMA_DEBUG
+ifdef LLAMA_CUDA_DEBUG
+	MK_NVCCFLAGS += --device-debug
+endif # LLAMA_CUDA_DEBUG
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(CCACHE) $(LLAMA_CUDA_NVCC)
 else
@@ -496,7 +503,10 @@ ifdef LLAMA_CUDA_NO_PEER_COPY
 endif # LLAMA_CUDA_NO_PEER_COPY
 ifdef LLAMA_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
-endif
+endif # LLAMA_CUDA_CCBIN
+ifdef LLAMA_CUDA_FA_ALL_QUANTS
+	MK_NVCCFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
+endif # LLAMA_CUDA_FA_ALL_QUANTS
 
 ifdef JETSON_EOL_MODULE_DETECT
 define NVCC_COMPILE
@@ -508,7 +518,7 @@ define NVCC_COMPILE
 endef # NVCC_COMPILE
 endif # JETSON_EOL_MODULE_DETECT
 
-ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
+ggml-cuda/%.o: ggml-cuda/%.cu ggml.h ggml-common.h ggml-cuda/common.cuh
 	$(NVCC_COMPILE)
 
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
@@ -560,10 +570,10 @@ endif # LLAMA_VULKAN
 ifdef LLAMA_HIPBLAS
 	ifeq ($(wildcard /opt/rocm),)
 		ROCM_PATH	?= /usr
-		GPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
+		AMDGPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
 	else
 		ROCM_PATH	?= /opt/rocm
-		GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+		AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 	endif
 	HIPCC                   ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
 	LLAMA_CUDA_DMMV_X       ?= 32
@@ -574,8 +584,9 @@ ifdef LLAMA_HIP_UMA
 	MK_CPPFLAGS += -DGGML_HIP_UMA
 endif # LLAMA_HIP_UMA
 	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
+	MK_LDFLAGS  += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
 	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
-	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
+	HIPFLAGS    += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
 	HIPFLAGS    += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 	HIPFLAGS    += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
 	HIPFLAGS    += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
@@ -587,11 +598,12 @@ ifdef LLAMA_CUDA_NO_PEER_COPY
 endif # LLAMA_CUDA_NO_PEER_COPY
 	OBJS        += ggml-cuda.o
 	OBJS        += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
+	OBJS        += $(OBJS_CUDA_TEMP_INST)
 
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
 
-ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
+ggml-cuda/%.o: ggml-cuda/%.cu ggml.h ggml-common.h ggml-cuda/common.cuh
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
 
 endif # LLAMA_HIPBLAS
@@ -629,11 +641,6 @@ ggml-metal-embed.o: ggml-metal.metal ggml-common.h
 endif
 endif # LLAMA_METAL
 
-ifdef LLAMA_MPI
-ggml-mpi.o: ggml-mpi.c ggml-mpi.h
-	$(CC) $(CFLAGS) -c $< -o $@
-endif # LLAMA_MPI
-
 ifndef LLAMA_NO_LLAMAFILE
 sgemm.o: sgemm.cpp sgemm.h ggml.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -756,6 +763,7 @@ libllama.a: llama.o ggml.o $(OBJS) $(COMMON_DEPS)
 clean:
 	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult lookup-create lookup-merge lookup-stats common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 	rm -vrf ggml-cuda/*.o
+	rm -vrf ggml-cuda/template-instances/*.o
 	find examples pocs -type f -name "*.o" -delete
 
 #

README-sycl.md

@@ -54,10 +54,10 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
 
 ## OS
 
-| OS      | Status  | Verified                           |
-|---------|---------|------------------------------------|
-| Linux   | Support | Ubuntu 22.04, Fedora Silverblue 39 |
-| Windows | Support | Windows 11                         |
+| OS      | Status  | Verified                                       |
+|---------|---------|------------------------------------------------|
+| Linux   | Support | Ubuntu 22.04, Fedora Silverblue 39, Arch Linux |
+| Windows | Support | Windows 11                                     |
 
 
 ## Hardware
@@ -70,7 +70,7 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
 |-------------------------------|---------|---------------------------------------|
 | Intel Data Center Max Series  | Support | Max 1550, 1100                        |
 | Intel Data Center Flex Series | Support | Flex 170                              |
-| Intel Arc Series              | Support | Arc 770, 730M                         |
+| Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
 | Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
 

README.md

@@ -2,7 +2,9 @@
 
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
-[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) [![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
+[![Conan Center](https://shields.io/conan/v/llama-cpp)](https://conan.io/center/llama-cpp)
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
 
@@ -20,7 +22,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
-- **Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021**
+- **`convert.py` has been deprecated and moved to `examples/convert-legacy-llama.py`, please use `convert-hf-to-gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430
+- Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021
 - BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920
 - MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387
 - Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404
@@ -107,7 +110,6 @@ Typically finetunes of the base models below are supported as well.
 - [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
 - [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
 - [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
-- [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
 - [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
 - [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
 - [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
@@ -128,6 +130,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))
 
@@ -141,6 +144,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
 - [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
 - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
+- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 
 **HTTP server**
 
@@ -199,9 +203,14 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
 - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
 - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
+- [AIKit](https://github.com/sozercan/aikit) (MIT)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
+**Tools:**
+
+- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
+
 ---
 
 Here is a typical run using LLaMA v2 13B on M2 Ultra:
@@ -301,7 +310,7 @@ cd llama.cpp
 
 ### Build
 
-In order to build llama.cpp you have three different options.
+In order to build llama.cpp you have four different options.
 
 - Using `make`:
   - On Linux or MacOS:
@@ -310,8 +319,6 @@ In order to build llama.cpp you have three different options.
       make
       ```
 
-      **Note**: for `Debug` builds, run `make LLAMA_DEBUG=1`
-
   - On Windows:
 
     1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
@@ -323,23 +330,32 @@ In order to build llama.cpp you have three different options.
         make
         ```
 
+  - Notes:
+    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
+    - For faster repeated compilation, install [ccache](https://ccache.dev/).
+    - For debug builds, run `make LLAMA_DEBUG=1`
+
 - Using `CMake`:
 
-    ```bash
-    cmake -B build
-    cmake --build build --config Release
-    ```
+  ```bash
+  cmake -B build
+  cmake --build build --config Release
+  ```
 
-    **Note**: for `Debug` builds, there are two cases:
+  **Notes**:
 
-    - Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
+    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
+    - For faster repeated compilation, install [ccache](https://ccache.dev/).
+    - For debug builds, there are two cases:
+
+      1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
 
       ```bash
       cmake -B build -DCMAKE_BUILD_TYPE=Debug
       cmake --build build
       ```
 
-    - Multi-config generators (`-G` param set to Visual Studio, XCode...):
+      2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
 
       ```bash
       cmake -B build -G "Xcode"
@@ -374,6 +390,14 @@ In order to build llama.cpp you have three different options.
     CLBLAST support for use OpenCL GPU acceleration in FreeBSD. Please read
     the instructions for use and activate this options in this document below.
 
+### Homebrew
+
+On Mac and Linux, the homebrew package manager can be used via
+```
+brew install llama.cpp
+```
+The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
+
 ### Metal Build
 
 On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
@@ -382,45 +406,6 @@ To disable the Metal build at compile time use the `LLAMA_NO_METAL=1` flag or th
 When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
 argument.
 
-### MPI Build
-
-MPI lets you distribute the computation over a cluster of machines. Because of the serial nature of LLM prediction, this won't yield any end-to-end speed-ups, but it will let you run larger models than would otherwise fit into RAM on a single machine.
-
-First you will need MPI libraries installed on your system. The two most popular (only?) options are [MPICH](https://www.mpich.org) and [OpenMPI](https://www.open-mpi.org). Either can be installed with a package manager (`apt`, Homebrew, MacPorts, etc).
-
-Next you will need to build the project with `LLAMA_MPI` set to true on all machines; if you're building with `make`, you will also need to specify an MPI-capable compiler (when building with CMake, this is configured automatically):
-
-- Using `make`:
-
-  ```bash
-  make CC=mpicc CXX=mpicxx LLAMA_MPI=1
-  ```
-
-- Using `CMake`:
-
-  ```bash
-  cmake -S . -B build -DLLAMA_MPI=ON
-  ```
-
-Once the programs are built, download/convert the weights on all of the machines in your cluster. The paths to the weights and programs should be identical on all machines.
-
-Next, ensure password-less SSH access to each machine from the primary host, and create a `hostfile` with a list of the hostnames and their relative "weights" (slots). If you want to use localhost for computation, use its local subnet IP address rather than the loopback address or "localhost".
-
-Here is an example hostfile:
-
-```
-192.168.0.1:2
-malvolio.local:1
-```
-
-The above will distribute the computation across 2 processes on the first host and 1 process on the second host. Each process will use roughly an equal amount of RAM. Try to keep these numbers small, as inter-process (intra-host) communication is expensive.
-
-Finally, you're ready to run a computation using `mpirun`:
-
-```bash
-mpirun -hostfile hostfile -n 3 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
-```
-
 ### BLAS Build
 
 Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS and CLBlast. There are currently several different BLAS implementations available for build and use:
@@ -511,10 +496,12 @@ Building the program with BLAS support may lead to some performance improvements
   |--------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
   | LLAMA_CUDA_FORCE_DMMV          | Boolean                | false   | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_X              | Positive integer >= 32 | 32      | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants.                                         |
-  | LLAMA_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                                                                                                                         |
+  | LLAMA_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                               |
+  | LLAMA_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of dequantization + matrix multiplication kernels instead of leveraging Math libraries. |                                                                                                                                         |
   | LLAMA_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
   | LLAMA_CUDA_KQUANTS_ITER        | 1 or 2                 | 2       | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                                                                     |
   | LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
+  | LLAMA_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
 
 - #### hipBLAS
 
@@ -528,13 +515,28 @@ Building the program with BLAS support may lead to some performance improvements
     ```
   - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
     ```bash
-    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ \
-        cmake -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
+        cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
         && cmake --build build --config Release -- -j 16
     ```
     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).
 
+    Note that if you get the following error:
+    ```
+    clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
+    ```
+    Try searching for a directory under `HIP_PATH` that contains the file
+    `oclc_abi_version_400.bc`. Then, add the following to the start of the
+    command: `HIP_DEVICE_LIB_PATH=<directory-you-just-found>`, so something
+    like:
+    ```bash
+    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
+    HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
+        cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+        && cmake --build build -- -j 16
+    ```
+
   - Using `make` (example for target gfx1030, build with 16 CPU threads):
     ```bash
     make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
@@ -543,10 +545,8 @@ Building the program with BLAS support may lead to some performance improvements
   - 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
-    cd build
-    cmake -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release ..
-    cmake --build .
+    cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
+    cmake --build 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`.
@@ -717,7 +717,8 @@ Building the program with BLAS support may lead to some performance improvements
 
 To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
 
-Note: `convert.py` does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face.
+Note: `convert.py` has been moved to `examples/convert-legacy-llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derievatives.
+It does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face.
 
 ```bash
 # obtain the official LLaMA model weights and place them in ./models
@@ -734,10 +735,10 @@ ls ./models
 python3 -m pip install -r requirements.txt
 
 # convert the model to ggml FP16 format
-python3 convert.py models/mymodel/
+python3 convert-hf-to-gguf.py models/mymodel/
 
 # [Optional] for models using BPE tokenizers
-python convert.py models/mymodel/ --vocab-type bpe
+python convert-hf-to-gguf.py models/mymodel/ --vocab-type bpe
 
 # quantize the model to 4-bits (using Q4_K_M method)
 ./quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M

build.zig

@@ -1,172 +0,0 @@
-// Compatible with Zig Version 0.11.0
-const std = @import("std");
-const ArrayList = std.ArrayList;
-const Compile = std.Build.Step.Compile;
-const ConfigHeader = std.Build.Step.ConfigHeader;
-const Mode = std.builtin.Mode;
-const CrossTarget = std.zig.CrossTarget;
-
-const Maker = struct {
-    builder: *std.build.Builder,
-    target: CrossTarget,
-    optimize: Mode,
-    enable_lto: bool,
-
-    include_dirs: ArrayList([]const u8),
-    cflags: ArrayList([]const u8),
-    cxxflags: ArrayList([]const u8),
-    objs: ArrayList(*Compile),
-
-    fn addInclude(m: *Maker, dir: []const u8) !void {
-        try m.include_dirs.append(dir);
-    }
-    fn addProjectInclude(m: *Maker, path: []const []const u8) !void {
-        try m.addInclude(try m.builder.build_root.join(m.builder.allocator, path));
-    }
-    fn addCFlag(m: *Maker, flag: []const u8) !void {
-        try m.cflags.append(flag);
-    }
-    fn addCxxFlag(m: *Maker, flag: []const u8) !void {
-        try m.cxxflags.append(flag);
-    }
-    fn addFlag(m: *Maker, flag: []const u8) !void {
-        try m.addCFlag(flag);
-        try m.addCxxFlag(flag);
-    }
-
-    fn init(builder: *std.build.Builder) !Maker {
-        const target = builder.standardTargetOptions(.{});
-        const zig_version = @import("builtin").zig_version_string;
-        const commit_hash = try std.ChildProcess.exec(
-            .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
-        );
-        try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt(
-            \\int LLAMA_BUILD_NUMBER = {};
-            \\char const *LLAMA_COMMIT = "{s}";
-            \\char const *LLAMA_COMPILER = "Zig {s}";
-            \\char const *LLAMA_BUILD_TARGET = "{s}";
-            \\
-        , .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) }));
-        var m = Maker{
-            .builder = builder,
-            .target = target,
-            .optimize = builder.standardOptimizeOption(.{}),
-            .enable_lto = false,
-            .include_dirs = ArrayList([]const u8).init(builder.allocator),
-            .cflags = ArrayList([]const u8).init(builder.allocator),
-            .cxxflags = ArrayList([]const u8).init(builder.allocator),
-            .objs = ArrayList(*Compile).init(builder.allocator),
-        };
-
-        try m.addCFlag("-std=c11");
-        try m.addCxxFlag("-std=c++11");
-        try m.addProjectInclude(&.{});
-        try m.addProjectInclude(&.{"common"});
-        return m;
-    }
-
-    fn obj(m: *const Maker, name: []const u8, src: []const u8) *Compile {
-        const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
-        if (o.target.getAbi() != .msvc)
-            o.defineCMacro("_GNU_SOURCE", null);
-
-        if (std.mem.endsWith(u8, src, ".c")) {
-            o.addCSourceFiles(&.{src}, m.cflags.items);
-            o.linkLibC();
-        } else {
-            o.addCSourceFiles(&.{src}, m.cxxflags.items);
-            if (o.target.getAbi() == .msvc) {
-                o.linkLibC(); // need winsdk + crt
-            } else {
-                // linkLibCpp already add (libc++ + libunwind + libc)
-                o.linkLibCpp();
-            }
-        }
-        for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
-        o.want_lto = m.enable_lto;
-        return o;
-    }
-
-    fn exe(m: *const Maker, name: []const u8, src: []const u8, deps: []const *Compile) *Compile {
-        const e = m.builder.addExecutable(.{ .name = name, .target = m.target, .optimize = m.optimize });
-        e.addCSourceFiles(&.{src}, m.cxxflags.items);
-        for (deps) |d| e.addObject(d);
-        for (m.objs.items) |o| e.addObject(o);
-        for (m.include_dirs.items) |i| e.addIncludePath(.{ .path = i });
-
-        // https://github.com/ziglang/zig/issues/15448
-        if (e.target.getAbi() == .msvc) {
-            e.linkLibC(); // need winsdk + crt
-        } else {
-            // linkLibCpp already add (libc++ + libunwind + libc)
-            e.linkLibCpp();
-        }
-        m.builder.installArtifact(e);
-        e.want_lto = m.enable_lto;
-        return e;
-    }
-};
-
-pub fn build(b: *std.build.Builder) !void {
-    var make = try Maker.init(b);
-    make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
-
-    const ggml = make.obj("ggml", "ggml.c");
-    const sgemm = make.obj("sgemm", "sgemm.cpp");
-    const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
-    const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
-    const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
-    const unicode = make.obj("unicode", "unicode.cpp");
-    const unicode_data = make.obj("unicode-data", "unicode-data.cpp");
-    const llama = make.obj("llama", "llama.cpp");
-    const buildinfo = make.obj("common", "common/build-info.cpp");
-    const common = make.obj("common", "common/common.cpp");
-    const console = make.obj("console", "common/console.cpp");
-    const sampling = make.obj("sampling", "common/sampling.cpp");
-    const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
-    const json_schema_to_grammar = make.obj("json-schema-to-grammar", "common/json-schema-to-grammar.cpp");
-    const train = make.obj("train", "common/train.cpp");
-    const clip = make.obj("clip", "examples/llava/clip.cpp");
-    const llava = make.obj("llava", "examples/llava/llava.cpp");
-
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
-
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, grammar_parser, clip, llava });
-    if (server.target.isWindows()) {
-        server.linkSystemLibrary("ws2_32");
-    }
-
-    const server_assets = [_][]const u8{ "index.html", "index.js", "completion.js", "json-schema-to-grammar.mjs" };
-    for (server_assets) |asset| {
-        const input_path = b.fmt("examples/server/public/{s}", .{asset});
-        const output_path = b.fmt("examples/server/{s}.hpp", .{asset});
-
-        // Portable equivalent of `b.addSystemCommand(&.{ "xxd", "-n", asset, "-i", input_path, output_path }) })`:
-
-        const input = try std.fs.cwd().readFileAlloc(b.allocator, input_path, std.math.maxInt(usize));
-        defer b.allocator.free(input);
-
-        var buf = std.ArrayList(u8).init(b.allocator);
-        defer buf.deinit();
-
-        for (input) |byte| {
-            try std.fmt.format(buf.writer(), "0x{X:0>2}, ", .{byte});
-        }
-
-        var name = try std.mem.replaceOwned(u8, b.allocator, asset, "-", "_");
-        defer b.allocator.free(name);
-        std.mem.replaceScalar(u8, name, '.', '_');
-
-        try std.fs.cwd().writeFile(output_path, b.fmt(
-            "unsigned char {s}[] = {{{s}}};\nunsigned int {s}_len = {d};\n",
-            .{ name, buf.items, name, input.len },
-        ));
-
-        std.debug.print("Dumped hex of \"{s}\" ({s}) to {s}\n", .{ input_path, name, output_path });
-    }
-}

ci/run.sh

@@ -202,12 +202,15 @@ function gg_sum_test_scripts_release {
 }
 
 function gg_get_model {
-    local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
-    local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
-    if [[ -s $gguf_3b ]]; then
-        echo -n "$gguf_3b"
-    elif [[ -s $gguf_7b ]]; then
-        echo -n "$gguf_7b"
+    local gguf_0="$MNT/models/pythia/1.4B/ggml-model-f16.gguf"
+    local gguf_1="$MNT/models/pythia/2.8B/ggml-model-f16.gguf"
+    local gguf_2="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
+    if [[ -s $gguf_0 ]]; then
+        echo -n "$gguf_0"
+    elif [[ -s $gguf_1 ]]; then
+        echo -n "$gguf_1"
+    elif [[ -s $gguf_2 ]]; then
+        echo -n "$gguf_2"
     else
         echo >&2 "No model found. Can't run gg_run_ctest_with_model."
         exit 1
@@ -256,139 +259,6 @@ function gg_sum_ctest_with_model_release {
     gg_printf '```\n'
 }
 
-# open_llama_3b_v2
-
-function gg_run_open_llama_3b_v2 {
-    cd ${SRC}
-
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/config.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/tokenizer.model
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/tokenizer_config.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/special_tokens_map.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
-
-    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
-    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
-
-    path_models="../models-mnt/open-llama/3B-v2"
-    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
-
-    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
-
-    set -e
-
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                                             ) 2>&1 | tee -a $OUT/${ci}-make.log
-
-    python3 ../convert.py ${path_models}
-
-    model_f16="${path_models}/ggml-model-f16.gguf"
-    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
-    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
-    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
-    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
-    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
-    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
-    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
-    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
-    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
-    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
-
-    wiki_test_60="${path_wiki}/wiki.test-60.raw"
-
-    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
-
-    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
-    (time ./bin/save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-
-    function check_ppl {
-        qnt="$1"
-        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
-            return 20
-        fi
-
-        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
-        return 0
-    }
-
-    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-
-    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
-
-    set +e
-}
-
-function gg_sum_open_llama_3b_v2 {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'OpenLLaMA 3B-v2:\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
-    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
-    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
-    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
-    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
-    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
-    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
-    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
-    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
-    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
-    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
-    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
-    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-}
-
 # open_llama_7b_v2
 # requires: GG_BUILD_CUDA
 
@@ -417,7 +287,7 @@ function gg_run_open_llama_7b_v2 {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert.py ${path_models}
+    python3 ../examples/convert-legacy-llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -526,6 +396,272 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
 }
 
+# pythia_1.4b
+
+function gg_run_pythia_1_4b {
+    cd ${SRC}
+
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/config.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/special_tokens_map.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/resolve/main/pytorch_model.bin
+
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
+
+    path_models="../models-mnt/pythia/1.4B"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (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
+
+    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
+
+    wiki_test_60="${path_wiki}/wiki.test-60.raw"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
+    set +e
+}
+
+function gg_sum_pythia_1_4b {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Pythia 1.4B:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
+}
+
+# pythia_2_8b
+# requires: GG_BUILD_CUDA
+
+function gg_run_pythia_2_8b {
+    cd ${SRC}
+
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/config.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/special_tokens_map.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/resolve/main/pytorch_model.bin
+
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+
+    path_models="../models-mnt/pythia/2.8B"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
+
+    wiki_test="${path_wiki}/wiki.test.raw"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/main --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
+    set +e
+}
+
+function gg_sum_pythia_2_8b {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Pythia 2.8B:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
+}
+
 # bge-small
 
 function gg_run_embd_bge_small {
@@ -552,7 +688,7 @@ function gg_run_embd_bge_small {
     (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
 
-    python3 ../convert-hf-to-gguf.py ${path_models}
+    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -606,9 +742,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
         if [ -z ${GG_BUILD_CUDA} ]; then
-            test $ret -eq 0 && gg_run open_llama_3b_v2
+            test $ret -eq 0 && gg_run pythia_1_4b
         else
-            test $ret -eq 0 && gg_run open_llama_7b_v2
+            test $ret -eq 0 && gg_run pythia_2_8b
+            #test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
         test $ret -eq 0 && gg_run ctest_with_model_debug
         test $ret -eq 0 && gg_run ctest_with_model_release

common/common.h

@@ -27,7 +27,7 @@
 #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
 
 #define print_build_info() do {                                                                     \
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);           \
+    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);      \
     fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
@@ -35,14 +35,18 @@
 
 // build info
 extern int LLAMA_BUILD_NUMBER;
-extern char const *LLAMA_COMMIT;
-extern char const *LLAMA_COMPILER;
-extern char const *LLAMA_BUILD_TARGET;
+extern char const * LLAMA_COMMIT;
+extern char const * LLAMA_COMPILER;
+extern char const * LLAMA_BUILD_TARGET;
 
 struct llama_control_vector_load_info;
 
-int get_math_cpu_count();
-int32_t get_num_physical_cores();
+//
+// CPU utils
+//
+
+int32_t cpu_get_num_physical_cores();
+int32_t cpu_get_num_math();
 
 //
 // CLI argument parsing
@@ -51,7 +55,7 @@ int32_t get_num_physical_cores();
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
-    int32_t n_threads             = get_math_cpu_count();
+    int32_t n_threads             = cpu_get_num_math();
     int32_t n_threads_draft       = -1;
     int32_t n_threads_batch       = -1;    // number of threads to use for batch processing (-1 = use n_threads)
     int32_t n_threads_batch_draft = -1;
@@ -142,6 +146,7 @@ struct gpt_params {
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
     bool interactive_specials = false; // whether to allow special tokens from user, during interactive mode
+    bool special           = false; // enable special token output
     bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
     bool chatml            = false; // chatml mode (used for models trained on chatml syntax)
     bool prompt_cache_all  = false; // save user input and generations to prompt cache
@@ -179,33 +184,34 @@ struct gpt_params {
 
 void gpt_params_handle_model_default(gpt_params & params);
 
-bool parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
+bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);
+bool gpt_params_parse      (int argc, char ** argv, gpt_params & params);
+bool gpt_params_find_arg   (int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
+void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);
 
-bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
-
-bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
-
-void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
-
-bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
-
-std::string get_system_info(const gpt_params & params);
-
-std::string gpt_random_prompt(std::mt19937 & rng);
-
-void process_escapes(std::string& input);
-
-bool validate_file_name(const std::string & filename);
+std::string gpt_params_get_system_info(const gpt_params & params);
 
 //
 // String utils
 //
 
-std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
-std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
 std::vector<std::string> string_split(std::string input, char separator);
+
 std::string string_strip(const std::string & str);
-std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
+std::string string_get_sortable_timestamp();
+std::string string_random_prompt(std::mt19937 & rng);
+
+bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
+void string_process_escapes(std::string & input);
+
+//
+// Filesystem utils
+//
+
+bool fs_validate_filename(const std::string & filename);
+bool fs_create_directory_with_parents(const std::string & path);
+
+std::string fs_get_cache_directory();
 
 //
 // Model utils
@@ -276,29 +282,15 @@ std::string llama_detokenize_bpe(
 // defaults to true when model type is SPM, otherwise false.
 bool llama_should_add_bos_token(const llama_model * model);
 
-//
-// YAML utils
-//
-
-bool create_directory_with_parents(const std::string & path);
-void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data);
-void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data);
-void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data);
-std::string get_sortable_timestamp();
-
-void dump_non_result_info_yaml(
-    FILE * stream, const gpt_params & params, const llama_context * lctx,
-    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
-
 //
 // KV cache utils
 //
 
 // Dump the KV cache view with the number of sequences per cell.
-void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
+void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
 
 // Dump the KV cache view showing individual sequences in each cell (long output).
-void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
+void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
 
 //
 // Embedding utils
@@ -332,6 +324,20 @@ llama_control_vector_data llama_control_vector_load(const std::vector<llama_cont
 //
 // Split utils
 //
+
 static const char * const LLM_KV_SPLIT_NO            = "split.no";
 static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
+
+//
+// YAML utils
+//
+
+void yaml_dump_vector_float    (FILE * stream, const char * prop_name, const std::vector<float> & data);
+void yaml_dump_vector_int      (FILE * stream, const char * prop_name, const std::vector<int> & data);
+void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
+
+void yaml_dump_non_result_info(
+    FILE * stream, const gpt_params & params, const llama_context * lctx,
+    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
+

common/sampling.cpp

@@ -125,7 +125,7 @@ std::string llama_sampling_order_print(const llama_sampling_params & params) {
     std::string result = "CFG -> Penalties ";
     if (params.mirostat == 0) {
         for (auto sampler_type : params.samplers_sequence) {
-            const auto sampler_type_name = sampler_type_to_name_string(sampler_type);
+            const auto sampler_type_name = llama_sampling_type_to_str(sampler_type);
             if (!sampler_type_name.empty()) {
                 result += "-> " + sampler_type_name + " ";
             }
@@ -137,6 +137,87 @@ std::string llama_sampling_order_print(const llama_sampling_params & params) {
     return result;
 }
 
+std::string llama_sampling_type_to_str(llama_sampler_type sampler_type) {
+    switch (sampler_type) {
+        case llama_sampler_type::TOP_K:       return "top_k";
+        case llama_sampler_type::TFS_Z:       return "tfs_z";
+        case llama_sampler_type::TYPICAL_P:   return "typical_p";
+        case llama_sampler_type::TOP_P:       return "top_p";
+        case llama_sampler_type::MIN_P:       return "min_p";
+        case llama_sampler_type::TEMPERATURE: return "temperature";
+        default : return "";
+    }
+}
+
+std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
+    std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
+        {"top_k",       llama_sampler_type::TOP_K},
+        {"top_p",       llama_sampler_type::TOP_P},
+        {"typical_p",   llama_sampler_type::TYPICAL_P},
+        {"min_p",       llama_sampler_type::MIN_P},
+        {"tfs_z",       llama_sampler_type::TFS_Z},
+        {"temperature", llama_sampler_type::TEMPERATURE}
+    };
+
+    // since samplers names are written multiple ways
+    // make it ready for both system names and input names
+    std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
+        {"top-k",       llama_sampler_type::TOP_K},
+        {"top-p",       llama_sampler_type::TOP_P},
+        {"nucleus",     llama_sampler_type::TOP_P},
+        {"typical-p",   llama_sampler_type::TYPICAL_P},
+        {"typical",     llama_sampler_type::TYPICAL_P},
+        {"min-p",       llama_sampler_type::MIN_P},
+        {"tfs-z",       llama_sampler_type::TFS_Z},
+        {"tfs",         llama_sampler_type::TFS_Z},
+        {"temp",        llama_sampler_type::TEMPERATURE}
+    };
+
+    std::vector<llama_sampler_type> sampler_types;
+    sampler_types.reserve(names.size());
+    for (const auto & name : names)
+    {
+        auto sampler_item = sampler_canonical_name_map.find(name);
+        if (sampler_item != sampler_canonical_name_map.end())
+        {
+            sampler_types.push_back(sampler_item->second);
+        }
+        else
+        {
+            if (allow_alt_names)
+            {
+                sampler_item = sampler_alt_name_map.find(name);
+                if (sampler_item != sampler_alt_name_map.end())
+                {
+                    sampler_types.push_back(sampler_item->second);
+                }
+            }
+        }
+    }
+    return sampler_types;
+}
+
+std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string) {
+    std::unordered_map<char, llama_sampler_type> sampler_name_map {
+        {'k', llama_sampler_type::TOP_K},
+        {'p', llama_sampler_type::TOP_P},
+        {'y', llama_sampler_type::TYPICAL_P},
+        {'m', llama_sampler_type::MIN_P},
+        {'f', llama_sampler_type::TFS_Z},
+        {'t', llama_sampler_type::TEMPERATURE}
+    };
+
+    std::vector<llama_sampler_type> sampler_types;
+    sampler_types.reserve(names_string.size());
+    for (const auto & c : names_string) {
+        const auto sampler_item = sampler_name_map.find(c);
+        if (sampler_item != sampler_name_map.end()) {
+            sampler_types.push_back(sampler_item->second);
+        }
+    }
+    return sampler_types;
+}
+
 // no reasons to expose this function in header
 static void sampler_queue(
                    struct llama_context * ctx_main,
@@ -179,7 +260,7 @@ static llama_token llama_sampling_sample_impl(
                   struct llama_context * ctx_main,
                   struct llama_context * ctx_cfg,
                   const int idx,
-                  bool is_resampling) {  // Add a parameter to indicate if we are resampling
+                  bool is_resampling) {
     const llama_sampling_params & params = ctx_sampling->params;
 
     const float   temp            = params.temp;
@@ -188,8 +269,8 @@ static llama_token llama_sampling_sample_impl(
     const float   mirostat_eta    = params.mirostat_eta;
 
     std::vector<float> original_logits;
-    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, !is_resampling, &original_logits);
-    if (!is_resampling) {
+    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, /* apply_grammar= */ is_resampling, &original_logits);
+    if (ctx_sampling->grammar != NULL && !is_resampling) {
         GGML_ASSERT(!original_logits.empty());
     }
     llama_token id = 0;
@@ -252,7 +333,7 @@ static llama_token llama_sampling_sample_impl(
             // 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 llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ true);
         }
     }
 
@@ -285,7 +366,8 @@ static llama_token_data_array llama_sampling_prepare_impl(
     // Get a pointer to the logits
     float * logits = llama_get_logits_ith(ctx_main, idx);
 
-    if (apply_grammar && original_logits != NULL) {
+    if (ctx_sampling->grammar != NULL && !apply_grammar) {
+        GGML_ASSERT(original_logits != NULL);
         // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
         *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
     }
@@ -342,7 +424,7 @@ llama_token llama_sampling_sample(
                   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);
+    return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ false);
 }
 
 llama_token_data_array llama_sampling_prepare(

common/sampling.h

@@ -116,6 +116,11 @@ std::string llama_sampling_print(const llama_sampling_params & params);
 // Print sampling order into a string
 std::string llama_sampling_order_print(const llama_sampling_params & params);
 
+std::string llama_sampling_type_to_str(llama_sampler_type sampler_type);
+
+std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string);
+
 // this is a common sampling function used across the examples for convenience
 // it can serve as a starting point for implementing your own sampling function
 // Note: When using multiple sequences, it is the caller's responsibility to call

common/train.cpp

@@ -1052,7 +1052,7 @@ struct train_params_common get_default_train_params_common() {
 
     params.custom_n_ctx = false;
 
-    params.use_flash              = true;
+    params.use_flash              = false;
     params.use_checkpointing      = true;
 
     params.sample_start           = "";
@@ -1380,7 +1380,7 @@ bool consume_common_train_arg(
 
 void finish_processing_train_args(struct train_params_common * params) {
     if (params->escape) {
-        process_escapes(params->sample_start);
+        string_process_escapes(params->sample_start);
     }
 }
 

convert-hf-to-gguf-update.py

@@ -72,6 +72,7 @@ models = [
     {"name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
     {"name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
     {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
+    {"name": "stablelm2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
     {"name": "refact",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
     {"name": "command-r",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
     {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
@@ -80,6 +81,7 @@ models = [
     {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
     {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
     {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
+    {"name": "smaug-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
 ]
 
 

convert-hf-to-gguf.py

@@ -14,6 +14,7 @@ from pathlib import Path
 from hashlib import sha256
 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Sequence, TypeVar, cast
 
+import math
 import numpy as np
 import torch
 
@@ -24,8 +25,6 @@ if 'NO_LOCAL_GGUF' not in os.environ:
     sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
 
-from convert import LlamaHfVocab
-
 logger = logging.getLogger("hf-to-gguf")
 
 
@@ -312,11 +311,10 @@ class Model:
                         data = data.astype(np.float32)
                     data_qtype = gguf.GGMLQuantizationType.F32
 
-                block_size, type_size = gguf.GGML_QUANT_SIZES[data_qtype]
+                shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
+
                 # reverse shape to make it similar to the internal ggml dimension order
-                shape_str = f"""{{{', '.join(str(n) for n in reversed(
-                    (*data.shape[:-1], data.shape[-1] * data.dtype.itemsize // type_size * block_size))
-                )}}}"""
+                shape_str = f"{{{', '.join(str(n) for n in reversed(shape))}}}"
 
                 # n_dims is implicit in the shape
                 logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
@@ -446,6 +444,9 @@ class Model:
         if chkhsh == "3ce83efda5659b07b1ad37ca97ca5797ea4285d9b9ab0dc679e4a720c9da7454":
             # ref: https://huggingface.co/openai-community/gpt2
             res = "gpt-2"
+        if chkhsh == "32d85c31273f8019248f2559fed492d929ea28b17e51d81d3bb36fff23ca72b3":
+            # ref: https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b
+            res = "stablelm2"
         if chkhsh == "6221ad2852e85ce96f791f476e0b390cf9b474c9e3d1362f53a24a06dc8220ff":
             # ref: https://huggingface.co/smallcloudai/Refact-1_6-base
             res = "refact"
@@ -470,6 +471,9 @@ class Model:
         if chkhsh == "27949a2493fc4a9f53f5b9b029c82689cfbe5d3a1929bb25e043089e28466de6":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-de
             res = "jina-v2-de"
+        if chkhsh == "c136ed14d01c2745d4f60a9596ae66800e2b61fa45643e72436041855ad4089d":
+            # ref: https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct
+            res = "smaug-bpe"
 
         if res is None:
             logger.warning("\n")
@@ -573,6 +577,10 @@ class Model:
 
         vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
 
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+
         for token_id in range(tokenizer.vocab_size()):
             piece = tokenizer.IdToPiece(token_id)
             text = piece.encode("utf-8")
@@ -588,21 +596,23 @@ class Model:
             elif tokenizer.IsByte(token_id):
                 toktype = SentencePieceTokenTypes.BYTE
 
-            tokens.append(text)
-            scores.append(score)
-            toktypes.append(toktype)
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
 
         added_tokens_file = self.dir_model / 'added_tokens.json'
         if added_tokens_file.is_file():
             with open(added_tokens_file, "r", encoding="utf-8") as f:
                 added_tokens_json = json.load(f)
-
                 for key in added_tokens_json:
-                    key = key.encode("utf-8")
-                    if key not in tokens:
-                        tokens.append(key)
-                        scores.append(-1000.0)
-                        toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+                    token_id = added_tokens_json[key]
+                    if (token_id >= vocab_size):
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
+
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
 
         if vocab_size > len(tokens):
             pad_count = vocab_size - len(tokens)
@@ -612,8 +622,6 @@ class Model:
                 scores.append(-1000.0)
                 toktypes.append(SentencePieceTokenTypes.UNUSED)
 
-        assert len(tokens) == vocab_size
-
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_tokenizer_pre("default")
         self.gguf_writer.add_token_list(tokens)
@@ -624,7 +632,7 @@ class Model:
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def _set_vocab_llama_hf(self):
-        vocab = LlamaHfVocab(self.dir_model)
+        vocab = gguf.LlamaHfVocab(self.dir_model)
         tokens = []
         scores = []
         toktypes = []
@@ -665,6 +673,44 @@ class GPTNeoXModel(Model):
         self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+
+        tensors: list[tuple[str, Tensor]] = []
+
+        if re.match(r"gpt_neox\.layers\.\d+\.attention\.query_key_value\.weight", name):
+            # Map bloom-style qkv_linear to gpt-style qkv_linear
+            # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
+            # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
+            qkv_weights = data_torch.reshape((n_head, 3, n_embed // n_head, n_embed))
+            data_torch = torch.cat(
+                (
+                    qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
+                    qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
+                    qkv_weights[:, 2, :, :].reshape((-1, n_embed)),
+                ),
+                dim=0,
+            )
+            logger.info("re-format attention.linear_qkv.weight")
+        elif re.match(r"gpt_neox\.layers\.\d+\.attention\.query_key_value\.bias", name):
+            qkv_bias = data_torch.reshape((n_head, 3, n_embed // n_head))
+            data_torch = torch.cat(
+                (
+                    qkv_bias[:, 0, :].reshape((n_embed,)),
+                    qkv_bias[:, 1, :].reshape((n_embed,)),
+                    qkv_bias[:, 2, :].reshape((n_embed,)),
+                ),
+                dim=0,
+            )
+            logger.info("re-format attention.linear_qkv.bias")
+
+        tensors.append((self.map_tensor_name(name), data_torch))
+
+        return tensors
+
 
 @Model.register("BloomForCausalLM")
 class BloomModel(Model):
@@ -1141,45 +1187,6 @@ class RefactModel(Model):
         return tensors
 
 
-@Model.register("PersimmonForCausalLM")
-class PersimmonModel(Model):
-    model_arch = gguf.MODEL_ARCH.PERSIMMON
-
-    def set_gguf_parameters(self):
-        block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
-        head_count = self.hparams["num_attention_heads"]
-        head_count_kv = head_count
-        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"])
-
-        # 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"])
-        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
-
-    def set_vocab(self):
-        self._set_vocab_sentencepiece()
-        # self.gguf_writer.add_bos_token_id(71013)
-        # self.gguf_writer.add_eos_token_id(71013)
-
-    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
-        del name, new_name, bid, n_dims  # unused
-
-        # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
-        return True
-
-
 @Model.register("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM")
 class StableLMModel(Model):
     model_arch = gguf.MODEL_ARCH.STABLELM
@@ -1308,6 +1315,17 @@ class LlamaModel(Model):
                 self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
                 self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
 
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if "add_prefix_space" in tokenizer_config_json:
+                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
+
+        # Apply to granite small models only
+        if self.hparams.get("vocab_size", 32000) == 49152:
+            self.gguf_writer.add_add_bos_token(False)
+
     @staticmethod
     def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
         if n_head_kv is not None and n_head != n_head_kv:
@@ -1322,9 +1340,9 @@ class LlamaModel(Model):
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
 
-        if name.endswith("q_proj.weight"):
+        if name.endswith(("q_proj.weight", "q_proj.bias")):
             data_torch = LlamaModel.permute(data_torch, n_head, n_head)
-        if name.endswith("k_proj.weight"):
+        if name.endswith(("k_proj.weight", "k_proj.bias")):
             data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
 
         # process the experts separately
@@ -1772,6 +1790,38 @@ class Phi3MiniModel(Model):
                     scores[token_id] = -1000.0
                     toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
 
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {})
+                for token_id, foken_data in added_tokens_decoder.items():
+                    token_id = int(token_id)
+                    token = foken_data["content"].encode("utf-8")
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
+                        assert tokens[token_id] == token
+                    tokens[token_id] = token
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+                    if foken_data.get("special"):
+                        toktypes[token_id] = SentencePieceTokenTypes.CONTROL
+
+        tokenizer_file = self.dir_model / 'tokenizer.json'
+        if tokenizer_file.is_file():
+            with open(tokenizer_file, "r", encoding="utf-8") as f:
+                tokenizer_json = json.load(f)
+                added_tokens = tokenizer_json.get("added_tokens", [])
+                for foken_data in added_tokens:
+                    token_id = int(foken_data["id"])
+                    token = foken_data["content"].encode("utf-8")
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
+                        assert tokens[token_id] == token
+                    tokens[token_id] = token
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+                    if foken_data.get("special"):
+                        toktypes[token_id] = SentencePieceTokenTypes.CONTROL
+
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_tokenizer_pre("default")
         self.gguf_writer.add_token_list(tokens)
@@ -1784,23 +1834,59 @@ class Phi3MiniModel(Model):
     def set_gguf_parameters(self):
         block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
 
-        rot_pct = 1.0
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        n_head_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
         rms_eps = self.find_hparam(["rms_norm_eps"])
+        max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
+        orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
+        rope_dims = n_embd // n_head
 
         self.gguf_writer.add_name("Phi3")
-        self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
-
+        self.gguf_writer.add_context_length(max_pos_embds)
+        self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds)
         self.gguf_writer.add_embedding_length(n_embd)
-        self.gguf_writer.add_feed_forward_length(8192)
+        self.gguf_writer.add_feed_forward_length(self.find_hparam(["intermediate_size"]))
         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_head_count_kv(n_head_kv)
         self.gguf_writer.add_layer_norm_rms_eps(rms_eps)
-        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
+        self.gguf_writer.add_rope_dimension_count(rope_dims)
+        self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"]))
         self.gguf_writer.add_file_type(self.ftype)
 
+        # write rope scaling for long context (128k) model
+        rope_scaling = self.find_hparam(['rope_scaling'], True)
+        if (rope_scaling is None):
+            return
+
+        scale = max_pos_embds / orig_max_pos_embds
+
+        rope_scaling_type = rope_scaling.get('type', '').lower()
+        if len(rope_scaling_type) == 0:
+            raise KeyError('Missing the required key rope_scaling.type')
+
+        if rope_scaling_type == 'su':
+            attn_factor = math.sqrt(1 + math.log(scale) / math.log(orig_max_pos_embds)) if scale > 1.0 else 1.0
+        elif rope_scaling_type == 'yarn':
+            attn_factor = 0.1 * math.log(scale) + 1.0 if scale > 1.0 else 1.0
+        else:
+            raise NotImplementedError(f'The rope scaling type {rope_scaling_type} is not supported yet')
+
+        self.gguf_writer.add_rope_scaling_attn_factors(attn_factor)
+
+        long_factors = rope_scaling.get('long_factor', None)
+        short_factors = rope_scaling.get('short_factor', None)
+
+        if long_factors is None or short_factors is None:
+            raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
+
+        if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
+            raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
+
+        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG]  + ".weight", np.array(long_factors, dtype=np.float32))
+        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
+
 
 @Model.register("PlamoForCausalLM")
 class PlamoModel(Model):
@@ -2318,7 +2404,8 @@ class CommandR2Model(Model):
 
         # max_position_embeddings = 8192 in config.json but model was actually
         # trained on 128k context length
-        self.hparams["max_position_embeddings"] = self.hparams["model_max_length"]
+        # aya-23 models don't have model_max_length specified
+        self.hparams["max_position_embeddings"] = self.find_hparam(["model_max_length", "max_position_embeddings"])
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -2391,6 +2478,236 @@ class JinaBertV2Model(BertModel):
         self.gguf_writer.add_add_eos_token(True)
 
 
+@Model.register("ArcticForCausalLM")
+class ArcticModel(Model):
+    model_arch = gguf.MODEL_ARCH.ARCTIC
+
+    def set_vocab(self):
+        # The reason for using a custom implementation here is that the
+        # snowflake-arctic-instruct model redefined tokens 31998 and 31999 from
+        # tokenizer.model and used them as BOS and EOS instead of adding new tokens.
+        from sentencepiece import SentencePieceProcessor
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        if not tokenizer_path.is_file():
+            logger.error(f'Error: Missing {tokenizer_path}')
+            sys.exit(1)
+
+        # Read the whole vocabulary from the tokenizer.model file
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        # Use the added_tokens_decoder field from tokeniser_config.json as the source
+        # of information about added/redefined tokens and modify them accordingly.
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+
+                if "added_tokens_decoder" in tokenizer_config_json:
+                    added_tokens_decoder = tokenizer_config_json["added_tokens_decoder"]
+                    for token_id, token_json in added_tokens_decoder.items():
+                        token_id = int(token_id)
+                        if (token_id >= vocab_size):
+                            logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                            continue
+
+                        token_content = token_json["content"]
+                        token_type = SentencePieceTokenTypes.USER_DEFINED
+                        token_score = -10000.0
+
+                        # Map unk_token to UNKNOWN, other special tokens to CONTROL
+                        # Set the score to 0.0 as in the original tokenizer.model
+                        if ("special" in token_json) and token_json["special"]:
+                            if token_content == tokenizer_config_json["unk_token"]:
+                                token_type = SentencePieceTokenTypes.UNKNOWN
+                            else:
+                                token_type = SentencePieceTokenTypes.CONTROL
+                            token_score = 0.0
+
+                        logger.info(f"Setting added token {token_id} to '{token_content}' (type: {token_type}, score: {token_score:.2f})")
+                        tokens[token_id] = token_content.encode("utf-8")
+                        toktypes[token_id] = token_type
+                        scores[token_id] = token_score
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+
+        if name.endswith("q_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith("k_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+
+        # process the experts separately
+        if name.find("block_sparse_moe.experts") != -1:
+            n_experts = self.hparams["num_local_experts"]
+
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for wid in ["w1", "w2", "w3"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"layers.{bid}.feed_forward.experts.{wid}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def write_tensors(self):
+        super().write_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
+@Model.register("DeepseekV2ForCausalLM")
+class DeepseekV2Model(Model):
+    model_arch = gguf.MODEL_ARCH.DEEPSEEK2
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+
+        self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None:
+            self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
+        self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
+        self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_value_length(hparams["v_head_dim"])
+        self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_count(hparams["n_routed_experts"])
+        self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"])
+        self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
+        self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
+
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "yarn":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+                self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * hparams["rope_scaling"]["mscale_all_dim"])
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # process the experts separately
+        if name.find("mlp.experts") != -1:
+            n_experts = self.hparams["n_routed_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def write_tensors(self):
+        super().write_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 ###### CONVERSION LOGIC ######
 
 
@@ -2523,7 +2840,12 @@ def main() -> None:
     hparams = Model.load_hparams(dir_model)
 
     with torch.inference_mode():
-        model_class = Model.from_model_architecture(hparams["architectures"][0])
+        try:
+            model_class = Model.from_model_architecture(hparams["architectures"][0])
+        except NotImplementedError:
+            logger.error(f"Model {hparams['architectures'][0]} is not supported")
+            sys.exit(1)
+
         model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file, args.no_lazy)
 
         logger.info("Set model parameters")

convert-persimmon-to-gguf.py

@@ -1,143 +0,0 @@
-#!/usr/bin/env python3
-from __future__ import annotations
-
-import logging
-import argparse
-import os
-import sys
-from pathlib import Path
-from pprint import pprint
-
-import torch
-from sentencepiece import SentencePieceProcessor
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
-import gguf
-
-logger = logging.getLogger("persimmon-to-gguf")
-
-
-def _flatten_dict(dct, tensors, prefix=None):
-    assert isinstance(dct, dict)
-    for key in dct.keys():
-        new_prefix = prefix + '.' + key if prefix is not None else key
-        if isinstance(dct[key], torch.Tensor):
-            tensors[new_prefix] = dct[key]
-        elif isinstance(dct[key], dict):
-            _flatten_dict(dct[key], tensors, new_prefix)
-        else:
-            raise ValueError(type(dct[key]))
-    return None
-
-
-def _get_sentencepiece_tokenizer_info(dir_model: Path):
-    tokenizer_path = dir_model / 'adept_vocab.model'
-    logger.info('getting sentencepiece tokenizer from', tokenizer_path)
-    tokenizer = SentencePieceProcessor(str(tokenizer_path))
-    logger.info('adding tokens')
-    tokens: list[bytes] = []
-    scores: list[float] = []
-    toktypes: list[int] = []
-
-    for i in range(tokenizer.vocab_size()):
-        text: bytes
-        score: float
-
-        piece = tokenizer.id_to_piece(i)
-        text = piece.encode("utf-8")
-        score = tokenizer.get_score(i)
-
-        toktype = 1
-        if tokenizer.is_unknown(i):
-            toktype = 2
-        if tokenizer.is_control(i):
-            toktype = 3
-        if tokenizer.is_unused(i):
-            toktype = 5
-        if tokenizer.is_byte(i):
-            toktype = 6
-
-        tokens.append(text)
-        scores.append(score)
-        toktypes.append(toktype)
-        pass
-    return tokens, scores, toktypes
-
-
-def main():
-    parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file")
-    parser.add_argument("--outfile",             type=Path, help="path to write to; default: based on input")
-    parser.add_argument("--ckpt-path",           type=Path, help="path to persimmon checkpoint .pt file")
-    parser.add_argument("--model-dir",           type=Path, help="directory containing model e.g. 8b_chat_model_release")
-    parser.add_argument("--adept-inference-dir", type=str,  help="path to adept-inference code directory")
-    parser.add_argument("--verbose",  action="store_true",  help="increase output verbosity")
-    args = parser.parse_args()
-    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
-    sys.path.append(str(args.adept_inference_dir))
-    persimmon_model = torch.load(args.ckpt_path)
-    hparams = persimmon_model['args']
-    pprint(hparams)
-    tensors: dict[str, torch.Tensor] = {}
-    _flatten_dict(persimmon_model['model'], tensors, None)
-
-    arch = gguf.MODEL_ARCH.PERSIMMON
-    gguf_writer = gguf.GGUFWriter(args.outfile, gguf.MODEL_ARCH_NAMES[arch])
-
-    block_count = hparams.num_layers
-    head_count = hparams.num_attention_heads
-    head_count_kv = head_count
-    ctx_length = hparams.seq_length
-    hidden_size = hparams.hidden_size
-
-    gguf_writer.add_name('persimmon-8b-chat')
-    gguf_writer.add_context_length(ctx_length)
-    gguf_writer.add_embedding_length(hidden_size)
-    gguf_writer.add_block_count(block_count)
-    gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size)
-    # ref: https://github.com/ggerganov/llama.cpp/pull/4889/commits/eea19039fc52ea2dbd1aab45b59ab4e3e29a3443
-    gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
-    gguf_writer.add_head_count(head_count)
-    gguf_writer.add_head_count_kv(head_count_kv)
-    gguf_writer.add_rope_freq_base(hparams.rotary_emb_base)
-    gguf_writer.add_layer_norm_eps(hparams.layernorm_epsilon)
-
-    tokens, scores, toktypes = _get_sentencepiece_tokenizer_info(args.model_dir)
-    gguf_writer.add_tokenizer_model('llama')
-    gguf_writer.add_tokenizer_pre('default')
-    gguf_writer.add_token_list(tokens)
-    gguf_writer.add_token_scores(scores)
-    gguf_writer.add_token_types(toktypes)
-    gguf_writer.add_bos_token_id(71013)
-    gguf_writer.add_eos_token_id(71013)
-
-    tensor_map = gguf.get_tensor_name_map(arch, block_count)
-    logger.info(tensor_map)
-    for name in tensors.keys():
-        data_torch = tensors[name]
-        if name.endswith(".self_attention.rotary_emb.inv_freq"):
-            continue
-        old_dtype = data_torch.dtype
-        # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
-        data = data_torch.to(torch.float32).squeeze().numpy()
-        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-        if new_name is None:
-            raise ValueError(f"Can not map tensor '{name}'")
-
-        n_dims = len(data.shape)
-        logger.debug(f"{new_name}, n_dims = {str(n_dims)}, {str(old_dtype)} --> {str(data.dtype)}")
-        gguf_writer.add_tensor(new_name, data)
-    logger.info("gguf: write header")
-    gguf_writer.write_header_to_file()
-    logger.info("gguf: write metadata")
-    gguf_writer.write_kv_data_to_file()
-    logger.info("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-    gguf_writer.close()
-
-    logger.info(f"gguf: model successfully exported to '{args.outfile}'")
-
-
-if __name__ == '__main__':
-    main()

docs/HOWTO-add-model.md

@@ -17,7 +17,7 @@ Also, it is important to check that the examples and main ggml backends (CUDA, M
 ### 1. Convert the model to GGUF
 
 This step is done in python with a `convert` script using the [gguf](https://pypi.org/project/gguf/) library.
-Depending on the model architecture, you can use either [convert.py](../convert.py) or [convert-hf-to-gguf.py](../convert-hf-to-gguf.py).
+Depending on the model architecture, you can use either [convert-hf-to-gguf.py](../convert-hf-to-gguf.py) or [examples/convert-legacy-llama.py](../examples/convert-legacy-llama.py) (for `llama/llama2` models in `.pth` format).
 
 The convert script reads the model configuration, tokenizer, tensor names+data and converts them to GGUF metadata and tensors.
 

examples/batched/batched.cpp

@@ -48,7 +48,7 @@ int main(int argc, char ** argv) {
         params.prompt = "Hello my name is";
     }
 
-    process_escapes(params.prompt);
+    string_process_escapes(params.prompt);
 
     // init LLM
 

examples/convert-legacy-llama.py

@@ -24,14 +24,16 @@ from abc import ABC, abstractmethod
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable, Optional
+from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar, Optional
 
 import numpy as np
-from sentencepiece import SentencePieceProcessor
 
 if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
+    # use .parent.parent since we are in "examples" directory
+    sys.path.insert(1, str(Path(__file__).parent.parent / 'gguf-py'))
+
 import gguf
+from gguf import BaseVocab, Vocab, NoVocab, BpeVocab, SentencePieceVocab, LlamaHfVocab
 
 if TYPE_CHECKING:
     from typing_extensions import Self, TypeAlias
@@ -380,306 +382,6 @@ class Metadata:
         return metadata
 
 
-#
-# vocab
-#
-
-
-@runtime_checkable
-class BaseVocab(Protocol):
-    tokenizer_model: ClassVar[str]
-    name: ClassVar[str]
-
-
-class NoVocab(BaseVocab):
-    tokenizer_model = "no_vocab"
-    name = "no_vocab"
-
-    def __repr__(self) -> str:
-        return "<NoVocab for a model without integrated vocabulary>"
-
-
-@runtime_checkable
-class Vocab(BaseVocab, Protocol):
-    vocab_size: int
-    added_tokens_dict: dict[str, int]
-    added_tokens_list: list[str]
-    fname_tokenizer: Path
-
-    def __init__(self, base_path: Path): ...
-    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: ...
-
-
-class BpeVocab(Vocab):
-    tokenizer_model = "gpt2"
-    name = "bpe"
-
-    def __init__(self, base_path: Path):
-        added_tokens: dict[str, int] = {}
-
-        if (fname_tokenizer := base_path / 'vocab.json').exists():
-            # "slow" tokenizer
-            with open(fname_tokenizer, encoding="utf-8") as f:
-                self.vocab = json.load(f)
-
-            try:
-                # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
-                with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f:
-                    added_tokens = json.load(f)
-            except FileNotFoundError:
-                pass
-        else:
-            # "fast" tokenizer
-            fname_tokenizer = base_path / FAST_TOKENIZER_FILE
-
-            # if this fails, FileNotFoundError propagates to caller
-            with open(fname_tokenizer, encoding="utf-8") as f:
-                tokenizer_json = json.load(f)
-
-            tokenizer_model: dict[str, Any] = tokenizer_json['model']
-            if (
-                tokenizer_model['type'] != 'BPE' or tokenizer_model.get('byte_fallback', False)
-                or tokenizer_json['decoder']['type'] != 'ByteLevel'
-            ):
-                raise FileNotFoundError('Cannot find GPT-2 BPE tokenizer')
-
-            self.vocab = tokenizer_model["vocab"]
-
-            if (added := tokenizer_json.get('added_tokens')) is not None:
-                # Added tokens here can be duplicates of the main vocabulary.
-                added_tokens = {item['content']: item['id']
-                                for item in added
-                                if item['content'] not in self.vocab}
-
-        vocab_size   = len(self.vocab)
-        expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
-        actual_ids   = sorted(added_tokens.values())
-        if expected_ids != actual_ids:
-            expected_end_id = vocab_size + len(actual_ids) - 1
-            raise ValueError(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range "
-                             f"{vocab_size} - {expected_end_id}; got {actual_ids}")
-
-        items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
-        self.added_tokens_dict    = added_tokens
-        self.added_tokens_list    = [text for (text, idx) in items]
-        self.vocab_size_base      = vocab_size
-        self.vocab_size           = self.vocab_size_base + len(self.added_tokens_list)
-        self.fname_tokenizer      = fname_tokenizer
-
-    def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()}
-
-        for i, _ in enumerate(self.vocab):
-            yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
-
-    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        for text in self.added_tokens_list:
-            score = -1000.0
-            yield text.encode("utf-8"), score, gguf.TokenType.CONTROL
-
-    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        yield from self.bpe_tokens()
-        yield from self.added_tokens()
-
-    def __repr__(self) -> str:
-        return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
-
-
-class SentencePieceVocab(Vocab):
-    tokenizer_model = "llama"
-    name = "spm"
-
-    def __init__(self, base_path: Path):
-        added_tokens: dict[str, int] = {}
-        if (fname_tokenizer := base_path / 'tokenizer.model').exists():
-            # normal location
-            try:
-                with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f:
-                    added_tokens = json.load(f)
-            except FileNotFoundError:
-                pass
-        elif not (fname_tokenizer := base_path.parent / 'tokenizer.model').exists():
-            # not found in alternate location either
-            raise FileNotFoundError('Cannot find tokenizer.model')
-
-        self.sentencepiece_tokenizer = SentencePieceProcessor()
-        self.sentencepiece_tokenizer.LoadFromFile(str(fname_tokenizer))
-        vocab_size = self.sentencepiece_tokenizer.vocab_size()
-
-        new_tokens       = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
-        expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
-        actual_new_ids   = sorted(new_tokens.keys())
-
-        if expected_new_ids != actual_new_ids:
-            raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
-
-        # Token pieces that were added to the base vocabulary.
-        self.added_tokens_dict  = added_tokens
-        self.added_tokens_list  = [new_tokens[id] for id in actual_new_ids]
-        self.vocab_size_base    = vocab_size
-        self.vocab_size         = self.vocab_size_base + len(self.added_tokens_list)
-        self.fname_tokenizer    = fname_tokenizer
-
-    def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        tokenizer = self.sentencepiece_tokenizer
-        for i in range(tokenizer.vocab_size()):
-            piece = tokenizer.IdToPiece(i)
-            text         = piece.encode("utf-8")
-            score: float = tokenizer.GetScore(i)
-
-            toktype = gguf.TokenType.NORMAL
-            if tokenizer.IsUnknown(i):
-                toktype = gguf.TokenType.UNKNOWN
-            if tokenizer.IsControl(i):
-                toktype = gguf.TokenType.CONTROL
-
-            # NOTE: I think added_tokens are user defined.
-            # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto
-            # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED
-
-            if tokenizer.IsUnused(i):
-                toktype = gguf.TokenType.UNUSED
-            if tokenizer.IsByte(i):
-                toktype = gguf.TokenType.BYTE
-
-            yield text, score, toktype
-
-    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        for text in self.added_tokens_list:
-            score = -1000.0
-            yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
-
-    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        yield from self.sentencepiece_tokens()
-        yield from self.added_tokens()
-
-    def __repr__(self) -> str:
-        return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
-
-
-class LlamaHfVocab(Vocab):
-    tokenizer_model = "llama"
-    name = "hfft"
-
-    def __init__(self, base_path: Path):
-        fname_tokenizer = base_path / FAST_TOKENIZER_FILE
-        # if this fails, FileNotFoundError propagates to caller
-        with open(fname_tokenizer, encoding='utf-8') as f:
-            tokenizer_json = json.load(f)
-
-        # pre-check so we know if we need transformers
-        tokenizer_model: dict[str, Any] = tokenizer_json['model']
-        is_llama3 = (
-            tokenizer_model['type'] == 'BPE' and tokenizer_model.get('ignore_merges', False)
-            and not tokenizer_model.get('byte_fallback', True)
-        )
-        if is_llama3:
-            raise TypeError('Llama 3 must be converted with BpeVocab')
-
-        if not is_llama3 and (
-            tokenizer_model['type'] != 'BPE' or not tokenizer_model.get('byte_fallback', False)
-            or tokenizer_json['decoder']['type'] != 'Sequence'
-        ):
-            raise FileNotFoundError('Cannot find Llama BPE tokenizer')
-
-        try:
-            from transformers import AutoTokenizer
-        except ImportError as e:
-            raise ImportError(
-                "To use LlamaHfVocab, please install the `transformers` package. "
-                "You can install it with `pip install transformers`."
-            ) from e
-
-        # Allow the tokenizer to default to slow or fast versions.
-        # Explicitly set tokenizer to use local paths.
-        self.tokenizer = AutoTokenizer.from_pretrained(
-            base_path,
-            cache_dir=base_path,
-            local_files_only=True,
-        )
-        assert self.tokenizer.is_fast  # assume tokenizer.json is used
-
-        # Initialize lists and dictionaries for added tokens
-        self.added_tokens_list = []
-        self.added_tokens_dict = dict()
-        self.added_tokens_ids  = set()
-
-        # Process added tokens
-        for tok, tokidx in sorted(
-            self.tokenizer.get_added_vocab().items(), key=lambda x: x[1]
-        ):
-            # Only consider added tokens that are not in the base vocabulary
-            if tokidx >= self.tokenizer.vocab_size:
-                self.added_tokens_list.append(tok)
-                self.added_tokens_dict[tok] = tokidx
-                self.added_tokens_ids.add(tokidx)
-
-        # Store special tokens and their IDs
-        self.specials = {
-            tok: self.tokenizer.get_vocab()[tok]
-            for tok in self.tokenizer.all_special_tokens
-        }
-        self.special_ids = set(self.tokenizer.all_special_ids)
-
-        # Set vocabulary sizes
-        self.vocab_size_base = self.tokenizer.vocab_size
-        self.vocab_size      = self.vocab_size_base + len(self.added_tokens_list)
-
-        self.fname_tokenizer = fname_tokenizer
-
-    def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        reverse_vocab = {
-            id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items()
-        }
-
-        for token_id in range(self.vocab_size_base):
-            # Skip processing added tokens here
-            if token_id in self.added_tokens_ids:
-                continue
-
-            # Convert token text to bytes
-            token_text = reverse_vocab[token_id].encode("utf-8")
-
-            # Yield token text, score, and type
-            yield token_text, self.get_token_score(token_id), self.get_token_type(
-                token_id, token_text, self.special_ids  # Reuse already stored special IDs
-            )
-
-    def get_token_type(self, token_id: int, token_text: bytes, special_ids: set[int]) -> gguf.TokenType:
-        # Special case for byte tokens
-        if re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):
-            return gguf.TokenType.BYTE
-
-        # Determine token type based on whether it's a special token
-        return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL
-
-    def get_token_score(self, token_id: int) -> float:
-        # Placeholder for actual logic to determine the token's score
-        # This needs to be implemented based on specific requirements
-        return -1000.0  # Default score
-
-    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        for text in self.added_tokens_list:
-            if text in self.specials:
-                toktype = self.get_token_type(self.specials[text], b'', self.special_ids)
-                score = self.get_token_score(self.specials[text])
-            else:
-                toktype = gguf.TokenType.USER_DEFINED
-                score = -1000.0
-
-            yield text.encode("utf-8"), score, toktype
-
-    def has_newline_token(self):
-        return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab
-
-    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        yield from self.hf_tokens()
-        yield from self.added_tokens()
-
-    def __repr__(self) -> str:
-        return f"<LlamaHfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
-
-
 #
 # data loading
 # TODO: reuse (probably move to gguf.py?)

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

@@ -774,7 +774,7 @@ static struct train_params get_default_train_params() {
 
     params.samples_start_after_nl = false;
     params.use_adam               = true;
-    params.use_flash              = true;
+    params.use_flash              = false;
     params.use_scratch            = true;
 
     // only adam

examples/embedding/embedding.cpp

@@ -80,7 +80,7 @@ int main(int argc, char ** argv) {
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
     }
 
     llama_backend_init();
@@ -107,7 +107,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     // split the prompt into lines

examples/eval-callback/eval-callback.cpp

@@ -152,7 +152,7 @@ int main(int argc, char ** argv) {
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
     }
 
     llama_backend_init();
@@ -176,7 +176,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     bool OK = run(ctx, params);

examples/finetune/finetune.cpp

@@ -563,8 +563,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         // not capturing these, to silcence warnings
         const int rope_mode = 0;
 
-        return ggml_rope_custom(ctx,
-            t, KQ_pos, n_rot, rope_mode, n_ctx, 0,
+        return ggml_rope_ext(ctx,
+            t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0,
             rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
         );
     };
@@ -643,7 +643,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct ggml_tensor * t15 = ggml_permute      (ctx, t12, 0, 3, 1, 2);                         set_name(t15, "t15");     assert_shape_4d(t15, N, n_embd_head, n_head_kv, n_batch);
         struct ggml_tensor * t16;
         if (enable_flash_attn) {
-            t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                         set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
+            GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported");
+            //t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                         set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
         } else {
             struct ggml_tensor * t16_0 = ggml_mul_mat              (ctx, t14, t13);                  set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
             struct ggml_tensor * t16_1 = ggml_scale_inplace        (ctx, t16_0, kv_scale);           set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);

examples/imatrix/imatrix.cpp

@@ -598,7 +598,7 @@ int main(int argc, char ** argv) {
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
     }
 
     sparams.dataset = params.prompt_file;
@@ -667,7 +667,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     bool OK = compute_imatrix(ctx, params, compute_ppl, from_chunk);

examples/infill/infill.cpp

@@ -50,9 +50,9 @@ static void write_logfile(
         return;
     }
 
-    const std::string timestamp = get_sortable_timestamp();
+    const std::string timestamp = string_get_sortable_timestamp();
 
-    const bool success = create_directory_with_parents(params.logdir);
+    const bool success = fs_create_directory_with_parents(params.logdir);
     if (!success) {
         fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
                 __func__, params.logdir.c_str());
@@ -70,7 +70,7 @@ static void write_logfile(
     fprintf(logfile, "binary: infill\n");
     char model_desc[128];
     llama_model_desc(model, model_desc, sizeof(model_desc));
-    dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
 
     fprintf(logfile, "\n");
     fprintf(logfile, "######################\n");
@@ -78,8 +78,8 @@ static void write_logfile(
     fprintf(logfile, "######################\n");
     fprintf(logfile, "\n");
 
-    dump_string_yaml_multiline(logfile, "output", output.c_str());
-    dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
+    yaml_dump_string_multiline(logfile, "output", output.c_str());
+    yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
 
     llama_dump_timing_info_yaml(logfile, ctx);
     fclose(logfile);
@@ -236,7 +236,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         LOG_TEE("\n");
-        LOG_TEE("%s\n", get_system_info(params).c_str());
+        LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
     }
     const bool add_bos = llama_should_add_bos_token(model);
     GGML_ASSERT(llama_add_eos_token(model) != 1);
@@ -621,8 +621,8 @@ int main(int argc, char ** argv) {
 
                 if (params.escape) {
                     //process escape sequences, for the initial prompt this is done in common.cpp when we load the params, but for the interactive mode we need to do it here
-                    process_escapes(params.input_prefix);
-                    process_escapes(params.input_suffix);
+                    string_process_escapes(params.input_prefix);
+                    string_process_escapes(params.input_suffix);
                 }
                 suff_rm_leading_spc = params.escape;
                 if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {

examples/llama-bench/llama-bench.cpp

@@ -178,6 +178,7 @@ struct cmd_params {
     std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
     std::vector<int> n_gpu_layers;
+    std::vector<std::string> rpc_servers;
     std::vector<llama_split_mode> split_mode;
     std::vector<int> main_gpu;
     std::vector<bool> no_kv_offload;
@@ -195,13 +196,14 @@ static const cmd_params cmd_params_defaults = {
     /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
-    /* n_pg          */ {{512, 128}},
+    /* n_pg          */ {},
     /* n_batch       */ {2048},
     /* n_ubatch      */ {512},
     /* type_k        */ {GGML_TYPE_F16},
     /* type_v        */ {GGML_TYPE_F16},
-    /* n_threads     */ {get_math_cpu_count()},
+    /* n_threads     */ {cpu_get_num_math()},
     /* n_gpu_layers  */ {99},
+    /* rpc_servers   */ {""},
     /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
     /* main_gpu      */ {0},
     /* no_kv_offload */ {false},
@@ -230,6 +232,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -ctv, --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("  -rpc, --rpc <rpc_servers>           (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").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());
@@ -384,6 +387,12 @@ 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 == "-rpc" || arg == "--rpc") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rpc_servers.push_back(argv[i]);
         } else if (arg == "-sm" || arg == "--split-mode") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -519,6 +528,7 @@ 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.rpc_servers.empty())  { params.rpc_servers = cmd_params_defaults.rpc_servers; }
     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; }
@@ -541,6 +551,7 @@ struct cmd_params_instance {
     ggml_type type_v;
     int n_threads;
     int n_gpu_layers;
+    std::string rpc_servers;
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
@@ -553,6 +564,9 @@ struct cmd_params_instance {
         llama_model_params mparams = llama_model_default_params();
 
         mparams.n_gpu_layers = n_gpu_layers;
+        if (!rpc_servers.empty()) {
+            mparams.rpc_servers = rpc_servers.c_str();
+        }
         mparams.split_mode = split_mode;
         mparams.main_gpu = main_gpu;
         mparams.tensor_split = tensor_split.data();
@@ -564,6 +578,7 @@ struct cmd_params_instance {
     bool equal_mparams(const cmd_params_instance & other) const {
         return model == other.model &&
                n_gpu_layers == other.n_gpu_layers &&
+               rpc_servers == other.rpc_servers &&
                split_mode == other.split_mode &&
                main_gpu == other.main_gpu &&
                use_mmap == other.use_mmap &&
@@ -592,6 +607,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     // 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 & rpc : params.rpc_servers)
     for (const auto & sm : params.split_mode)
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
@@ -618,6 +634,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
+                /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
                 /* .no_kv_offload= */ nkvo,
@@ -643,6 +660,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
+                /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
                 /* .no_kv_offload= */ nkvo,
@@ -668,6 +686,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
+                /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
                 /* .no_kv_offload= */ nkvo,
@@ -692,6 +711,7 @@ struct test {
     static const bool kompute;
     static const bool metal;
     static const bool sycl;
+    static const bool rpc;
     static const bool gpu_blas;
     static const bool blas;
     static const std::string cpu_info;
@@ -790,6 +810,9 @@ struct test {
         if (sycl) {
             return GGML_SYCL_NAME;
         }
+        if (rpc) {
+            return "RPC";
+        }
         if (gpu_blas) {
             return "GPU BLAS";
         }
@@ -803,7 +826,7 @@ struct test {
     static const std::vector<std::string> & get_fields() {
         static const std::vector<std::string> fields = {
             "build_commit", "build_number",
-            "cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "gpu_blas", "blas",
+            "cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "rpc", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_ubatch",
@@ -859,7 +882,7 @@ struct test {
         std::vector<std::string> values = {
             build_commit, std::to_string(build_number),
             std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(vulkan),
-            std::to_string(metal), std::to_string(sycl), std::to_string(gpu_blas), std::to_string(blas),
+            std::to_string(metal), std::to_string(sycl), std::to_string(rpc), std::to_string(gpu_blas), std::to_string(blas),
             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_ubatch),
@@ -894,6 +917,7 @@ const bool        test::metal        = !!ggml_cpu_has_metal();
 const bool        test::gpu_blas     = !!ggml_cpu_has_gpublas();
 const bool        test::blas         = !!ggml_cpu_has_blas();
 const bool        test::sycl         = !!ggml_cpu_has_sycl();
+const bool        test::rpc          = !!ggml_cpu_has_rpc();
 const std::string test::cpu_info     = get_cpu_info();
 const std::string test::gpu_info     = get_gpu_info();
 

examples/llama.android/app/build.gradle.kts

@@ -7,8 +7,6 @@ android {
     namespace = "com.example.llama"
     compileSdk = 34
 
-    ndkVersion = "26.1.10909125"
-
     defaultConfig {
         applicationId = "com.example.llama"
         minSdk = 33
@@ -20,17 +18,6 @@ android {
         vectorDrawables {
             useSupportLibrary = true
         }
-        ndk {
-            // Add NDK properties if wanted, e.g.
-            // abiFilters += listOf("arm64-v8a")
-        }
-        externalNativeBuild {
-            cmake {
-                arguments += "-DCMAKE_BUILD_TYPE=Release"
-                cppFlags += listOf()
-                arguments += listOf()
-            }
-        }
     }
 
     buildTypes {
@@ -55,17 +42,6 @@ android {
     composeOptions {
         kotlinCompilerExtensionVersion = "1.5.1"
     }
-    packaging {
-        resources {
-            excludes += "/META-INF/{AL2.0,LGPL2.1}"
-        }
-    }
-    externalNativeBuild {
-        cmake {
-            path = file("src/main/cpp/CMakeLists.txt")
-            version = "3.22.1"
-        }
-    }
 }
 
 dependencies {
@@ -78,6 +54,7 @@ dependencies {
     implementation("androidx.compose.ui:ui-graphics")
     implementation("androidx.compose.ui:ui-tooling-preview")
     implementation("androidx.compose.material3:material3")
+    implementation(project(":llama"))
     testImplementation("junit:junit:4.13.2")
     androidTestImplementation("androidx.test.ext:junit:1.1.5")
     androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")

examples/llama.android/app/src/main/java/com/example/llama/MainViewModel.kt

@@ -1,5 +1,6 @@
 package com.example.llama
 
+import android.llama.cpp.LLamaAndroid
 import android.util.Log
 import androidx.compose.runtime.getValue
 import androidx.compose.runtime.mutableStateOf
@@ -9,7 +10,7 @@ import androidx.lifecycle.viewModelScope
 import kotlinx.coroutines.flow.catch
 import kotlinx.coroutines.launch
 
-class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
+class MainViewModel(private val llamaAndroid: LLamaAndroid = LLamaAndroid.instance()): ViewModel() {
     companion object {
         @JvmStatic
         private val NanosPerSecond = 1_000_000_000.0
@@ -28,7 +29,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
 
         viewModelScope.launch {
             try {
-                llm.unload()
+                llamaAndroid.unload()
             } catch (exc: IllegalStateException) {
                 messages += exc.message!!
             }
@@ -44,7 +45,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
         messages += ""
 
         viewModelScope.launch {
-            llm.send(text)
+            llamaAndroid.send(text)
                 .catch {
                     Log.e(tag, "send() failed", it)
                     messages += it.message!!
@@ -57,7 +58,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
         viewModelScope.launch {
             try {
                 val start = System.nanoTime()
-                val warmupResult = llm.bench(pp, tg, pl, nr)
+                val warmupResult = llamaAndroid.bench(pp, tg, pl, nr)
                 val end = System.nanoTime()
 
                 messages += warmupResult
@@ -70,7 +71,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
                     return@launch
                 }
 
-                messages += llm.bench(512, 128, 1, 3)
+                messages += llamaAndroid.bench(512, 128, 1, 3)
             } catch (exc: IllegalStateException) {
                 Log.e(tag, "bench() failed", exc)
                 messages += exc.message!!
@@ -81,7 +82,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
     fun load(pathToModel: String) {
         viewModelScope.launch {
             try {
-                llm.load(pathToModel)
+                llamaAndroid.load(pathToModel)
                 messages += "Loaded $pathToModel"
             } catch (exc: IllegalStateException) {
                 Log.e(tag, "load() failed", exc)

examples/llama.android/build.gradle.kts

@@ -2,4 +2,5 @@
 plugins {
     id("com.android.application") version "8.2.0" apply false
     id("org.jetbrains.kotlin.android") version "1.9.0" apply false
+    id("com.android.library") version "8.2.0" apply false
 }

examples/llama.android/llama/.gitignore

@@ -0,0 +1 @@
+/build

examples/llama.android/llama/CMakeLists.txt

@@ -12,17 +12,20 @@ cmake_minimum_required(VERSION 3.22.1)
 # build script scope).
 project("llama-android")
 
+## Fetch latest llama.cpp from GitHub
 #include(FetchContent)
 #FetchContent_Declare(
 #        llama
 #        GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
-#        GIT_TAG        ci-android
+#        GIT_TAG        master
 #)
 #
 ## Also provides "common"
 #FetchContent_MakeAvailable(llama)
 
-add_subdirectory(../../../../../../ please-work)
+# llama.cpp CI uses the code from the current branch
+# ref: https://github.com/ggerganov/llama.cpp/pull/7341#issuecomment-2117617700
+add_subdirectory(../../../../../../ build-llama)
 
 # Creates and names a library, sets it as either STATIC
 # or SHARED, and provides the relative paths to its source code.
@@ -39,7 +42,7 @@ add_subdirectory(../../../../../../ please-work)
 # used in the AndroidManifest.xml file.
 add_library(${CMAKE_PROJECT_NAME} SHARED
     # List C/C++ source files with relative paths to this CMakeLists.txt.
-    llama-android.cpp)
+        llama-android.cpp)
 
 # Specifies libraries CMake should link to your target library. You
 # can link libraries from various origins, such as libraries defined in this

examples/llama.android/llama/build.gradle.kts

@@ -0,0 +1,68 @@
+plugins {
+    id("com.android.library")
+    id("org.jetbrains.kotlin.android")
+}
+
+android {
+    namespace = "android.llama.cpp"
+    compileSdk = 34
+
+    defaultConfig {
+        minSdk = 33
+
+        testInstrumentationRunner = "androidx.test.runner.AndroidJUnitRunner"
+        consumerProguardFiles("consumer-rules.pro")
+        ndk {
+            // Add NDK properties if wanted, e.g.
+            // abiFilters += listOf("arm64-v8a")
+        }
+        externalNativeBuild {
+            cmake {
+                arguments += "-DCMAKE_BUILD_TYPE=Release"
+                cppFlags += listOf()
+                arguments += listOf()
+
+                cppFlags("")
+            }
+        }
+    }
+
+    buildTypes {
+        release {
+            isMinifyEnabled = false
+            proguardFiles(
+                getDefaultProguardFile("proguard-android-optimize.txt"),
+                "proguard-rules.pro"
+            )
+        }
+    }
+    externalNativeBuild {
+        cmake {
+            path("src/main/cpp/CMakeLists.txt")
+            version = "3.22.1"
+        }
+    }
+    compileOptions {
+        sourceCompatibility = JavaVersion.VERSION_1_8
+        targetCompatibility = JavaVersion.VERSION_1_8
+    }
+    kotlinOptions {
+        jvmTarget = "1.8"
+    }
+
+    packaging {
+        resources {
+            excludes += "/META-INF/{AL2.0,LGPL2.1}"
+        }
+    }
+}
+
+dependencies {
+
+    implementation("androidx.core:core-ktx:1.12.0")
+    implementation("androidx.appcompat:appcompat:1.6.1")
+    implementation("com.google.android.material:material:1.11.0")
+    testImplementation("junit:junit:4.13.2")
+    androidTestImplementation("androidx.test.ext:junit:1.1.5")
+    androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")
+}

examples/llama.android/llama/proguard-rules.pro

@@ -0,0 +1,21 @@
+# Add project specific ProGuard rules here.
+# You can control the set of applied configuration files using the
+# proguardFiles setting in build.gradle.
+#
+# For more details, see
+#   http://developer.android.com/guide/developing/tools/proguard.html
+
+# If your project uses WebView with JS, uncomment the following
+# and specify the fully qualified class name to the JavaScript interface
+# class:
+#-keepclassmembers class fqcn.of.javascript.interface.for.webview {
+#   public *;
+#}
+
+# Uncomment this to preserve the line number information for
+# debugging stack traces.
+#-keepattributes SourceFile,LineNumberTable
+
+# If you keep the line number information, uncomment this to
+# hide the original source file name.
+#-renamesourcefileattribute SourceFile

examples/llama.android/llama/src/androidTest/java/android/llama/cpp/ExampleInstrumentedTest.kt

@@ -0,0 +1,24 @@
+package android.llama.cpp
+
+import androidx.test.platform.app.InstrumentationRegistry
+import androidx.test.ext.junit.runners.AndroidJUnit4
+
+import org.junit.Test
+import org.junit.runner.RunWith
+
+import org.junit.Assert.*
+
+/**
+ * Instrumented test, which will execute on an Android device.
+ *
+ * See [testing documentation](http://d.android.com/tools/testing).
+ */
+@RunWith(AndroidJUnit4::class)
+class ExampleInstrumentedTest {
+    @Test
+    fun useAppContext() {
+        // Context of the app under test.
+        val appContext = InstrumentationRegistry.getInstrumentation().targetContext
+        assertEquals("android.llama.cpp.test", appContext.packageName)
+    }
+}

examples/llama.android/llama/src/main/AndroidManifest.xml

@@ -0,0 +1,4 @@
+<?xml version="1.0" encoding="utf-8"?>
+<manifest xmlns:android="http://schemas.android.com/apk/res/android">
+
+</manifest>

examples/llama.android/llama/src/main/cpp/CMakeLists.txt

@@ -0,0 +1,49 @@
+# For more information about using CMake with Android Studio, read the
+# documentation: https://d.android.com/studio/projects/add-native-code.html.
+# For more examples on how to use CMake, see https://github.com/android/ndk-samples.
+
+# Sets the minimum CMake version required for this project.
+cmake_minimum_required(VERSION 3.22.1)
+
+# Declares the project name. The project name can be accessed via ${ PROJECT_NAME},
+# Since this is the top level CMakeLists.txt, the project name is also accessible
+# with ${CMAKE_PROJECT_NAME} (both CMake variables are in-sync within the top level
+# build script scope).
+project("llama-android")
+
+include(FetchContent)
+FetchContent_Declare(
+        llama
+        GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
+        GIT_TAG        master
+)
+
+# Also provides "common"
+FetchContent_MakeAvailable(llama)
+
+# Creates and names a library, sets it as either STATIC
+# or SHARED, and provides the relative paths to its source code.
+# You can define multiple libraries, and CMake builds them for you.
+# Gradle automatically packages shared libraries with your APK.
+#
+# In this top level CMakeLists.txt, ${CMAKE_PROJECT_NAME} is used to define
+# the target library name; in the sub-module's CMakeLists.txt, ${PROJECT_NAME}
+# is preferred for the same purpose.
+#
+# In order to load a library into your app from Java/Kotlin, you must call
+# System.loadLibrary() and pass the name of the library defined here;
+# for GameActivity/NativeActivity derived applications, the same library name must be
+# used in the AndroidManifest.xml file.
+add_library(${CMAKE_PROJECT_NAME} SHARED
+        # List C/C++ source files with relative paths to this CMakeLists.txt.
+        llama-android.cpp)
+
+# Specifies libraries CMake should link to your target library. You
+# can link libraries from various origins, such as libraries defined in this
+# build script, prebuilt third-party libraries, or Android system libraries.
+target_link_libraries(${CMAKE_PROJECT_NAME}
+        # List libraries link to the target library
+        llama
+        common
+        android
+        log)

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -81,7 +81,7 @@ static void log_callback(ggml_log_level level, const char * fmt, void * data) {
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
+Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring filename) {
     llama_model_params model_params = llama_model_default_params();
 
     auto path_to_model = env->GetStringUTFChars(filename, 0);
@@ -101,13 +101,13 @@ Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1model(JNIEnv *, jobject, jlong model) {
+Java_android_llama_cpp_LLamaAndroid_free_1model(JNIEnv *, jobject, jlong model) {
     llama_free_model(reinterpret_cast<llama_model *>(model));
 }
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
+Java_android_llama_cpp_LLamaAndroid_new_1context(JNIEnv *env, jobject, jlong jmodel) {
     auto model = reinterpret_cast<llama_model *>(jmodel);
 
     if (!model) {
@@ -139,25 +139,25 @@ Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1context(JNIEnv *, jobject, jlong context) {
+Java_android_llama_cpp_LLamaAndroid_free_1context(JNIEnv *, jobject, jlong context) {
     llama_free(reinterpret_cast<llama_context *>(context));
 }
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1free(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_backend_1free(JNIEnv *, jobject) {
     llama_backend_free();
 }
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_log_1to_1android(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_log_1to_1android(JNIEnv *, jobject) {
     llama_log_set(log_callback, NULL);
 }
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_bench_1model(
+Java_android_llama_cpp_LLamaAndroid_bench_1model(
         JNIEnv *env,
         jobject,
         jlong context_pointer,
@@ -271,13 +271,13 @@ Java_com_example_llama_Llm_bench_1model(
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
+Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
     llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
 }
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
+Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
 
     // Source: Copy of llama.cpp:llama_batch_init but heap-allocated.
 
@@ -313,19 +313,19 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_backend_1init(JNIEnv *, jobject) {
     llama_backend_init();
 }
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_system_1info(JNIEnv *env, jobject) {
+Java_android_llama_cpp_LLamaAndroid_system_1info(JNIEnv *env, jobject) {
     return env->NewStringUTF(llama_print_system_info());
 }
 
 extern "C"
 JNIEXPORT jint JNICALL
-Java_com_example_llama_Llm_completion_1init(
+Java_android_llama_cpp_LLamaAndroid_completion_1init(
         JNIEnv *env,
         jobject,
         jlong context_pointer,
@@ -376,7 +376,7 @@ Java_com_example_llama_Llm_completion_1init(
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_completion_1loop(
+Java_android_llama_cpp_LLamaAndroid_completion_1loop(
         JNIEnv * env,
         jobject,
         jlong context_pointer,
@@ -438,6 +438,6 @@ Java_com_example_llama_Llm_completion_1loop(
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
+Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
     llama_kv_cache_clear(reinterpret_cast<llama_context *>(context));
 }

examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt

@@ -1,4 +1,4 @@
-package com.example.llama
+package android.llama.cpp
 
 import android.util.Log
 import kotlinx.coroutines.CoroutineDispatcher
@@ -10,7 +10,7 @@ import kotlinx.coroutines.withContext
 import java.util.concurrent.Executors
 import kotlin.concurrent.thread
 
-class Llm {
+class LLamaAndroid {
     private val tag: String? = this::class.simpleName
 
     private val threadLocalState: ThreadLocal<State> = ThreadLocal.withInitial { State.Idle }
@@ -165,8 +165,8 @@ class Llm {
         }
 
         // Enforce only one instance of Llm.
-        private val _instance: Llm = Llm()
+        private val _instance: LLamaAndroid = LLamaAndroid()
 
-        fun instance(): Llm = _instance
+        fun instance(): LLamaAndroid = _instance
     }
 }

examples/llama.android/llama/src/test/java/android/llama/cpp/ExampleUnitTest.kt

@@ -0,0 +1,17 @@
+package android.llama.cpp
+
+import org.junit.Test
+
+import org.junit.Assert.*
+
+/**
+ * Example local unit test, which will execute on the development machine (host).
+ *
+ * See [testing documentation](http://d.android.com/tools/testing).
+ */
+class ExampleUnitTest {
+    @Test
+    fun addition_isCorrect() {
+        assertEquals(4, 2 + 2)
+    }
+}

examples/llama.android/settings.gradle.kts

@@ -15,3 +15,4 @@ dependencyResolutionManagement {
 
 rootProject.name = "LlamaAndroid"
 include(":app")
+include(":llama")

examples/llava/MobileVLM-README.md

@@ -54,10 +54,10 @@ python ./examples/llava/convert-image-encoder-to-gguf \
     --projector-type ldpv2
 ```
 
-4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
+4. Use `examples/convert-legacy-llama.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./convert.py path/to/MobileVLM-1.7B
+python ./examples/convert-legacy-llama.py path/to/MobileVLM-1.7B
 ```
 
 5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k`

examples/llava/README.md

@@ -50,10 +50,10 @@ python ./examples/llava/llava-surgery.py -m ../llava-v1.5-7b
 python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
 ```
 
-5. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
+5. Use `examples/convert-legacy-llama.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./convert.py ../llava-v1.5-7b --skip-unknown
+python ./examples/convert-legacy-llama.py ../llava-v1.5-7b --skip-unknown
 ```
 
 Now both the LLaMA part and the image encoder are in the `llava-v1.5-7b` directory.
@@ -92,7 +92,7 @@ python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projecto
 
 6) Then convert the model to gguf format:
 ```console
-python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
+python ./examples/convert-legacy-llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown
 ```
 
 7) And finally we can run the llava-cli using the 1.6 model version:

examples/llava/clip.h

@@ -68,7 +68,7 @@ CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8
 /** 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);
 
-/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
+/** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
 CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
 
 CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);

examples/llava/llava-cli.cpp

@@ -290,7 +290,7 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
-        gpt_print_usage(argc, argv, params);
+        gpt_params_print_usage(argc, argv, params);
         show_additional_info(argc, argv);
         return 1;
     }

examples/llava/requirements.txt

@@ -1,3 +1,3 @@
--r ../../requirements/requirements-convert.txt
+-r ../../requirements/requirements-convert-legacy-llama.txt
 pillow~=10.2.0
 torch~=2.1.1

examples/lookahead/lookahead.cpp

@@ -174,7 +174,7 @@ int main(int argc, char ** argv) {
         // debug
         if (dump_kv_cache) {
             llama_kv_cache_view_update(ctx, &kvc_view);
-            dump_kv_cache_view_seqs(kvc_view, 40);
+            llama_kv_cache_dump_view_seqs(kvc_view, 40);
         }
 
         // build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/

examples/lookup/lookup.cpp

@@ -121,7 +121,7 @@ int main(int argc, char ** argv){
         // debug
         if (dump_kv_cache) {
             llama_kv_cache_view_update(ctx, &kvc_view);
-            dump_kv_cache_view_seqs(kvc_view, 40);
+            llama_kv_cache_dump_view_seqs(kvc_view, 40);
         }
 
         // print current draft sequence

examples/main/README.md

@@ -325,3 +325,5 @@ These options provide extra functionality and customization when running the LLa
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
+
+-   `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable  or in an OS-specific local cache.

examples/main/main.cpp

@@ -60,9 +60,9 @@ static void write_logfile(
         return;
     }
 
-    const std::string timestamp = get_sortable_timestamp();
+    const std::string timestamp = string_get_sortable_timestamp();
 
-    const bool success = create_directory_with_parents(params.logdir);
+    const bool success = fs_create_directory_with_parents(params.logdir);
     if (!success) {
         fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
                 __func__, params.logdir.c_str());
@@ -80,7 +80,7 @@ static void write_logfile(
     fprintf(logfile, "binary: main\n");
     char model_desc[128];
     llama_model_desc(model, model_desc, sizeof(model_desc));
-    dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
 
     fprintf(logfile, "\n");
     fprintf(logfile, "######################\n");
@@ -88,8 +88,8 @@ static void write_logfile(
     fprintf(logfile, "######################\n");
     fprintf(logfile, "\n");
 
-    dump_string_yaml_multiline(logfile, "output", output.c_str());
-    dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
+    yaml_dump_string_multiline(logfile, "output", output.c_str());
+    yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
 
     llama_dump_timing_info_yaml(logfile, ctx);
     fclose(logfile);
@@ -181,7 +181,7 @@ int main(int argc, char ** argv) {
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
     }
 
     LOG("%s: llama backend init\n", __func__);
@@ -219,7 +219,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         LOG_TEE("\n");
-        LOG_TEE("%s\n", get_system_info(params).c_str());
+        LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     std::string path_session = params.path_prompt_cache;
@@ -474,12 +474,12 @@ int main(int argc, char ** argv) {
     LOG_TEE("\n\n");
 
     if (params.interactive) {
-        const char *control_message;
+        const char * control_message;
         if (params.multiline_input) {
-            control_message = " - To return control to LLaMa, end your input with '\\'.\n"
+            control_message = " - To return control to the AI, end your input with '\\'.\n"
                               " - To return control without starting a new line, end your input with '/'.\n";
         } else {
-            control_message = " - Press Return to return control to LLaMa.\n"
+            control_message = " - Press Return to return control to the AI.\n"
                               " - To return control without starting a new line, end your input with '/'.\n"
                               " - If you want to submit another line, end your input with '\\'.\n";
         }
@@ -707,7 +707,7 @@ int main(int argc, char ** argv) {
 
             const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            llama_sampling_accept(ctx_sampling, ctx, id, /* apply_grammar= */ true);
 
             LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
 
@@ -728,7 +728,7 @@ int main(int argc, char ** argv) {
 
                 // push the prompt in the sampling context in order to apply repetition penalties later
                 // for the prompt, we don't apply grammar rules
-                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
+                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], /* apply_grammar= */ false);
 
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
@@ -740,18 +740,26 @@ int main(int argc, char ** argv) {
         // display text
         if (input_echo && display) {
             for (auto id : embd) {
-                const std::string token_str = llama_token_to_piece(ctx, id, !params.conversation);
-                printf("%s", token_str.c_str());
+                const std::string token_str = llama_token_to_piece(ctx, id, params.special);
 
+                // Console/Stream Output
+                fprintf(stdout, "%s", token_str.c_str());
+
+                // Record Displayed Tokens To Log
+                // Note: Generated tokens are created one by one hence this check
                 if (embd.size() > 1) {
+                    // Incoming Requested Tokens
                     input_tokens.push_back(id);
                 } else {
+                    // Outgoing Generated Tokens
                     output_tokens.push_back(id);
                     output_ss << token_str;
                 }
+
+                fflush(stdout);
             }
-            fflush(stdout);
         }
+
         // reset color to default if there is no pending user input
         if (input_echo && (int) embd_inp.size() == n_consumed) {
             console::set_display(console::reset);
@@ -879,7 +887,7 @@ int main(int argc, char ** argv) {
                         embd_inp.insert(embd_inp.end(), cml_pfx.begin(), cml_pfx.end());
                     }
                     if (params.escape) {
-                        process_escapes(buffer);
+                        string_process_escapes(buffer);
                     }
 
                     const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);

examples/make-ggml.py

@@ -1,98 +0,0 @@
-#!/usr/bin/env python3
-"""
-This script converts Hugging Face Llama, StarCoder, Falcon, Baichuan, and GPT-NeoX models to GGUF and quantizes them.
-
-Usage:
-python make-ggml.py {model_dir_or_hf_repo_name} --model_type {model_type} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]
-
-Arguments:
-- model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
-- --model_type: (Required) The type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.
-- --outname: (Optional) The name of the output model. If not specified, the last part of the model directory path or the Hugging Face model repo name will be used.
-- --outdir: (Optional) The directory where the output model(s) will be stored. If not specified, '../models/{outname}' will be used.
-- --quants: (Optional) The types of quantization to apply. This should be a space-separated list. The default is 'Q4_K_M Q5_K_S'.
-- --keep_fp16: (Optional) If specified, the FP16 model will not be deleted after the quantized models are created.
-
-Old quant types (some base model types require these):
-- Q4_0: small, very high quality loss - legacy, prefer using Q3_K_M
-- Q4_1: small, substantial quality loss - legacy, prefer using Q3_K_L
-- Q5_0: medium, balanced quality - legacy, prefer using Q4_K_M
-- Q5_1: medium, low quality loss - legacy, prefer using Q5_K_M
-
-New quant types (recommended):
-- Q2_K: smallest, extreme quality loss - not recommended
-- Q3_K: alias for Q3_K_M
-- Q3_K_S: very small, very high quality loss
-- Q3_K_M: very small, very high quality loss
-- Q3_K_L: small, substantial quality loss
-- Q4_K: alias for Q4_K_M
-- Q4_K_S: small, significant quality loss
-- Q4_K_M: medium, balanced quality - recommended
-- Q5_K: alias for Q5_K_M
-- Q5_K_S: large, low quality loss - recommended
-- Q5_K_M: large, very low quality loss - recommended
-- Q6_K: very large, extremely low quality loss
-- Q8_0: very large, extremely low quality loss - not recommended
-- F16: extremely large, virtually no quality loss - not recommended
-- F32: absolutely huge, lossless - not recommended
-"""
-import subprocess
-subprocess.run(f"pip install huggingface-hub==0.16.4", shell=True, check=True)
-
-import argparse
-import os
-from huggingface_hub import snapshot_download
-
-def main(model, model_type, outname, outdir, quants, keep_fp16):
-    if not os.path.isdir(model):
-        print(f"Model not found at {model}. Downloading...")
-        try:
-            if outname is None:
-                outname = model.split('/')[-1]
-            model = snapshot_download(repo_id=model, cache_dir='../models/hf_cache')
-        except Exception as e:
-            raise Exception(f"Could not download the model: {e}")
-
-    if outdir is None:
-        outdir = f'../models/{outname}'
-
-    if not os.path.isfile(f"{model}/config.json"):
-        raise Exception(f"Could not find config.json in {model}")
-
-    os.makedirs(outdir, exist_ok=True)
-
-    print("Building llama.cpp")
-    subprocess.run(f"cd .. && make quantize", shell=True, check=True)
-
-    fp16 = f"{outdir}/{outname}.gguf.fp16.bin"
-
-    print(f"Making unquantised GGUF at {fp16}")
-    if not os.path.isfile(fp16):
-        if model_type != "llama":
-            subprocess.run(f"python3 ../convert-{model_type}-hf-to-gguf.py {model} 1 --outfile {fp16}", shell=True, check=True)
-        else:
-            subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
-    else:
-        print(f"Unquantised GGML already exists at: {fp16}")
-
-    print("Making quants")
-    for type in quants:
-        outfile = f"{outdir}/{outname}.gguf.{type}.bin"
-        print(f"Making {type} : {outfile}")
-        subprocess.run(f"../quantize {fp16} {outfile} {type}", shell=True, check=True)
-
-    if not keep_fp16:
-        os.remove(fp16)
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description='Convert/Quantize HF models to GGUF. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
-    parser.add_argument('model', help='Downloaded model dir or Hugging Face model repo name')
-    parser.add_argument('--model_type', required=True, choices=['llama', 'starcoder', 'falcon', 'baichuan', 'gptneox'], help='Type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.')
-    parser.add_argument('--outname', default=None, help='Output model(s) name')
-    parser.add_argument('--outdir', default=None, help='Output directory')
-    parser.add_argument('--quants', nargs='*', default=["Q4_K_M", "Q5_K_S"], help='Quant types')
-    parser.add_argument('--keep_fp16', action='store_true', help='Keep fp16 model', default=False)
-
-    args = parser.parse_args()
-
-    main(args.model, args.model_type, args.outname, args.outdir, args.quants, args.keep_fp16)

examples/parallel/parallel.cpp

@@ -210,7 +210,7 @@ int main(int argc, char ** argv) {
     while (true) {
         if (dump_kv_cache) {
             llama_kv_cache_view_update(ctx, &kvc_view);
-            dump_kv_cache_view_seqs(kvc_view, 40);
+            llama_kv_cache_dump_view_seqs(kvc_view, 40);
         }
 
         llama_batch_clear(batch);

examples/perplexity/README.md

@@ -42,10 +42,13 @@ In addition to the KL divergence the following statistics are calculated with `-
 
 Results were generated using the CUDA backend and are sorted by Kullback-Leibler divergence relative to FP16.
 The "WT" importance matrices were created using varying numbers of Wikitext tokens and can be found [here](https://huggingface.co/JohannesGaessler/llama.cpp_importance_matrices/blob/main/imatrix-llama_3-8b-f16-2.7m_tokens.dat).
+Note: the FP16 logits used for the calculation of all metrics other than perplexity are stored in a binary file between runs.
+In order to save space this file does **not** contain the exact same FP32 logits but instead casts them to 16 bit unsigned integers (with some scaling).
+So the "f16" results are to be understood as the difference resulting only from this downcast.
 
 | Quantization | imatrix | Model size [GiB] | PPL                    | ΔPPL                   | KLD                   | Mean Δp           | RMS Δp           |
 |--------------|---------|------------------|------------------------|------------------------|-----------------------|-------------------|------------------|
-| f16          | None    |            14.97 | 6.233160 ±   0.037828  | -                      | -                     | -                 | -                |
+| f16          | None    |            14.97 | 6.233160 ±   0.037828  | 0.001524 ±   0.000755  | 0.000551 ±   0.000002 |  0.001 ± 0.002 %  | 0.787 ± 0.004 %  |
 | q8_0         | None    |             7.96 | 6.234284 ±   0.037878  | 0.002650 ±   0.001006  | 0.001355 ±   0.000006 | -0.019 ± 0.003 %  | 1.198 ± 0.007 %  |
 | q6_K         | None    |             6.14 | 6.253382 ±   0.038078  | 0.021748 ±   0.001852  | 0.005452 ±   0.000035 | -0.007 ± 0.006 %  | 2.295 ± 0.019 %  |
 | q5_K_M       | None    |             5.33 | 6.288607 ±   0.038338  | 0.056974 ±   0.002598  | 0.010762 ±   0.000079 | -0.114 ± 0.008 %  | 3.160 ± 0.031 %  |

examples/perplexity/perplexity.cpp

@@ -44,9 +44,9 @@ static void write_logfile(
         return;
     }
 
-    const std::string timestamp = get_sortable_timestamp();
+    const std::string timestamp = string_get_sortable_timestamp();
 
-    const bool success = create_directory_with_parents(params.logdir);
+    const bool success = fs_create_directory_with_parents(params.logdir);
     if (!success) {
         fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
                 __func__, params.logdir.c_str());
@@ -64,7 +64,7 @@ static void write_logfile(
     fprintf(logfile, "binary: main\n");
     char model_desc[128];
     llama_model_desc(model, model_desc, sizeof(model_desc));
-    dump_non_result_info_yaml(logfile, params, ctx, timestamp, results.tokens, model_desc);
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, results.tokens, model_desc);
 
     fprintf(logfile, "\n");
     fprintf(logfile, "######################\n");
@@ -72,9 +72,9 @@ static void write_logfile(
     fprintf(logfile, "######################\n");
     fprintf(logfile, "\n");
 
-    dump_vector_float_yaml(logfile, "logits", results.logits);
+    yaml_dump_vector_float(logfile, "logits", results.logits);
     fprintf(logfile, "ppl_value: %f\n", results.ppl_value);
-    dump_vector_float_yaml(logfile, "probs", results.probs);
+    yaml_dump_vector_float(logfile, "probs", results.probs);
 
     llama_dump_timing_info_yaml(logfile, ctx);
     fclose(logfile);
@@ -1425,7 +1425,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         // Use all tasks
         tasks.resize(n_task);
         printf("%s: reading tasks", __func__);
-        int n_dot = n_task/100;
+        int n_dot = std::max((int) n_task/100, 1);
         int i = 0;
         for (auto& task : tasks) {
             ++i;
@@ -1675,7 +1675,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
 
     llama_batch_free(batch);
 
-    if (n_done < 100) return;
+    if (n_done < 100 && (params.multiple_choice_tasks != 0 && params.multiple_choice_tasks < (size_t)n_task)) return;
 
     float p = 1.f*n_correct/n_done;
     float sigma = sqrt(p*(1-p)/(n_done-1));
@@ -2007,7 +2007,7 @@ int main(int argc, char ** argv) {
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
     }
 
     llama_backend_init();
@@ -2035,7 +2035,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     struct results_perplexity results;

examples/quantize/quantize.cpp

@@ -259,7 +259,7 @@ int main(int argc, char ** argv) {
                 usage(argv[0]);
             }
         } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
-            if (arg_idx == argc-1 || !parse_kv_override(argv[++arg_idx], kv_overrides)) {
+            if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
                 usage(argv[0]);
             }
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
@@ -284,7 +284,7 @@ int main(int argc, char ** argv) {
             } else {
                 usage(argv[0]);
             }
-        } else if (strcmp(argv[arg_idx], "--keep-split")) {
+        } else if (strcmp(argv[arg_idx], "--keep-split") == 0) {
             params.keep_split = true;
         } else {
             usage(argv[0]);

examples/quantize/tests.sh

@@ -41,8 +41,8 @@ $SPLIT --split-max-tensors 28  $WORK_PATH/gemma-1.1-2b-it.Q8_0.gguf $WORK_PATH/g
 echo PASS
 echo
 
-# 3. Requant model with '--keep_split'
-$QUANTIZE --allow-requantize --keep_split $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant.gguf Q4_K
+# 3. Requant model with '--keep-split'
+$QUANTIZE --allow-requantize --keep-split $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant.gguf Q4_K
 echo PASS
 echo
 
@@ -51,7 +51,7 @@ $MAIN --model $WORK_PATH/ggml-model-requant-00001-of-00006.gguf --random-prompt
 echo PASS
 echo
 
-# 4. Requant mode without '--keep_split'
+# 4. Requant mode without '--keep-split'
 $QUANTIZE --allow-requantize $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant-merge.gguf Q4_K
 echo PASS
 echo

examples/retrieval/retrieval.cpp

@@ -11,7 +11,7 @@ struct retrieval_params {
 };
 
 static void retrieval_params_print_usage(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & params) {
-    gpt_print_usage(argc, argv, gpt_params);
+    gpt_params_print_usage(argc, argv, gpt_params);
     printf("retrieval options:\n");
     printf("  --context-file FNAME  file containing context to embed.\n");
     printf("                        specify multiple files by providing --context-file option multiple times.\n");
@@ -226,7 +226,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     // max batch size

examples/rpc/rpc-server.cpp

@@ -56,6 +56,10 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
         } else if (arg == "-h" || arg == "--help") {
             print_usage(argc, argv, params);
             exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            print_usage(argc, argv, params);
+            exit(0);
         }
     }
     return true;

examples/server/README.md

@@ -18,8 +18,8 @@ The project is under active development, and we are [looking for feedback and co
 **Command line options:**
 
 - `-v`, `--verbose`: Enable verbose server output. When using the `/completion` endpoint, this includes the tokenized prompt, the full request and the full response.
-- `-t N`, `--threads N`: Set the number of threads to use during generation. Not used if model layers are offloaded to GPU. The server is using batching. This parameter is used only if one token is to be processed on CPU backend.
-- `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation. Not used if model layers are offloaded to GPU.
+- `-t N`, `--threads N`: Set the number of threads to use by CPU layers during generation. Not used by model layers that are offloaded to GPU. This option has no effect when using the maximum number of GPU layers. Default: `std::thread::hardware_concurrency()` (number of CPU cores).
+- `-tb N, --threads-batch N`: Set the number of threads to use by CPU layers during batch and prompt processing (>= 32 tokens). This option has no effect if a GPU is available. Default: `--threads`.
 - `--threads-http N`: Number of threads in the http server pool to process requests. Default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`
 - `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
 - `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file. Default: unused
@@ -48,7 +48,7 @@ The project is under active development, and we are [looking for feedback and co
 - `--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.
 - `--embeddings`: Enable embedding vector output and the OAI compatible endpoint /v1/embeddings. Physical batch size (`--ubatch-size`) must be carefully defined. Default: disabled
-- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`
+- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`. Values > 1 will allow for higher throughput with multiple parallel requests but the results will **not** be deterministic due to differences in rounding error.
 - `-cb`, `--cont-batching`: Enable continuous batching (a.k.a dynamic batching).  Default: disabled
 - `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load a system prompt (initial prompt of all slots). This is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 - `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.

examples/server/public/index.html

@@ -594,7 +594,7 @@
           message = html`<${Probabilities} data=${data} />`
         } else {
           const text = isArrayMessage ?
-            data.map(msg => msg.content).join('').replace(/^\s+/, '') :
+            data.map(msg => msg.content).join('') :
             data;
           message = isCompletionMode ?
             text :
@@ -877,19 +877,30 @@
 
     // poor mans markdown replacement
     const Markdownish = (params) => {
-      const md = params.text
-        .replace(/&/g, '&amp;')
-        .replace(/</g, '&lt;')
-        .replace(/>/g, '&gt;')
-        .replace(/(^|\n)#{1,6} ([^\n]*)(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1<h3>$2</h3>')
-        .replace(/\*\*(.*?)\*\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '<strong>$1</strong>')
-        .replace(/__(.*?)__(?=([^`]*`[^`]*`)*[^`]*$)/g, '<strong>$1</strong>')
-        .replace(/\*(.*?)\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '<em>$1</em>')
-        .replace(/_(.*?)_(?=([^`]*`[^`]*`)*[^`]*$)/g, '<em>$1</em>')
-        .replace(/```.*?\n([\s\S]*?)```/g, '<pre><code>$1</code></pre>')
-        .replace(/`(.*?)`/g, '<code>$1</code>')
-        .replace(/\n/gim, '<br />');
-      return html`<span dangerouslySetInnerHTML=${{ __html: md }} />`;
+      const chunks = params.text.split('```');
+
+      for (let i = 0; i < chunks.length; i++) {
+        if (i % 2 === 0) { // outside code block
+          chunks[i] = chunks[i]
+          .replace(/&/g, '&amp;')
+          .replace(/</g, '&lt;')
+          .replace(/>/g, '&gt;')
+          .replace(/(^|\n)#{1,6} ([^\n]*)(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1<h3>$2</h3>')
+          .replace(/\*\*(.*?)\*\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '<strong>$1</strong>')
+          .replace(/__(.*?)__(?=([^`]*`[^`]*`)*[^`]*$)/g, '<strong>$1</strong>')
+          .replace(/\*(.*?)\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '<em>$1</em>')
+          .replace(/_(.*?)_(?=([^`]*`[^`]*`)*[^`]*$)/g, '<em>$1</em>')
+          .replace(/```.*?\n([\s\S]*?)```/g, '<pre><code>$1</code></pre>')
+          .replace(/`(.*?)`/g, '<code>$1</code>')
+          .replace(/\n/gim, '<br />');
+        } else { // inside code block
+          chunks[i] = `<pre><code>${chunks[i]}</code></pre>`;
+        }
+      }
+
+      const restoredText = chunks.join('');
+
+      return html`<span dangerouslySetInnerHTML=${{ __html: restoredText }} />`;
     };
 
     const ModelGenerationInfo = (params) => {
@@ -903,6 +914,7 @@
       `
     }
 
+
     // simple popover impl
     const Popover = (props) => {
       const isOpen = useSignal(false);
@@ -1054,4 +1066,3 @@
 </body>
 
 </html>
-

examples/server/public_simplechat/index.html

@@ -0,0 +1,49 @@
+<!DOCTYPE html>
+<html lang="en">
+    <head>
+        <title>SimpleChat LlamaCppEtal </title>
+        <meta charset="UTF-8" />
+        <meta name="viewport" content="width=device-width, initial-scale=1" />
+        <meta name="message" content="Save Nature Save Earth" />
+        <meta name="description" content="SimpleChat: trigger LLM web service endpoints /chat/completions and /completions, single/multi chat sessions" />
+        <meta name="author" content="by Humans for All" />
+        <meta http-equiv="Cache-Control" content="no-cache, no-store, must-revalidate" />
+        <script src="simplechat.js" defer></script>
+        <link rel="stylesheet" href="simplechat.css" />
+    </head>
+    <body>
+        <div class="samecolumn" id="fullbody">
+
+            <div class="sameline">
+                <p class="heading flex-grow" > <b> SimpleChat </b> </p>
+                <div class="sameline">
+                    <label for="api-ep">Mode:</label>
+                    <select name="api-ep" id="api-ep">
+                    <option value="chat" selected>Chat</option>
+                    <option value="completion">Completion</option>
+                    </select>
+                </div>
+            </div>
+
+            <div id="sessions-div" class="sameline"></div>
+
+            <hr>
+            <div class="sameline">
+                <label for="system-in">System</label>
+                <input type="text" name="system" id="system-in" placeholder="e.g. you are a helpful ai assistant, who provides concise answers" class="flex-grow"/>
+            </div>
+
+            <hr>
+            <div id="chat-div">
+                <p> You need to have javascript enabled.</p>
+            </div>
+
+            <hr>
+            <div class="sameline">
+                <textarea id="user-in" class="flex-grow" rows="3" placeholder="enter your query to the ai model here" ></textarea>
+                <button id="user-btn">submit</button>
+            </div>
+
+        </div>
+    </body>
+</html>

examples/server/public_simplechat/readme.md

@@ -0,0 +1,201 @@
+
+# SimpleChat
+
+by Humans for All.
+
+
+## overview
+
+This simple web frontend, allows triggering/testing the server's /completions or /chat/completions endpoints
+in a simple way with minimal code from a common code base. Inturn additionally it tries to allow single or
+multiple independent back and forth chatting to an extent, with the ai llm model at a basic level, with their
+own system prompts.
+
+The UI follows a responsive web design so that the layout can adapt to available display space in a usable
+enough manner, in general.
+
+Allows developer/end-user to control some of the behaviour by updating gMe members from browser's devel-tool
+console.
+
+NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and
+culling of old messages from the chat by default. However by enabling the sliding window chat logic, a crude
+form of old messages culling can be achieved.
+
+NOTE: It doesnt set any parameters other than temperature and max_tokens for now. However if someone wants
+they can update the js file or equivalent member in gMe as needed.
+
+
+## usage
+
+One could run this web frontend directly using server itself or if anyone is thinking of adding a built in web
+frontend to configure the server over http(s) or so, then run this web frontend using something like python's
+http module.
+
+### running using examples/server
+
+bin/server -m path/model.gguf --path ../examples/server/public_simplechat [--port PORT]
+
+### running using python3's server module
+
+first run examples/server
+* bin/server -m path/model.gguf
+
+next run this web front end in examples/server/public_simplechat
+* cd ../examples/server/public_simplechat
+* python3 -m http.server PORT
+
+### using the front end
+
+Open this simple web front end from your local browser
+
+* http://127.0.0.1:PORT/index.html
+
+Once inside
+
+* Select between chat and completion mode. By default it is set to chat mode.
+
+* In completion mode
+  * logic by default doesnt insert any role specific "ROLE: " prefix wrt each role's message.
+    If the model requires any prefix wrt user role messages, then the end user has to
+    explicitly add the needed prefix, when they enter their chat message.
+    Similarly if the model requires any prefix to trigger assistant/ai-model response,
+    then the end user needs to enter the same.
+    This keeps the logic simple, while still giving flexibility to the end user to
+    manage any templating/tagging requirement wrt their messages to the model.
+  * the logic doesnt insert newline at the begining and end wrt the prompt message generated.
+    However if the chat being sent to /completions end point has more than one role's message,
+    then insert newline when moving from one role's message to the next role's message, so
+    that it can be clearly identified/distinguished.
+  * given that /completions endpoint normally doesnt add additional chat-templating of its
+    own, the above ensures that end user can create a custom single/multi message combo with
+    any tags/special-tokens related chat templating to test out model handshake. Or enduser
+    can use it just for normal completion related/based query.
+
+* If you want to provide a system prompt, then ideally enter it first, before entering any user query.
+  Normally Completion mode doesnt need system prompt, while Chat mode can generate better/interesting
+  responses with a suitable system prompt.
+  * if chat.add_system_begin is used
+    * you cant change the system prompt, after it is has been submitted once along with user query.
+    * you cant set a system prompt, after you have submitted any user query
+  * if chat.add_system_anytime is used
+    * one can change the system prompt any time during chat, by changing the contents of system prompt.
+    * inturn the updated/changed system prompt will be inserted into the chat session.
+    * this allows for the subsequent user chatting to be driven by the new system prompt set above.
+
+* Enter your query and either press enter or click on the submit button.
+  If you want to insert enter (\n) as part of your chat/query to ai model, use shift+enter.
+
+* Wait for the logic to communicate with the server and get the response.
+  * the user is not allowed to enter any fresh query during this time.
+  * the user input box will be disabled and a working message will be shown in it.
+
+* just refresh the page, to reset wrt the chat history and or system prompt and start afresh.
+
+* Using NewChat one can start independent chat sessions.
+  * two independent chat sessions are setup by default.
+
+
+## Devel note
+
+### Reason behind this
+
+The idea is to be easy enough to use for basic purposes, while also being simple and easily discernable
+by developers who may not be from web frontend background (so inturn may not be familiar with template /
+end-use-specific-language-extensions driven flows) so that they can use it to explore/experiment things.
+
+And given that the idea is also to help explore/experiment for developers, some flexibility is provided
+to change behaviour easily using the devel-tools/console, for now. And skeletal logic has been implemented
+to explore some of the end points and ideas/implications around them.
+
+
+### General
+
+Me/gMe consolidates the settings which control the behaviour into one object.
+One can see the current settings, as well as change/update them using browsers devel-tool/console.
+
+  bCompletionFreshChatAlways - whether Completion mode collates complete/sliding-window history when
+  communicating with the server or only sends the latest user query/message.
+
+  bCompletionInsertStandardRolePrefix - whether Completion mode inserts role related prefix wrt the
+  messages that get inserted into prompt field wrt /Completion endpoint.
+
+  chatRequestOptions - maintains the list of options/fields to send along with chat request,
+  irrespective of whether /chat/completions or /completions endpoint.
+
+    If you want to add additional options/fields to send to the server/ai-model, and or
+    modify the existing options value or remove them, for now you can update this global var
+    using browser's development-tools/console.
+
+  iRecentUserMsgCnt - a simple minded SlidingWindow to limit context window load at Ai Model end.
+  This is disabled by default. However if enabled, then in addition to latest system message, only
+  the last/latest iRecentUserMsgCnt user messages after the latest system prompt and its responses
+  from the ai model will be sent to the ai-model, when querying for a new response. IE if enabled,
+  only user messages after the latest system message/prompt will be considered.
+
+    This specified sliding window user message count also includes the latest user query.
+    <0 : Send entire chat history to server
+     0 : Send only the system message if any to the server
+    >0 : Send the latest chat history from the latest system prompt, limited to specified cnt.
+
+
+By using gMe's iRecentUserMsgCnt and chatRequestOptions.max_tokens one can try to control the
+implications of loading of the ai-model's context window by chat history, wrt chat response to
+some extent in a simple crude way.
+
+
+Sometimes the browser may be stuborn with caching of the file, so your updates to html/css/js
+may not be visible. Also remember that just refreshing/reloading page in browser or for that
+matter clearing site data, dont directly override site caching in all cases. Worst case you may
+have to change port. Or in dev tools of browser, you may be able to disable caching fully.
+
+
+Concept of multiple chat sessions with different servers, as well as saving and restoring of
+those across browser usage sessions, can be woven around the SimpleChat/MultiChatUI class and
+its instances relatively easily, however given the current goal of keeping this simple, it has
+not been added, for now.
+
+
+By switching between chat.add_system_begin/anytime, one can control whether one can change
+the system prompt, anytime during the conversation or only at the beginning.
+
+
+read_json_early, is to experiment with reading json response data early on, if available,
+so that user can be shown generated data, as and when it is being generated, rather than
+at the end when full data is available.
+
+  the server flow doesnt seem to be sending back data early, atleast for request (inc options)
+  that is currently sent.
+
+  if able to read json data early on in future, as and when ai model is generating data, then
+  this helper needs to indirectly update the chat div with the recieved data, without waiting
+  for the overall data to be available.
+
+
+### Default setup
+
+By default things are setup to try and make the user experience a bit better, if possible.
+However a developer when testing the server of ai-model may want to change these value.
+
+Using iRecentUserMsgCnt reduce chat history context sent to the server/ai-model to be
+just the system-prompt, prev-user-request-and-ai-response and cur-user-request, instead of
+full chat history. This way if there is any response with garbage/repeatation, it doesnt
+mess with things beyond the next question/request/query, in some ways.
+
+Set max_tokens to 1024, so that a relatively large previous reponse doesnt eat up the space
+available wrt next query-response. However dont forget that the server when started should
+also be started with a model context size of 1k or more, to be on safe side.
+
+  The /completions endpoint of examples/server doesnt take max_tokens, instead it takes the
+  internal n_predict, for now add the same here on the client side, maybe later add max_tokens
+  to /completions endpoint handling code on server side.
+
+Frequency and presence penalty fields are set to 1.2 in the set of fields sent to server
+along with the user query. So that the model is partly set to try avoid repeating text in
+its response.
+
+A end-user can change these behaviour by editing gMe from browser's devel-tool/console.
+
+
+## At the end
+
+Also a thank you to all open source and open model developers, who strive for the common good.

examples/server/public_simplechat/simplechat.css

@@ -0,0 +1,68 @@
+/**
+ * the styling of the simplechat web frontend
+ * by Humans for All
+ */
+
+#fullbody {
+    height: 98vh;
+}
+
+.heading {
+    background-color: lightgray;
+}
+
+.session-selected {
+    background-color: lightblue;
+}
+
+.role-system {
+    background-color: lightblue;
+}
+.role-user {
+    background-color: lightgray;
+}
+
+.flex-grow {
+    flex-grow: 1;
+}
+.float-right {
+    float: right;
+}
+
+#chat-div {
+    overflow: scroll;
+    flex-grow: 1;
+    flex-shrink: 1;
+    min-height: 40vh;
+}
+button {
+    min-width: 8vw;
+}
+
+.sameline {
+    display: flex;
+    flex-direction: row;
+}
+.samecolumn {
+    display: flex;
+    flex-direction: column;
+}
+
+.ul1 {
+    padding-inline-start: 2vw;
+}
+.ul2 {
+    padding-inline-start: 2vw;
+}
+
+* {
+    margin: 0.6vmin;
+}
+
+@media print {
+
+    #fullbody {
+        height: auto;
+    }
+
+}

examples/server/public_simplechat/simplechat.js

@@ -0,0 +1,629 @@
+// @ts-check
+// A simple completions and chat/completions test related web front end logic
+// by Humans for All
+
+class Roles {
+    static System = "system";
+    static User = "user";
+    static Assistant = "assistant";
+}
+
+class ApiEP {
+    static Chat = "chat";
+    static Completion = "completion";
+}
+
+let gUsageMsg = `
+    <p class="role-system">Usage</p>
+    <ul class="ul1">
+    <li> Set system prompt above, to try control ai response charactersitic, if model supports same.</li>
+        <ul class="ul2">
+        <li> Completion mode normally wont have a system prompt.</li>
+        </ul>
+    <li> Enter your query to ai assistant below.</li>
+        <ul class="ul2">
+        <li> Completion mode doesnt insert user/role: prefix implicitly.</li>
+        <li> Use shift+enter for inserting enter/newline.</li>
+        </ul>
+    <li> Default ContextWindow = [System, Last Query+Resp, Cur Query].</li>
+        <ul class="ul2">
+        <li> experiment iRecentUserMsgCnt, max_tokens, model ctxt window to expand</li>
+        </ul>
+    </ul>
+`;
+
+/** @typedef {{role: string, content: string}[]} ChatMessages */
+
+class SimpleChat {
+
+    constructor() {
+        /**
+         * Maintain in a form suitable for common LLM web service chat/completions' messages entry
+         * @type {ChatMessages}
+         */
+        this.xchat = [];
+        this.iLastSys = -1;
+    }
+
+    clear() {
+        this.xchat = [];
+        this.iLastSys = -1;
+    }
+
+    /**
+     * Recent chat messages.
+     * If iRecentUserMsgCnt < 0
+     *   Then return the full chat history
+     * Else
+     *   Return chat messages from latest going back till the last/latest system prompt.
+     *   While keeping track that the number of user queries/messages doesnt exceed iRecentUserMsgCnt.
+     * @param {number} iRecentUserMsgCnt
+     */
+    recent_chat(iRecentUserMsgCnt) {
+        if (iRecentUserMsgCnt < 0) {
+            return this.xchat;
+        }
+        if (iRecentUserMsgCnt == 0) {
+            console.warn("WARN:SimpleChat:SC:RecentChat:iRecentUsermsgCnt of 0 means no user message/query sent");
+        }
+        /** @type{ChatMessages} */
+        let rchat = [];
+        let sysMsg = this.get_system_latest();
+        if (sysMsg.length != 0) {
+            rchat.push({role: Roles.System, content: sysMsg});
+        }
+        let iUserCnt = 0;
+        let iStart = this.xchat.length;
+        for(let i=this.xchat.length-1; i > this.iLastSys; i--) {
+            if (iUserCnt >= iRecentUserMsgCnt) {
+                break;
+            }
+            let msg = this.xchat[i];
+            if (msg.role == Roles.User) {
+                iStart = i;
+                iUserCnt += 1;
+            }
+        }
+        for(let i = iStart; i < this.xchat.length; i++) {
+            let msg = this.xchat[i];
+            if (msg.role == Roles.System) {
+                continue;
+            }
+            rchat.push({role: msg.role, content: msg.content});
+        }
+        return rchat;
+    }
+
+    /**
+     * Add an entry into xchat
+     * @param {string} role
+     * @param {string|undefined|null} content
+     */
+    add(role, content) {
+        if ((content == undefined) || (content == null) || (content == "")) {
+            return false;
+        }
+        this.xchat.push( {role: role, content: content} );
+        if (role == Roles.System) {
+            this.iLastSys = this.xchat.length - 1;
+        }
+        return true;
+    }
+
+    /**
+     * Show the contents in the specified div
+     * @param {HTMLDivElement} div
+     * @param {boolean} bClear
+     */
+    show(div, bClear=true) {
+        if (bClear) {
+            div.replaceChildren();
+        }
+        let last = undefined;
+        for(const x of this.recent_chat(gMe.iRecentUserMsgCnt)) {
+            let entry = document.createElement("p");
+            entry.className = `role-${x.role}`;
+            entry.innerText = `${x.role}: ${x.content}`;
+            div.appendChild(entry);
+            last = entry;
+        }
+        if (last !== undefined) {
+            last.scrollIntoView(false);
+        } else {
+            if (bClear) {
+                div.innerHTML = gUsageMsg;
+                gMe.show_info(div);
+            }
+        }
+    }
+
+    /**
+     * Add needed fields wrt json object to be sent wrt LLM web services completions endpoint.
+     * The needed fields/options are picked from a global object.
+     * Convert the json into string.
+     * @param {Object} obj
+     */
+    request_jsonstr(obj) {
+        for(let k in gMe.chatRequestOptions) {
+            obj[k] = gMe.chatRequestOptions[k];
+        }
+        return JSON.stringify(obj);
+    }
+
+    /**
+     * Return a string form of json object suitable for chat/completions
+     */
+    request_messages_jsonstr() {
+        let req = {
+            messages: this.recent_chat(gMe.iRecentUserMsgCnt),
+        }
+        return this.request_jsonstr(req);
+    }
+
+    /**
+     * Return a string form of json object suitable for /completions
+     * @param {boolean} bInsertStandardRolePrefix Insert "<THE_ROLE>: " as prefix wrt each role's message
+     */
+    request_prompt_jsonstr(bInsertStandardRolePrefix) {
+        let prompt = "";
+        let iCnt = 0;
+        for(const chat of this.recent_chat(gMe.iRecentUserMsgCnt)) {
+            iCnt += 1;
+            if (iCnt > 1) {
+                prompt += "\n";
+            }
+            if (bInsertStandardRolePrefix) {
+                prompt += `${chat.role}: `;
+            }
+            prompt += `${chat.content}`;
+        }
+        let req = {
+            prompt: prompt,
+        }
+        return this.request_jsonstr(req);
+    }
+
+    /**
+     * Allow setting of system prompt, but only at begining.
+     * @param {string} sysPrompt
+     * @param {string} msgTag
+     */
+    add_system_begin(sysPrompt, msgTag) {
+        if (this.xchat.length == 0) {
+            if (sysPrompt.length > 0) {
+                return this.add(Roles.System, sysPrompt);
+            }
+        } else {
+            if (sysPrompt.length > 0) {
+                if (this.xchat[0].role !== Roles.System) {
+                    console.error(`ERRR:SimpleChat:SC:${msgTag}:You need to specify system prompt before any user query, ignoring...`);
+                } else {
+                    if (this.xchat[0].content !== sysPrompt) {
+                        console.error(`ERRR:SimpleChat:SC:${msgTag}:You cant change system prompt, mid way through, ignoring...`);
+                    }
+                }
+            }
+        }
+        return false;
+    }
+
+    /**
+     * Allow setting of system prompt, at any time.
+     * @param {string} sysPrompt
+     * @param {string} msgTag
+     */
+    add_system_anytime(sysPrompt, msgTag) {
+        if (sysPrompt.length <= 0) {
+            return false;
+        }
+
+        if (this.iLastSys < 0) {
+            return this.add(Roles.System, sysPrompt);
+        }
+
+        let lastSys = this.xchat[this.iLastSys].content;
+        if (lastSys !== sysPrompt) {
+            return this.add(Roles.System, sysPrompt);
+        }
+        return false;
+    }
+
+    /**
+     * Retrieve the latest system prompt.
+     */
+    get_system_latest() {
+        if (this.iLastSys == -1) {
+            return "";
+        }
+        let sysPrompt = this.xchat[this.iLastSys].content;
+        return sysPrompt;
+    }
+
+}
+
+
+let gBaseURL = "http://127.0.0.1:8080";
+let gChatURL = {
+    'chat': `${gBaseURL}/chat/completions`,
+    'completion': `${gBaseURL}/completions`,
+}
+
+
+/**
+ * Set the class of the children, based on whether it is the idSelected or not.
+ * @param {HTMLDivElement} elBase
+ * @param {string} idSelected
+ * @param {string} classSelected
+ * @param {string} classUnSelected
+ */
+function el_children_config_class(elBase, idSelected, classSelected, classUnSelected="") {
+    for(let child of elBase.children) {
+        if (child.id == idSelected) {
+            child.className = classSelected;
+        } else {
+            child.className = classUnSelected;
+        }
+    }
+}
+
+/**
+ * Create button and set it up.
+ * @param {string} id
+ * @param {(this: HTMLButtonElement, ev: MouseEvent) => any} callback
+ * @param {string | undefined} name
+ * @param {string | undefined} innerText
+ */
+function el_create_button(id, callback, name=undefined, innerText=undefined) {
+    if (!name) {
+        name = id;
+    }
+    if (!innerText) {
+        innerText = id;
+    }
+    let btn = document.createElement("button");
+    btn.id = id;
+    btn.name = name;
+    btn.innerText = innerText;
+    btn.addEventListener("click", callback);
+    return btn;
+}
+
+
+class MultiChatUI {
+
+    constructor() {
+        /** @type {Object<string, SimpleChat>} */
+        this.simpleChats = {};
+        /** @type {string} */
+        this.curChatId = "";
+
+        // the ui elements
+        this.elInSystem = /** @type{HTMLInputElement} */(document.getElementById("system-in"));
+        this.elDivChat = /** @type{HTMLDivElement} */(document.getElementById("chat-div"));
+        this.elBtnUser = /** @type{HTMLButtonElement} */(document.getElementById("user-btn"));
+        this.elInUser = /** @type{HTMLInputElement} */(document.getElementById("user-in"));
+        this.elSelectApiEP = /** @type{HTMLSelectElement} */(document.getElementById("api-ep"));
+        this.elDivSessions = /** @type{HTMLDivElement} */(document.getElementById("sessions-div"));
+
+        this.validate_element(this.elInSystem, "system-in");
+        this.validate_element(this.elDivChat, "chat-div");
+        this.validate_element(this.elInUser, "user-in");
+        this.validate_element(this.elSelectApiEP, "api-ep");
+        this.validate_element(this.elDivChat, "sessions-div");
+    }
+
+    /**
+     * Check if the element got
+     * @param {HTMLElement | null} el
+     * @param {string} msgTag
+     */
+    validate_element(el, msgTag) {
+        if (el == null) {
+            throw Error(`ERRR:SimpleChat:MCUI:${msgTag} element missing in html...`);
+        } else {
+            console.debug(`INFO:SimpleChat:MCUI:${msgTag} Id[${el.id}] Name[${el["name"]}]`);
+        }
+    }
+
+    /**
+     * Reset user input ui.
+     * * clear user input
+     * * enable user input
+     * * set focus to user input
+     */
+    ui_reset_userinput() {
+        this.elInUser.value = "";
+        this.elInUser.disabled = false;
+        this.elInUser.focus();
+    }
+
+    /**
+     * Setup the needed callbacks wrt UI, curChatId to defaultChatId and
+     * optionally switch to specified defaultChatId.
+     * @param {string} defaultChatId
+     * @param {boolean} bSwitchSession
+     */
+    setup_ui(defaultChatId, bSwitchSession=false) {
+
+        this.curChatId = defaultChatId;
+        if (bSwitchSession) {
+            this.handle_session_switch(this.curChatId);
+        }
+
+        this.elBtnUser.addEventListener("click", (ev)=>{
+            if (this.elInUser.disabled) {
+                return;
+            }
+            this.handle_user_submit(this.curChatId, this.elSelectApiEP.value).catch((/** @type{Error} */reason)=>{
+                let msg = `ERRR:SimpleChat\nMCUI:HandleUserSubmit:${this.curChatId}\n${reason.name}:${reason.message}`;
+                console.debug(msg.replace("\n", ":"));
+                alert(msg);
+                this.ui_reset_userinput();
+            });
+        });
+
+        this.elInUser.addEventListener("keyup", (ev)=> {
+            // allow user to insert enter into their message using shift+enter.
+            // while just pressing enter key will lead to submitting.
+            if ((ev.key === "Enter") && (!ev.shiftKey)) {
+                let value = this.elInUser.value;
+                this.elInUser.value = value.substring(0,value.length-1);
+                this.elBtnUser.click();
+                ev.preventDefault();
+            }
+        });
+
+        this.elInSystem.addEventListener("keyup", (ev)=> {
+            // allow user to insert enter into the system prompt using shift+enter.
+            // while just pressing enter key will lead to setting the system prompt.
+            if ((ev.key === "Enter") && (!ev.shiftKey)) {
+                let chat = this.simpleChats[this.curChatId];
+                chat.add_system_anytime(this.elInSystem.value, this.curChatId);
+                chat.show(this.elDivChat);
+                ev.preventDefault();
+            }
+        });
+
+    }
+
+    /**
+     * Setup a new chat session and optionally switch to it.
+     * @param {string} chatId
+     * @param {boolean} bSwitchSession
+     */
+    new_chat_session(chatId, bSwitchSession=false) {
+        this.simpleChats[chatId] = new SimpleChat();
+        if (bSwitchSession) {
+            this.handle_session_switch(chatId);
+        }
+    }
+
+    /**
+     * Try read json response early, if available.
+     * @param {Response} resp
+     */
+    async read_json_early(resp) {
+        if (!resp.body) {
+            throw Error("ERRR:SimpleChat:MCUI:ReadJsonEarly:No body...");
+        }
+        let tdUtf8 = new TextDecoder("utf-8");
+        let rr = resp.body.getReader();
+        let gotBody = "";
+        while(true) {
+            let { value: cur,  done: done} = await rr.read();
+            let curBody = tdUtf8.decode(cur);
+            console.debug("DBUG:SC:PART:", curBody);
+            gotBody += curBody;
+            if (done) {
+                break;
+            }
+        }
+        return JSON.parse(gotBody);
+    }
+
+    /**
+     * Handle user query submit request, wrt specified chat session.
+     * @param {string} chatId
+     * @param {string} apiEP
+     */
+    async handle_user_submit(chatId, apiEP) {
+
+        let chat = this.simpleChats[chatId];
+
+        // In completion mode, if configured, clear any previous chat history.
+        // So if user wants to simulate a multi-chat based completion query,
+        // they will have to enter the full thing, as a suitable multiline
+        // user input/query.
+        if ((apiEP == ApiEP.Completion) && (gMe.bCompletionFreshChatAlways)) {
+            chat.clear();
+        }
+
+        chat.add_system_anytime(this.elInSystem.value, chatId);
+
+        let content = this.elInUser.value;
+        if (!chat.add(Roles.User, content)) {
+            console.debug(`WARN:SimpleChat:MCUI:${chatId}:HandleUserSubmit:Ignoring empty user input...`);
+            return;
+        }
+        chat.show(this.elDivChat);
+
+        let theBody;
+        let theUrl = gChatURL[apiEP]
+        if (apiEP == ApiEP.Chat) {
+            theBody = chat.request_messages_jsonstr();
+        } else {
+            theBody = chat.request_prompt_jsonstr(gMe.bCompletionInsertStandardRolePrefix);
+        }
+
+        this.elInUser.value = "working...";
+        this.elInUser.disabled = true;
+        console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:${theUrl}:ReqBody:${theBody}`);
+        let resp = await fetch(theUrl, {
+            method: "POST",
+            headers: {
+                "Content-Type": "application/json",
+            },
+            body: theBody,
+        });
+
+        let respBody = await resp.json();
+        //let respBody = await this.read_json_early(resp);
+        console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:RespBody:${JSON.stringify(respBody)}`);
+        let assistantMsg;
+        if (apiEP == ApiEP.Chat) {
+            assistantMsg = respBody["choices"][0]["message"]["content"];
+        } else {
+            try {
+                assistantMsg = respBody["choices"][0]["text"];
+            } catch {
+                assistantMsg = respBody["content"];
+            }
+        }
+        chat.add(Roles.Assistant, assistantMsg);
+        if (chatId == this.curChatId) {
+            chat.show(this.elDivChat);
+        } else {
+            console.debug(`DBUG:SimpleChat:MCUI:HandleUserSubmit:ChatId has changed:[${chatId}] [${this.curChatId}]`);
+        }
+        this.ui_reset_userinput();
+    }
+
+    /**
+     * Show buttons for NewChat and available chat sessions, in the passed elDiv.
+     * If elDiv is undefined/null, then use this.elDivSessions.
+     * Take care of highlighting the selected chat-session's btn.
+     * @param {HTMLDivElement | undefined} elDiv
+     */
+    show_sessions(elDiv=undefined) {
+        if (!elDiv) {
+            elDiv = this.elDivSessions;
+        }
+        elDiv.replaceChildren();
+        // Btn for creating new chat session
+        let btnNew = el_create_button("New CHAT", (ev)=> {
+            if (this.elInUser.disabled) {
+                console.error(`ERRR:SimpleChat:MCUI:NewChat:Current session [${this.curChatId}] awaiting response, ignoring request...`);
+                alert("ERRR:SimpleChat\nMCUI:NewChat\nWait for response to pending query, before starting new chat session");
+                return;
+            }
+            let chatId = `Chat${Object.keys(this.simpleChats).length}`;
+            let chatIdGot = prompt("INFO:SimpleChat\nMCUI:NewChat\nEnter id for new chat session", chatId);
+            if (!chatIdGot) {
+                console.error("ERRR:SimpleChat:MCUI:NewChat:Skipping based on user request...");
+                return;
+            }
+            this.new_chat_session(chatIdGot, true);
+            this.create_session_btn(elDiv, chatIdGot);
+            el_children_config_class(elDiv, chatIdGot, "session-selected", "");
+        });
+        elDiv.appendChild(btnNew);
+        // Btns for existing chat sessions
+        let chatIds = Object.keys(this.simpleChats);
+        for(let cid of chatIds) {
+            let btn = this.create_session_btn(elDiv, cid);
+            if (cid == this.curChatId) {
+                btn.className = "session-selected";
+            }
+        }
+    }
+
+    create_session_btn(elDiv, cid) {
+        let btn = el_create_button(cid, (ev)=>{
+            let target = /** @type{HTMLButtonElement} */(ev.target);
+            console.debug(`DBUG:SimpleChat:MCUI:SessionClick:${target.id}`);
+            if (this.elInUser.disabled) {
+                console.error(`ERRR:SimpleChat:MCUI:SessionClick:${target.id}:Current session [${this.curChatId}] awaiting response, ignoring switch...`);
+                alert("ERRR:SimpleChat\nMCUI:SessionClick\nWait for response to pending query, before switching");
+                return;
+            }
+            this.handle_session_switch(target.id);
+            el_children_config_class(elDiv, target.id, "session-selected", "");
+        });
+        elDiv.appendChild(btn);
+        return btn;
+    }
+
+    /**
+     * Switch ui to the specified chatId and set curChatId to same.
+     * @param {string} chatId
+     */
+    async handle_session_switch(chatId) {
+        let chat = this.simpleChats[chatId];
+        if (chat == undefined) {
+            console.error(`ERRR:SimpleChat:MCUI:HandleSessionSwitch:${chatId} missing...`);
+            return;
+        }
+        this.elInSystem.value = chat.get_system_latest();
+        this.elInUser.value = "";
+        chat.show(this.elDivChat);
+        this.elInUser.focus();
+        this.curChatId = chatId;
+        console.log(`INFO:SimpleChat:MCUI:HandleSessionSwitch:${chatId} entered...`);
+    }
+
+}
+
+
+class Me {
+
+    constructor() {
+        this.defaultChatIds = [ "Default", "Other" ];
+        this.multiChat = new MultiChatUI();
+        this.bCompletionFreshChatAlways = true;
+        this.bCompletionInsertStandardRolePrefix = false;
+        this.iRecentUserMsgCnt = 2;
+        // Add needed fields wrt json object to be sent wrt LLM web services completions endpoint.
+        this.chatRequestOptions = {
+            "temperature": 0.7,
+            "max_tokens": 1024,
+            "frequency_penalty": 1.2,
+            "presence_penalty": 1.2,
+            "n_predict": 1024
+        };
+    }
+
+    /**
+     * @param {HTMLDivElement} elDiv
+     */
+    show_info(elDiv) {
+
+        var p = document.createElement("p");
+        p.innerText = "Settings (devel-tools-console gMe)";
+        p.className = "role-system";
+        elDiv.appendChild(p);
+
+        var p = document.createElement("p");
+        p.innerText = `bCompletionFreshChatAlways:${this.bCompletionFreshChatAlways}`;
+        elDiv.appendChild(p);
+
+        p = document.createElement("p");
+        p.innerText = `bCompletionInsertStandardRolePrefix:${this.bCompletionInsertStandardRolePrefix}`;
+        elDiv.appendChild(p);
+
+        p = document.createElement("p");
+        p.innerText = `iRecentUserMsgCnt:${this.iRecentUserMsgCnt}`;
+        elDiv.appendChild(p);
+
+        p = document.createElement("p");
+        p.innerText = `chatRequestOptions:${JSON.stringify(this.chatRequestOptions)}`;
+        elDiv.appendChild(p);
+
+    }
+
+}
+
+
+/** @type {Me} */
+let gMe;
+
+function startme() {
+    console.log("INFO:SimpleChat:StartMe:Starting...");
+    gMe = new Me();
+    for (let cid of gMe.defaultChatIds) {
+        gMe.multiChat.new_chat_session(cid);
+    }
+    gMe.multiChat.setup_ui(gMe.defaultChatIds[0], true);
+    gMe.multiChat.show_sessions();
+}
+
+document.addEventListener("DOMContentLoaded", startme);

examples/server/server.cpp

@@ -102,7 +102,6 @@ struct slot_params {
     bool stream       = true;
     bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt
 
-    uint32_t seed      = -1; // RNG seed
     int32_t  n_keep    =  0; // number of tokens to keep from initial prompt
     int32_t  n_discard =  0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half
     int32_t  n_predict = -1; // new tokens to predict
@@ -1020,7 +1019,7 @@ struct server_context {
                         sampler_names.emplace_back(sampler_name);
                     }
                 }
-                slot.sparams.samplers_sequence = sampler_types_from_names(sampler_names, false);
+                slot.sparams.samplers_sequence = llama_sampling_types_from_names(sampler_names, false);
             } else {
                 slot.sparams.samplers_sequence = default_sparams.samplers_sequence;
             }
@@ -1257,14 +1256,14 @@ struct server_context {
         std::vector<std::string> samplers_sequence;
         samplers_sequence.reserve(slot.sparams.samplers_sequence.size());
         for (const auto & sampler_type : slot.sparams.samplers_sequence) {
-            samplers_sequence.emplace_back(sampler_type_to_name_string(sampler_type));
+            samplers_sequence.emplace_back(llama_sampling_type_to_str(sampler_type));
         }
 
         return json {
             {"n_ctx",                     slot.n_ctx},
             {"n_predict",                 slot.n_predict},
             {"model",                     params.model_alias},
-            {"seed",                      slot.params.seed},
+            {"seed",                      slot.sparams.seed},
             {"temperature",               slot.sparams.temp},
             {"dynatemp_range",            slot.sparams.dynatemp_range},
             {"dynatemp_exponent",         slot.sparams.dynatemp_exponent},
@@ -1982,8 +1981,7 @@ struct server_context {
                                 slot.state = SLOT_STATE_PROCESSING;
                                 slot.command = SLOT_COMMAND_NONE;
                                 slot.release();
-                                slot.print_timings();
-                                send_final_response(slot);
+                                send_error(slot, "input is too large to process. increase the physical batch size", ERROR_TYPE_SERVER);
                                 continue;
                             }
                         } else {
@@ -2854,7 +2852,7 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                 invalid_param = true;
                 break;
             }
-            if (!parse_kv_override(argv[i], params.kv_overrides)) {
+            if (!string_parse_kv_override(argv[i], params.kv_overrides)) {
                 fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
                 invalid_param = true;
                 break;
@@ -3312,7 +3310,7 @@ int main(int argc, char ** argv) {
     const auto handle_slots_save = [&ctx_server, &res_error, &sparams](const httplib::Request & req, httplib::Response & res, int id_slot) {
         json request_data = json::parse(req.body);
         std::string filename = request_data.at("filename");
-        if (!validate_file_name(filename)) {
+        if (!fs_validate_filename(filename)) {
             res_error(res, format_error_response("Invalid filename", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
@@ -3342,7 +3340,7 @@ int main(int argc, char ** argv) {
     const auto handle_slots_restore = [&ctx_server, &res_error, &sparams](const httplib::Request & req, httplib::Response & res, int id_slot) {
         json request_data = json::parse(req.body);
         std::string filename = request_data.at("filename");
-        if (!validate_file_name(filename)) {
+        if (!fs_validate_filename(filename)) {
             res_error(res, format_error_response("Invalid filename", ERROR_TYPE_INVALID_REQUEST));
             return;
         }

examples/server/tests/features/results.feature

@@ -13,6 +13,7 @@ Feature: Results
 
   Scenario Outline: consistent results with same seed
     Given <n_slots> slots
+    And   1.0 temperature
     Then  the server is starting
     Then  the server is healthy
 
@@ -26,10 +27,12 @@ Feature: Results
     Examples:
       | n_slots |
       | 1       |
-      | 2       |
+      # FIXME: unified KV cache nondeterminism
+      # | 2       |
 
   Scenario Outline: different results with different seed
     Given <n_slots> slots
+    And   1.0 temperature
     Then  the server is starting
     Then  the server is healthy
 
@@ -70,12 +73,46 @@ Feature: Results
     Then all predictions are equal
     Examples:
       | n_parallel | temp |
-      |  1         | 0.0  |
-      |  2         | 0.0  |
-      |  4         | 0.0  |
-      |  1         | 1.0  |
-      # FIXME: These tests fail on master. The problem seems to be the unified KV cache.
+      | 1          | 0.0  |
+      | 1          | 1.0  |
+      # FIXME: unified KV cache nondeterminism
       # See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
-      # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574 .
-      # |  2         | 1.0  |
-      # |  4         | 1.0  |
+      # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574
+      # and https://github.com/ggerganov/llama.cpp/pull/7347 .
+      # | 2          | 0.0  |
+      # | 4          | 0.0  |
+      # | 2          | 1.0  |
+      # | 4          | 1.0  |
+
+  Scenario Outline: consistent token probs with same seed and prompt
+    Given <n_slots> slots
+    And   <n_kv> KV cache size
+    And   1.0 temperature
+    And   <n_predict> max tokens to predict
+    Then  the server is starting
+    Then  the server is healthy
+
+    Given 1 prompts "The meaning of life is" with seed 42
+    And   concurrent completion requests
+    # Then the server is busy # Not all slots will be utilized.
+    Then  the server is idle
+    And   all slots are idle
+
+    Given <n_parallel> prompts "The meaning of life is" with seed 42
+    And   concurrent completion requests
+    # Then the server is busy # Not all slots will be utilized.
+    Then the server is idle
+    And  all slots are idle
+
+    Then all token probabilities are equal
+    Examples:
+      | n_slots | n_kv | n_predict | n_parallel |
+      | 4       | 1024 | 1         | 1          |
+      # FIXME: unified KV cache nondeterminism
+      # See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
+      # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574
+      # and https://github.com/ggerganov/llama.cpp/pull/7347 .
+      # | 4       | 1024 | 1         | 4          |
+      # | 4       | 1024 | 100       | 1          |
+      # This test still fails even the above patches; the first token probabilities are already different.
+      # | 4       | 1024 | 100       | 4          |

examples/server/tests/features/steps/steps.py

@@ -23,6 +23,7 @@ from prometheus_client import parser
 def step_server_config(context, server_fqdn, server_port):
     context.server_fqdn = server_fqdn
     context.server_port = int(server_port)
+    context.n_threads = None
     context.n_gpu_layer = None
     if 'PORT' in os.environ:
         context.server_port = int(os.environ['PORT'])
@@ -109,6 +110,11 @@ def step_n_gpu_layer(context, ngl):
     context.n_gpu_layer = ngl
 
 
+@step('{n_threads:d} threads')
+def step_n_threads(context, n_threads):
+    context.n_thread = n_threads
+
+
 @step('{draft:d} as draft')
 def step_draft(context, draft):
     context.draft = draft
@@ -193,7 +199,7 @@ async def step_wait_for_the_server_to_be_started(context, expecting_status):
 
         case 'ready' | 'idle':
             await wait_for_health_status(context, context.base_url, 200, 'ok',
-                                         timeout=10,
+                                         timeout=30,
                                          params={'fail_on_no_slot': 0, 'include_slots': 0},
                                          slots_idle=context.n_slots,
                                          slots_processing=0,
@@ -274,13 +280,22 @@ async def step_predictions_equal(context):
 
 @step('all predictions are different')
 @async_run_until_complete
-async def step_predictions_equal(context):
+async def step_predictions_different(context):
     n_completions = await gather_tasks_results(context)
     assert n_completions >= 2, "need at least 2 completions"
     assert_all_predictions_different(context.tasks_result)
     context.tasks_result = []
 
 
+@step('all token probabilities are equal')
+@async_run_until_complete
+async def step_token_probabilities_equal(context):
+    n_completions = await gather_tasks_results(context)
+    assert n_completions >= 2, "need at least 2 completions"
+    assert_all_token_probabilities_equal(context.tasks_result)
+    context.tasks_result = []
+
+
 @step('the completion is  truncated')
 def step_assert_completion_truncated(context):
     step_assert_completion_truncated(context, '')
@@ -868,7 +883,8 @@ async def request_completion(prompt,
                                     "cache_prompt": cache_prompt,
                                     "id_slot": id_slot,
                                     "seed": seed if seed is not None else 42,
-                                    "temperature": temperature if temperature is not None else "0.8f",
+                                    "temperature": temperature if temperature is not None else 0.8,
+                                    "n_probs": 2,
                                 },
                                 headers=headers,
                                 timeout=3600) as response:
@@ -1123,6 +1139,23 @@ def assert_all_predictions_different(completion_responses):
         assert content_i != content_j, "contents not different"
 
 
+def assert_all_token_probabilities_equal(completion_responses):
+    n_predict = len(completion_responses[0]['completion_probabilities'])
+    if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
+        for pos in range(n_predict):
+            for i, response_i in enumerate(completion_responses):
+                probs_i = response_i['completion_probabilities'][pos]['probs']
+                print(f"pos {pos}, probs {i}: {probs_i}")
+    for pos in range(n_predict):
+        for i, response_i in enumerate(completion_responses):
+            probs_i = response_i['completion_probabilities'][pos]['probs']
+            for j, response_j in enumerate(completion_responses):
+                if i == j:
+                    continue
+                probs_j = response_j['completion_probabilities'][pos]['probs']
+            assert probs_i == probs_j, "contents not equal"
+
+
 async def gather_tasks_results(context):
     n_tasks = len(context.concurrent_tasks)
     if context.debug:
@@ -1261,6 +1294,8 @@ def start_server_background(context):
         server_args.extend(['--batch-size', context.n_batch])
     if context.n_ubatch:
         server_args.extend(['--ubatch-size', context.n_ubatch])
+    if context.n_threads:
+        server_args.extend(['--threads', context.threads])
     if context.n_gpu_layer:
         server_args.extend(['--n-gpu-layers', context.n_gpu_layer])
     if context.draft is not None:

examples/sycl/win-build-sycl.bat

@@ -13,10 +13,10 @@ if %errorlevel% neq 0 goto ERROR
 
 ::  for FP16
 ::  faster for long-prompt inference
-::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
+::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
 ::  for FP32
-cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
 if %errorlevel% neq 0 goto ERROR
 ::  build example/main only
 ::  make main

examples/tokenize/tokenize.cpp

@@ -3,40 +3,390 @@
 
 #include <cmath>
 #include <cstdio>
+#include <fstream>
 #include <string>
 #include <vector>
 
-int main(int argc, char ** argv) {
-    if (argc < 3 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH PROMPT [--ids]\n" , argv[0]);
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#include <shellapi.h>   // For CommandLineToArgvW
+#endif
+
+static void print_usage_information(const char * argv0, FILE * stream) {
+    fprintf(stream, "usage: %s [options]\n\n", argv0);
+    fprintf(stream, "The tokenize program tokenizes a prompt using a given model,\n");
+    fprintf(stream, "and prints the resulting tokens to standard output.\n\n");
+    fprintf(stream, "It needs a model file, a prompt, and optionally other flags\n");
+    fprintf(stream, "to control the behavior of the tokenizer.\n\n");
+    fprintf(stream, "    The possible options are:\n");
+    fprintf(stream, "\n");
+    fprintf(stream, "    -h, --help                           print this help and exit\n");
+    fprintf(stream, "    -m MODEL_PATH, --model MODEL_PATH    path to model.\n");
+    fprintf(stream, "    --ids                                if given, only print numerical token IDs, and not token strings.\n");
+    fprintf(stream, "                                         The output format looks like [1, 2, 3], i.e. parseable by Python.\n");
+    fprintf(stream, "    -f PROMPT_FNAME, --file PROMPT_FNAME read prompt from a file.\n");
+    fprintf(stream, "    -p PROMPT, --prompt PROMPT           read prompt from the argument.\n");
+    fprintf(stream, "    --stdin                              read prompt from standard input.\n");
+    fprintf(stream, "    --no-bos                             do not ever add a BOS token to the prompt, even if normally the model uses a BOS token.\n");
+    fprintf(stream, "    --log-disable                        disable logs. Makes stderr quiet when loading the model.\n");
+}
+
+static void llama_log_callback_null(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) text;
+    (void) user_data;
+}
+
+static std::string read_prompt_from_file(const char * filepath, bool & success) {
+    success = false;
+
+    std::ifstream in(filepath, std::ios::binary);
+    if (!in) {
+        fprintf(stderr, "%s: could not open file '%s' for reading: %s\n", __func__, filepath, strerror(errno));
+        return std::string();
+    }
+    // do not assume the file is seekable (e.g. /dev/stdin)
+    std::stringstream buffer;
+    buffer << in.rdbuf();
+    if (in.fail()) {
+        fprintf(stderr, "%s: could not read the entire file '%s': %s\n", __func__, filepath, strerror(errno));
+        return std::string();
+    }
+
+    success = true;
+    return buffer.str();
+}
+
+//
+// Function: ingest_args(...) -> vector<string>
+//
+//  Takes argc and argv arguments, and converts them to a vector of UTF-8 encoded
+//  strings, as an STL vector<string>.
+//
+//  In particular, it handles character encoding shenanigans on Windows.
+//
+// Note: raw_argc and raw_argv are not actually read at all on Windows.
+//       On Windows we call GetCommandLineW to get the arguments in wchar_t
+//       format, ignoring the regular argc/argv arguments to main().
+//
+// TODO: potential opportunity to roll common stuff into common/console.cpp
+//       in relation to Windows wchar_t shenanigans.
+static std::vector<std::string> ingest_args(int raw_argc, char ** raw_argv) {
+    std::vector<std::string> argv;
+
+    // Handle Windows, if given non-ASCII arguments.
+    // We convert wchar_t arguments into UTF-8 char* on this platform.
+    // Lets you invoke 'tokenize' on Windows cmd.exe with non-ASCII characters
+    // without throwing tantrums.
+#if defined(_WIN32)
+    int argc;
+    const LPWSTR cmdline_wargv = GetCommandLineW();
+    LPWSTR * wargv = CommandLineToArgvW(cmdline_wargv, &argc);
+
+    // silence unused arg warnings
+    (void) raw_argc;
+    (void) raw_argv;
+
+    for (int i = 0; i < argc; ++i) {
+        int length_needed = WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), 0, 0, NULL, NULL);
+        char * output_buf = (char *) calloc(length_needed+1, sizeof(char));
+        GGML_ASSERT(output_buf);
+
+        WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), output_buf, length_needed, NULL, NULL);
+        output_buf[length_needed] = '\0';
+
+        argv.push_back(output_buf);
+        free(output_buf);
+    }
+
+    LocalFree((HLOCAL) wargv);
+#else
+    int argc = raw_argc;
+    for (int i = 0; i < argc; ++i) {
+        argv.push_back(raw_argv[i]);
+    }
+#endif
+
+    GGML_ASSERT((unsigned int) argc == argv.size());
+
+    return argv;
+}
+
+//
+// Function: write_utf8_cstr_to_stdout(const char *) -> <writes to stdout>
+//
+// writes a string to standard output; taking into account that on Windows
+// to display correctly you have to use special handling. Works even if the
+// user has not set a unicode code page on a Windows cmd.exe.
+//
+// In case of invalid UTF-8, invalid_utf8 is set to true on Windows, and something
+// a human-readable is written instead.
+//
+// On non-Windows systems, simply printfs() the string.
+static void write_utf8_cstr_to_stdout(const char * str, bool & invalid_utf8) {
+        invalid_utf8 = false;
+
+#if defined(_WIN32)
+        // Are we in a console?
+        HANDLE hConsole = GetStdHandle(STD_OUTPUT_HANDLE);
+        DWORD dwMode = 0;
+
+        // According to Microsoft docs:
+        // "WriteConsole fails if it is used with a standard handle that is redirected to a file."
+        // Also according to the docs, you can use GetConsoleMode to check for that.
+        if (hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(hConsole, &dwMode)) {
+            printf("%s", str);
+            return;
+        }
+
+        // MultiByteToWideChar reports an error if str is empty, don't report
+        // them as invalid_utf8.
+        if (*str == 0) {
+            return;
+        }
+        int length_needed = MultiByteToWideChar(CP_UTF8, MB_ERR_INVALID_CHARS, str, strlen(str), NULL, 0);
+        if (length_needed == 0) {
+            DWORD err = GetLastError();
+            if (err == ERROR_NO_UNICODE_TRANSLATION) {
+                invalid_utf8 = true;
+                int len = strlen(str);
+                printf("<");
+                for (int i = 0; i < len; ++i) {
+                    if (i > 0) {
+                        printf(" ");
+                    }
+                    printf("%02x", (uint8_t) str[i]);
+                }
+                printf(">");
+                return;
+            }
+            GGML_ASSERT(false && "MultiByteToWideChar() failed in an unexpected way.");
+        }
+
+        LPWSTR wstr = (LPWSTR) calloc(length_needed+1, sizeof(*wstr));
+        GGML_ASSERT(wstr);
+
+        MultiByteToWideChar(CP_UTF8, 0, str, strlen(str), wstr, length_needed);
+        WriteConsoleW(hConsole, wstr, length_needed, NULL, NULL);
+
+        free(wstr);
+#else
+        // TODO: reporting invalid_utf8 would be useful on non-Windows too.
+        // printf will silently just write bad unicode.
+        printf("%s", str);
+#endif
+}
+
+int main(int raw_argc, char ** raw_argv) {
+    const std::vector<std::string> argv = ingest_args(raw_argc, raw_argv);
+    const int argc = argv.size();
+
+    if (argc <= 1) {
+        print_usage_information(argv[0].c_str(), stderr);
         return 1;
     }
 
-    const char * model_path = argv[1];
-    const char * prompt     = argv[2];
+    //////
+    // Read out all the command line arguments.
+    //////
 
-    const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids";
+    // variables where to put any arguments we see.
+    bool printing_ids = false;
+    bool no_bos = false;
+    bool disable_logging = false;
+    const char * model_path = NULL;
+    const char * prompt_path = NULL;
+    const char * prompt_arg = NULL;
+
+    // track which arguments were explicitly given
+    // used for sanity checking down the line
+    bool model_path_set = false;
+    bool prompt_path_set = false;
+    bool prompt_set = false;
+    bool stdin_set = false;
+
+    int iarg = 1;
+    for (; iarg < argc; ++iarg) {
+        std::string arg{argv[iarg]};
+        if (arg == "-h" || arg == "--help") {
+            print_usage_information(argv[0].c_str(), stdout);
+            return 0;
+        }
+        else if (arg == "--ids") {
+            printing_ids = true;
+        }
+        else if (arg == "-m" || arg == "--model") {
+            if (model_path_set) {
+                fprintf(stderr, "Error: -m or --model specified multiple times.\n");
+                return 1;
+            }
+            model_path = argv[++iarg].c_str();
+            model_path_set = true;
+        }
+        else if (arg == "--no-bos") {
+            no_bos = true;
+        }
+        else if (arg == "-p" || arg == "--prompt") {
+            if (prompt_set) {
+                fprintf(stderr, "Error: -p or --prompt specified multiple times.\n");
+                return 1;
+            }
+            prompt_arg = argv[++iarg].c_str();
+            prompt_set = true;
+        }
+        else if (arg == "-f" || arg == "--file") {
+            if (prompt_path_set) {
+                fprintf(stderr, "Error: -f or --file specified multiple times.\n");
+                return 1;
+            }
+            prompt_path = argv[++iarg].c_str();
+            prompt_path_set = true;
+        }
+        else if (arg == "--stdin") {
+            stdin_set = true;
+        }
+        else if (arg == "--log-disable") {
+            disable_logging = true;
+        }
+        else {
+            fprintf(stderr, "Error: unknown option '%s'\n", argv[iarg].c_str());
+            return 1;
+        }
+    }
+
+    //////
+    // Sanity check the command line arguments.
+    //////
+
+    // Check that we have the required stuff set.
+    if (model_path_set && model_path == NULL) {
+        fprintf(stderr, "Error: --model requires an argument.\n");
+        return 1;
+    }
+    if (!model_path_set) {
+        fprintf(stderr, "Error: must specify --model.\n");
+        return 1;
+    }
+    if (prompt_path_set && prompt_path == NULL) {
+        fprintf(stderr, "Error: --file requires an argument.\n");
+        return 1;
+    }
+    if (prompt_set && prompt_arg == NULL) {
+        fprintf(stderr, "Error: --prompt requires an argument.\n");
+        return 1;
+    }
+    const int prompts_set = !!(prompt_path_set) + !!(prompt_set) + !!(stdin_set);
+    if (prompts_set > 1) {
+        fprintf(stderr, "Error: --stdin, --file and --prompt are mutually exclusive.\n");
+        return 1;
+    }
+    // Must have some prompt.
+    if (prompts_set == 0) {
+        fprintf(stderr, "Error: must specify one of: --stdin, --file or --prompt.\n");
+        return 1;
+    }
+
+    GGML_ASSERT(model_path);
+    GGML_ASSERT(prompt_path || prompt_arg || stdin_set);
+
+    //////
+    // Figure out where will the prompt come from.
+    //////
+
+    std::string prompt;
+    if (prompt_path_set) {
+        bool success = false;
+        prompt = read_prompt_from_file(prompt_path, success);
+        if (!success) {
+            return 1;
+        }
+    } else if (prompt_set) {
+        prompt = prompt_arg;
+    } else {
+        GGML_ASSERT(stdin_set);
+        // we read stdin *after* loading model (early exit if model cannot
+        // be loaded, which can be a nicer user experience)
+    }
+
+    //////
+    // Start actually doing the tokenizing stuff.
+    //////
+
+#ifdef LOG_DISABLE_LOGS
+    disable_logging = true;
+#endif
+
+    if (disable_logging) {
+        llama_log_set(llama_log_callback_null, NULL);
+    }
 
     llama_backend_init();
 
     llama_model_params model_params = llama_model_default_params();
     model_params.vocab_only = true;
     llama_model * model = llama_load_model_from_file(model_path, model_params);
+    if (!model) {
+        fprintf(stderr, "Error: could not load model from file '%s'.\n", model_path);
+        return 1;
+    }
 
     llama_context_params ctx_params = llama_context_default_params();
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+    if (!ctx) {
+        fprintf(stderr, "Error: could not create context.\n");
+        return 1;
+    }
+
+    // read entire prompt from stdin?
+    if (stdin_set) {
+        GGML_ASSERT(!prompt_path_set && !prompt_set);
+
+        std::stringstream stdin_buffer;
+        stdin_buffer << std::cin.rdbuf();
+        if (std::cin.fail()) {
+            fprintf(stderr, "Error: could not read the entire standard input.\n");
+            return 1;
+        }
+
+        prompt = stdin_buffer.str();
+    }
+
+    const bool model_wants_add_bos = llama_should_add_bos_token(model);
+    const bool add_bos = model_wants_add_bos && !no_bos;
 
     std::vector<llama_token> tokens;
+    tokens = ::llama_tokenize(model, prompt, add_bos, true);
 
-    tokens = ::llama_tokenize(model, prompt, true, true);
+    if (printing_ids) {
+        printf("[");
+    }
 
     for (int i = 0; i < (int) tokens.size(); i++) {
         if (printing_ids) {
-            printf("%d\n", tokens[i]);
+            if (i > 0) {
+                printf(", ");
+            }
+            printf("%d", tokens[i]);
         } else {
-            printf("%6d -> '%s'\n", tokens[i], llama_token_to_piece(ctx, tokens[i]).c_str());
+            bool invalid_utf8 = false;
+            printf("%6d -> '", tokens[i]);
+            write_utf8_cstr_to_stdout(llama_token_to_piece(ctx, tokens[i]).c_str(), invalid_utf8);
+            if (invalid_utf8) {
+                printf("' (utf-8 decode failure)\n");
+            } else {
+                printf("'\n");
+            }
         }
     }
 
+    if (printing_ids) {
+        printf("]\n");
+    }
+
+    // silence valgrind
+    llama_free(ctx);
+    llama_free_model(model);
+
     return 0;
 }

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

@@ -301,8 +301,8 @@ static struct ggml_tensor * llama_build_train_graphs(
         // not capturing these, to silcence warnings
         const int rope_mode = 0;
 
-        return ggml_rope_custom(
-            ctx, t, KQ_pos, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
+        return ggml_rope_ext(
+            ctx, t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
         );
     };
 
@@ -341,7 +341,8 @@ static struct ggml_tensor * llama_build_train_graphs(
         struct ggml_tensor * t15 = ggml_permute      (ctx, t12, 0, 3, 1, 2);                        set_name(t15, "t15");     assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch);
         struct ggml_tensor * t16;
         if (enable_flash_attn) {
-            t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                        set_name(t16, "t16");     assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
+            GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported");
+            //t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                        set_name(t16, "t16");     assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
         } else {
             struct ggml_tensor * t16_0 = ggml_mul_mat              (ctx, t14, t13);                 set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
             struct ggml_tensor * t16_1 = ggml_scale_inplace        (ctx, t16_0, kv_scale);          set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);

flake.lock

@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1714641030,
-        "narHash": "sha256-yzcRNDoyVP7+SCNX0wmuDju1NUCt8Dz9+lyUXEI0dbI=",
+        "lastModified": 1715865404,
+        "narHash": "sha256-/GJvTdTpuDjNn84j82cU6bXztE0MSkdnTWClUCRub78=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "e5d10a24b66c3ea8f150e47dfdb0416ab7c3390e",
+        "rev": "8dc45382d5206bd292f9c2768b8058a8fd8311d9",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1714635257,
-        "narHash": "sha256-4cPymbty65RvF1DWQfc+Bc8B233A1BWxJnNULJKQ1EY=",
+        "lastModified": 1716509168,
+        "narHash": "sha256-4zSIhSRRIoEBwjbPm3YiGtbd8HDWzFxJjw5DYSDy1n8=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "63c3a29ca82437c87573e4c6919b09a24ea61b0f",
+        "rev": "bfb7a882678e518398ce9a31a881538679f6f092",
         "type": "github"
       },
       "original": {

ggml-common.h

@@ -65,13 +65,8 @@ typedef sycl::half2 ggml_half2;
 // QK = number of values after dequantization
 // QK_K = super-block size
 
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
 #define QK_K 256
 #define K_SCALE_SIZE 12
-#endif // GGML_QKK_64
 
 #if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
 // QR = QK / number of values before dequantization
@@ -131,13 +126,8 @@ typedef sycl::half2 ggml_half2;
 #define QI4_NL (QK4_NL / (4*QR4_NL))
 #define QR4_NL 2
 
-#if QK_K == 64
-#define QI4_XS QI4_NL
-#define QR4_XS QR4_NL
-#else
 #define QI4_XS (QK_K / (4*QR4_XS))
 #define QR4_XS 8
-#endif
 
 #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
 
@@ -228,15 +218,6 @@ static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wro
 // weight is represented as x = a * q
 // 16 blocks of 16 elements each
 // Effectively 3.4375 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8]; // quants - high bit
-    uint8_t qs[QK_K/4];    // quants - low 2 bits
-    uint8_t scales[2];
-    ggml_half d;           // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
 typedef struct {
     uint8_t hmask[QK_K/8]; // quants - high bit
     uint8_t qs[QK_K/4];    // quants - low 2 bits
@@ -244,20 +225,11 @@ typedef struct {
     ggml_half d;           // super-block scale
 } block_q3_K;
 static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif
 
 // 4-bit quantization
 // 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 4.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d[2];     // super-block scales/mins
-    uint8_t scales[2];  // 4-bit block scales/mins
-    uint8_t qs[QK_K/2]; // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
 typedef struct {
     union {
         struct {
@@ -270,21 +242,11 @@ typedef struct {
     uint8_t qs[QK_K/2];           // 4--bit quants
 } block_q4_K;
 static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif
 
 // 5-bit quantization
 // 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 5.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d;             // super-block scale
-    int8_t  scales[QK_K/16]; // 8-bit block scales
-    uint8_t qh[QK_K/8];      // quants, high bit
-    uint8_t qs[QK_K/2];      // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
 typedef struct {
     union {
         struct {
@@ -298,7 +260,6 @@ typedef struct {
     uint8_t qs[QK_K/2];           // quants, low 4 bits
 } block_q5_K;
 static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
 
 // 6-bit quantization
 // weight is represented as x = a * q
@@ -356,11 +317,7 @@ typedef struct {
 static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
 
 // 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
 #define IQ3S_N_SCALE QK_K/64
-#endif
 typedef struct {
     ggml_half d;
     uint8_t qs[QK_K/4];
@@ -381,16 +338,9 @@ static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wron
 typedef struct {
     uint8_t  qs[QK_K/8];      // grid index, low 8 bits
     uint8_t  qh[QK_K/16];     // grid index, high 3 bits + grid shift bit (for two groups of 8)
-#if QK_K == 64
-    ggml_half d;
-#endif
     uint8_t  scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
 } block_iq1_m;
-#if QK_K == 64
-static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding");
-#else
 static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
-#endif
 
 // Used by IQ1_M quants
 typedef union {
@@ -406,9 +356,6 @@ typedef struct {
 } block_iq4_nl;
 static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
 
-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-#else
 typedef struct {
     ggml_half d;
     uint16_t scales_h;
@@ -416,7 +363,6 @@ typedef struct {
     uint8_t  qs[QK_K/2];
 } block_iq4_xs;
 static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
-#endif
 
 #endif // GGML_COMMON_DECL
 #endif // GGML_COMMON_DECL