naive whitespace detection

Add instructions to get public key on windows

.github/workflows/test.yaml

@@ -0,0 +1,162 @@
+name: test
+
+on:
+  pull_request:
+
+jobs:
+  generate:
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-latest]
+        arch: [amd64, arm64]
+        exclude:
+          - os: ubuntu-latest
+            arch: arm64
+          - os: windows-latest
+            arch: arm64
+    runs-on: ${{ matrix.os }}
+    env:
+      GOARCH: ${{ matrix.arch }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v5
+        with:
+          go-version: '1.21'
+          cache: true
+      - run: go get ./...
+      - run: go generate -x ./...
+      - uses: actions/upload-artifact@v4
+        with:
+          name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
+          path: llm/llama.cpp/build/**/lib/*
+  generate-cuda:
+    strategy:
+      matrix:
+        cuda-version:
+          - '11.8.0'
+    runs-on: linux
+    container: nvidia/cuda:${{ matrix.cuda-version }}-devel-ubuntu20.04
+    steps:
+      - run: |
+          apt-get update && apt-get install -y git build-essential curl
+          curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.28.1/cmake-3.28.1-linux-x86_64.tar.gz \
+            | tar -zx -C /usr --strip-components 1
+        env:
+          DEBIAN_FRONTEND: noninteractive
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v4
+        with:
+          go-version: '1.21'
+          cache: true
+      - run: go get ./...
+      - run: |
+          git config --global --add safe.directory /__w/ollama/ollama
+          go generate -x ./...
+        env:
+          OLLAMA_SKIP_CPU_GENERATE: '1'
+      - uses: actions/upload-artifact@v4
+        with:
+          name: cuda-${{ matrix.cuda-version }}-libraries
+          path: llm/llama.cpp/build/**/lib/*
+  generate-rocm:
+    strategy:
+      matrix:
+        rocm-version:
+          - '5.7.1'
+          - '6.0'
+    runs-on: linux
+    container: rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}
+    steps:
+      - run: |
+          apt-get update && apt-get install -y git build-essential curl rocm-libs
+          curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.28.1/cmake-3.28.1-linux-x86_64.tar.gz \
+            | tar -zx -C /usr --strip-components 1
+        env:
+          DEBIAN_FRONTEND: noninteractive
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v4
+        with:
+          go-version: '1.21'
+          cache: true
+      - run: go get ./...
+      - run: |
+          git config --global --add safe.directory /__w/ollama/ollama
+          go generate -x ./...
+        env:
+          OLLAMA_SKIP_CPU_GENERATE: '1'
+      - uses: actions/upload-artifact@v4
+        with:
+          name: rocm-${{ matrix.rocm-version }}-libraries
+          path: llm/llama.cpp/build/**/lib/*
+  lint:
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-latest]
+        arch: [amd64, arm64]
+        exclude:
+          - os: ubuntu-latest
+            arch: arm64
+          - os: windows-latest
+            arch: arm64
+          - os: macos-latest
+            arch: amd64
+    runs-on: ${{ matrix.os }}
+    env:
+      GOARCH: ${{ matrix.arch }}
+      CGO_ENABLED: "1"
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - uses: actions/setup-go@v5
+        with:
+          go-version: '1.21'
+          cache: false
+      - run: |
+          mkdir -p llm/llama.cpp/build/linux/${{ matrix.arch }}/stub/lib/
+          touch llm/llama.cpp/build/linux/${{ matrix.arch }}/stub/lib/stub.so
+        if: ${{ startsWith(matrix.os, 'ubuntu-') }}
+      - run: |
+          mkdir -p llm/llama.cpp/build/darwin/${{ matrix.arch }}/stub/lib/
+          touch llm/llama.cpp/build/darwin/${{ matrix.arch }}/stub/lib/stub.dylib
+          touch llm/llama.cpp/ggml-metal.metal
+        if: ${{ startsWith(matrix.os, 'macos-') }}
+      - run: |
+          mkdir -p llm/llama.cpp/build/windows/${{ matrix.arch }}/stub/lib/
+          touch llm/llama.cpp/build/windows/${{ matrix.arch }}/stub/lib/stub.dll
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+      - uses: golangci/golangci-lint-action@v3
+  test:
+    needs: generate
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-latest]
+        arch: [amd64]
+        exclude:
+          - os: ubuntu-latest
+            arch: arm64
+          - os: windows-latest
+            arch: arm64
+    runs-on: ${{ matrix.os }}
+    env:
+      GOARCH: ${{ matrix.arch }}
+      CGO_ENABLED: "1"
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - uses: actions/setup-go@v5
+        with:
+          go-version: '1.21'
+          cache: true
+      - run: go get
+      - uses: actions/download-artifact@v4
+        with:
+          name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
+          path: llm/llama.cpp/build
+      - run: go build
+      - run: go test -v ./...
+      - uses: actions/upload-artifact@v4
+        with:
+          name: ${{ matrix.os }}-binaries
+          path: ollama

.gitignore

@@ -9,4 +9,5 @@ ggml-metal.metal
 .cache
 *.exe
 .idea
-test_data
+test_data
+*.crt

.golangci.yaml

@@ -0,0 +1,27 @@
+run:
+  timeout: 5m
+linters:
+  enable:
+    - asasalint
+    - bidichk
+    - bodyclose
+    - containedctx
+    - contextcheck
+    - exportloopref
+    - gocheckcompilerdirectives
+    # FIXME: for some reason this errors on windows
+    # - gofmt
+    # - goimports
+    - misspell
+    - nilerr
+    - unused
+linters-settings:
+  errcheck:
+    # exclude the following functions since we don't generally
+    # need to be concerned with the returned errors
+    exclude-functions:
+      - encoding/binary.Read
+      - (*os.File).Seek
+      - (*bufio.Writer).WriteString
+      - (*github.com/spf13/pflag.FlagSet).Set
+      - (*github.com/jmorganca/ollama/llm.readSeekOffset).Seek

Dockerfile

@@ -1,27 +1,135 @@
-FROM nvidia/cuda:11.8.0-devel-ubuntu22.04
+ARG GOLANG_VERSION=1.21.3
+ARG CMAKE_VERSION=3.22.1
+ARG CUDA_VERSION=11.3.1
 
-ARG TARGETARCH
-ARG GOFLAGS="'-ldflags=-w -s'"
+# Copy the minimal context we need to run the generate scripts
+FROM scratch AS llm-code
+COPY .git .git
+COPY .gitmodules .gitmodules
+COPY llm llm
 
+FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION-devel-centos7 AS cuda-build-amd64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
+WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+ARG CGO_CFLAGS
+RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+
+FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION-devel-rockylinux8 AS cuda-build-arm64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
+COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
+WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+ARG CGO_CFLAGS
+RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+
+FROM --platform=linux/amd64 rocm/dev-centos-7:5.7.1-complete AS rocm-5-build-amd64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV LIBRARY_PATH /opt/amdgpu/lib64
+COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
+WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+ARG CGO_CFLAGS
+ARG AMDGPU_TARGETS
+RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+
+FROM --platform=linux/amd64 rocm/dev-centos-7:6.0-complete AS rocm-6-build-amd64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV LIBRARY_PATH /opt/amdgpu/lib64
+COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
+WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+ARG CGO_CFLAGS
+ARG AMDGPU_TARGETS
+RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+
+FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
+ARG CMAKE_VERSION
+ARG GOLANG_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
+ARG OLLAMA_CUSTOM_CPU_DEFS
+ARG CGO_CFLAGS
+WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+
+FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
+RUN OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
+FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx-build-amd64
+RUN OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh
+FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
+RUN OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh
+
+FROM --platform=linux/arm64 centos:7 AS cpu-build-arm64
+ARG CMAKE_VERSION
+ARG GOLANG_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
+WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+# Note, we only build the "base" CPU variant on arm since avx/avx2 are x86 features
+ARG OLLAMA_CUSTOM_CPU_DEFS
+ARG CGO_CFLAGS
+RUN OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
+
+# Intermediate stage used for ./scripts/build_linux.sh
+FROM --platform=linux/amd64 cpu-build-amd64 AS build-amd64
+ENV CGO_ENABLED 1
 WORKDIR /go/src/github.com/jmorganca/ollama
-RUN apt-get update && apt-get install -y git build-essential cmake
-ADD https://dl.google.com/go/go1.21.3.linux-$TARGETARCH.tar.gz /tmp/go1.21.3.tar.gz
-RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.3.tar.gz
-
 COPY . .
-ENV GOARCH=$TARGETARCH
-ENV GOFLAGS=$GOFLAGS
-RUN /usr/local/go/bin/go generate ./... \
-    && /usr/local/go/bin/go build .
+COPY --from=cpu_avx-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
+COPY --from=cpu_avx2-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
+COPY --from=cuda-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
+COPY --from=rocm-5-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
+COPY --from=rocm-6-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
+ARG GOFLAGS
+ARG CGO_CFLAGS
+RUN go build .
 
-FROM ubuntu:22.04
+# Intermediate stage used for ./scripts/build_linux.sh
+FROM --platform=linux/arm64 cpu-build-arm64 AS build-arm64
+ENV CGO_ENABLED 1
+ARG GOLANG_VERSION
+WORKDIR /go/src/github.com/jmorganca/ollama
+COPY . .
+COPY --from=cuda-build-arm64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
+ARG GOFLAGS
+ARG CGO_CFLAGS
+RUN go build .
+
+# Runtime stages
+FROM --platform=linux/amd64 ubuntu:22.04 as runtime-amd64
 RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=0 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+COPY --from=build-amd64 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+FROM --platform=linux/arm64 ubuntu:22.04 as runtime-arm64
+RUN apt-get update && apt-get install -y ca-certificates
+COPY --from=build-arm64 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+
+# Radeon images are much larger so we keep it distinct from the CPU/CUDA image
+FROM --platform=linux/amd64 rocm/dev-centos-7:5.7.1-complete as runtime-rocm
+RUN update-pciids
+COPY --from=build-amd64 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
 EXPOSE 11434
 ENV OLLAMA_HOST 0.0.0.0
 
-# set some environment variable for better NVIDIA compatibility
-ENV PATH=/usr/local/nvidia/bin:/usr/local/cuda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
+ENTRYPOINT ["/bin/ollama"]
+CMD ["serve"]
+
+FROM runtime-$TARGETARCH
+EXPOSE 11434
+ENV OLLAMA_HOST 0.0.0.0
+ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
 ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
 ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
 

Dockerfile.build

@@ -1,74 +0,0 @@
-# Ubuntu 20.04 amd64 dependencies
-FROM --platform=linux/amd64 ubuntu:20.04 AS base-amd64
-ARG CUDA_VERSION=11.3.1-1
-ARG CMAKE_VERSION=3.22.1
-# ROCm only supports amd64
-ARG ROCM_VERSION=6.0
-ARG CLBLAST_VER=1.6.1
-
-# Note: https://rocm.docs.amd.com/en/latest/release/user_kernel_space_compat_matrix.html
-RUN apt-get update && \
-    apt-get install -y wget gnupg && \
-    wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-ubuntu2004.pin && \
-    mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600 && \
-    apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/3bf863cc.pub && \
-    echo "deb [by-hash=no] https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/ /" > /etc/apt/sources.list.d/cuda.list && \
-    wget "https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-x86_64.sh" -O /tmp/cmake-installer.sh && \
-    chmod +x /tmp/cmake-installer.sh && /tmp/cmake-installer.sh --skip-license --prefix=/usr && \
-    mkdir --parents --mode=0755 /etc/apt/keyrings && \
-    wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor > /etc/apt/keyrings/rocm.gpg && \
-    echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${ROCM_VERSION} focal main" > /etc/apt/sources.list.d/rocm.list && \
-    echo "Package: *" > /etc/apt/preferences.d/rocm-pin-600 && \
-    echo "Pin: release o=repo.radeon.com" >> /etc/apt/preferences.d/rocm-pin-600 && \
-    echo "Pin-Priority: 600" >> /etc/apt/preferences.d/rocm-pin-600 && \
-    apt-get update && \
-    DEBIAN_FRONTEND=noninteractive apt-get -y install cuda=${CUDA_VERSION} rocm-hip-libraries rocm-device-libs rocm-libs rocm-ocl-icd rocm-hip-sdk rocm-hip-libraries rocm-cmake rocm-clang-ocl rocm-dev
-
-# CLBlast
-RUN wget -qO- https://github.com/CNugteren/CLBlast/archive/refs/tags/${CLBLAST_VER}.tar.gz | tar zxv -C /tmp/ && \
-    cd /tmp/CLBlast-${CLBLAST_VER} && mkdir build && cd build && cmake .. && make && make install
-
-ENV ROCM_PATH=/opt/rocm
-
-# Ubuntu 22.04 arm64 dependencies
-FROM --platform=linux/arm64 ubuntu:20.04 AS base-arm64
-ARG CUDA_VERSION=11.3.1-1
-ARG CMAKE_VERSION=3.27.6
-RUN apt-get update && \
-    apt-get install -y wget gnupg && \
-    wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/sbsa/cuda-ubuntu2004.pin && \
-    mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600 && \
-    apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/sbsa//3bf863cc.pub && \
-    echo "deb [by-hash=no] https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/sbsa/ /" > /etc/apt/sources.list.d/cuda.list && \
-    wget "https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-aarch64.sh" -O /tmp/cmake-installer.sh && \
-    chmod +x /tmp/cmake-installer.sh && /tmp/cmake-installer.sh --skip-license --prefix=/usr && \
-    apt-get update && \
-    apt-cache madison cuda && \
-    DEBIAN_FRONTEND=noninteractive apt-get -y install cuda=${CUDA_VERSION} 
-
-FROM base-${TARGETARCH}
-ARG TARGETARCH
-ARG GOFLAGS="'-ldflags -w -s'"
-ARG CGO_CFLAGS
-ARG GOLANG_VERSION=1.21.3
-
-# Common toolchain
-RUN apt-get update && \
-    DEBIAN_FRONTEND=noninteractive apt-get install -y gcc-10 g++-10 cpp-10 git ocl-icd-opencl-dev && \
-    update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 --slave /usr/bin/g++ g++ /usr/bin/g++-10 --slave /usr/bin/gcov gcov /usr/bin/gcov-10
-
-# install go
-ADD https://dl.google.com/go/go${GOLANG_VERSION}.linux-$TARGETARCH.tar.gz /tmp/go${GOLANG_VERSION}.tar.gz
-RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go${GOLANG_VERSION}.tar.gz
-
-# build the final binary
-WORKDIR /go/src/github.com/jmorganca/ollama
-COPY . .
-
-ENV GOOS=linux
-ENV GOARCH=$TARGETARCH
-ENV GOFLAGS=$GOFLAGS
-ENV CGO_CFLAGS=${CGO_CFLAGS}
-
-RUN /usr/local/go/bin/go generate ./... && \
-    /usr/local/go/bin/go build .

README.md

@@ -1,8 +1,5 @@
 <div align="center">
-  <picture>
-    <source media="(prefers-color-scheme: dark)" height="200px" srcset="https://github.com/jmorganca/ollama/assets/3325447/56ea1849-1284-4645-8970-956de6e51c3c">
-    <img alt="logo" height="200px" src="https://github.com/jmorganca/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
-  </picture>
+  <img alt="ollama" height="200px" src="https://github.com/jmorganca/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
 </div>
 
 # Ollama
@@ -13,16 +10,16 @@ Get up and running with large language models locally.
 
 ### macOS
 
-[Download](https://ollama.ai/download/Ollama-darwin.zip)
+[Download](https://ollama.com/download/Ollama-darwin.zip)
 
-### Windows
+### Windows preview
 
-Coming soon! For now, you can install Ollama on Windows via WSL2.
+[Download](https://ollama.com/download/OllamaSetup.exe)
 
-### Linux & WSL2
+### Linux
 
 ```
-curl https://ollama.ai/install.sh | sh
+curl -fsSL https://ollama.com/install.sh | sh
 ```
 
 [Manual install instructions](https://github.com/jmorganca/ollama/blob/main/docs/linux.md)
@@ -31,9 +28,14 @@ curl https://ollama.ai/install.sh | sh
 
 The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `ollama/ollama` is available on Docker Hub.
 
+### Libraries
+
+- [ollama-python](https://github.com/ollama/ollama-python)
+- [ollama-js](https://github.com/ollama/ollama-js)
+
 ## Quickstart
 
-To run and chat with [Llama 2](https://ollama.ai/library/llama2):
+To run and chat with [Llama 2](https://ollama.com/library/llama2):
 
 ```
 ollama run llama2
@@ -41,9 +43,9 @@ ollama run llama2
 
 ## Model library
 
-Ollama supports a list of open-source models available on [ollama.ai/library](https://ollama.ai/library 'ollama model library')
+Ollama supports a list of models available on [ollama.com/library](https://ollama.com/library 'ollama model library')
 
-Here are some example open-source models that can be downloaded:
+Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                       |
 | ------------------ | ---------- | ----- | ------------------------------ |
@@ -198,18 +200,21 @@ brew install cmake go
 ```
 
 Then generate dependencies:
+
 ```
 go generate ./...
 ```
+
 Then build the binary:
+
 ```
 go build .
 ```
 
 More detailed instructions can be found in the [developer guide](https://github.com/jmorganca/ollama/blob/main/docs/development.md)
 
-
 ### Running local builds
+
 Next, start the server:
 
 ```
@@ -251,19 +256,21 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ## Community Integrations
 
 ### Web & Desktop
+
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
 - [HTML UI](https://github.com/rtcfirefly/ollama-ui)
 - [Chatbot UI](https://github.com/ivanfioravanti/chatbot-ollama)
 - [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
 - [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)
-- [Web UI](https://github.com/ollama-webui/ollama-webui)
+- [Open WebUI](https://github.com/open-webui/open-webui)
 - [Ollamac](https://github.com/kevinhermawan/Ollamac)
 - [big-AGI](https://github.com/enricoros/big-agi/blob/main/docs/config-ollama.md)
 - [Cheshire Cat assistant framework](https://github.com/cheshire-cat-ai/core)
 - [Amica](https://github.com/semperai/amica)
 - [chatd](https://github.com/BruceMacD/chatd)
 - [Ollama-SwiftUI](https://github.com/kghandour/Ollama-SwiftUI)
-
+- [MindMac](https://mindmac.app)
+- [NextJS Web Interface for Ollama](https://github.com/jakobhoeg/nextjs-ollama-llm-ui)
 
 ### Terminal
 
@@ -272,10 +279,14 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Emacs client](https://github.com/zweifisch/ollama)
 - [gen.nvim](https://github.com/David-Kunz/gen.nvim)
 - [ollama.nvim](https://github.com/nomnivore/ollama.nvim)
+- [ollama-chat.nvim](https://github.com/gerazov/ollama-chat.nvim)
 - [ogpt.nvim](https://github.com/huynle/ogpt.nvim)
 - [gptel Emacs client](https://github.com/karthink/gptel)
 - [Oatmeal](https://github.com/dustinblackman/oatmeal)
 - [cmdh](https://github.com/pgibler/cmdh)
+- [tenere](https://github.com/pythops/tenere)
+- [llm-ollama](https://github.com/taketwo/llm-ollama) for [Datasette's LLM CLI](https://llm.datasette.io/en/stable/).
+- [ShellOracle](https://github.com/djcopley/ShellOracle)
 
 ### Database
 
@@ -284,14 +295,17 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Package managers
 
 - [Pacman](https://archlinux.org/packages/extra/x86_64/ollama/)
+- [Helm Chart](https://artifacthub.io/packages/helm/ollama-helm/ollama)
 
 ### Libraries
 
 - [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
 - [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
 - [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
+- [LangChain4j](https://github.com/langchain4j/langchain4j/tree/main/langchain4j-ollama)
 - [LiteLLM](https://github.com/BerriAI/litellm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
+- [Ollama for Ruby](https://github.com/gbaptista/ollama-ai)
 - [Ollama-rs for Rust](https://github.com/pepperoni21/ollama-rs)
 - [Ollama4j for Java](https://github.com/amithkoujalgi/ollama4j)
 - [ModelFusion Typescript Library](https://modelfusion.dev/integration/model-provider/ollama)
@@ -299,6 +313,10 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Ollama for Dart](https://github.com/breitburg/dart-ollama)
 - [Ollama for Laravel](https://github.com/cloudstudio/ollama-laravel)
 - [LangChainDart](https://github.com/davidmigloz/langchain_dart)
+- [Semantic Kernel - Python](https://github.com/microsoft/semantic-kernel/tree/main/python/semantic_kernel/connectors/ai/ollama)
+- [Haystack](https://github.com/deepset-ai/haystack-integrations/blob/main/integrations/ollama.md)
+- [Ollama for R - rollama](https://github.com/JBGruber/rollama)
+- [Ollama-ex for Elixir](https://github.com/lebrunel/ollama-ex)
 
 ### Mobile
 
@@ -319,3 +337,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Rivet plugin](https://github.com/abrenneke/rivet-plugin-ollama)
 - [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
 - [Obsidian BMO Chatbot plugin](https://github.com/longy2k/obsidian-bmo-chatbot)
+- [Open Interpreter](https://docs.openinterpreter.com/language-model-setup/local-models/ollama)
+- [twinny](https://github.com/rjmacarthy/twinny) (Copilot and Copilot chat alternative using Ollama)
+- [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and HuggingFace)

api/client.py

@@ -1,284 +0,0 @@
-import os
-import json
-import requests
-import os
-import hashlib
-import json
-from pathlib import Path
-
-BASE_URL = os.environ.get('OLLAMA_HOST', 'http://localhost:11434')
-
-# Generate a response for a given prompt with a provided model. This is a streaming endpoint, so will be a series of responses.
-# The final response object will include statistics and additional data from the request. Use the callback function to override
-# the default handler.
-def generate(model_name, prompt, system=None, template=None, format="", context=None, options=None, callback=None):
-    try:
-        url = f"{BASE_URL}/api/generate"
-        payload = {
-            "model": model_name, 
-            "prompt": prompt, 
-            "system": system, 
-            "template": template, 
-            "context": context, 
-            "options": options,
-            "format": format,
-        }
-        
-        # Remove keys with None values
-        payload = {k: v for k, v in payload.items() if v is not None}
-        
-        with requests.post(url, json=payload, stream=True) as response:
-            response.raise_for_status()
-            
-            # Creating a variable to hold the context history of the final chunk
-            final_context = None
-            
-            # Variable to hold concatenated response strings if no callback is provided
-            full_response = ""
-
-            # Iterating over the response line by line and displaying the details
-            for line in response.iter_lines():
-                if line:
-                    # Parsing each line (JSON chunk) and extracting the details
-                    chunk = json.loads(line)
-                    
-                    # If a callback function is provided, call it with the chunk
-                    if callback:
-                        callback(chunk)
-                    else:
-                        # If this is not the last chunk, add the "response" field value to full_response and print it
-                        if not chunk.get("done"):
-                            response_piece = chunk.get("response", "")
-                            full_response += response_piece
-                            print(response_piece, end="", flush=True)
-                    
-                    # Check if it's the last chunk (done is true)
-                    if chunk.get("done"):
-                        final_context = chunk.get("context")
-            
-            # Return the full response and the final context
-            return full_response, final_context
-    except requests.exceptions.RequestException as e:
-        print(f"An error occurred: {e}")
-        return None, None
-    
-
-# Create a blob file on the server if it doesn't exist.
-def create_blob(digest, file_path):
-    url = f"{BASE_URL}/api/blobs/{digest}"
-
-    # Check if the blob exists
-    response = requests.head(url)
-    if response.status_code != 404:
-        return  # Blob already exists, no need to upload
-    response.raise_for_status()
-
-    # Upload the blob
-    with open(file_path, 'rb') as file_data:
-        requests.post(url, data=file_data)
-
-
-# Create a model from a Modelfile. Use the callback function to override the default handler.
-def create(model_name, filename, callback=None):
-    try:
-        file_path = Path(filename).expanduser().resolve()
-        processed_lines = []
-
-        # Read and process the modelfile
-        with open(file_path, 'r') as f:
-            for line in f:            
-                # Skip empty or whitespace-only lines
-                if not line.strip():
-                    continue
-            
-                command, args = line.split(maxsplit=1)
-
-                if command.upper() in ["FROM", "ADAPTER"]:
-                    path = Path(args.strip()).expanduser()
-
-                    # Check if path is relative and resolve it
-                    if not path.is_absolute():
-                        path = (file_path.parent / path)
-
-                    # Skip if file does not exist for "model", this is handled by the server
-                    if not path.exists():
-                        processed_lines.append(line)
-                        continue
-
-                    # Calculate SHA-256 hash
-                    with open(path, 'rb') as bin_file:
-                        hash = hashlib.sha256()
-                        hash.update(bin_file.read())
-                        blob = f"sha256:{hash.hexdigest()}"
-                
-                    # Add the file to the remote server
-                    create_blob(blob, path)
-
-                    # Replace path with digest in the line
-                    line = f"{command} @{blob}\n"
-
-                processed_lines.append(line)
-
-        # Combine processed lines back into a single string
-        modelfile_content = '\n'.join(processed_lines)
-
-        url = f"{BASE_URL}/api/create"
-        payload = {"name": model_name, "modelfile": modelfile_content}
-
-        # Making a POST request with the stream parameter set to True to handle streaming responses
-        with requests.post(url, json=payload, stream=True) as response:
-            response.raise_for_status()
-            # Iterating over the response line by line and displaying the status
-            for line in response.iter_lines():
-                if line:
-                    chunk = json.loads(line)
-                    if callback:
-                        callback(chunk)
-                    else:
-                        print(f"Status: {chunk.get('status')}")
-
-    except Exception as e:
-        print(f"An error occurred: {e}")
-
-
-# Pull a model from a the model registry. Cancelled pulls are resumed from where they left off, and multiple
-# calls to will share the same download progress. Use the callback function to override the default handler.
-def pull(model_name, insecure=False, callback=None):
-    try:
-        url = f"{BASE_URL}/api/pull"
-        payload = {
-            "name": model_name,
-            "insecure": insecure
-        }
-
-        # Making a POST request with the stream parameter set to True to handle streaming responses
-        with requests.post(url, json=payload, stream=True) as response:
-            response.raise_for_status()
-
-            # Iterating over the response line by line and displaying the details
-            for line in response.iter_lines():
-                if line:
-                    # Parsing each line (JSON chunk) and extracting the details
-                    chunk = json.loads(line)
-
-                    # If a callback function is provided, call it with the chunk
-                    if callback:
-                        callback(chunk)
-                    else:
-                        # Print the status message directly to the console
-                        print(chunk.get('status', ''), end='', flush=True)
-                    
-                    # If there's layer data, you might also want to print that (adjust as necessary)
-                    if 'digest' in chunk:
-                        print(f" - Digest: {chunk['digest']}", end='', flush=True)
-                        print(f" - Total: {chunk['total']}", end='', flush=True)
-                        print(f" - Completed: {chunk['completed']}", end='\n', flush=True)
-                    else:
-                        print()
-    except requests.exceptions.RequestException as e:
-        print(f"An error occurred: {e}")
-
-# Push a model to the model registry. Use the callback function to override the default handler.
-def push(model_name, insecure=False, callback=None):
-    try:
-        url = f"{BASE_URL}/api/push"
-        payload = {
-            "name": model_name,
-            "insecure": insecure
-        }
-
-        # Making a POST request with the stream parameter set to True to handle streaming responses
-        with requests.post(url, json=payload, stream=True) as response:
-            response.raise_for_status()
-
-            # Iterating over the response line by line and displaying the details
-            for line in response.iter_lines():
-                if line:
-                    # Parsing each line (JSON chunk) and extracting the details
-                    chunk = json.loads(line)
-
-                    # If a callback function is provided, call it with the chunk
-                    if callback:
-                        callback(chunk)
-                    else:
-                        # Print the status message directly to the console
-                        print(chunk.get('status', ''), end='', flush=True)
-                    
-                    # If there's layer data, you might also want to print that (adjust as necessary)
-                    if 'digest' in chunk:
-                        print(f" - Digest: {chunk['digest']}", end='', flush=True)
-                        print(f" - Total: {chunk['total']}", end='', flush=True)
-                        print(f" - Completed: {chunk['completed']}", end='\n', flush=True)
-                    else:
-                        print()
-    except requests.exceptions.RequestException as e:
-        print(f"An error occurred: {e}")
-
-# List models that are available locally.
-def list():
-    try:
-        response = requests.get(f"{BASE_URL}/api/tags")
-        response.raise_for_status()
-        data = response.json()
-        models = data.get('models', [])
-        return models
-
-    except requests.exceptions.RequestException as e:
-        print(f"An error occurred: {e}")
-        return None
-
-# Copy a model. Creates a model with another name from an existing model.
-def copy(source, destination):
-    try:
-        # Create the JSON payload
-        payload = {
-            "source": source,
-            "destination": destination
-        }
-        
-        response = requests.post(f"{BASE_URL}/api/copy", json=payload)
-        response.raise_for_status()
-        
-        # If the request was successful, return a message indicating that the copy was successful
-        return "Copy successful"
-
-    except requests.exceptions.RequestException as e:
-        print(f"An error occurred: {e}")
-        return None
-
-# Delete a model and its data.
-def delete(model_name):
-    try:
-        url = f"{BASE_URL}/api/delete"
-        payload = {"name": model_name}
-        response = requests.delete(url, json=payload)
-        response.raise_for_status()
-        return "Delete successful"
-    except requests.exceptions.RequestException as e:
-        print(f"An error occurred: {e}")
-        return None
-
-# Show info about a model.
-def show(model_name):
-    try:
-        url = f"{BASE_URL}/api/show"
-        payload = {"name": model_name}
-        response = requests.post(url, json=payload)
-        response.raise_for_status()
-        
-        # Parse the JSON response and return it
-        data = response.json()
-        return data
-    except requests.exceptions.RequestException as e:
-        print(f"An error occurred: {e}")
-        return None
-
-def heartbeat():
-    try:
-        url = f"{BASE_URL}/"
-        response = requests.head(url)
-        response.raise_for_status()
-        return "Ollama is running"
-    except requests.exceptions.RequestException as e:
-        print(f"An error occurred: {e}")
-        return "Ollama is not running"

api/types.go

@@ -34,24 +34,26 @@ func (e StatusError) Error() string {
 type ImageData []byte
 
 type GenerateRequest struct {
-	Model    string      `json:"model"`
-	Prompt   string      `json:"prompt"`
-	System   string      `json:"system"`
-	Template string      `json:"template"`
-	Context  []int       `json:"context,omitempty"`
-	Stream   *bool       `json:"stream,omitempty"`
-	Raw      bool        `json:"raw,omitempty"`
-	Format   string      `json:"format"`
-	Images   []ImageData `json:"images,omitempty"`
+	Model     string      `json:"model"`
+	Prompt    string      `json:"prompt"`
+	System    string      `json:"system"`
+	Template  string      `json:"template"`
+	Context   []int       `json:"context,omitempty"`
+	Stream    *bool       `json:"stream,omitempty"`
+	Raw       bool        `json:"raw,omitempty"`
+	Format    string      `json:"format"`
+	KeepAlive *Duration   `json:"keep_alive,omitempty"`
+	Images    []ImageData `json:"images,omitempty"`
 
 	Options map[string]interface{} `json:"options"`
 }
 
 type ChatRequest struct {
-	Model    string    `json:"model"`
-	Messages []Message `json:"messages"`
-	Stream   *bool     `json:"stream,omitempty"`
-	Format   string    `json:"format"`
+	Model     string    `json:"model"`
+	Messages  []Message `json:"messages"`
+	Stream    *bool     `json:"stream,omitempty"`
+	Format    string    `json:"format"`
+	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
 	Options map[string]interface{} `json:"options"`
 }
@@ -126,8 +128,9 @@ type Runner struct {
 }
 
 type EmbeddingRequest struct {
-	Model  string `json:"model"`
-	Prompt string `json:"prompt"`
+	Model     string    `json:"model"`
+	Prompt    string    `json:"prompt"`
+	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
 	Options map[string]interface{} `json:"options"`
 }
@@ -137,17 +140,30 @@ type EmbeddingResponse struct {
 }
 
 type CreateRequest struct {
-	Name      string `json:"name"`
+	Model     string `json:"model"`
 	Path      string `json:"path"`
 	Modelfile string `json:"modelfile"`
 	Stream    *bool  `json:"stream,omitempty"`
+
+	// Name is deprecated, see Model
+	Name string `json:"name"`
 }
 
 type DeleteRequest struct {
+	Model string `json:"model"`
+
+	// Name is deprecated, see Model
 	Name string `json:"name"`
 }
 
 type ShowRequest struct {
+	Model    string `json:"model"`
+	System   string `json:"system"`
+	Template string `json:"template"`
+
+	Options map[string]interface{} `json:"options"`
+
+	// Name is deprecated, see Model
 	Name string `json:"name"`
 }
 
@@ -158,6 +174,7 @@ type ShowResponse struct {
 	Template   string       `json:"template,omitempty"`
 	System     string       `json:"system,omitempty"`
 	Details    ModelDetails `json:"details,omitempty"`
+	Messages   []Message    `json:"messages,omitempty"`
 }
 
 type CopyRequest struct {
@@ -166,11 +183,14 @@ type CopyRequest struct {
 }
 
 type PullRequest struct {
-	Name     string `json:"name"`
+	Model    string `json:"model"`
 	Insecure bool   `json:"insecure,omitempty"`
 	Username string `json:"username"`
 	Password string `json:"password"`
 	Stream   *bool  `json:"stream,omitempty"`
+
+	// Name is deprecated, see Model
+	Name string `json:"name"`
 }
 
 type ProgressResponse struct {
@@ -181,11 +201,14 @@ type ProgressResponse struct {
 }
 
 type PushRequest struct {
-	Name     string `json:"name"`
+	Model    string `json:"model"`
 	Insecure bool   `json:"insecure,omitempty"`
 	Username string `json:"username"`
 	Password string `json:"password"`
 	Stream   *bool  `json:"stream,omitempty"`
+
+	// Name is deprecated, see Model
+	Name string `json:"name"`
 }
 
 type ListResponse struct {
@@ -194,6 +217,7 @@ type ListResponse struct {
 
 type ModelResponse struct {
 	Name       string       `json:"name"`
+	Model      string       `json:"model"`
 	ModifiedAt time.Time    `json:"modified_at"`
 	Size       int64        `json:"size"`
 	Digest     string       `json:"digest"`
@@ -216,6 +240,7 @@ type GenerateResponse struct {
 }
 
 type ModelDetails struct {
+	ParentModel       string   `json:"parent_model"`
 	Format            string   `json:"format"`
 	Family            string   `json:"family"`
 	Families          []string `json:"families"`
@@ -390,15 +415,18 @@ func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 	switch t := v.(type) {
 	case float64:
 		if t < 0 {
-			t = math.MaxFloat64
+			d.Duration = time.Duration(math.MaxInt64)
+		} else {
+			d.Duration = time.Duration(t * float64(time.Second))
 		}
-
-		d.Duration = time.Duration(t)
 	case string:
 		d.Duration, err = time.ParseDuration(t)
 		if err != nil {
 			return err
 		}
+		if d.Duration < 0 {
+			d.Duration = time.Duration(math.MaxInt64)
+		}
 	}
 
 	return nil

app/.gitignore

@@ -1,92 +1 @@
-# Logs
-logs
-*.log
-npm-debug.log*
-yarn-debug.log*
-yarn-error.log*
-lerna-debug.log*
-
-# Diagnostic reports (https://nodejs.org/api/report.html)
-report.[0-9]*.[0-9]*.[0-9]*.[0-9]*.json
-
-# Runtime data
-pids
-*.pid
-*.seed
-*.pid.lock
-.DS_Store
-
-# Directory for instrumented libs generated by jscoverage/JSCover
-lib-cov
-
-# Coverage directory used by tools like istanbul
-coverage
-*.lcov
-
-# nyc test coverage
-.nyc_output
-
-# node-waf configuration
-.lock-wscript
-
-# Compiled binary addons (https://nodejs.org/api/addons.html)
-build/Release
-
-# Dependency directories
-node_modules/
-jspm_packages/
-
-# TypeScript v1 declaration files
-typings/
-
-# TypeScript cache
-*.tsbuildinfo
-
-# Optional npm cache directory
-.npm
-
-# Optional eslint cache
-.eslintcache
-
-# Optional REPL history
-.node_repl_history
-
-# Output of 'npm pack'
-*.tgz
-
-# Yarn Integrity file
-.yarn-integrity
-
-# dotenv environment variables file
-.env
-.env.test
-
-# parcel-bundler cache (https://parceljs.org/)
-.cache
-
-# next.js build output
-.next
-
-# nuxt.js build output
-.nuxt
-
-# vuepress build output
-.vuepress/dist
-
-# Serverless directories
-.serverless/
-
-# FuseBox cache
-.fusebox/
-
-# DynamoDB Local files
-.dynamodb/
-
-# Webpack
-.webpack/
-
-# Vite
-.vite/
-
-# Electron-Forge
-out/
+ollama.syso

app/README.md

@@ -1,21 +1,22 @@
-# Desktop
+# Ollama App
 
-This app builds upon Ollama to provide a desktop experience for running models.
+## Linux
 
-## Developing
+TODO
 
-First, build the `ollama` binary:
+## MacOS
+
+TODO
+
+## Windows
+
+If you want to build the installer, youll need to install
+- https://jrsoftware.org/isinfo.php
+
+
+In the top directory of this repo, run the following powershell script
+to build the ollama CLI, ollama app, and ollama installer.
 
 ```
-cd ..
-go build .
+powershell -ExecutionPolicy Bypass -File .\scripts\build_windows.ps1
 ```
-
-Then run the desktop app with `npm start`:
-
-```
-cd app
-npm install
-npm start
-```
-

app/assets/assets.go

@@ -0,0 +1,17 @@
+package assets
+
+import (
+	"embed"
+	"io/fs"
+)
+
+//go:embed *.ico
+var icons embed.FS
+
+func ListIcons() ([]string, error) {
+	return fs.Glob(icons, "*")
+}
+
+func GetIcon(filename string) ([]byte, error) {
+	return icons.ReadFile(filename)
+}

app/lifecycle/getstarted_nonwindows.go

@@ -0,0 +1,9 @@
+//go:build !windows
+
+package lifecycle
+
+import "fmt"
+
+func GetStarted() error {
+	return fmt.Errorf("GetStarted not implemented")
+}

app/lifecycle/getstarted_windows.go

@@ -0,0 +1,44 @@
+package lifecycle
+
+import (
+	"fmt"
+	"log/slog"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"syscall"
+)
+
+func GetStarted() error {
+	const CREATE_NEW_CONSOLE = 0x00000010
+	var err error
+	bannerScript := filepath.Join(AppDir, "ollama_welcome.ps1")
+	args := []string{
+		// TODO once we're signed, the execution policy bypass should be removed
+		"powershell", "-noexit", "-ExecutionPolicy", "Bypass", "-nologo", "-file", bannerScript,
+	}
+	args[0], err = exec.LookPath(args[0])
+	if err != nil {
+		return err
+	}
+
+	// Make sure the script actually exists
+	_, err = os.Stat(bannerScript)
+	if err != nil {
+		return fmt.Errorf("getting started banner script error %s", err)
+	}
+
+	slog.Info(fmt.Sprintf("opening getting started terminal with %v", args))
+	attrs := &os.ProcAttr{
+		Files: []*os.File{os.Stdin, os.Stdout, os.Stderr},
+		Sys:   &syscall.SysProcAttr{CreationFlags: CREATE_NEW_CONSOLE, HideWindow: false},
+	}
+	proc, err := os.StartProcess(args[0], args, attrs)
+
+	if err != nil {
+		return fmt.Errorf("unable to start getting started shell %w", err)
+	}
+
+	slog.Debug(fmt.Sprintf("getting started terminal PID: %d", proc.Pid))
+	return proc.Release()
+}

app/lifecycle/lifecycle.go

@@ -0,0 +1,92 @@
+package lifecycle
+
+import (
+	"context"
+	"fmt"
+	"log"
+	"log/slog"
+	"os"
+	"os/signal"
+	"syscall"
+
+	"github.com/jmorganca/ollama/app/store"
+	"github.com/jmorganca/ollama/app/tray"
+)
+
+func Run() {
+	InitLogging()
+
+	ctx, cancel := context.WithCancel(context.Background())
+	var done chan int
+
+	t, err := tray.NewTray()
+	if err != nil {
+		log.Fatalf("Failed to start: %s", err)
+	}
+	callbacks := t.GetCallbacks()
+
+	signals := make(chan os.Signal, 1)
+	signal.Notify(signals, syscall.SIGINT, syscall.SIGTERM)
+
+	go func() {
+		slog.Debug("starting callback loop")
+		for {
+			select {
+			case <-callbacks.Quit:
+				slog.Debug("quit called")
+				t.Quit()
+			case <-signals:
+				slog.Debug("shutting down due to signal")
+				t.Quit()
+			case <-callbacks.Update:
+				err := DoUpgrade(cancel, done)
+				if err != nil {
+					slog.Warn(fmt.Sprintf("upgrade attempt failed: %s", err))
+				}
+			case <-callbacks.ShowLogs:
+				ShowLogs()
+			case <-callbacks.DoFirstUse:
+				err := GetStarted()
+				if err != nil {
+					slog.Warn(fmt.Sprintf("Failed to launch getting started shell: %s", err))
+				}
+			}
+		}
+	}()
+
+	// Are we first use?
+	if !store.GetFirstTimeRun() {
+		slog.Debug("First time run")
+		err = t.DisplayFirstUseNotification()
+		if err != nil {
+			slog.Debug(fmt.Sprintf("XXX failed to display first use notification %v", err))
+		}
+		store.SetFirstTimeRun(true)
+	} else {
+		slog.Debug("Not first time, skipping first run notification")
+	}
+
+	if IsServerRunning(ctx) {
+		slog.Info("Detected another instance of ollama running, exiting")
+		os.Exit(1)
+	} else {
+		done, err = SpawnServer(ctx, CLIName)
+		if err != nil {
+			// TODO - should we retry in a backoff loop?
+			// TODO - should we pop up a warning and maybe add a menu item to view application logs?
+			slog.Error(fmt.Sprintf("Failed to spawn ollama server %s", err))
+			done = make(chan int, 1)
+			done <- 1
+		}
+	}
+
+	StartBackgroundUpdaterChecker(ctx, t.UpdateAvailable)
+
+	t.Run()
+	cancel()
+	slog.Info("Waiting for ollama server to shutdown...")
+	if done != nil {
+		<-done
+	}
+	slog.Info("Ollama app exiting")
+}

app/lifecycle/logging.go

@@ -0,0 +1,46 @@
+package lifecycle
+
+import (
+	"fmt"
+	"log/slog"
+	"os"
+	"path/filepath"
+)
+
+func InitLogging() {
+	level := slog.LevelInfo
+
+	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
+		level = slog.LevelDebug
+	}
+
+	var logFile *os.File
+	var err error
+	// Detect if we're a GUI app on windows, and if not, send logs to console
+	if os.Stderr.Fd() != 0 {
+		// Console app detected
+		logFile = os.Stderr
+		// TODO - write one-line to the app.log file saying we're running in console mode to help avoid confusion
+	} else {
+		logFile, err = os.OpenFile(AppLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
+		if err != nil {
+			slog.Error(fmt.Sprintf("failed to create server log %v", err))
+			return
+		}
+	}
+	handler := slog.NewTextHandler(logFile, &slog.HandlerOptions{
+		Level:     level,
+		AddSource: true,
+		ReplaceAttr: func(_ []string, attr slog.Attr) slog.Attr {
+			if attr.Key == slog.SourceKey {
+				source := attr.Value.Any().(*slog.Source)
+				source.File = filepath.Base(source.File)
+			}
+			return attr
+		},
+	})
+
+	slog.SetDefault(slog.New(handler))
+
+	slog.Info("ollama app started")
+}

app/lifecycle/logging_nonwindows.go

@@ -0,0 +1,9 @@
+//go:build !windows
+
+package lifecycle
+
+import "log/slog"
+
+func ShowLogs() {
+	slog.Warn("ShowLogs not yet implemented")
+}

app/lifecycle/logging_windows.go

@@ -0,0 +1,19 @@
+package lifecycle
+
+import (
+	"fmt"
+	"log/slog"
+	"os/exec"
+	"syscall"
+)
+
+func ShowLogs() {
+	cmd_path := "c:\\Windows\\system32\\cmd.exe"
+	slog.Debug(fmt.Sprintf("viewing logs with start %s", AppDataDir))
+	cmd := exec.Command(cmd_path, "/c", "start", AppDataDir)
+	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: false, CreationFlags: 0x08000000}
+	err := cmd.Start()
+	if err != nil {
+		slog.Error(fmt.Sprintf("Failed to open log dir: %s", err))
+	}
+}

app/lifecycle/paths.go

@@ -0,0 +1,79 @@
+package lifecycle
+
+import (
+	"errors"
+	"fmt"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+)
+
+var (
+	AppName    = "ollama app"
+	CLIName    = "ollama"
+	AppDir     = "/opt/Ollama"
+	AppDataDir = "/opt/Ollama"
+	// TODO - should there be a distinct log dir?
+	UpdateStageDir = "/tmp"
+	AppLogFile     = "/tmp/ollama_app.log"
+	ServerLogFile  = "/tmp/ollama.log"
+	UpgradeLogFile = "/tmp/ollama_update.log"
+	Installer      = "OllamaSetup.exe"
+)
+
+func init() {
+	if runtime.GOOS == "windows" {
+		AppName += ".exe"
+		CLIName += ".exe"
+		// Logs, configs, downloads go to LOCALAPPDATA
+		localAppData := os.Getenv("LOCALAPPDATA")
+		AppDataDir = filepath.Join(localAppData, "Ollama")
+		UpdateStageDir = filepath.Join(AppDataDir, "updates")
+		AppLogFile = filepath.Join(AppDataDir, "app.log")
+		ServerLogFile = filepath.Join(AppDataDir, "server.log")
+		UpgradeLogFile = filepath.Join(AppDataDir, "upgrade.log")
+
+		// Executables are stored in APPDATA
+		AppDir = filepath.Join(localAppData, "Programs", "Ollama")
+
+		// Make sure we have PATH set correctly for any spawned children
+		paths := strings.Split(os.Getenv("PATH"), ";")
+		// Start with whatever we find in the PATH/LD_LIBRARY_PATH
+		found := false
+		for _, path := range paths {
+			d, err := filepath.Abs(path)
+			if err != nil {
+				continue
+			}
+			if strings.EqualFold(AppDir, d) {
+				found = true
+			}
+		}
+		if !found {
+			paths = append(paths, AppDir)
+
+			pathVal := strings.Join(paths, ";")
+			slog.Debug("setting PATH=" + pathVal)
+			err := os.Setenv("PATH", pathVal)
+			if err != nil {
+				slog.Error(fmt.Sprintf("failed to update PATH: %s", err))
+			}
+		}
+
+		// Make sure our logging dir exists
+		_, err := os.Stat(AppDataDir)
+		if errors.Is(err, os.ErrNotExist) {
+			if err := os.MkdirAll(AppDataDir, 0o755); err != nil {
+				slog.Error(fmt.Sprintf("create ollama dir %s: %v", AppDataDir, err))
+			}
+		}
+
+	} else if runtime.GOOS == "darwin" {
+		// TODO
+		AppName += ".app"
+		// } else if runtime.GOOS == "linux" {
+		// TODO
+	}
+}

app/lifecycle/server.go

@@ -0,0 +1,139 @@
+package lifecycle
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"time"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+func getCLIFullPath(command string) string {
+	cmdPath := ""
+	appExe, err := os.Executable()
+	if err == nil {
+		cmdPath = filepath.Join(filepath.Dir(appExe), command)
+		_, err := os.Stat(cmdPath)
+		if err == nil {
+			return cmdPath
+		}
+	}
+	cmdPath, err = exec.LookPath(command)
+	if err == nil {
+		_, err := os.Stat(cmdPath)
+		if err == nil {
+			return cmdPath
+		}
+	}
+	pwd, err := os.Getwd()
+	if err == nil {
+		cmdPath = filepath.Join(pwd, command)
+		_, err = os.Stat(cmdPath)
+		if err == nil {
+			return cmdPath
+		}
+	}
+
+	return command
+}
+
+func SpawnServer(ctx context.Context, command string) (chan int, error) {
+	done := make(chan int)
+
+	logDir := filepath.Dir(ServerLogFile)
+	_, err := os.Stat(logDir)
+	if errors.Is(err, os.ErrNotExist) {
+		if err := os.MkdirAll(logDir, 0o755); err != nil {
+			return done, fmt.Errorf("create ollama server log dir %s: %v", logDir, err)
+		}
+	}
+
+	cmd := getCmd(ctx, getCLIFullPath(command))
+	// send stdout and stderr to a file
+	stdout, err := cmd.StdoutPipe()
+	if err != nil {
+		return done, fmt.Errorf("failed to spawn server stdout pipe %s", err)
+	}
+	stderr, err := cmd.StderrPipe()
+	if err != nil {
+		return done, fmt.Errorf("failed to spawn server stderr pipe %s", err)
+	}
+	stdin, err := cmd.StdinPipe()
+	if err != nil {
+		return done, fmt.Errorf("failed to spawn server stdin pipe %s", err)
+	}
+
+	// TODO - rotation
+	logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
+	if err != nil {
+		return done, fmt.Errorf("failed to create server log %w", err)
+	}
+	go func() {
+		defer logFile.Close()
+		io.Copy(logFile, stdout) //nolint:errcheck
+	}()
+	go func() {
+		defer logFile.Close()
+		io.Copy(logFile, stderr) //nolint:errcheck
+	}()
+
+	// run the command and wait for it to finish
+	if err := cmd.Start(); err != nil {
+		return done, fmt.Errorf("failed to start server %w", err)
+	}
+	if cmd.Process != nil {
+		slog.Info(fmt.Sprintf("started ollama server with pid %d", cmd.Process.Pid))
+	}
+	slog.Info(fmt.Sprintf("ollama server logs %s", ServerLogFile))
+
+	go func() {
+		// Keep the server running unless we're shuttind down the app
+		crashCount := 0
+		for {
+			cmd.Wait() //nolint:errcheck
+			stdin.Close()
+			var code int
+			if cmd.ProcessState != nil {
+				code = cmd.ProcessState.ExitCode()
+			}
+
+			select {
+			case <-ctx.Done():
+				slog.Debug(fmt.Sprintf("server shutdown with exit code %d", code))
+				done <- code
+				return
+			default:
+				crashCount++
+				slog.Warn(fmt.Sprintf("server crash %d - exit code %d - respawning", crashCount, code))
+				time.Sleep(500 * time.Millisecond)
+				if err := cmd.Start(); err != nil {
+					slog.Error(fmt.Sprintf("failed to restart server %s", err))
+					// Keep trying, but back off if we keep failing
+					time.Sleep(time.Duration(crashCount) * time.Second)
+				}
+			}
+		}
+	}()
+	return done, nil
+}
+
+func IsServerRunning(ctx context.Context) bool {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		slog.Info("unable to connect to server")
+		return false
+	}
+	err = client.Heartbeat(ctx)
+	if err != nil {
+		slog.Debug(fmt.Sprintf("heartbeat from server: %s", err))
+		slog.Info("unable to connect to server")
+		return false
+	}
+	return true
+}

app/lifecycle/server_unix.go

@@ -0,0 +1,12 @@
+//go:build !windows
+
+package lifecycle
+
+import (
+	"context"
+	"os/exec"
+)
+
+func getCmd(ctx context.Context, cmd string) *exec.Cmd {
+	return exec.CommandContext(ctx, cmd, "serve")
+}

app/lifecycle/server_windows.go

@@ -0,0 +1,13 @@
+package lifecycle
+
+import (
+	"context"
+	"os/exec"
+	"syscall"
+)
+
+func getCmd(ctx context.Context, exePath string) *exec.Cmd {
+	cmd := exec.CommandContext(ctx, exePath, "serve")
+	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: true, CreationFlags: 0x08000000}
+	return cmd
+}

app/lifecycle/updater.go

@@ -0,0 +1,238 @@
+package lifecycle
+
+import (
+	"context"
+	"crypto/rand"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"log/slog"
+	"mime"
+	"net/http"
+	"net/url"
+	"os"
+	"path"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"time"
+
+	"github.com/jmorganca/ollama/auth"
+	"github.com/jmorganca/ollama/version"
+)
+
+var (
+	UpdateCheckURLBase  = "https://ollama.com/api/update"
+	UpdateDownloaded    = false
+	UpdateCheckInterval = 60 * 60 * time.Second
+)
+
+// TODO - maybe move up to the API package?
+type UpdateResponse struct {
+	UpdateURL     string `json:"url"`
+	UpdateVersion string `json:"version"`
+}
+
+func getClient(req *http.Request) http.Client {
+	proxyURL, err := http.ProxyFromEnvironment(req)
+	if err != nil {
+		slog.Warn(fmt.Sprintf("failed to handle proxy: %s", err))
+		return http.Client{}
+	}
+
+	return http.Client{
+		Transport: &http.Transport{
+			Proxy: http.ProxyURL(proxyURL),
+		},
+	}
+}
+
+func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) {
+	var updateResp UpdateResponse
+
+	requestURL, err := url.Parse(UpdateCheckURLBase)
+	if err != nil {
+		return false, updateResp
+	}
+
+	query := requestURL.Query()
+	query.Add("os", runtime.GOOS)
+	query.Add("arch", runtime.GOARCH)
+	query.Add("version", version.Version)
+	query.Add("ts", fmt.Sprintf("%d", time.Now().Unix()))
+
+	nonce, err := auth.NewNonce(rand.Reader, 16)
+	if err != nil {
+		return false, updateResp
+	}
+
+	query.Add("nonce", nonce)
+	requestURL.RawQuery = query.Encode()
+
+	data := []byte(fmt.Sprintf("%s,%s", http.MethodGet, requestURL.RequestURI()))
+	signature, err := auth.Sign(ctx, data)
+	if err != nil {
+		return false, updateResp
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, requestURL.String(), nil)
+	if err != nil {
+		slog.Warn(fmt.Sprintf("failed to check for update: %s", err))
+		return false, updateResp
+	}
+	req.Header.Set("Authorization", signature)
+	req.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))
+	client := getClient(req)
+
+	slog.Debug("checking for available update", "requestURL", requestURL)
+	resp, err := client.Do(req)
+	if err != nil {
+		slog.Warn(fmt.Sprintf("failed to check for update: %s", err))
+		return false, updateResp
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode == 204 {
+		slog.Debug("check update response 204 (current version is up to date)")
+		return false, updateResp
+	}
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		slog.Warn(fmt.Sprintf("failed to read body response: %s", err))
+	}
+	err = json.Unmarshal(body, &updateResp)
+	if err != nil {
+		slog.Warn(fmt.Sprintf("malformed response checking for update: %s", err))
+		return false, updateResp
+	}
+	// Extract the version string from the URL in the github release artifact path
+	updateResp.UpdateVersion = path.Base(path.Dir(updateResp.UpdateURL))
+
+	slog.Info("New update available at " + updateResp.UpdateURL)
+	return true, updateResp
+}
+
+func DownloadNewRelease(ctx context.Context, updateResp UpdateResponse) error {
+	// Do a head first to check etag info
+	req, err := http.NewRequestWithContext(ctx, http.MethodHead, updateResp.UpdateURL, nil)
+	if err != nil {
+		return err
+	}
+	client := getClient(req)
+	resp, err := client.Do(req)
+	if err != nil {
+		return fmt.Errorf("error checking update: %w", err)
+	}
+	if resp.StatusCode != 200 {
+		return fmt.Errorf("unexpected status attempting to download update %d", resp.StatusCode)
+	}
+	resp.Body.Close()
+	etag := strings.Trim(resp.Header.Get("etag"), "\"")
+	if etag == "" {
+		slog.Debug("no etag detected, falling back to filename based dedup")
+		etag = "_"
+	}
+	filename := Installer
+	_, params, err := mime.ParseMediaType(resp.Header.Get("content-disposition"))
+	if err == nil {
+		filename = params["filename"]
+	}
+
+	stageFilename := filepath.Join(UpdateStageDir, etag, filename)
+
+	// Check to see if we already have it downloaded
+	_, err = os.Stat(stageFilename)
+	if err == nil {
+		slog.Info("update already downloaded")
+		return nil
+	}
+
+	cleanupOldDownloads()
+
+	req.Method = http.MethodGet
+	resp, err = client.Do(req)
+	if err != nil {
+		return fmt.Errorf("error checking update: %w", err)
+	}
+	defer resp.Body.Close()
+	etag = strings.Trim(resp.Header.Get("etag"), "\"")
+	if etag == "" {
+		slog.Debug("no etag detected, falling back to filename based dedup") // TODO probably can get rid of this redundant log
+		etag = "_"
+	}
+
+	stageFilename = filepath.Join(UpdateStageDir, etag, filename)
+
+	_, err = os.Stat(filepath.Dir(stageFilename))
+	if errors.Is(err, os.ErrNotExist) {
+		if err := os.MkdirAll(filepath.Dir(stageFilename), 0o755); err != nil {
+			return fmt.Errorf("create ollama dir %s: %v", filepath.Dir(stageFilename), err)
+		}
+	}
+
+	payload, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return fmt.Errorf("failed to read body response: %w", err)
+	}
+	fp, err := os.OpenFile(stageFilename, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+	if err != nil {
+		return fmt.Errorf("write payload %s: %w", stageFilename, err)
+	}
+	defer fp.Close()
+	if n, err := fp.Write(payload); err != nil || n != len(payload) {
+		return fmt.Errorf("write payload %s: %d vs %d -- %w", stageFilename, n, len(payload), err)
+	}
+	slog.Info("new update downloaded " + stageFilename)
+
+	UpdateDownloaded = true
+	return nil
+}
+
+func cleanupOldDownloads() {
+	files, err := os.ReadDir(UpdateStageDir)
+	if err != nil && errors.Is(err, os.ErrNotExist) {
+		// Expected behavior on first run
+		return
+	} else if err != nil {
+		slog.Warn(fmt.Sprintf("failed to list stage dir: %s", err))
+		return
+	}
+	for _, file := range files {
+		fullname := filepath.Join(UpdateStageDir, file.Name())
+		slog.Debug("cleaning up old download: " + fullname)
+		err = os.RemoveAll(fullname)
+		if err != nil {
+			slog.Warn(fmt.Sprintf("failed to cleanup stale update download %s", err))
+		}
+	}
+}
+
+func StartBackgroundUpdaterChecker(ctx context.Context, cb func(string) error) {
+	go func() {
+		// Don't blast an update message immediately after startup
+		// time.Sleep(30 * time.Second)
+		time.Sleep(3 * time.Second)
+
+		for {
+			available, resp := IsNewReleaseAvailable(ctx)
+			if available {
+				err := DownloadNewRelease(ctx, resp)
+				if err != nil {
+					slog.Error(fmt.Sprintf("failed to download new release: %s", err))
+				}
+				err = cb(resp.UpdateVersion)
+				if err != nil {
+					slog.Warn(fmt.Sprintf("failed to register update available with tray: %s", err))
+				}
+			}
+			select {
+			case <-ctx.Done():
+				slog.Debug("stopping background update checker")
+				return
+			default:
+				time.Sleep(UpdateCheckInterval)
+			}
+		}
+	}()
+}

app/lifecycle/updater_nonwindows.go

@@ -0,0 +1,12 @@
+//go:build !windows
+
+package lifecycle
+
+import (
+	"context"
+	"fmt"
+)
+
+func DoUpgrade(cancel context.CancelFunc, done chan int) error {
+	return fmt.Errorf("DoUpgrade not yet implemented")
+}

app/lifecycle/updater_windows.go

@@ -0,0 +1,80 @@
+package lifecycle
+
+import (
+	"context"
+	"fmt"
+	"log/slog"
+	"os"
+	"os/exec"
+	"path/filepath"
+)
+
+func DoUpgrade(cancel context.CancelFunc, done chan int) error {
+	files, err := filepath.Glob(filepath.Join(UpdateStageDir, "*", "*.exe")) // TODO generalize for multiplatform
+	if err != nil {
+		return fmt.Errorf("failed to lookup downloads: %s", err)
+	}
+	if len(files) == 0 {
+		return fmt.Errorf("no update downloads found")
+	} else if len(files) > 1 {
+		// Shouldn't happen
+		slog.Warn(fmt.Sprintf("multiple downloads found, using first one %v", files))
+	}
+	installerExe := files[0]
+
+	slog.Info("starting upgrade with " + installerExe)
+	slog.Info("upgrade log file " + UpgradeLogFile)
+
+	// When running in debug mode, we'll be "verbose" and let the installer pop up and prompt
+	installArgs := []string{
+		"/CLOSEAPPLICATIONS",                    // Quit the tray app if it's still running
+		"/LOG=" + filepath.Base(UpgradeLogFile), // Only relative seems reliable, so set pwd
+		"/FORCECLOSEAPPLICATIONS",               // Force close the tray app - might be needed
+	}
+	// When we're not in debug mode, make the upgrade as quiet as possible (no GUI, no prompts)
+	// TODO - temporarily disable since we're pinning in debug mode for the preview
+	// if debug := os.Getenv("OLLAMA_DEBUG"); debug == "" {
+	installArgs = append(installArgs,
+		"/SP", // Skip the "This will install... Do you wish to continue" prompt
+		"/SUPPRESSMSGBOXES",
+		"/SILENT",
+		"/VERYSILENT",
+	)
+	// }
+
+	// Safeguard in case we have requests in flight that need to drain...
+	slog.Info("Waiting for server to shutdown")
+	cancel()
+	if done != nil {
+		<-done
+	} else {
+		// Shouldn't happen
+		slog.Warn("done chan was nil, not actually waiting")
+	}
+
+	slog.Debug(fmt.Sprintf("starting installer: %s %v", installerExe, installArgs))
+	os.Chdir(filepath.Dir(UpgradeLogFile)) //nolint:errcheck
+	cmd := exec.Command(installerExe, installArgs...)
+
+	if err := cmd.Start(); err != nil {
+		return fmt.Errorf("unable to start ollama app %w", err)
+	}
+
+	if cmd.Process != nil {
+		err = cmd.Process.Release()
+		if err != nil {
+			slog.Error(fmt.Sprintf("failed to release server process: %s", err))
+		}
+	} else {
+		// TODO - some details about why it didn't start, or is this a pedantic error case?
+		return fmt.Errorf("installer process did not start")
+	}
+
+	// TODO should we linger for a moment and check to make sure it's actually running by checking the pid?
+
+	slog.Info("Installer started in background, exiting")
+
+	os.Exit(0)
+	// Not reached
+	return nil
+}

app/main.go

@@ -0,0 +1,12 @@
+package main
+
+// Compile with the following to get rid of the cmd pop up on windows
+// go build -ldflags="-H windowsgui" .
+
+import (
+	"github.com/jmorganca/ollama/app/lifecycle"
+)
+
+func main() {
+	lifecycle.Run()
+}

app/ollama.iss

@@ -0,0 +1,153 @@
+; Inno Setup Installer for Ollama
+;
+; To build the installer use the build script invoked from the top of the source tree
+; 
+; powershell -ExecutionPolicy Bypass -File .\scripts\build_windows.ps
+
+
+#define MyAppName "Ollama"
+#if GetEnv("PKG_VERSION") != ""
+  #define MyAppVersion GetEnv("PKG_VERSION")
+#else
+  #define MyAppVersion "0.0.0"
+#endif
+#define MyAppPublisher "Ollama"
+#define MyAppURL "https://ollama.com/"
+#define MyAppExeName "ollama app.exe"
+#define MyIcon ".\assets\app.ico"
+
+[Setup]
+; NOTE: The value of AppId uniquely identifies this application. Do not use the same AppId value in installers for other applications.
+; (To generate a new GUID, click Tools | Generate GUID inside the IDE.)
+AppId={{44E83376-CE68-45EB-8FC1-393500EB558C}
+AppName={#MyAppName}
+AppVersion={#MyAppVersion}
+VersionInfoVersion={#MyAppVersion}
+;AppVerName={#MyAppName} {#MyAppVersion}
+AppPublisher={#MyAppPublisher}
+AppPublisherURL={#MyAppURL}
+AppSupportURL={#MyAppURL}
+AppUpdatesURL={#MyAppURL}
+ArchitecturesAllowed=x64
+ArchitecturesInstallIn64BitMode=x64
+DefaultDirName={localappdata}\Programs\{#MyAppName}
+DefaultGroupName={#MyAppName}
+DisableProgramGroupPage=yes
+PrivilegesRequired=lowest
+OutputBaseFilename="OllamaSetup"
+SetupIconFile={#MyIcon}
+UninstallDisplayIcon={uninstallexe}
+Compression=lzma2
+SolidCompression=no
+WizardStyle=modern
+ChangesEnvironment=yes
+OutputDir=..\dist\
+
+; Disable logging once everything's battle tested
+; Filename will be %TEMP%\Setup Log*.txt
+SetupLogging=yes
+CloseApplications=yes
+RestartApplications=no
+
+; Make sure they can at least download llama2 as a minimum
+ExtraDiskSpaceRequired=3826806784
+
+; https://jrsoftware.org/ishelp/index.php?topic=setup_wizardimagefile
+WizardSmallImageFile=.\assets\setup.bmp
+
+; TODO verifty actual min windows version...
+; OG Win 10
+MinVersion=10.0.10240
+
+; First release that supports WinRT UI Composition for win32 apps
+; MinVersion=10.0.17134
+; First release with XAML Islands - possible UI path forward
+; MinVersion=10.0.18362
+
+; quiet...
+DisableDirPage=yes
+DisableFinishedPage=yes
+DisableReadyMemo=yes
+DisableReadyPage=yes
+DisableStartupPrompt=yes
+DisableWelcomePage=yes
+
+; TODO - percentage can't be set less than 100, so how to make it shorter?
+; WizardSizePercent=100,80
+
+#if GetEnv("KEY_CONTAINER")
+SignTool=MySignTool
+SignedUninstaller=yes
+#endif
+
+SetupMutex=OllamaSetupMutex
+
+[Languages]
+Name: "english"; MessagesFile: "compiler:Default.isl"
+
+[LangOptions]
+DialogFontSize=12
+
+[Files]
+Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit
+Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
+Source: "..\dist\windeps\*.dll"; DestDir: "{app}"; Flags: ignoreversion 64bit
+Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
+Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
+
+[Icons]
+Name: "{group}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
+Name: "{userstartup}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
+Name: "{userprograms}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
+
+[Run]
+Filename: "{cmd}"; Parameters: "/C set PATH={app};%PATH% & ""{app}\{#MyAppExeName}"""; Flags: postinstall nowait runhidden
+
+[UninstallRun]
+; Filename: "{cmd}"; Parameters: "/C ""taskkill /im ''{#MyAppExeName}'' /f /t"; Flags: runhidden
+; Filename: "{cmd}"; Parameters: "/C ""taskkill /im ollama.exe /f /t"; Flags: runhidden
+Filename: "taskkill"; Parameters: "/im ""{#MyAppExeName}"" /f /t"; Flags: runhidden
+Filename: "taskkill"; Parameters: "/im ""ollama.exe"" /f /t"; Flags: runhidden
+; HACK!  need to give the server and app enough time to exit
+; TODO - convert this to a Pascal code script so it waits until they're no longer running, then completes
+Filename: "{cmd}"; Parameters: "/c timeout 5"; Flags: runhidden
+
+[UninstallDelete]
+Type: filesandordirs; Name: "{%TEMP}\ollama*"
+Type: filesandordirs; Name: "{%LOCALAPPDATA}\Ollama"
+Type: filesandordirs; Name: "{%LOCALAPPDATA}\Programs\Ollama"
+Type: filesandordirs; Name: "{%USERPROFILE}\.ollama"
+; NOTE: if the user has a custom OLLAMA_MODELS it will be preserved
+
+[Messages]
+WizardReady=Ollama Windows Preview
+ReadyLabel1=%nLet's get you up and running with your own large language models.
+SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or finish the other installer, then click OK to continue with this install, or Cancel to exit.
+
+
+;FinishedHeadingLabel=Run your first model
+;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama2
+;ClickFinish=%n
+
+[Registry]
+Root: HKCU; Subkey: "Environment"; \
+    ValueType: expandsz; ValueName: "Path"; ValueData: "{olddata};{app}"; \
+    Check: NeedsAddPath('{app}')
+
+[Code]
+
+function NeedsAddPath(Param: string): boolean;
+var
+  OrigPath: string;
+begin
+  if not RegQueryStringValue(HKEY_CURRENT_USER,
+    'Environment',
+    'Path', OrigPath)
+  then begin
+    Result := True;
+    exit;
+  end;
+  { look for the path with leading and trailing semicolon }
+  { Pos() returns 0 if not found }
+  Result := Pos(';' + ExpandConstant(Param) + ';', ';' + OrigPath + ';') = 0;
+end;

app/ollama.rc

@@ -0,0 +1,29 @@
+#include <winver.h>
+
+VS_VERSION_INFO VERSIONINFO
+ FILEFLAGSMASK 0x3fL
+#ifdef _DEBUG
+ FILEFLAGS 0x1L
+#else
+ FILEFLAGS 0x0L
+#endif
+ FILEOS 0x40004L
+ FILETYPE 0x1L
+ FILESUBTYPE 0x0L
+BEGIN
+    BLOCK "StringFileInfo"
+    BEGIN
+        BLOCK "040904b0"
+        BEGIN
+            VALUE "FileDescription", "Ollama"
+            VALUE "InternalName", "Ollama"
+            VALUE "OriginalFilename", "ollama app.exe"
+            VALUE "ProductName", "Ollama"
+        END
+    END
+
+    BLOCK "VarFileInfo"
+    BEGIN
+        VALUE "Translation", 0x409, 1200
+    END
+END

app/ollama_welcome.ps1

@@ -0,0 +1,8 @@
+# TODO - consider ANSI colors and maybe ASCII art...
+write-host ""
+write-host "Welcome to Ollama!"
+write-host ""
+write-host "Run your first model:"
+write-host ""
+write-host "`tollama run llama2"
+write-host ""

app/store/store.go

@@ -0,0 +1,98 @@
+package store
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"sync"
+
+	"github.com/google/uuid"
+)
+
+type Store struct {
+	ID           string `json:"id"`
+	FirstTimeRun bool   `json:"first-time-run"`
+}
+
+var (
+	lock  sync.Mutex
+	store Store
+)
+
+func GetID() string {
+	lock.Lock()
+	defer lock.Unlock()
+	if store.ID == "" {
+		initStore()
+	}
+	return store.ID
+
+}
+
+func GetFirstTimeRun() bool {
+	lock.Lock()
+	defer lock.Unlock()
+	if store.ID == "" {
+		initStore()
+	}
+	return store.FirstTimeRun
+}
+
+func SetFirstTimeRun(val bool) {
+	lock.Lock()
+	defer lock.Unlock()
+	if store.FirstTimeRun == val {
+		return
+	}
+	store.FirstTimeRun = val
+	writeStore(getStorePath())
+}
+
+// lock must be held
+func initStore() {
+	storeFile, err := os.Open(getStorePath())
+	if err == nil {
+		defer storeFile.Close()
+		err = json.NewDecoder(storeFile).Decode(&store)
+		if err == nil {
+			slog.Debug(fmt.Sprintf("loaded existing store %s - ID: %s", getStorePath(), store.ID))
+			return
+		}
+	} else if !errors.Is(err, os.ErrNotExist) {
+		slog.Debug(fmt.Sprintf("unexpected error searching for store: %s", err))
+	}
+	slog.Debug("initializing new store")
+	store.ID = uuid.New().String()
+	writeStore(getStorePath())
+}
+
+func writeStore(storeFilename string) {
+	ollamaDir := filepath.Dir(storeFilename)
+	_, err := os.Stat(ollamaDir)
+	if errors.Is(err, os.ErrNotExist) {
+		if err := os.MkdirAll(ollamaDir, 0o755); err != nil {
+			slog.Error(fmt.Sprintf("create ollama dir %s: %v", ollamaDir, err))
+			return
+		}
+	}
+	payload, err := json.Marshal(store)
+	if err != nil {
+		slog.Error(fmt.Sprintf("failed to marshal store: %s", err))
+		return
+	}
+	fp, err := os.OpenFile(storeFilename, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+	if err != nil {
+		slog.Error(fmt.Sprintf("write store payload %s: %v", storeFilename, err))
+		return
+	}
+	defer fp.Close()
+	if n, err := fp.Write(payload); err != nil || n != len(payload) {
+		slog.Error(fmt.Sprintf("write store payload %s: %d vs %d -- %v", storeFilename, n, len(payload), err))
+		return
+	}
+	slog.Debug("Store contents: " + string(payload))
+	slog.Info(fmt.Sprintf("wrote store: %s", storeFilename))
+}

app/store/store_darwin.go

@@ -0,0 +1,13 @@
+package store
+
+import (
+	"os"
+	"path/filepath"
+)
+
+func getStorePath() string {
+	// TODO - system wide location?
+
+	home := os.Getenv("HOME")
+	return filepath.Join(home, "Library", "Application Support", "Ollama", "config.json")
+}

app/store/store_linux.go

@@ -0,0 +1,16 @@
+package store
+
+import (
+	"os"
+	"path/filepath"
+)
+
+func getStorePath() string {
+	if os.Geteuid() == 0 {
+		// TODO where should we store this on linux for system-wide operation?
+		return "/etc/ollama/config.json"
+	}
+
+	home := os.Getenv("HOME")
+	return filepath.Join(home, ".ollama", "config.json")
+}

app/store/store_windows.go

@@ -0,0 +1,11 @@
+package store
+
+import (
+	"os"
+	"path/filepath"
+)
+
+func getStorePath() string {
+	localAppData := os.Getenv("LOCALAPPDATA")
+	return filepath.Join(localAppData, "Ollama", "config.json")
+}

app/tray/commontray/types.go

@@ -0,0 +1,24 @@
+package commontray
+
+var (
+	Title   = "Ollama"
+	ToolTip = "Ollama"
+
+	UpdateIconName = "tray_upgrade"
+	IconName       = "tray"
+)
+
+type Callbacks struct {
+	Quit       chan struct{}
+	Update     chan struct{}
+	DoFirstUse chan struct{}
+	ShowLogs   chan struct{}
+}
+
+type OllamaTray interface {
+	GetCallbacks() Callbacks
+	Run()
+	UpdateAvailable(ver string) error
+	DisplayFirstUseNotification() error
+	Quit()
+}

app/tray/tray.go

@@ -0,0 +1,33 @@
+package tray
+
+import (
+	"fmt"
+	"runtime"
+
+	"github.com/jmorganca/ollama/app/assets"
+	"github.com/jmorganca/ollama/app/tray/commontray"
+)
+
+func NewTray() (commontray.OllamaTray, error) {
+	extension := ".png"
+	if runtime.GOOS == "windows" {
+		extension = ".ico"
+	}
+	iconName := commontray.UpdateIconName + extension
+	updateIcon, err := assets.GetIcon(iconName)
+	if err != nil {
+		return nil, fmt.Errorf("failed to load icon %s: %w", iconName, err)
+	}
+	iconName = commontray.IconName + extension
+	icon, err := assets.GetIcon(iconName)
+	if err != nil {
+		return nil, fmt.Errorf("failed to load icon %s: %w", iconName, err)
+	}
+
+	tray, err := InitPlatformTray(icon, updateIcon)
+	if err != nil {
+		return nil, err
+	}
+
+	return tray, nil
+}

app/tray/tray_nonwindows.go

@@ -0,0 +1,13 @@
+//go:build !windows
+
+package tray
+
+import (
+	"fmt"
+
+	"github.com/jmorganca/ollama/app/tray/commontray"
+)
+
+func InitPlatformTray(icon, updateIcon []byte) (commontray.OllamaTray, error) {
+	return nil, fmt.Errorf("NOT IMPLEMENTED YET")
+}

app/tray/tray_windows.go

@@ -0,0 +1,10 @@
+package tray
+
+import (
+	"github.com/jmorganca/ollama/app/tray/commontray"
+	"github.com/jmorganca/ollama/app/tray/wintray"
+)
+
+func InitPlatformTray(icon, updateIcon []byte) (commontray.OllamaTray, error) {
+	return wintray.InitTray(icon, updateIcon)
+}

app/tray/wintray/eventloop.go

@@ -0,0 +1,184 @@
+//go:build windows
+
+package wintray
+
+import (
+	"fmt"
+	"log/slog"
+	"sync"
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+)
+
+var (
+	quitOnce sync.Once
+)
+
+func (t *winTray) Run() {
+	nativeLoop()
+}
+
+func nativeLoop() {
+	// Main message pump.
+	slog.Debug("starting event handling loop")
+	m := &struct {
+		WindowHandle windows.Handle
+		Message      uint32
+		Wparam       uintptr
+		Lparam       uintptr
+		Time         uint32
+		Pt           point
+		LPrivate     uint32
+	}{}
+	for {
+		ret, _, err := pGetMessage.Call(uintptr(unsafe.Pointer(m)), 0, 0, 0)
+
+		// If the function retrieves a message other than WM_QUIT, the return value is nonzero.
+		// If the function retrieves the WM_QUIT message, the return value is zero.
+		// If there is an error, the return value is -1
+		// https://msdn.microsoft.com/en-us/library/windows/desktop/ms644936(v=vs.85).aspx
+		switch int32(ret) {
+		case -1:
+			slog.Error(fmt.Sprintf("get message failure: %v", err))
+			return
+		case 0:
+			return
+		default:
+			pTranslateMessage.Call(uintptr(unsafe.Pointer(m))) //nolint:errcheck
+			pDispatchMessage.Call(uintptr(unsafe.Pointer(m)))  //nolint:errcheck
+
+		}
+	}
+}
+
+// WindowProc callback function that processes messages sent to a window.
+// https://msdn.microsoft.com/en-us/library/windows/desktop/ms633573(v=vs.85).aspx
+func (t *winTray) wndProc(hWnd windows.Handle, message uint32, wParam, lParam uintptr) (lResult uintptr) {
+	const (
+		WM_RBUTTONUP   = 0x0205
+		WM_LBUTTONUP   = 0x0202
+		WM_COMMAND     = 0x0111
+		WM_ENDSESSION  = 0x0016
+		WM_CLOSE       = 0x0010
+		WM_DESTROY     = 0x0002
+		WM_MOUSEMOVE   = 0x0200
+		WM_LBUTTONDOWN = 0x0201
+	)
+	switch message {
+	case WM_COMMAND:
+		menuItemId := int32(wParam)
+		// https://docs.microsoft.com/en-us/windows/win32/menurc/wm-command#menus
+		switch menuItemId {
+		case quitMenuID:
+			select {
+			case t.callbacks.Quit <- struct{}{}:
+			// should not happen but in case not listening
+			default:
+				slog.Error("no listener on Quit")
+			}
+		case updateMenuID:
+			select {
+			case t.callbacks.Update <- struct{}{}:
+			// should not happen but in case not listening
+			default:
+				slog.Error("no listener on Update")
+			}
+		case diagLogsMenuID:
+			select {
+			case t.callbacks.ShowLogs <- struct{}{}:
+			// should not happen but in case not listening
+			default:
+				slog.Error("no listener on ShowLogs")
+			}
+		default:
+			slog.Debug(fmt.Sprintf("Unexpected menu item id: %d", menuItemId))
+		}
+	case WM_CLOSE:
+		boolRet, _, err := pDestroyWindow.Call(uintptr(t.window))
+		if boolRet == 0 {
+			slog.Error(fmt.Sprintf("failed to destroy window: %s", err))
+		}
+		err = t.wcex.unregister()
+		if err != nil {
+			slog.Error(fmt.Sprintf("failed to uregister windo %s", err))
+		}
+	case WM_DESTROY:
+		// same as WM_ENDSESSION, but throws 0 exit code after all
+		defer pPostQuitMessage.Call(uintptr(int32(0))) //nolint:errcheck
+		fallthrough
+	case WM_ENDSESSION:
+		t.muNID.Lock()
+		if t.nid != nil {
+			err := t.nid.delete()
+			if err != nil {
+				slog.Error(fmt.Sprintf("failed to delete nid: %s", err))
+			}
+		}
+		t.muNID.Unlock()
+	case t.wmSystrayMessage:
+		switch lParam {
+		case WM_MOUSEMOVE, WM_LBUTTONDOWN:
+			// Ignore these...
+		case WM_RBUTTONUP, WM_LBUTTONUP:
+			err := t.showMenu()
+			if err != nil {
+				slog.Error(fmt.Sprintf("failed to show menu: %s", err))
+			}
+		case 0x405: // TODO - how is this magic value derived for the notification left click
+			if t.pendingUpdate {
+				select {
+				case t.callbacks.Update <- struct{}{}:
+				// should not happen but in case not listening
+				default:
+					slog.Error("no listener on Update")
+				}
+			} else {
+				select {
+				case t.callbacks.DoFirstUse <- struct{}{}:
+				// should not happen but in case not listening
+				default:
+					slog.Error("no listener on DoFirstUse")
+				}
+			}
+		case 0x404: // Middle click or close notification
+			// slog.Debug("doing nothing on close of first time notification")
+		default:
+			// 0x402 also seems common - what is it?
+			slog.Debug(fmt.Sprintf("unmanaged app message, lParm: 0x%x", lParam))
+		}
+	case t.wmTaskbarCreated: // on explorer.exe restarts
+		t.muNID.Lock()
+		err := t.nid.add()
+		if err != nil {
+			slog.Error(fmt.Sprintf("failed to refresh the taskbar on explorer restart: %s", err))
+		}
+		t.muNID.Unlock()
+	default:
+		// Calls the default window procedure to provide default processing for any window messages that an application does not process.
+		// https://msdn.microsoft.com/en-us/library/windows/desktop/ms633572(v=vs.85).aspx
+		lResult, _, _ = pDefWindowProc.Call(
+			uintptr(hWnd),
+			uintptr(message),
+			uintptr(wParam),
+			uintptr(lParam),
+		)
+	}
+	return
+}
+
+func (t *winTray) Quit() {
+	quitOnce.Do(quit)
+}
+
+func quit() {
+	boolRet, _, err := pPostMessage.Call(
+		uintptr(wt.window),
+		WM_CLOSE,
+		0,
+		0,
+	)
+	if boolRet == 0 {
+		slog.Error(fmt.Sprintf("failed to post close message on shutdown %s", err))
+	}
+}

app/tray/wintray/menus.go

@@ -0,0 +1,71 @@
+//go:build windows
+
+package wintray
+
+import (
+	"fmt"
+	"log/slog"
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+)
+
+const (
+	updatAvailableMenuID = 1
+	updateMenuID         = updatAvailableMenuID + 1
+	separatorMenuID      = updateMenuID + 1
+	diagLogsMenuID       = separatorMenuID + 1
+	diagSeparatorMenuID  = diagLogsMenuID + 1
+	quitMenuID           = diagSeparatorMenuID + 1
+)
+
+func (t *winTray) initMenus() error {
+	if err := t.addOrUpdateMenuItem(diagLogsMenuID, 0, diagLogsMenuTitle, false); err != nil {
+		return fmt.Errorf("unable to create menu entries %w\n", err)
+	}
+	if err := t.addSeparatorMenuItem(diagSeparatorMenuID, 0); err != nil {
+		return fmt.Errorf("unable to create menu entries %w", err)
+	}
+	if err := t.addOrUpdateMenuItem(quitMenuID, 0, quitMenuTitle, false); err != nil {
+		return fmt.Errorf("unable to create menu entries %w\n", err)
+	}
+	return nil
+}
+
+func (t *winTray) UpdateAvailable(ver string) error {
+	if !t.updateNotified {
+		slog.Debug("updating menu and sending notification for new update")
+		if err := t.addOrUpdateMenuItem(updatAvailableMenuID, 0, updateAvailableMenuTitle, true); err != nil {
+			return fmt.Errorf("unable to create menu entries %w", err)
+		}
+		if err := t.addOrUpdateMenuItem(updateMenuID, 0, updateMenutTitle, false); err != nil {
+			return fmt.Errorf("unable to create menu entries %w", err)
+		}
+		if err := t.addSeparatorMenuItem(separatorMenuID, 0); err != nil {
+			return fmt.Errorf("unable to create menu entries %w", err)
+		}
+		iconFilePath, err := iconBytesToFilePath(wt.updateIcon)
+		if err != nil {
+			return fmt.Errorf("unable to write icon data to temp file: %w", err)
+		}
+		if err := wt.setIcon(iconFilePath); err != nil {
+			return fmt.Errorf("unable to set icon: %w", err)
+		}
+		t.updateNotified = true
+
+		t.pendingUpdate = true
+		// Now pop up the notification
+		t.muNID.Lock()
+		defer t.muNID.Unlock()
+		copy(t.nid.InfoTitle[:], windows.StringToUTF16(updateTitle))
+		copy(t.nid.Info[:], windows.StringToUTF16(fmt.Sprintf(updateMessage, ver)))
+		t.nid.Flags |= NIF_INFO
+		t.nid.Timeout = 10
+		t.nid.Size = uint32(unsafe.Sizeof(*wt.nid))
+		err = t.nid.modify()
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}

app/tray/wintray/messages.go

@@ -0,0 +1,15 @@
+//go:build windows
+
+package wintray
+
+const (
+	firstTimeTitle   = "Ollama is running"
+	firstTimeMessage = "Click here to get started"
+	updateTitle      = "Update available"
+	updateMessage    = "Ollama version %s is ready to install"
+
+	quitMenuTitle            = "Quit Ollama"
+	updateAvailableMenuTitle = "An update is available"
+	updateMenutTitle         = "Restart to update"
+	diagLogsMenuTitle        = "View logs"
+)

app/tray/wintray/notifyicon.go

@@ -0,0 +1,66 @@
+//go:build windows
+
+package wintray
+
+import (
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+)
+
+// Contains information that the system needs to display notifications in the notification area.
+// Used by Shell_NotifyIcon.
+// https://msdn.microsoft.com/en-us/library/windows/desktop/bb773352(v=vs.85).aspx
+// https://msdn.microsoft.com/en-us/library/windows/desktop/bb762159
+type notifyIconData struct {
+	Size                       uint32
+	Wnd                        windows.Handle
+	ID, Flags, CallbackMessage uint32
+	Icon                       windows.Handle
+	Tip                        [128]uint16
+	State, StateMask           uint32
+	Info                       [256]uint16
+	// Timeout, Version           uint32
+	Timeout uint32
+
+	InfoTitle   [64]uint16
+	InfoFlags   uint32
+	GuidItem    windows.GUID
+	BalloonIcon windows.Handle
+}
+
+func (nid *notifyIconData) add() error {
+	const NIM_ADD = 0x00000000
+	res, _, err := pShellNotifyIcon.Call(
+		uintptr(NIM_ADD),
+		uintptr(unsafe.Pointer(nid)),
+	)
+	if res == 0 {
+		return err
+	}
+	return nil
+}
+
+func (nid *notifyIconData) modify() error {
+	const NIM_MODIFY = 0x00000001
+	res, _, err := pShellNotifyIcon.Call(
+		uintptr(NIM_MODIFY),
+		uintptr(unsafe.Pointer(nid)),
+	)
+	if res == 0 {
+		return err
+	}
+	return nil
+}
+
+func (nid *notifyIconData) delete() error {
+	const NIM_DELETE = 0x00000002
+	res, _, err := pShellNotifyIcon.Call(
+		uintptr(NIM_DELETE),
+		uintptr(unsafe.Pointer(nid)),
+	)
+	if res == 0 {
+		return err
+	}
+	return nil
+}

app/tray/wintray/tray.go

@@ -0,0 +1,485 @@
+//go:build windows
+
+package wintray
+
+import (
+	"crypto/md5"
+	"encoding/hex"
+	"fmt"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"sort"
+	"sync"
+	"unsafe"
+
+	"github.com/jmorganca/ollama/app/tray/commontray"
+	"golang.org/x/sys/windows"
+)
+
+// Helpful sources: https://github.com/golang/exp/blob/master/shiny/driver/internal/win32
+
+// Contains information about loaded resources
+type winTray struct {
+	instance,
+	icon,
+	cursor,
+	window windows.Handle
+
+	loadedImages   map[string]windows.Handle
+	muLoadedImages sync.RWMutex
+
+	// menus keeps track of the submenus keyed by the menu item ID, plus 0
+	// which corresponds to the main popup menu.
+	menus    map[uint32]windows.Handle
+	muMenus  sync.RWMutex
+	menuOf   map[uint32]windows.Handle
+	muMenuOf sync.RWMutex
+	// menuItemIcons maintains the bitmap of each menu item (if applies). It's
+	// needed to show the icon correctly when showing a previously hidden menu
+	// item again.
+	// menuItemIcons   map[uint32]windows.Handle
+	// muMenuItemIcons sync.RWMutex
+	visibleItems   map[uint32][]uint32
+	muVisibleItems sync.RWMutex
+
+	nid   *notifyIconData
+	muNID sync.RWMutex
+	wcex  *wndClassEx
+
+	wmSystrayMessage,
+	wmTaskbarCreated uint32
+
+	pendingUpdate  bool
+	updateNotified bool // Only pop up the notification once - TODO consider daily nag?
+	// Callbacks
+	callbacks  commontray.Callbacks
+	normalIcon []byte
+	updateIcon []byte
+}
+
+var wt winTray
+
+func (t *winTray) GetCallbacks() commontray.Callbacks {
+	return t.callbacks
+}
+
+func InitTray(icon, updateIcon []byte) (*winTray, error) {
+	wt.callbacks.Quit = make(chan struct{})
+	wt.callbacks.Update = make(chan struct{})
+	wt.callbacks.ShowLogs = make(chan struct{})
+	wt.callbacks.DoFirstUse = make(chan struct{})
+	wt.normalIcon = icon
+	wt.updateIcon = updateIcon
+	if err := wt.initInstance(); err != nil {
+		return nil, fmt.Errorf("Unable to init instance: %w\n", err)
+	}
+
+	if err := wt.createMenu(); err != nil {
+		return nil, fmt.Errorf("Unable to create menu: %w\n", err)
+	}
+
+	iconFilePath, err := iconBytesToFilePath(wt.normalIcon)
+	if err != nil {
+		return nil, fmt.Errorf("Unable to write icon data to temp file: %w", err)
+	}
+	if err := wt.setIcon(iconFilePath); err != nil {
+		return nil, fmt.Errorf("Unable to set icon: %w", err)
+	}
+
+	return &wt, wt.initMenus()
+}
+
+func (t *winTray) initInstance() error {
+	const (
+		className  = "OllamaClass"
+		windowName = ""
+	)
+
+	t.wmSystrayMessage = WM_USER + 1
+	t.visibleItems = make(map[uint32][]uint32)
+	t.menus = make(map[uint32]windows.Handle)
+	t.menuOf = make(map[uint32]windows.Handle)
+
+	t.loadedImages = make(map[string]windows.Handle)
+
+	taskbarEventNamePtr, _ := windows.UTF16PtrFromString("TaskbarCreated")
+	// https://msdn.microsoft.com/en-us/library/windows/desktop/ms644947
+	res, _, err := pRegisterWindowMessage.Call(
+		uintptr(unsafe.Pointer(taskbarEventNamePtr)),
+	)
+	if res == 0 { // success 0xc000-0xfff
+		return fmt.Errorf("failed to register window: %w", err)
+	}
+	t.wmTaskbarCreated = uint32(res)
+
+	instanceHandle, _, err := pGetModuleHandle.Call(0)
+	if instanceHandle == 0 {
+		return err
+	}
+	t.instance = windows.Handle(instanceHandle)
+
+	// https://msdn.microsoft.com/en-us/library/windows/desktop/ms648072(v=vs.85).aspx
+	iconHandle, _, err := pLoadIcon.Call(0, uintptr(IDI_APPLICATION))
+	if iconHandle == 0 {
+		return err
+	}
+	t.icon = windows.Handle(iconHandle)
+
+	// https://msdn.microsoft.com/en-us/library/windows/desktop/ms648391(v=vs.85).aspx
+	cursorHandle, _, err := pLoadCursor.Call(0, uintptr(IDC_ARROW))
+	if cursorHandle == 0 {
+		return err
+	}
+	t.cursor = windows.Handle(cursorHandle)
+
+	classNamePtr, err := windows.UTF16PtrFromString(className)
+	if err != nil {
+		return err
+	}
+
+	windowNamePtr, err := windows.UTF16PtrFromString(windowName)
+	if err != nil {
+		return err
+	}
+
+	t.wcex = &wndClassEx{
+		Style:      CS_HREDRAW | CS_VREDRAW,
+		WndProc:    windows.NewCallback(t.wndProc),
+		Instance:   t.instance,
+		Icon:       t.icon,
+		Cursor:     t.cursor,
+		Background: windows.Handle(6), // (COLOR_WINDOW + 1)
+		ClassName:  classNamePtr,
+		IconSm:     t.icon,
+	}
+	if err := t.wcex.register(); err != nil {
+		return err
+	}
+
+	windowHandle, _, err := pCreateWindowEx.Call(
+		uintptr(0),
+		uintptr(unsafe.Pointer(classNamePtr)),
+		uintptr(unsafe.Pointer(windowNamePtr)),
+		uintptr(WS_OVERLAPPEDWINDOW),
+		uintptr(CW_USEDEFAULT),
+		uintptr(CW_USEDEFAULT),
+		uintptr(CW_USEDEFAULT),
+		uintptr(CW_USEDEFAULT),
+		uintptr(0),
+		uintptr(0),
+		uintptr(t.instance),
+		uintptr(0),
+	)
+	if windowHandle == 0 {
+		return err
+	}
+	t.window = windows.Handle(windowHandle)
+
+	pShowWindow.Call(uintptr(t.window), uintptr(SW_HIDE)) //nolint:errcheck
+
+	boolRet, _, err := pUpdateWindow.Call(uintptr(t.window))
+	if boolRet == 0 {
+		slog.Error(fmt.Sprintf("failed to update window: %s", err))
+	}
+
+	t.muNID.Lock()
+	defer t.muNID.Unlock()
+	t.nid = &notifyIconData{
+		Wnd:             windows.Handle(t.window),
+		ID:              100,
+		Flags:           NIF_MESSAGE,
+		CallbackMessage: t.wmSystrayMessage,
+	}
+	t.nid.Size = uint32(unsafe.Sizeof(*t.nid))
+
+	return t.nid.add()
+}
+
+func (t *winTray) createMenu() error {
+
+	menuHandle, _, err := pCreatePopupMenu.Call()
+	if menuHandle == 0 {
+		return err
+	}
+	t.menus[0] = windows.Handle(menuHandle)
+
+	// https://msdn.microsoft.com/en-us/library/windows/desktop/ms647575(v=vs.85).aspx
+	mi := struct {
+		Size, Mask, Style, Max uint32
+		Background             windows.Handle
+		ContextHelpID          uint32
+		MenuData               uintptr
+	}{
+		Mask: MIM_APPLYTOSUBMENUS,
+	}
+	mi.Size = uint32(unsafe.Sizeof(mi))
+
+	res, _, err := pSetMenuInfo.Call(
+		uintptr(t.menus[0]),
+		uintptr(unsafe.Pointer(&mi)),
+	)
+	if res == 0 {
+		return err
+	}
+	return nil
+}
+
+// Contains information about a menu item.
+// https://msdn.microsoft.com/en-us/library/windows/desktop/ms647578(v=vs.85).aspx
+type menuItemInfo struct {
+	Size, Mask, Type, State     uint32
+	ID                          uint32
+	SubMenu, Checked, Unchecked windows.Handle
+	ItemData                    uintptr
+	TypeData                    *uint16
+	Cch                         uint32
+	BMPItem                     windows.Handle
+}
+
+func (t *winTray) addOrUpdateMenuItem(menuItemId uint32, parentId uint32, title string, disabled bool) error {
+	titlePtr, err := windows.UTF16PtrFromString(title)
+	if err != nil {
+		return err
+	}
+
+	mi := menuItemInfo{
+		Mask:     MIIM_FTYPE | MIIM_STRING | MIIM_ID | MIIM_STATE,
+		Type:     MFT_STRING,
+		ID:       uint32(menuItemId),
+		TypeData: titlePtr,
+		Cch:      uint32(len(title)),
+	}
+	mi.Size = uint32(unsafe.Sizeof(mi))
+	if disabled {
+		mi.State |= MFS_DISABLED
+	}
+
+	var res uintptr
+	t.muMenus.RLock()
+	menu := t.menus[parentId]
+	t.muMenus.RUnlock()
+	if t.getVisibleItemIndex(parentId, menuItemId) != -1 {
+		// We set the menu item info based on the menuID
+		boolRet, _, err := pSetMenuItemInfo.Call(
+			uintptr(menu),
+			uintptr(menuItemId),
+			0,
+			uintptr(unsafe.Pointer(&mi)),
+		)
+		if boolRet == 0 {
+			return fmt.Errorf("failed to set menu item: %w", err)
+		}
+	}
+
+	if res == 0 {
+		// Menu item does not already exist, create it
+		t.muMenus.RLock()
+		submenu, exists := t.menus[menuItemId]
+		t.muMenus.RUnlock()
+		if exists {
+			mi.Mask |= MIIM_SUBMENU
+			mi.SubMenu = submenu
+		}
+		t.addToVisibleItems(parentId, menuItemId)
+		position := t.getVisibleItemIndex(parentId, menuItemId)
+		res, _, err = pInsertMenuItem.Call(
+			uintptr(menu),
+			uintptr(position),
+			1,
+			uintptr(unsafe.Pointer(&mi)),
+		)
+		if res == 0 {
+			t.delFromVisibleItems(parentId, menuItemId)
+			return err
+		}
+		t.muMenuOf.Lock()
+		t.menuOf[menuItemId] = menu
+		t.muMenuOf.Unlock()
+	}
+
+	return nil
+}
+
+func (t *winTray) addSeparatorMenuItem(menuItemId, parentId uint32) error {
+
+	mi := menuItemInfo{
+		Mask: MIIM_FTYPE | MIIM_ID | MIIM_STATE,
+		Type: MFT_SEPARATOR,
+		ID:   uint32(menuItemId),
+	}
+
+	mi.Size = uint32(unsafe.Sizeof(mi))
+
+	t.addToVisibleItems(parentId, menuItemId)
+	position := t.getVisibleItemIndex(parentId, menuItemId)
+	t.muMenus.RLock()
+	menu := uintptr(t.menus[parentId])
+	t.muMenus.RUnlock()
+	res, _, err := pInsertMenuItem.Call(
+		menu,
+		uintptr(position),
+		1,
+		uintptr(unsafe.Pointer(&mi)),
+	)
+	if res == 0 {
+		return err
+	}
+
+	return nil
+}
+
+// func (t *winTray) hideMenuItem(menuItemId, parentId uint32) error {
+// 	const ERROR_SUCCESS syscall.Errno = 0
+
+// 	t.muMenus.RLock()
+// 	menu := uintptr(t.menus[parentId])
+// 	t.muMenus.RUnlock()
+// 	res, _, err := pRemoveMenu.Call(
+// 		menu,
+// 		uintptr(menuItemId),
+// 		MF_BYCOMMAND,
+// 	)
+// 	if res == 0 && err.(syscall.Errno) != ERROR_SUCCESS {
+// 		return err
+// 	}
+// 	t.delFromVisibleItems(parentId, menuItemId)
+
+// 	return nil
+// }
+
+func (t *winTray) showMenu() error {
+	p := point{}
+	boolRet, _, err := pGetCursorPos.Call(uintptr(unsafe.Pointer(&p)))
+	if boolRet == 0 {
+		return err
+	}
+	boolRet, _, err = pSetForegroundWindow.Call(uintptr(t.window))
+	if boolRet == 0 {
+		slog.Warn(fmt.Sprintf("failed to bring menu to foreground: %s", err))
+	}
+
+	boolRet, _, err = pTrackPopupMenu.Call(
+		uintptr(t.menus[0]),
+		TPM_BOTTOMALIGN|TPM_LEFTALIGN,
+		uintptr(p.X),
+		uintptr(p.Y),
+		0,
+		uintptr(t.window),
+		0,
+	)
+	if boolRet == 0 {
+		return err
+	}
+
+	return nil
+}
+
+func (t *winTray) delFromVisibleItems(parent, val uint32) {
+	t.muVisibleItems.Lock()
+	defer t.muVisibleItems.Unlock()
+	visibleItems := t.visibleItems[parent]
+	for i, itemval := range visibleItems {
+		if val == itemval {
+			t.visibleItems[parent] = append(visibleItems[:i], visibleItems[i+1:]...)
+			break
+		}
+	}
+}
+
+func (t *winTray) addToVisibleItems(parent, val uint32) {
+	t.muVisibleItems.Lock()
+	defer t.muVisibleItems.Unlock()
+	if visibleItems, exists := t.visibleItems[parent]; !exists {
+		t.visibleItems[parent] = []uint32{val}
+	} else {
+		newvisible := append(visibleItems, val)
+		sort.Slice(newvisible, func(i, j int) bool { return newvisible[i] < newvisible[j] })
+		t.visibleItems[parent] = newvisible
+	}
+}
+
+func (t *winTray) getVisibleItemIndex(parent, val uint32) int {
+	t.muVisibleItems.RLock()
+	defer t.muVisibleItems.RUnlock()
+	for i, itemval := range t.visibleItems[parent] {
+		if val == itemval {
+			return i
+		}
+	}
+	return -1
+}
+
+func iconBytesToFilePath(iconBytes []byte) (string, error) {
+	bh := md5.Sum(iconBytes)
+	dataHash := hex.EncodeToString(bh[:])
+	iconFilePath := filepath.Join(os.TempDir(), "ollama_temp_icon_"+dataHash)
+
+	if _, err := os.Stat(iconFilePath); os.IsNotExist(err) {
+		if err := os.WriteFile(iconFilePath, iconBytes, 0644); err != nil {
+			return "", err
+		}
+	}
+	return iconFilePath, nil
+}
+
+// Loads an image from file and shows it in tray.
+// Shell_NotifyIcon: https://msdn.microsoft.com/en-us/library/windows/desktop/bb762159(v=vs.85).aspx
+func (t *winTray) setIcon(src string) error {
+
+	h, err := t.loadIconFrom(src)
+	if err != nil {
+		return err
+	}
+
+	t.muNID.Lock()
+	defer t.muNID.Unlock()
+	t.nid.Icon = h
+	t.nid.Flags |= NIF_ICON
+	t.nid.Size = uint32(unsafe.Sizeof(*t.nid))
+
+	return t.nid.modify()
+}
+
+// Loads an image from file to be shown in tray or menu item.
+// LoadImage: https://msdn.microsoft.com/en-us/library/windows/desktop/ms648045(v=vs.85).aspx
+func (t *winTray) loadIconFrom(src string) (windows.Handle, error) {
+
+	// Save and reuse handles of loaded images
+	t.muLoadedImages.RLock()
+	h, ok := t.loadedImages[src]
+	t.muLoadedImages.RUnlock()
+	if !ok {
+		srcPtr, err := windows.UTF16PtrFromString(src)
+		if err != nil {
+			return 0, err
+		}
+		res, _, err := pLoadImage.Call(
+			0,
+			uintptr(unsafe.Pointer(srcPtr)),
+			IMAGE_ICON,
+			0,
+			0,
+			LR_LOADFROMFILE|LR_DEFAULTSIZE,
+		)
+		if res == 0 {
+			return 0, err
+		}
+		h = windows.Handle(res)
+		t.muLoadedImages.Lock()
+		t.loadedImages[src] = h
+		t.muLoadedImages.Unlock()
+	}
+	return h, nil
+}
+
+func (t *winTray) DisplayFirstUseNotification() error {
+	t.muNID.Lock()
+	defer t.muNID.Unlock()
+	copy(t.nid.InfoTitle[:], windows.StringToUTF16(firstTimeTitle))
+	copy(t.nid.Info[:], windows.StringToUTF16(firstTimeMessage))
+	t.nid.Flags |= NIF_INFO
+	t.nid.Size = uint32(unsafe.Sizeof(*wt.nid))
+
+	return t.nid.modify()
+}

app/tray/wintray/w32api.go

@@ -0,0 +1,89 @@
+//go:build windows
+
+package wintray
+
+import (
+	"runtime"
+
+	"golang.org/x/sys/windows"
+)
+
+var (
+	k32 = windows.NewLazySystemDLL("Kernel32.dll")
+	u32 = windows.NewLazySystemDLL("User32.dll")
+	s32 = windows.NewLazySystemDLL("Shell32.dll")
+
+	pCreatePopupMenu       = u32.NewProc("CreatePopupMenu")
+	pCreateWindowEx        = u32.NewProc("CreateWindowExW")
+	pDefWindowProc         = u32.NewProc("DefWindowProcW")
+	pDestroyWindow         = u32.NewProc("DestroyWindow")
+	pDispatchMessage       = u32.NewProc("DispatchMessageW")
+	pGetCursorPos          = u32.NewProc("GetCursorPos")
+	pGetMessage            = u32.NewProc("GetMessageW")
+	pGetModuleHandle       = k32.NewProc("GetModuleHandleW")
+	pInsertMenuItem        = u32.NewProc("InsertMenuItemW")
+	pLoadCursor            = u32.NewProc("LoadCursorW")
+	pLoadIcon              = u32.NewProc("LoadIconW")
+	pLoadImage             = u32.NewProc("LoadImageW")
+	pPostMessage           = u32.NewProc("PostMessageW")
+	pPostQuitMessage       = u32.NewProc("PostQuitMessage")
+	pRegisterClass         = u32.NewProc("RegisterClassExW")
+	pRegisterWindowMessage = u32.NewProc("RegisterWindowMessageW")
+	pSetForegroundWindow   = u32.NewProc("SetForegroundWindow")
+	pSetMenuInfo           = u32.NewProc("SetMenuInfo")
+	pSetMenuItemInfo       = u32.NewProc("SetMenuItemInfoW")
+	pShellNotifyIcon       = s32.NewProc("Shell_NotifyIconW")
+	pShowWindow            = u32.NewProc("ShowWindow")
+	pTrackPopupMenu        = u32.NewProc("TrackPopupMenu")
+	pTranslateMessage      = u32.NewProc("TranslateMessage")
+	pUnregisterClass       = u32.NewProc("UnregisterClassW")
+	pUpdateWindow          = u32.NewProc("UpdateWindow")
+)
+
+const (
+	CS_HREDRAW          = 0x0002
+	CS_VREDRAW          = 0x0001
+	CW_USEDEFAULT       = 0x80000000
+	IDC_ARROW           = 32512 // Standard arrow
+	IDI_APPLICATION     = 32512
+	IMAGE_ICON          = 1          // Loads an icon
+	LR_DEFAULTSIZE      = 0x00000040 // Loads default-size icon for windows(SM_CXICON x SM_CYICON) if cx, cy are set to zero
+	LR_LOADFROMFILE     = 0x00000010 // Loads the stand-alone image from the file
+	MF_BYCOMMAND        = 0x00000000
+	MFS_DISABLED        = 0x00000003
+	MFT_SEPARATOR       = 0x00000800
+	MFT_STRING          = 0x00000000
+	MIIM_BITMAP         = 0x00000080
+	MIIM_FTYPE          = 0x00000100
+	MIIM_ID             = 0x00000002
+	MIIM_STATE          = 0x00000001
+	MIIM_STRING         = 0x00000040
+	MIIM_SUBMENU        = 0x00000004
+	MIM_APPLYTOSUBMENUS = 0x80000000
+	NIF_ICON            = 0x00000002
+	NIF_INFO            = 0x00000010
+	NIF_MESSAGE         = 0x00000001
+	SW_HIDE             = 0
+	TPM_BOTTOMALIGN     = 0x0020
+	TPM_LEFTALIGN       = 0x0000
+	WM_CLOSE            = 0x0010
+	WM_USER             = 0x0400
+	WS_CAPTION          = 0x00C00000
+	WS_MAXIMIZEBOX      = 0x00010000
+	WS_MINIMIZEBOX      = 0x00020000
+	WS_OVERLAPPED       = 0x00000000
+	WS_OVERLAPPEDWINDOW = WS_OVERLAPPED | WS_CAPTION | WS_SYSMENU | WS_THICKFRAME | WS_MINIMIZEBOX | WS_MAXIMIZEBOX
+	WS_SYSMENU          = 0x00080000
+	WS_THICKFRAME       = 0x00040000
+)
+
+// Not sure if this is actually needed on windows
+func init() {
+	runtime.LockOSThread()
+}
+
+// The POINT structure defines the x- and y- coordinates of a point.
+// https://msdn.microsoft.com/en-us/library/windows/desktop/dd162805(v=vs.85).aspx
+type point struct {
+	X, Y int32
+}

app/tray/wintray/winclass.go

@@ -0,0 +1,45 @@
+//go:build windows
+
+package wintray
+
+import (
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+)
+
+// Contains window class information.
+// It is used with the RegisterClassEx and GetClassInfoEx functions.
+// https://msdn.microsoft.com/en-us/library/ms633577.aspx
+type wndClassEx struct {
+	Size, Style                        uint32
+	WndProc                            uintptr
+	ClsExtra, WndExtra                 int32
+	Instance, Icon, Cursor, Background windows.Handle
+	MenuName, ClassName                *uint16
+	IconSm                             windows.Handle
+}
+
+// Registers a window class for subsequent use in calls to the CreateWindow or CreateWindowEx function.
+// https://msdn.microsoft.com/en-us/library/ms633587.aspx
+func (w *wndClassEx) register() error {
+	w.Size = uint32(unsafe.Sizeof(*w))
+	res, _, err := pRegisterClass.Call(uintptr(unsafe.Pointer(w)))
+	if res == 0 {
+		return err
+	}
+	return nil
+}
+
+// Unregisters a window class, freeing the memory required for the class.
+// https://msdn.microsoft.com/en-us/library/ms644899.aspx
+func (w *wndClassEx) unregister() error {
+	res, _, err := pUnregisterClass.Call(
+		uintptr(unsafe.Pointer(w.ClassName)),
+		uintptr(w.Instance),
+	)
+	if res == 0 {
+		return err
+	}
+	return nil
+}

auth/auth.go

@@ -0,0 +1,61 @@
+package auth
+
+import (
+	"bytes"
+	"context"
+	"crypto/rand"
+	"encoding/base64"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"path/filepath"
+
+	"golang.org/x/crypto/ssh"
+)
+
+const defaultPrivateKey = "id_ed25519"
+
+func NewNonce(r io.Reader, length int) (string, error) {
+	nonce := make([]byte, length)
+	if _, err := io.ReadFull(r, nonce); err != nil {
+		return "", err
+	}
+
+	return base64.RawURLEncoding.EncodeToString(nonce), nil
+}
+
+func Sign(ctx context.Context, bts []byte) (string, error) {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return "", err
+	}
+
+	keyPath := filepath.Join(home, ".ollama", defaultPrivateKey)
+
+	privateKeyFile, err := os.ReadFile(keyPath)
+	if err != nil {
+		slog.Info(fmt.Sprintf("Failed to load private key: %v", err))
+		return "", err
+	}
+
+	privateKey, err := ssh.ParsePrivateKey(privateKeyFile)
+	if err != nil {
+		return "", err
+	}
+
+	// get the pubkey, but remove the type
+	publicKey := ssh.MarshalAuthorizedKey(privateKey.PublicKey())
+	parts := bytes.Split(publicKey, []byte(" "))
+	if len(parts) < 2 {
+		return "", fmt.Errorf("malformed public key")
+	}
+
+	signedData, err := privateKey.Sign(rand.Reader, bts)
+	if err != nil {
+		return "", err
+	}
+
+	// signature is <pubkey>:<signature>
+	return fmt.Sprintf("%s:%s", bytes.TrimSpace(parts[1]), base64.StdEncoding.EncodeToString(signedData.Blob)), nil
+}

cmd/cmd.go

@@ -14,15 +14,15 @@ import (
 	"net"
 	"net/http"
 	"os"
-	"os/exec"
 	"os/signal"
 	"path/filepath"
-	"regexp"
 	"runtime"
 	"strings"
 	"syscall"
 	"time"
 
+	"github.com/containerd/console"
+
 	"github.com/olekukonko/tablewriter"
 	"github.com/spf13/cobra"
 	"golang.org/x/crypto/ssh"
@@ -33,13 +33,10 @@ import (
 	"github.com/jmorganca/ollama/format"
 	"github.com/jmorganca/ollama/parser"
 	"github.com/jmorganca/ollama/progress"
-	"github.com/jmorganca/ollama/readline"
 	"github.com/jmorganca/ollama/server"
 	"github.com/jmorganca/ollama/version"
 )
 
-type ImageData []byte
-
 func CreateHandler(cmd *cobra.Command, args []string) error {
 	filename, _ := cmd.Flags().GetString("file")
 	filename, err := filepath.Abs(filename)
@@ -151,19 +148,68 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 
 	name := args[0]
+
 	// check if the model exists on the server
-	_, err = client.Show(cmd.Context(), &api.ShowRequest{Name: name})
+	show, err := client.Show(cmd.Context(), &api.ShowRequest{Name: name})
 	var statusError api.StatusError
 	switch {
 	case errors.As(err, &statusError) && statusError.StatusCode == http.StatusNotFound:
-		if err := PullHandler(cmd, args); err != nil {
+		if err := PullHandler(cmd, []string{name}); err != nil {
+			return err
+		}
+
+		show, err = client.Show(cmd.Context(), &api.ShowRequest{Name: name})
+		if err != nil {
 			return err
 		}
 	case err != nil:
 		return err
 	}
 
-	return RunGenerate(cmd, args)
+	interactive := true
+
+	opts := runOptions{
+		Model:       args[0],
+		WordWrap:    os.Getenv("TERM") == "xterm-256color",
+		Options:     map[string]interface{}{},
+		MultiModal:  slices.Contains(show.Details.Families, "clip"),
+		ParentModel: show.Details.ParentModel,
+	}
+
+	format, err := cmd.Flags().GetString("format")
+	if err != nil {
+		return err
+	}
+	opts.Format = format
+
+	prompts := args[1:]
+	// prepend stdin to the prompt if provided
+	if !term.IsTerminal(int(os.Stdin.Fd())) {
+		in, err := io.ReadAll(os.Stdin)
+		if err != nil {
+			return err
+		}
+
+		prompts = append([]string{string(in)}, prompts...)
+		opts.WordWrap = false
+		interactive = false
+	}
+	opts.Prompt = strings.Join(prompts, " ")
+	if len(prompts) > 0 {
+		interactive = false
+	}
+
+	nowrap, err := cmd.Flags().GetBool("nowordwrap")
+	if err != nil {
+		return err
+	}
+	opts.WordWrap = !nowrap
+
+	if !interactive {
+		return generate(cmd, opts)
+	}
+
+	return generateInteractive(cmd, opts)
 }
 
 func PushHandler(cmd *cobra.Command, args []string) error {
@@ -415,66 +461,139 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 	return nil
 }
 
-func RunGenerate(cmd *cobra.Command, args []string) error {
-	interactive := true
-
-	opts := generateOptions{
-		Model:    args[0],
-		WordWrap: os.Getenv("TERM") == "xterm-256color",
-		Options:  map[string]interface{}{},
-		Images:   []ImageData{},
-	}
-
-	format, err := cmd.Flags().GetString("format")
-	if err != nil {
-		return err
-	}
-	opts.Format = format
-
-	prompts := args[1:]
-	// prepend stdin to the prompt if provided
-	if !term.IsTerminal(int(os.Stdin.Fd())) {
-		in, err := io.ReadAll(os.Stdin)
-		if err != nil {
-			return err
-		}
-
-		prompts = append([]string{string(in)}, prompts...)
-		opts.WordWrap = false
-		interactive = false
-	}
-	opts.Prompt = strings.Join(prompts, " ")
-	if len(prompts) > 0 {
-		interactive = false
-	}
-
-	nowrap, err := cmd.Flags().GetBool("nowordwrap")
-	if err != nil {
-		return err
-	}
-	opts.WordWrap = !nowrap
-
-	if !interactive {
-		return generate(cmd, opts)
-	}
-
-	return generateInteractive(cmd, opts)
-}
-
 type generateContextKey string
 
-type generateOptions struct {
-	Model    string
-	Prompt   string
-	WordWrap bool
-	Format   string
-	System   string
-	Template string
-	Images   []ImageData
-	Options  map[string]interface{}
+type runOptions struct {
+	Model       string
+	ParentModel string
+	Prompt      string
+	Messages    []api.Message
+	WordWrap    bool
+	Format      string
+	System      string
+	Template    string
+	Images      []api.ImageData
+	Options     map[string]interface{}
+	MultiModal  bool
 }
 
-func generate(cmd *cobra.Command, opts generateOptions) error {
+type displayResponseState struct {
+	lineLength int
+	wordBuffer string
+}
+
+func displayResponse(content string, wordWrap bool, state *displayResponseState) {
+	termWidth, _, _ := term.GetSize(int(os.Stdout.Fd()))
+	if wordWrap && termWidth >= 10 {
+		for _, ch := range content {
+			if state.lineLength+1 > termWidth-5 {
+				if len(state.wordBuffer) > termWidth-10 {
+					fmt.Printf("%s%c", state.wordBuffer, ch)
+					state.wordBuffer = ""
+					state.lineLength = 0
+					continue
+				}
+
+				// backtrack the length of the last word and clear to the end of the line
+				fmt.Printf("\x1b[%dD\x1b[K\n", len(state.wordBuffer))
+				fmt.Printf("%s%c", state.wordBuffer, ch)
+				state.lineLength = len(state.wordBuffer) + 1
+			} else {
+				fmt.Print(string(ch))
+				state.lineLength += 1
+
+				switch ch {
+				case ' ':
+					state.wordBuffer = ""
+				case '\n':
+					state.lineLength = 0
+				default:
+					state.wordBuffer += string(ch)
+				}
+			}
+		}
+	} else {
+		fmt.Printf("%s%s", state.wordBuffer, content)
+		if len(state.wordBuffer) > 0 {
+			state.wordBuffer = ""
+		}
+	}
+}
+
+func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return nil, err
+	}
+
+	p := progress.NewProgress(os.Stderr)
+	defer p.StopAndClear()
+
+	spinner := progress.NewSpinner("")
+	p.Add("", spinner)
+
+	cancelCtx, cancel := context.WithCancel(cmd.Context())
+	defer cancel()
+
+	sigChan := make(chan os.Signal, 1)
+	signal.Notify(sigChan, syscall.SIGINT)
+
+	go func() {
+		<-sigChan
+		cancel()
+	}()
+
+	var state *displayResponseState = &displayResponseState{}
+	var latest api.ChatResponse
+	var fullResponse strings.Builder
+	var role string
+
+	fn := func(response api.ChatResponse) error {
+		p.StopAndClear()
+
+		latest = response
+
+		role = response.Message.Role
+		content := response.Message.Content
+		fullResponse.WriteString(content)
+
+		displayResponse(content, opts.WordWrap, state)
+
+		return nil
+	}
+
+	req := &api.ChatRequest{
+		Model:    opts.Model,
+		Messages: opts.Messages,
+		Format:   opts.Format,
+		Options:  opts.Options,
+	}
+
+	if err := client.Chat(cancelCtx, req, fn); err != nil {
+		if errors.Is(err, context.Canceled) {
+			return nil, nil
+		}
+		return nil, err
+	}
+
+	if len(opts.Messages) > 0 {
+		fmt.Println()
+		fmt.Println()
+	}
+
+	verbose, err := cmd.Flags().GetBool("verbose")
+	if err != nil {
+		return nil, err
+	}
+
+	if verbose {
+		latest.Summary()
+	}
+
+	return &api.Message{Role: role, Content: fullResponse.String()}, nil
+}
+
+func generate(cmd *cobra.Command, opts runOptions) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
@@ -493,11 +612,6 @@ func generate(cmd *cobra.Command, opts generateOptions) error {
 		generateContext = []int{}
 	}
 
-	termWidth, _, err := term.GetSize(int(os.Stdout.Fd()))
-	if err != nil {
-		opts.WordWrap = false
-	}
-
 	ctx, cancel := context.WithCancel(cmd.Context())
 	defer cancel()
 
@@ -509,94 +623,44 @@ func generate(cmd *cobra.Command, opts generateOptions) error {
 		cancel()
 	}()
 
-	var currentLineLength int
-	var wordBuffer string
+	var state *displayResponseState = &displayResponseState{}
 
 	fn := func(response api.GenerateResponse) error {
 		p.StopAndClear()
 
 		latest = response
+		content := response.Response
 
-		termWidth, _, _ = term.GetSize(int(os.Stdout.Fd()))
-		if opts.WordWrap && termWidth >= 10 {
-			for _, ch := range response.Response {
-				if currentLineLength+1 > termWidth-5 {
-					if len(wordBuffer) > termWidth-10 {
-						fmt.Printf("%s%c", wordBuffer, ch)
-						wordBuffer = ""
-						currentLineLength = 0
-						continue
-					}
-
-					// backtrack the length of the last word and clear to the end of the line
-					fmt.Printf("\x1b[%dD\x1b[K\n", len(wordBuffer))
-					fmt.Printf("%s%c", wordBuffer, ch)
-					currentLineLength = len(wordBuffer) + 1
-				} else {
-					fmt.Print(string(ch))
-					currentLineLength += 1
-
-					switch ch {
-					case ' ':
-						wordBuffer = ""
-					case '\n':
-						currentLineLength = 0
-					default:
-						wordBuffer += string(ch)
-					}
-				}
-			}
-		} else {
-			fmt.Printf("%s%s", wordBuffer, response.Response)
-			if len(wordBuffer) > 0 {
-				wordBuffer = ""
-			}
-		}
+		displayResponse(content, opts.WordWrap, state)
 
 		return nil
 	}
 
-	images := make([]api.ImageData, 0)
-	for _, i := range opts.Images {
-		images = append(images, api.ImageData(i))
+	if opts.MultiModal {
+		opts.Prompt, opts.Images, err = extractFileData(opts.Prompt)
+		if err != nil {
+			return err
+		}
 	}
+
 	request := api.GenerateRequest{
 		Model:    opts.Model,
 		Prompt:   opts.Prompt,
 		Context:  generateContext,
+		Images:   opts.Images,
 		Format:   opts.Format,
 		System:   opts.System,
 		Template: opts.Template,
 		Options:  opts.Options,
-		Images:   images,
 	}
 
 	if err := client.Generate(ctx, &request, fn); err != nil {
-		switch {
-		case errors.Is(err, context.Canceled):
+		if errors.Is(err, context.Canceled) {
 			return nil
-		case strings.Contains(err.Error(), "unsupported model format"):
-			// pull and retry to see if the model has been updated
-			parts := strings.Split(opts.Model, string(os.PathSeparator))
-			if len(parts) == 1 {
-				// this is a library model, log some info
-				fmt.Fprintln(os.Stderr, "This model is no longer compatible with Ollama. Pulling a new version...")
-			}
-			if err := PullHandler(cmd, []string{opts.Model}); err != nil {
-				fmt.Printf("Error: %s\n", err)
-				return fmt.Errorf("unsupported model, please update this model to gguf format") // relay the original error
-			}
-			// retry
-			if err := client.Generate(ctx, &request, fn); err != nil {
-				if errors.Is(err, context.Canceled) {
-					return nil
-				}
-				return err
-			}
-		default:
-			return err
 		}
+		return err
 	}
+
 	if opts.Prompt != "" {
 		fmt.Println()
 		fmt.Println()
@@ -621,461 +685,8 @@ func generate(cmd *cobra.Command, opts generateOptions) error {
 	return nil
 }
 
-type MultilineState int
-
-const (
-	MultilineNone MultilineState = iota
-	MultilinePrompt
-	MultilineSystem
-	MultilineTemplate
-)
-
-func modelIsMultiModal(cmd *cobra.Command, name string) bool {
-	// get model details
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		fmt.Println("error: couldn't connect to ollama server")
-		return false
-	}
-
-	req := api.ShowRequest{Name: name}
-	resp, err := client.Show(cmd.Context(), &req)
-	if err != nil {
-		return false
-	}
-
-	return slices.Contains(resp.Details.Families, "clip")
-}
-
-func generateInteractive(cmd *cobra.Command, opts generateOptions) error {
-	multiModal := modelIsMultiModal(cmd, opts.Model)
-
-	// load the model
-	loadOpts := generateOptions{
-		Model:  opts.Model,
-		Prompt: "",
-		Images: []ImageData{},
-	}
-	if err := generate(cmd, loadOpts); err != nil {
-		return err
-	}
-
-	usage := func() {
-		fmt.Fprintln(os.Stderr, "Available Commands:")
-		fmt.Fprintln(os.Stderr, "  /set          Set session variables")
-		fmt.Fprintln(os.Stderr, "  /show         Show model information")
-		fmt.Fprintln(os.Stderr, "  /bye          Exit")
-		fmt.Fprintln(os.Stderr, "  /?, /help     Help for a command")
-		fmt.Fprintln(os.Stderr, "  /? shortcuts  Help for keyboard shortcuts")
-		fmt.Fprintln(os.Stderr, "")
-		fmt.Fprintln(os.Stderr, "Use \"\"\" to begin a multi-line message.")
-		fmt.Fprintln(os.Stderr, "")
-	}
-
-	usageSet := func() {
-		fmt.Fprintln(os.Stderr, "Available Commands:")
-		fmt.Fprintln(os.Stderr, "  /set parameter ...     Set a parameter")
-		fmt.Fprintln(os.Stderr, "  /set system <string>   Set system message")
-		fmt.Fprintln(os.Stderr, "  /set template <string> Set prompt template")
-		fmt.Fprintln(os.Stderr, "  /set history           Enable history")
-		fmt.Fprintln(os.Stderr, "  /set nohistory         Disable history")
-		fmt.Fprintln(os.Stderr, "  /set wordwrap          Enable wordwrap")
-		fmt.Fprintln(os.Stderr, "  /set nowordwrap        Disable wordwrap")
-		fmt.Fprintln(os.Stderr, "  /set format json       Enable JSON mode")
-		fmt.Fprintln(os.Stderr, "  /set noformat          Disable formatting")
-		fmt.Fprintln(os.Stderr, "  /set verbose           Show LLM stats")
-		fmt.Fprintln(os.Stderr, "  /set quiet             Disable LLM stats")
-		fmt.Fprintln(os.Stderr, "")
-	}
-
-	usageShortcuts := func() {
-		fmt.Fprintln(os.Stderr, "Available keyboard shortcuts:")
-		fmt.Fprintln(os.Stderr, "  Ctrl + a            Move to the beginning of the line (Home)")
-		fmt.Fprintln(os.Stderr, "  Ctrl + e            Move to the end of the line (End)")
-		fmt.Fprintln(os.Stderr, "   Alt + b            Move back (left) one word")
-		fmt.Fprintln(os.Stderr, "   Alt + f            Move forward (right) one word")
-		fmt.Fprintln(os.Stderr, "  Ctrl + k            Delete the sentence after the cursor")
-		fmt.Fprintln(os.Stderr, "  Ctrl + u            Delete the sentence before the cursor")
-		fmt.Fprintln(os.Stderr, "")
-		fmt.Fprintln(os.Stderr, "  Ctrl + l            Clear the screen")
-		fmt.Fprintln(os.Stderr, "  Ctrl + c            Stop the model from responding")
-		fmt.Fprintln(os.Stderr, "  Ctrl + d            Exit ollama (/bye)")
-		fmt.Fprintln(os.Stderr, "")
-	}
-
-	usageShow := func() {
-		fmt.Fprintln(os.Stderr, "Available Commands:")
-		fmt.Fprintln(os.Stderr, "  /show license      Show model license")
-		fmt.Fprintln(os.Stderr, "  /show modelfile    Show Modelfile for this model")
-		fmt.Fprintln(os.Stderr, "  /show parameters   Show parameters for this model")
-		fmt.Fprintln(os.Stderr, "  /show system       Show system message")
-		fmt.Fprintln(os.Stderr, "  /show template     Show prompt template")
-		fmt.Fprintln(os.Stderr, "")
-	}
-
-	// only list out the most common parameters
-	usageParameters := func() {
-		fmt.Fprintln(os.Stderr, "Available Parameters:")
-		fmt.Fprintln(os.Stderr, "  /set parameter seed <int>             Random number seed")
-		fmt.Fprintln(os.Stderr, "  /set parameter num_predict <int>      Max number of tokens to predict")
-		fmt.Fprintln(os.Stderr, "  /set parameter top_k <int>            Pick from top k num of tokens")
-		fmt.Fprintln(os.Stderr, "  /set parameter top_p <float>          Pick token based on sum of probabilities")
-		fmt.Fprintln(os.Stderr, "  /set parameter num_ctx <int>          Set the context size")
-		fmt.Fprintln(os.Stderr, "  /set parameter temperature <float>    Set creativity level")
-		fmt.Fprintln(os.Stderr, "  /set parameter repeat_penalty <float> How strongly to penalize repetitions")
-		fmt.Fprintln(os.Stderr, "  /set parameter repeat_last_n <int>    Set how far back to look for repetitions")
-		fmt.Fprintln(os.Stderr, "  /set parameter num_gpu <int>          The number of layers to send to the GPU")
-		fmt.Fprintln(os.Stderr, "  /set parameter stop \"<string>\", ...   Set the stop parameters")
-		fmt.Fprintln(os.Stderr, "")
-	}
-
-	scanner, err := readline.New(readline.Prompt{
-		Prompt:         ">>> ",
-		AltPrompt:      "... ",
-		Placeholder:    "Send a message (/? for help)",
-		AltPlaceholder: `Use """ to end multi-line input`,
-	})
-	if err != nil {
-		return err
-	}
-
-	fmt.Print(readline.StartBracketedPaste)
-	defer fmt.Printf(readline.EndBracketedPaste)
-
-	var multiline MultilineState
-	var prompt string
-
-	for {
-		line, err := scanner.Readline()
-		switch {
-		case errors.Is(err, io.EOF):
-			fmt.Println()
-			return nil
-		case errors.Is(err, readline.ErrInterrupt):
-			if line == "" {
-				fmt.Println("\nUse Ctrl + d or /bye to exit.")
-			}
-
-			scanner.Prompt.UseAlt = false
-			prompt = ""
-
-			continue
-		case err != nil:
-			return err
-		}
-
-		switch {
-		case strings.HasPrefix(prompt, `"""`):
-			// if the prompt so far starts with """ then we're in multiline mode
-			// and we need to keep reading until we find a line that ends with """
-			cut, found := strings.CutSuffix(line, `"""`)
-			prompt += cut
-
-			if !found {
-				prompt += "\n"
-				continue
-			}
-
-			prompt = strings.TrimPrefix(prompt, `"""`)
-			scanner.Prompt.UseAlt = false
-
-			switch multiline {
-			case MultilineSystem:
-				opts.System = prompt
-				prompt = ""
-				fmt.Println("Set system message.")
-			case MultilineTemplate:
-				opts.Template = prompt
-				prompt = ""
-				fmt.Println("Set prompt template.")
-			}
-			multiline = MultilineNone
-		case strings.HasPrefix(line, `"""`) && len(prompt) == 0:
-			scanner.Prompt.UseAlt = true
-			multiline = MultilinePrompt
-			prompt += line + "\n"
-			continue
-		case scanner.Pasting:
-			prompt += line + "\n"
-			continue
-		case strings.HasPrefix(line, "/list"):
-			args := strings.Fields(line)
-			if err := ListHandler(cmd, args[1:]); err != nil {
-				return err
-			}
-		case strings.HasPrefix(line, "/set"):
-			args := strings.Fields(line)
-			if len(args) > 1 {
-				switch args[1] {
-				case "history":
-					scanner.HistoryEnable()
-				case "nohistory":
-					scanner.HistoryDisable()
-				case "wordwrap":
-					opts.WordWrap = true
-					fmt.Println("Set 'wordwrap' mode.")
-				case "nowordwrap":
-					opts.WordWrap = false
-					fmt.Println("Set 'nowordwrap' mode.")
-				case "verbose":
-					cmd.Flags().Set("verbose", "true")
-					fmt.Println("Set 'verbose' mode.")
-				case "quiet":
-					cmd.Flags().Set("verbose", "false")
-					fmt.Println("Set 'quiet' mode.")
-				case "format":
-					if len(args) < 3 || args[2] != "json" {
-						fmt.Println("Invalid or missing format. For 'json' mode use '/set format json'")
-					} else {
-						opts.Format = args[2]
-						fmt.Printf("Set format to '%s' mode.\n", args[2])
-					}
-				case "noformat":
-					opts.Format = ""
-					fmt.Println("Disabled format.")
-				case "parameter":
-					if len(args) < 4 {
-						usageParameters()
-						continue
-					}
-					var params []string
-					for _, p := range args[3:] {
-						params = append(params, p)
-					}
-					fp, err := api.FormatParams(map[string][]string{args[2]: params})
-					if err != nil {
-						fmt.Printf("Couldn't set parameter: %q\n\n", err)
-						continue
-					}
-					fmt.Printf("Set parameter '%s' to '%s'\n\n", args[2], strings.Join(params, ", "))
-					opts.Options[args[2]] = fp[args[2]]
-				case "system", "template":
-					if len(args) < 3 {
-						usageSet()
-						continue
-					}
-					line := strings.Join(args[2:], " ")
-					line = strings.TrimPrefix(line, `"""`)
-					if strings.HasPrefix(args[2], `"""`) {
-						cut, found := strings.CutSuffix(line, `"""`)
-						prompt += cut
-						if found {
-							if args[1] == "system" {
-								opts.System = prompt
-								fmt.Println("Set system message.")
-							} else {
-								opts.Template = prompt
-								fmt.Println("Set prompt template.")
-							}
-							prompt = ""
-						} else {
-							prompt = `"""` + prompt + "\n"
-							if args[1] == "system" {
-								multiline = MultilineSystem
-							} else {
-								multiline = MultilineTemplate
-							}
-							scanner.Prompt.UseAlt = true
-						}
-					} else {
-						opts.System = line
-						fmt.Println("Set system message.")
-					}
-				default:
-					fmt.Printf("Unknown command '/set %s'. Type /? for help\n", args[1])
-				}
-			} else {
-				usageSet()
-			}
-		case strings.HasPrefix(line, "/show"):
-			args := strings.Fields(line)
-			if len(args) > 1 {
-				client, err := api.ClientFromEnvironment()
-				if err != nil {
-					fmt.Println("error: couldn't connect to ollama server")
-					return err
-				}
-				resp, err := client.Show(cmd.Context(), &api.ShowRequest{Name: opts.Model})
-				if err != nil {
-					fmt.Println("error: couldn't get model")
-					return err
-				}
-
-				switch args[1] {
-				case "license":
-					if resp.License == "" {
-						fmt.Print("No license was specified for this model.\n\n")
-					} else {
-						fmt.Println(resp.License)
-					}
-				case "modelfile":
-					fmt.Println(resp.Modelfile)
-				case "parameters":
-					if resp.Parameters == "" {
-						fmt.Print("No parameters were specified for this model.\n\n")
-					} else {
-						if len(opts.Options) > 0 {
-							fmt.Println("User defined parameters:")
-							for k, v := range opts.Options {
-								fmt.Printf("%-*s %v\n", 30, k, v)
-							}
-							fmt.Println()
-						}
-						fmt.Println("Model defined parameters:")
-						fmt.Println(resp.Parameters)
-					}
-				case "system":
-					switch {
-					case opts.System != "":
-						fmt.Println(opts.System + "\n")
-					case resp.System != "":
-						fmt.Println(resp.System + "\n")
-					default:
-						fmt.Print("No system message was specified for this model.\n\n")
-					}
-				case "template":
-					switch {
-					case opts.Template != "":
-						fmt.Println(opts.Template + "\n")
-					case resp.Template != "":
-						fmt.Println(resp.Template)
-					default:
-						fmt.Print("No prompt template was specified for this model.\n\n")
-					}
-				default:
-					fmt.Printf("Unknown command '/show %s'. Type /? for help\n", args[1])
-				}
-			} else {
-				usageShow()
-			}
-		case strings.HasPrefix(line, "/help"), strings.HasPrefix(line, "/?"):
-			args := strings.Fields(line)
-			if len(args) > 1 {
-				switch args[1] {
-				case "set", "/set":
-					usageSet()
-				case "show", "/show":
-					usageShow()
-				case "shortcut", "shortcuts":
-					usageShortcuts()
-				}
-			} else {
-				usage()
-			}
-		case line == "/exit", line == "/bye":
-			return nil
-		case strings.HasPrefix(line, "/"):
-			args := strings.Fields(line)
-			isFile := false
-
-			if multiModal {
-				for _, f := range extractFileNames(line) {
-					if strings.HasPrefix(f, args[0]) {
-						isFile = true
-						break
-					}
-				}
-			}
-
-			if isFile {
-				prompt += line
-			} else {
-				fmt.Printf("Unknown command '%s'. Type /? for help\n", args[0])
-				continue
-			}
-		default:
-			prompt += line
-		}
-
-		if len(prompt) > 0 && multiline == MultilineNone {
-			opts.Prompt = prompt
-			if multiModal {
-				newPrompt, images, err := extractFileData(prompt)
-				if err != nil {
-					return err
-				}
-				opts.Prompt = newPrompt
-
-				// reset the context if we find another image
-				if len(images) > 0 {
-					opts.Images = images
-					ctx := cmd.Context()
-					ctx = context.WithValue(ctx, generateContextKey("context"), []int{})
-					cmd.SetContext(ctx)
-				}
-				if len(opts.Images) == 0 {
-					fmt.Println("This model requires you to add a jpeg, png, or svg image.")
-					fmt.Println()
-					prompt = ""
-					continue
-				}
-			}
-			if err := generate(cmd, opts); err != nil {
-				return err
-			}
-
-			prompt = ""
-		}
-	}
-}
-
-func normalizeFilePath(fp string) string {
-	// Define a map of escaped characters and their replacements
-	replacements := map[string]string{
-		"\\ ":  " ",  // Escaped space
-		"\\(":  "(",  // Escaped left parenthesis
-		"\\)":  ")",  // Escaped right parenthesis
-		"\\[":  "[",  // Escaped left square bracket
-		"\\]":  "]",  // Escaped right square bracket
-		"\\{":  "{",  // Escaped left curly brace
-		"\\}":  "}",  // Escaped right curly brace
-		"\\$":  "$",  // Escaped dollar sign
-		"\\&":  "&",  // Escaped ampersand
-		"\\;":  ";",  // Escaped semicolon
-		"\\'":  "'",  // Escaped single quote
-		"\\\\": "\\", // Escaped backslash
-		"\\*":  "*",  // Escaped asterisk
-		"\\?":  "?",  // Escaped question mark
-	}
-
-	for escaped, actual := range replacements {
-		fp = strings.ReplaceAll(fp, escaped, actual)
-	}
-	return fp
-}
-
-func extractFileNames(input string) []string {
-	// Regex to match file paths starting with / or ./ and include escaped spaces (\ or %20)
-	// and followed by more characters and a file extension
-	regexPattern := `(?:\./|/)[\S\\ ]+?\.(?i:jpg|jpeg|png|svg)\b`
-	re := regexp.MustCompile(regexPattern)
-
-	return re.FindAllString(input, -1)
-}
-
-func extractFileData(input string) (string, []ImageData, error) {
-	filePaths := extractFileNames(input)
-	var imgs []ImageData
-
-	for _, fp := range filePaths {
-		nfp := normalizeFilePath(fp)
-		data, err := getImageData(nfp)
-		if err != nil {
-			if os.IsNotExist(err) {
-				continue
-			}
-			fmt.Printf("Couldn't process image: %q\n", err)
-			return "", imgs, err
-		}
-		fmt.Printf("Added image '%s'\n", nfp)
-		input = strings.ReplaceAll(input, fp, "")
-		imgs = append(imgs, data)
-	}
-	return input, imgs, nil
-}
-
 func RunServer(cmd *cobra.Command, _ []string) error {
-	host, port, err := net.SplitHostPort(os.Getenv("OLLAMA_HOST"))
+	host, port, err := net.SplitHostPort(strings.Trim(os.Getenv("OLLAMA_HOST"), "\"'"))
 	if err != nil {
 		host, port = "127.0.0.1", "11434"
 		if ip := net.ParseIP(strings.Trim(os.Getenv("OLLAMA_HOST"), "[]")); ip != nil {
@@ -1095,50 +706,6 @@ func RunServer(cmd *cobra.Command, _ []string) error {
 	return server.Serve(ln)
 }
 
-func getImageData(filePath string) ([]byte, error) {
-	file, err := os.Open(filePath)
-	if err != nil {
-		return nil, err
-	}
-	defer file.Close()
-
-	buf := make([]byte, 512)
-	_, err = file.Read(buf)
-	if err != nil {
-		return nil, err
-	}
-
-	contentType := http.DetectContentType(buf)
-	allowedTypes := []string{"image/jpeg", "image/jpg", "image/svg+xml", "image/png"}
-	if !slices.Contains(allowedTypes, contentType) {
-		return nil, fmt.Errorf("invalid image type: %s", contentType)
-	}
-
-	info, err := file.Stat()
-	if err != nil {
-		return nil, err
-	}
-
-	// Check if the file size exceeds 100MB
-	var maxSize int64 = 100 * 1024 * 1024 // 100MB in bytes
-	if info.Size() > maxSize {
-		return nil, fmt.Errorf("file size exceeds maximum limit (100MB)")
-	}
-
-	buf = make([]byte, info.Size())
-	_, err = file.Seek(0, 0)
-	if err != nil {
-		return nil, err
-	}
-
-	_, err = io.ReadFull(file, buf)
-	if err != nil {
-		return nil, err
-	}
-
-	return buf, nil
-}
-
 func initializeKeypair() error {
 	home, err := os.UserHomeDir()
 	if err != nil {
@@ -1188,22 +755,8 @@ func initializeKeypair() error {
 	return nil
 }
 
-func startMacApp(ctx context.Context, client *api.Client) error {
-	exe, err := os.Executable()
-	if err != nil {
-		return err
-	}
-	link, err := os.Readlink(exe)
-	if err != nil {
-		return err
-	}
-	if !strings.Contains(link, "Ollama.app") {
-		return fmt.Errorf("could not find ollama app")
-	}
-	path := strings.Split(link, "Ollama.app")
-	if err := exec.Command("/usr/bin/open", "-a", path[0]+"Ollama.app").Run(); err != nil {
-		return err
-	}
+//nolint:unused
+func waitForServer(ctx context.Context, client *api.Client) error {
 	// wait for the server to start
 	timeout := time.After(5 * time.Second)
 	tick := time.Tick(500 * time.Millisecond)
@@ -1217,6 +770,7 @@ func startMacApp(ctx context.Context, client *api.Client) error {
 			}
 		}
 	}
+
 }
 
 func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
@@ -1225,15 +779,11 @@ func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
 		return err
 	}
 	if err := client.Heartbeat(cmd.Context()); err != nil {
-		if !strings.Contains(err.Error(), "connection refused") {
+		if !strings.Contains(err.Error(), " refused") {
 			return err
 		}
-		if runtime.GOOS == "darwin" {
-			if err := startMacApp(cmd.Context(), client); err != nil {
-				return fmt.Errorf("could not connect to ollama app, is it running?")
-			}
-		} else {
-			return fmt.Errorf("could not connect to ollama server, run 'ollama serve' to start it")
+		if err := startApp(cmd.Context(), client); err != nil {
+			return fmt.Errorf("could not connect to ollama app, is it running?")
 		}
 	}
 	return nil
@@ -1263,6 +813,11 @@ func NewCLI() *cobra.Command {
 	log.SetFlags(log.LstdFlags | log.Lshortfile)
 	cobra.EnableCommandSorting = false
 
+	if runtime.GOOS == "windows" {
+		// Enable colorful ANSI escape code in Windows terminal (disabled by default)
+		console.ConsoleFromFile(os.Stdout) //nolint:errcheck
+	}
+
 	rootCmd := &cobra.Command{
 		Use:           "ollama",
 		Short:         "Large language model runner",

cmd/interactive.go

@@ -0,0 +1,663 @@
+package cmd
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"path/filepath"
+	"regexp"
+	"sort"
+	"strings"
+
+	"github.com/spf13/cobra"
+	"golang.org/x/exp/slices"
+
+	"github.com/jmorganca/ollama/api"
+	"github.com/jmorganca/ollama/progress"
+	"github.com/jmorganca/ollama/readline"
+)
+
+type MultilineState int
+
+const (
+	MultilineNone MultilineState = iota
+	MultilinePrompt
+	MultilineSystem
+	MultilineTemplate
+)
+
+func loadModel(cmd *cobra.Command, opts *runOptions) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	p := progress.NewProgress(os.Stderr)
+	defer p.StopAndClear()
+
+	spinner := progress.NewSpinner("")
+	p.Add("", spinner)
+
+	showReq := api.ShowRequest{Name: opts.Model}
+	showResp, err := client.Show(cmd.Context(), &showReq)
+	if err != nil {
+		return err
+	}
+	opts.MultiModal = slices.Contains(showResp.Details.Families, "clip")
+	opts.ParentModel = showResp.Details.ParentModel
+
+	if len(showResp.Messages) > 0 {
+		opts.Messages = append(opts.Messages, showResp.Messages...)
+	}
+
+	chatReq := &api.ChatRequest{
+		Model:    opts.Model,
+		Messages: []api.Message{},
+	}
+	err = client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error {
+		p.StopAndClear()
+		if len(opts.Messages) > 0 {
+			for _, msg := range opts.Messages {
+				switch msg.Role {
+				case "user":
+					fmt.Printf(">>> %s\n", msg.Content)
+				case "assistant":
+					state := &displayResponseState{}
+					displayResponse(msg.Content, opts.WordWrap, state)
+					fmt.Println()
+					fmt.Println()
+				}
+			}
+		}
+		return nil
+	})
+	if err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func generateInteractive(cmd *cobra.Command, opts runOptions) error {
+	opts.Messages = make([]api.Message, 0)
+
+	err := loadModel(cmd, &opts)
+	if err != nil {
+		return err
+	}
+
+	usage := func() {
+		fmt.Fprintln(os.Stderr, "Available Commands:")
+		fmt.Fprintln(os.Stderr, "  /set            Set session variables")
+		fmt.Fprintln(os.Stderr, "  /show           Show model information")
+		fmt.Fprintln(os.Stderr, "  /load <model>   Load a session or model")
+		fmt.Fprintln(os.Stderr, "  /save <model>   Save your current session")
+		fmt.Fprintln(os.Stderr, "  /bye            Exit")
+		fmt.Fprintln(os.Stderr, "  /?, /help       Help for a command")
+		fmt.Fprintln(os.Stderr, "  /? shortcuts    Help for keyboard shortcuts")
+		fmt.Fprintln(os.Stderr, "")
+		fmt.Fprintln(os.Stderr, "Use \"\"\" to begin a multi-line message.")
+
+		if opts.MultiModal {
+			fmt.Fprintf(os.Stderr, "Use %s to include .jpg or .png images.\n", filepath.FromSlash("/path/to/file"))
+		}
+
+		fmt.Fprintln(os.Stderr, "")
+	}
+
+	usageSet := func() {
+		fmt.Fprintln(os.Stderr, "Available Commands:")
+		fmt.Fprintln(os.Stderr, "  /set parameter ...     Set a parameter")
+		fmt.Fprintln(os.Stderr, "  /set system <string>   Set system message")
+		fmt.Fprintln(os.Stderr, "  /set template <string> Set prompt template")
+		fmt.Fprintln(os.Stderr, "  /set history           Enable history")
+		fmt.Fprintln(os.Stderr, "  /set nohistory         Disable history")
+		fmt.Fprintln(os.Stderr, "  /set wordwrap          Enable wordwrap")
+		fmt.Fprintln(os.Stderr, "  /set nowordwrap        Disable wordwrap")
+		fmt.Fprintln(os.Stderr, "  /set format json       Enable JSON mode")
+		fmt.Fprintln(os.Stderr, "  /set noformat          Disable formatting")
+		fmt.Fprintln(os.Stderr, "  /set verbose           Show LLM stats")
+		fmt.Fprintln(os.Stderr, "  /set quiet             Disable LLM stats")
+		fmt.Fprintln(os.Stderr, "")
+	}
+
+	usageShortcuts := func() {
+		fmt.Fprintln(os.Stderr, "Available keyboard shortcuts:")
+		fmt.Fprintln(os.Stderr, "  Ctrl + a            Move to the beginning of the line (Home)")
+		fmt.Fprintln(os.Stderr, "  Ctrl + e            Move to the end of the line (End)")
+		fmt.Fprintln(os.Stderr, "   Alt + b            Move back (left) one word")
+		fmt.Fprintln(os.Stderr, "   Alt + f            Move forward (right) one word")
+		fmt.Fprintln(os.Stderr, "  Ctrl + k            Delete the sentence after the cursor")
+		fmt.Fprintln(os.Stderr, "  Ctrl + u            Delete the sentence before the cursor")
+		fmt.Fprintln(os.Stderr, "")
+		fmt.Fprintln(os.Stderr, "  Ctrl + l            Clear the screen")
+		fmt.Fprintln(os.Stderr, "  Ctrl + c            Stop the model from responding")
+		fmt.Fprintln(os.Stderr, "  Ctrl + d            Exit ollama (/bye)")
+		fmt.Fprintln(os.Stderr, "")
+	}
+
+	usageShow := func() {
+		fmt.Fprintln(os.Stderr, "Available Commands:")
+		fmt.Fprintln(os.Stderr, "  /show info         Show details for this model")
+		fmt.Fprintln(os.Stderr, "  /show license      Show model license")
+		fmt.Fprintln(os.Stderr, "  /show modelfile    Show Modelfile for this model")
+		fmt.Fprintln(os.Stderr, "  /show parameters   Show parameters for this model")
+		fmt.Fprintln(os.Stderr, "  /show system       Show system message")
+		fmt.Fprintln(os.Stderr, "  /show template     Show prompt template")
+		fmt.Fprintln(os.Stderr, "")
+	}
+
+	// only list out the most common parameters
+	usageParameters := func() {
+		fmt.Fprintln(os.Stderr, "Available Parameters:")
+		fmt.Fprintln(os.Stderr, "  /set parameter seed <int>             Random number seed")
+		fmt.Fprintln(os.Stderr, "  /set parameter num_predict <int>      Max number of tokens to predict")
+		fmt.Fprintln(os.Stderr, "  /set parameter top_k <int>            Pick from top k num of tokens")
+		fmt.Fprintln(os.Stderr, "  /set parameter top_p <float>          Pick token based on sum of probabilities")
+		fmt.Fprintln(os.Stderr, "  /set parameter num_ctx <int>          Set the context size")
+		fmt.Fprintln(os.Stderr, "  /set parameter temperature <float>    Set creativity level")
+		fmt.Fprintln(os.Stderr, "  /set parameter repeat_penalty <float> How strongly to penalize repetitions")
+		fmt.Fprintln(os.Stderr, "  /set parameter repeat_last_n <int>    Set how far back to look for repetitions")
+		fmt.Fprintln(os.Stderr, "  /set parameter num_gpu <int>          The number of layers to send to the GPU")
+		fmt.Fprintln(os.Stderr, "  /set parameter stop \"<string>\", ...   Set the stop parameters")
+		fmt.Fprintln(os.Stderr, "")
+	}
+
+	scanner, err := readline.New(readline.Prompt{
+		Prompt:         ">>> ",
+		AltPrompt:      "... ",
+		Placeholder:    "Send a message (/? for help)",
+		AltPlaceholder: `Use """ to end multi-line input`,
+	})
+	if err != nil {
+		return err
+	}
+
+	fmt.Print(readline.StartBracketedPaste)
+	defer fmt.Printf(readline.EndBracketedPaste)
+
+	var sb strings.Builder
+	var multiline MultilineState
+
+	for {
+		line, err := scanner.Readline()
+		switch {
+		case errors.Is(err, io.EOF):
+			fmt.Println()
+			return nil
+		case errors.Is(err, readline.ErrInterrupt):
+			if line == "" {
+				fmt.Println("\nUse Ctrl + d or /bye to exit.")
+			}
+
+			scanner.Prompt.UseAlt = false
+			sb.Reset()
+
+			continue
+		case err != nil:
+			return err
+		}
+
+		switch {
+		case multiline != MultilineNone:
+			// check if there's a multiline terminating string
+			before, ok := strings.CutSuffix(line, `"""`)
+			sb.WriteString(before)
+			if !ok {
+				fmt.Fprintln(&sb)
+				continue
+			}
+
+			switch multiline {
+			case MultilineSystem:
+				opts.System = sb.String()
+				opts.Messages = append(opts.Messages, api.Message{Role: "system", Content: opts.System})
+				fmt.Println("Set system message.")
+				sb.Reset()
+			case MultilineTemplate:
+				opts.Template = sb.String()
+				fmt.Println("Set prompt template.")
+				sb.Reset()
+			}
+
+			multiline = MultilineNone
+			scanner.Prompt.UseAlt = false
+		case strings.HasPrefix(line, `"""`):
+			line := strings.TrimPrefix(line, `"""`)
+			line, ok := strings.CutSuffix(line, `"""`)
+			sb.WriteString(line)
+			if !ok {
+				// no multiline terminating string; need more input
+				fmt.Fprintln(&sb)
+				multiline = MultilinePrompt
+				scanner.Prompt.UseAlt = true
+			}
+		case scanner.Pasting:
+			fmt.Fprintln(&sb, line)
+			continue
+		case strings.HasPrefix(line, "/list"):
+			args := strings.Fields(line)
+			if err := ListHandler(cmd, args[1:]); err != nil {
+				return err
+			}
+		case strings.HasPrefix(line, "/load"):
+			args := strings.Fields(line)
+			if len(args) != 2 {
+				fmt.Println("Usage:\n  /load <modelname>")
+				continue
+			}
+			opts.Model = args[1]
+			opts.Messages = []api.Message{}
+			fmt.Printf("Loading model '%s'\n", opts.Model)
+			if err := loadModel(cmd, &opts); err != nil {
+				return err
+			}
+			continue
+		case strings.HasPrefix(line, "/save"):
+			args := strings.Fields(line)
+			if len(args) != 2 {
+				fmt.Println("Usage:\n  /save <modelname>")
+				continue
+			}
+
+			client, err := api.ClientFromEnvironment()
+			if err != nil {
+				fmt.Println("error: couldn't connect to ollama server")
+				return err
+			}
+
+			req := &api.CreateRequest{
+				Name:      args[1],
+				Modelfile: buildModelfile(opts),
+			}
+			fn := func(resp api.ProgressResponse) error { return nil }
+			err = client.Create(cmd.Context(), req, fn)
+			if err != nil {
+				fmt.Println("error: couldn't save model")
+				return err
+			}
+			fmt.Printf("Created new model '%s'\n", args[1])
+			continue
+		case strings.HasPrefix(line, "/set"):
+			args := strings.Fields(line)
+			if len(args) > 1 {
+				switch args[1] {
+				case "history":
+					scanner.HistoryEnable()
+				case "nohistory":
+					scanner.HistoryDisable()
+				case "wordwrap":
+					opts.WordWrap = true
+					fmt.Println("Set 'wordwrap' mode.")
+				case "nowordwrap":
+					opts.WordWrap = false
+					fmt.Println("Set 'nowordwrap' mode.")
+				case "verbose":
+					cmd.Flags().Set("verbose", "true")
+					fmt.Println("Set 'verbose' mode.")
+				case "quiet":
+					cmd.Flags().Set("verbose", "false")
+					fmt.Println("Set 'quiet' mode.")
+				case "format":
+					if len(args) < 3 || args[2] != "json" {
+						fmt.Println("Invalid or missing format. For 'json' mode use '/set format json'")
+					} else {
+						opts.Format = args[2]
+						fmt.Printf("Set format to '%s' mode.\n", args[2])
+					}
+				case "noformat":
+					opts.Format = ""
+					fmt.Println("Disabled format.")
+				case "parameter":
+					if len(args) < 4 {
+						usageParameters()
+						continue
+					}
+					params := args[3:]
+					fp, err := api.FormatParams(map[string][]string{args[2]: params})
+					if err != nil {
+						fmt.Printf("Couldn't set parameter: %q\n", err)
+						continue
+					}
+					fmt.Printf("Set parameter '%s' to '%s'\n", args[2], strings.Join(params, ", "))
+					opts.Options[args[2]] = fp[args[2]]
+				case "system", "template":
+					if len(args) < 3 {
+						usageSet()
+						continue
+					}
+
+					if args[1] == "system" {
+						multiline = MultilineSystem
+					} else if args[1] == "template" {
+						multiline = MultilineTemplate
+					}
+
+					line := strings.Join(args[2:], " ")
+					line, ok := strings.CutPrefix(line, `"""`)
+					if !ok {
+						multiline = MultilineNone
+					} else {
+						// only cut suffix if the line is multiline
+						line, ok = strings.CutSuffix(line, `"""`)
+						if ok {
+							multiline = MultilineNone
+						}
+					}
+
+					sb.WriteString(line)
+					if multiline != MultilineNone {
+						scanner.Prompt.UseAlt = true
+						continue
+					}
+
+					if args[1] == "system" {
+						opts.System = sb.String() // for display in modelfile
+						newMessage := api.Message{Role: "system", Content: sb.String()}
+						// Check if the slice is not empty and the last message is from 'system'
+						if len(opts.Messages) > 0 && opts.Messages[len(opts.Messages)-1].Role == "system" {
+							// Replace the last message
+							opts.Messages[len(opts.Messages)-1] = newMessage
+						} else {
+							opts.Messages = append(opts.Messages, newMessage)
+						}
+						fmt.Println("Set system message.")
+						sb.Reset()
+					} else if args[1] == "template" {
+						opts.Template = sb.String()
+						fmt.Println("Set prompt template.")
+						sb.Reset()
+					}
+
+					sb.Reset()
+					continue
+				default:
+					fmt.Printf("Unknown command '/set %s'. Type /? for help\n", args[1])
+				}
+			} else {
+				usageSet()
+			}
+		case strings.HasPrefix(line, "/show"):
+			args := strings.Fields(line)
+			if len(args) > 1 {
+				client, err := api.ClientFromEnvironment()
+				if err != nil {
+					fmt.Println("error: couldn't connect to ollama server")
+					return err
+				}
+				req := &api.ShowRequest{
+					Name:     opts.Model,
+					System:   opts.System,
+					Template: opts.Template,
+					Options:  opts.Options,
+				}
+				resp, err := client.Show(cmd.Context(), req)
+				if err != nil {
+					fmt.Println("error: couldn't get model")
+					return err
+				}
+
+				switch args[1] {
+				case "info":
+					fmt.Println("Model details:")
+					if len(resp.Details.Families) > 0 {
+						fmt.Printf("Family              %s\n", strings.Join(resp.Details.Families, ", "))
+					} else if resp.Details.Family != "" {
+						fmt.Printf("Family              %s\n", resp.Details.Family)
+					}
+					fmt.Printf("Parameter Size      %s\n", resp.Details.ParameterSize)
+					fmt.Printf("Quantization Level  %s\n", resp.Details.QuantizationLevel)
+					fmt.Println("")
+				case "license":
+					if resp.License == "" {
+						fmt.Println("No license was specified for this model.")
+					} else {
+						fmt.Println(resp.License)
+					}
+				case "modelfile":
+					fmt.Println(resp.Modelfile)
+				case "parameters":
+					if resp.Parameters == "" {
+						fmt.Println("No parameters were specified for this model.")
+					} else {
+						if len(opts.Options) > 0 {
+							fmt.Println("User defined parameters:")
+							for k, v := range opts.Options {
+								fmt.Printf("%-*s %v\n", 30, k, v)
+							}
+							fmt.Println()
+						}
+						fmt.Println("Model defined parameters:")
+						fmt.Println(resp.Parameters)
+					}
+				case "system":
+					switch {
+					case opts.System != "":
+						fmt.Println(opts.System + "\n")
+					case resp.System != "":
+						fmt.Println(resp.System + "\n")
+					default:
+						fmt.Println("No system message was specified for this model.")
+					}
+				case "template":
+					switch {
+					case opts.Template != "":
+						fmt.Println(opts.Template + "\n")
+					case resp.Template != "":
+						fmt.Println(resp.Template)
+					default:
+						fmt.Println("No prompt template was specified for this model.")
+					}
+				default:
+					fmt.Printf("Unknown command '/show %s'. Type /? for help\n", args[1])
+				}
+			} else {
+				usageShow()
+			}
+		case strings.HasPrefix(line, "/help"), strings.HasPrefix(line, "/?"):
+			args := strings.Fields(line)
+			if len(args) > 1 {
+				switch args[1] {
+				case "set", "/set":
+					usageSet()
+				case "show", "/show":
+					usageShow()
+				case "shortcut", "shortcuts":
+					usageShortcuts()
+				}
+			} else {
+				usage()
+			}
+		case strings.HasPrefix(line, "/exit"), strings.HasPrefix(line, "/bye"):
+			return nil
+		case strings.HasPrefix(line, "/"):
+			args := strings.Fields(line)
+			isFile := false
+
+			if opts.MultiModal {
+				for _, f := range extractFileNames(line) {
+					if strings.HasPrefix(f, args[0]) {
+						isFile = true
+						break
+					}
+				}
+			}
+
+			if !isFile {
+				fmt.Printf("Unknown command '%s'. Type /? for help\n", args[0])
+				continue
+			}
+
+			sb.WriteString(line)
+		default:
+			sb.WriteString(line)
+		}
+
+		if sb.Len() > 0 && multiline == MultilineNone {
+			newMessage := api.Message{Role: "user", Content: sb.String()}
+
+			if opts.MultiModal {
+				msg, images, err := extractFileData(sb.String())
+				if err != nil {
+					return err
+				}
+
+				// clear all previous images for better responses
+				if len(images) > 0 {
+					for i := range opts.Messages {
+						opts.Messages[i].Images = nil
+					}
+				}
+
+				newMessage.Content = msg
+				newMessage.Images = images
+			}
+
+			opts.Messages = append(opts.Messages, newMessage)
+
+			assistant, err := chat(cmd, opts)
+			if err != nil {
+				return err
+			}
+			if assistant != nil {
+				opts.Messages = append(opts.Messages, *assistant)
+			}
+
+			sb.Reset()
+		}
+	}
+}
+
+func buildModelfile(opts runOptions) string {
+	var mf strings.Builder
+	model := opts.ParentModel
+	if model == "" {
+		model = opts.Model
+	}
+	fmt.Fprintf(&mf, "FROM %s\n", model)
+	if opts.System != "" {
+		fmt.Fprintf(&mf, "SYSTEM \"\"\"%s\"\"\"\n", opts.System)
+	}
+
+	if opts.Template != "" {
+		fmt.Fprintf(&mf, "TEMPLATE \"\"\"%s\"\"\"\n", opts.Template)
+	}
+
+	keys := make([]string, 0)
+	for k := range opts.Options {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+	for _, k := range keys {
+		fmt.Fprintf(&mf, "PARAMETER %s %v\n", k, opts.Options[k])
+	}
+	fmt.Fprintln(&mf)
+
+	for _, msg := range opts.Messages {
+		fmt.Fprintf(&mf, "MESSAGE %s \"\"\"%s\"\"\"\n", msg.Role, msg.Content)
+	}
+
+	return mf.String()
+}
+
+func normalizeFilePath(fp string) string {
+	// Define a map of escaped characters and their replacements
+	replacements := map[string]string{
+		"\\ ":  " ",  // Escaped space
+		"\\(":  "(",  // Escaped left parenthesis
+		"\\)":  ")",  // Escaped right parenthesis
+		"\\[":  "[",  // Escaped left square bracket
+		"\\]":  "]",  // Escaped right square bracket
+		"\\{":  "{",  // Escaped left curly brace
+		"\\}":  "}",  // Escaped right curly brace
+		"\\$":  "$",  // Escaped dollar sign
+		"\\&":  "&",  // Escaped ampersand
+		"\\;":  ";",  // Escaped semicolon
+		"\\'":  "'",  // Escaped single quote
+		"\\\\": "\\", // Escaped backslash
+		"\\*":  "*",  // Escaped asterisk
+		"\\?":  "?",  // Escaped question mark
+	}
+
+	for escaped, actual := range replacements {
+		fp = strings.ReplaceAll(fp, escaped, actual)
+	}
+	return fp
+}
+
+func extractFileNames(input string) []string {
+	// Regex to match file paths starting with optional drive letter, / ./ \ or .\ and include escaped or unescaped spaces (\ or %20)
+	// and followed by more characters and a file extension
+	// This will capture non filename strings, but we'll check for file existence to remove mismatches
+	regexPattern := `(?:[a-zA-Z]:)?(?:\./|/|\\)[\S\\ ]+?\.(?i:jpg|jpeg|png|svg)\b`
+	re := regexp.MustCompile(regexPattern)
+
+	return re.FindAllString(input, -1)
+}
+
+func extractFileData(input string) (string, []api.ImageData, error) {
+	filePaths := extractFileNames(input)
+	var imgs []api.ImageData
+
+	for _, fp := range filePaths {
+		nfp := normalizeFilePath(fp)
+		data, err := getImageData(nfp)
+		if err != nil {
+			if os.IsNotExist(err) {
+				continue
+			}
+			fmt.Fprintf(os.Stderr, "Couldn't process image: %q\n", err)
+			return "", imgs, err
+		}
+		fmt.Fprintf(os.Stderr, "Added image '%s'\n", nfp)
+		input = strings.ReplaceAll(input, fp, "")
+		imgs = append(imgs, data)
+	}
+	return input, imgs, nil
+}
+
+func getImageData(filePath string) ([]byte, error) {
+	file, err := os.Open(filePath)
+	if err != nil {
+		return nil, err
+	}
+	defer file.Close()
+
+	buf := make([]byte, 512)
+	_, err = file.Read(buf)
+	if err != nil {
+		return nil, err
+	}
+
+	contentType := http.DetectContentType(buf)
+	allowedTypes := []string{"image/jpeg", "image/jpg", "image/png"}
+	if !slices.Contains(allowedTypes, contentType) {
+		return nil, fmt.Errorf("invalid image type: %s", contentType)
+	}
+
+	info, err := file.Stat()
+	if err != nil {
+		return nil, err
+	}
+
+	// Check if the file size exceeds 100MB
+	var maxSize int64 = 100 * 1024 * 1024 // 100MB in bytes
+	if info.Size() > maxSize {
+		return nil, fmt.Errorf("file size exceeds maximum limit (100MB)")
+	}
+
+	buf = make([]byte, info.Size())
+	_, err = file.Seek(0, 0)
+	if err != nil {
+		return nil, err
+	}
+
+	_, err = io.ReadFull(file, buf)
+	if err != nil {
+		return nil, err
+	}
+
+	return buf, nil
+}

cmd/interactive_test.go

@@ -0,0 +1,116 @@
+package cmd
+
+import (
+	"bytes"
+	"testing"
+	"text/template"
+
+	"github.com/stretchr/testify/assert"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+func TestExtractFilenames(t *testing.T) {
+	// Unix style paths
+	input := ` some preamble 
+ ./relative\ path/one.png inbetween1 ./not a valid two.jpg inbetween2
+/unescaped space /three.jpeg inbetween3 /valid\ path/dir/four.png "./quoted with spaces/five.svg`
+	res := extractFileNames(input)
+	assert.Len(t, res, 5)
+	assert.Contains(t, res[0], "one.png")
+	assert.Contains(t, res[1], "two.jpg")
+	assert.Contains(t, res[2], "three.jpeg")
+	assert.Contains(t, res[3], "four.png")
+	assert.Contains(t, res[4], "five.svg")
+	assert.NotContains(t, res[4], '"')
+	assert.NotContains(t, res, "inbtween")
+
+	// Windows style paths
+	input = ` some preamble
+ c:/users/jdoe/one.png inbetween1 c:/program files/someplace/two.jpg inbetween2 
+ /absolute/nospace/three.jpeg inbetween3 /absolute/with space/four.png inbetween4
+./relative\ path/five.svg inbetween5 "./relative with/spaces/six.png inbetween6
+d:\path with\spaces\seven.svg inbetween7 c:\users\jdoe\eight.png inbetween8 
+ d:\program files\someplace\nine.png inbetween9 "E:\program files\someplace\ten.svg some ending
+`
+	res = extractFileNames(input)
+	assert.Len(t, res, 10)
+	assert.NotContains(t, res, "inbtween")
+	assert.Contains(t, res[0], "one.png")
+	assert.Contains(t, res[0], "c:")
+	assert.Contains(t, res[1], "two.jpg")
+	assert.Contains(t, res[1], "c:")
+	assert.Contains(t, res[2], "three.jpeg")
+	assert.Contains(t, res[3], "four.png")
+	assert.Contains(t, res[4], "five.svg")
+	assert.Contains(t, res[5], "six.png")
+	assert.Contains(t, res[6], "seven.svg")
+	assert.Contains(t, res[6], "d:")
+	assert.Contains(t, res[7], "eight.png")
+	assert.Contains(t, res[7], "c:")
+	assert.Contains(t, res[8], "nine.png")
+	assert.Contains(t, res[8], "d:")
+	assert.Contains(t, res[9], "ten.svg")
+	assert.Contains(t, res[9], "E:")
+}
+
+func TestModelfileBuilder(t *testing.T) {
+	opts := runOptions{
+		Model:    "hork",
+		System:   "You are part horse and part shark, but all hork. Do horklike things",
+		Template: "This is a template.",
+		Messages: []api.Message{
+			{Role: "user", Content: "Hey there hork!"},
+			{Role: "assistant", Content: "Yes it is true, I am half horse, half shark."},
+		},
+		Options: map[string]interface{}{},
+	}
+
+	opts.Options["temperature"] = 0.9
+	opts.Options["seed"] = 42
+	opts.Options["penalize_newline"] = false
+	opts.Options["stop"] = []string{"hi", "there"}
+
+	mf := buildModelfile(opts)
+	expectedModelfile := `FROM {{.Model}}
+SYSTEM """{{.System}}"""
+TEMPLATE """{{.Template}}"""
+PARAMETER penalize_newline false
+PARAMETER seed 42
+PARAMETER stop [hi there]
+PARAMETER temperature 0.9
+
+MESSAGE user """Hey there hork!"""
+MESSAGE assistant """Yes it is true, I am half horse, half shark."""
+`
+
+	tmpl, err := template.New("").Parse(expectedModelfile)
+	assert.Nil(t, err)
+
+	var buf bytes.Buffer
+	err = tmpl.Execute(&buf, opts)
+	assert.Nil(t, err)
+	assert.Equal(t, buf.String(), mf)
+
+	opts.ParentModel = "horseshark"
+	mf = buildModelfile(opts)
+	expectedModelfile = `FROM {{.ParentModel}}
+SYSTEM """{{.System}}"""
+TEMPLATE """{{.Template}}"""
+PARAMETER penalize_newline false
+PARAMETER seed 42
+PARAMETER stop [hi there]
+PARAMETER temperature 0.9
+
+MESSAGE user """Hey there hork!"""
+MESSAGE assistant """Yes it is true, I am half horse, half shark."""
+`
+
+	tmpl, err = template.New("").Parse(expectedModelfile)
+	assert.Nil(t, err)
+
+	var parentBuf bytes.Buffer
+	err = tmpl.Execute(&parentBuf, opts)
+	assert.Nil(t, err)
+	assert.Equal(t, parentBuf.String(), mf)
+}

cmd/start_darwin.go

@@ -0,0 +1,30 @@
+package cmd
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"os/exec"
+	"strings"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+func startApp(ctx context.Context, client *api.Client) error {
+	exe, err := os.Executable()
+	if err != nil {
+		return err
+	}
+	link, err := os.Readlink(exe)
+	if err != nil {
+		return err
+	}
+	if !strings.Contains(link, "Ollama.app") {
+		return fmt.Errorf("could not find ollama app")
+	}
+	path := strings.Split(link, "Ollama.app")
+	if err := exec.Command("/usr/bin/open", "-a", path[0]+"Ollama.app").Run(); err != nil {
+		return err
+	}
+	return waitForServer(ctx, client)
+}

cmd/start_default.go

@@ -0,0 +1,14 @@
+//go:build !windows && !darwin
+
+package cmd
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+func startApp(ctx context.Context, client *api.Client) error {
+	return fmt.Errorf("could not connect to ollama server, run 'ollama serve' to start it")
+}

cmd/start_windows.go

@@ -0,0 +1,58 @@
+package cmd
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"syscall"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+func startApp(ctx context.Context, client *api.Client) error {
+	// log.Printf("XXX Attempting to find and start ollama app")
+	AppName := "ollama app.exe"
+	exe, err := os.Executable()
+	if err != nil {
+		return err
+	}
+	appExe := filepath.Join(filepath.Dir(exe), AppName)
+	_, err = os.Stat(appExe)
+	if errors.Is(err, os.ErrNotExist) {
+		// Try the standard install location
+		localAppData := os.Getenv("LOCALAPPDATA")
+		appExe = filepath.Join(localAppData, "Ollama", AppName)
+		_, err := os.Stat(appExe)
+		if errors.Is(err, os.ErrNotExist) {
+			// Finally look in the path
+			appExe, err = exec.LookPath(AppName)
+			if err != nil {
+				return fmt.Errorf("could not locate ollama app")
+			}
+		}
+	}
+	// log.Printf("XXX attempting to start app %s", appExe)
+
+	cmd_path := "c:\\Windows\\system32\\cmd.exe"
+	cmd := exec.Command(cmd_path, "/c", appExe)
+	// TODO - these hide flags aren't working - still pops up a command window for some reason
+	cmd.SysProcAttr = &syscall.SysProcAttr{CreationFlags: 0x08000000, HideWindow: true}
+
+	// TODO this didn't help either...
+	cmd.Stdin = strings.NewReader("")
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+
+	if err := cmd.Start(); err != nil {
+		return fmt.Errorf("unable to start ollama app %w", err)
+	}
+
+	if cmd.Process != nil {
+		defer cmd.Process.Release() //nolint:errcheck
+	}
+	return waitForServer(ctx, client)
+}

docs/README.md

@@ -10,9 +10,9 @@ Create new models or modify models already in the library using the Modelfile. L
 
 Import models using source model weights found on Hugging Face and similar sites by referring to the **[Import Documentation](./import.md)**.
 
-Installing on Linux in most cases is easy using the script on Ollama.ai. To get more detail about the install, including CUDA drivers, see the **[Linux Documentation](./linux.md)**.
+Installing on Linux in most cases is easy using the script on [ollama.com/download](ollama.com/download). To get more detail about the install, including CUDA drivers, see the **[Linux Documentation](./linux.md)**.
 
-Many of our users like the flexibility of using our official Docker Image. Learn more about using Docker with Ollama using the **[Docker Documentation](./docker.md)**.
+Many of our users like the flexibility of using our official Docker Image. Learn more about using Docker with Ollama using the **[Docker Documentation](https://hub.docker.com/r/ollama/ollama)**.
 
 It is easy to install on Linux and Mac, but many users will choose to build Ollama on their own. To do this, refer to the **[Development Documentation](./development.md)**.
 

docs/api.md

@@ -49,7 +49,8 @@ Advanced parameters (optional):
 - `template`: the prompt template to use (overrides what is defined in the `Modelfile`)
 - `context`: the context parameter returned from a previous request to `/generate`, this can be used to keep a short conversational memory
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
-- `raw`: if `true` no formatting will be applied to the prompt. You may choose to use the `raw` parameter if you are specifying a full templated prompt in your request to the API.
+- `raw`: if `true` no formatting will be applied to the prompt. You may choose to use the `raw` parameter if you are specifying a full templated prompt in your request to the API
+- `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
 #### JSON mode
 
@@ -246,6 +247,23 @@ curl http://localhost:11434/api/generate -d '{
 }'
 ```
 
+#### Request (Reproducible outputs)
+
+For reproducible outputs, set `temperature` to 0 and `seed` to a number:
+
+##### Request
+
+```shell
+curl http://localhost:11434/api/generate -d '{
+  "model": "mistral",
+  "prompt": "[INST] why is the sky blue? [/INST]",
+  "options": {
+    "seed": 101,
+    "temperature": 0
+  }
+}'
+```
+
 ##### Response
 
 ```json
@@ -379,6 +397,7 @@ Advanced parameters (optional):
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `template`: the prompt template to use (overrides what is defined in the `Modelfile`)
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
+- `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
 ### Examples
 
@@ -409,7 +428,7 @@ A stream of JSON objects is returned:
   "model": "llama2",
   "created_at": "2023-08-04T08:52:19.385406455-07:00",
   "message": {
-    "role": "assisant",
+    "role": "assistant",
     "content": "The",
     "images": null
   },
@@ -505,7 +524,7 @@ A stream of JSON objects is returned:
   "model": "llama2",
   "created_at": "2023-08-04T08:52:19.385406455-07:00",
   "message": {
-    "role": "assisant",
+    "role": "assistant",
     "content": "The"
   },
   "done": false
@@ -542,7 +561,7 @@ curl http://localhost:11434/api/chat -d '{
       "role": "user",
       "content": "what is in this image?",
       "images": ["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"]
-    },
+    }
   ]
 }'
 ```
@@ -568,6 +587,46 @@ curl http://localhost:11434/api/chat -d '{
 }
 ```
 
+#### Chat request (Reproducible outputs)
+
+##### Request
+
+```shell
+curl http://localhost:11434/api/chat -d '{
+  "model": "llama2",
+  "messages": [
+    {
+      "role": "user",
+      "content": "Hello!"
+    }
+  ],
+  "options": {
+    "seed": 101,
+    "temperature": 0
+  }
+}'
+```
+
+##### Response
+
+```json
+{
+  "model": "registry.ollama.ai/library/llama2:latest",
+  "created_at": "2023-12-12T14:13:43.416799Z",
+  "message": {
+    "role": "assistant",
+    "content": "Hello! How are you today?"
+  },
+  "done": true,
+  "total_duration": 5191566416,
+  "load_duration": 2154458,
+  "prompt_eval_count": 26,
+  "prompt_eval_duration": 383809000,
+  "eval_count": 298,
+  "eval_duration": 4799921000
+}
+```
+
 ## Create a Model
 
 ```shell
@@ -958,6 +1017,7 @@ Generate embeddings from a model
 Advanced parameters:
 
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
 ### Examples
 

docs/development.md

@@ -1,13 +1,9 @@
 # Development
 
-- Install cmake or (optionally, required tools for GPUs)
-- run `go generate ./...`
-- run `go build .`
-
 Install required tools:
 
 - cmake version 3.24 or higher
-- go version 1.20 or higher
+- go version 1.21 or higher
 - gcc version 11.4.0 or higher
 
 ```bash
@@ -17,7 +13,11 @@ brew install go cmake gcc
 Optionally enable debugging and more verbose logging:
 
 ```bash
+# At build time
 export CGO_CFLAGS="-g"
+
+# At runtime
+export OLLAMA_DEBUG=1
 ```
 
 Get the required libraries and build the native LLM code:
@@ -38,37 +38,101 @@ Now you can run `ollama`:
 ./ollama
 ```
 
-## Building on Linux with GPU support
+### Linux
 
+#### Linux CUDA (NVIDIA)
 
-### Linux/Windows CUDA (NVIDIA)
 *Your operating system distribution may already have packages for NVIDIA CUDA. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!*
 
-Install `cmake` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads) development and runtime packages.
+Install `cmake` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads)
+development and runtime packages. 
+
+Typically the build scripts will auto-detect CUDA, however, if your Linux distro
+or installation approach uses unusual paths, you can specify the location by
+specifying an environment variable `CUDA_LIB_DIR` to the location of the shared
+libraries, and `CUDACXX` to the location of the nvcc compiler.  You can customize
+set set of target CUDA architectues by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")
+
 Then generate dependencies:
+
 ```
 go generate ./...
 ```
+
 Then build the binary:
+
 ```
 go build .
 ```
 
-### Linux ROCm (AMD)
+#### Linux ROCm (AMD)
+
 *Your operating system distribution may already have packages for AMD ROCm and CLBlast. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!*
 
 Install [CLBlast](https://github.com/CNugteren/CLBlast/blob/master/doc/installation.md) and [ROCm](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html) developement packages first, as well as `cmake` and `golang`.
-Adjust the paths below (correct for Arch) as appropriate for your distributions install locations and generate dependencies:
+
+Typically the build scripts will auto-detect ROCm, however, if your Linux distro
+or installation approach uses unusual paths, you can specify the location by
+specifying an environment variable `ROCM_PATH` to the location of the ROCm
+install (typically `/opt/rocm`), and `CLBlast_DIR` to the location of the
+CLBlast install (typically `/usr/lib/cmake/CLBlast`).  You can also customize
+the AMD GPU targets by setting AMDGPU_TARGETS (e.g. `AMDGPU_TARGETS="gfx1101;gfx1102"`)
+
 ```
-CLBlast_DIR=/usr/lib/cmake/CLBlast ROCM_PATH=/opt/rocm go generate ./...
+go generate ./...
 ```
+
 Then build the binary:
+
 ```
 go build .
 ```
 
 ROCm requires elevated privileges to access the GPU at runtime.  On most distros you can add your user account to the `render` group, or run as root.
 
-## Containerized Build
+#### Advanced CPU Settings
 
-If you have Docker available, you can build linux binaries with `./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included.
+By default, running `go generate ./...` will compile a few different variations
+of the LLM library based on common CPU families and vector math capabilities,
+including a lowest-common-denominator which should run on almost any 64 bit CPU
+somewhat slowly.  At runtime, Ollama will auto-detect the optimal variation to
+load.  If you would like to build a CPU-based build customized for your
+processor, you can set `OLLAMA_CUSTOM_CPU_DEFS` to the llama.cpp flags you would
+like to use.  For example, to compile an optimized binary for an Intel i9-9880H,
+you might use:
+
+```
+OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go generate ./...
+go build .
+```
+
+#### Containerized Linux Build
+
+If you have Docker available, you can build linux binaries with `./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included.  The resulting binary is placed in `./dist`
+
+
+### Windows
+
+Note: The windows build for Ollama is still under development.
+
+Install required tools:
+
+- MSVC toolchain - C/C++ and cmake as minimal requirements
+- go version 1.21 or higher
+- MinGW (pick one variant) with GCC.
+  - <https://www.mingw-w64.org/>
+  - <https://www.msys2.org/>
+
+```powershell
+$env:CGO_ENABLED="1"
+
+go generate ./...
+
+go build .
+```
+
+#### Windows CUDA (NVIDIA)
+
+In addition to the common Windows development tools described above, install:
+
+- [NVIDIA CUDA](https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html)

docs/faq.md

@@ -2,69 +2,116 @@
 
 ## How can I upgrade Ollama?
 
-To upgrade Ollama, run the installation process again. On the Mac, click the Ollama icon in the menubar and choose the restart option if an update is available.
+Ollama on macOS and Windows will automatically download updates. Click on the taskbar or menubar item and then click "Restart to update" to apply the update. Updates can also be installed by downloading the latest version [manually](https://ollama.com/download/).
+
+On Linux, re-run the install script:
+
+```
+curl -fsSL https://ollama.com/install.sh | sh
+```
 
 ## How can I view the logs?
 
 Review the [Troubleshooting](./troubleshooting.md) docs for more about using logs.
 
-## How do I use Ollama server environment variables on Mac
+## How can I specify the context window size?
 
-On macOS, Ollama runs in the background and is managed by the menubar app. If adding environment variables, Ollama will need to be run manually.
+By default, Ollama uses a context window size of 2048 tokens.
 
-1. Click the menubar icon for Ollama and choose **Quit Ollama**.
-2. Open a new terminal window and run the following command (this example uses `OLLAMA_HOST` with an IP address of `123.1.1.1`):
+To change this when using `ollama run`, use `/set parameter`:
 
-   ```bash
-   OLLAMA_HOST=123.1.1.1 ollama serve
-   ```
+```
+/set parameter num_ctx 4096
+```
 
-## How do I use Ollama server environment variables on Linux?
+When using the API, specify the `num_ctx` parameter:
 
-If Ollama is installed with the install script, a systemd service was created, running as the Ollama user. To add an environment variable, such as OLLAMA_HOST, follow these steps:
+```
+curl http://localhost:11434/api/generate -d '{
+  "model": "llama2",
+  "prompt": "Why is the sky blue?",
+  "options": {
+    "num_ctx": 4096
+  }
+}'
+```
 
-1. Create a `systemd` drop-in directory and add a config file. This is only needed once.
+## How do I configure Ollama server?
 
-   ```bash
-   mkdir -p /etc/systemd/system/ollama.service.d
-   echo '[Service]' >>/etc/systemd/system/ollama.service.d/environment.conf
-   ```
+Ollama server can be configured with environment variables.
 
-2. For each environment variable, add it to the config file:
+### Setting environment variables on Mac
 
-   ```bash
-   echo 'Environment="OLLAMA_HOST=0.0.0.0:11434"' >>/etc/systemd/system/ollama.service.d/environment.conf
-   ```
+If Ollama is run as a macOS application, environment variables should be set using `launchctl`:
 
-3. Reload `systemd` and restart Ollama:
+1. For each environment variable, call `launchctl setenv`.
+
+    ```bash
+    launchctl setenv OLLAMA_HOST "0.0.0.0"
+    ```
+
+2. Restart Ollama application.
+
+### Setting environment variables on Linux
+
+If Ollama is run as a systemd service, environment variables should be set using `systemctl`:
+
+1. Edit the systemd service by calling `systemctl edit ollama.service`. This will open an editor.
+
+2. For each environment variable, add a line `Environment` under section `[Service]`:
+
+    ```ini
+    [Service]
+    Environment="OLLAMA_HOST=0.0.0.0"
+    ```
+
+3. Save and exit.
+
+4. Reload `systemd` and restart Ollama:
 
    ```bash
    systemctl daemon-reload
    systemctl restart ollama
    ```
 
+### Setting environment variables on Windows
+
+On windows, Ollama inherits your user and system environment variables.
+
+1. First Quit Ollama by clicking on it in the task bar
+
+2. Edit system environment variables from the control panel
+
+3. Edit or create New variable(s) for your user account for `OLLAMA_HOST`, `OLLAMA_MODELS`, etc.
+
+4. Click OK/Apply to save 
+
+5. Run `ollama` from a new terminal window 
+
+
 ## How can I expose Ollama on my network?
 
-Ollama binds to 127.0.0.1 port 11434 by default. Change the bind address with the `OLLAMA_HOST` environment variable. Refer to the section above for how to use environment variables on your platform.
+Ollama binds 127.0.0.1 port 11434 by default. Change the bind address with the `OLLAMA_HOST` environment variable.
+
+Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
 
 ## How can I allow additional web origins to access Ollama?
 
-Ollama allows cross-origin requests from `127.0.0.1` and `0.0.0.0` by default. Add additional origins with the `OLLAMA_ORIGINS` environment variable. For example, to add all ports on 192.168.1.1 and https://example.com, use:
+Ollama allows cross-origin requests from `127.0.0.1` and `0.0.0.0` by default. Additional origins can be configured with `OLLAMA_ORIGINS`.
 
-```shell
-OLLAMA_ORIGINS=http://192.168.1.1:*,https://example.com
-```
-
-Refer to the section above for how to use environment variables on your platform.
+Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
 
 ## Where are models stored?
 
-- macOS: `~/.ollama/models`.
+- macOS: `~/.ollama/models`
 - Linux: `/usr/share/ollama/.ollama/models`
+- Windows: `C:\Users\<username>\.ollama\models`
 
-## How do I set them to a different location?
+### How do I set them to a different location?
 
-If a different directory needs to be used, set the environment variable `OLLAMA_MODELS` to the chosen directory. Refer to the section above for how to use environment variables on your platform.
+If a different directory needs to be used, set the environment variable `OLLAMA_MODELS` to the chosen directory.
+
+Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
 
 ## Does Ollama send my prompts and answers back to Ollama.ai to use in any way?
 
@@ -112,3 +159,37 @@ This can impact both installing Ollama, as well as downloading models.
 Open `Control Panel > Networking and Internet > View network status and tasks` and click on `Change adapter settings` on the left panel. Find the `vEthernel (WSL)` adapter, right click and select `Properties`.
 Click on `Configure` and open the `Advanced` tab. Search through each of the properties until you find `Large Send Offload Version 2 (IPv4)` and `Large Send Offload Version 2 (IPv6)`. *Disable* both of these
 properties.
+
+## How can I pre-load a model to get faster response times?
+
+If you are using the API you can preload a model by sending the Ollama server an empty request. This works with both the `/api/generate` and `/api/chat` API endpoints.
+
+To preload the mistral model using the generate endpoint, use:
+```shell
+curl http://localhost:11434/api/generate -d '{"model": "mistral"}'
+```
+
+To use the chat completions endpoint, use:
+```shell
+curl http://localhost:11434/api/chat -d '{"model": "mistral"}'
+```
+
+## How do I keep a model loaded in memory or make it unload immediately?
+
+By default models are kept in memory for 5 minutes before being unloaded. This allows for quicker response times if you are making numerous requests to the LLM. You may, however, want to free up the memory before the 5 minutes have elapsed or keep the model loaded indefinitely. Use the `keep_alive` parameter with either the `/api/generate` and `/api/chat` API endpoints to control how long the model is left in memory.
+
+The `keep_alive` parameter can be set to:
+* a duration string (such as "10m" or "24h")
+* a number in seconds (such as 3600)
+* any negative number which will keep the model loaded in memory (e.g. -1 or "-1m")
+* '0' which will unload the model immediately after generating a response
+
+For example, to preload a model and leave it in memory use:
+```shell
+curl http://localhost:11434/api/generate -d '{"model": "llama2", "keep_alive": -1}'
+```
+
+To unload the model and free up memory use:
+```shell
+curl http://localhost:11434/api/generate -d '{"model": "llama2", "keep_alive": 0}'
+```

docs/import.md

@@ -15,7 +15,7 @@ FROM ./mistral-7b-v0.1.Q4_0.gguf
 (Optional) many chat models require a prompt template in order to answer correctly. A default prompt template can be specified with the `TEMPLATE` instruction in the `Modelfile`:
 
 ```
-FROM ./q4_0.bin
+FROM ./mistral-7b-v0.1.Q4_0.gguf
 TEMPLATE "[INST] {{ .Prompt }} [/INST]"
 ```
 
@@ -37,55 +37,69 @@ ollama run example "What is your favourite condiment?"
 
 ## Importing (PyTorch & Safetensors)
 
-### Supported models
+> Importing from PyTorch and Safetensors is a longer process than importing from GGUF. Improvements that make it easier are a work in progress.
 
-Ollama supports a set of model architectures, with support for more coming soon:
+### Setup
 
-- Llama & Mistral
-- Falcon & RW
-- BigCode
+First, clone the `ollama/ollama` repo:
 
-To view a model's architecture, check the `config.json` file in its HuggingFace repo. You should see an entry under `architectures` (e.g. `LlamaForCausalLM`).
+```
+git clone git@github.com:ollama/ollama.git ollama
+cd ollama
+```
 
-### Step 1: Clone the HuggingFace repository (optional)
+and then fetch its `llama.cpp` submodule:
+
+```shell
+git submodule init
+git submodule update llm/llama.cpp
+```
+
+Next, install the Python dependencies:
+
+```
+python3 -m venv llm/llama.cpp/.venv
+source llm/llama.cpp/.venv/bin/activate
+pip install -r llm/llama.cpp/requirements.txt
+```
+
+Then build the `quantize` tool:
+
+```
+make -C llm/llama.cpp quantize
+```
+
+### Clone the HuggingFace repository (optional)
 
 If the model is currently hosted in a HuggingFace repository, first clone that repository to download the raw model.
 
+Install [Git LFS](https://docs.github.com/en/repositories/working-with-files/managing-large-files/installing-git-large-file-storage), verify it's installed, and then clone the model's repository:
+
 ```
 git lfs install
-git clone https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1
-cd Mistral-7B-Instruct-v0.1
+git clone https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1 model
 ```
 
-### Step 2: Convert and quantize to a `.bin` file (optional, for PyTorch and Safetensors)
+### Convert the model
 
-If the model is in PyTorch or Safetensors format, a [Docker image](https://hub.docker.com/r/ollama/quantize) with the tooling required to convert and quantize models is available.
-
-First, Install [Docker](https://www.docker.com/get-started/).
-
-Next, to convert and quantize your model, run:
+> Note: some model architectures require using specific convert scripts. For example, Qwen models require running `convert-hf-to-gguf.py` instead of `convert.py`
 
 ```
-docker run --rm -v .:/model ollama/quantize -q q4_0 /model
+python llm/llama.cpp/convert.py ./model --outtype f16 --outfile converted.bin
 ```
 
-This will output two files into the directory:
+### Quantize the model
 
-- `f16.bin`: the model converted to GGUF
-- `q4_0.bin` the model quantized to a 4-bit quantization (Ollama will use this file to create the Ollama model)
+```
+llm/llama.cpp/quantize converted.bin quantized.bin q4_0
+```
 
 ### Step 3: Write a `Modelfile`
 
 Next, create a `Modelfile` for your model:
 
 ```
-FROM ./q4_0.bin
-```
-
-(Optional) many chat models require a prompt template in order to answer correctly. A default prompt template can be specified with the `TEMPLATE` instruction in the `Modelfile`:
-
-```
-FROM ./q4_0.bin
+FROM quantized.bin
 TEMPLATE "[INST] {{ .Prompt }} [/INST]"
 ```
 
@@ -109,9 +123,9 @@ ollama run example "What is your favourite condiment?"
 
 Publishing models is in early alpha. If you'd like to publish your model to share with others, follow these steps:
 
-1. Create [an account](https://ollama.ai/signup)
-2. Run `cat ~/.ollama/id_ed25519.pub` to view your Ollama public key. Copy this to the clipboard.
-3. Add your public key to your [Ollama account](https://ollama.ai/settings/keys)
+1. Create [an account](https://ollama.com/signup)
+2. Run `cat ~/.ollama/id_ed25519.pub` (or `type %USERPROFILE%\.ollama\id_ed25519.pub` on Windows) to view your Ollama public key. Copy this to the clipboard.
+3. Add your public key to your [Ollama account](https://ollama.com/settings/keys)
 
 Next, copy your model to your username's namespace:
 
@@ -125,7 +139,7 @@ Then push the model:
 ollama push <your username>/example
 ```
 
-After publishing, your model will be available at `https://ollama.ai/<your username>/example`.
+After publishing, your model will be available at `https://ollama.com/<your username>/example`.
 
 ## Quantization reference
 
@@ -149,47 +163,3 @@ The quantization options are as follow (from highest highest to lowest levels of
 - `q6_K`
 - `q8_0`
 - `f16`
-
-## Manually converting & quantizing models
-
-### Prerequisites
-
-Start by cloning the `llama.cpp` repo to your machine in another directory:
-
-```
-git clone https://github.com/ggerganov/llama.cpp.git
-cd llama.cpp
-```
-
-Next, install the Python dependencies:
-
-```
-pip install -r requirements.txt
-```
-
-Finally, build the `quantize` tool:
-
-```
-make quantize
-```
-
-### Convert the model
-
-Run the correct conversion script for your model architecture:
-
-```shell
-# LlamaForCausalLM or MistralForCausalLM
-python convert.py <path to model directory>
-
-# FalconForCausalLM
-python convert-falcon-hf-to-gguf.py <path to model directory>
-
-# GPTBigCodeForCausalLM
-python convert-starcoder-hf-to-gguf.py <path to model directory>
-```
-
-### Quantize the model
-
-```
-quantize <path to model dir>/ggml-model-f32.bin <path to model dir>/q4_0.bin q4_0
-```

docs/linux.md

@@ -3,9 +3,11 @@
 ## Install
 
 Install Ollama running this one-liner:
+
 >
+
 ```bash
-curl https://ollama.ai/install.sh | sh
+curl -fsSL https://ollama.com/install.sh | sh
 ```
 
 ## Manual install
@@ -15,7 +17,7 @@ curl https://ollama.ai/install.sh | sh
 Ollama is distributed as a self-contained binary. Download it to a directory in your PATH:
 
 ```bash
-sudo curl -L https://ollama.ai/download/ollama-linux-amd64 -o /usr/bin/ollama
+sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama
 sudo chmod +x /usr/bin/ollama
 ```
 
@@ -75,13 +77,13 @@ sudo systemctl start ollama
 Update ollama by running the install script again:
 
 ```bash
-curl https://ollama.ai/install.sh | sh
+curl -fsSL https://ollama.com/install.sh | sh
 ```
 
 Or by downloading the ollama binary:
 
 ```bash
-sudo curl -L https://ollama.ai/download/ollama-linux-amd64 -o /usr/bin/ollama
+sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama
 sudo chmod +x /usr/bin/ollama
 ```
 
@@ -109,8 +111,10 @@ Remove the ollama binary from your bin directory (either `/usr/local/bin`, `/usr
 sudo rm $(which ollama)
 ```
 
-Remove the downloaded models and Ollama service user:
+Remove the downloaded models and Ollama service user and group:
+
 ```bash
 sudo rm -r /usr/share/ollama
 sudo userdel ollama
+sudo groupdel ollama
 ```

docs/modelfile.md

@@ -19,6 +19,7 @@ A model file is the blueprint to create and share models with Ollama.
   - [SYSTEM](#system)
   - [ADAPTER](#adapter)
   - [LICENSE](#license)
+  - [MESSAGE](#message)
 - [Notes](#notes)
 
 ## Format
@@ -38,6 +39,7 @@ INSTRUCTION arguments
 | [`SYSTEM`](#system)                 | Specifies the system message that will be set in the template. |
 | [`ADAPTER`](#adapter)               | Defines the (Q)LoRA adapters to apply to the model.            |
 | [`LICENSE`](#license)               | Specifies the legal license.                                   |
+| [`MESSAGE`](#message)               | Specify message history.                                       |
 
 ## Examples
 
@@ -65,13 +67,13 @@ To use this:
 
 More examples are available in the [examples directory](../examples).
 
-### `Modelfile`s in [ollama.ai/library][1]
+### `Modelfile`s in [ollama.com/library][1]
 
-There are two ways to view `Modelfile`s underlying the models in [ollama.ai/library][1]:
+There are two ways to view `Modelfile`s underlying the models in [ollama.com/library][1]:
 
 - Option 1: view a details page from a model's tags page:
-  1.  Go to a particular model's tags (e.g. https://ollama.ai/library/llama2/tags)
-  2.  Click on a tag (e.g. https://ollama.ai/library/llama2:13b)
+  1.  Go to a particular model's tags (e.g. https://ollama.com/library/llama2/tags)
+  2.  Click on a tag (e.g. https://ollama.com/library/llama2:13b)
   3.  Scroll down to "Layers"
       - Note: if the [`FROM` instruction](#from-required) is not present,
         it means the model was created from a local file
@@ -84,7 +86,7 @@ There are two ways to view `Modelfile`s underlying the models in [ollama.ai/libr
   # FROM llama2:13b
 
   FROM /root/.ollama/models/blobs/sha256:123abc
-  TEMPLATE """[INST] {{ if and .First .System }}<<SYS>>{{ .System }}<</SYS>>
+  TEMPLATE """[INST] {{ if .System }}<<SYS>>{{ .System }}<</SYS>>
 
   {{ end }}{{ .Prompt }} [/INST] """
   SYSTEM """"""
@@ -152,30 +154,23 @@ PARAMETER <parameter> <parametervalue>
 
 ### TEMPLATE
 
-`TEMPLATE` of the full prompt template to be passed into the model. It may include (optionally) a system message and a user's prompt. This is used to create a full custom prompt, and syntax may be model specific. You can usually find the template for a given model in the readme for that model.
+`TEMPLATE` of the full prompt template to be passed into the model. It may include (optionally) a system message, a user's message and the response from the model. Note: syntax may be model specific. Templates use Go [template syntax](https://pkg.go.dev/text/template).
 
 #### Template Variables
 
-| Variable        | Description                                                                                                   |
-| --------------- | ------------------------------------------------------------------------------------------------------------- |
-| `{{ .System }}` | The system message used to specify custom behavior, this must also be set in the Modelfile as an instruction. |
-| `{{ .Prompt }}` | The incoming prompt, this is not specified in the model file and will be set based on input.                  |
-| `{{ .First }}`  | A boolean value used to render specific template information for the first generation of a session.           |
+| Variable          | Description                                                                                   |
+| ----------------- | --------------------------------------------------------------------------------------------- |
+| `{{ .System }}`   | The system message used to specify custom behavior.                                           |
+| `{{ .Prompt }}`   | The user prompt message.                                                                      |
+| `{{ .Response }}` | The response from the model. When generating a response, text after this variable is omitted. |
 
-```modelfile
-TEMPLATE """
-{{- if .First }}
-### System:
-{{ .System }}
-{{- end }}
-
-### User:
-{{ .Prompt }}
-
-### Response:
+```
+TEMPLATE """{{ if .System }}<|im_start|>system
+{{ .System }}<|im_end|>
+{{ end }}{{ if .Prompt }}<|im_start|>user
+{{ .Prompt }}<|im_end|>
+{{ end }}<|im_start|>assistant
 """
-
-SYSTEM """<system message>"""
 ```
 
 ### SYSTEM
@@ -204,9 +199,22 @@ LICENSE """
 """
 ```
 
+### MESSAGE
+
+The `MESSAGE` instruction allows you to specify a message history for the model to use when responding:
+
+```modelfile
+MESSAGE user Is Toronto in Canada?
+MESSAGE assistant yes
+MESSAGE user Is Sacramento in Canada?
+MESSAGE assistant no
+MESSAGE user Is Ontario in Canada?
+MESSAGE assistant yes
+```
+
 ## Notes
 
 - the **`Modelfile` is not case sensitive**. In the examples, uppercase instructions are used to make it easier to distinguish it from arguments.
 - Instructions can be in any order. In the examples, the `FROM` instruction is first to keep it easily readable.
 
-[1]: https://ollama.ai/library
+[1]: https://ollama.com/library

docs/openai.md

@@ -0,0 +1,141 @@
+# OpenAI compatibility
+
+> **Note:** OpenAI compatibility is experimental and is subject to major adjustments including breaking changes. For fully-featured access to the Ollama API, see the Ollama [Python library](https://github.com/ollama/ollama-python), [JavaScript library](https://github.com/ollama/ollama-js) and [REST API](https://github.com/jmorganca/ollama/blob/main/docs/api.md).
+
+Ollama provides experimental compatibility with parts of the [OpenAI API](https://platform.openai.com/docs/api-reference) to help connect existing applications to Ollama.
+
+## Usage
+
+### OpenAI Python library
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    base_url='http://localhost:11434/v1/',
+
+    # required but ignored
+    api_key='ollama',
+)
+
+chat_completion = client.chat.completions.create(
+    messages=[
+        {
+            'role': 'user',
+            'content': 'Say this is a test',
+        }
+    ],
+    model='llama2',
+)
+```
+
+### OpenAI JavaScript library
+
+```javascript
+import OpenAI from 'openai'
+
+const openai = new OpenAI({
+  baseURL: 'http://localhost:11434/v1/',
+
+  // required but ignored
+  apiKey: 'ollama',
+})
+
+const chatCompletion = await openai.chat.completions.create({
+  messages: [{ role: 'user', content: 'Say this is a test' }],
+  model: 'llama2',
+})
+```
+
+### `curl`
+
+```
+curl http://localhost:11434/v1/chat/completions \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "llama2",
+        "messages": [
+            {
+                "role": "system",
+                "content": "You are a helpful assistant."
+            },
+            {
+                "role": "user",
+                "content": "Hello!"
+            }
+        ]
+    }'
+```
+
+## Endpoints
+
+### `/v1/chat/completions`
+
+#### Supported features
+
+- [x] Chat completions
+- [x] Streaming
+- [x] JSON mode
+- [x] Reproducible outputs
+- [ ] Vision
+- [ ] Function calling
+- [ ] Logprobs
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `messages`
+  - [x] Text `content`
+  - [ ] Array of `content` parts
+- [x] `frequency_penalty`
+- [x] `presence_penalty`
+- [x] `response_format`
+- [x] `seed`
+- [x] `stop`
+- [x] `stream`
+- [x] `temperature`
+- [x] `top_p`
+- [x] `max_tokens`
+- [ ] `logit_bias`
+- [ ] `tools`
+- [ ] `tool_choice`
+- [ ] `user`
+- [ ] `n`
+
+#### Notes
+
+- Setting `seed` will always set `temperature` to `0`
+- `finish_reason` will always be `stop`
+- `usage.prompt_tokens` will be 0 for completions where prompt evaluation is cached
+
+## Models
+
+Before using a model, pull it locally `ollama pull`:
+
+```shell
+ollama pull llama2
+```
+
+### Default model names
+
+For tooling that relies on default OpenAI model names such as `gpt-3.5-turbo`, use `ollama cp` to copy an existing model name to a temporary name:
+
+```
+ollama cp llama2 gpt-3.5-turbo
+```
+
+Afterwards, this new model name can be specified the `model` field:
+
+```shell
+curl http://localhost:11434/v1/chat/completions \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "gpt-3.5-turbo",
+        "messages": [
+            {
+                "role": "user",
+                "content": "Hello!"
+            }
+        ]
+    }'
+```

docs/troubleshooting.md

@@ -1,22 +1,72 @@
-# How to troubleshoot issues
-
-Sometimes Ollama may not perform as expected. One of the best ways to figure out what happened is to take a look at the logs. Find the logs on Mac by running the command:
-
-```shell
-cat ~/.ollama/logs/server.log
-```
-
-On Linux systems with systemd, the logs can be found with this command:
-
-```shell
-journalctl -u ollama
-```
-
-If manually running `ollama serve` in a terminal, the logs will be on that terminal.
-
-Join the [Discord](https://discord.gg/ollama) for help interpreting the logs.
-
-## Known issues
-
-
-* `signal: illegal instruction (core dumped)`: Ollama requires AVX support from the CPU. This was introduced in 2011 and CPUs started offering it in 2012. CPUs from before that and some lower end CPUs after that may not have AVX support and thus are not supported by Ollama. Some users have had luck with building Ollama on their machines disabling the need for AVX.
+# How to troubleshoot issues
+
+Sometimes Ollama may not perform as expected. One of the best ways to figure out what happened is to take a look at the logs. Find the logs on **Mac** by running the command:
+
+```shell
+cat ~/.ollama/logs/server.log
+```
+
+On **Linux** systems with systemd, the logs can be found with this command:
+
+```shell
+journalctl -u ollama
+```
+
+When you run Ollama in a **container**, the logs go to stdout/stderr in the container:
+
+```shell
+docker logs <container-name>
+```
+(Use `docker ps` to find the container name)
+
+If manually running `ollama serve` in a terminal, the logs will be on that terminal.
+
+When you run Ollama on **Windows**, there are a few different locations.  You can view them in the explorer window by hitting `<cmd>+R` and type in:
+- `explorer %LOCALAPPDATA%\Ollama` to view logs
+- `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
+- `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
+- `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
+
+To enable additional debug logging to help troubleshoot problems, first **Quit the running app from the tray menu** then in a powershell terminal
+```powershell
+$env:OLLAMA_DEBUG="1"
+& "ollama app.exe"
+```
+
+Join the [Discord](https://discord.gg/ollama) for help interpreting the logs.
+
+## LLM libraries
+
+Ollama includes multiple LLM libraries compiled for different GPUs and CPU
+vector features.  Ollama tries to pick the best one based on the capabilities of
+your system.  If this autodetection has problems, or you run into other problems
+(e.g. crashes in your GPU) you can workaround this by forcing a specific LLM
+library.  `cpu_avx2` will perform the best, followed by `cpu_avx` an the slowest
+but most compatible is `cpu`.  Rosetta emulation under MacOS will work with the
+`cpu` library. 
+
+In the server log, you will see a message that looks something like this (varies
+from release to release):
+
+```
+Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v11 rocm_v5]
+```
+
+**Experimental LLM Library Override**
+
+You can set OLLAMA_LLM_LIBRARY to any of the available LLM libraries to bypass
+autodetection, so for example, if you have a CUDA card, but want to force the
+CPU LLM library with AVX2 vector support, use:
+
+```
+OLLAMA_LLM_LIBRARY="cpu_avx2" ollama serve
+```
+
+You can see what features your CPU has with the following.  
+```
+cat /proc/cpuinfo| grep flags  | head -1
+```
+
+## Known issues
+
+* N/A

docs/tutorials/nvidia-jetson.md

@@ -17,7 +17,7 @@ Prerequisites:
 
 Here are the steps:
 
-- Install Ollama via standard Linux command (ignore the 404 error): `curl https://ollama.ai/install.sh | sh`
+- Install Ollama via standard Linux command (ignore the 404 error): `curl https://ollama.com/install.sh | sh`
 - Stop the Ollama service: `sudo systemctl stop ollama`
 - Start Ollama serve in a tmux session called ollama_jetson and reference the CUDA libraries path: `tmux has-session -t ollama_jetson 2>/dev/null || tmux new-session -d -s ollama_jetson 
 'LD_LIBRARY_PATH=/usr/local/cuda/lib64 ollama serve'`

docs/windows.md

@@ -0,0 +1,46 @@
+# Ollama Windows Preview
+
+Welcome to the Ollama Windows preview.
+
+No more WSL required!
+
+Ollama now runs as a native Windows application, including NVIDIA GPU support.
+After installing Ollama Windows Preview, Ollama will run in the background and
+the `ollama` command line is available in `cmd`, `powershell` or your favorite
+terminal application. As usual the Ollama [api](./api.md) will be served on
+`http://localhost:11434`.
+
+As this is a preview release, you should expect a few bugs here and there.  If
+you run into a problem you can reach out on
+[Discord](https://discord.gg/ollama), or file an 
+[issue](https://github.com/ollama/ollama/issues).
+Logs will often be helpful in dianosing the problem (see
+[Troubleshooting](#troubleshooting) below)
+
+## System Requirements
+
+* Windows 10 or newer, Home or Pro
+* NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
+
+## API Access
+
+Here's a quick example showing API access from `powershell`
+```powershell
+(Invoke-WebRequest -method POST -Body '{"model":"llama2", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
+```
+
+## Troubleshooting
+
+While we're in preview, `OLLAMA_DEBUG` is always enabled, which adds
+a "view logs" menu item to the app, and increses logging for the GUI app and
+server.
+
+Ollama on Windows stores files in a few different locations.  You can view them in
+the explorer window by hitting `<cmd>+R` and type in:
+- `explorer %LOCALAPPDATA%\Ollama` contains logs, and downloaded updates
+    - *app.log* contains logs from the GUI application
+    - *server.log* contains the server logs
+    - *upgrade.log* contains log output for upgrades
+- `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
+- `explorer %HOMEPATH%\.ollama` contains models and configuration
+- `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories

examples/jupyter-notebook/ollama.ipynb

@@ -8,7 +8,7 @@
    "outputs": [],
    "source": [
     "# Download and run the Ollama Linux install script\n",
-    "!curl https://ollama.ai/install.sh | sh\n",
+    "!curl -fsSL https://ollama.com/install.sh | sh\n",
     "!command -v systemctl >/dev/null && sudo systemctl stop ollama"
    ]
   },

examples/kubernetes/README.md

@@ -2,28 +2,28 @@
 
 ## Prerequisites
 
-- Ollama: https://ollama.ai/download
+- Ollama: https://ollama.com/download
 - Kubernetes cluster. This example will use Google Kubernetes Engine.
 
 ## Steps
 
 1. Create the Ollama namespace, daemon set, and service
 
-    ```bash
-    kubectl apply -f cpu.yaml
-    ```
+   ```bash
+   kubectl apply -f cpu.yaml
+   ```
 
 1. Port forward the Ollama service to connect and use it locally
 
-    ```bash
-    kubectl -n ollama port-forward service/ollama 11434:80
-    ```
+   ```bash
+   kubectl -n ollama port-forward service/ollama 11434:80
+   ```
 
 1. Pull and run a model, for example `orca-mini:3b`
 
-    ```bash
-    ollama run orca-mini:3b
-    ```
+   ```bash
+   ollama run orca-mini:3b
+   ```
 
 ## (Optional) Hardware Acceleration
 

examples/langchain-python-rag-websummary/README.md

@@ -1,6 +1,6 @@
 # LangChain Web Summarization
 
-This example summarizes the website, [https://ollama.ai/blog/run-llama2-uncensored-locally](https://ollama.ai/blog/run-llama2-uncensored-locally)
+This example summarizes the website, [https://ollama.com/blog/run-llama2-uncensored-locally](https://ollama.com/blog/run-llama2-uncensored-locally)
 
 ## Running the Example
 

examples/langchain-python-rag-websummary/main.py

@@ -2,7 +2,7 @@ from langchain.llms import Ollama
 from langchain.document_loaders import WebBaseLoader
 from langchain.chains.summarize import load_summarize_chain
 
-loader = WebBaseLoader("https://ollama.ai/blog/run-llama2-uncensored-locally")
+loader = WebBaseLoader("https://ollama.com/blog/run-llama2-uncensored-locally")
 docs = loader.load()
 
 llm = Ollama(model="llama2")

examples/python-loganalysis/readme.md

@@ -40,13 +40,13 @@ You are a log file analyzer. You will receive a set of lines from a log file for
 """
 ```
 
-This model is available at https://ollama.ai/mattw/loganalyzer. You can customize it and add to your own namespace using the command `ollama create <namespace/modelname> -f <path-to-modelfile>` then `ollama push <namespace/modelname>`.
+This model is available at https://ollama.com/mattw/loganalyzer. You can customize it and add to your own namespace using the command `ollama create <namespace/modelname> -f <path-to-modelfile>` then `ollama push <namespace/modelname>`.
 
 Then loganalysis.py scans all the lines in the given log file and searches for the word 'error'. When the word is found, the 10 lines before and after are set as the prompt for a call to the Generate API.
 
 ```python
 data = {
-  "prompt": "\n".join(error_logs), 
+  "prompt": "\n".join(error_logs),
   "model": "mattw/loganalyzer"
 }
 ```

examples/typescript-mentors/README.md

@@ -29,9 +29,9 @@ You can also add your own character to be chosen at random when you ask a questi
    ```bash
    ollama pull stablebeluga2:70b-q4_K_M
    ```
-  
+
 2. Create a new character:
-  
+
    ```bash
    npm run charactergen "Lorne Greene"
    ```
@@ -41,15 +41,15 @@ You can also add your own character to be chosen at random when you ask a questi
 3. Now you can create a model with this command:
 
    ```bash
-   ollama create <YourNamespace>/lornegreene -f lornegreene/Modelfile
+   ollama create <username>/lornegreene -f lornegreene/Modelfile
    ```
 
-   `YourNamespace` is whatever name you set up when you signed up at [https://ollama.ai/signup](https://ollama.ai/signup).
+   `username` is whatever name you set up when you signed up at [https://ollama.com/signup](https://ollama.com/signup).
 
-4. To add this to your mentors, you will have to update the code as follows. On line 8 of `mentors.ts`, add an object to the array, replacing `<YourNamespace>` with the namespace you used above.
+4. To add this to your mentors, you will have to update the code as follows. On line 8 of `mentors.ts`, add an object to the array, replacing `<username>` with the username you used above.
 
    ```bash
-   {ns: "<YourNamespace>", char: "Lorne Greene"}
+   {ns: "<username>", char: "Lorne Greene"}
    ```
 
 ## Review the Code

go.mod

@@ -1,6 +1,6 @@
 module github.com/jmorganca/ollama
 
-go 1.20
+go 1.21
 
 require (
 	github.com/emirpasic/gods v1.18.1
@@ -12,12 +12,26 @@ require (
 )
 
 require (
+	github.com/containerd/console v1.0.3 // indirect
+	github.com/cratonica/2goarray v0.0.0-20190331194516-514510793eaa // indirect
 	github.com/davecgh/go-spew v1.1.1 // indirect
+	github.com/getlantern/context v0.0.0-20190109183933-c447772a6520 // indirect
+	github.com/getlantern/errors v0.0.0-20190325191628-abdb3e3e36f7 // indirect
+	github.com/getlantern/golog v0.0.0-20190830074920-4ef2e798c2d7 // indirect
+	github.com/getlantern/hex v0.0.0-20190417191902-c6586a6fe0b7 // indirect
+	github.com/getlantern/hidden v0.0.0-20190325191715-f02dbb02be55 // indirect
+	github.com/getlantern/ops v0.0.0-20190325191751-d70cb0d6f85f // indirect
+	github.com/getlantern/systray v1.2.2 // indirect
+	github.com/go-stack/stack v1.8.0 // indirect
+	github.com/google/uuid v1.0.0 // indirect
 	github.com/mattn/go-runewidth v0.0.14 // indirect
+	github.com/oxtoacart/bpool v0.0.0-20190530202638-03653db5a59c // indirect
+	github.com/pborman/uuid v1.2.1 // indirect
 	github.com/pmezard/go-difflib v1.0.0 // indirect
 	github.com/rivo/uniseg v0.2.0 // indirect
 )
 
+
 require (
 	github.com/bytedance/sonic v1.9.1 // indirect
 	github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311 // indirect
@@ -45,7 +59,7 @@ require (
 	golang.org/x/crypto v0.14.0
 	golang.org/x/exp v0.0.0-20230817173708-d852ddb80c63
 	golang.org/x/net v0.17.0 // indirect
-	golang.org/x/sys v0.13.0 // indirect
+	golang.org/x/sys v0.13.0
 	golang.org/x/term v0.13.0
 	golang.org/x/text v0.13.0 // indirect
 	google.golang.org/protobuf v1.30.0 // indirect

go.sum

@@ -4,7 +4,11 @@ github.com/bytedance/sonic v1.9.1/go.mod h1:i736AoUSYt75HyZLoJW9ERYxcy6eaN6h4BZX
 github.com/chenzhuoyu/base64x v0.0.0-20211019084208-fb5309c8db06/go.mod h1:DH46F32mSOjUmXrMHnKwZdA8wcEefY7UVqBKYGjpdQY=
 github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311 h1:qSGYFH7+jGhDF8vLC+iwCD4WpbV1EBDSzWkJODFLams=
 github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311/go.mod h1:b583jCggY9gE99b6G5LEC39OIiVsWj+R97kbl5odCEk=
+github.com/containerd/console v1.0.3 h1:lIr7SlA5PxZyMV30bDW0MGbiOPXwc63yRuCP0ARubLw=
+github.com/containerd/console v1.0.3/go.mod h1:7LqA/THxQ86k76b8c/EMSiaJ3h1eZkMkXar0TQ1gf3U=
 github.com/cpuguy83/go-md2man/v2 v2.0.2/go.mod h1:tgQtvFlXSQOSOSIRvRPT7W67SCa46tRHOmNcaadrF8o=
+github.com/cratonica/2goarray v0.0.0-20190331194516-514510793eaa h1:Wg+722vs7a2zQH5lR9QWYsVbplKeffaQFIs5FTdfNNo=
+github.com/cratonica/2goarray v0.0.0-20190331194516-514510793eaa/go.mod h1:6Arca19mRx58CA7OWEd7Wu1NpC1rd3uDnNs6s1pj/DI=
 github.com/creack/pty v1.1.9/go.mod h1:oKZEueFk5CKHvIhNR5MUki03XCEU+Q6VDXinZuGJ33E=
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
@@ -13,6 +17,20 @@ github.com/emirpasic/gods v1.18.1 h1:FXtiHYKDGKCW2KzwZKx0iC0PQmdlorYgdFG9jPXJ1Bc
 github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FMNAnJvWQ=
 github.com/gabriel-vasile/mimetype v1.4.2 h1:w5qFW6JKBz9Y393Y4q372O9A7cUSequkh1Q7OhCmWKU=
 github.com/gabriel-vasile/mimetype v1.4.2/go.mod h1:zApsH/mKG4w07erKIaJPFiX0Tsq9BFQgN3qGY5GnNgA=
+github.com/getlantern/context v0.0.0-20190109183933-c447772a6520 h1:NRUJuo3v3WGC/g5YiyF790gut6oQr5f3FBI88Wv0dx4=
+github.com/getlantern/context v0.0.0-20190109183933-c447772a6520/go.mod h1:L+mq6/vvYHKjCX2oez0CgEAJmbq1fbb/oNJIWQkBybY=
+github.com/getlantern/errors v0.0.0-20190325191628-abdb3e3e36f7 h1:6uJ+sZ/e03gkbqZ0kUG6mfKoqDb4XMAzMIwlajq19So=
+github.com/getlantern/errors v0.0.0-20190325191628-abdb3e3e36f7/go.mod h1:l+xpFBrCtDLpK9qNjxs+cHU6+BAdlBaxHqikB6Lku3A=
+github.com/getlantern/golog v0.0.0-20190830074920-4ef2e798c2d7 h1:guBYzEaLz0Vfc/jv0czrr2z7qyzTOGC9hiQ0VC+hKjk=
+github.com/getlantern/golog v0.0.0-20190830074920-4ef2e798c2d7/go.mod h1:zx/1xUUeYPy3Pcmet8OSXLbF47l+3y6hIPpyLWoR9oc=
+github.com/getlantern/hex v0.0.0-20190417191902-c6586a6fe0b7 h1:micT5vkcr9tOVk1FiH8SWKID8ultN44Z+yzd2y/Vyb0=
+github.com/getlantern/hex v0.0.0-20190417191902-c6586a6fe0b7/go.mod h1:dD3CgOrwlzca8ed61CsZouQS5h5jIzkK9ZWrTcf0s+o=
+github.com/getlantern/hidden v0.0.0-20190325191715-f02dbb02be55 h1:XYzSdCbkzOC0FDNrgJqGRo8PCMFOBFL9py72DRs7bmc=
+github.com/getlantern/hidden v0.0.0-20190325191715-f02dbb02be55/go.mod h1:6mmzY2kW1TOOrVy+r41Za2MxXM+hhqTtY3oBKd2AgFA=
+github.com/getlantern/ops v0.0.0-20190325191751-d70cb0d6f85f h1:wrYrQttPS8FHIRSlsrcuKazukx/xqO/PpLZzZXsF+EA=
+github.com/getlantern/ops v0.0.0-20190325191751-d70cb0d6f85f/go.mod h1:D5ao98qkA6pxftxoqzibIBBrLSUli+kYnJqrgBf9cIA=
+github.com/getlantern/systray v1.2.2 h1:dCEHtfmvkJG7HZ8lS/sLklTH4RKUcIsKrAD9sThoEBE=
+github.com/getlantern/systray v1.2.2/go.mod h1:pXFOI1wwqwYXEhLPm9ZGjS2u/vVELeIgNMY5HvhHhcE=
 github.com/gin-contrib/cors v1.4.0 h1:oJ6gwtUl3lqV0WEIwM/LxPF1QZ5qe2lGWdY2+bz7y0g=
 github.com/gin-contrib/cors v1.4.0/go.mod h1:bs9pNM0x/UsmHPBWT2xZz9ROh8xYjYkiURUfmBoMlcs=
 github.com/gin-contrib/sse v0.1.0 h1:Y/yl/+YNO8GZSjAhjMsSuLt29uWRFHdHYUb5lYOV9qE=
@@ -31,6 +49,8 @@ github.com/go-playground/universal-translator v0.18.1/go.mod h1:xekY+UJKNuX9WP91
 github.com/go-playground/validator/v10 v10.10.0/go.mod h1:74x4gJWsvQexRdW8Pn3dXSGrTK4nAUsbPlLADvpJkos=
 github.com/go-playground/validator/v10 v10.14.0 h1:vgvQWe3XCz3gIeFDm/HnTIbj6UGmg/+t63MyGU2n5js=
 github.com/go-playground/validator/v10 v10.14.0/go.mod h1:9iXMNT7sEkjXb0I+enO7QXmzG6QCsPWY4zveKFVRSyU=
+github.com/go-stack/stack v1.8.0 h1:5SgMzNM5HxrEjV0ww2lTmX6E2Izsfxas4+YHWRs3Lsk=
+github.com/go-stack/stack v1.8.0/go.mod h1:v0f6uXyyMGvRgIKkXu+yp6POWl0qKG85gN/melR3HDY=
 github.com/goccy/go-json v0.9.7/go.mod h1:6MelG93GURQebXPDq3khkgXZkazVtN9CRI+MGFi0w8I=
 github.com/goccy/go-json v0.10.2 h1:CrxCmQqYDkv1z7lO7Wbh2HN93uovUHgrECaO5ZrCXAU=
 github.com/goccy/go-json v0.10.2/go.mod h1:6MelG93GURQebXPDq3khkgXZkazVtN9CRI+MGFi0w8I=
@@ -39,6 +59,8 @@ github.com/google/go-cmp v0.5.5/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/
 github.com/google/go-cmp v0.5.9 h1:O2Tfq5qg4qc4AmwVlvv0oLiVAGB7enBSJ2x2DqQFi38=
 github.com/google/go-cmp v0.5.9/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
 github.com/google/gofuzz v1.0.0/go.mod h1:dBl0BpW6vV/+mYPU4Po3pmUjxk6FQPldtuIdl/M65Eg=
+github.com/google/uuid v1.0.0 h1:b4Gk+7WdP/d3HZH8EJsZpvV7EtDOgaZLtnaNGIu1adA=
+github.com/google/uuid v1.0.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
 github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2s0bqwp9tc8=
 github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLfsEA9PFc4w1p2J65bw=
 github.com/json-iterator/go v1.1.12 h1:PV8peI4a0ysnczrg+LtxykD8LfKY9ML6u2jnxaEnrnM=
@@ -57,6 +79,8 @@ github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=
 github.com/leodido/go-urn v1.2.1/go.mod h1:zt4jvISO2HfUBqxjfIshjdMTYS56ZS/qv49ictyFfxY=
 github.com/leodido/go-urn v1.2.4 h1:XlAE/cm/ms7TE/VMVoduSpNBoyc2dOxHs5MZSwAN63Q=
 github.com/leodido/go-urn v1.2.4/go.mod h1:7ZrI8mTSeBSHl/UaRyKQW1qZeMgak41ANeCNaVckg+4=
+github.com/lxn/walk v0.0.0-20210112085537-c389da54e794/go.mod h1:E23UucZGqpuUANJooIbHWCufXvOcT6E7Stq81gU+CSQ=
+github.com/lxn/win v0.0.0-20210218163916-a377121e959e/go.mod h1:KxxjdtRkfNoYDCUP5ryK7XJJNTnpC8atvtmTheChOtk=
 github.com/mattn/go-isatty v0.0.14/go.mod h1:7GGIvUiUoEMVVmxf/4nioHXj79iQHKdU27kJ6hsGG94=
 github.com/mattn/go-isatty v0.0.19 h1:JITubQf0MOLdlGRuRq+jtsDlekdYPia9ZFsB8h/APPA=
 github.com/mattn/go-isatty v0.0.19/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
@@ -70,8 +94,12 @@ github.com/modern-go/reflect2 v1.0.2 h1:xBagoLtFs94CBntxluKeaWgTMpvLxC4ur3nMaC9G
 github.com/modern-go/reflect2 v1.0.2/go.mod h1:yWuevngMOJpCy52FWWMvUC8ws7m/LJsjYzDa0/r8luk=
 github.com/olekukonko/tablewriter v0.0.5 h1:P2Ga83D34wi1o9J6Wh1mRuqd4mF/x/lgBS7N7AbDhec=
 github.com/olekukonko/tablewriter v0.0.5/go.mod h1:hPp6KlRPjbx+hW8ykQs1w3UBbZlj6HuIJcUGPhkA7kY=
+github.com/oxtoacart/bpool v0.0.0-20190530202638-03653db5a59c h1:rp5dCmg/yLR3mgFuSOe4oEnDDmGLROTvMragMUXpTQw=
+github.com/oxtoacart/bpool v0.0.0-20190530202638-03653db5a59c/go.mod h1:X07ZCGwUbLaax7L0S3Tw4hpejzu63ZrrQiUe6W0hcy0=
 github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58 h1:onHthvaw9LFnH4t2DcNVpwGmV9E1BkGknEliJkfwQj0=
 github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58/go.mod h1:DXv8WO4yhMYhSNPKjeNKa5WY9YCIEBRbNzFFPJbWO6Y=
+github.com/pborman/uuid v1.2.1 h1:+ZZIw58t/ozdjRaXh/3awHfmWRbzYxJoAdNJxe/3pvw=
+github.com/pborman/uuid v1.2.1/go.mod h1:X/NO0urCmaxf9VXbdlT7C2Yzkj2IKimNn4k+gtPdI/k=
 github.com/pelletier/go-toml/v2 v2.0.1/go.mod h1:r9LEWfGN8R5k0VXJ+0BkIe7MYkRdwZOjgMj2KwnJFUo=
 github.com/pelletier/go-toml/v2 v2.0.8 h1:0ctb6s9mE31h0/lhu+J6OPmVeDxJn+kYnJc2jZR9tGQ=
 github.com/pelletier/go-toml/v2 v2.0.8/go.mod h1:vuYfssBdrU2XDZ9bYydBu6t+6a6PYNcZljzZR9VXg+4=
@@ -84,6 +112,7 @@ github.com/rogpeppe/go-internal v1.6.1/go.mod h1:xXDCJY+GAPziupqXw64V24skbSoqbTE
 github.com/rogpeppe/go-internal v1.8.0 h1:FCbCCtXNOY3UtUuHUYaghJg4y7Fd14rXifAYUAtL9R8=
 github.com/rogpeppe/go-internal v1.8.0/go.mod h1:WmiCO8CzOY8rg0OYDC4/i/2WRWAB6poM+XZ2dLUbcbE=
 github.com/russross/blackfriday/v2 v2.1.0/go.mod h1:+Rmxgy9KzJVeS9/2gXHxylqXiyQDYRxCVz55jmeOWTM=
+github.com/skratchdot/open-golang v0.0.0-20200116055534-eef842397966/go.mod h1:sUM3LWHvSMaG192sy56D9F7CNvL7jUJVXoqM1QKLnog=
 github.com/spf13/cobra v1.7.0 h1:hyqWnYt1ZQShIddO5kBpj3vu05/++x6tJ6dg8EC572I=
 github.com/spf13/cobra v1.7.0/go.mod h1:uLxZILRyS/50WlhOIKD7W6V5bgeIt+4sICxh6uRMrb0=
 github.com/spf13/pflag v1.0.5 h1:iy+VFUOCP1a+8yFto/drg2CJ5u0yRoB7fZw3DKv/JXA=
@@ -120,11 +149,14 @@ golang.org/x/net v0.17.0 h1:pVaXccu2ozPjCXewfr1S7xza/zcXTity9cCdXQYSjIM=
 golang.org/x/net v0.17.0/go.mod h1:NxSsAGuq816PNPmqtQdLE42eU2Fs7NoRIZrHJAlaCOE=
 golang.org/x/sync v0.3.0 h1:ftCYgMx6zT/asHUrPw8BLLscYtGznsLAnjq5RH9P66E=
 golang.org/x/sync v0.3.0/go.mod h1:FU7BRWz2tNW+3quACPkgCx/L+uEAv1htQ0V83Z9Rj+Y=
+golang.org/x/sys v0.0.0-20201018230417-eeed37f84f13/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20201119102817-f84b799fce68/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20210124154548-22da62e12c0c/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20210615035016-665e8c7367d1/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20210806184541-e5e7981a1069/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20220704084225-05e143d24a9e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.1.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.13.0 h1:Af8nKPmuFypiUBjVoU9V20FiaFXOcuZI21p0ycVYYGE=
 golang.org/x/sys v0.13.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
@@ -141,6 +173,7 @@ google.golang.org/protobuf v1.26.0-rc.1/go.mod h1:jlhhOSvTdKEhbULTjvd4ARK9grFBp0
 google.golang.org/protobuf v1.28.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I=
 google.golang.org/protobuf v1.30.0 h1:kPPoIgf3TsEvrm0PFe15JQ+570QVxYzEvvHqChK+cng=
 google.golang.org/protobuf v1.30.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I=
+gopkg.in/Knetic/govaluate.v3 v3.0.0/go.mod h1:csKLBORsPbafmSCGTEh3U7Ozmsuq8ZSIlKk1bcqph0E=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c h1:Hei/4ADfdWqJk1ZMxUNpqntNwaWcugrBjAiHlqqRiVk=

gpu/amd.go

@@ -0,0 +1,101 @@
+package gpu
+
+import (
+	"bufio"
+	"errors"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+)
+
+// TODO - windows vs. non-windows vs darwin
+
+// Discovery logic for AMD/ROCm GPUs
+
+const (
+	DriverVersionFile     = "/sys/module/amdgpu/version"
+	GPUPropertiesFileGlob = "/sys/class/kfd/kfd/topology/nodes/*/properties"
+	// TODO probably break these down per GPU to make the logic simpler
+	GPUTotalMemoryFileGlob = "/sys/class/kfd/kfd/topology/nodes/*/mem_banks/*/properties" // size_in_bytes line
+	GPUUsedMemoryFileGlob  = "/sys/class/kfd/kfd/topology/nodes/*/mem_banks/*/used_memory"
+)
+
+func AMDDetected() bool {
+	// Some driver versions (older?) don't have a version file, so just lookup the parent dir
+	sysfsDir := filepath.Dir(DriverVersionFile)
+	_, err := os.Stat(sysfsDir)
+	if errors.Is(err, os.ErrNotExist) {
+		slog.Debug("amd driver not detected " + sysfsDir)
+		return false
+	} else if err != nil {
+		slog.Debug(fmt.Sprintf("error looking up amd driver %s %s", sysfsDir, err))
+		return false
+	}
+	return true
+}
+
+func AMDDriverVersion() (string, error) {
+	_, err := os.Stat(DriverVersionFile)
+	if err != nil {
+		return "", fmt.Errorf("amdgpu file stat error: %s %w", DriverVersionFile, err)
+	}
+	fp, err := os.Open(DriverVersionFile)
+	if err != nil {
+		return "", err
+	}
+	defer fp.Close()
+	verString, err := io.ReadAll(fp)
+	if err != nil {
+		return "", err
+	}
+	return strings.TrimSpace(string(verString)), nil
+}
+
+func AMDGFXVersions() []Version {
+	res := []Version{}
+	matches, _ := filepath.Glob(GPUPropertiesFileGlob)
+	for _, match := range matches {
+		fp, err := os.Open(match)
+		if err != nil {
+			slog.Debug(fmt.Sprintf("failed to open sysfs node file %s: %s", match, err))
+			continue
+		}
+		defer fp.Close()
+
+		scanner := bufio.NewScanner(fp)
+		// optionally, resize scanner's capacity for lines over 64K, see next example
+		for scanner.Scan() {
+			line := strings.TrimSpace(scanner.Text())
+			if strings.HasPrefix(line, "gfx_target_version") {
+				ver := strings.Fields(line)
+				if len(ver) != 2 || len(ver[1]) < 5 {
+					slog.Debug("malformed " + line)
+					continue
+				}
+				l := len(ver[1])
+				patch, err1 := strconv.ParseUint(ver[1][l-2:l], 10, 32)
+				minor, err2 := strconv.ParseUint(ver[1][l-4:l-2], 10, 32)
+				major, err3 := strconv.ParseUint(ver[1][:l-4], 10, 32)
+				if err1 != nil || err2 != nil || err3 != nil {
+					slog.Debug("malformed int " + line)
+					continue
+				}
+
+				res = append(res, Version{
+					Major: uint(major),
+					Minor: uint(minor),
+					Patch: uint(patch),
+				})
+			}
+		}
+	}
+	return res
+}
+
+func (v Version) ToGFXString() string {
+	return fmt.Sprintf("gfx%d%d%d", v.Major, v.Minor, v.Patch)
+}

gpu/cpu_common.go

@@ -0,0 +1,21 @@
+package gpu
+
+import (
+	"log/slog"
+
+	"golang.org/x/sys/cpu"
+)
+
+func GetCPUVariant() string {
+	if cpu.X86.HasAVX2 {
+		slog.Info("CPU has AVX2")
+		return "avx2"
+	}
+	if cpu.X86.HasAVX {
+		slog.Info("CPU has AVX")
+		return "avx"
+	}
+	slog.Info("CPU does not have vector extensions")
+	// else LCD
+	return ""
+}

gpu/gpu.go

@@ -12,12 +12,14 @@ package gpu
 import "C"
 import (
 	"fmt"
-	"log"
+	"log/slog"
+	"os"
+	"path/filepath"
 	"runtime"
+	"strconv"
+	"strings"
 	"sync"
 	"unsafe"
-
-	"github.com/jmorganca/ollama/api"
 )
 
 type handles struct {
@@ -28,31 +30,86 @@ type handles struct {
 var gpuMutex sync.Mutex
 var gpuHandles *handles = nil
 
+// With our current CUDA compile flags, older than 5.0 will not work properly
+var CudaComputeMin = [2]C.int{5, 0}
+
+// Possible locations for the nvidia-ml library
+var CudaLinuxGlobs = []string{
+	"/usr/local/cuda/lib64/libnvidia-ml.so*",
+	"/usr/lib/x86_64-linux-gnu/nvidia/current/libnvidia-ml.so*",
+	"/usr/lib/x86_64-linux-gnu/libnvidia-ml.so*",
+	"/usr/lib/wsl/lib/libnvidia-ml.so*",
+	"/usr/lib/wsl/drivers/*/libnvidia-ml.so*",
+	"/opt/cuda/lib64/libnvidia-ml.so*",
+	"/usr/lib*/libnvidia-ml.so*",
+	"/usr/local/lib*/libnvidia-ml.so*",
+	"/usr/lib/aarch64-linux-gnu/nvidia/current/libnvidia-ml.so*",
+	"/usr/lib/aarch64-linux-gnu/libnvidia-ml.so*",
+
+	// TODO: are these stubs ever valid?
+	"/opt/cuda/targets/x86_64-linux/lib/stubs/libnvidia-ml.so*",
+}
+
+var CudaWindowsGlobs = []string{
+	"c:\\Windows\\System32\\nvml.dll",
+}
+
+var RocmLinuxGlobs = []string{
+	"/opt/rocm*/lib*/librocm_smi64.so*",
+}
+
+var RocmWindowsGlobs = []string{
+	"c:\\Windows\\System32\\rocm_smi64.dll",
+}
+
 // Note: gpuMutex must already be held
 func initGPUHandles() {
-	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
-	log.Printf("Detecting GPU type")
-	gpuHandles = &handles{nil, nil}
-	var resp C.cuda_init_resp_t
-	C.cuda_init(&resp)
-	if resp.err != nil {
-		log.Printf("CUDA not detected: %s", C.GoString(resp.err))
-		C.free(unsafe.Pointer(resp.err))
 
-		var resp C.rocm_init_resp_t
-		C.rocm_init(&resp)
-		if resp.err != nil {
-			log.Printf("ROCm not detected: %s", C.GoString(resp.err))
-			C.free(unsafe.Pointer(resp.err))
-		} else {
-			log.Printf("Radeon GPU detected")
-			rocm := resp.rh
-			gpuHandles.rocm = &rocm
+	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
+
+	gpuHandles = &handles{nil, nil}
+	var cudaMgmtName string
+	var cudaMgmtPatterns []string
+	var rocmMgmtName string
+	var rocmMgmtPatterns []string
+	switch runtime.GOOS {
+	case "windows":
+		cudaMgmtName = "nvml.dll"
+		cudaMgmtPatterns = make([]string, len(CudaWindowsGlobs))
+		copy(cudaMgmtPatterns, CudaWindowsGlobs)
+		rocmMgmtName = "rocm_smi64.dll"
+		rocmMgmtPatterns = make([]string, len(RocmWindowsGlobs))
+		copy(rocmMgmtPatterns, RocmWindowsGlobs)
+	case "linux":
+		cudaMgmtName = "libnvidia-ml.so"
+		cudaMgmtPatterns = make([]string, len(CudaLinuxGlobs))
+		copy(cudaMgmtPatterns, CudaLinuxGlobs)
+		rocmMgmtName = "librocm_smi64.so"
+		rocmMgmtPatterns = make([]string, len(RocmLinuxGlobs))
+		copy(rocmMgmtPatterns, RocmLinuxGlobs)
+	default:
+		return
+	}
+
+	slog.Info("Detecting GPU type")
+	cudaLibPaths := FindGPULibs(cudaMgmtName, cudaMgmtPatterns)
+	if len(cudaLibPaths) > 0 {
+		cuda := LoadCUDAMgmt(cudaLibPaths)
+		if cuda != nil {
+			slog.Info("Nvidia GPU detected")
+			gpuHandles.cuda = cuda
+			return
+		}
+	}
+
+	rocmLibPaths := FindGPULibs(rocmMgmtName, rocmMgmtPatterns)
+	if len(rocmLibPaths) > 0 {
+		rocm := LoadROCMMgmt(rocmLibPaths)
+		if rocm != nil {
+			slog.Info("Radeon GPU detected")
+			gpuHandles.rocm = rocm
+			return
 		}
-	} else {
-		log.Printf("Nvidia GPU detected")
-		cuda := resp.ch
-		gpuHandles.cuda = &cuda
 	}
 }
 
@@ -65,39 +122,107 @@ func GetGPUInfo() GpuInfo {
 		initGPUHandles()
 	}
 
+	// All our GPU builds on x86 have AVX enabled, so fallback to CPU if we don't detect at least AVX
+	cpuVariant := GetCPUVariant()
+	if cpuVariant == "" && runtime.GOARCH == "amd64" {
+		slog.Warn("CPU does not have AVX or AVX2, disabling GPU support.")
+	}
+
 	var memInfo C.mem_info_t
 	resp := GpuInfo{}
-	if gpuHandles.cuda != nil {
+	if gpuHandles.cuda != nil && (cpuVariant != "" || runtime.GOARCH != "amd64") {
 		C.cuda_check_vram(*gpuHandles.cuda, &memInfo)
 		if memInfo.err != nil {
-			log.Printf("error looking up CUDA GPU memory: %s", C.GoString(memInfo.err))
+			slog.Info(fmt.Sprintf("error looking up CUDA GPU memory: %s", C.GoString(memInfo.err)))
 			C.free(unsafe.Pointer(memInfo.err))
-		} else {
-			resp.Library = "cuda"
+		} else if memInfo.count > 0 {
+			// Verify minimum compute capability
+			var cc C.cuda_compute_capability_t
+			C.cuda_compute_capability(*gpuHandles.cuda, &cc)
+			if cc.err != nil {
+				slog.Info(fmt.Sprintf("error looking up CUDA GPU compute capability: %s", C.GoString(cc.err)))
+				C.free(unsafe.Pointer(cc.err))
+			} else if cc.major > CudaComputeMin[0] || (cc.major == CudaComputeMin[0] && cc.minor >= CudaComputeMin[1]) {
+				slog.Info(fmt.Sprintf("CUDA Compute Capability detected: %d.%d", cc.major, cc.minor))
+				resp.Library = "cuda"
+			} else {
+				slog.Info(fmt.Sprintf("CUDA GPU is too old. Falling back to CPU mode. Compute Capability detected: %d.%d", cc.major, cc.minor))
+			}
 		}
-	} else if gpuHandles.rocm != nil {
-		C.rocm_check_vram(*gpuHandles.rocm, &memInfo)
-		if memInfo.err != nil {
-			log.Printf("error looking up ROCm GPU memory: %s", C.GoString(memInfo.err))
-			C.free(unsafe.Pointer(memInfo.err))
+	} else if AMDDetected() && gpuHandles.rocm != nil && (cpuVariant != "" || runtime.GOARCH != "amd64") {
+		ver, err := AMDDriverVersion()
+		if err == nil {
+			slog.Info("AMD Driver: " + ver)
 		} else {
-			resp.Library = "rocm"
+			// For now this is benign, but we may eventually need to fail compatibility checks
+			slog.Debug("error looking up amd driver version: %s", err)
+		}
+		gfx := AMDGFXVersions()
+		tooOld := false
+		for _, v := range gfx {
+			if v.Major < 9 {
+				slog.Info("AMD GPU too old, falling back to CPU " + v.ToGFXString())
+				tooOld = true
+				break
+			}
+
+			// TODO - remap gfx strings for unsupporetd minor/patch versions to supported for the same major
+			// e.g. gfx1034 works if we map it to gfx1030 at runtime
+
+		}
+		if !tooOld {
+			// TODO - this algo can be shifted over to use sysfs instead of the rocm info library...
+			C.rocm_check_vram(*gpuHandles.rocm, &memInfo)
+			if memInfo.err != nil {
+				slog.Info(fmt.Sprintf("error looking up ROCm GPU memory: %s", C.GoString(memInfo.err)))
+				C.free(unsafe.Pointer(memInfo.err))
+			} else if memInfo.igpu_index >= 0 && memInfo.count == 1 {
+				// Only one GPU detected and it appears to be an integrated GPU - skip it
+				slog.Info("ROCm unsupported integrated GPU detected")
+			} else if memInfo.count > 0 {
+				if memInfo.igpu_index >= 0 {
+					// We have multiple GPUs reported, and one of them is an integrated GPU
+					// so we have to set the env var to bypass it
+					// If the user has specified their own ROCR_VISIBLE_DEVICES, don't clobber it
+					val := os.Getenv("ROCR_VISIBLE_DEVICES")
+					if val == "" {
+						devices := []string{}
+						for i := 0; i < int(memInfo.count); i++ {
+							if i == int(memInfo.igpu_index) {
+								continue
+							}
+							devices = append(devices, strconv.Itoa(i))
+						}
+						val = strings.Join(devices, ",")
+						os.Setenv("ROCR_VISIBLE_DEVICES", val)
+					}
+					slog.Info(fmt.Sprintf("ROCm integrated GPU detected - ROCR_VISIBLE_DEVICES=%s", val))
+				}
+				resp.Library = "rocm"
+				var version C.rocm_version_resp_t
+				C.rocm_get_version(*gpuHandles.rocm, &version)
+				verString := C.GoString(version.str)
+				if version.status == 0 {
+					resp.Variant = "v" + verString
+				} else {
+					slog.Info(fmt.Sprintf("failed to look up ROCm version: %s", verString))
+				}
+				C.free(unsafe.Pointer(version.str))
+			}
 		}
 	}
 	if resp.Library == "" {
 		C.cpu_check_ram(&memInfo)
-		// In the future we may offer multiple CPU variants to tune CPU features
-		if runtime.GOOS == "windows" {
-			resp.Library = "cpu"
-		} else {
-			resp.Library = "default"
-		}
+		resp.Library = "cpu"
+		resp.Variant = cpuVariant
 	}
 	if memInfo.err != nil {
-		log.Printf("error looking up CPU memory: %s", C.GoString(memInfo.err))
+		slog.Info(fmt.Sprintf("error looking up CPU memory: %s", C.GoString(memInfo.err)))
 		C.free(unsafe.Pointer(memInfo.err))
 		return resp
 	}
+
+	resp.DeviceCount = uint32(memInfo.count)
 	resp.FreeMemory = uint64(memInfo.free)
 	resp.TotalMemory = uint64(memInfo.total)
 	return resp
@@ -119,31 +244,111 @@ func getCPUMem() (memInfo, error) {
 func CheckVRAM() (int64, error) {
 	gpuInfo := GetGPUInfo()
 	if gpuInfo.FreeMemory > 0 && (gpuInfo.Library == "cuda" || gpuInfo.Library == "rocm") {
-		return int64(gpuInfo.FreeMemory), nil
+		// leave 10% or 1024MiB of VRAM free per GPU to handle unaccounted for overhead
+		overhead := gpuInfo.FreeMemory / 10
+		gpus := uint64(gpuInfo.DeviceCount)
+		if overhead < gpus*1024*1024*1024 {
+			overhead = gpus * 1024 * 1024 * 1024
+		}
+		avail := int64(gpuInfo.FreeMemory - overhead)
+		slog.Debug(fmt.Sprintf("%s detected %d devices with %dM available memory", gpuInfo.Library, gpuInfo.DeviceCount, avail/1024/1024))
+		return avail, nil
 	}
+
 	return 0, fmt.Errorf("no GPU detected") // TODO - better handling of CPU based memory determiniation
 }
 
-func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
-	if opts.NumGPU != -1 {
-		return opts.NumGPU
+func FindGPULibs(baseLibName string, patterns []string) []string {
+	// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
+	var ldPaths []string
+	gpuLibPaths := []string{}
+	slog.Info(fmt.Sprintf("Searching for GPU management library %s", baseLibName))
+
+	switch runtime.GOOS {
+	case "windows":
+		ldPaths = strings.Split(os.Getenv("PATH"), ";")
+	case "linux":
+		ldPaths = strings.Split(os.Getenv("LD_LIBRARY_PATH"), ":")
+	default:
+		return gpuLibPaths
 	}
-	info := GetGPUInfo()
-	if info.Library == "cpu" || info.Library == "default" {
-		return 0
+	// Start with whatever we find in the PATH/LD_LIBRARY_PATH
+	for _, ldPath := range ldPaths {
+		d, err := filepath.Abs(ldPath)
+		if err != nil {
+			continue
+		}
+		patterns = append(patterns, filepath.Join(d, baseLibName+"*"))
 	}
-
-	/*
-		Calculate bytes per layer, this will roughly be the size of the model file divided by the number of layers.
-		We can store the model weights and the kv cache in vram,
-		to enable kv chache vram storage add two additional layers to the number of layers retrieved from the model file.
-	*/
-	bytesPerLayer := uint64(fileSizeBytes / numLayer)
-
-	// 75% of the absolute max number of layers we can fit in available VRAM, off-loading too many layers to the GPU can cause OOM errors
-	layers := int(info.FreeMemory/bytesPerLayer) * 3 / 4
-
-	log.Printf("%d MB VRAM available, loading up to %d %s GPU layers out of %d", info.FreeMemory/(1024*1024), layers, info.Library, numLayer)
-
-	return layers
+	slog.Debug(fmt.Sprintf("gpu management search paths: %v", patterns))
+	for _, pattern := range patterns {
+		// Ignore glob discovery errors
+		matches, _ := filepath.Glob(pattern)
+		for _, match := range matches {
+			// Resolve any links so we don't try the same lib multiple times
+			// and weed out any dups across globs
+			libPath := match
+			tmp := match
+			var err error
+			for ; err == nil; tmp, err = os.Readlink(libPath) {
+				if !filepath.IsAbs(tmp) {
+					tmp = filepath.Join(filepath.Dir(libPath), tmp)
+				}
+				libPath = tmp
+			}
+			new := true
+			for _, cmp := range gpuLibPaths {
+				if cmp == libPath {
+					new = false
+					break
+				}
+			}
+			if new {
+				gpuLibPaths = append(gpuLibPaths, libPath)
+			}
+		}
+	}
+	slog.Info(fmt.Sprintf("Discovered GPU libraries: %v", gpuLibPaths))
+	return gpuLibPaths
+}
+
+func LoadCUDAMgmt(cudaLibPaths []string) *C.cuda_handle_t {
+	var resp C.cuda_init_resp_t
+	resp.ch.verbose = getVerboseState()
+	for _, libPath := range cudaLibPaths {
+		lib := C.CString(libPath)
+		defer C.free(unsafe.Pointer(lib))
+		C.cuda_init(lib, &resp)
+		if resp.err != nil {
+			slog.Info(fmt.Sprintf("Unable to load CUDA management library %s: %s", libPath, C.GoString(resp.err)))
+			C.free(unsafe.Pointer(resp.err))
+		} else {
+			return &resp.ch
+		}
+	}
+	return nil
+}
+
+func LoadROCMMgmt(rocmLibPaths []string) *C.rocm_handle_t {
+	var resp C.rocm_init_resp_t
+	resp.rh.verbose = getVerboseState()
+	for _, libPath := range rocmLibPaths {
+		lib := C.CString(libPath)
+		defer C.free(unsafe.Pointer(lib))
+		C.rocm_init(lib, &resp)
+		if resp.err != nil {
+			slog.Info(fmt.Sprintf("Unable to load ROCm management library %s: %s", libPath, C.GoString(resp.err)))
+			C.free(unsafe.Pointer(resp.err))
+		} else {
+			return &resp.rh
+		}
+	}
+	return nil
+}
+
+func getVerboseState() C.uint16_t {
+	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
+		return C.uint16_t(1)
+	}
+	return C.uint16_t(0)
 }

gpu/gpu_darwin.go

@@ -6,21 +6,41 @@ import "C"
 import (
 	"runtime"
 
-	"github.com/jmorganca/ollama/api"
+	"github.com/pbnjay/memory"
 )
 
 // CheckVRAM returns the free VRAM in bytes on Linux machines with NVIDIA GPUs
 func CheckVRAM() (int64, error) {
-	// TODO - assume metal, and return free memory?
-	return 0, nil
+	if runtime.GOARCH == "amd64" {
+		// gpu not supported, this may not be metal
+		return 0, nil
+	}
 
+	// on macOS, there's already buffer for available vram (see below) so just return the total
+	systemMemory := int64(memory.TotalMemory())
+
+	// macOS limits how much memory is available to the GPU based on the amount of system memory
+	// TODO: handle case where iogpu.wired_limit_mb is set to a higher value
+	if systemMemory <= 36*1024*1024*1024 {
+		systemMemory = systemMemory * 2 / 3
+	} else {
+		systemMemory = systemMemory * 3 / 4
+	}
+
+	return systemMemory, nil
 }
 
 func GetGPUInfo() GpuInfo {
-	// TODO - Metal vs. x86 macs...
 	mem, _ := getCPUMem()
+	if runtime.GOARCH == "amd64" {
+		return GpuInfo{
+			Library: "cpu",
+			Variant: GetCPUVariant(),
+			memInfo: mem,
+		}
+	}
 	return GpuInfo{
-		Library: "default",
+		Library: "metal",
 		memInfo: mem,
 	}
 }
@@ -29,22 +49,6 @@ func getCPUMem() (memInfo, error) {
 	return memInfo{
 		TotalMemory: 0,
 		FreeMemory:  0,
+		DeviceCount: 0,
 	}, nil
 }
-
-func NumGPU(numLayer, fileSizeBytes int64, opts api.Options) int {
-	if opts.NumGPU != -1 {
-		return opts.NumGPU
-	}
-
-	// metal only supported on arm64
-	if runtime.GOARCH == "arm64" {
-		return 1
-	}
-
-	return 0
-}
-
-func nativeInit() error {
-	return nil
-}

gpu/gpu_info.h

@@ -27,6 +27,13 @@
 
 #endif
 
+#define LOG(verbose, ...) \
+  do { \
+    if (verbose) { \
+      fprintf(stderr, __VA_ARGS__); \
+    } \
+  } while (0)
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -34,6 +41,8 @@ extern "C" {
 typedef struct mem_info {
   uint64_t total;
   uint64_t free;
+  unsigned int count;
+  int igpu_index; // If >= 0, we detected an integrated GPU to ignore
   char *err;  // If non-nill, caller responsible for freeing
 } mem_info_t;
 

gpu/gpu_info_cpu.c

@@ -8,6 +8,7 @@ void cpu_check_ram(mem_info_t *resp) {
   MEMORYSTATUSEX info;
   info.dwLength = sizeof(info);
   if (GlobalMemoryStatusEx(&info) != 0) {
+    resp->count = 1;
     resp->total = info.ullTotalPhys;
     resp->free = info.ullAvailPhys;
   } else {
@@ -26,6 +27,7 @@ void cpu_check_ram(mem_info_t *resp) {
   if (sysinfo(&info) != 0) {
     resp->err = strdup(strerror(errno));
   } else {
+    resp->count = 1;
     resp->total = info.totalram * info.mem_unit;
     resp->free = info.freeram * info.mem_unit;
   }

gpu/gpu_info_cuda.c

@@ -4,23 +4,7 @@
 
 #include <string.h>
 
-#ifndef _WIN32
-const char *cuda_lib_paths[] = {
-    "libnvidia-ml.so",
-    "/usr/local/cuda/lib64/libnvidia-ml.so",
-    "/usr/lib/x86_64-linux-gnu/nvidia/current/libnvidia-ml.so",
-    "/usr/lib/wsl/lib/libnvidia-ml.so.1",  // TODO Maybe glob?
-    NULL,
-};
-#else
-const char *cuda_lib_paths[] = {
-    "nvml.dll",
-    "",
-    NULL,
-};
-#endif
-
-void cuda_init(cuda_init_resp_t *resp) {
+void cuda_init(char *cuda_lib_path, cuda_init_resp_t *resp) {
   nvmlReturn_t ret;
   resp->err = NULL;
   const int buflen = 256;
@@ -30,34 +14,47 @@ void cuda_init(cuda_init_resp_t *resp) {
   struct lookup {
     char *s;
     void **p;
-  } l[4] = {
-      {"nvmlInit_v2", (void *)&resp->ch.initFn},
-      {"nvmlShutdown", (void *)&resp->ch.shutdownFn},
-      {"nvmlDeviceGetHandleByIndex", (void *)&resp->ch.getHandle},
-      {"nvmlDeviceGetMemoryInfo", (void *)&resp->ch.getMemInfo},
+  } l[] = {
+      {"nvmlInit_v2", (void *)&resp->ch.nvmlInit_v2},
+      {"nvmlShutdown", (void *)&resp->ch.nvmlShutdown},
+      {"nvmlDeviceGetHandleByIndex", (void *)&resp->ch.nvmlDeviceGetHandleByIndex},
+      {"nvmlDeviceGetMemoryInfo", (void *)&resp->ch.nvmlDeviceGetMemoryInfo},
+      {"nvmlDeviceGetCount_v2", (void *)&resp->ch.nvmlDeviceGetCount_v2},
+      {"nvmlDeviceGetCudaComputeCapability", (void *)&resp->ch.nvmlDeviceGetCudaComputeCapability},
+      {"nvmlSystemGetDriverVersion", (void *)&resp->ch.nvmlSystemGetDriverVersion},
+      {"nvmlDeviceGetName", (void *)&resp->ch.nvmlDeviceGetName},
+      {"nvmlDeviceGetSerial", (void *)&resp->ch.nvmlDeviceGetSerial},
+      {"nvmlDeviceGetVbiosVersion", (void *)&resp->ch.nvmlDeviceGetVbiosVersion},
+      {"nvmlDeviceGetBoardPartNumber", (void *)&resp->ch.nvmlDeviceGetBoardPartNumber},
+      {"nvmlDeviceGetBrand", (void *)&resp->ch.nvmlDeviceGetBrand},
+      {NULL, NULL},
   };
 
-  for (i = 0; cuda_lib_paths[i] != NULL && resp->ch.handle == NULL; i++) {
-    resp->ch.handle = LOAD_LIBRARY(cuda_lib_paths[i], RTLD_LAZY);
-  }
+  resp->ch.handle = LOAD_LIBRARY(cuda_lib_path, RTLD_LAZY);
   if (!resp->ch.handle) {
-    // TODO improve error message, as the LOAD_ERR will have typically have the
-    // final path that was checked which might be confusing.
     char *msg = LOAD_ERR();
+    LOG(resp->ch.verbose, "library %s load err: %s\n", cuda_lib_path, msg);
     snprintf(buf, buflen,
              "Unable to load %s library to query for Nvidia GPUs: %s",
-             cuda_lib_paths[0], msg);
+             cuda_lib_path, msg);
     free(msg);
     resp->err = strdup(buf);
     return;
   }
 
-  for (i = 0; i < 4; i++) {  // TODO - fix this to use a null terminated list
+  // TODO once we've squashed the remaining corner cases remove this log
+  LOG(resp->ch.verbose, "wiring nvidia management library functions in %s\n", cuda_lib_path);
+  
+  for (i = 0; l[i].s != NULL; i++) {
+    // TODO once we've squashed the remaining corner cases remove this log
+    LOG(resp->ch.verbose, "dlsym: %s\n", l[i].s);
+
     *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
     if (!l[i].p) {
-      UNLOAD_LIBRARY(resp->ch.handle);
       resp->ch.handle = NULL;
       char *msg = LOAD_ERR();
+      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
+      UNLOAD_LIBRARY(resp->ch.handle);
       snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
                msg);
       free(msg);
@@ -66,13 +63,23 @@ void cuda_init(cuda_init_resp_t *resp) {
     }
   }
 
-  ret = (*resp->ch.initFn)();
+  ret = (*resp->ch.nvmlInit_v2)();
   if (ret != NVML_SUCCESS) {
+    LOG(resp->ch.verbose, "nvmlInit_v2 err: %d\n", ret);
+    UNLOAD_LIBRARY(resp->ch.handle);
+    resp->ch.handle = NULL;
     snprintf(buf, buflen, "nvml vram init failure: %d", ret);
     resp->err = strdup(buf);
+    return;
   }
 
-  return;
+  // Report driver version if we're in verbose mode, ignore errors
+  ret = (*resp->ch.nvmlSystemGetDriverVersion)(buf, buflen);
+  if (ret != NVML_SUCCESS) {
+    LOG(resp->ch.verbose, "nvmlSystemGetDriverVersion failed: %d\n", ret);
+  } else {
+    LOG(resp->ch.verbose, "CUDA driver version: %s\n", buf);
+  }
 }
 
 void cuda_check_vram(cuda_handle_t h, mem_info_t *resp) {
@@ -89,22 +96,119 @@ void cuda_check_vram(cuda_handle_t h, mem_info_t *resp) {
     return;
   }
 
-  // TODO - handle multiple GPUs
-  ret = (*h.getHandle)(0, &device);
+  ret = (*h.nvmlDeviceGetCount_v2)(&resp->count);
   if (ret != NVML_SUCCESS) {
-    snprintf(buf, buflen, "unable to get device handle: %d", ret);
+    snprintf(buf, buflen, "unable to get device count: %d", ret);
     resp->err = strdup(buf);
     return;
   }
 
-  ret = (*h.getMemInfo)(device, &memInfo);
+  resp->total = 0;
+  resp->free = 0;
+  for (i = 0; i < resp->count; i++) {
+    ret = (*h.nvmlDeviceGetHandleByIndex)(i, &device);
+    if (ret != NVML_SUCCESS) {
+      snprintf(buf, buflen, "unable to get device handle %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+
+    ret = (*h.nvmlDeviceGetMemoryInfo)(device, &memInfo);
+    if (ret != NVML_SUCCESS) {
+      snprintf(buf, buflen, "device memory info lookup failure %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+    if (h.verbose) {
+      nvmlBrandType_t brand = 0;
+      // When in verbose mode, report more information about
+      // the card we discover, but don't fail on error
+      ret = (*h.nvmlDeviceGetName)(device, buf, buflen);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "nvmlDeviceGetName failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] CUDA device name: %s\n", i, buf);
+      }
+      ret = (*h.nvmlDeviceGetBoardPartNumber)(device, buf, buflen);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "nvmlDeviceGetBoardPartNumber failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] CUDA part number: %s\n", i, buf);
+      }
+      ret = (*h.nvmlDeviceGetSerial)(device, buf, buflen);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "nvmlDeviceGetSerial failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] CUDA S/N: %s\n", i, buf);
+      }
+      ret = (*h.nvmlDeviceGetVbiosVersion)(device, buf, buflen);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "nvmlDeviceGetVbiosVersion failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] CUDA vbios version: %s\n", i, buf);
+      }
+      ret = (*h.nvmlDeviceGetBrand)(device, &brand);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "nvmlDeviceGetBrand failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] CUDA brand: %d\n", i, brand);
+      }
+    }
+
+    LOG(h.verbose, "[%d] CUDA totalMem %ld\n", i, memInfo.total);
+    LOG(h.verbose, "[%d] CUDA usedMem %ld\n", i, memInfo.free);
+
+    resp->total += memInfo.total;
+    resp->free += memInfo.free;
+  }
+}
+
+void cuda_compute_capability(cuda_handle_t h, cuda_compute_capability_t *resp) {
+  resp->err = NULL;
+  resp->major = 0;
+  resp->minor = 0;
+  nvmlDevice_t device;
+  int major = 0;
+  int minor = 0;
+  nvmlReturn_t ret;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  if (h.handle == NULL) {
+    resp->err = strdup("nvml handle not initialized");
+    return;
+  }
+
+  unsigned int devices;
+  ret = (*h.nvmlDeviceGetCount_v2)(&devices);
   if (ret != NVML_SUCCESS) {
-    snprintf(buf, buflen, "device memory info lookup failure: %d", ret);
+    snprintf(buf, buflen, "unable to get device count: %d", ret);
     resp->err = strdup(buf);
     return;
   }
-  resp->total = memInfo.total;
-  resp->free = memInfo.free;
-  return;
+
+  for (i = 0; i < devices; i++) {
+    ret = (*h.nvmlDeviceGetHandleByIndex)(i, &device);
+    if (ret != NVML_SUCCESS) {
+      snprintf(buf, buflen, "unable to get device handle %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+
+    ret = (*h.nvmlDeviceGetCudaComputeCapability)(device, &major, &minor);
+    if (ret != NVML_SUCCESS) {
+      snprintf(buf, buflen, "device compute capability lookup failure %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+    // Report the lowest major.minor we detect as that limits our compatibility
+    if (resp->major == 0 || resp->major > major ) {
+      resp->major = major;
+      resp->minor = minor;
+    } else if ( resp->major == major && resp->minor > minor ) {
+      resp->minor = minor;
+    }
+  }
 }
 #endif  // __APPLE__

gpu/gpu_info_cuda.h

@@ -15,12 +15,26 @@ typedef struct nvmlMemory_st {
   unsigned long long used;
 } nvmlMemory_t;
 
+typedef enum nvmlBrandType_enum
+{
+    NVML_BRAND_UNKNOWN          = 0,
+} nvmlBrandType_t;
+
 typedef struct cuda_handle {
   void *handle;
-  nvmlReturn_t (*initFn)(void);
-  nvmlReturn_t (*shutdownFn)(void);
-  nvmlReturn_t (*getHandle)(unsigned int, nvmlDevice_t *);
-  nvmlReturn_t (*getMemInfo)(nvmlDevice_t, nvmlMemory_t *);
+  uint16_t verbose;
+  nvmlReturn_t (*nvmlInit_v2)(void);
+  nvmlReturn_t (*nvmlShutdown)(void);
+  nvmlReturn_t (*nvmlDeviceGetHandleByIndex)(unsigned int, nvmlDevice_t *);
+  nvmlReturn_t (*nvmlDeviceGetMemoryInfo)(nvmlDevice_t, nvmlMemory_t *);
+  nvmlReturn_t (*nvmlDeviceGetCount_v2)(unsigned int *);
+  nvmlReturn_t (*nvmlDeviceGetCudaComputeCapability)(nvmlDevice_t, int* major, int* minor);
+  nvmlReturn_t (*nvmlSystemGetDriverVersion) (char* version, unsigned int  length);
+  nvmlReturn_t (*nvmlDeviceGetName) (nvmlDevice_t device, char* name, unsigned int  length);
+  nvmlReturn_t (*nvmlDeviceGetSerial) (nvmlDevice_t device, char* serial, unsigned int  length);
+  nvmlReturn_t (*nvmlDeviceGetVbiosVersion) (nvmlDevice_t device, char* version, unsigned int  length);
+  nvmlReturn_t (*nvmlDeviceGetBoardPartNumber) (nvmlDevice_t device, char* partNumber, unsigned int  length);
+  nvmlReturn_t (*nvmlDeviceGetBrand) (nvmlDevice_t device, nvmlBrandType_t* type);
 } cuda_handle_t;
 
 typedef struct cuda_init_resp {
@@ -28,8 +42,15 @@ typedef struct cuda_init_resp {
   cuda_handle_t ch;
 } cuda_init_resp_t;
 
-void cuda_init(cuda_init_resp_t *resp);
+typedef struct cuda_compute_capability {
+  char *err;
+  int major;
+  int minor;
+} cuda_compute_capability_t;
+
+void cuda_init(char *cuda_lib_path, cuda_init_resp_t *resp);
 void cuda_check_vram(cuda_handle_t ch, mem_info_t *resp);
+void cuda_compute_capability(cuda_handle_t ch, cuda_compute_capability_t *cc);
 
 #endif  // __GPU_INFO_CUDA_H__
 #endif  // __APPLE__

gpu/gpu_info_rocm.c

@@ -4,22 +4,7 @@
 
 #include <string.h>
 
-#ifndef _WIN32
-const char *rocm_lib_paths[] = {
-    "librocm_smi64.so",
-    "/opt/rocm/lib/librocm_smi64.so",
-    NULL,
-};
-#else
-// TODO untested
-const char *rocm_lib_paths[] = {
-    "rocm_smi64.dll",
-    "/opt/rocm/lib/rocm_smi64.dll",
-    NULL,
-};
-#endif
-
-void rocm_init(rocm_init_resp_t *resp) {
+void rocm_init(char *rocm_lib_path, rocm_init_resp_t *resp) {
   rsmi_status_t ret;
   resp->err = NULL;
   const int buflen = 256;
@@ -28,32 +13,48 @@ void rocm_init(rocm_init_resp_t *resp) {
   struct lookup {
     char *s;
     void **p;
-  } l[4] = {
-      {"rsmi_init", (void *)&resp->rh.initFn},
-      {"rsmi_shut_down", (void *)&resp->rh.shutdownFn},
-      {"rsmi_dev_memory_total_get", (void *)&resp->rh.totalMemFn},
-      {"rsmi_dev_memory_usage_get", (void *)&resp->rh.usageMemFn},
-      // { "rsmi_dev_id_get", (void*)&resp->rh.getHandle },
+  } l[] = {
+      {"rsmi_init", (void *)&resp->rh.rsmi_init},
+      {"rsmi_shut_down", (void *)&resp->rh.rsmi_shut_down},
+      {"rsmi_dev_memory_total_get", (void *)&resp->rh.rsmi_dev_memory_total_get},
+      {"rsmi_dev_memory_usage_get", (void *)&resp->rh.rsmi_dev_memory_usage_get},
+      {"rsmi_version_get", (void *)&resp->rh.rsmi_version_get},
+      {"rsmi_num_monitor_devices", (void*)&resp->rh.rsmi_num_monitor_devices},
+      {"rsmi_dev_id_get", (void*)&resp->rh.rsmi_dev_id_get},
+      {"rsmi_dev_name_get", (void *)&resp->rh.rsmi_dev_name_get},
+      {"rsmi_dev_brand_get", (void *)&resp->rh.rsmi_dev_brand_get},
+      {"rsmi_dev_vendor_name_get", (void *)&resp->rh.rsmi_dev_vendor_name_get},
+      {"rsmi_dev_vram_vendor_get", (void *)&resp->rh.rsmi_dev_vram_vendor_get},
+      {"rsmi_dev_serial_number_get", (void *)&resp->rh.rsmi_dev_serial_number_get},
+      {"rsmi_dev_subsystem_name_get", (void *)&resp->rh.rsmi_dev_subsystem_name_get},
+      {"rsmi_dev_vbios_version_get", (void *)&resp->rh.rsmi_dev_vbios_version_get},
+      {NULL, NULL},
   };
 
-  for (i = 0; rocm_lib_paths[i] != NULL && resp->rh.handle == NULL; i++) {
-    resp->rh.handle = LOAD_LIBRARY(rocm_lib_paths[i], RTLD_LAZY);
-  }
+  resp->rh.handle = LOAD_LIBRARY(rocm_lib_path, RTLD_LAZY);
   if (!resp->rh.handle) {
     char *msg = LOAD_ERR();
     snprintf(buf, buflen,
              "Unable to load %s library to query for Radeon GPUs: %s\n",
-             rocm_lib_paths[0], msg);
+             rocm_lib_path, msg);
     free(msg);
     resp->err = strdup(buf);
     return;
   }
 
-  for (i = 0; i < 4; i++) {
+  // TODO once we've squashed the remaining corner cases remove this log
+  LOG(resp->rh.verbose, "wiring rocm management library functions in %s\n", rocm_lib_path);
+
+  for (i = 0; l[i].s != NULL; i++) {
+    // TODO once we've squashed the remaining corner cases remove this log
+    LOG(resp->rh.verbose, "dlsym: %s\n", l[i].s);
+
     *l[i].p = LOAD_SYMBOL(resp->rh.handle, l[i].s);
     if (!l[i].p) {
-      UNLOAD_LIBRARY(resp->rh.handle);
+      resp->rh.handle = NULL;
       char *msg = LOAD_ERR();
+      LOG(resp->rh.verbose, "dlerr: %s\n", msg);
+      UNLOAD_LIBRARY(resp->rh.handle);
       snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
                msg);
       free(msg);
@@ -62,8 +63,11 @@ void rocm_init(rocm_init_resp_t *resp) {
     }
   }
 
-  ret = (*resp->rh.initFn)(0);
+  ret = (*resp->rh.rsmi_init)(0);
   if (ret != RSMI_STATUS_SUCCESS) {
+    LOG(resp->rh.verbose, "rsmi_init err: %d\n", ret);
+    UNLOAD_LIBRARY(resp->rh.handle);
+    resp->rh.handle = NULL;
     snprintf(buf, buflen, "rocm vram init failure: %d", ret);
     resp->err = strdup(buf);
   }
@@ -73,8 +77,7 @@ void rocm_init(rocm_init_resp_t *resp) {
 
 void rocm_check_vram(rocm_handle_t h, mem_info_t *resp) {
   resp->err = NULL;
-  // uint32_t num_devices;
-  // uint16_t device;
+  resp->igpu_index = -1;
   uint64_t totalMem = 0;
   uint64_t usedMem = 0;
   rsmi_status_t ret;
@@ -83,36 +86,113 @@ void rocm_check_vram(rocm_handle_t h, mem_info_t *resp) {
   int i;
 
   if (h.handle == NULL) {
-    resp->err = strdup("nvml handle sn't initialized");
+    resp->err = strdup("rocm handle not initialized");
     return;
   }
 
-  // TODO - iterate through devices...  ret =
-  // rsmi_num_monitor_devices(&num_devices);
-
-  // ret = (*h.getHandle)(0, &device);
-  // if (ret != RSMI_STATUS_SUCCESS) {
-  //     printf("rocm vram device lookup failure: %d\n", ret);
-  //     return -1;
-  // }
-
-  // Get total memory - used memory for available memory
-  ret = (*h.totalMemFn)(0, RSMI_MEM_TYPE_VRAM, &totalMem);
+  ret = (*h.rsmi_num_monitor_devices)(&resp->count);
   if (ret != RSMI_STATUS_SUCCESS) {
-    snprintf(buf, buflen, "rocm total mem lookup failure: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-  ret = (*h.usageMemFn)(0, RSMI_MEM_TYPE_VRAM, &usedMem);
-  if (ret != RSMI_STATUS_SUCCESS) {
-    snprintf(buf, buflen, "rocm usage mem lookup failure: %d", ret);
+    snprintf(buf, buflen, "unable to get device count: %d", ret);
     resp->err = strdup(buf);
     return;
   }
+  LOG(h.verbose, "discovered %d ROCm GPU Devices\n", resp->count);
 
-  resp->total = totalMem;
-  resp->free = totalMem - usedMem;
-  return;
+  resp->total = 0;
+  resp->free = 0;
+  for (i = 0; i < resp->count; i++) {
+    if (h.verbose) {
+      // When in verbose mode, report more information about
+      // the card we discover, but don't fail on error
+      ret = (*h.rsmi_dev_name_get)(i, buf, buflen);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "rsmi_dev_name_get failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] ROCm device name: %s\n", i, buf);
+      }
+      ret = (*h.rsmi_dev_brand_get)(i, buf, buflen);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "rsmi_dev_brand_get failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] ROCm brand: %s\n", i, buf);
+      }
+      ret = (*h.rsmi_dev_vendor_name_get)(i, buf, buflen);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "rsmi_dev_vendor_name_get failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] ROCm vendor: %s\n", i, buf);
+      }
+      ret = (*h.rsmi_dev_vram_vendor_get)(i, buf, buflen);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "rsmi_dev_vram_vendor_get failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] ROCm VRAM vendor: %s\n", i, buf);
+      }
+      ret = (*h.rsmi_dev_serial_number_get)(i, buf, buflen);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "rsmi_dev_serial_number_get failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] ROCm S/N: %s\n", i, buf);
+      }
+      ret = (*h.rsmi_dev_subsystem_name_get)(i, buf, buflen);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "rsmi_dev_subsystem_name_get failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] ROCm subsystem name: %s\n", i, buf);
+      }
+      ret = (*h.rsmi_dev_vbios_version_get)(i, buf, buflen);
+      if (ret != RSMI_STATUS_SUCCESS) {
+        LOG(h.verbose, "rsmi_dev_vbios_version_get failed: %d\n", ret);
+      } else {
+        LOG(h.verbose, "[%d] ROCm vbios version: %s\n", i, buf);
+      }
+    }
+
+    // Get total memory - used memory for available memory
+    ret = (*h.rsmi_dev_memory_total_get)(i, RSMI_MEM_TYPE_VRAM, &totalMem);
+    if (ret != RSMI_STATUS_SUCCESS) {
+      snprintf(buf, buflen, "rocm total mem lookup failure: %d", ret);
+      resp->err = strdup(buf);
+      return;
+    }
+    ret = (*h.rsmi_dev_memory_usage_get)(i, RSMI_MEM_TYPE_VRAM, &usedMem);
+    if (ret != RSMI_STATUS_SUCCESS) {
+      snprintf(buf, buflen, "rocm usage mem lookup failure: %d", ret);
+      resp->err = strdup(buf);
+      return;
+    }
+    LOG(h.verbose, "[%d] ROCm totalMem %ld\n", i, totalMem);
+    LOG(h.verbose, "[%d] ROCm usedMem %ld\n", i, usedMem);
+    if (totalMem < 1024 * 1024 * 1024) {
+      // Do not add up integrated GPU memory capacity, it's a bogus 512M, and actually uses system memory
+      LOG(h.verbose, "[%d] ROCm integrated GPU\n", i);
+      resp->igpu_index = i;
+    } else {
+      resp->total += totalMem;
+      resp->free += totalMem - usedMem;
+    }
+  }
 }
 
-#endif  // __APPLE__
+void rocm_get_version(rocm_handle_t h, rocm_version_resp_t *resp) {
+  const int buflen = 256;
+  char buf[buflen + 1];
+  if (h.handle == NULL) {
+    resp->str = strdup("rocm handle not initialized");
+    resp->status = 1;
+    return;
+  }
+  rsmi_version_t ver;
+  rsmi_status_t ret;
+  ret = h.rsmi_version_get(&ver);
+  if (ret != RSMI_STATUS_SUCCESS) {
+    snprintf(buf, buflen, "unexpected response on version lookup %d", ret);
+    resp->status = 1;
+  } else {
+    snprintf(buf, buflen, "%d", ver.major);
+    resp->status = 0;
+  }
+  resp->str = strdup(buf);
+}
+
+#endif  // __APPLE__

gpu/gpu_info_rocm.h

@@ -15,13 +15,30 @@ typedef enum rsmi_memory_type {
   RSMI_MEM_TYPE_GTT,
 } rsmi_memory_type_t;
 
+ typedef struct {
+     uint32_t major;     
+     uint32_t minor;     
+     uint32_t patch;     
+     const char *build;  
+ } rsmi_version_t;
+
 typedef struct rocm_handle {
   void *handle;
-  rsmi_status_t (*initFn)(uint64_t);
-  rsmi_status_t (*shutdownFn)(void);
-  rsmi_status_t (*totalMemFn)(uint32_t, rsmi_memory_type_t, uint64_t *);
-  rsmi_status_t (*usageMemFn)(uint32_t, rsmi_memory_type_t, uint64_t *);
-  // rsmi_status_t (*getHandle)(uint32_t, uint16_t *);
+  uint16_t verbose;
+  rsmi_status_t (*rsmi_init)(uint64_t);
+  rsmi_status_t (*rsmi_shut_down)(void);
+  rsmi_status_t (*rsmi_dev_memory_total_get)(uint32_t, rsmi_memory_type_t, uint64_t *);
+  rsmi_status_t (*rsmi_dev_memory_usage_get)(uint32_t, rsmi_memory_type_t, uint64_t *);
+  rsmi_status_t (*rsmi_version_get) (rsmi_version_t *version);
+  rsmi_status_t (*rsmi_num_monitor_devices) (uint32_t *);
+  rsmi_status_t (*rsmi_dev_id_get)(uint32_t, uint16_t *);
+  rsmi_status_t (*rsmi_dev_name_get) (uint32_t,char *,size_t);
+  rsmi_status_t (*rsmi_dev_brand_get) (uint32_t, char *, uint32_t);		
+  rsmi_status_t (*rsmi_dev_vendor_name_get) (uint32_t, char *, uint32_t);		
+  rsmi_status_t (*rsmi_dev_vram_vendor_get) (uint32_t, char *, uint32_t);		
+  rsmi_status_t (*rsmi_dev_serial_number_get) (uint32_t, char *, uint32_t);		
+  rsmi_status_t (*rsmi_dev_subsystem_name_get) (uint32_t, char *, uint32_t);		
+  rsmi_status_t (*rsmi_dev_vbios_version_get) (uint32_t, char *, uint32_t);		
 } rocm_handle_t;
 
 typedef struct rocm_init_resp {
@@ -29,8 +46,14 @@ typedef struct rocm_init_resp {
   rocm_handle_t rh;
 } rocm_init_resp_t;
 
-void rocm_init(rocm_init_resp_t *resp);
+typedef struct rocm_version_resp {
+  rsmi_status_t status;
+  char *str; // Contains version or error string if status != 0 
+} rocm_version_resp_t;
+
+void rocm_init(char *rocm_lib_path, rocm_init_resp_t *resp);
 void rocm_check_vram(rocm_handle_t rh, mem_info_t *resp);
+void rocm_get_version(rocm_handle_t rh, rocm_version_resp_t *resp);
 
 #endif  // __GPU_INFO_ROCM_H__
 #endif  // __APPLE__

gpu/gpu_test.go

@@ -9,7 +9,7 @@ import (
 
 func TestBasicGetGPUInfo(t *testing.T) {
 	info := GetGPUInfo()
-	assert.Contains(t, "cuda rocm cpu default", info.Library)
+	assert.Contains(t, "cuda rocm cpu metal", info.Library)
 
 	switch runtime.GOOS {
 	case "darwin":
@@ -18,6 +18,7 @@ func TestBasicGetGPUInfo(t *testing.T) {
 	case "linux", "windows":
 		assert.Greater(t, info.TotalMemory, uint64(0))
 		assert.Greater(t, info.FreeMemory, uint64(0))
+		assert.Greater(t, info.DeviceCount, uint32(0))
 	default:
 		return
 	}
@@ -35,7 +36,6 @@ func TestCPUMemInfo(t *testing.T) {
 	default:
 		return
 	}
-
 }
 
 // TODO - add some logic to figure out card type through other means and actually verify we got back what we expected

gpu/types.go

@@ -3,6 +3,7 @@ package gpu
 type memInfo struct {
 	TotalMemory uint64 `json:"total_memory,omitempty"`
 	FreeMemory  uint64 `json:"free_memory,omitempty"`
+	DeviceCount uint32 `json:"device_count,omitempty"`
 }
 
 // Beginning of an `ollama info` command
@@ -10,5 +11,14 @@ type GpuInfo struct {
 	memInfo
 	Library string `json:"library,omitempty"`
 
+	// Optional variant to select (e.g. versions, cpu feature flags)
+	Variant string `json:"variant,omitempty"`
+
 	// TODO add other useful attributes about the card here for discovery information
 }
+
+type Version struct {
+	Major uint
+	Minor uint
+	Patch uint
+}

llm/dyn_ext_server.c

@@ -1,11 +1,11 @@
-#include "dynamic_shim.h"
+#include "dyn_ext_server.h"
 
 #include <stdio.h>
 #include <string.h>
 
 #ifdef __linux__
 #include <dlfcn.h>
-#define LOAD_LIBRARY(lib, flags) dlopen(lib, flags | RTLD_DEEPBIND)
+#define LOAD_LIBRARY(lib, flags) dlopen(lib, flags)
 #define LOAD_SYMBOL(handle, sym) dlsym(handle, sym)
 #define LOAD_ERR() strdup(dlerror())
 #define UNLOAD_LIBRARY(handle) dlclose(handle)
@@ -33,7 +33,7 @@ inline char *LOAD_ERR() {
 #define UNLOAD_LIBRARY(handle) dlclose(handle)
 #endif
 
-void dynamic_shim_init(const char *libPath, struct dynamic_llama_server *s,
+void dyn_init(const char *libPath, struct dynamic_llama_server *s,
                        ext_server_resp_t *err) {
   int i = 0;
   struct lookup {
@@ -58,8 +58,8 @@ void dynamic_shim_init(const char *libPath, struct dynamic_llama_server *s,
       {"", NULL},
   };
 
-  printf("Lazy loading %s library\n", libPath);
-  s->handle = LOAD_LIBRARY(libPath, RTLD_NOW);
+  printf("loading library %s\n", libPath);
+  s->handle = LOAD_LIBRARY(libPath, RTLD_LOCAL|RTLD_NOW);
   if (!s->handle) {
     err->id = -1;
     char *msg = LOAD_ERR();
@@ -83,63 +83,63 @@ void dynamic_shim_init(const char *libPath, struct dynamic_llama_server *s,
   }
 }
 
-inline void dynamic_shim_llama_server_init(struct dynamic_llama_server s,
+inline void dyn_llama_server_init(struct dynamic_llama_server s,
                                            ext_server_params_t *sparams,
                                            ext_server_resp_t *err) {
   s.llama_server_init(sparams, err);
 }
 
-inline void dynamic_shim_llama_server_start(struct dynamic_llama_server s) {
+inline void dyn_llama_server_start(struct dynamic_llama_server s) {
   s.llama_server_start();
 }
 
-inline void dynamic_shim_llama_server_stop(struct dynamic_llama_server s) {
+inline void dyn_llama_server_stop(struct dynamic_llama_server s) {
   s.llama_server_stop();
 }
 
-inline void dynamic_shim_llama_server_completion(struct dynamic_llama_server s,
+inline void dyn_llama_server_completion(struct dynamic_llama_server s,
                                                  const char *json_req,
                                                  ext_server_resp_t *resp) {
   s.llama_server_completion(json_req, resp);
 }
 
-inline void dynamic_shim_llama_server_completion_next_result(
+inline void dyn_llama_server_completion_next_result(
     struct dynamic_llama_server s, const int task_id,
     ext_server_task_result_t *result) {
   s.llama_server_completion_next_result(task_id, result);
 }
 
-inline void dynamic_shim_llama_server_completion_cancel(
+inline void dyn_llama_server_completion_cancel(
     struct dynamic_llama_server s, const int task_id, ext_server_resp_t *err) {
   s.llama_server_completion_cancel(task_id, err);
 }
-inline void dynamic_shim_llama_server_release_task_result(
+inline void dyn_llama_server_release_task_result(
     struct dynamic_llama_server s, ext_server_task_result_t *result) {
   s.llama_server_release_task_result(result);
 }
 
-inline void dynamic_shim_llama_server_tokenize(struct dynamic_llama_server s,
+inline void dyn_llama_server_tokenize(struct dynamic_llama_server s,
                                                const char *json_req,
                                                char **json_resp,
                                                ext_server_resp_t *err) {
   s.llama_server_tokenize(json_req, json_resp, err);
 }
 
-inline void dynamic_shim_llama_server_detokenize(struct dynamic_llama_server s,
+inline void dyn_llama_server_detokenize(struct dynamic_llama_server s,
                                                  const char *json_req,
                                                  char **json_resp,
                                                  ext_server_resp_t *err) {
   s.llama_server_detokenize(json_req, json_resp, err);
 }
 
-inline void dynamic_shim_llama_server_embedding(struct dynamic_llama_server s,
+inline void dyn_llama_server_embedding(struct dynamic_llama_server s,
                                                 const char *json_req,
                                                 char **json_resp,
                                                 ext_server_resp_t *err) {
   s.llama_server_embedding(json_req, json_resp, err);
 }
 
-inline void dynamic_shim_llama_server_release_json_resp(
+inline void dyn_llama_server_release_json_resp(
     struct dynamic_llama_server s, char **json_resp) {
   s.llama_server_release_json_resp(json_resp);
 }

llm/dyn_ext_server.go

@@ -3,61 +3,43 @@ package llm
 /*
 #cgo CFLAGS: -I${SRCDIR}/ext_server -I${SRCDIR}/llama.cpp -I${SRCDIR}/llama.cpp/common -I${SRCDIR}/llama.cpp/examples/server
 #cgo CFLAGS: -DNDEBUG -DLLAMA_SERVER_LIBRARY=1 -D_XOPEN_SOURCE=600 -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64
-#cgo CFLAGS: -Wmissing-noreturn -Wall -Wextra -Wcast-qual -Wno-unused-function -Wno-array-bounds
-#cgo CPPFLAGS: -Ofast -Wall -Wextra -Wno-unused-function -Wno-unused-variable -Wno-deprecated-declarations -Wno-unused-but-set-variable
+#cgo CFLAGS: -Wmissing-noreturn -Wextra -Wcast-qual -Wno-unused-function -Wno-array-bounds
+#cgo CPPFLAGS: -Ofast -Wextra -Wno-unused-function -Wno-unused-variable -Wno-deprecated-declarations
 #cgo darwin CFLAGS: -D_DARWIN_C_SOURCE
 #cgo darwin CPPFLAGS:  -DGGML_USE_ACCELERATE
 #cgo darwin CPPFLAGS: -DGGML_USE_METAL -DGGML_METAL_NDEBUG
 #cgo darwin LDFLAGS: -lc++ -framework Accelerate
 #cgo darwin LDFLAGS: -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
-#cgo darwin LDFLAGS: ${SRCDIR}/llama.cpp/build/darwin/metal/lib/libcommon.a
-#cgo darwin LDFLAGS: ${SRCDIR}/llama.cpp/build/darwin/metal/lib/libext_server.a
-#cgo darwin LDFLAGS: ${SRCDIR}/llama.cpp/build/darwin/metal/lib/libllama.a
-#cgo darwin LDFLAGS: ${SRCDIR}/llama.cpp/build/darwin/metal/lib/libggml_static.a
 #cgo linux CFLAGS: -D_GNU_SOURCE
-#cgo linux windows CFLAGS: -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_MMV_Y=1 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_USE_CUBLAS
-#cgo linux LDFLAGS: -L/usr/local/cuda/targets/x86_64-linux/lib -L/usr/local/cuda/lib64 -L/usr/local/cuda/targets/x86_64-linux/lib/stubs
-#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/build/linux/cpu/lib/libext_server.a
-#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/build/linux/cpu/lib/libcommon.a
-#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/build/linux/cpu/lib/libllama.a
-#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/build/linux/cpu/lib/libggml_static.a
 #cgo linux LDFLAGS: -lrt -ldl -lstdc++ -lm
 #cgo linux windows LDFLAGS: -lpthread
 
 #include <stdlib.h>
-#include "ext_server.h"
+#include "dyn_ext_server.h"
 
 */
 import "C"
+
 import (
 	"bytes"
 	"context"
 	"encoding/json"
 	"fmt"
-	"log"
+	"log/slog"
 	"os"
+	"path/filepath"
+	"runtime"
 	"strings"
 	"sync"
 	"time"
 	"unsafe"
 
 	"github.com/jmorganca/ollama/api"
-	"github.com/jmorganca/ollama/gpu"
 )
 
-type extServer interface {
-	LLM
-	llama_server_init(sparams *C.ext_server_params_t, err *C.ext_server_resp_t)
-	llama_server_start()
-	llama_server_stop()
-	llama_server_completion(json_req *C.char, resp *C.ext_server_resp_t)
-	llama_server_completion_next_result(task_id C.int, resp *C.ext_server_task_result_t)
-	llama_server_completion_cancel(task_id C.int, err *C.ext_server_resp_t)
-	llama_server_release_task_result(result *C.ext_server_task_result_t)
-	llama_server_tokenize(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t)
-	llama_server_detokenize(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t)
-	llama_server_embedding(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t)
-	llama_server_release_json_resp(json_resp **C.char)
+type dynExtServer struct {
+	s       C.struct_dynamic_llama_server
+	options api.Options
 }
 
 // Note: current implementation does not support concurrent instantiations
@@ -82,25 +64,39 @@ func extServerResponseToErr(resp C.ext_server_resp_t) error {
 	return fmt.Errorf(C.GoString(resp.msg))
 }
 
-func newExtServer(server extServer, model string, adapters, projectors []string, numLayers int64, opts api.Options) (extServer, error) {
+// Note: current implementation does not support concurrent instantiations
+var llm *dynExtServer
+
+func newDynExtServer(library, model string, adapters, projectors []string, opts api.Options) (LLM, error) {
 	if !mutex.TryLock() {
-		log.Printf("concurrent llm servers not yet supported, waiting for prior server to complete")
+		slog.Info("concurrent llm servers not yet supported, waiting for prior server to complete")
 		mutex.Lock()
 	}
-	fileInfo, err := os.Stat(model)
-	if err != nil {
-		return nil, err
+	updatePath(filepath.Dir(library))
+	libPath := C.CString(library)
+	defer C.free(unsafe.Pointer(libPath))
+	resp := newExtServerResp(512)
+	defer freeExtServerResp(resp)
+	var srv C.struct_dynamic_llama_server
+	C.dyn_init(libPath, &srv, &resp)
+	if resp.id < 0 {
+		mutex.Unlock()
+		return nil, fmt.Errorf("Unable to load dynamic library: %s", C.GoString(resp.msg))
 	}
+	llm = &dynExtServer{
+		s:       srv,
+		options: opts,
+	}
+	slog.Info(fmt.Sprintf("Loading Dynamic llm server: %s", library))
+
 	var sparams C.ext_server_params_t
 	sparams.model = C.CString(model)
 	defer C.free(unsafe.Pointer(sparams.model))
 
-	numGPU := gpu.NumGPU(numLayers, fileInfo.Size(), opts)
-
 	sparams.embedding = true
 	sparams.n_ctx = C.uint(opts.NumCtx)
 	sparams.n_batch = C.uint(opts.NumBatch)
-	sparams.n_gpu_layers = C.int(numGPU)
+	sparams.n_gpu_layers = C.int(opts.NumGPU)
 	sparams.main_gpu = C.int(opts.MainGPU)
 	sparams.n_parallel = 1 // TODO - wire up concurrency
 
@@ -140,29 +136,35 @@ func newExtServer(server extServer, model string, adapters, projectors []string,
 
 	sparams.n_threads = C.uint(opts.NumThread)
 
-	log.Printf("Initializing internal llama server")
-	resp := newExtServerResp(128)
-	defer freeExtServerResp(resp)
-	server.llama_server_init(&sparams, &resp)
-	if resp.id < 0 {
-		return nil, extServerResponseToErr(resp)
+	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
+		sparams.verbose_logging = C.bool(true)
+	} else {
+		sparams.verbose_logging = C.bool(false)
 	}
 
-	log.Printf("Starting internal llama main loop")
-	server.llama_server_start()
-	return server, nil
+	slog.Info("Initializing llama server")
+	initResp := newExtServerResp(128)
+	defer freeExtServerResp(initResp)
+	C.dyn_llama_server_init(llm.s, &sparams, &initResp)
+	if initResp.id < 0 {
+		mutex.Unlock()
+		err := extServerResponseToErr(initResp)
+		slog.Debug(fmt.Sprintf("failure during initialization: %s", err))
+		return nil, err
+	}
+
+	slog.Info("Starting llama main loop")
+	C.dyn_llama_server_start(llm.s)
+	return llm, nil
 }
 
-func predict(ctx context.Context, llm extServer, predict PredictOpts, fn func(PredictResult)) error {
+func (llm *dynExtServer) Predict(ctx context.Context, predict PredictOpts, fn func(PredictResult)) error {
 	resp := newExtServerResp(128)
 	defer freeExtServerResp(resp)
-	var imageData []ImageData
+
 	if len(predict.Images) > 0 {
-		for cnt, i := range predict.Images {
-			imageData = append(imageData, ImageData{Data: i, ID: cnt})
-		}
+		slog.Info(fmt.Sprintf("loaded %d images", len(predict.Images)))
 	}
-	log.Printf("loaded %d images", len(imageData))
 
 	request := map[string]any{
 		"prompt":            predict.Prompt,
@@ -184,7 +186,7 @@ func predict(ctx context.Context, llm extServer, predict PredictOpts, fn func(Pr
 		"penalize_nl":       predict.Options.PenalizeNewline,
 		"seed":              predict.Options.Seed,
 		"stop":              predict.Options.Stop,
-		"image_data":        imageData,
+		"image_data":        predict.Images,
 		"cache_prompt":      true,
 	}
 
@@ -192,6 +194,7 @@ func predict(ctx context.Context, llm extServer, predict PredictOpts, fn func(Pr
 		request["grammar"] = jsonGrammar
 	}
 
+	var whitespace int
 	retryDelay := 100 * time.Microsecond
 	for retries := 0; retries < maxRetries; retries++ {
 		if retries > 0 {
@@ -211,7 +214,7 @@ func predict(ctx context.Context, llm extServer, predict PredictOpts, fn func(Pr
 		req := C.CString(buffer.String())
 		defer C.free(unsafe.Pointer(req))
 
-		llm.llama_server_completion(req, &resp)
+		C.dyn_llama_server_completion(llm.s, req, &resp)
 		if resp.id < 0 {
 			return extServerResponseToErr(resp)
 		}
@@ -222,7 +225,7 @@ func predict(ctx context.Context, llm extServer, predict PredictOpts, fn func(Pr
 			select {
 			case <-ctx.Done():
 				// This handles the request cancellation
-				llm.llama_server_completion_cancel(resp.id, &resp)
+				C.dyn_llama_server_completion_cancel(llm.s, resp.id, &resp)
 				if resp.id < 0 {
 					return extServerResponseToErr(resp)
 				} else {
@@ -230,13 +233,13 @@ func predict(ctx context.Context, llm extServer, predict PredictOpts, fn func(Pr
 				}
 			default:
 				var result C.ext_server_task_result_t
-				llm.llama_server_completion_next_result(resp.id, &result)
+				C.dyn_llama_server_completion_next_result(llm.s, resp.id, &result)
 				json_resp := C.GoString(result.json_resp)
-				llm.llama_server_release_task_result(&result)
+				C.dyn_llama_server_release_task_result(llm.s, &result)
 
 				var p prediction
 				if err := json.Unmarshal([]byte(json_resp), &p); err != nil {
-					llm.llama_server_completion_cancel(resp.id, &resp)
+					C.dyn_llama_server_completion_cancel(llm.s, resp.id, &resp)
 					if resp.id < 0 {
 						return fmt.Errorf("error unmarshaling llm prediction response: %w and cancel %s", err, C.GoString(resp.msg))
 					} else {
@@ -250,13 +253,31 @@ func predict(ctx context.Context, llm extServer, predict PredictOpts, fn func(Pr
 					break out
 				}
 
+				// detect if p.Content is entirely whitespace
+				if predict.Format == "json" && strings.TrimSpace(p.Content) == "" {
+					whitespace++
+
+					// if we get 100 consecutive whitespace responses, cancel
+					if whitespace > 100 {
+						slog.Debug("cancelling due to excessive whitespace")
+						C.dyn_llama_server_completion_cancel(llm.s, resp.id, &resp)
+						if resp.id < 0 {
+							return extServerResponseToErr(resp)
+						}
+
+						return nil
+					}
+				} else {
+					whitespace = 0
+				}
+
 				if p.Content != "" {
 					fn(PredictResult{
 						Content: p.Content,
 					})
 				}
 
-				if p.Stop {
+				if p.Stop || bool(result.stop) {
 					fn(PredictResult{
 						Done:               true,
 						PromptEvalCount:    p.Timings.PromptN,
@@ -277,7 +298,7 @@ func predict(ctx context.Context, llm extServer, predict PredictOpts, fn func(Pr
 	return fmt.Errorf("max retries exceeded")
 }
 
-func encode(llm extServer, ctx context.Context, prompt string) ([]int, error) {
+func (llm *dynExtServer) Encode(ctx context.Context, prompt string) ([]int, error) {
 	data, err := json.Marshal(TokenizeRequest{Content: prompt})
 	if err != nil {
 		return nil, fmt.Errorf("marshaling encode data: %w", err)
@@ -287,11 +308,11 @@ func encode(llm extServer, ctx context.Context, prompt string) ([]int, error) {
 	var json_resp *C.char
 	resp := newExtServerResp(128)
 	defer freeExtServerResp(resp)
-	llm.llama_server_tokenize(req, &json_resp, &resp)
+	C.dyn_llama_server_tokenize(llm.s, req, &json_resp, &resp)
 	if resp.id < 0 {
 		return nil, extServerResponseToErr(resp)
 	}
-	defer llm.llama_server_release_json_resp(&json_resp)
+	defer C.dyn_llama_server_release_json_resp(llm.s, &json_resp)
 
 	var encoded TokenizeResponse
 	if err2 := json.Unmarshal([]byte(C.GoString(json_resp)), &encoded); err2 != nil {
@@ -301,7 +322,7 @@ func encode(llm extServer, ctx context.Context, prompt string) ([]int, error) {
 	return encoded.Tokens, err
 }
 
-func decode(llm extServer, ctx context.Context, tokens []int) (string, error) {
+func (llm *dynExtServer) Decode(ctx context.Context, tokens []int) (string, error) {
 	if len(tokens) == 0 {
 		return "", nil
 	}
@@ -315,11 +336,11 @@ func decode(llm extServer, ctx context.Context, tokens []int) (string, error) {
 	var json_resp *C.char
 	resp := newExtServerResp(128)
 	defer freeExtServerResp(resp)
-	llm.llama_server_detokenize(req, &json_resp, &resp)
+	C.dyn_llama_server_detokenize(llm.s, req, &json_resp, &resp)
 	if resp.id < 0 {
 		return "", extServerResponseToErr(resp)
 	}
-	defer llm.llama_server_release_json_resp(&json_resp)
+	defer C.dyn_llama_server_release_json_resp(llm.s, &json_resp)
 
 	var decoded DetokenizeResponse
 	if err2 := json.Unmarshal([]byte(C.GoString(json_resp)), &decoded); err2 != nil {
@@ -329,7 +350,7 @@ func decode(llm extServer, ctx context.Context, tokens []int) (string, error) {
 	return decoded.Content, err
 }
 
-func embedding(llm extServer, ctx context.Context, input string) ([]float64, error) {
+func (llm *dynExtServer) Embedding(ctx context.Context, input string) ([]float64, error) {
 	data, err := json.Marshal(TokenizeRequest{Content: input})
 	if err != nil {
 		return nil, fmt.Errorf("error marshaling embed data: %w", err)
@@ -340,11 +361,11 @@ func embedding(llm extServer, ctx context.Context, input string) ([]float64, err
 	var json_resp *C.char
 	resp := newExtServerResp(128)
 	defer freeExtServerResp(resp)
-	llm.llama_server_embedding(req, &json_resp, &resp)
+	C.dyn_llama_server_embedding(llm.s, req, &json_resp, &resp)
 	if resp.id < 0 {
 		return nil, extServerResponseToErr(resp)
 	}
-	defer llm.llama_server_release_json_resp(&json_resp)
+	defer C.dyn_llama_server_release_json_resp(llm.s, &json_resp)
 
 	var embedding EmbeddingResponse
 	if err := json.Unmarshal([]byte(C.GoString(json_resp)), &embedding); err != nil {
@@ -354,7 +375,29 @@ func embedding(llm extServer, ctx context.Context, input string) ([]float64, err
 	return embedding.Embedding, nil
 }
 
-func close(llm extServer) {
-	llm.llama_server_stop()
+func (llm *dynExtServer) Close() {
+	C.dyn_llama_server_stop(llm.s)
 	mutex.Unlock()
 }
+
+func updatePath(dir string) {
+	if runtime.GOOS == "windows" {
+		tmpDir := filepath.Dir(dir)
+		pathComponents := strings.Split(os.Getenv("PATH"), ";")
+		i := 0
+		for _, comp := range pathComponents {
+			if strings.EqualFold(comp, dir) {
+				return
+			}
+			// Remove any other prior paths to our temp dir
+			if !strings.HasPrefix(strings.ToLower(comp), strings.ToLower(tmpDir)) {
+				pathComponents[i] = comp
+				i++
+			}
+		}
+		newPath := strings.Join(append([]string{dir}, pathComponents...), ";")
+		slog.Info(fmt.Sprintf("Updating PATH to %s", newPath))
+		os.Setenv("PATH", newPath)
+	}
+	// linux and darwin rely on rpath
+}

llm/dyn_ext_server.h

@@ -27,46 +27,46 @@ struct dynamic_llama_server {
   void (*llama_server_release_json_resp)(char **json_resp);
 };
 
-void dynamic_shim_init(const char *libPath, struct dynamic_llama_server *s,
+void dyn_init(const char *libPath, struct dynamic_llama_server *s,
                        ext_server_resp_t *err);
 
 // No good way to call C function pointers from Go so inline the indirection
-void dynamic_shim_llama_server_init(struct dynamic_llama_server s,
+void dyn_llama_server_init(struct dynamic_llama_server s,
                                     ext_server_params_t *sparams,
                                     ext_server_resp_t *err);
 
-void dynamic_shim_llama_server_start(struct dynamic_llama_server s);
+void dyn_llama_server_start(struct dynamic_llama_server s);
 
-void dynamic_shim_llama_server_stop(struct dynamic_llama_server s);
+void dyn_llama_server_stop(struct dynamic_llama_server s);
 
-void dynamic_shim_llama_server_completion(struct dynamic_llama_server s,
+void dyn_llama_server_completion(struct dynamic_llama_server s,
                                           const char *json_req,
                                           ext_server_resp_t *resp);
 
-void dynamic_shim_llama_server_completion_next_result(
+void dyn_llama_server_completion_next_result(
     struct dynamic_llama_server s, const int task_id,
     ext_server_task_result_t *result);
 
-void dynamic_shim_llama_server_completion_cancel(struct dynamic_llama_server s,
+void dyn_llama_server_completion_cancel(struct dynamic_llama_server s,
                                                  const int task_id,
                                                  ext_server_resp_t *err);
 
-void dynamic_shim_llama_server_release_task_result(
+void dyn_llama_server_release_task_result(
     struct dynamic_llama_server s, ext_server_task_result_t *result);
 
-void dynamic_shim_llama_server_tokenize(struct dynamic_llama_server s,
+void dyn_llama_server_tokenize(struct dynamic_llama_server s,
                                         const char *json_req, char **json_resp,
                                         ext_server_resp_t *err);
 
-void dynamic_shim_llama_server_detokenize(struct dynamic_llama_server s,
+void dyn_llama_server_detokenize(struct dynamic_llama_server s,
                                           const char *json_req,
                                           char **json_resp,
                                           ext_server_resp_t *err);
 
-void dynamic_shim_llama_server_embedding(struct dynamic_llama_server s,
+void dyn_llama_server_embedding(struct dynamic_llama_server s,
                                          const char *json_req, char **json_resp,
                                          ext_server_resp_t *err);
-void dynamic_shim_llama_server_release_json_resp(struct dynamic_llama_server s,
+void dyn_llama_server_release_json_resp(struct dynamic_llama_server s,
                                                  char **json_resp);
 
 #ifdef __cplusplus

llm/ext_server/CMakeLists.txt

@@ -2,28 +2,24 @@
 
 set(TARGET ext_server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
-add_library(${TARGET} STATIC ../../../ext_server/ext_server.cpp)
+if (WIN32)
+    add_library(${TARGET} SHARED ../../../ext_server/ext_server.cpp ../../llama.cpp)
+else()
+    add_library(${TARGET} STATIC ../../../ext_server/ext_server.cpp ../../llama.cpp)
+endif()
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_include_directories(${TARGET} PRIVATE ../..)
 target_include_directories(${TARGET} PRIVATE ../../..)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 target_compile_definitions(${TARGET} PUBLIC LLAMA_SERVER_LIBRARY=1)
-target_link_libraries(${TARGET} PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT})
-target_compile_definitions(${TARGET} PRIVATE
-    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
-)
-
-if (BUILD_SHARED_LIBS)
-    set_target_properties(ext_server PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_compile_definitions(ext_server PRIVATE LLAMA_SHARED LLAMA_BUILD)
-    add_library(ext_server_shared SHARED $<TARGET_OBJECTS:ext_server>)
-    target_link_libraries(ext_server_shared PRIVATE ggml llama llava common ${CMAKE_THREAD_LIBS_INIT})
-    install(TARGETS ext_server_shared LIBRARY)
-endif()
+target_link_libraries(${TARGET} PRIVATE ggml llava common )
+set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+target_compile_definitions(${TARGET} PRIVATE SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>)
+install(TARGETS ext_server LIBRARY)
 
 if (CUDAToolkit_FOUND)
     target_include_directories(${TARGET} PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
     if (WIN32)
-        target_link_libraries(ext_server_shared PRIVATE nvml)
+        target_link_libraries(${TARGET} PRIVATE nvml)
     endif()
 endif()

llm/ext_server/README.md

@@ -1,4 +1,18 @@
 # Extern C Server
 
-This directory contains a thin facade we layer on top of the Llama.cpp server
-to expose `extern C` interfaces to access the functionality through direct API calls in-process
+This directory contains a thin facade we layer on top of the Llama.cpp server to
+expose `extern C` interfaces to access the functionality through direct API
+calls in-process.  The llama.cpp code uses compile time macros to configure GPU
+type along with other settings.  During the `go generate ./...` execution, the
+build will generate one or more copies of the llama.cpp `extern C` server based
+on what GPU libraries are detected to support multiple GPU types as well as CPU
+only support. The Ollama go build then embeds these different servers to support
+different GPUs and settings at runtime.
+
+If you are making changes to the code in this directory, make sure to disable
+caching during your go build to ensure you pick up your changes.  A typical
+iteration cycle from the top of the source tree looks like:
+
+```
+go generate ./... && go build -a .
+```

llm/ext_server/ext_server.cpp

@@ -1,23 +1,63 @@
 #include "ext_server.h"
+#include <atomic>
 
 // Necessary evil since the server types are not defined in a header
 #include "server.cpp"
 
+// Low level API access to verify GPU access
+#if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_HIPBLAS)
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#include <hip/hip_fp16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif // __HIP_PLATFORM_AMD__
+#define cudaGetDevice hipGetDevice
+#define cudaError_t hipError_t
+#define cudaSuccess hipSuccess
+#define cudaGetErrorString hipGetErrorString
+#else
+#include <cuda_runtime.h>
+#include <cublas_v2.h>
+#include <cuda_fp16.h>
+#endif // defined(GGML_USE_HIPBLAS)
+#endif // GGML_USE_CUBLAS
+
 // Expose the llama server as a callable extern "C" API
 llama_server_context *llama = NULL;
-std::atomic<bool> ext_server_running(false);
 std::thread ext_server_thread;
+bool shutting_down = false;
+std::atomic_int recv_counter;
 
+// RAII wrapper for tracking in-flight recv calls
+class atomicRecv {
+  public:
+    atomicRecv(std::atomic<int> &atomic) : atomic(atomic) {
+      ++this->atomic;
+    }
+    ~atomicRecv() {
+      --this->atomic;
+    }
+  private:
+    std::atomic<int> &atomic;
+};
+ 
 void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err) {
-#if SERVER_VERBOSE != 1
-  log_disable();
-#endif
+  recv_counter = 0;
   assert(err != NULL && sparams != NULL);
+  log_set_target(stderr);
+  if (!sparams->verbose_logging) {
+    server_verbose = true;
+    log_disable();
+  }
+
+  LOG_TEE("system info: %s\n", llama_print_system_info());
   err->id = 0;
   err->msg[0] = '\0';
   try {
     llama = new llama_server_context;
-    log_set_target(stdout);
     gpt_params params;
     params.n_ctx = sparams->n_ctx;
     params.n_batch = sparams->n_batch;
@@ -46,15 +86,31 @@ void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err) {
       params.model = sparams->model;
     }
 
-    for (ext_server_lora_adapter *la = sparams->lora_adapters; la != NULL;
-         la = la->next) {
-      params.lora_adapter.push_back(std::make_tuple(la->adapter, la->scale));
+    if (sparams->lora_adapters != NULL) {
+      for (ext_server_lora_adapter *la = sparams->lora_adapters; la != NULL;
+          la = la->next) {
+        params.lora_adapter.push_back(std::make_tuple(la->adapter, la->scale));
+      }
+
+      params.use_mmap = false;
     }
 
     if (sparams->mmproj != NULL) {
       params.mmproj = std::string(sparams->mmproj);
     }
 
+#if defined(GGML_USE_CUBLAS)
+    // Before attempting to init the backend which will assert on error, verify the CUDA/ROCM GPU is accessible
+    LOG_TEE("Performing pre-initialization of GPU\n");
+    int id;
+    cudaError_t cudaErr = cudaGetDevice(&id);
+    if (cudaErr != cudaSuccess) {
+      err->id = -1;
+      snprintf(err->msg, err->msg_len, "Unable to init GPU: %s", cudaGetErrorString(cudaErr));
+      return;
+    }
+#endif
+
     llama_backend_init(params.numa);
 
     // load the model
@@ -83,18 +139,23 @@ void llama_server_start() {
   assert(llama != NULL);
   // TODO mutex to protect thread creation
   ext_server_thread = std::thread([&]() {
-    ext_server_running = true;
     try {
       LOG_TEE("llama server main loop starting\n");
       ggml_time_init();
-      while (ext_server_running.load()) {
-        if (!llama->update_slots()) {
-          LOG_TEE(
-              "unexpected error in llama server update_slots - exiting main "
-              "loop\n");
-          break;
-        }
-      }
+      llama->queue_tasks.on_new_task(std::bind(
+        &llama_server_context::process_single_task, llama, std::placeholders::_1));
+      llama->queue_tasks.on_finish_multitask(std::bind(
+          &llama_server_context::on_finish_multitask, llama, std::placeholders::_1));
+      llama->queue_tasks.on_all_tasks_finished(std::bind(
+          &llama_server_context::run_on_all_tasks_finished, llama));
+      llama->queue_results.on_multitask_update(std::bind(
+          &llama_server_queue::update_multitask,
+          &llama->queue_tasks,
+          std::placeholders::_1,
+          std::placeholders::_2,
+          std::placeholders::_3
+        ));
+      llama->queue_tasks.start_loop();
     } catch (std::exception &e) {
       LOG_TEE("caught exception in llama server main loop: %s\n", e.what());
     } catch (...) {
@@ -107,13 +168,22 @@ void llama_server_start() {
 
 void llama_server_stop() {
   assert(llama != NULL);
-  // TODO - too verbose, remove once things are solid
-  LOG_TEE("requesting llama server shutdown\n");
-  ext_server_running = false;
+  // Shutdown any in-flight requests and block incoming requests.
+  LOG_TEE("\ninitiating shutdown - draining remaining tasks...\n");
+  shutting_down = true;
+
+  while (recv_counter.load() > 0) {
+    std::this_thread::sleep_for(std::chrono::milliseconds(50));
+  }
+
+  // This may take a while for any pending tasks to drain
+  // TODO - consider a timeout to cancel tasks if it's taking too long
+  llama->queue_tasks.terminate();
   ext_server_thread.join();
   delete llama;
   llama = NULL;
   LOG_TEE("llama server shutdown complete\n");
+  shutting_down = false;
 }
 
 void llama_server_completion(const char *json_req, ext_server_resp_t *resp) {
@@ -121,8 +191,13 @@ void llama_server_completion(const char *json_req, ext_server_resp_t *resp) {
   resp->id = -1;
   resp->msg[0] = '\0';
   try {
+    if (shutting_down) {
+      throw std::runtime_error("server shutting down");
+    }
     json data = json::parse(json_req);
-    resp->id = llama->request_completion(data, false, false, -1);
+    resp->id = llama->queue_tasks.get_new_id();
+    llama->queue_results.add_waiting_task_id(resp->id);
+    llama->request_completion(resp->id, data, false, false, -1);
   } catch (std::exception &e) {
     snprintf(resp->msg, resp->msg_len, "exception %s", e.what());
   } catch (...) {
@@ -140,16 +215,28 @@ void llama_server_completion_next_result(const int task_id,
   resp->json_resp = NULL;
   std::string result_json;
   try {
-    task_result result = llama->next_result(task_id);
+    atomicRecv ar(recv_counter);
+    task_result result = llama->queue_results.recv(task_id);
     result_json =
         result.result_json.dump(-1, ' ', false, json::error_handler_t::replace);
     resp->id = result.id;
     resp->stop = result.stop;
     resp->error = result.error;
     if (result.error) {
+      LOG_TEE("next result cancel on error\n");
       llama->request_cancel(task_id);
+      LOG_TEE("next result removing waiting tak ID: %d\n", task_id);
+      llama->queue_results.remove_waiting_task_id(task_id);
     } else if (result.stop) {
+      LOG_TEE("next result cancel on stop\n");
       llama->request_cancel(task_id);
+      LOG_TEE("next result removing waiting task ID: %d\n", task_id);
+      llama->queue_results.remove_waiting_task_id(task_id);
+    } else if (shutting_down) {
+      LOG_TEE("aborting completion due to shutdown %d\n", task_id);
+      llama->request_cancel(task_id);
+      llama->queue_results.remove_waiting_task_id(task_id);
+      resp->stop = true;
     }
   } catch (std::exception &e) {
     resp->error = true;
@@ -180,6 +267,7 @@ void llama_server_completion_cancel(const int task_id, ext_server_resp_t *err) {
   err->msg[0] = '\0';
   try {
     llama->request_cancel(task_id);
+    llama->queue_results.remove_waiting_task_id(task_id);
   } catch (std::exception &e) {
     err->id = -1;
     snprintf(err->msg, err->msg_len, "exception %s", e.what());
@@ -197,6 +285,9 @@ void llama_server_tokenize(const char *json_req, char **json_resp,
   err->id = 0;
   err->msg[0] = '\0';
   try {
+    if (shutting_down) {
+      throw std::runtime_error("server shutting down");
+    }
     const json body = json::parse(json_req);
     std::vector<llama_token> tokens;
     if (body.count("content") != 0) {
@@ -230,6 +321,9 @@ void llama_server_detokenize(const char *json_req, char **json_resp,
   err->id = 0;
   err->msg[0] = '\0';
   try {
+    if (shutting_down) {
+      throw std::runtime_error("server shutting down");
+    }
     const json body = json::parse(json_req);
     std::string content;
     if (body.count("tokens") != 0) {
@@ -257,6 +351,9 @@ void llama_server_embedding(const char *json_req, char **json_resp,
   err->id = 0;
   err->msg[0] = '\0';
   try {
+    if (shutting_down) {
+      throw std::runtime_error("server shutting down");
+    }
     const json body = json::parse(json_req);
     json prompt;
     if (body.count("content") != 0) {
@@ -264,13 +361,16 @@ void llama_server_embedding(const char *json_req, char **json_resp,
     } else {
       prompt = "";
     }
-    const int task_id = llama->request_completion(
-        {{"prompt", prompt}, {"n_predict", 0}}, false, true, -1);
-    task_result result = llama->next_result(task_id);
+    const int task_id = llama->queue_tasks.get_new_id();
+    llama->queue_results.add_waiting_task_id(task_id);
+    llama->request_completion(task_id, {{"prompt", prompt}, {"n_predict", 0}}, false, true, -1);
+    atomicRecv ar(recv_counter);
+    task_result result = llama->queue_results.recv(task_id);
     std::string result_json = result.result_json.dump();
     const std::string::size_type size = result_json.size() + 1;
     *json_resp = new char[size];
     snprintf(*json_resp, size, "%s", result_json.c_str());
+    llama->queue_results.remove_waiting_task_id(task_id);
   } catch (std::exception &e) {
     err->id = -1;
     snprintf(err->msg, err->msg_len, "exception %s", e.what());

llm/ext_server/ext_server.h

@@ -45,6 +45,7 @@ typedef struct ext_server_params {
   bool embedding;        // get only sentence embedding
   ext_server_lora_adapter_t *lora_adapters;
   char *mmproj;
+  bool verbose_logging;  // Enable verbose logging of the server
 } ext_server_params_t;
 
 typedef struct ext_server_task_result {

llm/ext_server_default.go

@@ -1,80 +0,0 @@
-//go:build !windows
-
-package llm
-
-/*
-#include <stdlib.h>
-#include "ext_server.h"
-
-*/
-import "C"
-import (
-	"context"
-
-	"github.com/jmorganca/ollama/api"
-)
-
-type llamaExtServer struct {
-	api.Options
-}
-
-func (llm *llamaExtServer) llama_server_init(sparams *C.ext_server_params_t, err *C.ext_server_resp_t) {
-	C.llama_server_init(sparams, err)
-}
-func (llm *llamaExtServer) llama_server_start() {
-	C.llama_server_start()
-}
-func (llm *llamaExtServer) llama_server_stop() {
-	C.llama_server_stop()
-}
-
-func (llm *llamaExtServer) llama_server_completion(json_req *C.char, resp *C.ext_server_resp_t) {
-	C.llama_server_completion(json_req, resp)
-}
-func (llm *llamaExtServer) llama_server_completion_next_result(task_id C.int, resp *C.ext_server_task_result_t) {
-	C.llama_server_completion_next_result(task_id, resp)
-}
-func (llm *llamaExtServer) llama_server_completion_cancel(task_id C.int, err *C.ext_server_resp_t) {
-	C.llama_server_completion_cancel(task_id, err)
-}
-func (llm *llamaExtServer) llama_server_release_task_result(result *C.ext_server_task_result_t) {
-	C.llama_server_release_task_result(result)
-}
-
-func (llm *llamaExtServer) llama_server_tokenize(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t) {
-	C.llama_server_tokenize(json_req, json_resp, err)
-}
-func (llm *llamaExtServer) llama_server_detokenize(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t) {
-	C.llama_server_detokenize(json_req, json_resp, err)
-}
-func (llm *llamaExtServer) llama_server_embedding(json_req *C.char, json_resp **C.char, err *C.ext_server_resp_t) {
-	C.llama_server_embedding(json_req, json_resp, err)
-}
-func (llm *llamaExtServer) llama_server_release_json_resp(json_resp **C.char) {
-	C.llama_server_release_json_resp(json_resp)
-}
-
-func newDefaultExtServer(model string, adapters, projectors []string, numLayers int64, opts api.Options) (extServer, error) {
-	server := &llamaExtServer{opts}
-	return newExtServer(server, model, adapters, projectors, numLayers, opts)
-}
-
-func (llm *llamaExtServer) Predict(ctx context.Context, pred PredictOpts, fn func(PredictResult)) error {
-	return predict(ctx, llm, pred, fn)
-}
-
-func (llm *llamaExtServer) Encode(ctx context.Context, prompt string) ([]int, error) {
-	return encode(llm, ctx, prompt)
-}
-
-func (llm *llamaExtServer) Decode(ctx context.Context, tokens []int) (string, error) {
-	return decode(llm, ctx, tokens)
-}
-
-func (llm *llamaExtServer) Embedding(ctx context.Context, input string) ([]float64, error) {
-	return embedding(llm, ctx, input)
-}
-
-func (llm *llamaExtServer) Close() {
-	close(llm)
-}

llm/ext_server_windows.go

@@ -1,12 +0,0 @@
-package llm
-
-import (
-	"github.com/jmorganca/ollama/api"
-)
-
-func newDefaultExtServer(model string, adapters, projectors []string, numLayers int64, opts api.Options) (extServer, error) {
-	// On windows we always load the llama.cpp libraries dynamically to avoid startup DLL dependencies
-	// This ensures we can update the PATH at runtime to get everything loaded
-
-	return newDynamicShimExtServer(AvailableShims["cpu"], model, adapters, projectors, numLayers, opts)
-}

llm/generate/gen_common.sh

@@ -1,15 +1,46 @@
 # common logic accross linux and darwin
 
 init_vars() {
+    case "${GOARCH}" in
+    "amd64")
+        ARCH="x86_64"
+        ;;
+    "arm64")
+        ARCH="arm64"
+        ;;
+    *)
+        ARCH=$(uname -m | sed -e "s/aarch64/arm64/g")
+    esac
+
     LLAMACPP_DIR=../llama.cpp
-    PATCHES="0001-Expose-callable-API-for-server.patch"
     CMAKE_DEFS=""
-    CMAKE_TARGETS="--target ggml --target ggml_static --target llama --target build_info --target common --target ext_server --target llava_static"
+    CMAKE_TARGETS="--target ext_server"
     if echo "${CGO_CFLAGS}" | grep -- '-g' >/dev/null; then
-        CMAKE_DEFS="-DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_VERBOSE_MAKEFILE=on -DLLAMA_GPROF=on -DLLAMA_SERVER_VERBOSE=on"
+        CMAKE_DEFS="-DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_VERBOSE_MAKEFILE=on -DLLAMA_GPROF=on -DLLAMA_SERVER_VERBOSE=on ${CMAKE_DEFS}"
     else
         # TODO - add additional optimization flags...
-        CMAKE_DEFS="-DCMAKE_BUILD_TYPE=Release -DLLAMA_SERVER_VERBOSE=off"
+        CMAKE_DEFS="-DCMAKE_BUILD_TYPE=Release -DLLAMA_SERVER_VERBOSE=off ${CMAKE_DEFS}"
+    fi
+    case $(uname -s) in 
+    "Darwin")
+        LIB_EXT="dylib"
+        WHOLE_ARCHIVE="-Wl,-force_load"
+        NO_WHOLE_ARCHIVE=""
+        GCC_ARCH="-arch ${ARCH}"
+        ;;
+    "Linux")
+        LIB_EXT="so"
+        WHOLE_ARCHIVE="-Wl,--whole-archive"
+        NO_WHOLE_ARCHIVE="-Wl,--no-whole-archive"
+
+        # Cross compiling not supported on linux - Use docker
+        GCC_ARCH=""
+        ;;
+    *)
+        ;;
+    esac
+    if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then 
+        CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
     fi
 }
 
@@ -33,6 +64,19 @@ apply_patches() {
     if ! grep ollama ${LLAMACPP_DIR}/examples/server/CMakeLists.txt; then
         echo 'include (../../../ext_server/CMakeLists.txt) # ollama' >>${LLAMACPP_DIR}/examples/server/CMakeLists.txt
     fi
+
+    if [ -n "$(ls -A ../patches/*.diff)" ]; then
+        # apply temporary patches until fix is upstream
+        for patch in ../patches/*.diff; do
+            for file in $(grep "^+++ " ${patch} | cut -f2 -d' ' | cut -f2- -d/); do
+                (cd ${LLAMACPP_DIR}; git checkout ${file})
+            done
+        done
+        for patch in ../patches/*.diff; do
+            (cd ${LLAMACPP_DIR} && git apply ${patch})
+        done
+    fi
+
     # Avoid duplicate main symbols when we link into the cgo binary
     sed -e 's/int main(/int __main(/g' <${LLAMACPP_DIR}/examples/server/server.cpp >${LLAMACPP_DIR}/examples/server/server.cpp.tmp &&
         mv ${LLAMACPP_DIR}/examples/server/server.cpp.tmp ${LLAMACPP_DIR}/examples/server/server.cpp
@@ -41,18 +85,41 @@ apply_patches() {
 build() {
     cmake -S ${LLAMACPP_DIR} -B ${BUILD_DIR} ${CMAKE_DEFS}
     cmake --build ${BUILD_DIR} ${CMAKE_TARGETS} -j8
+    mkdir -p ${BUILD_DIR}/lib/
+    g++ -fPIC -g -shared -o ${BUILD_DIR}/lib/libext_server.${LIB_EXT} \
+        ${GCC_ARCH} \
+        ${WHOLE_ARCHIVE} ${BUILD_DIR}/examples/server/libext_server.a ${NO_WHOLE_ARCHIVE} \
+        ${BUILD_DIR}/common/libcommon.a \
+        ${BUILD_DIR}/libllama.a \
+        -Wl,-rpath,\$ORIGIN \
+        -lpthread -ldl -lm \
+        ${EXTRA_LIBS}
 }
 
-install() {
-    rm -rf ${BUILD_DIR}/lib
-    mkdir -p ${BUILD_DIR}/lib
-    cp ${BUILD_DIR}/examples/server/libext_server.a ${BUILD_DIR}/lib
-    cp ${BUILD_DIR}/common/libcommon.a ${BUILD_DIR}/lib
-    cp ${BUILD_DIR}/libllama.a ${BUILD_DIR}/lib
-    cp ${BUILD_DIR}/libggml_static.a ${BUILD_DIR}/lib
+compress_libs() {
+    echo "Compressing payloads to reduce overall binary size..."
+    pids=""
+    rm -rf ${BUILD_DIR}/lib/*.${LIB_EXT}*.gz
+    for lib in ${BUILD_DIR}/lib/*.${LIB_EXT}* ; do
+        gzip --best -f ${lib} &
+        pids+=" $!"
+    done
+    echo 
+    for pid in ${pids}; do
+        wait $pid
+    done
+    echo "Finished compression"
 }
 
 # Keep the local tree clean after we're done with the build
 cleanup() {
     (cd ${LLAMACPP_DIR}/examples/server/ && git checkout CMakeLists.txt server.cpp)
+
+    if [ -n "$(ls -A ../patches/*.diff)" ]; then
+        for patch in ../patches/*.diff; do
+            for file in $(grep "^+++ " ${patch} | cut -f2 -d' ' | cut -f2- -d/); do
+                (cd ${LLAMACPP_DIR}; git checkout ${file})
+            done
+        done
+    fi
 }

llm/generate/gen_darwin.sh

@@ -9,14 +9,63 @@ set -o pipefail
 echo "Starting darwin generate script"
 source $(dirname $0)/gen_common.sh
 init_vars
-CMAKE_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.0 -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on ${CMAKE_DEFS}"
-BUILD_DIR="${LLAMACPP_DIR}/build/darwin/metal"
+git_module_setup
+apply_patches
+
+sign() {
+    if [ -n "$APPLE_IDENTITY" ]; then
+        codesign -f --timestamp --deep --options=runtime --sign "$APPLE_IDENTITY" --identifier ai.ollama.ollama $1
+    fi
+}
+
+COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.0 -DCMAKE_SYSTEM_NAME=Darwin"
+
 case "${GOARCH}" in
 "amd64")
-    CMAKE_DEFS="-DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_NATIVE=off"
+
+    #
+    # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
+    #
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu"
+    echo "Building LCD CPU"
+    build
+    sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu/lib/libext_server.dylib
+    compress_libs
+
+    #
+    # ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
+    # Approximately 400% faster than LCD on same CPU
+    #
+    init_vars
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx"
+    echo "Building AVX CPU"
+    build
+    sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx/lib/libext_server.dylib
+    compress_libs
+
+    #
+    # ~2013 CPU Dynamic library
+    # Approximately 10% faster than AVX on same CPU
+    #
+    init_vars
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+    BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx2"
+    echo "Building AVX2 CPU"
+    EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
+    build
+    sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx2/lib/libext_server.dylib
+    compress_libs
     ;;
 "arm64")
-    CMAKE_DEFS="-DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 ${CMAKE_DEFS}"
+    CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
+    BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/metal"
+    EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
+    build
+    sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/metal/lib/libext_server.dylib
+    compress_libs
     ;;
 *)
     echo "GOARCH must be set"
@@ -25,8 +74,4 @@ case "${GOARCH}" in
     ;;
 esac
 
-git_module_setup
-apply_patches
-build
-install
 cleanup

llm/generate/gen_linux.sh

@@ -2,24 +2,25 @@
 # This script is intended to run inside the go generate
 # working directory must be llm/generate/
 
-# First we build our default built-in library which will be linked into the CGO
-# binary as a normal dependency. This default build is CPU based.
+# First we build one or more CPU based LLM libraries
 #
-# Then we build a CUDA dynamic library (although statically linked with the CUDA
-# library dependencies for maximum portability)
+# Then if we detect CUDA, we build a CUDA dynamic library, and carry the required
+# library dependencies
 #
-# Then if we detect ROCm, we build a dynamically loaded ROCm lib.  ROCm is particularly
-# important to be a dynamic lib even if it's the only GPU library detected because
-# we can't redistribute the objectfiles but must rely on dynamic libraries at
-# runtime, which could lead the server not to start if not present.
+# Then if we detect ROCm, we build a dynamically loaded ROCm lib.  The ROCM
+# libraries are quite large, and also dynamically load data files at runtime
+# which in turn are large, so we don't attempt to cary them as payload
 
 set -ex
 set -o pipefail
 
 # See https://llvm.org/docs/AMDGPUUsage.html#processors for reference
 amdGPUs() {
+    if [ -n "${AMDGPU_TARGETS}" ]; then
+        echo "${AMDGPU_TARGETS}"
+        return
+    fi
     GPU_LIST=(
-        "gfx803"
         "gfx900"
         "gfx906:xnack-"
         "gfx908:xnack-"
@@ -39,8 +40,13 @@ amdGPUs() {
 }
 
 echo "Starting linux generate script"
-if [ -z "${CUDACXX}" -a -x /usr/local/cuda/bin/nvcc ]; then
-    export CUDACXX=/usr/local/cuda/bin/nvcc
+if [ -z "${CUDACXX}" ]; then
+    if [ -x /usr/local/cuda/bin/nvcc ]; then
+        export CUDACXX=/usr/local/cuda/bin/nvcc
+    else
+        # Try the default location in case it exists
+        export CUDACXX=$(command -v nvcc)
+    fi
 fi
 COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off"
 source $(dirname $0)/gen_common.sh
@@ -48,38 +54,115 @@ init_vars
 git_module_setup
 apply_patches
 
-#
-# CPU first for the default library
-#
-CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS}"
-BUILD_DIR="${LLAMACPP_DIR}/build/linux/cpu"
+if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
+    # Users building from source can tune the exact flags we pass to cmake for configuring
+    # llama.cpp, and we'll build only 1 CPU variant in that case as the default.
+    if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
+        echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
+        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
+        BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cpu"
+        echo "Building custom CPU"
+        build
+        compress_libs
+    else
+        # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
+        # -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+        # -DLLAMA_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
+        # -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+        # -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+        # Note: the following seem to yield slower results than AVX2 - ymmv
+        # -DLLAMA_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
+        # -DLLAMA_AVX512_VBMI -- 2018 Intel Cannon Lake
+        # -DLLAMA_AVX512_VNNI -- 2021 Intel Alder Lake
 
-build
-install
+        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off"
+        if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
+            #
+            # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
+            #
+            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+            BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cpu"
+            echo "Building LCD CPU"
+            build
+            compress_libs
+        fi
 
-# Placeholder to keep go embed happy until we start building dynamic CPU lib variants
-touch ${BUILD_DIR}/lib/dummy.so
+        if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx" ]; then
+            #
+            # ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
+            # Approximately 400% faster than LCD on same CPU
+            #
+            init_vars
+            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+            BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cpu_avx"
+            echo "Building AVX CPU"
+            build
+            compress_libs
+        fi
 
-if [ -d /usr/local/cuda/lib64/ ]; then
+        if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx2" ]; then
+            #
+            # ~2013 CPU Dynamic library
+            # Approximately 10% faster than AVX on same CPU
+            #
+            init_vars
+            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+            BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cpu_avx2"
+            echo "Building AVX2 CPU"
+            build
+            compress_libs
+        fi
+    fi
+else
+    echo "Skipping CPU generation step as requested"
+fi
+
+# If needed, look for the default CUDA toolkit location
+if [ -z "${CUDA_LIB_DIR}" ] && [ -d /usr/local/cuda/lib64 ]; then
+    CUDA_LIB_DIR=/usr/local/cuda/lib64
+fi
+
+# If needed, look for CUDA on Arch Linux
+if [ -z "${CUDA_LIB_DIR}" ] && [ -d /opt/cuda/targets/x86_64-linux/lib ]; then
+    CUDA_LIB_DIR=/opt/cuda/targets/x86_64-linux/lib
+fi
+
+# Allow override in case libcudart is in the wrong place
+if [ -z "${CUDART_LIB_DIR}" ]; then
+    CUDART_LIB_DIR="${CUDA_LIB_DIR}"
+fi
+
+if [ -d "${CUDA_LIB_DIR}" ]; then
     echo "CUDA libraries detected - building dynamic CUDA library"
     init_vars
-    CMAKE_DEFS="-DLLAMA_CUBLAS=on ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS}"
-    BUILD_DIR="${LLAMACPP_DIR}/build/linux/cuda"
-    CUDA_LIB_DIR=/usr/local/cuda/lib64
+    CUDA_MAJOR=$(ls "${CUDA_LIB_DIR}"/libcudart.so.* | head -1 | cut -f3 -d. || true)
+    if [ -n "${CUDA_MAJOR}" ]; then
+        CUDA_VARIANT=_v${CUDA_MAJOR}
+    fi
+    CMAKE_DEFS="-DLLAMA_CUBLAS=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS}"
+    BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cuda${CUDA_VARIANT}"
+    EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
     build
-    install
-    gcc -fPIC -g -shared -o ${BUILD_DIR}/lib/libext_server.so \
-        -Wl,--whole-archive \
-        ${BUILD_DIR}/lib/libext_server.a \
-        ${BUILD_DIR}/lib/libcommon.a \
-        ${BUILD_DIR}/lib/libllama.a \
-        -Wl,--no-whole-archive \
-        ${CUDA_LIB_DIR}/libcudart_static.a \
-        ${CUDA_LIB_DIR}/libcublas_static.a \
-        ${CUDA_LIB_DIR}/libcublasLt_static.a \
-        ${CUDA_LIB_DIR}/libcudadevrt.a \
-        ${CUDA_LIB_DIR}/libculibos.a \
-        -lrt -lpthread -ldl -lstdc++ -lm
+
+    # Cary the CUDA libs as payloads to help reduce dependency burden on users
+    #
+    # TODO - in the future we may shift to packaging these separately and conditionally
+    #        downloading them in the install script.
+    DEPS="$(ldd ${BUILD_DIR}/lib/libext_server.so )"
+    for lib in libcudart.so libcublas.so libcublasLt.so ; do
+        DEP=$(echo "${DEPS}" | grep ${lib} | cut -f1 -d' ' | xargs || true)
+        if [ -n "${DEP}" -a -e "${CUDA_LIB_DIR}/${DEP}" ]; then
+            cp "${CUDA_LIB_DIR}/${DEP}" "${BUILD_DIR}/lib/"
+        elif [ -e "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" ]; then
+            cp "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" "${BUILD_DIR}/lib/"
+        elif [ -e "${CUDART_LIB_DIR}/${lib}" ]; then
+            cp -d ${CUDART_LIB_DIR}/${lib}* "${BUILD_DIR}/lib/"
+        else
+            cp -d "${CUDA_LIB_DIR}/${lib}*" "${BUILD_DIR}/lib/"
+        fi
+    done
+    compress_libs
+
 fi
 
 if [ -z "${ROCM_PATH}" ]; then
@@ -96,21 +179,18 @@ fi
 
 if [ -d "${ROCM_PATH}" ]; then
     echo "ROCm libraries detected - building dynamic ROCm library"
+    if [ -f ${ROCM_PATH}/lib/librocm_smi64.so.? ]; then
+        ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocm_smi64.so.? | cut -f3 -d. || true)
+    fi
     init_vars
     CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
-    BUILD_DIR="${LLAMACPP_DIR}/build/linux/rocm"
+    BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/rocm${ROCM_VARIANT}"
+    EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,${ROCM_PATH}/lib,-rpath,/opt/amdgpu/lib/x86_64-linux-gnu/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
     build
-    install
-    gcc -fPIC -g -shared -o ${BUILD_DIR}/lib/libext_server.so \
-        -Wl,--whole-archive \
-        ${BUILD_DIR}/lib/libext_server.a \
-        ${BUILD_DIR}/lib/libcommon.a \
-        ${BUILD_DIR}/lib/libllama.a \
-        -Wl,--no-whole-archive \
-        -lrt -lpthread -ldl -lstdc++ -lm \
-        -L/opt/rocm/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ \
-        -Wl,-rpath,/opt/rocm/lib,-rpath,/opt/amdgpu/lib/x86_64-linux-gnu/ \
-        -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu
+
+    # Note: the ROCM libs and runtime library files are too large to embed, so we depend on
+    #       them being present at runtime on the host
+    compress_libs
 fi
 
 cleanup

llm/generate/gen_windows.ps1

@@ -3,10 +3,11 @@
 $ErrorActionPreference = "Stop"
 
 function init_vars {
+    $script:SRC_DIR = $(resolve-path "..\..\")
     $script:llamacppDir = "../llama.cpp"
-    $script:patches = @("0001-Expose-callable-API-for-server.patch")
-    $script:cmakeDefs = @("-DBUILD_SHARED_LIBS=on", "-DLLAMA_NATIVE=off", "-DLLAMA_F16C=off", "-DLLAMA_FMA=off", "-DLLAMA_AVX512=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX=on", "-A","x64")
-    $script:cmakeTargets = @("ggml", "ggml_static", "llama", "build_info", "common", "ext_server_shared", "llava_static")
+    $script:cmakeDefs = @("-DBUILD_SHARED_LIBS=on", "-DLLAMA_NATIVE=off",  "-A", "x64")
+    $script:cmakeTargets = @("ext_server")
+    $script:ARCH = "amd64" # arm not yet supported.
     if ($env:CGO_CFLAGS -contains "-g") {
         $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on")
         $script:config = "RelWithDebInfo"
@@ -14,6 +15,28 @@ function init_vars {
         $script:cmakeDefs += @("-DLLAMA_SERVER_VERBOSE=off")
         $script:config = "Release"
     }
+    # Try to find the CUDA dir
+    if ($env:CUDA_LIB_DIR -eq $null) {
+        $d=(get-command -ea 'silentlycontinue' nvcc).path
+        if ($d -ne $null) {
+            $script:CUDA_LIB_DIR=($d| split-path -parent)
+            $script:CUDA_INCLUDE_DIR=($script:CUDA_LIB_DIR|split-path -parent)+"\include"
+        }
+    } else {
+        $script:CUDA_LIB_DIR=$env:CUDA_LIB_DIR
+    }
+    $script:GZIP=(get-command -ea 'silentlycontinue' gzip).path
+    $script:DUMPBIN=(get-command -ea 'silentlycontinue' dumpbin).path
+    if ($null -eq $env:CMAKE_CUDA_ARCHITECTURES) {
+        $script:CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
+    } else {
+        $script:CMAKE_CUDA_ARCHITECTURES=$env:CMAKE_CUDA_ARCHITECTURES
+    }
+    # Note: 10 Windows Kit signtool crashes with GCP's plugin
+    ${script:SignTool}="C:\Program Files (x86)\Windows Kits\8.1\bin\x64\signtool.exe"
+    if ("${env:KEY_CONTAINER}") {
+        ${script:OLLAMA_CERT}=$(resolve-path "${script:SRC_DIR}\ollama_inc.crt")
+    }
 }
 
 function git_module_setup {
@@ -29,6 +52,29 @@ function apply_patches {
     if (!(Select-String -Path "${script:llamacppDir}/examples/server/CMakeLists.txt" -Pattern 'ollama')) {
         Add-Content -Path "${script:llamacppDir}/examples/server/CMakeLists.txt" -Value 'include (../../../ext_server/CMakeLists.txt) # ollama'
     }
+
+    # Apply temporary patches until fix is upstream
+    $patches = Get-ChildItem "../patches/*.diff"
+    foreach ($patch in $patches) {
+        # Extract file paths from the patch file
+        $filePaths = Get-Content $patch.FullName | Where-Object { $_ -match '^\+\+\+ ' } | ForEach-Object {
+            $parts = $_ -split ' '
+            ($parts[1] -split '/', 2)[1]
+        }
+
+        # Checkout each file
+        Set-Location -Path ${script:llamacppDir}
+        foreach ($file in $filePaths) {
+            git checkout $file
+        }
+    }
+
+    # Apply each patch
+    foreach ($patch in $patches) {
+        Set-Location -Path ${script:llamacppDir}
+        git apply $patch.FullName
+    }
+
     # Avoid duplicate main symbols when we link into the cgo binary
     $content = Get-Content -Path "${script:llamacppDir}/examples/server/server.cpp"
     $content = $content -replace 'int main\(', 'int __main('
@@ -48,41 +94,113 @@ function build {
 function install {
     rm -ea 0 -recurse -force -path "${script:buildDir}/lib"
     md "${script:buildDir}/lib" -ea 0 > $null
-    cp "${script:buildDir}/bin/${script:config}/ext_server_shared.dll" "${script:buildDir}/lib"
+    cp "${script:buildDir}/bin/${script:config}/ext_server.dll" "${script:buildDir}/lib"
     cp "${script:buildDir}/bin/${script:config}/llama.dll" "${script:buildDir}/lib"
-
     # Display the dll dependencies in the build log
-    dumpbin /dependents "${script:buildDir}/bin/${script:config}/ext_server_shared.dll" | select-string ".dll"
+    if ($script:DUMPBIN -ne $null) {
+        & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/${script:config}/ext_server.dll" | select-string ".dll"
+    }
+}
+
+function sign {
+    if ("${env:KEY_CONTAINER}") {
+        write-host "Signing ${script:buildDir}/lib/*.dll"
+        foreach ($file in (get-childitem "${script:buildDir}/lib/*.dll")){
+            & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
+                /csp "Google Cloud KMS Provider" /kc "${env:KEY_CONTAINER}" $file
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+        }
+    }
+}
+
+function compress_libs {
+    if ($script:GZIP -eq $null) {
+        write-host "gzip not installed, not compressing files"
+        return
+    }
+    write-host "Compressing dlls..."
+    $libs = dir "${script:buildDir}/lib/*.dll"
+    foreach ($file in $libs) {
+        & "$script:GZIP" --best -f $file
+    }
 }
 
 function cleanup {
+    $patches = Get-ChildItem "../patches/*.diff"
+    foreach ($patch in $patches) {
+        # Extract file paths from the patch file
+        $filePaths = Get-Content $patch.FullName | Where-Object { $_ -match '^\+\+\+ ' } | ForEach-Object {
+            $parts = $_ -split ' '
+            ($parts[1] -split '/', 2)[1]
+        }
+
+        # Checkout each file
+        Set-Location -Path ${script:llamacppDir}
+        foreach ($file in $filePaths) {            
+            git checkout $file
+        }
+    }
     Set-Location "${script:llamacppDir}/examples/server"
     git checkout CMakeLists.txt server.cpp
+
 }
 
 init_vars
 git_module_setup
 apply_patches
 
-# first build CPU based
-$script:buildDir="${script:llamacppDir}/build/windows/cpu"
+# -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+# -DLLAMA_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
+# -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+# -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
 
+$script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on", "-DLLAMA_NATIVE=off")
+
+$script:cmakeDefs = $script:commonCpuDefs + @("-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+$script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu"
+write-host "Building LCD CPU"
 build
 install
+sign
+compress_libs
 
-# Then build cuda as a dynamically loaded library
-init_vars
-$script:buildDir="${script:llamacppDir}/build/windows/cuda"
-$script:cmakeDefs += @("-DLLAMA_CUBLAS=ON")
+$script:cmakeDefs = $script:commonCpuDefs + @("-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+$script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu_avx"
+write-host "Building AVX CPU"
 build
 install
+sign
+compress_libs
 
+$script:cmakeDefs = $script:commonCpuDefs + @("-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
+$script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu_avx2"
+write-host "Building AVX2 CPU"
+build
+install
+sign
+compress_libs
+
+if ($null -ne $script:CUDA_LIB_DIR) {
+    # Then build cuda as a dynamically loaded library
+    $nvcc = "$script:CUDA_LIB_DIR\nvcc.exe"
+    $script:CUDA_VERSION=(get-item ($nvcc | split-path | split-path)).Basename
+    if ($null -ne $script:CUDA_VERSION) {
+        $script:CUDA_VARIANT="_"+$script:CUDA_VERSION
+    }
+    init_vars
+    $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
+    $script:cmakeDefs += @("-DLLAMA_CUBLAS=ON", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR", "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}")
+    build
+    install
+    sign
+    compress_libs
+}
 # TODO - actually implement ROCm support on windows
-$script:buildDir="${script:llamacppDir}/build/windows/rocm"
+$script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/rocm"
 
 rm -ea 0 -recurse -force -path "${script:buildDir}/lib"
 md "${script:buildDir}/lib" -ea 0 > $null
 echo $null >> "${script:buildDir}/lib/.generated"
 
 cleanup
-write-host "`ngo generate completed"
+write-host "`ngo generate completed"

llm/ggml.go

@@ -78,7 +78,12 @@ type model interface {
 	ModelFamily() string
 	ModelType() string
 	FileType() string
-	NumLayers() int64
+	NumLayers() uint32
+	NumGQA() uint32
+	NumEmbed() uint32
+	NumHead() uint32
+	NumHeadKv() uint32
+	NumCtx() uint32
 }
 
 type container interface {
@@ -94,9 +99,9 @@ func (c *containerLORA) Name() string {
 	return "ggla"
 }
 
-func (c *containerLORA) Decode(ro *readSeekOffset) (model, error) {
+func (c *containerLORA) Decode(rso *readSeekOffset) (model, error) {
 	var version uint32
-	binary.Read(ro, binary.LittleEndian, &version)
+	binary.Read(rso, binary.LittleEndian, &version)
 
 	switch version {
 	case 1:
@@ -107,7 +112,7 @@ func (c *containerLORA) Decode(ro *readSeekOffset) (model, error) {
 	c.version = version
 
 	// remaining file contents aren't decoded
-	ro.Seek(0, io.SeekEnd)
+	rso.Seek(0, io.SeekEnd)
 
 	return nil, nil
 }

llm/gguf.go

@@ -69,12 +69,65 @@ type tensor struct {
 	name   string
 	kind   uint32
 	offset uint64
-	size   uint64
 
 	// shape is the number of elements in each dimension
 	shape [4]uint64
 }
 
+func (t tensor) blockSize() uint64 {
+	switch {
+	case t.kind < 2:
+		return 1
+	case t.kind < 10:
+		return 32
+	default:
+		return 256
+	}
+}
+
+func (t tensor) typeSize() uint64 {
+	blockSize := t.blockSize()
+
+	switch t.kind {
+	case 0: // FP32
+		return 4
+	case 1: // FP16
+		return 2
+	case 2: // Q4_0
+		return 2 + blockSize/2
+	case 3: // Q4_1
+		return 2 + 2 + blockSize/2
+	case 6: // Q5_0
+		return 2 + 4 + blockSize/2
+	case 7: // Q5_1
+		return 2 + 2 + 4 + blockSize/2
+	case 8: // Q8_0
+		return 2 + blockSize
+	case 9: // Q8_1
+		return 4 + 4 + blockSize
+	case 10: // Q2_K
+		return blockSize/16 + blockSize/4 + 2 + 2
+	case 11: // Q3_K
+		return blockSize/8 + blockSize/4 + 12 + 2
+	case 12: // Q4_K
+		return 2 + 2 + 12 + blockSize/2
+	case 13: // Q5_K
+		return 2 + 2 + 12 + blockSize/8 + blockSize/2
+	case 14: // Q6_K
+		return blockSize/2 + blockSize/4 + blockSize/16 + 2
+	default:
+		return 0
+	}
+}
+
+func (t tensor) parameters() uint64 {
+	return t.shape[0] * t.shape[1] * t.shape[2] * t.shape[3]
+}
+
+func (t tensor) size() uint64 {
+	return t.parameters() * t.typeSize() / t.blockSize()
+}
+
 type ggufModel struct {
 	*containerGGUF
 
@@ -201,61 +254,15 @@ func (llm *ggufModel) Decode(rso *readSeekOffset) error {
 			shape[i] = llm.readU64(rso)
 		}
 
-		kind := llm.readU32(rso)
-		offset := llm.readU64(rso)
-
-		var blockSize uint64
-		switch {
-		case kind < 2:
-			blockSize = 1
-		case kind < 10:
-			blockSize = 32
-		default:
-			blockSize = 256
-		}
-
-		var typeSize uint64
-		switch kind {
-		case 0: // FP32
-			typeSize = 4
-		case 1: // FP16
-			typeSize = 2
-		case 2: // Q4_0
-			typeSize = 2 + blockSize/2
-		case 3: // Q4_1
-			typeSize = 2 + 2 + blockSize/2
-		case 6: // Q5_0
-			typeSize = 2 + 4 + blockSize/2
-		case 7: // Q5_1
-			typeSize = 2 + 2 + 4 + blockSize/2
-		case 8: // Q8_0
-			typeSize = 2 + blockSize
-		case 9: // Q8_1
-			typeSize = 4 + 4 + blockSize
-		case 10: // Q2_K
-			typeSize = blockSize/16 + blockSize/4 + 2 + 2
-		case 11: // Q3_K
-			typeSize = blockSize/8 + blockSize/4 + 12 + 2
-		case 12: // Q4_K
-			typeSize = 2 + 2 + 12 + blockSize/2
-		case 13: // Q5_K
-			typeSize = 2 + 2 + 12 + blockSize/8 + blockSize/2
-		case 14: // Q6_K
-			typeSize = blockSize/2 + blockSize/4 + blockSize/16 + 2
-		}
-
-		parameters := shape[0] * shape[1] * shape[2] * shape[3]
-		size := parameters * typeSize / blockSize
-
-		llm.tensors = append(llm.tensors, tensor{
+		tensor := tensor{
 			name:   name,
-			kind:   kind,
-			offset: offset,
-			size:   size,
+			kind:   llm.readU32(rso),
+			offset: llm.readU64(rso),
 			shape:  shape,
-		})
+		}
 
-		llm.parameters += parameters
+		llm.tensors = append(llm.tensors, tensor)
+		llm.parameters += tensor.parameters()
 	}
 
 	alignment, ok := llm.kv["general.alignment"].(uint32)
@@ -265,21 +272,65 @@ func (llm *ggufModel) Decode(rso *readSeekOffset) error {
 
 	rso.Seek(int64(alignment)-rso.offset%int64(alignment), io.SeekCurrent)
 	for _, tensor := range llm.tensors {
-		padded := (int64(tensor.size) + int64(alignment) - 1) & ^(int64(alignment) - 1)
+		padded := (int64(tensor.size()) + int64(alignment) - 1) & ^(int64(alignment) - 1)
 		rso.Seek(padded, io.SeekCurrent)
 	}
 
 	return nil
 }
 
-func (llm *ggufModel) NumLayers() int64 {
+func (llm *ggufModel) NumLayers() uint32 {
 	value, exists := llm.kv[fmt.Sprintf("%s.block_count", llm.ModelFamily())]
 	if !exists {
 		return 0
 	}
 
-	v := value.(uint32)
-	return int64(v)
+	return value.(uint32)
+}
+
+func (llm *ggufModel) NumHead() uint32 {
+	value, exists := llm.kv[fmt.Sprintf("%s.attention.head_count", llm.ModelFamily())]
+	if !exists {
+		return 0
+	}
+
+	return value.(uint32)
+}
+
+func (llm *ggufModel) NumEmbed() uint32 {
+	value, exists := llm.kv[fmt.Sprintf("%s.embedding_length", llm.ModelFamily())]
+	if !exists {
+		return 0
+	}
+
+	return value.(uint32)
+}
+
+func (llm *ggufModel) NumHeadKv() uint32 {
+	value, exists := llm.kv[fmt.Sprintf("%s.attention.head_count_kv", llm.ModelFamily())]
+	if !exists {
+		return 0
+	}
+
+	return value.(uint32)
+}
+
+func (llm *ggufModel) NumCtx() uint32 {
+	value, exists := llm.kv[fmt.Sprintf("%s.context_length", llm.ModelFamily())]
+	if !exists {
+		return 0
+	}
+
+	return value.(uint32)
+}
+
+func (llm *ggufModel) NumGQA() uint32 {
+	numHeadKv := llm.NumHeadKv()
+	if numHeadKv == 0 {
+		return 0
+	}
+
+	return llm.NumHead() / numHeadKv
 }
 
 func (llm ggufModel) readU8(r io.Reader) uint8 {