fix ref

The numa flag may be having a performance impact on multi-socket systems with GPU loads

.gitattributes

@@ -1 +1,3 @@
 llm/ext_server/* linguist-vendored
+* text=auto
+*.go text eol=lf

.github/workflows/release.yaml

@@ -31,7 +31,7 @@ jobs:
           security set-keychain-settings -lut 3600 build.keychain
       - uses: actions/setup-go@v5
         with:
-          go-version: "stable"
+          go-version-file: go.mod
           cache: true
       - name: Build Darwin
         env:
@@ -87,7 +87,7 @@ jobs:
           write-host "plugin installed"
       - uses: actions/setup-go@v5
         with:
-          go-version: "stable"
+          go-version-file: go.mod
           cache: true
       - run: go get ./...
       - run: |
@@ -141,7 +141,7 @@ jobs:
           write-host "plugin installed"
       - uses: actions/setup-go@v5
         with:
-          go-version: "stable"
+          go-version-file: go.mod
           cache: true
       - name: 'Install ROCm'
         run: |
@@ -187,6 +187,13 @@ jobs:
   generate-windows-cuda:
     environment: release
     runs-on: windows
+    strategy:
+      matrix:
+        cuda:
+          - version: "11"
+            url: 'https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe'
+          - version: "12"
+            url: 'https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_551.61_windows.exe'
     env:
       KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
     steps:
@@ -218,13 +225,13 @@ jobs:
           write-host "plugin installed"
       - uses: actions/setup-go@v5
         with:
-          go-version: "stable"
+          go-version-file: go.mod
           cache: true
-      - name: 'Install CUDA'
+      - name: 'Install CUDA ${{ matrix.cuda.version }}'
         run: |
           $ErrorActionPreference = "Stop"
           write-host "downloading CUDA Installer"
-          Invoke-WebRequest -Uri "https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
+          Invoke-WebRequest -Uri "${{ matrix.cuda.url }}" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
           write-host "Installing CUDA"
           Start-Process "${env:RUNNER_TEMP}\cuda-install.exe" -ArgumentList '-s' -NoNewWindow -Wait
           write-host "Completed CUDA"
@@ -256,15 +263,16 @@ jobs:
           cp "${NVIDIA_DIR}\cublasLt64_*.dll" "dist\deps\"
       - uses: actions/upload-artifact@v4
         with:
-          name: generate-windows-cuda
+          name: generate-windows-cuda-${{ matrix.cuda.version }}
           path: |
             llm/build/**/bin/*
             dist/windows-amd64/**
       - uses: actions/upload-artifact@v4
         with:
-          name: windows-cuda-deps
+          name: windows-cuda-deps-${{ matrix.cuda.version }}
           path: dist/deps/*
 
+
   # Import the prior generation steps and build the final windows assets
   build-windows:
     environment: release
@@ -306,7 +314,7 @@ jobs:
           write-host "plugin installed"
       - uses: actions/setup-go@v5
         with:
-          go-version: "stable"
+          go-version-file: go.mod
           cache: true
       - run: go get
       - uses: actions/download-artifact@v4
@@ -314,10 +322,16 @@ jobs:
           name: generate-windows-cpu
       - uses: actions/download-artifact@v4
         with:
-          name: generate-windows-cuda
+          name: generate-windows-cuda-11
       - uses: actions/download-artifact@v4
         with:
-          name: windows-cuda-deps
+          name: generate-windows-cuda-12
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-cuda-deps-11
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-cuda-deps-12
       - uses: actions/download-artifact@v4
         with:
           name: windows-rocm-deps
@@ -363,7 +377,6 @@ jobs:
       - run: |
           ./scripts/build_linux.sh
           ./scripts/build_docker.sh
-          mv dist/deps/* dist/
       - uses: actions/upload-artifact@v4
         with:
           name: dist-linux-amd64
@@ -459,7 +472,10 @@ jobs:
           merge-multiple: true
       - run: |
           ls -lh dist/
-          (cd dist; sha256sum * > sha256sum.txt)
+          (cd dist; find . -type f | xargs sha256sum > ../sha256sum.txt)
+          mv sha256sum.txt dist/
+          mv dist/linux-???64 .
+          mv dist/linux-amd64-rocm .
           cat dist/sha256sum.txt
       - name: Create or update Release
         run: |

.github/workflows/test.yaml

@@ -63,7 +63,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
         with:
-          go-version: "stable"
+          go-version-file: go.mod
           cache: true
       - run: go get ./...
       - run: |
@@ -163,7 +163,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
         with:
-          go-version: "stable"
+          go-version-file: go.mod
           cache: true
       - name: 'Install ROCm'
         run: |
@@ -200,7 +200,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
         with:
-          go-version: "stable"
+          go-version-file: go.mod
           cache: true
       - name: 'Install CUDA'
         run: |
@@ -255,7 +255,7 @@ jobs:
           submodules: recursive
       - uses: actions/setup-go@v5
         with:
-          go-version: "stable"
+          go-version-file: go.mod
           cache: false
       - run: |
           case ${{ matrix.arch }} in
@@ -273,7 +273,7 @@ jobs:
         if: ${{ startsWith(matrix.os, 'macos-') }}
       - uses: golangci/golangci-lint-action@v6
         with:
-          args: --timeout 8m0s -v ${{ startsWith(matrix.os, 'windows-') && '' || '--disable gofmt --disable goimports' }}
+          args: --timeout 8m0s -v
   test:
     strategy:
       matrix:
@@ -297,7 +297,7 @@ jobs:
           submodules: recursive
       - uses: actions/setup-go@v5
         with:
-          go-version: "stable"
+          go-version-file: go.mod
           cache: true
       - run: |
           case ${{ matrix.arch }} in

.golangci.yaml

@@ -7,22 +7,31 @@ linters:
     - bodyclose
     - containedctx
     - contextcheck
+    - errcheck
     - exportloopref
+    - gci
     - gocheckcompilerdirectives
-    # conditionally enable this on linux/macos
-    # - gofmt
-    # - goimports
+    - gofmt
+    - gofumpt
+    - gosimple
+    - govet
+    - ineffassign
     - intrange
+    - makezero
     - misspell
     - nilerr
     - nolintlint
     - nosprintfhostport
-    - testifylint
+    - staticcheck
+    - tenv
     - unconvert
     - unused
+    - usestdlibvars
     - wastedassign
     - whitespace
-    - usestdlibvars
+linters-settings:
+  gci:
+    sections: [standard, default, localmodule]
 severity:
   default-severity: error
   rules:

CONTRIBUTING.md

@@ -0,0 +1,37 @@
+# Contributing to Ollama
+
+Thank you for your interest in contributing to Ollama! Here are a few guidelines to help get you started.
+
+## Set up
+
+See the [development documentation](./docs/development.md) for instructions on how to build and run Ollama locally.
+
+## Pull requests
+
+### Ideal issues
+
+* [Bugs](https://github.com/ollama/ollama/issues?q=is%3Aissue+is%3Aopen+label%3Abug): issues where Ollama stops working or where it results in an unexpected error.
+* [Performance](https://github.com/ollama/ollama/issues?q=is%3Aissue+is%3Aopen+label%3Aperformance): issues to make Ollama faster at model inference, downloading or uploading.
+* [Security](https://github.com/ollama/ollama/blob/main/SECURITY.md): issues that could lead to a security vulnerability. As mentioned in [SECURITY.md](https://github.com/ollama/ollama/blob/main/SECURITY.md), please do not disclose security vulnerabilities publicly.
+
+### Issues that are harder to review
+
+* New features: new features (e.g. API fields, environment variables) add surface area to Ollama and make it harder to maintain in the long run as they cannot be removed without potentially breaking users in the future.
+* Refactoring: large code improvements are important, but can be harder or take longer to review and merge.
+* Documentation: small updates to fill in or dorrect missing documentation is helpful, however large documentation additions can be hard to maintain over time.
+
+### Issues that may not be accepted
+
+* Changes that break backwards compatibility in Ollama's API (including the OpenAI-compatible API)
+* Changes that add significant friction to the user experience
+* Changes that create a large future maintenance burden for maintainers and contributors
+
+### Best practices
+
+* Commit messages: please leave both a title and a description in your commit messages. The title should be a short summary of the changes, with a leading word that explains the section of the code being changed (e.g. `api: fix parsing of prompt field`) . In the description, leave a short 2-3 sentences that explain more about the change and its impact.
+* Tests: please add test coverage to changes where possible.
+* Minimize dependencies: avoid adding new dependencies unless absolutely necessary.
+
+## Need help?
+
+If you need help with anything, feel free to reach out to us on our [Discord server](https://discord.gg/ollama).

Dockerfile

@@ -1,7 +1,9 @@
 ARG GOLANG_VERSION=1.22.5
 ARG CMAKE_VERSION=3.22.1
-# this CUDA_VERSION corresponds with the one specified in docs/gpu.md
-ARG CUDA_VERSION=11.3.1
+ARG CUDA_VERSION_11=11.3.1
+ARG CUDA_V11_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
+ARG CUDA_VERSION_12=12.4.0
+ARG CUDA_V12_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
 ARG ROCM_VERSION=6.1.2
 
 # Copy the minimal context we need to run the generate scripts
@@ -10,7 +12,7 @@ COPY .git .git
 COPY .gitmodules .gitmodules
 COPY llm llm
 
-FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION-devel-centos7 AS cuda-build-amd64
+FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_11-devel-centos7 AS cuda-11-build-amd64
 ARG CMAKE_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
@@ -18,9 +20,34 @@ ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+ARG CUDA_V11_ARCHITECTURES
+ENV GOARCH amd64 
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
+    CUDA_VARIANT="_v11" \
+    bash gen_linux.sh
 
-FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION-devel-rockylinux8 AS cuda-build-arm64
+FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_12-devel-centos7 AS cuda-12-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/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+ARG CGO_CFLAGS
+ARG CUDA_V12_ARCHITECTURES
+ENV GOARCH amd64 
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
+    CUDA_VARIANT="_v12" \
+    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
+    bash gen_linux.sh
+
+FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_11-devel-rockylinux8 AS cuda-11-build-server-arm64
 ARG CMAKE_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
@@ -28,7 +55,32 @@ ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+ARG CUDA_V11_ARCHITECTURES
+ENV GOARCH arm64 
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
+    CUDA_VARIANT="_v11" \
+    bash gen_linux.sh
+
+FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_12-devel-rockylinux8 AS cuda-12-build-server-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/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+ARG CGO_CFLAGS
+ARG CUDA_V12_ARCHITECTURES
+ENV GOARCH arm64 
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
+    CUDA_VARIANT="_v12" \
+    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
+    bash gen_linux.sh
+
 
 FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS rocm-build-amd64
 ARG CMAKE_VERSION
@@ -40,15 +92,11 @@ COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
 ARG AMDGPU_TARGETS
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
-RUN mkdir /tmp/scratch && \
-    for dep in $(zcat /go/src/github.com/ollama/ollama/llm/build/linux/x86_64/rocm*/bin/deps.txt.gz) ; do \
-        cp ${dep} /tmp/scratch/ || exit 1 ; \
-    done && \
-    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd /tmp/scratch/ && tar xf - ) && \
-    mkdir -p /go/src/github.com/ollama/ollama/dist/deps/ && \
-    (cd /tmp/scratch/ && tar czvf /go/src/github.com/ollama/ollama/dist/deps/ollama-linux-amd64-rocm.tgz . )
-
+ENV GOARCH amd64 
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 bash gen_linux.sh
+RUN mkdir -p ../../dist/linux-amd64-rocm/lib/ollama && \
+    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd ../../dist/linux-amd64-rocm/lib/ollama && tar xf - )
 
 FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
 ARG CMAKE_VERSION
@@ -59,16 +107,21 @@ ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 ARG OLLAMA_CUSTOM_CPU_DEFS
 ARG CGO_CFLAGS
+ENV GOARCH amd64 
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 
 FROM --platform=linux/amd64 cpu-builder-amd64 AS static-build-amd64
-RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx-build-amd64
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" bash gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" bash gen_linux.sh
 
 FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64
 ARG CMAKE_VERSION
@@ -79,12 +132,15 @@ ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 ARG OLLAMA_CUSTOM_CPU_DEFS
 ARG CGO_CFLAGS
+ENV GOARCH arm64
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 
 FROM --platform=linux/arm64 cpu-builder-arm64 AS static-build-arm64
-RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
 FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
 
 
 # Intermediate stage used for ./scripts/build_linux.sh
@@ -95,12 +151,16 @@ COPY . .
 COPY --from=static-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
 COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
 COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
 COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/deps/ ./dist/deps/
 ARG GOFLAGS
 ARG CGO_CFLAGS
-RUN go build -trimpath .
+RUN --mount=type=cache,target=/root/.ccache \
+    go build -trimpath -o dist/linux-amd64/bin/ollama .
 
 # Intermediate stage used for ./scripts/build_linux.sh
 FROM --platform=linux/arm64 cpu-build-arm64 AS build-arm64
@@ -109,23 +169,36 @@ ARG GOLANG_VERSION
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
 COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cuda-11-build-server-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-11-build-server-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cuda-12-build-server-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-12-build-server-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
 ARG GOFLAGS
 ARG CGO_CFLAGS
-RUN go build -trimpath .
+RUN --mount=type=cache,target=/root/.ccache \
+    go build -trimpath -o dist/linux-arm64/bin/ollama .
+
+# Strip out ROCm dependencies to keep the primary image lean
+FROM --platform=linux/amd64 ubuntu:22.04 as amd64-libs-without-rocm
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /scratch/
+RUN cd /scratch/ollama/ && rm -rf rocblas libamd* libdrm* libroc* libhip* libhsa* 
 
 # Runtime stages
 FROM --platform=linux/amd64 ubuntu:22.04 as runtime-amd64
+COPY --from=amd64-libs-without-rocm /scratch/ /lib/
 RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+
 FROM --platform=linux/arm64 ubuntu:22.04 as runtime-arm64
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
 RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-arm64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
 
 # Radeon images are much larger so we keep it distinct from the CPU/CUDA image
 FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete as runtime-rocm
 RUN update-pciids
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+RUN ln -s /opt/rocm/lib /lib/ollama
 EXPOSE 11434
 ENV OLLAMA_HOST 0.0.0.0
 

README.md

@@ -54,6 +54,7 @@ Here are some example models that can be downloaded:
 | Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`     |
 | Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`              |
 | Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`       |
+| Gemma 2            | 2B         | 1.6GB | `ollama run gemma2:2b`         |
 | Gemma 2            | 9B         | 5.5GB | `ollama run gemma2`            |
 | Gemma 2            | 27B        | 16GB  | `ollama run gemma2:27b`        |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
@@ -300,6 +301,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Sidellama](https://github.com/gyopak/sidellama) (browser-based LLM client)
 - [LLMStack](https://github.com/trypromptly/LLMStack) (No-code multi-agent framework to build LLM agents and workflows)
 - [BoltAI for Mac](https://boltai.com) (AI Chat Client for Mac)
+- [Harbor](https://github.com/av/harbor) (Containerized LLM Toolkit with Ollama as default backend)
 
 ### Terminal
 
@@ -323,6 +325,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [tlm](https://github.com/yusufcanb/tlm)
 - [podman-ollama](https://github.com/ericcurtin/podman-ollama)
 - [gollama](https://github.com/sammcj/gollama)
+- [Ollama eBook Summary](https://github.com/cognitivetech/ollama-ebook-summary/)
 
 ### Database
 

api/client.go

@@ -18,6 +18,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"net/http"
@@ -172,7 +173,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		}
 
 		if errorResponse.Error != "" {
-			return fmt.Errorf(errorResponse.Error)
+			return errors.New(errorResponse.Error)
 		}
 
 		if response.StatusCode >= http.StatusBadRequest {
@@ -297,7 +298,7 @@ func (c *Client) List(ctx context.Context) (*ListResponse, error) {
 	return &lr, nil
 }
 
-// List running models.
+// ListRunning lists running models.
 func (c *Client) ListRunning(ctx context.Context) (*ProcessResponse, error) {
 	var lr ProcessResponse
 	if err := c.do(ctx, http.MethodGet, "/api/ps", nil, &lr); err != nil {
@@ -332,7 +333,7 @@ func (c *Client) Show(ctx context.Context, req *ShowRequest) (*ShowResponse, err
 	return &resp, nil
 }
 
-// Hearbeat checks if the server has started and is responsive; if yes, it
+// Heartbeat checks if the server has started and is responsive; if yes, it
 // returns nil, otherwise an error.
 func (c *Client) Heartbeat(ctx context.Context) error {
 	if err := c.do(ctx, http.MethodHead, "/", nil, nil); err != nil {

api/types.go

@@ -231,7 +231,6 @@ type Options struct {
 
 // Runner options which must be set when the model is loaded into memory
 type Runner struct {
-	UseNUMA   bool  `json:"numa,omitempty"`
 	NumCtx    int   `json:"num_ctx,omitempty"`
 	NumBatch  int   `json:"num_batch,omitempty"`
 	NumGPU    int   `json:"num_gpu,omitempty"`
@@ -505,7 +504,7 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 	for key, val := range m {
 		opt, ok := jsonOpts[key]
 		if !ok {
-			slog.Warn("invalid option provided", "option", opt.Name)
+			slog.Warn("invalid option provided", "option", key)
 			continue
 		}
 
@@ -615,7 +614,6 @@ func DefaultOptions() Options {
 			F16KV:     true,
 			UseMLock:  false,
 			UseMMap:   nil,
-			UseNUMA:   false,
 		},
 	}
 }

api/types_test.go

@@ -2,7 +2,7 @@ package api
 
 import (
 	"encoding/json"
-	"fmt"
+	"errors"
 	"math"
 	"testing"
 	"time"
@@ -192,7 +192,7 @@ func TestUseMmapFormatParams(t *testing.T) {
 				"use_mmap": {"foo"},
 			},
 			exp: nil,
-			err: fmt.Errorf("invalid bool value [foo]"),
+			err: errors.New("invalid bool value [foo]"),
 		},
 	}
 

app/lifecycle/getstarted_nonwindows.go

@@ -2,8 +2,8 @@
 
 package lifecycle
 
-import "fmt"
+import "errors"
 
 func GetStarted() error {
-	return fmt.Errorf("GetStarted not implemented")
+	return errors.New("not implemented")
 }

app/lifecycle/getstarted_windows.go

@@ -34,7 +34,6 @@ func GetStarted() error {
 		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)
 	}

app/lifecycle/logging.go

@@ -27,7 +27,7 @@ func InitLogging() {
 		// TODO - write one-line to the app.log file saying we're running in console mode to help avoid confusion
 	} else {
 		rotateLogs(AppLogFile)
-		logFile, err = os.OpenFile(AppLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
+		logFile, err = os.OpenFile(AppLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0o755)
 		if err != nil {
 			slog.Error(fmt.Sprintf("failed to create server log %v", err))
 			return

app/lifecycle/logging_nonwindows.go

@@ -5,5 +5,5 @@ package lifecycle
 import "log/slog"
 
 func ShowLogs() {
-	slog.Warn("ShowLogs not yet implemented")
+	slog.Warn("not implemented")
 }

app/lifecycle/logging_test.go

@@ -17,7 +17,7 @@ func TestRotateLogs(t *testing.T) {
 	// No log exists
 	rotateLogs(logFile)
 
-	require.NoError(t, os.WriteFile(logFile, []byte("1"), 0644))
+	require.NoError(t, os.WriteFile(logFile, []byte("1"), 0o644))
 	assert.FileExists(t, logFile)
 	// First rotation
 	rotateLogs(logFile)
@@ -32,7 +32,7 @@ func TestRotateLogs(t *testing.T) {
 	assert.NoFileExists(t, logFile)
 
 	for i := 2; i <= LogRotationCount+1; i++ {
-		require.NoError(t, os.WriteFile(logFile, []byte(strconv.Itoa(i)), 0644))
+		require.NoError(t, os.WriteFile(logFile, []byte(strconv.Itoa(i)), 0o644))
 		assert.FileExists(t, logFile)
 		rotateLogs(logFile)
 		assert.NoFileExists(t, logFile)

app/lifecycle/server.go

@@ -55,7 +55,7 @@ func start(ctx context.Context, command string) (*exec.Cmd, error) {
 	}
 
 	rotateLogs(ServerLogFile)
-	logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
+	logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0o755)
 	if err != nil {
 		return nil, fmt.Errorf("failed to create server log: %w", err)
 	}

app/lifecycle/updater.go

@@ -15,6 +15,7 @@ import (
 	"path"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
 	"time"
 
@@ -46,7 +47,7 @@ func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) {
 	query.Add("os", runtime.GOOS)
 	query.Add("arch", runtime.GOARCH)
 	query.Add("version", version.Version)
-	query.Add("ts", fmt.Sprintf("%d", time.Now().Unix()))
+	query.Add("ts", strconv.FormatInt(time.Now().Unix(), 10))
 
 	nonce, err := auth.NewNonce(rand.Reader, 16)
 	if err != nil {

app/lifecycle/updater_nonwindows.go

@@ -4,9 +4,9 @@ package lifecycle
 
 import (
 	"context"
-	"fmt"
+	"errors"
 )
 
 func DoUpgrade(cancel context.CancelFunc, done chan int) error {
-	return fmt.Errorf("DoUpgrade not yet implemented")
+	return errors.New("not implemented")
 }

app/lifecycle/updater_windows.go

@@ -2,6 +2,7 @@ package lifecycle
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"log/slog"
 	"os"
@@ -15,7 +16,7 @@ func DoUpgrade(cancel context.CancelFunc, done chan int) error {
 		return fmt.Errorf("failed to lookup downloads: %s", err)
 	}
 	if len(files) == 0 {
-		return fmt.Errorf("no update downloads found")
+		return errors.New("no update downloads found")
 	} else if len(files) > 1 {
 		// Shouldn't happen
 		slog.Warn(fmt.Sprintf("multiple downloads found, using first one %v", files))
@@ -64,7 +65,7 @@ func DoUpgrade(cancel context.CancelFunc, done chan int) error {
 		}
 	} 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")
+		return errors.New("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?

app/ollama.iss

@@ -87,20 +87,11 @@ DialogFontSize=12
 
 [Files]
 Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit
-Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
-Source: "..\dist\windows-{#ARCH}\ollama_runners\*"; DestDir: "{app}\ollama_runners"; Flags: ignoreversion 64bit recursesubdirs
+Source: "..\ollama.exe"; DestDir: "{app}\bin"; Flags: ignoreversion 64bit
+Source: "..\dist\windows-{#ARCH}\lib\ollama\runners\*"; DestDir: "{app}\lib\ollama\runners"; Flags: ignoreversion 64bit recursesubdirs
 Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
 Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
-#if DirExists("..\dist\windows-amd64\cuda")
-  Source: "..\dist\windows-amd64\cuda\*"; DestDir: "{app}\cuda\"; Flags: ignoreversion recursesubdirs
-#endif
-#if DirExists("..\dist\windows-amd64\oneapi")
-  Source: "..\dist\windows-amd64\oneapi\*"; DestDir: "{app}\oneapi\"; Flags: ignoreversion recursesubdirs
-#endif
-#if DirExists("..\dist\windows-amd64\rocm")
-  Source: "..\dist\windows-amd64\rocm\*"; DestDir: "{app}\rocm\"; Flags: ignoreversion recursesubdirs
-#endif
-
+Source: "..\dist\windows-amd64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Flags: ignoreversion recursesubdirs
 
 [Icons]
 Name: "{group}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
@@ -108,7 +99,7 @@ Name: "{userstartup}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilen
 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
+Filename: "{cmd}"; Parameters: "/C set PATH={app}\bin;%PATH% & ""{app}\{#MyAppExeName}"""; Flags: postinstall nowait runhidden
 
 [UninstallRun]
 ; Filename: "{cmd}"; Parameters: "/C ""taskkill /im ''{#MyAppExeName}'' /f /t"; Flags: runhidden
@@ -143,8 +134,8 @@ SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or fi
 
 [Registry]
 Root: HKCU; Subkey: "Environment"; \
-    ValueType: expandsz; ValueName: "Path"; ValueData: "{olddata};{app}"; \
-    Check: NeedsAddPath('{app}')
+    ValueType: expandsz; ValueName: "Path"; ValueData: "{olddata};{app}\bin"; \
+    Check: NeedsAddPath('{app}\bin')
 
 [Code]
 

app/tray/tray_nonwindows.go

@@ -3,11 +3,11 @@
 package tray
 
 import (
-	"fmt"
+	"errors"
 
 	"github.com/ollama/ollama/app/tray/commontray"
 )
 
 func InitPlatformTray(icon, updateIcon []byte) (commontray.OllamaTray, error) {
-	return nil, fmt.Errorf("NOT IMPLEMENTED YET")
+	return nil, errors.New("not implemented")
 }

app/tray/wintray/eventloop.go

@@ -11,9 +11,7 @@ import (
 	"golang.org/x/sys/windows"
 )
 
-var (
-	quitOnce sync.Once
-)
+var quitOnce sync.Once
 
 func (t *winTray) Run() {
 	nativeLoop()

app/tray/wintray/menus.go

@@ -11,12 +11,12 @@ import (
 )
 
 const (
-	updatAvailableMenuID = 1
-	updateMenuID         = updatAvailableMenuID + 1
-	separatorMenuID      = updateMenuID + 1
-	diagLogsMenuID       = separatorMenuID + 1
-	diagSeparatorMenuID  = diagLogsMenuID + 1
-	quitMenuID           = diagSeparatorMenuID + 1
+	updateAvailableMenuID = 1
+	updateMenuID          = updateAvailableMenuID + 1
+	separatorMenuID       = updateMenuID + 1
+	diagLogsMenuID        = separatorMenuID + 1
+	diagSeparatorMenuID   = diagLogsMenuID + 1
+	quitMenuID            = diagSeparatorMenuID + 1
 )
 
 func (t *winTray) initMenus() error {
@@ -35,7 +35,7 @@ func (t *winTray) initMenus() error {
 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 {
+		if err := t.addOrUpdateMenuItem(updateAvailableMenuID, 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 {

app/tray/wintray/tray.go

@@ -11,10 +11,12 @@ import (
 	"path/filepath"
 	"sort"
 	"sync"
+	"syscall"
 	"unsafe"
 
-	"github.com/ollama/ollama/app/tray/commontray"
 	"golang.org/x/sys/windows"
+
+	"github.com/ollama/ollama/app/tray/commontray"
 )
 
 // Helpful sources: https://github.com/golang/exp/blob/master/shiny/driver/internal/win32
@@ -414,7 +416,7 @@ func iconBytesToFilePath(iconBytes []byte) (string, error) {
 	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 {
+		if err := os.WriteFile(iconFilePath, iconBytes, 0o644); err != nil {
 			return "", err
 		}
 	}
@@ -432,7 +434,12 @@ func (t *winTray) setIcon(src string) error {
 	t.muNID.Lock()
 	defer t.muNID.Unlock()
 	t.nid.Icon = h
-	t.nid.Flags |= NIF_ICON
+	t.nid.Flags |= NIF_ICON | NIF_TIP
+	if toolTipUTF16, err := syscall.UTF16FromString(commontray.ToolTip); err == nil {
+		copy(t.nid.Tip[:], toolTipUTF16)
+	} else {
+		return err
+	}
 	t.nid.Size = uint32(unsafe.Sizeof(*t.nid))
 
 	return t.nid.modify()

app/tray/wintray/w32api.go

@@ -61,6 +61,7 @@ const (
 	MIIM_SUBMENU        = 0x00000004
 	MIM_APPLYTOSUBMENUS = 0x80000000
 	NIF_ICON            = 0x00000002
+	NIF_TIP             = 0x00000004
 	NIF_INFO            = 0x00000010
 	NIF_MESSAGE         = 0x00000001
 	SW_HIDE             = 0

auth/auth.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"crypto/rand"
 	"encoding/base64"
+	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -78,7 +79,7 @@ func Sign(ctx context.Context, bts []byte) (string, error) {
 	publicKey := ssh.MarshalAuthorizedKey(privateKey.PublicKey())
 	parts := bytes.Split(publicKey, []byte(" "))
 	if len(parts) < 2 {
-		return "", fmt.Errorf("malformed public key")
+		return "", errors.New("malformed public key")
 	}
 
 	signedData, err := privateKey.Sign(rand.Reader, bts)

cmd/cmd.go

@@ -22,6 +22,7 @@ import (
 	"runtime"
 	"slices"
 	"strings"
+	"sync/atomic"
 	"syscall"
 	"time"
 
@@ -78,6 +79,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	status := "transferring model data"
 	spinner := progress.NewSpinner(status)
 	p.Add(status, spinner)
+	defer p.Stop()
 
 	for i := range modelfile.Commands {
 		switch modelfile.Commands[i].Name {
@@ -112,7 +114,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				path = tempfile
 			}
 
-			digest, err := createBlob(cmd, client, path)
+			digest, err := createBlob(cmd, client, path, spinner)
 			if err != nil {
 				return err
 			}
@@ -202,6 +204,12 @@ func tempZipFiles(path string) (string, error) {
 		// safetensors files might be unresolved git lfs references; skip if they are
 		// covers model-x-of-y.safetensors, model.fp32-x-of-y.safetensors, model.safetensors
 		files = append(files, st...)
+	} else if st, _ := glob(filepath.Join(path, "adapters.safetensors"), "application/octet-stream"); len(st) > 0 {
+		// covers adapters.safetensors
+		files = append(files, st...)
+	} else if st, _ := glob(filepath.Join(path, "adapter_model.safetensors"), "application/octet-stream"); len(st) > 0 {
+		// covers adapter_model.safetensors
+		files = append(files, st...)
 	} else if pt, _ := glob(filepath.Join(path, "pytorch_model*.bin"), "application/zip"); len(pt) > 0 {
 		// pytorch files might also be unresolved git lfs references; skip if they are
 		// covers pytorch_model-x-of-y.bin, pytorch_model.fp32-x-of-y.bin, pytorch_model.bin
@@ -221,6 +229,14 @@ func tempZipFiles(path string) (string, error) {
 	}
 	files = append(files, js...)
 
+	// bert models require a nested config.json
+	// TODO(mxyng): merge this with the glob above
+	js, err = glob(filepath.Join(path, "**/*.json"), "text/plain")
+	if err != nil {
+		return "", err
+	}
+	files = append(files, js...)
+
 	if tks, _ := glob(filepath.Join(path, "tokenizer.model"), "application/octet-stream"); len(tks) > 0 {
 		// add tokenizer.model if it exists, tokenizer.json is automatically picked up by the previous glob
 		// tokenizer.model might be a unresolved git lfs reference; error if it is
@@ -250,6 +266,11 @@ func tempZipFiles(path string) (string, error) {
 			return "", err
 		}
 
+		zfi.Name, err = filepath.Rel(path, file)
+		if err != nil {
+			return "", err
+		}
+
 		zf, err := zipfile.CreateHeader(zfi)
 		if err != nil {
 			return "", err
@@ -263,13 +284,20 @@ func tempZipFiles(path string) (string, error) {
 	return tempfile.Name(), nil
 }
 
-func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, error) {
+func createBlob(cmd *cobra.Command, client *api.Client, path string, spinner *progress.Spinner) (string, error) {
 	bin, err := os.Open(path)
 	if err != nil {
 		return "", err
 	}
 	defer bin.Close()
 
+	// Get file info to retrieve the size
+	fileInfo, err := bin.Stat()
+	if err != nil {
+		return "", err
+	}
+	fileSize := fileInfo.Size()
+
 	hash := sha256.New()
 	if _, err := io.Copy(hash, bin); err != nil {
 		return "", err
@@ -279,13 +307,43 @@ func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, er
 		return "", err
 	}
 
+	var pw progressWriter
+	status := "transferring model data 0%"
+	spinner.SetMessage(status)
+
+	done := make(chan struct{})
+	defer close(done)
+
+	go func() {
+		ticker := time.NewTicker(60 * time.Millisecond)
+		defer ticker.Stop()
+		for {
+			select {
+			case <-ticker.C:
+				spinner.SetMessage(fmt.Sprintf("transferring model data %d%%", int(100*pw.n.Load()/fileSize)))
+			case <-done:
+				spinner.SetMessage("transferring model data 100%")
+				return
+			}
+		}
+	}()
+
 	digest := fmt.Sprintf("sha256:%x", hash.Sum(nil))
-	if err = client.CreateBlob(cmd.Context(), digest, bin); err != nil {
+	if err = client.CreateBlob(cmd.Context(), digest, io.TeeReader(bin, &pw)); err != nil {
 		return "", err
 	}
 	return digest, nil
 }
 
+type progressWriter struct {
+	n atomic.Int64
+}
+
+func (w *progressWriter) Write(p []byte) (n int, err error) {
+	w.n.Add(int64(len(p)))
+	return len(p), nil
+}
+
 func RunHandler(cmd *cobra.Command, args []string) error {
 	interactive := true
 
@@ -1086,7 +1144,7 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 	return nil
 }
 
-func RunServer(cmd *cobra.Command, _ []string) error {
+func RunServer(_ *cobra.Command, _ []string) error {
 	if err := initializeKeypair(); err != nil {
 		return err
 	}
@@ -1160,7 +1218,7 @@ func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
 			return err
 		}
 		if err := startApp(cmd.Context(), client); err != nil {
-			return fmt.Errorf("could not connect to ollama app, is it running?")
+			return errors.New("could not connect to ollama app, is it running?")
 		}
 	}
 	return nil

cmd/interactive.go

@@ -604,7 +604,7 @@ func getImageData(filePath string) ([]byte, error) {
 	// 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)")
+		return nil, errors.New("file size exceeds maximum limit (100MB)")
 	}
 
 	buf = make([]byte, info.Size())

cmd/start_darwin.go

@@ -2,7 +2,7 @@ package cmd
 
 import (
 	"context"
-	"fmt"
+	"errors"
 	"os"
 	"os/exec"
 	"strings"
@@ -20,7 +20,7 @@ func startApp(ctx context.Context, client *api.Client) error {
 		return err
 	}
 	if !strings.Contains(link, "Ollama.app") {
-		return fmt.Errorf("could not find ollama app")
+		return errors.New("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 {

cmd/start_default.go

@@ -4,11 +4,11 @@ package cmd
 
 import (
 	"context"
-	"fmt"
+	"errors"
 
 	"github.com/ollama/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")
+	return errors.New("could not connect to ollama server, run 'ollama serve' to start it")
 }

cmd/start_windows.go

@@ -31,7 +31,7 @@ func startApp(ctx context.Context, client *api.Client) error {
 			// Finally look in the path
 			appExe, err = exec.LookPath(AppName)
 			if err != nil {
-				return fmt.Errorf("could not locate ollama app")
+				return errors.New("could not locate ollama app")
 			}
 		}
 	}

convert/convert.go

@@ -7,16 +7,27 @@ import (
 	"io"
 	"io/fs"
 	"log/slog"
+	"strings"
 
 	"github.com/ollama/ollama/llm"
 )
 
-type Parameters struct {
+type ModelParameters struct {
 	Architectures []string `json:"architectures"`
 	VocabSize     uint32   `json:"vocab_size"`
 }
 
-func (Parameters) KV(t *Tokenizer) llm.KV {
+type AdapterParameters struct {
+	Alpha          uint32 `json:"lora_alpha"`
+	LoraLayers     uint32 `json:"lora_layers"`
+	LoraParameters struct {
+		Rank  uint32  `json:"rank"`
+		Alpha float32 `json:"alpha"`
+		Scale float32 `json:"scale"`
+	} `json:"lora_parameters"`
+}
+
+func (ModelParameters) KV(t *Tokenizer) llm.KV {
 	kv := llm.KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
@@ -27,6 +38,10 @@ func (Parameters) KV(t *Tokenizer) llm.KV {
 		"tokenizer.ggml.token_type":    t.Vocabulary.Types,
 	}
 
+	if len(t.Merges) > 0 {
+		kv["tokenizer.ggml.merges"] = t.Merges
+	}
+
 	if t.Template != "" {
 		kv["tokenizer.chat_template"] = t.Template
 	}
@@ -39,40 +54,119 @@ func (Parameters) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (Parameters) specialTokenTypes() []string {
+func (p AdapterParameters) KV() llm.KV {
+	var alpha float32
+	if p.LoraParameters.Alpha == 0 {
+		alpha = float32(p.Alpha)
+	} else {
+		alpha = p.LoraParameters.Alpha
+	}
+
+	kv := llm.KV{
+		"adapter.lora.alpha": alpha,
+		"adapter.type":       "lora",
+		"general.file_type":  uint32(1),
+		"general.type":       "adapter",
+		"general.version":    "v0.2",
+	}
+
+	return kv
+}
+
+func (ModelParameters) specialTokenTypes() []string {
 	return []string{
 		"bos", "eos", "unk", "sep", "pad", "cls", "mask",
 	}
 }
 
-func (Parameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+func (ModelParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
 	return llm.WriteGGUF(ws, kv, ts)
 }
 
-type Converter interface {
+func (AdapterParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+	return llm.WriteGGUF(ws, kv, ts)
+}
+
+type ModelConverter interface {
 	// KV maps parameters to LLM key-values
 	KV(*Tokenizer) llm.KV
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
 	Tensors([]Tensor) []llm.Tensor
+	// Replacements returns a list of string pairs to replace in tensor names.
+	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
+	Replacements() []string
 
-	// tensorName returns the LLM tensor name for a specific input name
-	tensorName(string) string
 	// specialTokenTypes returns any special token types the model uses
 	specialTokenTypes() []string
+	// writeFile writes the model to the provided io.WriteSeeker
 	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
 }
 
+type moreParser interface {
+	parseMore(fs.FS) error
+}
+
+type AdapterConverter interface {
+	// KV maps parameters to LLM key-values
+	KV(llm.KV) llm.KV
+	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
+	Tensors([]Tensor) []llm.Tensor
+	// Replacements returns a list of string pairs to replace in tensor names.
+	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
+	Replacements() []string
+
+	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
+}
+
+func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV llm.KV) error {
+	bts, err := fs.ReadFile(fsys, "adapter_config.json")
+	if err != nil {
+		return err
+	}
+
+	var p AdapterParameters
+	if err := json.Unmarshal(bts, &p); err != nil {
+		return err
+	}
+
+	arch, ok := baseKV["general.architecture"]
+	if !ok {
+		return errors.New("architecture not set for the base model")
+	}
+
+	var conv AdapterConverter
+	switch arch {
+	case "llama":
+		conv = &llamaAdapter{}
+	case "gemma2":
+		conv = &gemma2Adapter{}
+	default:
+		return errors.New("unsupported architecture")
+	}
+
+	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
+	if err != nil {
+		return err
+	}
+
+	if err := json.Unmarshal(bts, conv); err != nil {
+		return err
+	}
+
+	return conv.writeFile(ws, conv.KV(baseKV), conv.Tensors(ts))
+}
+
 // Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
 // and files it finds in the input path.
 // Supported input model formats include safetensors.
 // Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
-func Convert(fsys fs.FS, ws io.WriteSeeker) error {
+func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 	bts, err := fs.ReadFile(fsys, "config.json")
 	if err != nil {
 		return err
 	}
 
-	var p Parameters
+	var p ModelParameters
 	if err := json.Unmarshal(bts, &p); err != nil {
 		return err
 	}
@@ -81,14 +175,20 @@ func Convert(fsys fs.FS, ws io.WriteSeeker) error {
 		return errors.New("unknown architecture")
 	}
 
-	var conv Converter
+	var conv ModelConverter
 	switch p.Architectures[0] {
 	case "LlamaForCausalLM", "MistralForCausalLM":
-		conv = &llama{}
+		conv = &llamaModel{}
 	case "MixtralForCausalLM":
-		conv = &mixtral{}
+		conv = &mixtralModel{}
 	case "GemmaForCausalLM":
-		conv = &gemma{}
+		conv = &gemmaModel{}
+	case "Gemma2ForCausalLM":
+		conv = &gemma2Model{}
+	case "Phi3ForCausalLM":
+		conv = &phi3Model{}
+	case "BertModel":
+		conv = &bertModel{}
 	default:
 		return errors.New("unsupported architecture")
 	}
@@ -97,6 +197,12 @@ func Convert(fsys fs.FS, ws io.WriteSeeker) error {
 		return err
 	}
 
+	if t, ok := conv.(moreParser); ok {
+		if err := t.parseMore(fsys); err != nil {
+			return err
+		}
+	}
+
 	t, err := parseTokenizer(fsys, conv.specialTokenTypes())
 	if err != nil {
 		return err
@@ -113,7 +219,7 @@ func Convert(fsys fs.FS, ws io.WriteSeeker) error {
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}
 
-	ts, err := parseTensors(fsys)
+	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
 	if err != nil {
 		return err
 	}

convert/convert_bert.go

@@ -0,0 +1,174 @@
+package convert
+
+import (
+	"cmp"
+	"encoding/json"
+	"io/fs"
+	"path/filepath"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type bertModel struct {
+	ModelParameters
+	NLayers               uint32  `json:"n_layers"`
+	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
+	NLayer                uint32  `json:"n_layer"`
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	NCtx                  uint32  `json:"n_ctx"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	NEmbd                 uint32  `json:"n_embd"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NInner                uint32  `json:"n_inner"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NHead                 uint32  `json:"n_head"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	LayerNormEPS          float32 `json:"layer_norm_eps"`
+	LayerNormEpsilon      float32 `json:"layer_norm_epsilon"`
+	NormEpsilon           float32 `json:"norm_epsilon"`
+
+	PoolingType uint32
+}
+
+var (
+	_ ModelConverter = (*bertModel)(nil)
+	_ moreParser     = (*bertModel)(nil)
+)
+
+func (p *bertModel) parseMore(fsys fs.FS) error {
+	bts, err := fs.ReadFile(fsys, "modules.json")
+	if err != nil {
+		return err
+	}
+
+	var modules []struct {
+		Type string `json:"type"`
+		Path string `json:"path"`
+	}
+
+	if err := json.Unmarshal(bts, &modules); err != nil {
+		return err
+	}
+
+	var pooling string
+	for _, m := range modules {
+		if m.Type == "sentence_transformers.models.Pooling" {
+			pooling = m.Path
+			break
+		}
+	}
+
+	if pooling != "" {
+		bts, err := fs.ReadFile(fsys, filepath.Join(pooling, "config.json"))
+		if err != nil {
+			return err
+		}
+
+		var pc struct {
+			PoolingModeCLSToken   bool `json:"pooling_mode_cls_token"`
+			PoolingModeMeanTokens bool `json:"pooling_mode_mean_tokens"`
+		}
+
+		if err := json.Unmarshal(bts, &pc); err != nil {
+			return err
+		}
+
+		if pc.PoolingModeMeanTokens {
+			p.PoolingType = 1
+		} else if pc.PoolingModeCLSToken {
+			p.PoolingType = 2
+		}
+	}
+
+	return nil
+}
+
+func (p *bertModel) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "bert"
+	kv["bert.attention.causal"] = false
+	kv["bert.pooling_type"] = p.PoolingType
+
+	kv["bert.block_count"] = cmp.Or(p.NLayers, p.NumHiddenLayers, p.NLayer)
+
+	if contextLength := cmp.Or(p.MaxPositionEmbeddings, p.NCtx); contextLength > 0 {
+		kv["bert.context_length"] = contextLength
+	}
+
+	if embeddingLength := cmp.Or(p.HiddenSize, p.NEmbd); embeddingLength > 0 {
+		kv["bert.embedding_length"] = cmp.Or(p.HiddenSize, p.NEmbd)
+	}
+
+	if feedForwardLength := cmp.Or(p.IntermediateSize, p.NInner); feedForwardLength > 0 {
+		kv["bert.feed_forward_length"] = cmp.Or(p.IntermediateSize, p.NInner)
+	}
+
+	if headCount := cmp.Or(p.NumAttentionHeads, p.NHead); headCount > 0 {
+		kv["bert.attention.head_count"] = cmp.Or(p.NumAttentionHeads, p.NHead)
+	}
+
+	if layerNormEpsilon := cmp.Or(p.LayerNormEPS, p.LayerNormEpsilon, p.NormEpsilon); layerNormEpsilon > 0 {
+		kv["bert.attention.layer_norm_epsilon"] = layerNormEpsilon
+	}
+
+	kv["tokenizer.ggml.model"] = "bert"
+	kv["tokenizer.ggml.token_type_count"] = uint32(2)
+
+	// convert to phantom space tokens
+	for i, e := range t.Tokens {
+		if strings.HasPrefix(e, "[") && strings.HasSuffix(e, "]") {
+			// noop
+		} else if strings.HasPrefix(e, "##") {
+			t.Tokens[i] = e[2:]
+		} else {
+			t.Tokens[i] = "\u2581" + e
+		}
+	}
+
+	kv["tokenizer.ggml.tokens"] = t.Tokens
+
+	return kv
+}
+
+func (p *bertModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+	for _, t := range ts {
+		if slices.Contains([]string{
+			"embeddings.position_ids",
+			"pooler.dense.weight",
+			"pooler.dense.bias",
+		}, t.Name()) {
+			continue
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (bertModel) Replacements() []string {
+	return []string{
+		"encoder.layer", "blk",
+		"encoder.layers", "blk",
+		"embeddings.word_embeddings", "token_embd",
+		"embeddings.token_type_embeddings", "token_types",
+		"embeddings.LayerNorm", "token_embd_norm",
+		"embeddings.position_embeddings", "position_embd",
+		"attention.self.query", "attn_q",
+		"attention.self.key", "attn_k",
+		"attention.self.value", "attn_v",
+		"attention.output.dense", "attn_output",
+		"attention.output.LayerNorm", "attn_output_norm",
+		"intermediate.dense", "ffn_up",
+		"output.dense", "ffn_down",
+		"output.LayerNorm", "layer_output_norm",
+	}
+}

convert/convert_gemma.go

@@ -9,8 +9,8 @@ import (
 	"github.com/ollama/ollama/llm"
 )
 
-type gemma struct {
-	Parameters
+type gemmaModel struct {
+	ModelParameters
 	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
 	HiddenSize            uint32  `json:"hidden_size"`
 	HiddenLayers          uint32  `json:"num_hidden_layers"`
@@ -21,12 +21,11 @@ type gemma struct {
 	HeadDim               uint32  `json:"head_dim"`
 }
 
-var _ Converter = (*gemma)(nil)
+var _ ModelConverter = (*gemmaModel)(nil)
 
-func (p *gemma) KV(t *Tokenizer) llm.KV {
-	kv := p.Parameters.KV(t)
+func (p *gemmaModel) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma"
-	kv["general.name"] = "gemma"
 	kv["gemma.context_length"] = p.MaxPositionEmbeddings
 	kv["gemma.embedding_length"] = p.HiddenSize
 	kv["gemma.block_count"] = p.HiddenLayers
@@ -43,16 +42,15 @@ func (p *gemma) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *gemma) Tensors(ts []Tensor) []llm.Tensor {
+func (p *gemmaModel) Tensors(ts []Tensor) []llm.Tensor {
 	var out []llm.Tensor
 	for _, t := range ts {
-		name := p.tensorName(t.Name())
-		if strings.HasSuffix(name, "_norm.weight") {
+		if strings.HasSuffix(t.Name(), "_norm.weight") {
 			t.SetRepacker(p.addOne)
 		}
 
 		out = append(out, llm.Tensor{
-			Name:     name,
+			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),
 			WriterTo: t,
@@ -62,8 +60,8 @@ func (p *gemma) Tensors(ts []Tensor) []llm.Tensor {
 	return out
 }
 
-func (p *gemma) tensorName(n string) string {
-	return strings.NewReplacer(
+func (p *gemmaModel) Replacements() []string {
+	return []string{
 		"model.embed_tokens", "token_embd",
 		"model.norm", "output_norm",
 		"model.layers", "blk",
@@ -76,11 +74,10 @@ func (p *gemma) tensorName(n string) string {
 		"mlp.down_proj", "ffn_down",
 		"mlp.up_proj", "ffn_up",
 		"post_attention_layernorm", "ffn_norm",
-		"block_sparse_moe.gate", "ffn_inp",
-	).Replace(n)
+	}
 }
 
-func (*gemma) addOne(_ string, data []float32, shape []uint64) ([]float32, error) {
+func (*gemmaModel) addOne(_ string, data []float32, shape []uint64) ([]float32, error) {
 	n := tensor.New(tensor.WithShape(int(shape[0])), tensor.WithBacking(data))
 	ones := tensor.Ones(tensor.Float32, int(shape[0]))
 

convert/convert_gemma2.go

@@ -0,0 +1,43 @@
+package convert
+
+import (
+	"github.com/ollama/ollama/llm"
+)
+
+type gemma2Model struct {
+	gemmaModel
+	SlidingWindow         uint32  `json:"sliding_window"`
+	AttentionLogitSoftcap float32 `json:"attn_logit_softcapping"`
+	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
+}
+
+func (p *gemma2Model) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "gemma2"
+	kv["gemma2.context_length"] = p.MaxPositionEmbeddings
+	kv["gemma2.embedding_length"] = p.HiddenSize
+	kv["gemma2.block_count"] = p.HiddenLayers
+	kv["gemma2.feed_forward_length"] = p.IntermediateSize
+	kv["gemma2.attention.head_count"] = p.NumAttentionHeads
+	kv["gemma2.attention.head_count_kv"] = p.NumKeyValueHeads
+	kv["gemma2.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["gemma2.attention.key_length"] = p.HeadDim
+	kv["gemma2.attention.value_length"] = p.HeadDim
+	kv["gemma2.attention.sliding_window"] = p.SlidingWindow
+	kv["gemma2.attn_logit_softcapping"] = p.AttentionLogitSoftcap
+	kv["gemma2.final_logit_softcapping"] = p.FinalLogitSoftcap
+	kv["tokenizer.ggml.eot_token_id"] = uint32(107)
+	kv["tokenizer.ggml.middle_token_id"] = uint32(68)
+	kv["tokenizer.ggml.prefix_token_id"] = uint32(67)
+	kv["tokenizer.ggml.suffix_token_id"] = uint32(69)
+	return kv
+}
+
+func (p *gemma2Model) Replacements() []string {
+	return append(
+		p.gemmaModel.Replacements(),
+		"post_attention_layernorm", "post_attention_norm",
+		"pre_feedforward_layernorm", "ffn_norm",
+		"post_feedforward_layernorm", "post_ffw_norm",
+	)
+}

convert/convert_gemma2_adapter.go

@@ -0,0 +1,91 @@
+package convert
+
+import (
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type gemma2Adapter struct {
+	AdapterParameters
+}
+
+var _ AdapterConverter = (*gemma2Adapter)(nil)
+
+func (p *gemma2Adapter) KV(baseKV llm.KV) llm.KV {
+	kv := p.AdapterParameters.KV()
+	kv["general.architecture"] = "gemma2"
+	return kv
+}
+
+func (p *gemma2Adapter) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+	for _, t := range ts {
+		shape := t.Shape()
+		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
+			(strings.HasSuffix(t.Name(), "weight.lora_b") && shape[0] < shape[1]) {
+			shape[0], shape[1] = shape[1], shape[0]
+			t.SetRepacker(p.repack)
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *gemma2Adapter) Replacements() []string {
+	return []string{
+		"base_model.model.", "",
+		"model.layers", "blk",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"lora_A.weight", "weight.lora_a",
+		"lora_B.weight", "weight.lora_b",
+		"lora_a", "weight.lora_a",
+		"lora_b", "weight.lora_b",
+	}
+}
+
+func (p *gemma2Adapter) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	dims := []int{int(shape[1]), int(shape[0])}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+
+	if err := n.T(1, 0); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}

convert/convert_llama.go

@@ -3,15 +3,17 @@ package convert
 import (
 	"cmp"
 	"fmt"
+	"math"
 	"strings"
 
-	"github.com/ollama/ollama/llm"
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
 )
 
-type llama struct {
-	Parameters
+type llamaModel struct {
+	ModelParameters
 	NLayers               uint32  `json:"n_layers"`
 	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
 	NLayer                uint32  `json:"n_layer"`
@@ -26,8 +28,14 @@ type llama struct {
 	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
 	RopeTheta             float32 `json:"rope_theta"`
 	RopeScaling           struct {
-		Type   string  `json:"type"`
-		Factor float32 `json:"factor"`
+		Type                            string  `json:"type"`
+		RopeType                        string  `json:"rope_type"`
+		Factor                          float32 `json:"factor"`
+		LowFrequencyFactor              float32 `json:"low_freq_factor"`
+		HighFrequencyFactor             float32 `json:"high_freq_factor"`
+		OriginalMaxPositionalEmbeddings uint32  `json:"original_max_positional_embeddings"`
+
+		factors ropeFactor
 	} `json:"rope_scaling"`
 	RMSNormEPS       float32 `json:"rms_norm_eps"`
 	LayerNormEPS     float32 `json:"layer_norm_eps"`
@@ -36,12 +44,11 @@ type llama struct {
 	HeadDim          uint32  `json:"head_dim"`
 }
 
-var _ Converter = (*llama)(nil)
+var _ ModelConverter = (*llamaModel)(nil)
 
-func (p *llama) KV(t *Tokenizer) llm.KV {
-	kv := p.Parameters.KV(t)
+func (p *llamaModel) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama"
-	kv["general.name"] = "llama"
 	kv["llama.vocab_size"] = p.VocabSize
 
 	kv["llama.block_count"] = cmp.Or(p.NLayers, p.NumHiddenLayers, p.NLayer)
@@ -70,6 +77,27 @@ func (p *llama) KV(t *Tokenizer) llm.KV {
 	if p.RopeScaling.Type == "linear" {
 		kv["llama.rope.scaling.type"] = p.RopeScaling.Type
 		kv["llama.rope.scaling.factor"] = p.RopeScaling.Factor
+	} else if p.RopeScaling.RopeType == "llama3" {
+		dim := p.HiddenSize / p.NumAttentionHeads
+		for i := uint32(0); i < dim; i += 2 {
+			factor := cmp.Or(p.RopeScaling.Factor, 8.0)
+			factorLow := cmp.Or(p.RopeScaling.LowFrequencyFactor, 1.0)
+			factorHigh := cmp.Or(p.RopeScaling.HighFrequencyFactor, 4.0)
+
+			original := cmp.Or(p.RopeScaling.OriginalMaxPositionalEmbeddings, 8192)
+			lambdaLow := float32(original) / factorLow
+			lambdaHigh := float32(original) / factorHigh
+
+			lambda := 2 * math.Pi * math.Pow(float64(p.RopeTheta), float64(i)/float64(dim))
+			if lambda < float64(lambdaHigh) {
+				p.RopeScaling.factors = append(p.RopeScaling.factors, 1.0)
+			} else if lambda > float64(lambdaLow) {
+				p.RopeScaling.factors = append(p.RopeScaling.factors, factor)
+			} else {
+				smooth := (float32(original)/float32(lambda) - factorLow) / (factorHigh - factorLow)
+				p.RopeScaling.factors = append(p.RopeScaling.factors, 1.0/((1-smooth)/factor+smooth))
+			}
+		}
 	}
 
 	if p.NumKeyValueHeads > 0 {
@@ -89,24 +117,29 @@ func (p *llama) KV(t *Tokenizer) llm.KV {
 		kv["llama.attention.value_length"] = p.HeadDim
 	}
 
-	if len(t.Merges) > 0 {
-		kv["tokenizer.ggml.merges"] = t.Merges
-	}
-
 	return kv
 }
 
-func (p *llama) Tensors(ts []Tensor) []llm.Tensor {
+func (p *llamaModel) Tensors(ts []Tensor) []llm.Tensor {
 	var out []llm.Tensor
+
+	if p.RopeScaling.factors != nil {
+		out = append(out, llm.Tensor{
+			Name:     "rope_freqs.weight",
+			Kind:     0,
+			Shape:    []uint64{uint64(len(p.RopeScaling.factors))},
+			WriterTo: p.RopeScaling.factors,
+		})
+	}
+
 	for _, t := range ts {
-		name := p.tensorName(t.Name())
-		if strings.HasSuffix(name, "attn_q.weight") ||
-			strings.HasSuffix(name, "attn_k.weight") {
+		if strings.HasSuffix(t.Name(), "attn_q.weight") ||
+			strings.HasSuffix(t.Name(), "attn_k.weight") {
 			t.SetRepacker(p.repack)
 		}
 
 		out = append(out, llm.Tensor{
-			Name:     name,
+			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),
 			WriterTo: t,
@@ -116,8 +149,8 @@ func (p *llama) Tensors(ts []Tensor) []llm.Tensor {
 	return out
 }
 
-func (p *llama) tensorName(n string) string {
-	return strings.NewReplacer(
+func (p *llamaModel) Replacements() []string {
+	return []string{
 		"lm_head", "output",
 		"model.embed_tokens", "token_embd",
 		"model.norm", "output_norm",
@@ -131,21 +164,19 @@ func (p *llama) tensorName(n string) string {
 		"mlp.down_proj", "ffn_down",
 		"mlp.up_proj", "ffn_up",
 		"post_attention_layernorm", "ffn_norm",
-		// mixtral
-		"block_sparse_moe.gate", "ffn_gate_inp",
-	).Replace(n)
+	}
 }
 
-func (p *llama) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+func (p *llamaModel) repack(name string, data []float32, shape []uint64) ([]float32, error) {
 	var dims []int
 	for _, dim := range shape {
 		dims = append(dims, int(dim))
 	}
 
 	var heads uint32
-	if strings.HasSuffix(name, "q_proj.weight") {
+	if strings.HasSuffix(name, "attn_q.weight") {
 		heads = p.NumAttentionHeads
-	} else if strings.HasSuffix(name, "k_proj.weight") {
+	} else if strings.HasSuffix(name, "attn_k.weight") {
 		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
 	} else {
 		return nil, fmt.Errorf("unknown tensor for repack: %s", name)

convert/convert_llama_adapter.go

@@ -0,0 +1,169 @@
+package convert
+
+import (
+	"cmp"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type llamaAdapter struct {
+	AdapterParameters
+	NumAttentionHeads uint32 `json:"num_attention_heads"`
+	NumKeyValueHeads  uint32 `json:"num_key_value_heads"`
+}
+
+var _ AdapterConverter = (*llamaAdapter)(nil)
+
+func (p *llamaAdapter) KV(baseKV llm.KV) llm.KV {
+	kv := p.AdapterParameters.KV()
+	kv["general.architecture"] = "llama"
+	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
+	kv["llama.attention.head_count_kv"] = baseKV["llama.attention.head_count_kv"]
+
+	p.NumAttentionHeads = baseKV["llama.attention.head_count"].(uint32)
+
+	return kv
+}
+
+func (p *llamaAdapter) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+	for _, t := range ts {
+		shape := t.Shape()
+		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
+			(strings.HasSuffix(t.Name(), "weight.lora_b") && shape[0] < shape[1]) {
+			shape[0], shape[1] = shape[1], shape[0]
+			t.SetRepacker(p.repackAndTranspose)
+		} else {
+			t.SetRepacker(p.repack)
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    shape,
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *llamaAdapter) Replacements() []string {
+	return []string{
+		"base_model.model.", "",
+		"model.layers", "blk",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"lora_A.weight", "weight.lora_a",
+		"lora_B.weight", "weight.lora_b",
+		"lora_a", "weight.lora_a",
+		"lora_b", "weight.lora_b",
+	}
+}
+
+func (p *llamaAdapter) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	dims := []int{int(shape[1]), int(shape[0])}
+
+	var heads uint32
+	if strings.HasSuffix(name, "attn_q.weight.lora_a") {
+		heads = p.NumAttentionHeads
+	} else if strings.HasSuffix(name, "attn_k.weight.lora_a") {
+		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+	} else {
+		return data, nil
+	}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+
+	if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}
+
+func (p *llamaAdapter) repackAndTranspose(name string, data []float32, shape []uint64) ([]float32, error) {
+	dims := []int{int(shape[1]), int(shape[0])}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+
+	var heads uint32
+	if strings.HasSuffix(name, "attn_q.weight.lora_a") {
+		heads = p.NumAttentionHeads
+	} else if strings.HasSuffix(name, "attn_k.weight.lora_a") {
+		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+	}
+
+	if heads > 0 {
+		if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
+			return nil, err
+		}
+
+		if err := n.T(0, 2, 1, 3); err != nil {
+			return nil, err
+		}
+
+		if err := n.Reshape(dims...); err != nil {
+			return nil, err
+		}
+
+		if err := n.Transpose(); err != nil {
+			return nil, err
+		}
+	}
+
+	if err := n.T(1, 0); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}

convert/convert_mixtral.go

@@ -9,16 +9,14 @@ import (
 	"github.com/ollama/ollama/llm"
 )
 
-type mixtral struct {
-	llama
+type mixtralModel struct {
+	llamaModel
 	NumLocalExperts    uint32 `json:"num_local_experts"`
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }
 
-var _ Converter = (*mixtral)(nil)
-
-func (p *mixtral) KV(t *Tokenizer) llm.KV {
-	kv := p.llama.KV(t)
+func (p *mixtralModel) KV(t *Tokenizer) llm.KV {
+	kv := p.llamaModel.KV(t)
 
 	if p.NumLocalExperts > 0 {
 		kv["llama.expert_count"] = p.NumLocalExperts
@@ -31,7 +29,7 @@ func (p *mixtral) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *mixtral) Tensors(ts []Tensor) []llm.Tensor {
+func (p *mixtralModel) Tensors(ts []Tensor) []llm.Tensor {
 	oldnew := []string{
 		"model.layers", "blk",
 		"w1", "ffn_gate_exps",
@@ -69,7 +67,14 @@ func (p *mixtral) Tensors(ts []Tensor) []llm.Tensor {
 		})
 	}
 
-	return append(out, p.llama.Tensors(ts)...)
+	return append(out, p.llamaModel.Tensors(ts)...)
+}
+
+func (p *mixtralModel) Replacements() []string {
+	return append(
+		p.llamaModel.Replacements(),
+		"block_sparse_moe.gate", "ffn_gate_inp",
+	)
 }
 
 type experts []Tensor

convert/convert_phi3.go

@@ -0,0 +1,123 @@
+package convert
+
+import (
+	"cmp"
+	"encoding/binary"
+	"io"
+	"math"
+	"strings"
+	"sync"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type phi3Model struct {
+	ModelParameters
+	NumHiddenLayers   uint32  `json:"num_hidden_layers"`
+	NLayers           uint32  `json:"n_layers"`
+	HiddenSize        uint32  `json:"hidden_size"`
+	NEmbd             uint32  `json:"n_embd"`
+	IntermediateSize  uint32  `json:"intermediate_size"`
+	NumAttentionHeads uint32  `json:"num_attention_heads"`
+	NHead             uint32  `json:"n_head"`
+	NumKeyValueHeads  uint32  `json:"num_key_value_heads"`
+	NHeadKV           uint32  `json:"n_head_kv"`
+	RopeTheta         float32 `json:"rope_theta"`
+	RopeScaling       struct {
+		Type        string     `json:"type"`
+		LongFactor  ropeFactor `json:"long_factor"`
+		ShortFactor ropeFactor `json:"short_factor"`
+	} `json:"rope_scaling"`
+	RMSNormEPS                    float32 `json:"rms_norm_eps"`
+	NPositions                    uint32  `json:"n_positions"`
+	MaxPositionEmbeddings         uint32  `json:"max_position_embeddings"`
+	OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
+	SlidingWindow                 uint32  `json:"sliding_window"`
+}
+
+var _ ModelConverter = (*phi3Model)(nil)
+
+func (p *phi3Model) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "phi3"
+	kv["phi3.context_length"] = p.MaxPositionEmbeddings
+	kv["phi3.embedding_length"] = cmp.Or(p.HiddenSize, p.NEmbd)
+	kv["phi3.feed_forward_length"] = p.IntermediateSize
+	kv["phi3.block_count"] = cmp.Or(p.NumHiddenLayers, p.NLayers)
+	kv["phi3.attention.head_count"] = cmp.Or(p.NumAttentionHeads, p.NHead)
+	kv["phi3.attention.head_count_kv"] = cmp.Or(p.NumKeyValueHeads, p.NHeadKV)
+	kv["phi3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["phi3.rope.dimension_count"] = p.HiddenSize / cmp.Or(p.NumAttentionHeads, p.NHead)
+	kv["phi3.rope.freq_base"] = p.RopeTheta
+	kv["phi3.rope.scaling.original_context_length"] = p.OriginalMaxPositionEmbeddings
+	kv["phi3.attention.sliding_window"] = p.SlidingWindow
+
+	scale := float64(p.MaxPositionEmbeddings) / float64(p.OriginalMaxPositionEmbeddings)
+
+	switch p.RopeScaling.Type {
+	case "":
+		// no scaling
+	case "su", "longrope":
+		kv["phi3.rope.scaling.attn_factor"] = float32(max(math.Sqrt(1+math.Log(scale)/math.Log(float64(p.OriginalMaxPositionEmbeddings))), 1.0))
+	case "yarn":
+		kv["phi3.rope.scaling.attn_factor"] = float32(max(0.1*math.Log(scale)+1.0, 1.0))
+	default:
+		panic("unknown rope scaling type")
+	}
+
+	return kv
+}
+
+func (p *phi3Model) Tensors(ts []Tensor) []llm.Tensor {
+	var addRopeFactors sync.Once
+
+	out := make([]llm.Tensor, 0, len(ts)+2)
+	for _, t := range ts {
+		if strings.HasPrefix(t.Name(), "blk.0.") {
+			addRopeFactors.Do(func() {
+				out = append(out, llm.Tensor{
+					Name:     "rope_factors_long.weight",
+					Kind:     0,
+					Shape:    []uint64{uint64(len(p.RopeScaling.LongFactor))},
+					WriterTo: p.RopeScaling.LongFactor,
+				}, llm.Tensor{
+					Name:     "rope_factors_short.weight",
+					Kind:     0,
+					Shape:    []uint64{uint64(len(p.RopeScaling.ShortFactor))},
+					WriterTo: p.RopeScaling.ShortFactor,
+				})
+			})
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *phi3Model) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.qkv_proj", "attn_qkv",
+		"self_attn.o_proj", "attn_output",
+		"mlp.down_proj", "ffn_down",
+		"mlp.gate_up_proj", "ffn_up",
+		"post_attention_layernorm", "ffn_norm",
+	}
+}
+
+type ropeFactor []float32
+
+func (r ropeFactor) WriteTo(w io.Writer) (int64, error) {
+	err := binary.Write(w, binary.LittleEndian, r)
+	return 0, err
+}

convert/convert_test.go

@@ -1,7 +1,10 @@
 package convert
 
 import (
+	"bytes"
 	"crypto/sha256"
+	"encoding/binary"
+	"encoding/hex"
 	"encoding/json"
 	"flag"
 	"fmt"
@@ -14,8 +17,9 @@ import (
 	"slices"
 	"testing"
 
-	"github.com/ollama/ollama/llm"
 	"golang.org/x/exp/maps"
+
+	"github.com/ollama/ollama/llm"
 )
 
 func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
@@ -27,7 +31,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
 	}
 	defer f.Close()
 
-	if err := Convert(fsys, f); err != nil {
+	if err := ConvertModel(fsys, f); err != nil {
 		t.Fatal(err)
 	}
 
@@ -49,6 +53,34 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
 	return r, m.KV(), m.Tensors()
 }
 
+func generateResultsJSON(t *testing.T, f *os.File, kv llm.KV, tensors llm.Tensors) map[string]string {
+	actual := make(map[string]string)
+	for k, v := range kv {
+		if s, ok := v.(json.Marshaler); !ok {
+			actual[k] = fmt.Sprintf("%v", v)
+		} else {
+			bts, err := json.Marshal(s)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			actual[k] = fmt.Sprintf("%x", sha256.Sum256(bts))
+		}
+	}
+
+	for _, tensor := range tensors.Items {
+		sha256sum := sha256.New()
+		sr := io.NewSectionReader(f, int64(tensors.Offset+tensor.Offset), int64(tensor.Size()))
+		if _, err := io.Copy(sha256sum, sr); err != nil {
+			t.Fatal(err)
+		}
+
+		actual[tensor.Name] = hex.EncodeToString(sha256sum.Sum(nil))
+	}
+
+	return actual
+}
+
 func TestMain(m *testing.M) {
 	var level slog.Level
 	flag.TextVar(&level, "level", slog.LevelInfo, "log level")
@@ -60,9 +92,14 @@ func TestMain(m *testing.M) {
 func TestConvertFull(t *testing.T) {
 	cases := []string{
 		"Meta-Llama-3-8B-Instruct",
+		"Meta-Llama-3.1-8B-Instruct",
 		"Mistral-7B-Instruct-v0.2",
 		"Mixtral-8x7B-Instruct-v0.1",
 		"gemma-2b-it",
+		// microsoft/Phi-3-mini-128-instruct@d548c233192db00165d842bf8edff054bb3212f8
+		"Phi-3-mini-128k-instruct",
+		"all-MiniLM-L6-v2",
+		"gemma-2-9b-it",
 	}
 
 	for i := range cases {
@@ -78,29 +115,7 @@ func TestConvertFull(t *testing.T) {
 			}
 
 			f, kv, tensors := convertFull(t, os.DirFS(p))
-			actual := make(map[string]string)
-			for k, v := range kv {
-				if s, ok := v.(json.Marshaler); !ok {
-					actual[k] = fmt.Sprintf("%v", v)
-				} else {
-					bts, err := json.Marshal(s)
-					if err != nil {
-						t.Fatal(err)
-					}
-
-					actual[k] = fmt.Sprintf("%x", sha256.Sum256(bts))
-				}
-			}
-
-			for _, tensor := range tensors.Items {
-				sha256sum := sha256.New()
-				sr := io.NewSectionReader(f, int64(tensors.Offset+tensor.Offset), int64(tensor.Size()))
-				if _, err := io.Copy(sha256sum, sr); err != nil {
-					t.Fatal(err)
-				}
-
-				actual[tensor.Name] = fmt.Sprintf("%x", sha256sum.Sum(nil))
-			}
+			actual := generateResultsJSON(t, f, kv, tensors)
 
 			expectFile, err := os.Open(filepath.Join("testdata", fmt.Sprintf("%s.json", tt)))
 			if err != nil {
@@ -124,3 +139,209 @@ func TestConvertFull(t *testing.T) {
 		})
 	}
 }
+
+func TestConvertAdapter(t *testing.T) {
+	type AdapterCase struct {
+		Name     string
+		BaseKV   map[string]any
+		Expected map[string]string
+	}
+
+	cases := []AdapterCase{
+		{
+			Name: "discollama",
+			BaseKV: map[string]any{
+				"general.architecture":          "llama",
+				"llama.attention.head_count":    uint32(32),
+				"llama.attention.head_count_kv": uint32(8),
+			},
+			Expected: map[string]string{
+				"general.architecture":          "llama",
+				"general.file_type":             "1",
+				"general.parameter_count":       "106496",
+				"general.type":                  "adapter",
+				"general.version":               "v0.2",
+				"adapter.lora.alpha":            "16",
+				"adapter.type":                  "lora",
+				"llama.attention.head_count":    "32",
+				"llama.attention.head_count_kv": "8",
+				"blk.31.attn_q.weight.lora_a":   "0eb3318b02cd313429bcc7621b539fdbb10240fea190c56c9e5f93fcd37a4e50",
+				"blk.31.attn_q.weight.lora_b":   "0eb3318b02cd313429bcc7621b539fdbb10240fea190c56c9e5f93fcd37a4e50",
+				"blk.31.attn_v.weight.lora_a":   "0eb3318b02cd313429bcc7621b539fdbb10240fea190c56c9e5f93fcd37a4e50",
+				"blk.31.attn_v.weight.lora_b":   "071dcafe89df065d6e1c935ecb8fdf6479b3c202eb912e7da938597673ff5857",
+			},
+		},
+	}
+
+	for _, c := range cases {
+		t.Run(c.Name, func(t *testing.T) {
+			t.Parallel()
+
+			f, err := os.CreateTemp(t.TempDir(), "f16")
+			if err != nil {
+				t.Fatal(err)
+			}
+			defer f.Close()
+
+			tempDir := t.TempDir()
+			generateLoraTestData(t, tempDir)
+
+			if err = ConvertAdapter(os.DirFS(tempDir), f, c.BaseKV); err != nil {
+				t.Fatal(err)
+			}
+
+			r, err := os.Open(f.Name())
+			if err != nil {
+				t.Fatal(err)
+			}
+			defer r.Close()
+
+			m, _, err := llm.DecodeGGML(r, math.MaxInt)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			if _, err := r.Seek(0, io.SeekStart); err != nil {
+				t.Fatal(err)
+			}
+
+			actual := generateResultsJSON(t, r, m.KV(), m.Tensors())
+
+			keys := maps.Keys(c.Expected)
+			slices.Sort(keys)
+			for _, k := range keys {
+				if v, ok := actual[k]; !ok {
+					t.Errorf("missing %s", k)
+				} else if v != c.Expected[k] {
+					t.Errorf("unexpected %s: want %s, got %s", k, c.Expected[k], v)
+				}
+			}
+		})
+	}
+}
+
+func generateLoraTestData(t *testing.T, tempDir string) {
+	type tensorData struct {
+		Offsets []int  `json:"data_offsets"`
+		Type    string `json:"dtype"`
+		Shape   []int  `json:"shape"`
+	}
+	offset := 4096 * 8 * 4
+
+	td := map[string]*tensorData{"__metadata__": nil}
+	td["model.layers.31.self_attn.q_proj.lora_a"] = &tensorData{
+		Offsets: []int{0, offset},
+		Type:    "F32",
+		Shape:   []int{4096, 8},
+	}
+	td["model.layers.31.self_attn.q_proj.lora_b"] = &tensorData{
+		Offsets: []int{offset, offset * 2},
+		Type:    "F32",
+		Shape:   []int{8, 4096},
+	}
+	td["model.layers.31.self_attn.v_proj.lora_a"] = &tensorData{
+		Offsets: []int{offset * 2, offset * 3},
+		Type:    "F32",
+		Shape:   []int{4096, 8},
+	}
+	td["model.layers.31.self_attn.v_proj.lora_b"] = &tensorData{
+		Offsets: []int{offset * 3, offset*3 + 8*1024*4},
+		Type:    "F32",
+		Shape:   []int{8, 1024},
+	}
+
+	data, err := json.Marshal(td)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	var buf bytes.Buffer
+
+	l := int64(len(data))
+	err = binary.Write(&buf, binary.LittleEndian, l)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = buf.Write(data)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// write some data for the tensors
+
+	ones := make([]float32, 4096*8)
+	for i := range ones {
+		ones[i] = float32(1)
+	}
+
+	for range 3 {
+		err = binary.Write(&buf, binary.LittleEndian, ones)
+		if err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	ones = make([]float32, 1024*8)
+	for i := range ones {
+		ones[i] = float32(1)
+	}
+
+	err = binary.Write(&buf, binary.LittleEndian, ones)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	fdata, err := os.Create(filepath.Join(tempDir, "adapters.safetensors"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer fdata.Close()
+
+	_, err = fdata.Write(buf.Bytes())
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	configData := `
+{
+    "adapter_path": "adapters-test",
+    "batch_size": 8,
+    "config": "config-tiny.json",
+    "data": "../discollama-completion",
+    "grad_checkpoint": null,
+    "iters": 1000,
+    "learning_rate": 1e-05,
+    "lora_layers": 1,
+    "lora_parameters": {
+        "rank": 8,
+        "alpha": 16,
+        "dropout": 0.0,
+        "scale": 2.0
+    },
+    "lr_schedule": null,
+    "max_seq_length": 2048,
+    "model": "/Users/pdevine/git/Meta-Llama-3-8B-Instruct",
+    "resume_adapter_file": null,
+    "save_every": 100,
+    "seed": 0,
+    "steps_per_eval": 200,
+    "steps_per_report": 10,
+    "test": false,
+    "test_batches": 500,
+    "train": true,
+    "use_dora": false,
+    "val_batches": 25
+}
+`
+	f, err := os.Create(filepath.Join(tempDir, "adapter_config.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	_, err = f.WriteString(configData)
+	if err != nil {
+		t.Fatal(err)
+	}
+}

convert/fs.go

@@ -10,8 +10,8 @@ import (
 )
 
 type ZipReader struct {
-	r     *zip.Reader
-	p     string
+	r *zip.Reader
+	p string
 
 	// limit is the maximum size of a file that can be read directly
 	// from the zip archive. Files larger than this size will be extracted

convert/reader.go

@@ -35,7 +35,9 @@ const (
 )
 
 func (t tensorBase) Kind() uint32 {
-	if strings.HasSuffix(t.name, ".block_sparse_moe.gate.weight") {
+	if strings.HasSuffix(t.name, ".ffn_gate_inp.weight") ||
+		t.name == "token_types.weight" {
+		// these tensors are always F32
 		return 0
 	}
 
@@ -55,13 +57,15 @@ func (t *tensorBase) SetRepacker(fn repacker) {
 
 type repacker func(string, []float32, []uint64) ([]float32, error)
 
-func parseTensors(fsys fs.FS) ([]Tensor, error) {
+func parseTensors(fsys fs.FS, replacer *strings.Replacer) ([]Tensor, error) {
 	patterns := []struct {
 		Pattern string
-		Func    func(fs.FS, ...string) ([]Tensor, error)
+		Func    func(fs.FS, *strings.Replacer, ...string) ([]Tensor, error)
 	}{
 		{"model-*-of-*.safetensors", parseSafetensors},
 		{"model.safetensors", parseSafetensors},
+		{"adapters.safetensors", parseSafetensors},
+		{"adapter_model.safetensors", parseSafetensors},
 		{"pytorch_model-*-of-*.bin", parseTorch},
 		{"pytorch_model.bin", parseTorch},
 		{"consolidated.*.pth", parseTorch},
@@ -74,7 +78,7 @@ func parseTensors(fsys fs.FS) ([]Tensor, error) {
 		}
 
 		if len(matches) > 0 {
-			return pattern.Func(fsys, matches...)
+			return pattern.Func(fsys, replacer, matches...)
 		}
 	}
 

convert/reader_safetensors.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"io/fs"
 	"slices"
+	"strings"
 
 	"github.com/d4l3k/go-bfloat16"
 	"github.com/x448/float16"
@@ -20,7 +21,7 @@ type safetensorMetadata struct {
 	Offsets []int64  `json:"data_offsets"`
 }
 
-func parseSafetensors(fsys fs.FS, ps ...string) ([]Tensor, error) {
+func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]Tensor, error) {
 	var ts []Tensor
 	for _, p := range ps {
 		f, err := fsys.Open(p)
@@ -56,7 +57,7 @@ func parseSafetensors(fsys fs.FS, ps ...string) ([]Tensor, error) {
 					offset: safetensorsPad(n, value.Offsets[0]),
 					size:   safetensorsPad(n, value.Offsets[1]) - safetensorsPad(n, value.Offsets[0]),
 					tensorBase: &tensorBase{
-						name:  key,
+						name:  replacer.Replace(key),
 						shape: value.Shape,
 					},
 				})
@@ -111,8 +112,9 @@ func (st safetensor) WriteTo(w io.Writer) (int64, error) {
 			return 0, err
 		}
 
-		for _, b := range u16s {
-			f32s = append(f32s, float16.Frombits(b).Float32())
+		f32s = make([]float32, len(u16s))
+		for i := range u16s {
+			f32s[i] = float16.Frombits(u16s[i]).Float32()
 		}
 
 	case "BF16":

convert/reader_torch.go

@@ -3,12 +3,13 @@ package convert
 import (
 	"io"
 	"io/fs"
+	"strings"
 
 	"github.com/nlpodyssey/gopickle/pytorch"
 	"github.com/nlpodyssey/gopickle/types"
 )
 
-func parseTorch(fsys fs.FS, ps ...string) ([]Tensor, error) {
+func parseTorch(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]Tensor, error) {
 	var ts []Tensor
 	for _, p := range ps {
 		pt, err := pytorch.Load(p)
@@ -27,7 +28,7 @@ func parseTorch(fsys fs.FS, ps ...string) ([]Tensor, error) {
 			ts = append(ts, torch{
 				storage: t.(*pytorch.Tensor).Source,
 				tensorBase: &tensorBase{
-					name:  k.(string),
+					name:  replacer.Replace(k.(string)),
 					shape: shape,
 				},
 			})

convert/testdata/Meta-Llama-3.1-8B-Instruct.json

@@ -0,0 +1,3 @@
+{
+  "rope_freqs.weight": "80fd5efb2f729381785b293a091a268cfeceb0079167f6ece9b07070e662b222"
+}

convert/testdata/Phi-3-mini-128k-instruct.json

@@ -0,0 +1,225 @@
+{
+  "general.architecture": "phi3",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "phi3.block_count": "32",
+  "phi3.context_length": "131072",
+  "phi3.embedding_length": "3072",
+  "phi3.feed_forward_length": "8192",
+  "phi3.rope.scaling.original_context_length": "4096",
+  "phi3.rope.dimension_count": "96",
+  "phi3.rope.freq_base": "10000",
+  "phi3.rope.scaling.attn_factor": "1.1902381",
+  "phi3.attention.head_count": "32",
+  "phi3.attention.head_count_kv": "32",
+  "phi3.attention.layer_norm_rms_epsilon": "1e-05",
+  "phi3.attention.sliding_window": "262144",
+  "tokenizer.ggml.model": "llama",
+  "tokenizer.ggml.pre": "default",
+  "tokenizer.ggml.add_bos_token": "false",
+  "tokenizer.ggml.add_eos_token": "false",
+  "tokenizer.ggml.bos_token_id": "1",
+  "tokenizer.ggml.eos_token_id": "32000",
+  "tokenizer.ggml.unknown_token_id": "0",
+  "tokenizer.ggml.padding_token_id": "32000",
+  "tokenizer.ggml.scores": "6e37bcde2adc7e350e87c496eddd7a2124329c1dc66c5bf3ad3997253e4f7a62",
+  "tokenizer.ggml.token_type": "b6ecf55ec64ee67d87750bdb8d757a2c58bf78377e9f4219f5689a6c4dea57ce",
+  "tokenizer.ggml.tokens": "d168da3ddd3eee820916945fcb9baf24dd3cde42f606cffa2d19e7c8a8743918",
+  "blk.0.attn_norm.weight": "216aeb2c9e0c271f899e1ef2a63cceeb8f41e97642e84fada54b1d3c1c11cf25",
+  "blk.0.attn_output.weight": "b597d56f7188ffc1fafc273fadc59d41738cffd677ae98c61a62c3285b3a3099",
+  "blk.0.attn_qkv.weight": "d28a6b44e13f59be5483e4be2bedb544e346168d720aca27f47d1a5a722be91e",
+  "blk.0.ffn_down.weight": "4a691370e5a61fcbbf540fbcbf4c0f1d15dec0364528c0e916d0744f6262b63b",
+  "blk.0.ffn_norm.weight": "0c00af2b4a3128bec64a0cbb1084b042fdbe13d9ad0d03bd577f9449dfead338",
+  "blk.0.ffn_up.weight": "b32b52f790c1c083bfb8a3126dc1111cfeeb28dc8c584a930a1e5334cb176bf4",
+  "blk.1.attn_norm.weight": "68748011503c6c029e8e69a84a8e5a89338f378769627b6dbf7f93d715c292e1",
+  "blk.1.attn_output.weight": "2267344add13b048ca59e4377c86dc512be8046a57156901fa32a20fa74e4ee0",
+  "blk.1.attn_qkv.weight": "9109d2e3d7a2eacfda5226587b8be124a3bf44b972da7ebb17aa15795897eacc",
+  "blk.1.ffn_down.weight": "d675df4df4dd039c0c339ad6445d39eddd2004db6bf35bed6314c7497245a633",
+  "blk.1.ffn_norm.weight": "3b5767ae977bc8baaa06b06efdbea193b6b3ba605ce76d77a76ce317e935500c",
+  "blk.1.ffn_up.weight": "80dfd6d9d234b00334c89b8e0a02f81899c2efd377321c34ba5ba51a5f61b5ff",
+  "blk.2.attn_norm.weight": "6a6743b057e5088f145bc179e92c9bfb41163e7295d7b81c62e23dd89d2b59c4",
+  "blk.2.attn_output.weight": "bc5491ea54e0db81462d7d9b7d25cbdda380c2db8de041bd1c4ab7b76a1d19c3",
+  "blk.2.attn_qkv.weight": "a61287a9852e2f5aca9c100b471d98398b2913a3497c743de3c70ec9ddd7087f",
+  "blk.2.ffn_down.weight": "4fddcc382c8dceeab027fe43d8d44e67edb5e8ce4b9a1b7f773c87770380ade1",
+  "blk.2.ffn_norm.weight": "07e05f82b3f63f711db3b684ca79aed25c0657917e66f88af47348a82065c227",
+  "blk.2.ffn_up.weight": "4835a682ef1826c12df01ae7663fc45f9c82bc8e64b665f13fb7da8e201ec0fb",
+  "blk.3.attn_norm.weight": "f22aba7c03999ba7136f39cda747a39715e498699dc1716cd97fc5dfc58d1b1c",
+  "blk.3.attn_output.weight": "53b579855366fd786c5126b2b30aac4d583ca7bda56833c4865f5cadb5c18c6d",
+  "blk.3.attn_qkv.weight": "bb56aba78158123140fcea59c69ac562ca208f6d3086819417cdad8c50f333ad",
+  "blk.3.ffn_down.weight": "97280897a7cd86db2830c004bccc5bc094f50e293baded0189159a2019145a6e",
+  "blk.3.ffn_norm.weight": "10a8c99f8b57a960e8e0a1133c4a26f9148403d1b9bff2eff114917de996f3b5",
+  "blk.3.ffn_up.weight": "7324046c915e75d621b2043597a245a428d8eea31869135e6257a861491d8dcc",
+  "blk.4.attn_norm.weight": "507d8e164de94646edbfe33def8e8fbf7c9a6ee3fbaedb5000f72d9f51ec5e36",
+  "blk.4.attn_output.weight": "bbb3429e6efa98c150e0fdbf48c16180cbf0d0cbc1b3c253c6c319d78f4593a2",
+  "blk.4.attn_qkv.weight": "b95ee5be0786d3901273d806c339fe6c20e6bfffd2a20672a9f56af80921e8ab",
+  "blk.4.ffn_down.weight": "806bbf91df92a5a22bd5aa1ffb7fc2869f7293ffc7704771c290ecc583b27975",
+  "blk.4.ffn_norm.weight": "cfc2930a81df7aee3a5e7f726a15c1182233e868bf0d9d37f6b6ae6d8c15c234",
+  "blk.4.ffn_up.weight": "c3390c69533de2c8424e8069323ccc5d0c4543111535da04cf2c7d26745576aa",
+  "blk.5.attn_norm.weight": "0d71c4fbcefabbd021569442853d2fe90668b19409ae2805a718a829ca60beab",
+  "blk.5.attn_output.weight": "10ebd93629112bf2df5c30dd0953a4a5e9020306768283181ed426934d47e14f",
+  "blk.5.attn_qkv.weight": "5cb05633369f12d4b00e0ff787736bd846856682115720ebc6cce05270c334f6",
+  "blk.5.ffn_down.weight": "e28bcc5094212eafc7476dbc5b7a520d25b79578cbf4229d698e2655956a80ad",
+  "blk.5.ffn_norm.weight": "b6f2c4cf9f34bb4d59989f96165c14a67dc1e266ad0a6d0fcc49f1add929e6ff",
+  "blk.5.ffn_up.weight": "0f9ef99423cc07ebedc0e9cfa95809f2d7108d910bb4ef97ebc0b0309c440750",
+  "blk.6.attn_norm.weight": "b3edcc47a42218234f7564d7470611b49401a41ae8cd42123f86557c69f5d7f2",
+  "blk.6.attn_output.weight": "eb9b7d257b388bb5b8fe0515e5c6873317239cb94cda236e4b6ada2a6c57c65c",
+  "blk.6.attn_qkv.weight": "eb968081f478c52f07bd9c2761741e982dba33cc4eeadeea3557d391b9ac2106",
+  "blk.6.ffn_down.weight": "1b8588bb7463206290322695577dcfced300895d6e6f4b26966c53a9ae2f0f84",
+  "blk.6.ffn_norm.weight": "1219c04b7770983c77814200eefe743f46d15328ea2b12711e44f8103eab08d3",
+  "blk.6.ffn_up.weight": "197ef287239fec47c55677f0fbb66eaf0644f775bc382de843971730721394f6",
+  "blk.7.attn_norm.weight": "b630ad08c80d564ed1c024384818e9fd3f22a36cd7a14aa96e7e2759a8285099",
+  "blk.7.attn_output.weight": "970255aa750828a47d6b9d399f9612b5bf25aefe7dadbcba41fc416d0d4067c1",
+  "blk.7.attn_qkv.weight": "ebb157c880293e6de8d629f263ba8853ed1dbdc02c311d43432bb8cfbb310739",
+  "blk.7.ffn_down.weight": "24bcd4db4cba844c89f878b81843c373dbbc0675e889d32c5b12e63384a7b670",
+  "blk.7.ffn_norm.weight": "b9c6f71001808ee873ce7db8056e4b53fb4cccec8b7f0f312899b575fae39d39",
+  "blk.7.ffn_up.weight": "979f1828d227455c26015a2a11afe9dd05f2bb97a8ba6b38c8dab3f50e627401",
+  "blk.8.attn_norm.weight": "4e8e347e3775010b7112ee630f2f4f2383be7ff64e6ca6154b9b22566552eaa6",
+  "blk.8.attn_output.weight": "65a44babf44a435a1829945211b3168f9ec78ac3cb7a049a733e93d11f0d6659",
+  "blk.8.attn_qkv.weight": "343ed07671da400b040812a4058482fa38284b5d9af9becfed07417fe26ce747",
+  "blk.8.ffn_down.weight": "7fb7e073e3c2c503c4e9d60efa0988fed7398d900cc003695fe3fffd3e188b82",
+  "blk.8.ffn_norm.weight": "b07c1f655d8593e3892a2cf73f8a0c19ce8e5cb613fafbe7cbd430da8ce4c57d",
+  "blk.8.ffn_up.weight": "8b26e14de54b3fdc2e2d3ea41720f9d9c236a93688c3b7fd7bf43f5fbb327c9b",
+  "blk.9.attn_norm.weight": "46394d408a8e316916177e6aa261de32e137a82d729c0b1800b072f0c38c39b6",
+  "blk.9.attn_output.weight": "d57f3d46107947a7073373a0b35d6ecf7759b5df15406f4a3590a60666af6b16",
+  "blk.9.attn_qkv.weight": "14bb8ace8c5453148f4b536e9f4279c813f31136716947256f5cca333448639c",
+  "blk.9.ffn_down.weight": "2b8d98e2b5ed68338f6e4de43bf7de0c4858cc69103cd5177725f7444eec7694",
+  "blk.9.ffn_norm.weight": "41a499dfd418cc4c6b8c12313f673f7e2cd4a3f9c4065eb6c4feb5eed02fb542",
+  "blk.9.ffn_up.weight": "143aab7533a64b17fbe201490a6f674bc7f0bd370c094500b2e100419073d1c2",
+  "blk.10.attn_norm.weight": "ebb670aafd36816a794347287269d8f1a5b19c1e3c0a1e38023bc19fdba9b073",
+  "blk.10.attn_output.weight": "b5d65bbc0ed5e49fdd9d754bc18163cd042a285024d0cf6f954c503bc8c877cb",
+  "blk.10.attn_qkv.weight": "f06b15bac88da798fa34a62b03eaac0dbe8b846020516603c387541f2d8dd672",
+  "blk.10.ffn_down.weight": "fb091fcd1b4de25d1bea94d1755e255cb02914a030d23e3a234e57b8d46bde6e",
+  "blk.10.ffn_norm.weight": "eb347bdf9c40414af87e13a8e72e40b31f004b50f7cb366f1a219ced60a61355",
+  "blk.10.ffn_up.weight": "ed2d52fc881a173f404fe8a1067862c9856d6c3e0d2e90a330a7aa394e3f84d1",
+  "blk.11.attn_norm.weight": "64e252603cf010a0e502ca39fdf8d0a196a79aec67c0d2bb9213fc0cb80c47d4",
+  "blk.11.attn_output.weight": "228e33e21c69f52efc74fdfc831bc9af271e44b2a29a3dced1d64e667ce36eb5",
+  "blk.11.attn_qkv.weight": "ab9ce6d4ef9e42ee0da3f20a7708a3bbc5e79e967b05fa86ba946a05e2eb63eb",
+  "blk.11.ffn_down.weight": "0ca133b7835c98dc77c25d64e4eb7873778bdb5e4d22d8b80f920f46865b43bd",
+  "blk.11.ffn_norm.weight": "02455741a0dfd161c79aa1ecc381901721f229fdcda5615622a629631fb61cfd",
+  "blk.11.ffn_up.weight": "9fecdcc099fbb8e23c6b1ea9294702a027f4a58d265543ec5e7be79b8f63b354",
+  "blk.12.attn_norm.weight": "783bb459911b1b3609a9b2bdfe272f1670add73b5471da738e07ac47e2e07dfd",
+  "blk.12.attn_output.weight": "1e1a914c9e48b857206ac5a1f7cead994bc1ea91d5d4fff8c834d73f2e38ef5d",
+  "blk.12.attn_qkv.weight": "5953e7185ccb87fb4dae8f9426ec86315d4c7794326e8ab59b3a95d4af2189f0",
+  "blk.12.ffn_down.weight": "a3eecf0f394f86e2cfb48a5940a5c50ca86d71883b2f79fcc642a935fabce0d4",
+  "blk.12.ffn_norm.weight": "0a4272e41373c23bd72f10d2d82930aa3a1480aac75832bfbf01cebf0b86b6a4",
+  "blk.12.ffn_up.weight": "06f42776de3a7ceac3025f26a7a8bd20e062233cce2bdaa2183470dc4b30b87d",
+  "blk.13.attn_norm.weight": "5915da60fb03e201fa649faba780e5fdf1c761c262b206e5415cf83181f65780",
+  "blk.13.attn_output.weight": "4dbf6eab074fa3835fd32bd631a8208e511037d5056d2fd3015735cca7674ef7",
+  "blk.13.attn_qkv.weight": "d3d8339a1c4782d9e73d77fdebe154d3c5b83ac40c9175b3e91a4977d08f876b",
+  "blk.13.ffn_down.weight": "de6772b46a55e1fd42b007637dfbf68b6598e5d5b61622da0935002e1e192d3a",
+  "blk.13.ffn_norm.weight": "5a640ea3b8c7be49c95a58a2327e10d8e8d9d142504bde5c8091613e5b961d7a",
+  "blk.13.ffn_up.weight": "f35e3545e4bd3531b2e843b5efd31dee0c13c807ee6386e65473ba67bbec30d0",
+  "blk.14.attn_norm.weight": "9b34986450b7c98b4927e81e61a816f9e84b1addc7c14926402100037aad6678",
+  "blk.14.attn_output.weight": "155d52efb23d366016d861a251d4d1f4a0c13699188c50d50dba016a0d8bfcd9",
+  "blk.14.attn_qkv.weight": "8e1415084e1f33c73a777f19e752489f4dd312cca047733e5ea643cd4a955e04",
+  "blk.14.ffn_down.weight": "a2a142226b94baa01ccb65bdea2b7418e49085c1d9c3c63e544e3112c58a25da",
+  "blk.14.ffn_norm.weight": "8aecfd9b0ae6affaea31a80c5c9a4a14b31deaa0db7bd8f6da2a64d23447921c",
+  "blk.14.ffn_up.weight": "0c1407237b8c1bd02f193346b5681926fe698a5055eac6a7450451b0f991707c",
+  "blk.15.attn_norm.weight": "e037bd19880bfa83d983200fb0c7866f8ad16c3ff5cc4b4f3a37ca7373870ff6",
+  "blk.15.attn_output.weight": "045fe4fc95cc129a1b92771b179c11b12845c4c088786c607f17bd98857e68e1",
+  "blk.15.attn_qkv.weight": "7621b7559705cab1d4dea1c69f76dbf9dc1c8837a203b656f484703b9c1b70ce",
+  "blk.15.ffn_down.weight": "7e5ac20e290bc60761e1cd972354fde225b7fa861048d44d9a0dd9b046d55f58",
+  "blk.15.ffn_norm.weight": "b6d830d88f1db1825687973c8c2b1a24c6fa84f07af8d0e3ef9c86009baca0b2",
+  "blk.15.ffn_up.weight": "dcda0957cd04fc45476774dba2bbf9aa89d6b05d5ca7b10ae6f73ad2c49b1cd3",
+  "blk.16.attn_norm.weight": "4ee9b70ba15cb2a08240f93990e90f5068c48fceb481f8e2186bec8b7214eb3f",
+  "blk.16.attn_output.weight": "315cfe5536658d2498192b2980eade15b2c9a4ff220e4011911457b1727fa103",
+  "blk.16.attn_qkv.weight": "3c8122e3ad637583b9dcde8ff3a323267d3014bb1f0f9771e5322260ca9ecc8d",
+  "blk.16.ffn_down.weight": "3b5fbebd5ee2b86cad96fb8a9b45a8770d08f82c1c8b74d7061e866f7020a18d",
+  "blk.16.ffn_norm.weight": "ffab69f20bda372de6e5878f0539163e2fc6ba113621ded95705fc3b1465c9f0",
+  "blk.16.ffn_up.weight": "0935ea3d258da42d6258406365f39f58ddaabfe97ea5977580db3635188f24a1",
+  "blk.17.attn_norm.weight": "f030441733f3d147b4a06a1eb4aeb8465c7c24d9c53bf4c48fe7e134d3629803",
+  "blk.17.attn_output.weight": "07a955ef09e8dc766ac0df647d0b2c69f23c4c69a7137654b4aad80303ed0eda",
+  "blk.17.attn_qkv.weight": "1c10688061e21e2fe12ad0cb54bf03895c1f83c3b0df743a42f548b52cbca1b2",
+  "blk.17.ffn_down.weight": "ebb9cc9836f41d88fdae2aa9a4355514e4edaec8d1577ffeb947a35204e77f52",
+  "blk.17.ffn_norm.weight": "50aff44f6528b13db5389f2ddcdb7676244947610bd7ffbff3f881c968c2a0d4",
+  "blk.17.ffn_up.weight": "d716537949582be33bde6b02e38f5a70081c9642a9fb05a61312126718b8d148",
+  "blk.18.attn_norm.weight": "0ea695c4e53d637902f46663a6ee42adc493c36794476acc7dbddaa05b13840d",
+  "blk.18.attn_output.weight": "5fd35b500221a612eb4f4bddf0e9b6b7db4d7733032a75f8802fb2d884647c2e",
+  "blk.18.attn_qkv.weight": "b0da37fd030fe69581f990bf23bfd35467a1bbe558af6de7c0924f6b72e92317",
+  "blk.18.ffn_down.weight": "b355c33f44b328f4bb977567de8f7544db4b005d7a8fbded658518ecf3c5a153",
+  "blk.18.ffn_norm.weight": "58b3fe9094079989a86e0387143259e1cc35952d24dc3df290c4ba6df44f5c51",
+  "blk.18.ffn_up.weight": "2ce530954c342c30ed2ead5353f931960bfae1d278868504c0efb973560fabbe",
+  "blk.19.attn_norm.weight": "533e9aed66feea8f0392aa81f9e293240e1f009a5334253915fb60c2749b615d",
+  "blk.19.attn_output.weight": "84f2d00f98a4113a779d3b5d1c3e7c914eb47784d3ab13b290367c124c2994aa",
+  "blk.19.attn_qkv.weight": "fbe6b9f53b07fa7537d3b3d452d20a9bc666f9fd41ec2091dd28bc2f70fc668f",
+  "blk.19.ffn_down.weight": "b30199e098c8bb3f890183d8b18471e80b62b604729b277ad62488dd71e1206b",
+  "blk.19.ffn_norm.weight": "c81373e41cd340b7badb19f9517c77c4250b4eb9a02dc758b8b49b652487d7ff",
+  "blk.19.ffn_up.weight": "5a5cb083ca7725720e3a890f7fa46354760e8007a8188849a092e305694a75e3",
+  "blk.20.attn_norm.weight": "4953091b4477e354357a8e743ba0a1900633e52f1599ee082a0c9b0b2b5cd978",
+  "blk.20.attn_output.weight": "62d54f7749cd6856097b2632066a322b0296df915fe66f382c5b5981be0d4f23",
+  "blk.20.attn_qkv.weight": "406de9e35b0729ebe902d7a47905cc7fb29a921431ed35dbef0c03e5690a1329",
+  "blk.20.ffn_down.weight": "62fb678b0d1261e19a4903a2b347d67afcc8acff01feb33a687a35a2d1e6f9a5",
+  "blk.20.ffn_norm.weight": "cd9d36b7e71e55c8925b97bb09c28219f182626bcff094878ae39c3db887a14b",
+  "blk.20.ffn_up.weight": "b9276771d79d3e932e73ccc520c3f8476342b9ef312ed2ee1e0da822e6e3ad18",
+  "blk.21.attn_norm.weight": "66d8c8a35e13ce9c2a0e75b670150e2c31484a55c2316df46075312196178ed3",
+  "blk.21.attn_output.weight": "12ab46c9382648f9b3350fdd92a6be6352743d62d6b520d7e2024e0c838588f5",
+  "blk.21.attn_qkv.weight": "a7909676ee1675ca23cd29a5fdd226df8dd9d68f94c6c9bbb51dd9fd38504008",
+  "blk.21.ffn_down.weight": "6fb317279c6542e82f97d5a12a60fac1bd0fa0405154f9fbe265e2fe39bd49cc",
+  "blk.21.ffn_norm.weight": "c0f703eb3ff161b5ba4490d87d8684b8a6c47a8f433e12f418333b9db439010a",
+  "blk.21.ffn_up.weight": "6dbdb80ef0c35e364bbce12d40d5e74c7963c7b55d58d9579567a07ffce7b863",
+  "blk.22.attn_norm.weight": "f94237433bf03d675cb2f655b81ca91a1ce2447bc6b00b13d6b0ccfe2d411eff",
+  "blk.22.attn_output.weight": "e821f95995ce497c01e63ca64f737713b1b65f11df1903e51d444aa516f33f71",
+  "blk.22.attn_qkv.weight": "1b0f717c73afb5eb4c82a1708c4e85c969e8a2a8770d9ddb78b1870a2d8a781e",
+  "blk.22.ffn_down.weight": "0f33f7a3cdc685484be99aa0c03642b0b20850a27d1fddbe054b13a9382f3ccb",
+  "blk.22.ffn_norm.weight": "9df285cf211ddd7df2b36a50489af574755c7d4d98b29a05cd04566ae613c8dc",
+  "blk.22.ffn_up.weight": "63ac300e1efb34041dd0136cf43ea622fac6f0caccce1cd9262f5e08d2cf179c",
+  "blk.23.attn_norm.weight": "5f72d9e88689b4027b28f5f8f26cd3abb03635ceea7ec98a4c91a9fc691f6707",
+  "blk.23.attn_output.weight": "6ecf04ff61125c5fc768f8656497152149373daf321ee9c957e8f7245a1184d1",
+  "blk.23.attn_qkv.weight": "a9d9978806724c2959f2cf386c233831f08e1e933dbf2b32665e788d9d512ea4",
+  "blk.23.ffn_down.weight": "72c7d17886a3da17fa0daa456aa5e877b2ef5b8b403182b870d9ca5ca9c70347",
+  "blk.23.ffn_norm.weight": "971e4b712e3025a13419b5b57d674b5e4ab7f18f74b57b9afc4671623da90c4b",
+  "blk.23.ffn_up.weight": "df2b5c7dbd5834545b815073af0c7355b065124e6d6f0fee78d8fa5b2076dc3e",
+  "blk.24.attn_norm.weight": "c41957c4a79ad3b16f6e11daec1c7f530b9f3f4b618e1e4367c3b67787ac4ab6",
+  "blk.24.attn_output.weight": "ef7d61f5fc88ac6f31bf60cb5f4d2d6b8df42d38825807112361a7224b0dee3b",
+  "blk.24.attn_qkv.weight": "3e6a58fe7d49c90bb6971efbad3371c32256881173ea5aee4b0c296cb206490f",
+  "blk.24.ffn_down.weight": "f43619144047de42fed81dfa495f1815d3cb771330e574043e2b67620819292c",
+  "blk.24.ffn_norm.weight": "5501d4a2a98c8ca6b42e77b53b221dbc08f530f6a067256d787534ec6fe028bd",
+  "blk.24.ffn_up.weight": "d64c8b0e509e2b1118f6000176f8956cacecdbb200c7e95ed93fb78b6e26c84a",
+  "blk.25.attn_norm.weight": "502fa3c302d371f61c5791f4615b73018ffb1daa09b6499b227116581244c5d4",
+  "blk.25.attn_output.weight": "ad8391d4e9c980856f2547aa945b2b6a407a6382158dc1ddd4f08d94ecc24be6",
+  "blk.25.attn_qkv.weight": "42e8983780d4a01a02c54ad23d4df21eea437f119a10af5a9c12a76a42d308c1",
+  "blk.25.ffn_down.weight": "302dd010d4e0ab4eeaee89090409ea0dddeeeed3236415eb8f97c942497eea91",
+  "blk.25.ffn_norm.weight": "fb34c1ee5bca96986c08834df0a0c047ba041c1123ac1f563e9d64312bf82d6a",
+  "blk.25.ffn_up.weight": "10739a8de156816d93c92b935386540bfa976bdbef204f0312960f6fc657582f",
+  "blk.26.attn_norm.weight": "7036c711609128c4e55968ff3681d3043338879a5737efd6c2ac9e1a2a61f1a0",
+  "blk.26.attn_output.weight": "db5db45dead5cb911fa01da59832f121b7c18b2d167bf53741c40819f24d346c",
+  "blk.26.attn_qkv.weight": "cae34c6b7f82ed14348d5ed30a79919c383737c1694a9cb9c0de609d3b0c1d0a",
+  "blk.26.ffn_down.weight": "491ec3a4da9b4f49f8ebc6be658ce397a9b801ae9fb35e82177e47808c65e5d0",
+  "blk.26.ffn_norm.weight": "fd7059d75d7f0e5288511ddeeb0f772eb3cae3ccfe4226b877015834edc3c386",
+  "blk.26.ffn_up.weight": "ea1ee1274c56458ce056d2205e5bb6e5422ce4cb0ad58006b8141749b97a0c39",
+  "blk.27.attn_norm.weight": "cc362c9a937609265052cd38544af17a1a7448cea086d4c801139e1fc865832d",
+  "blk.27.attn_output.weight": "ba757a81dabde9cb1b069d1bb616fe79649a1724f756567ec61caed1304fe6cf",
+  "blk.27.attn_qkv.weight": "1ab8d7d02d87756c12c2275636823aa5ede3d683178225c4cac4bd892c319bd4",
+  "blk.27.ffn_down.weight": "deb1c711c8a66acf4dcd2d088e1548f8e08f296f755e4067d6557fa55afde88c",
+  "blk.27.ffn_norm.weight": "fc6242d8cb8a4a37a8ddb7e41e7e60a63d4a89edf36acb35df052f10b9c91ece",
+  "blk.27.ffn_up.weight": "8df39b09c4801f343aca78f2918a1f6db78c8c55e591eda4c69eadb74c26e180",
+  "blk.28.attn_norm.weight": "75b539308f77e3cefdc6d98484d8b5cbf0538f0c2869a77b7373a145a18bc850",
+  "blk.28.attn_output.weight": "ae128940eb60a6d2e121762ef4b3e9dcf9eb3e105b249507fa7f12de0e19822c",
+  "blk.28.attn_qkv.weight": "bdda781c288e9326c240e33905f8e621b6a2ad902e620739d34f93fcd6f933de",
+  "blk.28.ffn_down.weight": "f1d6e6d1c286b1138bfd7e53fe477f399ae93bc2c04e35416f84218ed7247965",
+  "blk.28.ffn_norm.weight": "3f837ce82c8b9bde0d61d08b6f5fe5574886ea5328dbdc53f2929f18da8b4087",
+  "blk.28.ffn_up.weight": "2af027002e31d1b6cfedbdb30a2b9d7213f3aa691167c353913adfd48fda31e4",
+  "blk.29.attn_norm.weight": "61e8003b5329462ffe0fe172f2b160260de006aed858332d49d75504b6b6aa7a",
+  "blk.29.attn_output.weight": "ca44542a72a37476dc73dbdcc01f5b7497cb3ebc4ea230a55c9634ccd8e56ad4",
+  "blk.29.attn_qkv.weight": "abb3d9d6abe57872ae3daa51935d43264093ded5ce63b49d1e280ee5758be0e4",
+  "blk.29.ffn_down.weight": "6764b895fce881df097489c263446f0106de36217997660c15984b3ee22a5a06",
+  "blk.29.ffn_norm.weight": "89e03e9a33fc0e6e31ba9f0c2bd7c5734a118c5602bb90148793e08a80e8d0ae",
+  "blk.29.ffn_up.weight": "fa7ad57a84954f4121653152efed1a871d8adb20a1ea9086e3e849ce359d7d2e",
+  "blk.30.attn_norm.weight": "91a697aca1e42af54f806a20211031c3369e8d0bd58df1b0147fe24954e1f5a4",
+  "blk.30.attn_output.weight": "36063fcf766c89ac75be56f688cc63cefe5f2c733fbf4378ea9956ad386fa148",
+  "blk.30.attn_qkv.weight": "2cacd1161f1121a2c0b979930134f4666f73fb8d7237b3b0659ae091b15955a6",
+  "blk.30.ffn_down.weight": "9f3fcb6217100595850c05dc98f9ab2a263afdb6ab28df2fcb08aeff512057d7",
+  "blk.30.ffn_norm.weight": "6c600bc1fc7de39d4f8917b81fc7d1d5ed2a9b56492234c13a4bd6028c30d880",
+  "blk.30.ffn_up.weight": "73cabd1bb011956b2689ea3338bb76642ef3a57c197377d666d2ab5f56317668",
+  "blk.31.attn_norm.weight": "72d3e1cc771380645fa75a899858c95f39857a4f3f1ed60fe1578df383b8bc53",
+  "blk.31.attn_output.weight": "40089cdd29994dc19a1d89fa15902a89cfeca3540f12dc9bf4d00ef82506e456",
+  "blk.31.attn_qkv.weight": "1d0bb40e9258071ae14290a53c619a8e331dda07354d2a02ef45766c029ae5e4",
+  "blk.31.ffn_down.weight": "8defa0e06335b793fa8be03883f0a322d6c5b33f52c69c943c35c60d16e42c0a",
+  "blk.31.ffn_norm.weight": "33c55d9d0c496ccfb130361fe131649346e098abaaac39c0519507e5d846721d",
+  "blk.31.ffn_up.weight": "599f6503f61c692c1f82001973d35119f9688db5e6be9d9c298411491c93f09b",
+  "output.weight": "14b8dc662bfa3308ebb2e102c562d8e52c15670e538f20f3216a9c310ca9dd41",
+  "output_norm.weight": "7f2294ba94ce65681df6c7ddd8698799199b9d77dc83c10bdad5c3999f0fdb82",
+  "rope_factors_long.weight": "e34d378664e354652c38f47d10dafb0498ccc2fb042d39ff7fef768146fff22b",
+  "rope_factors_short.weight": "9379146a4988f373d362fe47b06c75e7fe7c54aa4dc9558758df79b7a87471fd",
+  "token_embd.weight": "19a03c1fb5ac0baee93b0a7d8b0f26e9a9b011e229b694afc50ebfc13d84f8bf"
+}

convert/testdata/all-MiniLM-L6-v2.json

@@ -0,0 +1,124 @@
+{
+  "general.architecture": "bert",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "bert.attention.causal": "false",
+  "bert.attention.head_count": "12",
+  "bert.attention.layer_norm_epsilon": "1e-12",
+  "bert.block_count": "6",
+  "bert.context_length": "512",
+  "bert.embedding_length": "384",
+  "bert.feed_forward_length": "1536",
+  "bert.pooling_type": "1",
+  "tokenizer.ggml.model": "bert",
+  "tokenizer.ggml.padding_token_id": "0",
+  "tokenizer.ggml.unknown_token_id": "100",
+  "tokenizer.ggml.cls_token_id": "101",
+  "tokenizer.ggml.seperator_token_id": "102",
+  "tokenizer.ggml.mask_token_id": "103",
+  "tokenizer.ggml.token_type_count": "2",
+  "tokenizer.ggml.scores": "6db964fe67338aca57790481a390121ff3dd643eebe49f7dd308029ad99abb6f",
+  "tokenizer.ggml.token_type": "98d247c5404b6b18f05f133b92dd56edf6efefefac326794b00d7b351f6c5aa1",
+  "tokenizer.ggml.tokens": "9efe405e229a45ff9916f54c475d151d2200cd2ab0006f347abfb069cf096c86",
+  "token_embd.weight": "8c1ee80a9ea4f65aa385ba30112010068af3d209bebc6e149d3d4589c2cd0a5a",
+  "position_embd.weight": "6c516f0b1c4e2388ab90394dd80ad69e4e4509b890982fc3408108ae66210eb6",
+  "token_types.weight": "f879f8e422ed211948f28b560d3c5e17aae7993f063b51196a28cf5c0fb3da21",
+  "token_embd_norm.weight": "75076e095d717aab96f8b6beeee503c27940d9a76f2b891a0e3de72f8a6043e4",
+  "token_embd_norm.bias": "298735285ffe944e1bf03e5d35c7280326b85cf121bde9874f1af5dc51ab939d",
+  "blk.0.attn_q.weight": "ab0923ce4c1549175112dcdfcc860fe30137f991e03ea6857fb5993670adaf6c",
+  "blk.0.attn_q.bias": "a3ec29551dabf976e1d34256b8ab5ab7b758f3ed9742c3cafdbd984d5441df62",
+  "blk.0.attn_k.weight": "4c1038a6d035c3e9ffed7fa672b614627814752503755fbad0cfb76a41ad71ba",
+  "blk.0.attn_k.bias": "e0363930eb588d91816aa3d230bb03b6e2551c165117b80b8d60397413819ef9",
+  "blk.0.attn_v.weight": "425e2e53e3f00ce98d29c3e6a161eb55d3e6ae0d96fdb9f6242d1c4fd6eef4b3",
+  "blk.0.attn_v.bias": "6579173a1e65ee124fbd0bd53cbdca4225515b4f2c5f18fb1bfd000f5978f9bb",
+  "blk.0.attn_output.weight": "a6d70a08cd7164de5d12af65d86d657c3db35aaecde778b2b3fda9193c4c9802",
+  "blk.0.attn_output.bias": "2b8d12c4f9a9c5bfaa29c597839568f6e0525cb41eeaf64ddeb6bd84dfeb9701",
+  "blk.0.attn_output_norm.weight": "bbe6e502a473228b525aeed26cc31b7db123ad63bdc5a6eebac6ea70b8b51d62",
+  "blk.0.attn_output_norm.bias": "36eaacaf0007c5c62daea97aab0115390c0682914f78482e37eb76885f4b7a50",
+  "blk.0.ffn_up.weight": "24654561c76ce387d125759ba843f06b904ef721fcceaeff6ccc62180a48e874",
+  "blk.0.ffn_up.bias": "fd3f0126aa1d95768fa60eb6f4ab8a2763cfcb7e5405f35b92353031d86f4d34",
+  "blk.0.ffn_down.weight": "97a829763a6a5bf3329ceb4d39c424ba4787d61653a5b0bbd1f84782e4d4e0ca",
+  "blk.0.ffn_down.bias": "7aa980c30ae8b4ee7f69df28808dbf5c431f56ccc4a80340f644a0419f16c054",
+  "blk.0.layer_output_norm.weight": "ef30dad4c2a083ae1ff5039a2a6cda60ecc89bf1e486a6f8c0d15f50589603f8",
+  "blk.0.layer_output_norm.bias": "8b1b77e67568b1bce43fc476de1b177c53ff688d66beb66995e8eb3dc290da8a",
+  "blk.1.attn_q.weight": "284331622a1f6f9b87ccee4f652bd66a394ca493c4d93be4d1844e4f6159ad10",
+  "blk.1.attn_q.bias": "e24ebd4860330e08f6bfdd077a82db0bee33f4c8846cf1db26327a34754c7069",
+  "blk.1.attn_k.weight": "729dd0d555544b5bd0f7580b3c8b384256b974605f0e7487b95f295aa032997d",
+  "blk.1.attn_k.bias": "2aa51a828a858f35473f54477583fea54ce2ccc34ea60fbd1d228fbe9bca827f",
+  "blk.1.attn_v.weight": "6be304671cc311d5ca5c103f2b51467ee800c589bc5b8101e09ff5aed1f68c21",
+  "blk.1.attn_v.bias": "43bcbab78a8819e07f723bc9e5b737b71e87a7594f15234e882b63e327a64199",
+  "blk.1.attn_output.weight": "15ec8a1a12b26c9976445308a09f748ab0e4bef0f583d13ab08c3129f8738d73",
+  "blk.1.attn_output.bias": "dac2146f4baa6ed16f6c0dc7443831fb7ec79bedcceafd80d1a4b628a1bb072d",
+  "blk.1.attn_output_norm.weight": "d2151eb33bffac536787a4c9a5d2b31c7a80b17c4611877842a3cce2cd6e98d8",
+  "blk.1.attn_output_norm.bias": "31e1b779716dafb855d2cf5631ee168a0ccf372eb9c6ea6091f66fa97a9b9d2d",
+  "blk.1.ffn_up.weight": "a57547fc3fc3b77406f5cdcb0c87af9bc184701f175c39c1f35297826fce3cc7",
+  "blk.1.ffn_up.bias": "123be6d541d086202913c75d878c54d59a749f3af7b58f7ef9eb9e7c62a24c9a",
+  "blk.1.ffn_down.weight": "cfdb79788377e5cbded8790cd41b9e66c397ecab75474071fcd7cf32d30f9613",
+  "blk.1.ffn_down.bias": "bcb58315519a573097960891c9ae41cf4c685ab78c3e0e77471471758a7eae88",
+  "blk.1.layer_output_norm.weight": "819b554271452bfb1d84c2603b90377b2e41a0ac1e3aa8b417ccf9dce63375bd",
+  "blk.1.layer_output_norm.bias": "47a3433ac27f5ce8947fb38dd491f3706df4ef6adb0ddf74612bf0f54b19e164",
+  "blk.2.attn_q.weight": "1557a9ea852b1880551f7290e00aded4f35e6c4180fdcbed1b0039bf805f639e",
+  "blk.2.attn_q.bias": "c3bfe5f3066f655fd36b055530997b59ff33ef013563aaeb3cb8ff07dabd59a9",
+  "blk.2.attn_k.weight": "cfd08eb69c61ae2f9f14f9b7ff5c5394ca264b1a9f3d48156677f90dd1766289",
+  "blk.2.attn_k.bias": "9b839bc0e79974a0b3f5d1895972bc6f5c9a1bc16052e1af786e6a530758152d",
+  "blk.2.attn_v.weight": "02b26b1208480eaeeb00e7b4cf8b690006ca14759357fc44ed4a2a8924ead993",
+  "blk.2.attn_v.bias": "e7e6f0089fded1659a867ab736c220d9653ea7da6b1b94baf5c8d30a748b63ab",
+  "blk.2.attn_output.weight": "a1db121c7d33806b349cadd050300a57db49fdc91224fd07c9ac43bf4299dc79",
+  "blk.2.attn_output.bias": "7675128b6a92555cd955c820311e91e9417d31f48848f45d047b4100c62148b3",
+  "blk.2.attn_output_norm.weight": "5b4595e0fbcba67a700c4331adf746d2fba3546364a4db5607ae241947bb1a21",
+  "blk.2.attn_output_norm.bias": "7b8e16826ea30e5a2ba0b02e0095a901775981a296e98819625320e983060d08",
+  "blk.2.ffn_up.weight": "a0d815d946ac07a65095c4ae4df77b818845e6d97795c7d82f55e689d944db59",
+  "blk.2.ffn_up.bias": "ce37c0a4174d6bf773ded7bd016ede627ad3bdb8bc99b9992a18dc8e8898f252",
+  "blk.2.ffn_down.weight": "f6231d2a25426fbd45b9f1160aa484220eb227ceef0348c4a6a6de890606e5ef",
+  "blk.2.ffn_down.bias": "429e00556e8dc63a785238b309b9d83738500c1ef6d736fe6526ad88ea496d27",
+  "blk.2.layer_output_norm.weight": "651457a573adf3f7dd9ee5dfe1c8e89389e94443993aab77ec6a0b05aa621e35",
+  "blk.2.layer_output_norm.bias": "41fbbeda7fd89b0cef5f945ae44011c316982390401d6f75ba8c6d365e185247",
+  "blk.3.attn_q.weight": "95a43f32949d2cb8d22815bb27a44abfc6665ba96221af817dfe058cb6ca72c6",
+  "blk.3.attn_q.bias": "f4e34385e75d8108b6b3bd336106e2133a8c9be0cc343dfe5dc48c32a823c7cb",
+  "blk.3.attn_k.weight": "6b892da6a17d4d3265265a15f695864a31813ee8c8e710ae9bc9e1adbc6c9a18",
+  "blk.3.attn_k.bias": "40b8067b641a56014cee42548240aa8930820958b1933004892b5f04fbaef39e",
+  "blk.3.attn_v.weight": "9fcd5922319dd2a461082a5ce040c1dfe65d87d70ca6547dd0b46eeecc3eeb2b",
+  "blk.3.attn_v.bias": "b528c56212e66931fdbe267ac327a9c2f87cd03baff3ea719e30afe681da15f1",
+  "blk.3.attn_output.weight": "e3b178c1b03981e75510e0d277af23ea59cc404b5394e61bd32291825719b502",
+  "blk.3.attn_output.bias": "712c84d39a6a5a9c06a09da8fd9939ba0d5525524a4bba61ea4de09b48f45cae",
+  "blk.3.attn_output_norm.weight": "d1ffac88e675592ff72f8a617be32b4a381d443b2f8f2645dbe44a1e5745aac0",
+  "blk.3.attn_output_norm.bias": "ea31a1c73146234c50e0e43f485c458413714867b8e2703af66482f7db2d6c40",
+  "blk.3.ffn_up.weight": "4ef4f3b9a1ea6ab2ef2eb6e8b008e06a44790d099d97482a05a51e39a29afac0",
+  "blk.3.ffn_up.bias": "06a4296dda16f452675c51f108079fe7722552d6521c737d97734943818b9a2b",
+  "blk.3.ffn_down.weight": "f114b2bebe392c7d80433bb880c6730293aa4561b0b0370dcdaf7472daebd847",
+  "blk.3.ffn_down.bias": "2c8e67831d28a3bf613fc7912ae3259b63d72abcaf4d30efd8800758400158de",
+  "blk.3.layer_output_norm.weight": "a1dfeb7b5a51dd56447312ca41e2ad2f361a3ea12ddc355127f5f4219fb0a482",
+  "blk.3.layer_output_norm.bias": "1ed630021b25c6c6fc93fd32988b9907df966d4982a93081f639aac3044618ab",
+  "blk.4.attn_q.weight": "b5fae4c1f9a5f33a2a2e816ac0c01c25f422e4efdd59ef1ed93da2610e5370fc",
+  "blk.4.attn_q.bias": "c2e376524ea98ac3b10d9eee19ecb1b1e261fa5149efe0232844c923dfb428fb",
+  "blk.4.attn_k.weight": "a4632f5ebf9321d9d08f9112a4e5dda2efe5671df4a4e67fee24845f5b14af16",
+  "blk.4.attn_k.bias": "a9a02ffb8b8b4f6dfe487a7e0341f1d5318c9d2b793a688f34cb1b22fc66ef60",
+  "blk.4.attn_v.weight": "10ad8deb81d9fa093b1e5c0f24ea82aa7df43e6aca49e260fcbea56eab8cc86a",
+  "blk.4.attn_v.bias": "7326813e181e021130bd33ac136293fcffccce2d1d8cb59041e5b13a8cceacf6",
+  "blk.4.attn_output.weight": "c92573088c7437c2b3cda51490e152c27fb19e5468df591eabba5a49d5398d44",
+  "blk.4.attn_output.bias": "14e10b419e5859af1eb685af5c330aee67048cd704dcead9217840c6f5393222",
+  "blk.4.attn_output_norm.weight": "02b6831c0e0fb0edbc579a92812a1dd972cb15d14fcd382d4427c5a7b300ac44",
+  "blk.4.attn_output_norm.bias": "7eed5cd503bb6bb6ceb1bc8b07cc077903a4f14fb8b9d6cdf39644815ecf1374",
+  "blk.4.ffn_up.weight": "8d0c91d62e74d6431321116a37cf3339e630bd50ba164d3304fc4fe8dd831223",
+  "blk.4.ffn_up.bias": "d325f07f73c005a273c484c7be8e7abb4d6e8a5c4fd093f5869133b97629d017",
+  "blk.4.ffn_down.weight": "7ba7bd81143f40537b84f938e403e19f30e4928625eb371de052b9025beb4d21",
+  "blk.4.ffn_down.bias": "2853d9c2a75288214a4bf4907dc19d04d01926f4913d302b1aa7bdbfcce0f7a1",
+  "blk.4.layer_output_norm.weight": "a4ed1885fa77b90fed5300c355ef0aa0c876a8c747151d9d790939d464d57d4f",
+  "blk.4.layer_output_norm.bias": "62142a81e813a9e636333b2b805d6bc3b17c5e7cd4b15adce1ada6bc9a32563c",
+  "blk.5.attn_q.weight": "afc1dff080a72c3daad01384b1448d476aaf789871017c8ff8e144788887995d",
+  "blk.5.attn_q.bias": "748a820371c1d4f872c84545b36358d239c35bf6c99e2812c237d88c3292763b",
+  "blk.5.attn_k.weight": "59e30c1ed8acd2cbb01de5f62e7804015b9ecf98ba157d98cab016344639eda5",
+  "blk.5.attn_k.bias": "f839520078f9e589496e982e86d0126c7aa14196047339abffcf49a696229f77",
+  "blk.5.attn_v.weight": "3e21fb874e21b90308e1f46af034a3c32d3eba1628d62ae5f2246d6af5818923",
+  "blk.5.attn_v.bias": "5cd4852bf95c1444d10d756750f6bf49f842c0b39e9953c7f408bb67c325ac8c",
+  "blk.5.attn_output.weight": "636ce6a7752895f204b9d01ba0aedd9a294f908b42f372c22a16d9dd590d7471",
+  "blk.5.attn_output.bias": "82d924d4b0d2b94f2bbff91619216d6967a3541ce9b1531a6a60457a67b5d219",
+  "blk.5.attn_output_norm.weight": "5e7bd0a8d3396080f3360d7c4700bf094a06216431bd014c4479eef72ecf4271",
+  "blk.5.attn_output_norm.bias": "66c6de5edda5466d029c6753780be81ccd4218bf8bc00680000e0f06856ab712",
+  "blk.5.ffn_up.weight": "5bbf6e7ea380e216e33f8bee06d25f2265359d3876a300e92bc6e41d48e33430",
+  "blk.5.ffn_up.bias": "9d795388bb36fb33ad3a37fea3ccb4937838e02800a608fb47d363cd06b47370",
+  "blk.5.ffn_down.weight": "2fd628974e7f075479dd227b46fbd48ae8d3ca34d735b36f391ac06410730368",
+  "blk.5.ffn_down.bias": "cd213ba9eaa75fa541648097fbe9c96e58077e6c3ad6ad2fb1f21f8350f44291",
+  "blk.5.layer_output_norm.weight": "159a9df41d15b7022d136f86a2a2631c4635f9816e957472217077b522bcf52a",
+  "blk.5.layer_output_norm.bias": "24c1f27ffd1eb4e5be7e3a2909943e6f0980635d761fa1efdd0c19645da23766"
+}

convert/testdata/gemma-2-9b-it.json

@@ -0,0 +1,6 @@
+{
+  "general.architecture": "gemma2",
+  "gemma2.attention.sliding_window": "4096",
+  "gemma2.attn_logit_softcapping": "50",
+  "gemma2.final_logit_softcapping": "30"
+}

convert/tokenizer.go

@@ -1,7 +1,6 @@
 package convert
 
 import (
-	"cmp"
 	"crypto/sha256"
 	"encoding/hex"
 	"encoding/json"
@@ -11,6 +10,8 @@ import (
 	"log/slog"
 	"os"
 	"slices"
+
+	"golang.org/x/exp/maps"
 )
 
 const (
@@ -184,32 +185,32 @@ func parseVocabularyFromTokenizer(fsys fs.FS) (*Vocabulary, error) {
 		return nil, err
 	}
 
-	var tokens []token
+	tokens := make(map[int]token, len(t.Model.Vocab))
 	for k, v := range t.Model.Vocab {
-		tokens = append(tokens, token{
+		tokens[v] = token{
 			ID:      v,
 			Content: k,
-		})
+		}
 	}
 
-	for _, t := range t.AddedTokens {
-		t.UserDefined = true
-		tokens = append(tokens, t)
+	for _, token := range t.AddedTokens {
+		token.UserDefined = true
+		tokens[token.ID] = token
 	}
 
-	slices.SortFunc(tokens, func(i, j token) int {
-		return cmp.Compare(i.ID, j.ID)
-	})
+	keys := maps.Keys(tokens)
+	slices.Sort(keys)
 
 	v := Vocabulary{Model: "gpt2"}
-	for _, t := range tokens {
-		v.Tokens = append(v.Tokens, t.Content)
-		v.Scores = append(v.Scores, float32(t.ID))
+	for _, k := range keys {
+		token := tokens[k]
+		v.Tokens = append(v.Tokens, token.Content)
+		v.Scores = append(v.Scores, float32(token.ID))
 
 		switch {
-		case t.Special:
+		case token.Special:
 			v.Types = append(v.Types, tokenTypeControl)
-		case t.UserDefined:
+		case token.UserDefined:
 			v.Types = append(v.Types, tokenTypeUserDefined)
 		default:
 			v.Types = append(v.Types, tokenTypeNormal)

convert/tokenizer_spm.go

@@ -15,6 +15,11 @@ import (
 )
 
 func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
+	ast, err := parseAdditionalSpecialTokens(fsys)
+	if err != nil {
+		return nil, err
+	}
+
 	bts, err := fs.ReadFile(fsys, "tokenizer.model")
 	if err != nil {
 		return nil, err
@@ -37,7 +42,12 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 			sentencepiece.ModelProto_SentencePiece_BYTE:
 			v.Types = append(v.Types, int32(t))
 		default:
-			v.Types = append(v.Types, int32(sentencepiece.ModelProto_SentencePiece_NORMAL))
+			tt := int32(sentencepiece.ModelProto_SentencePiece_NORMAL)
+			if slices.Contains(ast, piece.GetPiece()) {
+				tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)
+			}
+
+			v.Types = append(v.Types, tt)
 		}
 	}
 
@@ -81,3 +91,23 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 
 	return &v, nil
 }
+
+func parseAdditionalSpecialTokens(fsys fs.FS) ([]string, error) {
+	f, err := fsys.Open("special_tokens_map.json")
+	if errors.Is(err, os.ErrNotExist) {
+		return nil, nil
+	} else if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	var m struct {
+		AdditionalSpecialTokens []string `json:"additional_special_tokens"`
+	}
+
+	if err := json.NewDecoder(f).Decode(&m); err != nil {
+		return nil, err
+	}
+
+	return m.AdditionalSpecialTokens, nil
+}

docs/api.md

@@ -669,7 +669,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "mistral",
+  "model": "llama3.1",
   "messages": [
     {
       "role": "user",
@@ -708,7 +708,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```json
 {
-  "model": "mistral:7b-instruct-v0.3-q4_K_M",
+  "model": "llama3.1",
   "created_at": "2024-07-22T20:33:28.123648Z",
   "message": {
     "role": "assistant",
@@ -1175,7 +1175,10 @@ curl http://localhost:11434/api/embed -d '{
   "embeddings": [[
     0.010071029, -0.0017594862, 0.05007221, 0.04692972, 0.054916814,
     0.008599704, 0.105441414, -0.025878139, 0.12958129, 0.031952348
-  ]]
+  ]],
+  "total_duration": 14143917,
+  "load_duration": 1019500,
+  "prompt_eval_count": 8
 }
 ```
 

docs/docker.md

@@ -1,71 +1,71 @@
-# Ollama Docker image
-
-### CPU only
-
-```bash
-docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
-```
-
-### Nvidia GPU
-Install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html#installation).
-
-#### Install with Apt
-1.  Configure the repository
-```bash
-curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
-    | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
-curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
-    | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
-    | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
-sudo apt-get update
-```
-2.  Install the NVIDIA Container Toolkit packages
-```bash
-sudo apt-get install -y nvidia-container-toolkit
-```
-
-#### Install with Yum or Dnf
-1.  Configure the repository
-    
-```bash
-curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo \
-    | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
-```
-    
-2. Install the NVIDIA Container Toolkit packages
-    
-```bash
-sudo yum install -y nvidia-container-toolkit
-```
-
-#### Configure Docker to use Nvidia driver 
-```
-sudo nvidia-ctk runtime configure --runtime=docker
-sudo systemctl restart docker
-```
-
-#### Start the container
-
-```bash
-docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
-```
-
-### AMD GPU
-
-To run Ollama using Docker with AMD GPUs, use the `rocm` tag and the following command:
-
-```
-docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama:rocm
-```
-
-### Run model locally
-
-Now you can run a model:
-
-```
-docker exec -it ollama ollama run llama3.1
-```
-
-### Try different models
-
-More models can be found on the [Ollama library](https://ollama.com/library).
+# Ollama Docker image
+
+### CPU only
+
+```bash
+docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
+```
+
+### Nvidia GPU
+Install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html#installation).
+
+#### Install with Apt
+1.  Configure the repository
+```bash
+curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
+    | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
+curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
+    | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
+    | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
+sudo apt-get update
+```
+2.  Install the NVIDIA Container Toolkit packages
+```bash
+sudo apt-get install -y nvidia-container-toolkit
+```
+
+#### Install with Yum or Dnf
+1.  Configure the repository
+
+```bash
+curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo \
+    | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
+```
+
+2. Install the NVIDIA Container Toolkit packages
+
+```bash
+sudo yum install -y nvidia-container-toolkit
+```
+
+#### Configure Docker to use Nvidia driver
+```
+sudo nvidia-ctk runtime configure --runtime=docker
+sudo systemctl restart docker
+```
+
+#### Start the container
+
+```bash
+docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
+```
+
+### AMD GPU
+
+To run Ollama using Docker with AMD GPUs, use the `rocm` tag and the following command:
+
+```
+docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama:rocm
+```
+
+### Run model locally
+
+Now you can run a model:
+
+```
+docker exec -it ollama ollama run llama3.1
+```
+
+### Try different models
+
+More models can be found on the [Ollama library](https://ollama.com/library).

docs/faq.md

@@ -111,7 +111,10 @@ On Windows, Ollama inherits your user and system environment variables.
 
 ## How do I use Ollama behind a proxy?
 
-Ollama is compatible with proxy servers if `HTTP_PROXY` or `HTTPS_PROXY` are configured. When using either variables, ensure it is set where `ollama serve` can access the values. When using `HTTPS_PROXY`, ensure the proxy certificate is installed as a system certificate. Refer to the section above for how to use environment variables on your platform.
+Ollama pulls models from the Internet and may require a proxy server to access the models. Use `HTTPS_PROXY` to redirect outbound requests through the proxy. Ensure the proxy certificate is installed as a system certificate. Refer to the section above for how to use environment variables on your platform.
+
+> [!NOTE]
+> Avoid setting `HTTP_PROXY`. Ollama does not use HTTP for model pulls, only HTTPS. Setting `HTTP_PROXY` may interrupt client connections to the server.
 
 ### How do I use Ollama behind a proxy in Docker?
 
@@ -276,4 +279,4 @@ Note: Windows with Radeon GPUs currently default to 1 model maximum due to limit
 
 ## How does Ollama load models on multiple GPUs?
 
-Installing multiple GPUs of the same brand can be a great way to increase your available VRAM to load larger models.  When you load a new model, Ollama evaluates the required VRAM for the model against what is currently available.  If the model will entirely fit on any single GPU, Ollama will load the model on that GPU.  This typically provides the best performance as it reduces the amount of data transfering across the PCI bus during inference.  If the model does not fit entirely on one GPU, then it will be spread across all the available GPUs.
+Installing multiple GPUs of the same brand can be a great way to increase your available VRAM to load larger models.  When you load a new model, Ollama evaluates the required VRAM for the model against what is currently available.  If the model will entirely fit on any single GPU, Ollama will load the model on that GPU.  This typically provides the best performance as it reduces the amount of data transfering across the PCI bus during inference.  If the model does not fit entirely on one GPU, then it will be spread across all the available GPUs.

docs/import.md

@@ -1,42 +1,129 @@
-# Import
+# Importing a model
 
-GGUF models and select Safetensors models can be imported directly into Ollama.
+## Table of Contents
 
-## Import GGUF
+  * [Importing a Safetensors adapter](#Importing-a-fine-tuned-adapter-from-Safetensors-weights)
+  * [Importing a Safetensors model](#Importing-a-model-from-Safetensors-weights)
+  * [Importing a GGUF file](#Importing-a-GGUF-based-model-or-adapter)
+  * [Sharing models on ollama.com](#Sharing-your-model-on-ollamacom)
 
-A binary GGUF file can be imported directly into Ollama through a Modelfile.
+## Importing a fine tuned adapter from Safetensors weights
+
+First, create a `Modelfile` with a `FROM` command pointing at the base model you used for fine tuning, and an `ADAPTER` command which points to the directory with your Safetensors adapter:
 
 ```dockerfile
-FROM /path/to/file.gguf
+FROM <base model name>
+ADAPTER /path/to/safetensors/adapter/directory
 ```
 
-## Import Safetensors
+Make sure that you use the same base model in the `FROM` command as you used to create the adapter otherwise you will get erratic results. Most frameworks use different quantization methods, so it's best to use non-quantized (i.e. non-QLoRA) adapters. If your adapter is in the same directory as your `Modelfile`, use `ADAPTER .` to specify the adapter path.
 
-If the model being imported is one of these architectures, it can be imported directly into Ollama through a Modelfile:
+Now run `ollama create` from the directory where the `Modelfile` was created:
 
- - LlamaForCausalLM
- - MistralForCausalLM
- - GemmaForCausalLM
+```bash
+ollama create my-model
+```
+
+Lastly, test the model:
+
+```bash
+ollama run my-model
+```
+
+Ollama supports importing adapters based on several different model architectures including:
+
+  * Llama (including Llama 2, Llama 3, and Llama 3.1);
+  * Mistral (including Mistral 1, Mistral 2, and Mixtral); and
+  * Gemma (including Gemma 1 and Gemma 2)
+
+You can create the adapter using a fine tuning framework or tool which can output adapters in the Safetensors format, such as:
+
+  * Hugging Face [fine tuning framework] (https://huggingface.co/docs/transformers/en/training)
+  * [Unsloth](https://github.com/unslothai/unsloth)
+  * [MLX](https://github.com/ml-explore/mlx)
+
+
+## Importing a model from Safetensors weights
+
+First, create a `Modelfile` with a `FROM` command which points to the directory containing your Safetensors weights:
 
 ```dockerfile
 FROM /path/to/safetensors/directory
 ```
 
-For architectures not directly convertable by Ollama, see llama.cpp's [guide](https://github.com/ggerganov/llama.cpp/blob/master/README.md#prepare-and-quantize) on conversion. After conversion, see [Import GGUF](#import-gguf).
+If you create the Modelfile in the same directory as the weights, you can use the command `FROM .`.
 
-## Automatic Quantization
+Now run the `ollama create` command from the directory where you created the `Modelfile`:
 
-> [!NOTE]
-> Automatic quantization requires v0.1.35 or higher.
+```shell
+ollama create my-model
+```
 
-Ollama is capable of quantizing FP16 or FP32 models to any of the supported quantizations with the `-q/--quantize` flag in `ollama create`.
+Lastly, test the model:
+
+```shell
+ollama run my-model
+```
+
+Ollama supports importing models for several different architectures including:
+
+  * Llama (including Llama 2, Llama 3, and Llama 3.1);
+  * Mistral (including Mistral 1, Mistral 2, and Mixtral);
+  * Gemma (including Gemma 1 and Gemma 2); and
+  * Phi3
+
+This includes importing foundation models as well as any fine tuned models which which have been _fused_ with a foundation model.
+
+
+## Importing a GGUF based model or adapter
+
+If you have a GGUF based model or adapter it is possible to import it into Ollama. You can obtain a GGUF model or adapter by:
+
+  * converting a Safetensors model with the `convert_hf_to_gguf.py` from Llama.cpp; 
+  * converting a Safetensors adapter with the `convert_lora_to_gguf.py` from Llama.cpp; or
+  * downloading a model or adapter from a place such as HuggingFace
+
+To import a GGUF model, create a `Modelfile` containg:
+
+```dockerfile
+FROM /path/to/file.gguf
+```
+
+For a GGUF adapter, create the `Modelfile` with:
+
+```dockerfile
+FROM <model name>
+ADAPTER /path/to/file.gguf
+```
+
+When importing a GGUF adapter, it's important to use the same base model as the base model that the adapter was created with. You can use:
+
+ * a model from Ollama
+ * a GGUF file
+ * a Safetensors based model 
+
+Once you have created your `Modelfile`, use the `ollama create` command to build the model.
+
+```shell
+ollama create my-model
+```
+
+## Quantizing a Model
+
+Quantizing a model allows you to run models faster and with less memory consumption but at reduced accuracy. This allows you to run a model on more modest hardware.
+
+Ollama can quantize FP16 and FP32 based models into different quantization levels using the `-q/--quantize` flag with the `ollama create` command.
+
+First, create a Modelfile with the FP16 or FP32 based model you wish to quantize.
 
 ```dockerfile
 FROM /path/to/my/gemma/f16/model
 ```
 
+Use `ollama create` to then create the quantized model.
+
 ```shell
-$ ollama create -q Q4_K_M mymodel
+$ ollama create --quantize q4_K_M mymodel
 transferring model data
 quantizing F16 model to Q4_K_M
 creating new layer sha256:735e246cc1abfd06e9cdcf95504d6789a6cd1ad7577108a70d9902fef503c1bd
@@ -47,42 +134,53 @@ success
 
 ### Supported Quantizations
 
-- `Q4_0`
-- `Q4_1`
-- `Q5_0`
-- `Q5_1`
-- `Q8_0`
+- `q4_0`
+- `q4_1`
+- `q5_0`
+- `q5_1`
+- `q8_0`
 
 #### K-means Quantizations
 
-- `Q3_K_S`
-- `Q3_K_M`
-- `Q3_K_L`
-- `Q4_K_S`
-- `Q4_K_M`
-- `Q5_K_S`
-- `Q5_K_M`
-- `Q6_K`
+- `q3_K_S`
+- `q3_K_M`
+- `q3_K_L`
+- `q4_K_S`
+- `q4_K_M`
+- `q5_K_S`
+- `q5_K_M`
+- `q6_K`
 
-## Template Detection
 
-> [!NOTE]
-> Template detection requires v0.1.42 or higher.
+## Sharing your model on ollama.com
 
-Ollama uses model metadata, specifically `tokenizer.chat_template`, to automatically create a template appropriate for the model you're importing.
+You can share any model you have created by pushing it to [ollama.com](https://ollama.com) so that other users can try it out.
 
-```dockerfile
-FROM /path/to/my/gemma/model
-```
+First, use your browser to go to the [Ollama Sign-Up](https://ollama.com/signup) page. If you already have an account, you can skip this step.
+
+![Sign-Up](images/signup.png)
+
+The `Username` field will be used as part of your model's name (e.g. `jmorganca/mymodel`), so make sure you are comfortable with the username that you have selected.
+
+Now that you have created an account and are signed-in, go to the [Ollama Keys Settings](https://ollama.com/settings/keys) page.
+
+Follow the directions on the page to determine where your Ollama Public Key is located.
+
+![Ollama Key](images/ollama-keys.png)
+
+Click on the `Add Ollama Public Key` button, and copy and paste the contents of your Ollama Public Key into the text field.
+
+To push a model to [ollama.com](https://ollama.com), first make sure that it is named correctly with your username. You may have to use the `ollama cp` command to copy
+your model to give it the correct name. Once you're happy with your model's name, use the `ollama push` command to push it to [ollama.com](https://ollama.com).
 
 ```shell
-$ ollama create mymodel
-transferring model data
-using autodetected template gemma-instruct
-creating new layer sha256:baa2a0edc27d19cc6b7537578a9a7ba1a4e3214dc185ed5ae43692b319af7b84
-creating new layer sha256:ba66c3309914dbef07e5149a648fd1877f030d337a4f240d444ea335008943cb
-writing manifest
-success
+ollama cp mymodel myuser/mymodel
+ollama push myuser/mymodel
+```
+
+Once your model has been pushed, other users can pull and run it by using the command:
+
+```shell
+ollama run myuser/mymodel
 ```
 
-Defining a template in the Modelfile will disable this feature which may be useful if you want to use a different template than the autodetected one.

docs/linux.md

@@ -20,13 +20,12 @@ GPU.
 
 ## Manual install
 
-### Download the `ollama` binary
+### Download `ollama`
 
-Ollama is distributed as a self-contained binary. Download it to a directory in your PATH:
+Download and extract the Linux package:
 
 ```bash
-sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama
-sudo chmod +x /usr/bin/ollama
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
 ```
 
 ### Adding Ollama as a startup service (recommended)
@@ -96,8 +95,7 @@ curl -fsSL https://ollama.com/install.sh | sh
 Or by downloading the ollama binary:
 
 ```bash
-sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama
-sudo chmod +x /usr/bin/ollama
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
 ```
 
 ## Installing specific versions

docs/openai.md

@@ -28,13 +28,35 @@ chat_completion = client.chat.completions.create(
     model='llama3',
 )
 
+response = client.chat.completions.create(
+    model="llava",
+    messages=[
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What's in this image?"},
+                {
+                    "type": "image_url",
+                    "image_url": "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",
+                },
+            ],
+        }
+    ],
+    max_tokens=300,
+)
+
+completion = client.completions.create(
+    model="llama3",
+    prompt="Say this is a test",
+)
+
 list_completion = client.models.list()
 
 model = client.models.retrieve("llama3")
 
 embeddings = client.embeddings.create(
     model="all-minilm",
-    input=["why is the sky blue?", "why is the grass green?"]
+    input=["why is the sky blue?", "why is the grass green?"],
 )
 ```
 
@@ -51,23 +73,44 @@ const openai = new OpenAI({
 })
 
 const chatCompletion = await openai.chat.completions.create({
-  messages: [{ role: 'user', content: 'Say this is a test' }],
-  model: 'llama3',
+    messages: [{ role: 'user', content: 'Say this is a test' }],
+    model: 'llama3',
+})
+
+const response = await openai.chat.completions.create({
+    model: "llava",
+    messages: [
+        {
+        role: "user",
+        content: [
+            { type: "text", text: "What's in this image?" },
+            {
+            type: "image_url",
+            image_url: "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",
+            },
+        ],
+        },
+    ],
+})
+
+const completion = await openai.completions.create({
+    model: "llama3",
+    prompt: "Say this is a test.",
 })
 
 const listCompletion = await openai.models.list()
 
-const model = await openai.models.retrieve("llama3");
+const model = await openai.models.retrieve("llama3")
 
 const embedding = await openai.embeddings.create({
   model: "all-minilm",
   input: ["why is the sky blue?", "why is the grass green?"],
-});
+})
 ```
 
 ### `curl`
 
-```
+``` shell
 curl http://localhost:11434/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{
@@ -84,6 +127,37 @@ curl http://localhost:11434/v1/chat/completions \
         ]
     }'
 
+curl http://localhost:11434/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "llava",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {
+            "type": "text",
+            "text": "What'\''s in this image?"
+          },
+          {
+            "type": "image_url",
+            "image_url": {
+               "url": "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"
+            }
+          }
+        ]
+      }
+    ],
+    "max_tokens": 300
+  }'
+
+curl http://localhost:11434/v1/completions \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "llama3",
+        "prompt": "Say this is a test"
+    }'
+
 curl http://localhost:11434/v1/models
 
 curl http://localhost:11434/v1/models/llama3
@@ -106,8 +180,8 @@ curl http://localhost:11434/v1/embeddings \
 - [x] Streaming
 - [x] JSON mode
 - [x] Reproducible outputs
+- [x] Vision
 - [x] Tools (streaming support coming soon)
-- [ ] Vision
 - [ ] Logprobs
 
 #### Supported request fields
@@ -115,7 +189,10 @@ curl http://localhost:11434/v1/embeddings \
 - [x] `model`
 - [x] `messages`
   - [x] Text `content`
-  - [ ] Array of `content` parts
+  - [x] Image `content`
+    - [x] Base64 encoded image
+    - [ ] Image URL
+  - [x] Array of `content` parts
 - [x] `frequency_penalty`
 - [x] `presence_penalty`
 - [x] `response_format`
@@ -131,6 +208,39 @@ curl http://localhost:11434/v1/embeddings \
 - [ ] `user`
 - [ ] `n`
 
+### `/v1/completions`
+
+#### Supported features
+
+- [x] Completions
+- [x] Streaming
+- [x] JSON mode
+- [x] Reproducible outputs
+- [ ] Logprobs
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `prompt`
+- [x] `frequency_penalty`
+- [x] `presence_penalty`
+- [x] `seed`
+- [x] `stop`
+- [x] `stream`
+- [x] `temperature`
+- [x] `top_p`
+- [x] `max_tokens`
+- [x] `suffix`
+- [ ] `best_of`
+- [ ] `echo`
+- [ ] `logit_bias`
+- [ ] `user`
+- [ ] `n`
+
+#### Notes
+
+- `prompt` currently only accepts a string
+
 ### `/v1/models`
 
 #### Notes

docs/template.md

@@ -112,15 +112,9 @@ Keep the following tips and best practices in mind when working with Go template
 ChatML is a popular template format. It can be used for models such as Databrick's DBRX, Intel's Neural Chat, and Microsoft's Orca 2.
 
 ```gotmpl
-{{- if .System }}<|im_start|>system
-{{ .System }}<|im_end|>
-{{ end }}
 {{- range .Messages }}<|im_start|>{{ .Role }}
 {{ .Content }}<|im_end|>
 {{ end }}<|im_start|>assistant
-{{ else }}
-{{ if .System }}<|im_start|>system
-{{ .System }}<|im_end|>
 ```
 
 ### Example Tools

docs/troubleshooting.md

@@ -9,7 +9,7 @@ cat ~/.ollama/logs/server.log
 On **Linux** systems with systemd, the logs can be found with this command:
 
 ```shell
-journalctl -u ollama
+journalctl -u ollama --no-pager
 ```
 
 When you run Ollama in a **container**, the logs go to stdout/stderr in the container:

envconfig/config.go

@@ -174,7 +174,7 @@ func RunnersDir() (p string) {
 
 	defer func() {
 		if p == "" {
-			slog.Error("unable to locate llm runner directory. Set OLLAMA_RUNNERS_DIR to the location of 'ollama_runners'")
+			slog.Error("unable to locate llm runner directory. Set OLLAMA_RUNNERS_DIR to the location of 'ollama/runners'")
 		}
 	}()
 
@@ -190,17 +190,17 @@ func RunnersDir() (p string) {
 	}
 
 	var paths []string
-	for _, root := range []string{filepath.Dir(exe), cwd} {
+	for _, root := range []string{filepath.Dir(exe), filepath.Join(filepath.Dir(exe), ".."), cwd} {
 		paths = append(paths,
 			root,
-			filepath.Join(root, "windows-"+runtime.GOARCH),
-			filepath.Join(root, "dist", "windows-"+runtime.GOARCH),
+			filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH),
+			filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH),
 		)
 	}
 
 	// Try a few variations to improve developer experience when building from source in the local tree
 	for _, path := range paths {
-		candidate := filepath.Join(path, "ollama_runners")
+		candidate := filepath.Join(path, "lib", "ollama", "runners")
 		if _, err := os.Stat(candidate); err == nil {
 			p = candidate
 			break

format/format.go

@@ -3,6 +3,7 @@ package format
 import (
 	"fmt"
 	"math"
+	"strconv"
 )
 
 const (
@@ -28,6 +29,6 @@ func HumanNumber(b uint64) string {
 	case b >= Thousand:
 		return fmt.Sprintf("%.0fK", float64(b)/Thousand)
 	default:
-		return fmt.Sprintf("%d", b)
+		return strconv.FormatUint(b, 10)
 	}
 }

go.mod

@@ -1,6 +1,6 @@
 module github.com/ollama/ollama
 
-go 1.22.0
+go 1.22.5
 
 require (
 	github.com/containerd/console v1.0.3

gpu/amd_common.go

@@ -3,7 +3,7 @@
 package gpu
 
 import (
-	"fmt"
+	"errors"
 	"log/slog"
 	"os"
 	"path/filepath"
@@ -54,7 +54,7 @@ func commonAMDValidateLibDir() (string, error) {
 	// Installer payload location if we're running the installed binary
 	exe, err := os.Executable()
 	if err == nil {
-		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
+		rocmTargetDir := filepath.Join(filepath.Dir(exe), "..", "lib", "ollama")
 		if rocmLibUsable(rocmTargetDir) {
 			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
 			return rocmTargetDir, nil
@@ -95,5 +95,5 @@ func commonAMDValidateLibDir() (string, error) {
 		}
 	}
 
-	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
+	return "", errors.New("no suitable rocm found, falling back to CPU")
 }

gpu/amd_hip_windows.go

@@ -1,6 +1,7 @@
 package gpu
 
 import (
+	"errors"
 	"fmt"
 	"log/slog"
 	"syscall"
@@ -76,7 +77,7 @@ func (hl *HipLib) Release() {
 
 func (hl *HipLib) AMDDriverVersion() (driverMajor, driverMinor int, err error) {
 	if hl.dll == 0 {
-		return 0, 0, fmt.Errorf("dll has been unloaded")
+		return 0, 0, errors.New("dll has been unloaded")
 	}
 	var version int
 	status, _, err := syscall.SyscallN(hl.hipDriverGetVersion, uintptr(unsafe.Pointer(&version)))
@@ -110,7 +111,7 @@ func (hl *HipLib) HipGetDeviceCount() int {
 
 func (hl *HipLib) HipSetDevice(device int) error {
 	if hl.dll == 0 {
-		return fmt.Errorf("dll has been unloaded")
+		return errors.New("dll has been unloaded")
 	}
 	status, _, err := syscall.SyscallN(hl.hipSetDevice, uintptr(device))
 	if status != hipSuccess {
@@ -121,7 +122,7 @@ func (hl *HipLib) HipSetDevice(device int) error {
 
 func (hl *HipLib) HipGetDeviceProperties(device int) (*hipDevicePropMinimal, error) {
 	if hl.dll == 0 {
-		return nil, fmt.Errorf("dll has been unloaded")
+		return nil, errors.New("dll has been unloaded")
 	}
 	var props hipDevicePropMinimal
 	status, _, err := syscall.SyscallN(hl.hipGetDeviceProperties, uintptr(unsafe.Pointer(&props)), uintptr(device))
@@ -134,7 +135,7 @@ func (hl *HipLib) HipGetDeviceProperties(device int) (*hipDevicePropMinimal, err
 // free, total, err
 func (hl *HipLib) HipMemGetInfo() (uint64, uint64, error) {
 	if hl.dll == 0 {
-		return 0, 0, fmt.Errorf("dll has been unloaded")
+		return 0, 0, errors.New("dll has been unloaded")
 	}
 	var totalMemory uint64
 	var freeMemory uint64

gpu/amd_linux.go

@@ -393,7 +393,7 @@ func AMDValidateLibDir() (string, error) {
 
 	// If we still haven't found a usable rocm, the user will have to install it on their own
 	slog.Warn("amdgpu detected, but no compatible rocm library found.  Either install rocm v6, or follow manual install instructions at https://github.com/ollama/ollama/blob/main/docs/linux.md#manual-install")
-	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
+	return "", errors.New("no suitable rocm found, falling back to CPU")
 }
 
 func AMDDriverVersion() (driverMajor, driverMinor int, err error) {

gpu/amd_windows.go

@@ -2,7 +2,7 @@ package gpu
 
 import (
 	"bytes"
-	"fmt"
+	"errors"
 	"log/slog"
 	"os"
 	"path/filepath"
@@ -85,7 +85,7 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 		n = bytes.IndexByte(props.GcnArchName[:], 0)
 		gfx := string(props.GcnArchName[:n])
 		slog.Debug("hip device", "id", i, "name", name, "gfx", gfx)
-		//slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY!  Always 0
+		// slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY!  Always 0
 		// TODO  Why isn't props.iGPU accurate!?
 		if strings.EqualFold(name, iGPUName) {
 			slog.Info("unsupported Radeon iGPU detected skipping", "id", i, "name", name, "gfx", gfx)
@@ -153,7 +153,7 @@ func AMDValidateLibDir() (string, error) {
 	// Installer payload (if we're running from some other location)
 	localAppData := os.Getenv("LOCALAPPDATA")
 	appDir := filepath.Join(localAppData, "Programs", "Ollama")
-	rocmTargetDir := filepath.Join(appDir, "rocm")
+	rocmTargetDir := filepath.Join(appDir, "..", "lib", "ollama")
 	if rocmLibUsable(rocmTargetDir) {
 		slog.Debug("detected ollama installed ROCm at " + rocmTargetDir)
 		return rocmTargetDir, nil
@@ -161,7 +161,7 @@ func AMDValidateLibDir() (string, error) {
 
 	// Should not happen on windows since we include it in the installer, but stand-alone binary might hit this
 	slog.Warn("amdgpu detected, but no compatible rocm library found.  Please install ROCm")
-	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
+	return "", errors.New("no suitable rocm found, falling back to CPU")
 }
 
 func (gpus RocmGPUInfoList) RefreshFreeMemory() error {

gpu/assets.go

@@ -42,20 +42,16 @@ func PayloadsDir() (string, error) {
 				return "", fmt.Errorf("failed to generate tmp dir: %w", err)
 			}
 		} else {
-			err = os.MkdirAll(tmpDir, 0755)
+			err = os.MkdirAll(tmpDir, 0o755)
 			if err != nil {
 				return "", fmt.Errorf("failed to generate tmp dir %s: %w", tmpDir, err)
 			}
 		}
 
 		// Track our pid so we can clean up orphaned tmpdirs
-		pidFilePath := filepath.Join(tmpDir, "ollama.pid")
-		pidFile, err := os.OpenFile(pidFilePath, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, os.ModePerm)
-		if err != nil {
-			return "", err
-		}
-		if _, err := pidFile.Write([]byte(fmt.Sprint(os.Getpid()))); err != nil {
-			return "", err
+		n := filepath.Join(tmpDir, "ollama.pid")
+		if err := os.WriteFile(n, []byte(strconv.Itoa(os.Getpid())), 0o644); err != nil {
+			return "", fmt.Errorf("failed to write pid file %s: %w", n, err)
 		}
 
 		// We create a distinct subdirectory for payloads within the tmpdir
@@ -67,37 +63,44 @@ func PayloadsDir() (string, error) {
 
 // Best effort to clean up prior tmpdirs
 func cleanupTmpDirs() {
-	dirs, err := filepath.Glob(filepath.Join(os.TempDir(), "ollama*"))
+	matches, err := filepath.Glob(filepath.Join(os.TempDir(), "ollama*", "ollama.pid"))
 	if err != nil {
 		return
 	}
-	for _, d := range dirs {
-		info, err := os.Stat(d)
-		if err != nil || !info.IsDir() {
+
+	for _, match := range matches {
+		raw, err := os.ReadFile(match)
+		if errors.Is(err, os.ErrNotExist) {
+			slog.Debug("not a ollama runtime directory, skipping", "path", match)
 			continue
-		}
-		raw, err := os.ReadFile(filepath.Join(d, "ollama.pid"))
-		if err != nil {
-			slog.Warn("failed to read ollama.pid", "path", d, "error", err)
-			// No pid, ignore this tmpdir
+		} else if err != nil {
+			slog.Warn("could not read ollama.pid, skipping", "path", match, "error", err)
 			continue
 		}
 
 		pid, err := strconv.Atoi(string(raw))
 		if err != nil {
-			slog.Warn("failed to parse pid", "path", d, "error", err)
+			slog.Warn("invalid pid, skipping", "path", match, "error", err)
 			continue
 		}
 
-		proc, err := os.FindProcess(pid)
-		if err == nil && !errors.Is(proc.Signal(syscall.Signal(0)), os.ErrProcessDone) {
-			slog.Warn("found running ollama", "pid", pid, "path", d)
-			// Another running ollama, ignore this tmpdir
+		p, err := os.FindProcess(pid)
+		if err == nil && !errors.Is(p.Signal(syscall.Signal(0)), os.ErrProcessDone) {
+			slog.Warn("process still running, skipping", "pid", pid, "path", match)
 			continue
 		}
 
-		if err := os.Remove(d); err != nil {
-			slog.Warn("unable to cleanup stale tmpdir", "path", d, "error", err)
+		if err := os.Remove(match); err != nil {
+			slog.Warn("could not cleanup stale pidfile", "path", match, "error", err)
+		}
+
+		runners := filepath.Join(filepath.Dir(match), "runners")
+		if err := os.RemoveAll(runners); err != nil {
+			slog.Warn("could not cleanup stale runners", "path", runners, "error", err)
+		}
+
+		if err := os.Remove(filepath.Dir(match)); err != nil {
+			slog.Warn("could not cleanup stale tmpdir", "path", filepath.Dir(match), "error", err)
 		}
 	}
 }

gpu/cpu_common.go

@@ -1,6 +1,11 @@
 package gpu
 
 import (
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+
 	"golang.org/x/sys/cpu"
 )
 
@@ -14,3 +19,19 @@ func GetCPUCapability() CPUCapability {
 	// else LCD
 	return CPUCapabilityNone
 }
+
+func IsNUMA() bool {
+	if runtime.GOOS != "linux" {
+		// numa support in llama.cpp is linux only
+		return false
+	}
+	ids := map[string]interface{}{}
+	packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")
+	for _, packageId := range packageIds {
+		id, err := os.ReadFile(packageId)
+		if err == nil {
+			ids[strings.TrimSpace(string(id))] = struct{}{}
+		}
+	}
+	return len(ids) > 1
+}

gpu/cuda_common.go

@@ -4,9 +4,17 @@ package gpu
 
 import (
 	"log/slog"
+	"os"
+	"regexp"
+	"runtime"
+	"strconv"
 	"strings"
 )
 
+// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
+// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
+var CudaTegra string = os.Getenv("JETSON_JETPACK")
+
 func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 	ids := []string{}
 	for _, info := range gpuInfo {
@@ -19,3 +27,38 @@ func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 	}
 	return "CUDA_VISIBLE_DEVICES", strings.Join(ids, ",")
 }
+
+func cudaVariant(gpuInfo CudaGPUInfo) string {
+	if runtime.GOARCH == "arm64" && runtime.GOOS == "linux" {
+		if CudaTegra != "" {
+			ver := strings.Split(CudaTegra, ".")
+			if len(ver) > 0 {
+				return "jetpack" + ver[0]
+			}
+		} else if data, err := os.ReadFile("/etc/nv_tegra_release"); err == nil {
+			r := regexp.MustCompile(` R(\d+) `)
+			m := r.FindSubmatch(data)
+			if len(m) != 2 {
+				slog.Info("Unexpected format for /etc/nv_tegra_release.  Set JETSON_JETPACK to select version")
+			} else {
+				if l4t, err := strconv.Atoi(string(m[1])); err == nil {
+					// Note: mapping from L4t -> JP is inconsistent (can't just subtract 30)
+					// https://developer.nvidia.com/embedded/jetpack-archive
+					switch l4t {
+					case 35:
+						return "jetpack5"
+					case 36:
+						return "jetpack6"
+					default:
+						slog.Info("unsupported L4T version", "nv_tegra_release", string(data))
+					}
+				}
+			}
+		}
+	}
+
+	if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 {
+		return "v11"
+	}
+	return "v12"
+}

gpu/gpu.go

@@ -7,9 +7,9 @@ package gpu
 #cgo windows LDFLAGS: -lpthread
 
 #include "gpu_info.h"
-
 */
 import "C"
+
 import (
 	"fmt"
 	"log/slog"
@@ -64,13 +64,8 @@ var RocmComputeMin = 9
 // TODO find a better way to detect iGPU instead of minimum memory
 const IGPUMemLimit = 1 * format.GibiByte // 512G is what they typically report, so anything less than 1G must be iGPU
 
-// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
-// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
-var CudaTegra string = os.Getenv("JETSON_JETPACK")
-
 // Note: gpuMutex must already be held
 func initCudaHandles() *cudaHandles {
-
 	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
 
 	cHandles := &cudaHandles{}
@@ -211,14 +206,16 @@ func GetGPUInfo() GpuInfoList {
 		if err != nil {
 			slog.Warn("error looking up system memory", "error", err)
 		}
-		cpus = []CPUInfo{CPUInfo{
-			GpuInfo: GpuInfo{
-				memInfo: mem,
-				Library: "cpu",
-				Variant: cpuCapability,
-				ID:      "0",
+		cpus = []CPUInfo{
+			{
+				GpuInfo: GpuInfo{
+					memInfo: mem,
+					Library: "cpu",
+					Variant: cpuCapability.String(),
+					ID:      "0",
+				},
 			},
-		}}
+		}
 
 		// Fallback to CPU mode if we're lacking required vector extensions on x86
 		if cpuCapability < GPURunnerCPUCapability && runtime.GOARCH == "amd64" {
@@ -228,11 +225,7 @@ func GetGPUInfo() GpuInfoList {
 			return GpuInfoList{cpus[0].GpuInfo}
 		}
 
-		// On windows we bundle the nvidia library one level above the runner dir
-		depPath := ""
-		if runtime.GOOS == "windows" && envconfig.RunnersDir() != "" {
-			depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir()), "cuda")
-		}
+		depPath := LibraryDir()
 
 		// Load ALL libraries
 		cHandles = initCudaHandles()
@@ -268,11 +261,23 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.FreeMemory = uint64(memInfo.free)
 				gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
 				gpuInfo.Compute = fmt.Sprintf("%d.%d", memInfo.major, memInfo.minor)
+				gpuInfo.computeMajor = int(memInfo.major)
+				gpuInfo.computeMinor = int(memInfo.minor)
 				gpuInfo.MinimumMemory = cudaMinimumMemory
-				gpuInfo.DependencyPath = depPath
-				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
 				gpuInfo.DriverMajor = driverMajor
 				gpuInfo.DriverMinor = driverMinor
+				variant := cudaVariant(gpuInfo)
+				if depPath != "" {
+					gpuInfo.DependencyPath = depPath
+					// Check for variant specific directory
+					if variant != "" {
+						if _, err := os.Stat(filepath.Join(depPath, "cuda_"+variant)); err == nil {
+							gpuInfo.DependencyPath = filepath.Join(depPath, "cuda_"+variant)
+						}
+					}
+				}
+				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
+				gpuInfo.Variant = variant
 
 				// query the management library as well so we can record any skew between the two
 				// which represents overhead on the GPU we must set aside on subsequent updates
@@ -304,38 +309,34 @@ func GetGPUInfo() GpuInfoList {
 		// Intel
 		if envconfig.IntelGPU() {
 			oHandles = initOneAPIHandles()
-			// On windows we bundle the oneapi library one level above the runner dir
-			depPath = ""
-			if runtime.GOOS == "windows" && envconfig.RunnersDir() != "" {
-				depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir()), "oneapi")
-			}
-
-			for d := range oHandles.oneapi.num_drivers {
-				if oHandles.oneapi == nil {
-					// shouldn't happen
-					slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers))
-					continue
-				}
-				devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
-				for i := range devCount {
-					gpuInfo := OneapiGPUInfo{
-						GpuInfo: GpuInfo{
-							Library: "oneapi",
-						},
-						driverIndex: int(d),
-						gpuIndex:    int(i),
+			if oHandles != nil && oHandles.oneapi != nil {
+				for d := range oHandles.oneapi.num_drivers {
+					if oHandles.oneapi == nil {
+						// shouldn't happen
+						slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers))
+						continue
+					}
+					devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
+					for i := range devCount {
+						gpuInfo := OneapiGPUInfo{
+							GpuInfo: GpuInfo{
+								Library: "oneapi",
+							},
+							driverIndex: int(d),
+							gpuIndex:    int(i),
+						}
+						// TODO - split bootstrapping from updating free memory
+						C.oneapi_check_vram(*oHandles.oneapi, C.int(d), i, &memInfo)
+						// TODO - convert this to MinimumMemory based on testing...
+						var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
+						memInfo.free = C.uint64_t(totalFreeMem)
+						gpuInfo.TotalMemory = uint64(memInfo.total)
+						gpuInfo.FreeMemory = uint64(memInfo.free)
+						gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
+						gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
+						gpuInfo.DependencyPath = depPath
+						oneapiGPUs = append(oneapiGPUs, gpuInfo)
 					}
-					// TODO - split bootstrapping from updating free memory
-					C.oneapi_check_vram(*oHandles.oneapi, C.int(d), i, &memInfo)
-					// TODO - convert this to MinimumMemory based on testing...
-					var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
-					memInfo.free = C.uint64_t(totalFreeMem)
-					gpuInfo.TotalMemory = uint64(memInfo.total)
-					gpuInfo.FreeMemory = uint64(memInfo.free)
-					gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
-					gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-					gpuInfo.DependencyPath = depPath
-					oneapiGPUs = append(oneapiGPUs, gpuInfo)
 				}
 			}
 		}
@@ -463,10 +464,12 @@ func GetGPUInfo() GpuInfoList {
 func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
 	var ldPaths []string
-	var patterns []string
 	gpuLibPaths := []string{}
 	slog.Debug("Searching for GPU library", "name", baseLibName)
 
+	// Start with our bundled libraries
+	patterns := []string{filepath.Join(LibraryDir(), baseLibName)}
+
 	switch runtime.GOOS {
 	case "windows":
 		ldPaths = strings.Split(os.Getenv("PATH"), ";")
@@ -475,13 +478,14 @@ func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	default:
 		return gpuLibPaths
 	}
-	// Start with whatever we find in the PATH/LD_LIBRARY_PATH
+
+	// Then 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+"*"))
+		patterns = append(patterns, filepath.Join(d, baseLibName))
 	}
 	patterns = append(patterns, defaultPatterns...)
 	slog.Debug("gpu library search", "globs", patterns)
@@ -637,3 +641,31 @@ func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
 		return "", ""
 	}
 }
+
+func LibraryDir() string {
+	// On Windows/linux we bundle the dependencies at the same level as the executable
+	appExe, err := os.Executable()
+	if err != nil {
+		slog.Warn("failed to lookup executable path", "error", err)
+	}
+	cwd, err := os.Getwd()
+	if err != nil {
+		slog.Warn("failed to lookup working directory", "error", err)
+	}
+	// Scan for any of our dependeices, and pick first match
+	for _, root := range []string{filepath.Dir(appExe), filepath.Join(filepath.Dir(appExe), ".."), cwd} {
+		libDep := filepath.Join("lib", "ollama")
+		if _, err := os.Stat(filepath.Join(root, libDep)); err == nil {
+			return filepath.Join(root, libDep)
+		}
+		// Developer mode, local build
+		if _, err := os.Stat(filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH, libDep)); err == nil {
+			return filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH, libDep)
+		}
+		if _, err := os.Stat(filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH, libDep)); err == nil {
+			return filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH, libDep)
+		}
+	}
+	slog.Warn("unable to locate gpu dependency libraries")
+	return ""
+}

gpu/gpu_darwin.go

@@ -8,6 +8,7 @@ package gpu
 #include "gpu_info_darwin.h"
 */
 import "C"
+
 import (
 	"runtime"
 
@@ -24,7 +25,7 @@ func GetGPUInfo() GpuInfoList {
 		return []GpuInfo{
 			{
 				Library: "cpu",
-				Variant: GetCPUCapability(),
+				Variant: GetCPUCapability().String(),
 				memInfo: mem,
 			},
 		}
@@ -47,7 +48,7 @@ func GetCPUInfo() GpuInfoList {
 	return []GpuInfo{
 		{
 			Library: "cpu",
-			Variant: GetCPUCapability(),
+			Variant: GetCPUCapability().String(),
 			memInfo: mem,
 		},
 	}

gpu/gpu_info.h

@@ -67,4 +67,4 @@ void cpu_check_ram(mem_info_t *resp);
 #include "gpu_info_oneapi.h"
 
 #endif  // __GPU_INFO_H__
-#endif  // __APPLE__
+#endif  // __APPLE__

gpu/gpu_linux.go

@@ -43,10 +43,12 @@ var OneapiGlobs = []string{
 	"/usr/lib*/libze_intel_gpu.so*",
 }
 
-var CudartMgmtName = "libcudart.so*"
-var NvcudaMgmtName = "libcuda.so*"
-var NvmlMgmtName = "" // not currently wired on linux
-var OneapiMgmtName = "libze_intel_gpu.so"
+var (
+	CudartMgmtName = "libcudart.so*"
+	NvcudaMgmtName = "libcuda.so*"
+	NvmlMgmtName   = "" // not currently wired on linux
+	OneapiMgmtName = "libze_intel_gpu.so*"
+)
 
 func GetCPUMem() (memInfo, error) {
 	var mem memInfo

gpu/gpu_test.go

@@ -32,4 +32,29 @@ func TestCPUMemInfo(t *testing.T) {
 	}
 }
 
+func TestByLibrary(t *testing.T) {
+	type testCase struct {
+		input  []GpuInfo
+		expect int
+	}
+
+	testCases := map[string]*testCase{
+		"empty":                    {input: []GpuInfo{}, expect: 0},
+		"cpu":                      {input: []GpuInfo{{Library: "cpu"}}, expect: 1},
+		"cpu + GPU":                {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda"}}, expect: 2},
+		"cpu + 2 GPU no variant":   {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda"}, {Library: "cuda"}}, expect: 2},
+		"cpu + 2 GPU same variant": {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda", Variant: "v11"}, {Library: "cuda", Variant: "v11"}}, expect: 2},
+		"cpu + 2 GPU diff variant": {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda", Variant: "v11"}, {Library: "cuda", Variant: "v12"}}, expect: 3},
+	}
+
+	for k, v := range testCases {
+		t.Run(k, func(t *testing.T) {
+			resp := (GpuInfoList)(v.input).ByLibrary()
+			if len(resp) != v.expect {
+				t.Fatalf("expected length %d, got %d => %+v", v.expect, len(resp), resp)
+			}
+		})
+	}
+}
+
 // TODO - add some logic to figure out card type through other means and actually verify we got back what we expected

gpu/gpu_windows.go

@@ -40,10 +40,12 @@ var OneapiGlobs = []string{
 	"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
 }
 
-var CudartMgmtName = "cudart64_*.dll"
-var NvcudaMgmtName = "nvcuda.dll"
-var NvmlMgmtName = "nvml.dll"
-var OneapiMgmtName = "ze_intel_gpu64.dll"
+var (
+	CudartMgmtName = "cudart64_*.dll"
+	NvcudaMgmtName = "nvcuda.dll"
+	NvmlMgmtName   = "nvml.dll"
+	OneapiMgmtName = "ze_intel_gpu64.dll"
+)
 
 func GetCPUMem() (memInfo, error) {
 	memStatus := MEMORYSTATUSEX{length: sizeofMemoryStatusEx}

gpu/types.go

@@ -19,7 +19,7 @@ type GpuInfo struct {
 	Library string `json:"library,omitempty"`
 
 	// Optional variant to select (e.g. versions, cpu feature flags)
-	Variant CPUCapability `json:"variant"`
+	Variant string `json:"variant"`
 
 	// MinimumMemory represents the minimum memory required to use the GPU
 	MinimumMemory uint64 `json:"-"`
@@ -53,8 +53,10 @@ type CPUInfo struct {
 
 type CudaGPUInfo struct {
 	GpuInfo
-	OSOverhead uint64 // Memory overhead between the driver library and management library
-	index      int    //nolint:unused,nolintlint
+	OSOverhead   uint64 // Memory overhead between the driver library and management library
+	index        int    //nolint:unused,nolintlint
+	computeMajor int    //nolint:unused,nolintlint
+	computeMinor int    //nolint:unused,nolintlint
 }
 type CudaGPUInfoList []CudaGPUInfo
 
@@ -81,8 +83,8 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	for _, info := range l {
 		found := false
 		requested := info.Library
-		if info.Variant != CPUCapabilityNone {
-			requested += "_" + info.Variant.String()
+		if info.Variant != CPUCapabilityNone.String() {
+			requested += "_" + info.Variant
 		}
 		for i, lib := range libs {
 			if lib == requested {
@@ -92,7 +94,7 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 			}
 		}
 		if !found {
-			libs = append(libs, info.Library)
+			libs = append(libs, requested)
 			resp = append(resp, []GpuInfo{info})
 		}
 	}
@@ -105,6 +107,7 @@ func (l GpuInfoList) LogDetails() {
 		slog.Info("inference compute",
 			"id", g.ID,
 			"library", g.Library,
+			"variant", g.Variant,
 			"compute", g.Compute,
 			"driver", fmt.Sprintf("%d.%d", g.DriverMajor, g.DriverMinor),
 			"name", g.Name,

integration/concurrency_test.go

@@ -5,6 +5,7 @@ package integration
 import (
 	"context"
 	"log/slog"
+	"os"
 	"strconv"
 	"sync"
 	"testing"
@@ -13,7 +14,6 @@ import (
 	"github.com/stretchr/testify/require"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 )
 
@@ -41,8 +41,8 @@ func TestMultiModelConcurrency(t *testing.T) {
 			},
 		}
 		resp = [2][]string{
-			[]string{"sunlight"},
-			[]string{"england", "english", "massachusetts", "pilgrims", "british"},
+			{"sunlight"},
+			{"england", "english", "massachusetts", "pilgrims", "british"},
 		}
 	)
 	var wg sync.WaitGroup
@@ -71,12 +71,11 @@ func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 	reqLimit := len(req)
 	iterLimit := 5
 
-	vram := os.Getenv("OLLAMA_MAX_VRAM") // TODO - discover actual VRAM
-	if vram != "" {
-		max, err := strconv.ParseUint(vram, 10, 64)
+	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
+		maxVram, err := strconv.ParseUint(s, 10, 64)
 		require.NoError(t, err)
 		// Don't hammer on small VRAM cards...
-		if max < 4*1024*1024*1024 {
+		if maxVram < 4*format.GibiByte {
 			reqLimit = min(reqLimit, 2)
 			iterLimit = 2
 		}
@@ -233,12 +232,12 @@ func TestMultiModelStress(t *testing.T) {
 	consumed := uint64(256 * format.MebiByte) // Assume some baseline usage
 	for i := 0; i < len(req); i++ {
 		// Always get at least 2 models, but dont' overshoot VRAM too much or we'll take too long
-		if i > 1 && consumed > vram {
-			slog.Info("achieved target vram exhaustion", "count", i, "vram", format.HumanBytes2(vram), "models", format.HumanBytes2(consumed))
+		if i > 1 && consumed > maxVram {
+			slog.Info("achieved target vram exhaustion", "count", i, "vram", format.HumanBytes2(maxVram), "models", format.HumanBytes2(consumed))
 			break
 		}
 		consumed += chosenModels[i].size
-		slog.Info("target vram", "count", i, "vram", format.HumanBytes2(vram), "models", format.HumanBytes2(consumed))
+		slog.Info("target vram", "count", i, "vram", format.HumanBytes2(maxVram), "models", format.HumanBytes2(consumed))
 
 		wg.Add(1)
 		go func(i int) {

integration/embed_test.go

@@ -70,8 +70,8 @@ func TestAllMiniLMEmbed(t *testing.T) {
 		t.Fatalf("expected 0.010071031, got %.8f", res.Embeddings[0][0])
 	}
 
-	if res.PromptEvalCount != 8 {
-		t.Fatalf("expected 8 prompt tokens, got %d", res.PromptEvalCount)
+	if res.PromptEvalCount != 6 {
+		t.Fatalf("expected 6 prompt tokens, got %d", res.PromptEvalCount)
 	}
 }
 
@@ -102,8 +102,8 @@ func TestAllMiniLMBatchEmbed(t *testing.T) {
 		t.Fatalf("expected 0.010071031 and -0.009802706, got %.8f and %.8f", res.Embeddings[0][0], res.Embeddings[1][0])
 	}
 
-	if res.PromptEvalCount != 16 {
-		t.Fatalf("expected 16 prompt tokens, got %d", res.PromptEvalCount)
+	if res.PromptEvalCount != 12 {
+		t.Fatalf("expected 12 prompt tokens, got %d", res.PromptEvalCount)
 	}
 }
 

integration/llm_test.go

@@ -35,8 +35,8 @@ var (
 		},
 	}
 	resp = [2][]string{
-		[]string{"sunlight"},
-		[]string{"england", "english", "massachusetts", "pilgrims"},
+		{"sunlight"},
+		{"england", "english", "massachusetts", "pilgrims"},
 	}
 )
 

integration/max_queue_test.go

@@ -29,7 +29,7 @@ func TestMaxQueue(t *testing.T) {
 	// Also note that by default Darwin can't sustain > ~128 connections without adjusting limits
 	threadCount := 32
 	if maxQueue := envconfig.MaxQueue(); maxQueue != 0 {
-		threadCount = maxQueue
+		threadCount = int(maxQueue)
 	} else {
 		t.Setenv("OLLAMA_MAX_QUEUE", strconv.Itoa(threadCount))
 	}

integration/utils_test.go

@@ -162,7 +162,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 	fn := func(resp api.ProgressResponse) error {
 		// fmt.Print(".")
 		if !stallTimer.Reset(stallDuration) {
-			return fmt.Errorf("stall was detected, aborting status reporting")
+			return errors.New("stall was detected, aborting status reporting")
 		}
 		return nil
 	}
@@ -180,7 +180,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 
 	select {
 	case <-stallTimer.C:
-		return fmt.Errorf("download stalled")
+		return errors.New("download stalled")
 	case <-done:
 		return pullError
 	}
@@ -243,7 +243,7 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 		// fmt.Print(".")
 		buf.Write([]byte(response.Response))
 		if !stallTimer.Reset(streamTimeout) {
-			return fmt.Errorf("stall was detected while streaming response, aborting")
+			return errors.New("stall was detected while streaming response, aborting")
 		}
 		return nil
 	}
@@ -334,10 +334,10 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 			},
 		},
 		[][]string{
-			[]string{"sunlight"},
-			[]string{"soil", "organic", "earth", "black", "tan"},
-			[]string{"england", "english", "massachusetts", "pilgrims", "british"},
-			[]string{"fourth", "july", "declaration", "independence"},
-			[]string{"nitrogen", "oxygen", "carbon", "dioxide"},
+			{"sunlight"},
+			{"soil", "organic", "earth", "black", "tan"},
+			{"england", "english", "massachusetts", "pilgrims", "british"},
+			{"fourth", "july", "declaration", "independence"},
+			{"nitrogen", "oxygen", "carbon", "dioxide"},
 		}
 }

llm/ext_server/CMakeLists.txt

@@ -1,13 +1,14 @@
-set(TARGET ollama_llama_server)
-option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
-include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
-install(TARGETS ${TARGET} RUNTIME)
-target_compile_definitions(${TARGET} PRIVATE
-    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
-)
-target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT})
-if (WIN32)
-    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
-endif()
+set(TARGET ollama_llama_server)
+option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
+set(LLAMA_SERVER_LDFLAGS $ENV{LLAMA_SERVER_LDFLAGS})
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
+install(TARGETS ${TARGET} RUNTIME)
+target_compile_definitions(${TARGET} PRIVATE
+    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
+)
+target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT} ${LLAMA_SERVER_LDFLAGS})
+if (WIN32)
+    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
+endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

llm/ext_server/server.cpp

@@ -44,6 +44,7 @@
 #include <errhandlingapi.h>
 #endif
 
+#include <algorithm>
 #include <cstddef>
 #include <thread>
 #include <chrono>
@@ -402,7 +403,9 @@ struct llama_server_context
             }
         }
 
-        std::tie(model, ctx) = llama_init_from_gpt_params(params);
+        auto init_result = llama_init_from_gpt_params(params);
+        model = init_result.model;
+        ctx = init_result.context;
         if (model == nullptr)
         {
             LOG_ERROR("unable to load model", {{"model", params.model}});
@@ -1221,7 +1224,6 @@ struct llama_server_context
                 res.result_json = json
                 {
                     {"embedding", std::vector<float>(embd, embd + n_embd)},
-                    {"timings",             slot.get_formated_timings()},
                 };
             }
         }
@@ -1427,7 +1429,13 @@ struct llama_server_context
         switch (task.type)
         {
             case TASK_TYPE_COMPLETION: {
-                server_slot *slot = prefix_slot(task.data["prompt"]);
+                server_slot *slot = nullptr;
+                if (task.embedding_mode) {
+                    // Embedding seq_id (aka slot id) must always be <= token length, so always use slot 0
+                    slot = slots[0].available() ? &slots[0] : nullptr;
+                } else {
+                    slot = prefix_slot(task.data["prompt"]);
+                }
                 if (slot == nullptr)
                 {
                     // if no slot is available, we defer this task for processing later
@@ -2420,7 +2428,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter.emplace_back(argv[i], 1.0f);
+            params.lora_adapters.push_back({
+                std::string(argv[i]),
+                1.0,
+            });
             params.use_mmap = false;
         }
         else if (arg == "--lora-scaled")
@@ -2436,7 +2447,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
+            params.lora_adapters.push_back({
+                lora_adapter,
+                std::stof(argv[i])
+            });
             params.use_mmap = false;
         }
         else if (arg == "-v" || arg == "--verbose")
@@ -3184,37 +3198,17 @@ int main(int argc, char **argv) {
                     prompt = "";
                 }
 
-                if (prompt.size() == 1) {
-                    prompt = prompt[0];
-                }
-
                 // create and queue the task
-                json responses;
-                {
-                    const int id_task = llama.queue_tasks.get_new_id();
-                    llama.queue_results.add_waiting_task_id(id_task);
-                    llama.request_completion(id_task, {{"prompt", prompt}}, true, -1);
+                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}}, true, -1);
 
-                    // get the result
-                    task_result result = llama.queue_results.recv(id_task);
-                    llama.queue_results.remove_waiting_task_id(id_task);
-                    if (result.error) {
-                        return res.set_content(result.result_json.dump(), "application/json; charset=utf-8");
-                    }
+                // get the result
+                task_result result = llama.queue_results.recv(task_id);
+                llama.queue_results.remove_waiting_task_id(task_id);
 
-                    responses = result.result_json.value("results", std::vector<json>{result.result_json});
-                    json embeddings = json::array();
-
-                    int prompt_n = 0;
-                    for (auto & elem : responses) {
-                        embeddings.push_back(elem.at("embedding"));
-                        prompt_n += elem.at("timings").at("prompt_n").get<int>();
-                    }
-
-                    // send the result
-                    json embedding_res = json{{"embedding", embeddings}, {"prompt_n", prompt_n}};
-                    return res.set_content(embedding_res.dump(), "application/json; charset=utf-8");
-                }
+                // send the result
+                return res.set_content(result.result_json.dump(), "application/json; charset=utf-8");
             });
 
     // GG: if I put the main loop inside a thread, it crashes on the first request when build in Debug!?

llm/generate/gen_common.sh

@@ -9,11 +9,14 @@ init_vars() {
         ARCH="arm64"
         ;;
     *)
-        ARCH=$(uname -m | sed -e "s/aarch64/arm64/g")
+        echo "GOARCH must be set"
+        echo "this script is meant to be run from within go generate"
+        exit 1
+        ;;
     esac
 
     LLAMACPP_DIR=../llama.cpp
-    CMAKE_DEFS=""
+    CMAKE_DEFS="-DCMAKE_SKIP_RPATH=on"
     CMAKE_TARGETS="--target ollama_llama_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}"
@@ -27,6 +30,7 @@ init_vars() {
         WHOLE_ARCHIVE="-Wl,-force_load"
         NO_WHOLE_ARCHIVE=""
         GCC_ARCH="-arch ${ARCH}"
+        DIST_BASE=../../dist/darwin-${GOARCH}/
         ;;
     "Linux")
         LIB_EXT="so"
@@ -35,6 +39,7 @@ init_vars() {
 
         # Cross compiling not supported on linux - Use docker
         GCC_ARCH=""
+        DIST_BASE=../../dist/linux-${GOARCH}/
         ;;
     *)
         ;;
@@ -42,6 +47,7 @@ init_vars() {
     if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then
         CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
     fi
+    GZIP=$(which pigz 2>/dev/null || echo "gzip")
 }
 
 git_module_setup() {
@@ -85,26 +91,36 @@ build() {
 
 compress() {
     echo "Compressing payloads to reduce overall binary size..."
-    pids=""
     rm -rf ${BUILD_DIR}/bin/*.gz
     for f in ${BUILD_DIR}/bin/* ; do
-        gzip -n --best -f ${f} &
-        pids+=" $!"
+        ${GZIP} -n --best -f ${f} &
+        compress_pids+=" $!"
     done
     # check for lib directory
     if [ -d ${BUILD_DIR}/lib ]; then
         for f in ${BUILD_DIR}/lib/* ; do
-            gzip -n --best -f ${f} &
-            pids+=" $!"
+            ${GZIP} -n --best -f ${f} &
+            compress_pids+=" $!"
         done
     fi
     echo
-    for pid in ${pids}; do
+}
+
+wait_for_compress() {
+    for pid in ${compress_pids}; do
         wait $pid
     done
     echo "Finished compression"
 }
 
+install() {
+    echo "Installing libraries to bin dir ${BUILD_DIR}/bin/"
+    for lib in $(find ${BUILD_DIR} -name \*.${LIB_EXT}); do
+        rm -f "${BUILD_DIR}/bin/$(basename ${lib})"
+        cp -af "${lib}" "${BUILD_DIR}/bin/"
+    done
+}
+
 # Keep the local tree clean after we're done with the build
 cleanup() {
     (cd ${LLAMACPP_DIR}/ && git checkout CMakeLists.txt)

llm/generate/gen_darwin.sh

@@ -6,6 +6,7 @@
 
 set -ex
 set -o pipefail
+compress_pids=""
 echo "Starting darwin generate script"
 source $(dirname $0)/gen_common.sh
 init_vars
@@ -98,4 +99,5 @@ case "${GOARCH}" in
 esac
 
 cleanup
+wait_for_compress
 echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"

llm/generate/gen_linux.sh

@@ -13,6 +13,7 @@
 
 set -ex
 set -o pipefail
+compress_pids=""
 
 # See https://llvm.org/docs/AMDGPUUsage.html#processors for reference
 amdGPUs() {
@@ -51,7 +52,7 @@ if [ -z "${CUDACXX}" ]; then
         export CUDACXX=$(command -v nvcc)
     fi
 fi
-COMMON_CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off"
+COMMON_CMAKE_DEFS="-DCMAKE_SKIP_RPATH=on -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off"
 source $(dirname $0)/gen_common.sh
 init_vars
 git_module_setup
@@ -77,10 +78,11 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
     if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
         init_vars
         echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
-        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
         BUILD_DIR="../build/linux/${ARCH}/cpu"
         echo "Building custom CPU"
         build
+        install
         compress
     else
         # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
@@ -93,7 +95,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
         # -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
         # -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake
 
-        COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
+        COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=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)
@@ -103,6 +105,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
             BUILD_DIR="../build/linux/${ARCH}/cpu"
             echo "Building LCD CPU"
             build
+            install
             compress
         fi
 
@@ -120,6 +123,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                 BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
                 echo "Building AVX CPU"
                 build
+                install
                 compress
             fi
 
@@ -133,6 +137,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                 BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
                 echo "Building AVX2 CPU"
                 build
+                install
                 compress
             fi
         fi
@@ -160,7 +165,7 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
     echo "CUDA libraries detected - building dynamic CUDA library"
     init_vars
     CUDA_MAJOR=$(ls "${CUDA_LIB_DIR}"/libcudart.so.* | head -1 | cut -f3 -d. || true)
-    if [ -n "${CUDA_MAJOR}" ]; then
+    if [ -n "${CUDA_MAJOR}" -a -z "${CUDA_VARIANT}" ]; then
         CUDA_VARIANT=_v${CUDA_MAJOR}
     fi
     if [ "${ARCH}" == "arm64" ]; then
@@ -178,29 +183,19 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
         CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
         echo "Building custom CUDA GPU"
     else
-        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
+        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
     fi
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
+    export CUDAFLAGS="-t8"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS} -DGGML_STATIC=off"
     BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
-    EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
+    export LLAMA_SERVER_LDFLAGS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
+    CUDA_DIST_DIR="${CUDA_DIST_DIR:-${DIST_BASE}/lib/ollama}"
     build
-
-    # Carry the CUDA libs as payloads to help reduce dependency burden on users
-    #
-    # TODO - in the future we may shift to packaging these separately and conditionally
-    #        downloading them in the install script.
-    DEPS="$(ldd ${BUILD_DIR}/bin/ollama_llama_server )"
-    for lib in libcudart.so libcublas.so libcublasLt.so ; do
-        DEP=$(echo "${DEPS}" | grep ${lib} | cut -f1 -d' ' | xargs || true)
-        if [ -n "${DEP}" -a -e "${CUDA_LIB_DIR}/${DEP}" ]; then
-            cp "${CUDA_LIB_DIR}/${DEP}" "${BUILD_DIR}/bin/"
-        elif [ -e "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" ]; then
-            cp "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" "${BUILD_DIR}/bin/"
-        elif [ -e "${CUDART_LIB_DIR}/${lib}" ]; then
-            cp -d ${CUDART_LIB_DIR}/${lib}* "${BUILD_DIR}/bin/"
-        else
-            cp -d "${CUDA_LIB_DIR}/${lib}*" "${BUILD_DIR}/bin/"
-        fi
+    install
+    echo "Installing CUDA dependencies in ${CUDA_DIST_DIR}"
+    mkdir -p "${CUDA_DIST_DIR}"
+    for lib in ${CUDA_LIB_DIR}/libcudart.so* ${CUDA_LIB_DIR}/libcublas.so* ${CUDA_LIB_DIR}/libcublasLt.so* ; do
+        cp -a "${lib}" "${CUDA_DIST_DIR}"
     done
     compress
 
@@ -218,21 +213,24 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
     CC=icx
     CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
     BUILD_DIR="../build/linux/${ARCH}/oneapi"
-    EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
+    ONEAPI_DIST_DIR="${DIST_BASE}/lib/ollama"
+    export LLAMA_SERVER_LDFLAGS="-fsycl -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
     DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
     build
 
     # copy oneAPI dependencies
+    mkdir -p "${ONEAPI_DIST_DIR}"
     for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e sycl -e mkl -e tbb); do
-        cp "${dep}" "${BUILD_DIR}/bin/"
+        cp -a "${dep}" "${ONEAPI_DIST_DIR}"
     done
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libOpenCL.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libimf.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libintlc.so.5" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libirng.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libpi_level_zero.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libOpenCL.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libimf.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libintlc.so.5" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libirng.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libpi_level_zero.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${ONEAPI_DIST_DIR}"
+    install
     compress
 fi
 
@@ -254,7 +252,7 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
         ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
     fi
     init_vars
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DLLAMA_CUDA_NO_PEER_COPY=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DGGML_CUDA_NO_PEER_COPY=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
     if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
@@ -262,23 +260,22 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
         echo "Building custom ROCM GPU"
     fi
     BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
-    EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
+    # ROCm dependencies are too large to fit into a unified bundle
+    ROCM_DIST_DIR="${DIST_BASE}/../linux-${GOARCH}-rocm/lib/ollama"
+    # TODO figure out how to disable runpath (rpath)
+    # export CMAKE_HIP_FLAGS="-fno-rtlib-add-rpath" # doesn't work
+    export LLAMA_SERVER_LDFLAGS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
     build
 
-    # Record the ROCM dependencies
-    rm -f "${BUILD_DIR}/bin/deps.txt"
-    touch "${BUILD_DIR}/bin/deps.txt"
-    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e rocm -e amdgpu -e libtinfo ); do
-        echo "${dep}" >> "${BUILD_DIR}/bin/deps.txt"
+    # copy the ROCM dependencies
+    mkdir -p "${ROCM_DIST_DIR}"
+    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -v "${ARCH}/rocm${ROCM_VARIANT}" | grep -e rocm -e amdgpu -e libtinfo ); do
+        cp -a "${dep}"* "${ROCM_DIST_DIR}"
     done
-    # bomb out if for some reason we didn't get a few deps
-    if [ $(cat "${BUILD_DIR}/bin/deps.txt" | wc -l ) -lt 8 ] ; then
-        cat "${BUILD_DIR}/bin/deps.txt"
-        echo "ERROR: deps file short"
-        exit 1
-    fi
+    install
     compress
 fi
 
 cleanup
+wait_for_compress
 echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"

llm/generate/gen_windows.ps1

@@ -35,7 +35,7 @@ function init_vars {
         )
     $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
     $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
-    $script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\ollama_runners"
+    $script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\runners"
     md "$script:DIST_BASE" -ea 0 > $null
     if ($env:CGO_CFLAGS -contains "-g") {
         $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on", "-DCMAKE_BUILD_TYPE=RelWithDebInfo")
@@ -117,7 +117,7 @@ function build {
     if ($cmakeDefs -contains "-G") {
         $extra=@("-j8")
     } else {
-        $extra= @("--", "/p:CL_MPcount=8")
+        $extra= @("--", "/maxCpuCount:8")
     }
     write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ }) $extra"
     & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ }) $extra
@@ -261,7 +261,7 @@ function build_cuda() {
     if ((-not "${env:OLLAMA_SKIP_CUDA_GENERATE}") -and ("${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
+        $script:CUDA_VERSION=((get-item ($nvcc | split-path | split-path)).Basename -Split "\.")[0]
         if ($null -ne $script:CUDA_VERSION) {
             $script:CUDA_VARIANT="_"+$script:CUDA_VERSION
         }
@@ -273,9 +273,9 @@ function build_cuda() {
             "-DGGML_CUDA=ON",
             "-DGGML_AVX=on",
             "-DGGML_AVX2=off",
-            "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR",
-            "-DCMAKE_CUDA_FLAGS=-t8",
-            "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}"
+            "-DCMAKE_CUDA_FLAGS=-t6",
+            "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}",
+            "-DCMAKE_CUDA_COMPILER_TOOLKIT_ROOT=$env:CUDA_PATH"
             )
         if ($null -ne $env:OLLAMA_CUSTOM_CUDA_DEFS) {
             write-host "OLLAMA_CUSTOM_CUDA_DEFS=`"${env:OLLAMA_CUSTOM_CUDA_DEFS}`""
@@ -286,12 +286,11 @@ function build_cuda() {
         sign
         install
 
-        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\" -ea 0 > $null
-        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
+        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\" -ea 0 > $null
+        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
     } else {
         write-host "Skipping CUDA generation step"
     }
@@ -325,18 +324,17 @@ function build_oneapi() {
     sign
     install
 
-    rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    md "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\" -ea 0 > $null
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\" -ea 0 > $null
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
   } else {
     Write-Host "Skipping oneAPI generation step"
   }
@@ -357,7 +355,7 @@ function build_rocm() {
             "-DCMAKE_C_COMPILER=clang.exe",
             "-DCMAKE_CXX_COMPILER=clang++.exe",
             "-DGGML_HIPBLAS=on",
-            "-DLLAMA_CUDA_NO_PEER_COPY=on",
+            "-DGGML_CUDA_NO_PEER_COPY=on",
             "-DHIP_PLATFORM=amd",
             "-DGGML_AVX=on",
             "-DGGML_AVX2=off",
@@ -386,12 +384,11 @@ function build_rocm() {
         sign
         install
 
-        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
-        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\" -ea 0 > $null
-        cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
-        cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\rocblas\library\" -ea 0 > $null
+        cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
         # amdhip64.dll dependency comes from the driver and must be installed on the host to use AMD GPUs
-        cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\"
+        cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\rocblas\library\"
     } else {
         write-host "Skipping ROCm generation step"
     }

llm/ggml.go

@@ -43,6 +43,14 @@ func (kv KV) Architecture() string {
 	return "unknown"
 }
 
+func (kv KV) Kind() string {
+	if s, ok := kv["general.type"].(string); ok {
+		return s
+	}
+
+	return "unknown"
+}
+
 func (kv KV) ParameterCount() uint64 {
 	return kv.u64("general.parameter_count")
 }
@@ -157,6 +165,14 @@ type Tensor struct {
 	io.WriterTo `json:"-"`
 }
 
+func (t Tensor) block() (n int) {
+	if _, err := fmt.Sscanf(t.Name, "blk.%d.", &n); err != nil {
+		return -1
+	}
+
+	return
+}
+
 func (t Tensor) blockSize() uint64 {
 	switch t.Kind {
 	case 0, 1, 24, 25, 26, 27, 28, 30: // F32, F16, I8, I16, I32, I64, F64, BF16

llm/gguf.go

@@ -532,15 +532,14 @@ func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
 		}
 	}
 
-	slices.SortFunc(ts, func(a, b Tensor) int {
-		var i, j int
-		if n, err := fmt.Sscanf(a.Name, "blk.%d", &i); err != nil || n != 1 {
-			return cmp.Compare(a.Name, b.Name)
-		} else if n, err := fmt.Sscanf(b.Name, "blk.%d", &j); err != nil || n != 1 {
-			return cmp.Compare(a.Name, b.Name)
+	slices.SortStableFunc(ts, func(a, b Tensor) int {
+		if i, j := a.block(), b.block(); i < 0 && j > 0 {
+			return 1
+		} else if i > 0 && j < 0 {
+			return -1
+		} else {
+			return cmp.Compare(i, j)
 		}
-
-		return cmp.Compare(i, j)
 	})
 
 	var s uint64

llm/llm.go

@@ -11,8 +11,9 @@ package llm
 // #include <stdlib.h>
 // #include "llama.h"
 import "C"
+
 import (
-	"fmt"
+	"errors"
 	"unsafe"
 )
 
@@ -33,7 +34,7 @@ func Quantize(infile, outfile string, ftype fileType) error {
 	params.ftype = ftype.Value()
 
 	if rc := C.llama_model_quantize(cinfile, coutfile, &params); rc != 0 {
-		return fmt.Errorf("failed to quantize model. This model architecture may not be supported, or you may need to upgrade Ollama to the latest version")
+		return errors.New("failed to quantize model. This model architecture may not be supported, or you may need to upgrade Ollama to the latest version")
 	}
 
 	return nil

llm/memory_test.go

@@ -6,10 +6,11 @@ import (
 	"os"
 	"testing"
 
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/gpu"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/gpu"
 )
 
 func TestEstimateGPULayers(t *testing.T) {
@@ -32,7 +33,6 @@ func TestEstimateGPULayers(t *testing.T) {
 	assert.Len(t, tensors, inputLayerCount+1)
 	err = WriteGGUF(f, KV{
 		"general.architecture":          "llama",
-		"general.name":                  "name",
 		"llama.context_length":          uint32(32),
 		"llama.embedding_length":        uint32(4096),
 		"llama.block_count":             uint32(inputLayerCount),

llm/patches/08-pooling.diff

@@ -1,60 +0,0 @@
-diff --git a/src/llama.cpp b/src/llama.cpp
-index 721b8f4e..cfe7ac40 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -8420,14 +8420,14 @@ struct llm_build_context {
-     }
- 
-     struct ggml_tensor * build_inp_mean() {
--        lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens);
-+        lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, cparams.n_seq_max);
-         cb(lctx.inp_mean, "inp_mean", -1);
-         ggml_set_input(lctx.inp_mean);
-         return lctx.inp_mean;
-     }
- 
-     struct ggml_tensor * build_inp_cls() {
--        lctx.inp_cls = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
-+        lctx.inp_cls = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, cparams.n_seq_max);
-         cb(lctx.inp_cls, "inp_cls", -1);
-         ggml_set_input(lctx.inp_cls);
-         return lctx.inp_cls;
-@@ -13847,19 +13847,16 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
-         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer));
- 
-         float * data = (float *) lctx.inp_mean->data;
--        memset(lctx.inp_mean->data, 0, n_tokens * n_tokens * ggml_element_size(lctx.inp_mean));
-+        memset(lctx.inp_mean->data, 0, n_tokens * cparams.n_seq_max * ggml_element_size(lctx.inp_mean));
- 
-         std::vector<uint64_t> sum(n_tokens, 0);
-         for (int i = 0; i < n_tokens; ++i) {
-             const llama_seq_id seq_id = batch.seq_id[i][0];
--
--            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN");
--
-             sum[seq_id] += 1;
-         }
- 
--        std::vector<float> div(n_tokens, 0.0f);
--        for (int i = 0; i < n_tokens; ++i) {
-+        std::vector<float> div(cparams.n_seq_max, 0.0f);
-+        for (uint32_t i = 0; i < cparams.n_seq_max; ++i) {
-             const uint64_t s = sum[i];
-             if (s > 0) {
-                 div[i] = 1.0f/float(s);
-@@ -13879,14 +13876,11 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
-         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
- 
-         uint32_t * data = (uint32_t *) lctx.inp_cls->data;
--        memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls));
-+        memset(lctx.inp_cls->data, 0, cparams.n_seq_max * ggml_element_size(lctx.inp_cls));
- 
-         for (int i = 0; i < n_tokens; ++i) {
-             const llama_seq_id seq_id = batch.seq_id[i][0];
-             const llama_pos    pos    = batch.pos[i];
--
--            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS");
--
-             if (pos == 0) {
-                 data[seq_id] = i;
-             }

llm/patches/09-lora.diff

@@ -1,40 +1,32 @@
 diff --git a/common/common.cpp b/common/common.cpp
-index dbb724fb..c26fe6ee 100644
+index 2e8374d5..70d0afde 100644
 --- a/common/common.cpp
 +++ b/common/common.cpp
-@@ -2087,14 +2087,27 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
-     for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
-         const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
-         float lora_scale = std::get<1>(params.lora_adapter[i]);
-+
-+        // try to load as gguf
-         auto adapter = llama_lora_adapter_init(model, lora_adapter.c_str());
-         if (adapter == nullptr) {
--            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+@@ -2110,9 +2110,21 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
+         loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
+         if (loaded_la.adapter == nullptr) {
+             fprintf(stderr, "%s: error: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
 -            llama_free(lctx);
 -            llama_free_model(model);
--            return std::make_tuple(nullptr, nullptr);
-+            fprintf(stderr, "%s: error: failed to apply lora adapter, trying ggla\n", __func__);
+-            return iparams;
 +
 +            // if that fails, try loading as ggla for compatibility
 +            int err = llama_model_apply_lora_from_file(model,
-+                                                    lora_adapter.c_str(),
-+                                                    lora_scale,
++                                                    la.path.c_str(),
++                                                    la.scale,
 +                                                    nullptr,
 +                                                    params.n_threads);
 +            if (err != 0) {
 +                fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
 +                llama_free(lctx);
 +                llama_free_model(model);
-+                return std::make_tuple(nullptr, nullptr);
++                return iparams;
++            } else {
++                break;
 +            }
-+        } else {
-+            llama_lora_adapter_set(lctx, adapter, lora_scale);
          }
--        llama_lora_adapter_set(lctx, adapter, lora_scale);
+         iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
      }
- 
-     if (params.ignore_eos) {
 diff --git a/include/llama.h b/include/llama.h
 index 93fd77ca..b0fb37a6 100644
 --- a/include/llama.h
@@ -355,4 +347,4 @@ index 80a0dd0f..9d7b0e17 100644
 +        return 1;
 +    }
 +}
-\ No newline at end of file
+\ No newline at end of file

llm/patches/10-params.diff

@@ -1,20 +0,0 @@
-diff --git a/src/llama.cpp b/src/llama.cpp
-index a207451f..fba6b175 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -4969,6 +4969,7 @@ static void llm_load_hparams(
-                 hparams.attn_soft_cap = true;
- 
-                 switch (hparams.n_layer) {
-+                    case 26: model.type = e_model::MODEL_2B; break;
-                     case 42: model.type = e_model::MODEL_9B; break;
-                     case 46: model.type = e_model::MODEL_27B; break;
-                     default: model.type = e_model::MODEL_UNKNOWN;
-@@ -11736,6 +11737,7 @@ struct llm_build_context {
- 
-                 // ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e
-                 switch (model.type) {
-+                    case e_model::MODEL_2B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); break;
-                     case e_model::MODEL_9B:  Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));   break;
-                     case e_model::MODEL_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break;
-                     default: GGML_ABORT("fatal error");

llm/payload.go

@@ -82,8 +82,8 @@ func serversForGpu(info gpu.GpuInfo) []string {
 	// glob workDir for files that start with ollama_
 	availableServers := getAvailableServers()
 	requested := info.Library
-	if info.Variant != gpu.CPUCapabilityNone {
-		requested += "_" + info.Variant.String()
+	if info.Variant != gpu.CPUCapabilityNone.String() {
+		requested += "_" + info.Variant
 	}
 
 	servers := []string{}

llm/server.go

@@ -33,7 +33,7 @@ type LlamaServer interface {
 	Ping(ctx context.Context) error
 	WaitUntilRunning(ctx context.Context) error
 	Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error
-	Embed(ctx context.Context, input []string) (*EmbedResponse, error)
+	Embedding(ctx context.Context, input string) ([]float32, error)
 	Tokenize(ctx context.Context, content string) ([]int, error)
 	Detokenize(ctx context.Context, tokens []int) (string, error)
 	Close() error
@@ -44,11 +44,12 @@ type LlamaServer interface {
 
 // llmServer is an instance of the llama.cpp server
 type llmServer struct {
-	port    int
-	cmd     *exec.Cmd
-	done    chan error // Channel to signal when the process exits
-	status  *StatusWriter
-	options api.Options
+	port        int
+	cmd         *exec.Cmd
+	done        chan error // Channel to signal when the process exits
+	status      *StatusWriter
+	options     api.Options
+	numParallel int
 
 	estimate    MemoryEstimate
 	totalLayers uint64
@@ -124,8 +125,9 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 	}
 
-	// On linux, over-allocating CPU memory will almost always result in an error
-	if runtime.GOOS == "linux" {
+	// On linux and windows, over-allocating CPU memory will almost always result in an error
+	// Darwin has fully dynamic swap so has no direct concept of free swap space
+	if runtime.GOOS != "darwin" {
 		systemMemoryRequired := estimate.TotalSize - estimate.VRAMSize
 		available := systemFreeMemory + systemSwapFreeMemory
 		if systemMemoryRequired > available {
@@ -184,15 +186,15 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 
 	params := []string{
 		"--model", model,
-		"--ctx-size", fmt.Sprintf("%d", opts.NumCtx),
-		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
+		"--ctx-size", strconv.Itoa(opts.NumCtx),
+		"--batch-size", strconv.Itoa(opts.NumBatch),
 		"--embedding",
 	}
 
 	params = append(params, "--log-disable")
 
 	if opts.NumGPU >= 0 {
-		params = append(params, "--n-gpu-layers", fmt.Sprintf("%d", opts.NumGPU))
+		params = append(params, "--n-gpu-layers", strconv.Itoa(opts.NumGPU))
 	}
 
 	if envconfig.Debug() {
@@ -200,7 +202,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	}
 
 	if opts.MainGPU > 0 {
-		params = append(params, "--main-gpu", fmt.Sprintf("%d", opts.MainGPU))
+		params = append(params, "--main-gpu", strconv.Itoa(opts.MainGPU))
 	}
 
 	if len(adapters) > 0 {
@@ -214,7 +216,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	}
 
 	if opts.NumThread > 0 {
-		params = append(params, "--threads", fmt.Sprintf("%d", opts.NumThread))
+		params = append(params, "--threads", strconv.Itoa(opts.NumThread))
 	}
 
 	if !opts.F16KV {
@@ -256,11 +258,17 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--mlock")
 	}
 
-	if opts.UseNUMA {
-		params = append(params, "--numa")
+	if gpu.IsNUMA() && gpus[0].Library == "cpu" {
+		numaMode := "distribute"
+		if runtime.GOOS == "linux" {
+			if _, err := exec.LookPath("numactl"); err == nil {
+				numaMode = "numactl"
+			}
+		}
+		params = append(params, "--numa", numaMode)
 	}
 
-	params = append(params, "--parallel", fmt.Sprintf("%d", numParallel))
+	params = append(params, "--parallel", strconv.Itoa(numParallel))
 
 	if estimate.TensorSplit != "" {
 		params = append(params, "--tensor-split", estimate.TensorSplit)
@@ -298,20 +306,18 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		if runtime.GOOS == "windows" {
 			pathEnv = "PATH"
 		}
-		// prepend the server directory to LD_LIBRARY_PATH/PATH and the parent dir for common dependencies
-		libraryPaths := []string{dir, filepath.Dir(dir)}
+		// Start with the server directory for the LD_LIBRARY_PATH/PATH
+		libraryPaths := []string{dir}
 
 		if libraryPath, ok := os.LookupEnv(pathEnv); ok {
-			// Append our runner directory to the path
-			// This will favor system libraries over our bundled library dependencies
+			// favor our bundled library dependencies over system libraries
 			libraryPaths = append(libraryPaths, filepath.SplitList(libraryPath)...)
 		}
 
 		// Note: we always put the dependency path first
-		// since this was the exact version we verified for AMD GPUs
-		// and we favor what the user had in their path
+		// since this was the exact version we compiled/linked against
 		if gpus[0].DependencyPath != "" {
-			// TODO refine for multi-gpu support
+			// assume gpus from the same library have the same dependency path
 			libraryPaths = append([]string{gpus[0].DependencyPath}, libraryPaths...)
 		}
 
@@ -337,6 +343,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			status:      NewStatusWriter(os.Stderr),
 			options:     opts,
 			estimate:    estimate,
+			numParallel: numParallel,
 			sem:         semaphore.NewWeighted(int64(numParallel)),
 			totalLayers: ggml.KV().BlockCount() + 1,
 			gpus:        gpus,
@@ -425,7 +432,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 				if strings.Contains(s.status.LastErrMsg, "unknown model") {
 					s.status.LastErrMsg = "this model is not supported by your version of Ollama. You may need to upgrade"
 				}
-				s.done <- fmt.Errorf(s.status.LastErrMsg)
+				s.done <- errors.New(s.status.LastErrMsg)
 			} else {
 				s.done <- err
 			}
@@ -874,16 +881,15 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 	return nil
 }
 
-type EmbedRequest struct {
-	Content []string `json:"content"`
+type EmbeddingRequest struct {
+	Content string `json:"content"`
 }
 
-type EmbedResponse struct {
-	Embedding       [][]float32 `json:"embedding"`
-	PromptEvalCount int         `json:"prompt_n"`
+type EmbeddingResponse struct {
+	Embedding []float32 `json:"embedding"`
 }
 
-func (s *llmServer) Embed(ctx context.Context, input []string) (*EmbedResponse, error) {
+func (s *llmServer) Embedding(ctx context.Context, input string) ([]float32, error) {
 	if err := s.sem.Acquire(ctx, 1); err != nil {
 		slog.Error("Failed to acquire semaphore", "error", err)
 		return nil, err
@@ -898,18 +904,18 @@ func (s *llmServer) Embed(ctx context.Context, input []string) (*EmbedResponse,
 		return nil, fmt.Errorf("unexpected server status: %s", status.ToString())
 	}
 
-	data, err := json.Marshal(EmbedRequest{Content: input})
+	data, err := json.Marshal(EmbeddingRequest{Content: input})
 	if err != nil {
 		return nil, fmt.Errorf("error marshaling embed data: %w", err)
 	}
 
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), bytes.NewBuffer(data))
+	r, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), bytes.NewBuffer(data))
 	if err != nil {
 		return nil, fmt.Errorf("error creating embed request: %w", err)
 	}
-	req.Header.Set("Content-Type", "application/json")
+	r.Header.Set("Content-Type", "application/json")
 
-	resp, err := http.DefaultClient.Do(req)
+	resp, err := http.DefaultClient.Do(r)
 	if err != nil {
 		return nil, fmt.Errorf("do embedding request: %w", err)
 	}
@@ -925,12 +931,12 @@ func (s *llmServer) Embed(ctx context.Context, input []string) (*EmbedResponse,
 		return nil, fmt.Errorf("%s", body)
 	}
 
-	var e EmbedResponse
+	var e EmbeddingResponse
 	if err := json.Unmarshal(body, &e); err != nil {
 		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
 	}
 
-	return &e, nil
+	return e.Embedding, nil
 }
 
 type TokenizeRequest struct {

llm/status.go

@@ -26,6 +26,7 @@ var errorPrefixes = []string{
 	"cudaMalloc failed",
 	"\"ERR\"",
 	"error loading model",
+	"GGML_ASSERT",
 }
 
 func (w *StatusWriter) Write(b []byte) (int, error) {