Merge https://github.com/ollama/ollama

Signed-off-by: baalajimaestro <baalajimaestro@ptr.moe>
2024-09-15 23:49:24 +05:30 · 2024-09-15 23:49:24 +05:30 · 7f1565721c
commit 7f1565721c
parent 76c9dc57fd d889c6fd07
51 changed files with 1383 additions and 868 deletions
--- a/.dockerignore
+++ b/.dockerignore
@ -7,3 +7,5 @@ llm/llama.cpp
 .env
 .cache
 test_data
+llm/build
+llama/build
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@ -102,8 +102,8 @@ jobs:
        with:
          name: generate-windows-cpu
          path: |
-            llm/build/**/bin/*
-            llm/build/**/*.a
+            build/**/*
+            build/**/*.a
            dist/windows-amd64/**

  # ROCm generation step
@ -176,7 +176,7 @@ jobs:
        with:
          name: generate-windows-rocm
          path: |
-            llm/build/**/bin/*
+            build/**/*
            dist/windows-amd64/**
      - uses: actions/upload-artifact@v4
        with:
@ -265,7 +265,7 @@ jobs:
        with:
          name: generate-windows-cuda-${{ matrix.cuda.version }}
          path: |
-            llm/build/**/bin/*
+            build/**/*
            dist/windows-amd64/**
      - uses: actions/upload-artifact@v4
        with:
@ -338,7 +338,7 @@ jobs:
      - uses: actions/download-artifact@v4
        with:
          name: generate-windows-rocm
-      - run: dir llm/build
+      - run: dir build
      - run: |
          $gopath=(get-command go).source | split-path -parent
          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
@ -359,9 +359,7 @@ jobs:
    environment: release
    runs-on: linux
    env:
-      OLLAMA_SKIP_MANIFEST_CREATE: '1'
      BUILD_ARCH: amd64
-      PUSH: '1'
    steps:
      - uses: actions/checkout@v4
        with:
@ -369,14 +367,8 @@ jobs:
      - name: Set Version
        shell: bash
        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
-      - name: Login to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ vars.DOCKER_USER }}
-          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
      - run: |
          ./scripts/build_linux.sh
-          ./scripts/build_docker.sh
      - uses: actions/upload-artifact@v4
        with:
          name: dist-linux-amd64
@ -390,9 +382,7 @@ jobs:
    environment: release
    runs-on: linux-arm64
    env:
-      OLLAMA_SKIP_MANIFEST_CREATE: '1'
      BUILD_ARCH: arm64
-      PUSH: '1'
    steps:
      - uses: actions/checkout@v4
        with:
@ -421,14 +411,8 @@ jobs:
          sudo usermod -aG docker $USER
          sudo apt-get install acl
          sudo setfacl --modify user:$USER:rw /var/run/docker.sock
-      - name: Login to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ vars.DOCKER_USER }}
-          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
      - run: |
          ./scripts/build_linux.sh
-          ./scripts/build_docker.sh
      - uses: actions/upload-artifact@v4
        with:
          name: dist-linux-arm64
@ -436,6 +420,181 @@ jobs:
            dist/*linux*
            !dist/*-cov

+  # Container image build
+  build-linux:
+    environment: release
+    strategy:
+      matrix:
+        runner:
+          - linux
+          - linux-arm64
+    runs-on: ${{ matrix.runner }}
+    env:
+      FINAL_IMAGE_REPO: ollama/ollama
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: 'Install Docker'
+        if: ${{ startsWith(matrix.runner, 'linux-arm64') }}
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y ca-certificates curl
+          sudo install -m 0755 -d /etc/apt/keyrings
+          sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
+          sudo chmod a+r /etc/apt/keyrings/docker.asc
+          echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/ubuntu \
+            $(. /etc/os-release && echo "$VERSION_CODENAME") stable" | \
+            sudo tee /etc/apt/sources.list.d/docker.list > /dev/null
+          sudo apt-get update
+          sudo apt-get install -y docker-ce docker-ce-cli containerd.io
+          sudo usermod -aG docker $USER
+          sudo apt-get install acl
+          sudo setfacl --modify user:$USER:rw /var/run/docker.sock
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.FINAL_IMAGE_REPO }}
+          flavor: |
+            latest=false
+          tags: |
+            type=ref,event=tag
+            type=ref,enable=true,priority=600,prefix=0.0.0-pr,suffix=,event=pr
+            type=semver,pattern={{version}}
+      - name: Set Version
+        shell: bash
+        run: |
+          machine=$(uname -m)
+          case ${machine} in
+            x86_64) echo ARCH=amd64; echo PLATFORM_PAIR=linux-amd64 ;;
+            aarch64) echo ARCH=arm64; echo PLATFORM_PAIR=linux-arm64 ;;
+          esac >>$GITHUB_ENV
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${{ env.DOCKER_METADATA_OUTPUT_VERSION }}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_ENV
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - name: Build and push by digest
+        id: build
+        uses: docker/build-push-action@v6
+        with:
+          context: "."
+          platforms: linux/${{ env.ARCH }}
+          build-args: |
+            GOFLAGS
+          outputs: type=image,name=${{ env.FINAL_IMAGE_REPO }},push-by-digest=true,name-canonical=true,push=true
+      - name: Export digest
+        run: |
+          mkdir -p /tmp/digests
+          digest="${{ steps.build.outputs.digest }}"
+          touch "/tmp/digests/${digest#sha256:}"
+      - name: Upload digest
+        uses: actions/upload-artifact@v4
+        with:
+          name: digests-${{ env.PLATFORM_PAIR }}
+          path: /tmp/digests/*
+          if-no-files-found: error
+          retention-days: 1
+  merge:
+    environment: release
+    runs-on: linux
+    needs:
+      - build-linux
+    env:
+      FINAL_IMAGE_REPO: ollama/ollama
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Download digests
+        uses: actions/download-artifact@v4
+        with:
+          path: /tmp/digests
+          pattern: digests-*
+          merge-multiple: true
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.FINAL_IMAGE_REPO }}
+          flavor: |
+            latest=false
+          tags: |
+            type=ref,event=tag
+            type=ref,enable=true,priority=600,prefix=0.0.0-pr,suffix=,event=pr
+            type=semver,pattern={{version}}
+      - name: Set Version
+        shell: bash
+        run: |
+          machine=$(uname -m)
+          case ${machine} in
+            x86_64) echo ARCH=amd64; echo PLATFORM_PAIR=linux-amd64 ;;
+            aarch64) echo ARCH=arm64; echo PLATFORM_PAIR=linux-arm64 ;;
+          esac >>$GITHUB_ENV
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${{ env.DOCKER_METADATA_OUTPUT_VERSION }}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_ENV
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - name: Create manifest list and push
+        working-directory: /tmp/digests
+        run: |
+          docker buildx imagetools create $(jq -cr '.tags | map("-t " + .) | join(" ")' <<< "$DOCKER_METADATA_OUTPUT_JSON") \
+            $(printf '${{ env.FINAL_IMAGE_REPO }}@sha256:%s ' *)
+      - name: Inspect image
+        run: |
+          docker buildx imagetools inspect ${{ env.FINAL_IMAGE_REPO }}:${{ steps.meta.outputs.version }}          
+  build-linux-rocm:
+    environment: release
+    runs-on: linux
+    env:
+      FINAL_IMAGE_REPO: ollama/ollama
+      ARCH: amd64
+      PLATFORM_PAIR: linux-amd64
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.FINAL_IMAGE_REPO }}
+          flavor: |
+            latest=false
+          tags: |
+            type=ref,event=tag
+            type=ref,enable=true,priority=600,prefix=0.0.0-pr,suffix=,event=pr
+            type=semver,pattern={{version}}
+      - name: Set Version
+        shell: bash
+        run: |
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${{ env.DOCKER_METADATA_OUTPUT_VERSION }}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_ENV
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - name: Build and push by digest
+        id: build
+        uses: docker/build-push-action@v6
+        with:
+          context: "."
+          target: runtime-rocm
+          build-args: |
+            GOFLAGS
+          tags: ${{ env.FINAL_IMAGE_REPO }}:${{ env.DOCKER_METADATA_OUTPUT_VERSION}}-rocm,${{ env.FINAL_IMAGE_REPO }}:rocm
+          push: true
+
  # Aggregate all the assets and ship a release
  release:
    needs:
@ -448,8 +607,6 @@ jobs:
    permissions:
      contents: write
    env:
-      OLLAMA_SKIP_IMAGE_BUILD: '1'
-      PUSH: '1'
      GH_TOKEN: ${{ github.token }}
    steps:
      - uses: actions/checkout@v4
@ -458,12 +615,6 @@ jobs:
        run: |
          echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
          echo "RELEASE_VERSION=$(echo ${GITHUB_REF_NAME} | cut -f1 -d-)" >> $GITHUB_ENV
-      - name: Login to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ vars.DOCKER_USER }}
-          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
-      - run: ./scripts/build_docker.sh
      - name: Retrieve built artifact
        uses: actions/download-artifact@v4
        with:
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@ -81,12 +81,6 @@ jobs:
        if: ${{ ! startsWith(matrix.os, 'windows-') }}
        name: 'Unix Go Generate'
      - run: go build .
-      - uses: actions/upload-artifact@v4
-        with:
-          name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
-          path: |
-            llm/build/**/bin/*
-            llm/build/**/*.a
  generate-cuda:
    needs: [changes]
    if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
@ -114,12 +108,6 @@ jobs:
          go generate -x ./...
        env:
          OLLAMA_SKIP_CPU_GENERATE: '1'
-      - uses: actions/upload-artifact@v4
-        with:
-          name: cuda-${{ matrix.cuda-version }}-libraries
-          path: |
-            llm/build/**/bin/*
-            dist/windows-amd64/**
  generate-rocm:
    needs: [changes]
    if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
@ -147,12 +135,6 @@ jobs:
          go generate -x ./...
        env:
          OLLAMA_SKIP_CPU_GENERATE: '1'
-      - uses: actions/upload-artifact@v4
-        with:
-          name: rocm-${{ matrix.rocm-version }}-libraries
-          path: |
-            llm/build/**/bin/*
-            dist/windows-amd64/**

  # ROCm generation step
  generate-windows-rocm:
@ -189,7 +171,6 @@ jobs:
        name: go generate
        env:
          OLLAMA_SKIP_CPU_GENERATE: '1'
-      # TODO - do we need any artifacts?

  # CUDA generation step
  generate-windows-cuda:
@ -231,7 +212,6 @@ jobs:
          go generate -x ./...
        env:
          OLLAMA_SKIP_CPU_GENERATE: '1'
-      # TODO - do we need any artifacts?

  lint:
    strategy:
@ -263,14 +243,6 @@ jobs:
            arm64) echo ARCH=arm64 ;;
          esac >>$GITHUB_ENV
        shell: bash
-      - run: |
-          mkdir -p llm/build/linux/$ARCH/stub/bin
-          touch llm/build/linux/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'ubuntu-') }}
-      - run: |
-          mkdir -p llm/build/darwin/$ARCH/stub/bin
-          touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'macos-') }}
      - uses: golangci/golangci-lint-action@v6
        with:
          args: --timeout 8m0s -v
@ -301,23 +273,10 @@ jobs:
          cache: true
      - run: |
          case ${{ matrix.arch }} in
-            amd64) echo ARCH=x86_64 ;;
+            amd64) echo ARCH=amd64 ;;
            arm64) echo ARCH=arm64 ;;
          esac >>$GITHUB_ENV
        shell: bash
-      - run: |
-          mkdir -p llm/build/linux/$ARCH/stub/bin
-          touch llm/build/linux/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'ubuntu-') }}
-      - run: |
-          mkdir -p llm/build/darwin/$ARCH/stub/bin
-          touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'macos-') }}
-        shell: bash
      - run: go generate ./...
      - run: go build
      - run: go test -v ./...
-      - uses: actions/upload-artifact@v4
-        with:
-          name: ${{ matrix.os }}-binaries
-          path: ollama
--- a/.gitignore
+++ b/.gitignore
@ -12,4 +12,7 @@ ggml-metal.metal
 test_data
 *.crt
 llm/build
+build/*/*/*
+!build/**/placeholder
+llama/build
 __debug_bin*
--- a/README.md
+++ b/README.md
@ -312,6 +312,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Cherry Studio](https://github.com/kangfenmao/cherry-studio) (Desktop client with Ollama support)
 - [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
 - [Archyve](https://github.com/nickthecook/archyve) (RAG-enabling document library)
+- [crewAI with Mesop](https://github.com/rapidarchitect/ollama-crew-mesop) (Mesop Web Interface to run crewAI with Ollama)

 ### Terminal

@ -336,6 +337,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [podman-ollama](https://github.com/ericcurtin/podman-ollama)
 - [gollama](https://github.com/sammcj/gollama)
 - [Ollama eBook Summary](https://github.com/cognitivetech/ollama-ebook-summary/)
+- [Ollama Mixture of Experts (MOE) in 50 lines of code](https://github.com/rapidarchitect/ollama_moe)

 ### Apple Vision Pro
 - [Enchanted](https://github.com/AugustDev/enchanted)
@ -358,6 +360,7 @@ See the [API documentation](./docs/api.md) for all endpoints.

 - [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
 - [Firebase Genkit](https://firebase.google.com/docs/genkit/plugins/ollama)
+- [crewAI](https://github.com/crewAIInc/crewAI)
 - [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
 - [LangChain4j](https://github.com/langchain4j/langchain4j) with [example](https://github.com/langchain4j/langchain4j-examples/tree/main/ollama-examples/src/main/java)
 - [LangChainRust](https://github.com/Abraxas-365/langchain-rust) with [example](https://github.com/Abraxas-365/langchain-rust/blob/main/examples/llm_ollama.rs)
@ -427,6 +430,8 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Headless Ollama](https://github.com/nischalj10/headless-ollama) (Scripts to automatically install ollama client & models on any OS for apps that depends on ollama server)
 - [vnc-lm](https://github.com/jk011ru/vnc-lm) (A containerized Discord bot with support for attachments and web links)
 - [LSP-AI](https://github.com/SilasMarvin/lsp-ai) (Open-source language server for AI-powered functionality)
+- [QodeAssist](https://github.com/Palm1r/QodeAssist) (AI-powered coding assistant plugin for Qt Creator)
+- [Obsidian Quiz Generator plugin](https://github.com/ECuiDev/obsidian-quiz-generator)

 ### Supported backends

--- a/build/darwin/amd64/placeholder
+++ b/build/darwin/amd64/placeholder
@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
--- a/build/darwin/arm64/placeholder
+++ b/build/darwin/arm64/placeholder
@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
--- a/build/embed_darwin_amd64.go
+++ b/build/embed_darwin_amd64.go
@ -0,0 +1,8 @@
+package build
+
+import "embed"
+
+// Darwin payloads separated by architecture to avoid duplicate payloads when cross compiling
+
+//go:embed darwin/amd64/*
+var EmbedFS embed.FS
--- a/build/embed_darwin_arm64.go
+++ b/build/embed_darwin_arm64.go
@ -0,0 +1,8 @@
+package build
+
+import "embed"
+
+// Darwin payloads separated by architecture to avoid duplicate payloads when cross compiling
+
+//go:embed darwin/arm64/*
+var EmbedFS embed.FS
--- a/build/embed_linux.go
+++ b/build/embed_linux.go
@ -0,0 +1,6 @@
+package build
+
+import "embed"
+
+//go:embed linux/*
+var EmbedFS embed.FS
--- a/build/embed_unused.go
+++ b/build/embed_unused.go
@ -0,0 +1,8 @@
+//go:build !linux && !darwin
+
+package build
+
+import "embed"
+
+// unused on windows
+var EmbedFS embed.FS
--- a/build/linux/amd64/placeholder
+++ b/build/linux/amd64/placeholder
@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
--- a/build/linux/arm64/placeholder
+++ b/build/linux/arm64/placeholder
@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@ -2,6 +2,7 @@ package cmd

 import (
 	"archive/zip"
+	"bufio"
 	"bytes"
 	"context"
 	"crypto/ed25519"
@ -21,6 +22,7 @@ import (
 	"regexp"
 	"runtime"
 	"slices"
+	"strconv"
 	"strings"
 	"sync/atomic"
 	"syscall"
@ -344,6 +346,39 @@ func (w *progressWriter) Write(p []byte) (n int, err error) {
 	return len(p), nil
 }

+func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
+	p := progress.NewProgress(os.Stderr)
+	defer p.StopAndClear()
+
+	spinner := progress.NewSpinner("")
+	p.Add("", spinner)
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	req := &api.GenerateRequest{
+		Model:     opts.Model,
+		KeepAlive: opts.KeepAlive,
+	}
+
+	return client.Generate(cmd.Context(), req, func(api.GenerateResponse) error { return nil })
+}
+
+func StopHandler(cmd *cobra.Command, args []string) error {
+	opts := &runOptions{
+		Model:     args[0],
+		KeepAlive: &api.Duration{Duration: 0},
+	}
+	if err := loadOrUnloadModel(cmd, opts); err != nil {
+		if strings.Contains(err.Error(), "not found") {
+			return fmt.Errorf("couldn't find model \"%s\" to stop", args[0])
+		}
+	}
+	return nil
+}
+
 func RunHandler(cmd *cobra.Command, args []string) error {
 	interactive := true

@ -422,7 +457,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	opts.ParentModel = info.Details.ParentModel

 	if interactive {
-		if err := loadModel(cmd, &opts); err != nil {
+		if err := loadOrUnloadModel(cmd, &opts); err != nil {
 			return err
 		}

@ -578,7 +613,7 @@ func ListHandler(cmd *cobra.Command, args []string) error {
 	table.SetHeaderLine(false)
 	table.SetBorder(false)
 	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("\t")
+	table.SetTablePadding("    ")
 	table.AppendBulk(data)
 	table.Render()

@ -613,7 +648,15 @@ func ListRunningHandler(cmd *cobra.Command, args []string) error {
 				cpuPercent := math.Round(float64(sizeCPU) / float64(m.Size) * 100)
 				procStr = fmt.Sprintf("%d%%/%d%% CPU/GPU", int(cpuPercent), int(100-cpuPercent))
 			}
-			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, format.HumanTime(m.ExpiresAt, "Never")})
+
+			var until string
+			delta := time.Since(m.ExpiresAt)
+			if delta > 0 {
+				until = "Stopping..."
+			} else {
+				until = format.HumanTime(m.ExpiresAt, "Never")
+			}
+			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, until})
 		}
 	}

@ -624,7 +667,7 @@ func ListRunningHandler(cmd *cobra.Command, args []string) error {
 	table.SetHeaderLine(false)
 	table.SetBorder(false)
 	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("\t")
+	table.SetTablePadding("    ")
 	table.AppendBulk(data)
 	table.Render()

@ -720,125 +763,89 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}

-	showInfo(resp)
-
-	return nil
+	return showInfo(resp, os.Stdout)
 }

-func showInfo(resp *api.ShowResponse) {
-	modelData := [][]string{
-		{"parameters", resp.Details.ParameterSize},
-		{"quantization", resp.Details.QuantizationLevel},
+func showInfo(resp *api.ShowResponse, w io.Writer) error {
+	tableRender := func(header string, rows func() [][]string) {
+		fmt.Fprintln(w, " ", header)
+		table := tablewriter.NewWriter(w)
+		table.SetAlignment(tablewriter.ALIGN_LEFT)
+		table.SetBorder(false)
+		table.SetNoWhiteSpace(true)
+		table.SetTablePadding("    ")
+
+		switch header {
+		case "Template", "System", "License":
+			table.SetColWidth(100)
 		}
+
+		table.AppendBulk(rows())
+		table.Render()
+		fmt.Fprintln(w)
+	}
+
+	tableRender("Model", func() (rows [][]string) {
 		if resp.ModelInfo != nil {
 			arch := resp.ModelInfo["general.architecture"].(string)
-		modelData = append(modelData,
-			[]string{"arch", arch},
-			[]string{"context length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))},
-			[]string{"embedding length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64))},
-		)
-	}
-
-	mainTableData := [][]string{
-		{"Model"},
-		{renderSubTable(modelData, false)},
+			rows = append(rows, []string{"", "architecture", arch})
+			rows = append(rows, []string{"", "parameters", format.HumanNumber(uint64(resp.ModelInfo["general.parameter_count"].(float64)))})
+			rows = append(rows, []string{"", "context length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64), 'f', -1, 64)})
+			rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64), 'f', -1, 64)})
+		} else {
+			rows = append(rows, []string{"", "architecture", resp.Details.Family})
+			rows = append(rows, []string{"", "parameters", resp.Details.ParameterSize})
 		}
+		rows = append(rows, []string{"", "quantization", resp.Details.QuantizationLevel})
+		return
+	})

 	if resp.ProjectorInfo != nil {
-		projectorData := [][]string{
-			{"arch", "clip"},
-			{"parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))},
-		}
-
-		if projectorType, ok := resp.ProjectorInfo["clip.projector_type"]; ok {
-			projectorData = append(projectorData, []string{"projector type", projectorType.(string)})
-		}
-
-		projectorData = append(projectorData,
-			[]string{"embedding length", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.embedding_length"].(float64))},
-			[]string{"projection dimensionality", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.projection_dim"].(float64))},
-		)
-
-		mainTableData = append(mainTableData,
-			[]string{"Projector"},
-			[]string{renderSubTable(projectorData, false)},
-		)
+		tableRender("Projector", func() (rows [][]string) {
+			arch := resp.ProjectorInfo["general.architecture"].(string)
+			rows = append(rows, []string{"", "architecture", arch})
+			rows = append(rows, []string{"", "parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))})
+			rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(resp.ProjectorInfo[fmt.Sprintf("%s.vision.embedding_length", arch)].(float64), 'f', -1, 64)})
+			rows = append(rows, []string{"", "dimensions", strconv.FormatFloat(resp.ProjectorInfo[fmt.Sprintf("%s.vision.projection_dim", arch)].(float64), 'f', -1, 64)})
+			return
+		})
 	}

 	if resp.Parameters != "" {
-		mainTableData = append(mainTableData, []string{"Parameters"}, []string{formatParams(resp.Parameters)})
+		tableRender("Parameters", func() (rows [][]string) {
+			scanner := bufio.NewScanner(strings.NewReader(resp.Parameters))
+			for scanner.Scan() {
+				if text := scanner.Text(); text != "" {
+					rows = append(rows, append([]string{""}, strings.Fields(text)...))
+				}
+			}
+			return
+		})
+	}
+
+	head := func(s string, n int) (rows [][]string) {
+		scanner := bufio.NewScanner(strings.NewReader(s))
+		for scanner.Scan() && (len(rows) < n || n < 0) {
+			if text := scanner.Text(); text != "" {
+				rows = append(rows, []string{"", strings.TrimSpace(text)})
+			}
+		}
+		return
 	}

 	if resp.System != "" {
-		mainTableData = append(mainTableData, []string{"System"}, []string{renderSubTable(twoLines(resp.System), true)})
+		tableRender("System", func() [][]string {
+			return head(resp.System, 2)
+		})
 	}

 	if resp.License != "" {
-		mainTableData = append(mainTableData, []string{"License"}, []string{renderSubTable(twoLines(resp.License), true)})
+		tableRender("License", func() [][]string {
+			return head(resp.License, 2)
+		})
 	}

-	table := tablewriter.NewWriter(os.Stdout)
-	table.SetAutoWrapText(false)
-	table.SetBorder(false)
-	table.SetAlignment(tablewriter.ALIGN_LEFT)
-
-	for _, v := range mainTableData {
-		table.Append(v)
-	}
-
-	table.Render()
-}
-
-func renderSubTable(data [][]string, file bool) string {
-	var buf bytes.Buffer
-	table := tablewriter.NewWriter(&buf)
-	table.SetAutoWrapText(!file)
-	table.SetBorder(false)
-	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("\t")
-	table.SetAlignment(tablewriter.ALIGN_LEFT)
-
-	for _, v := range data {
-		table.Append(v)
-	}
-
-	table.Render()
-
-	renderedTable := buf.String()
-	lines := strings.Split(renderedTable, "\n")
-	for i, line := range lines {
-		lines[i] = "\t" + line
-	}
-
-	return strings.Join(lines, "\n")
-}
-
-func twoLines(s string) [][]string {
-	lines := strings.Split(s, "\n")
-	res := [][]string{}
-
-	count := 0
-	for _, line := range lines {
-		line = strings.TrimSpace(line)
-		if line != "" {
-			count++
-			res = append(res, []string{line})
-			if count == 2 {
-				return res
-			}
-		}
-	}
-	return res
-}
-
-func formatParams(s string) string {
-	lines := strings.Split(s, "\n")
-	table := [][]string{}
-
-	for _, line := range lines {
-		table = append(table, strings.Fields(line))
-	}
-	return renderSubTable(table, false)
+	return nil
 }

 func CopyHandler(cmd *cobra.Command, args []string) error {
@ -1328,6 +1335,15 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("insecure", false, "Use an insecure registry")
 	runCmd.Flags().Bool("nowordwrap", false, "Don't wrap words to the next line automatically")
 	runCmd.Flags().String("format", "", "Response format (e.g. json)")
+
+	stopCmd := &cobra.Command{
+		Use:     "stop MODEL",
+		Short:   "Stop a running model",
+		Args:    cobra.ExactArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE:    StopHandler,
+	}
+
 	serveCmd := &cobra.Command{
 		Use:     "serve",
 		Aliases: []string{"start"},
@ -1395,6 +1411,7 @@ func NewCLI() *cobra.Command {
 		createCmd,
 		showCmd,
 		runCmd,
+		stopCmd,
 		pullCmd,
 		pushCmd,
 		listCmd,
@ -1434,6 +1451,7 @@ func NewCLI() *cobra.Command {
 		createCmd,
 		showCmd,
 		runCmd,
+		stopCmd,
 		pullCmd,
 		pushCmd,
 		listCmd,
--- a/cmd/cmd_test.go
+++ b/cmd/cmd_test.go
@ -0,0 +1,206 @@
+package cmd
+
+import (
+	"bytes"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestShowInfo(t *testing.T) {
+	t.Run("bare details", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+`
+
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("bare model info", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			ModelInfo: map[string]any{
+				"general.architecture":    "test",
+				"general.parameter_count": float64(7_000_000_000),
+				"test.context_length":     float64(0),
+				"test.embedding_length":   float64(0),
+			},
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture        test    
+    parameters          7B      
+    context length      0       
+    embedding length    0       
+    quantization        FP16    
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("parameters", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			Parameters: `
+			stop never
+			stop gonna
+			stop give
+			stop you
+			stop up
+			temperature 99`,
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  Parameters
+    stop           never    
+    stop           gonna    
+    stop           give     
+    stop           you      
+    stop           up       
+    temperature    99       
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("project info", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			ProjectorInfo: map[string]any{
+				"general.architecture":         "clip",
+				"general.parameter_count":      float64(133_700_000),
+				"clip.vision.embedding_length": float64(0),
+				"clip.vision.projection_dim":   float64(0),
+			},
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  Projector
+    architecture        clip       
+    parameters          133.70M    
+    embedding length    0          
+    dimensions          0          
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("system", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			System: `You are a pirate!
+Ahoy, matey!
+Weigh anchor!
+			`,
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  System
+    You are a pirate!    
+    Ahoy, matey!         
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("license", func(t *testing.T) {
+		var b bytes.Buffer
+		license, err := os.ReadFile(filepath.Join("..", "LICENSE"))
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			License: string(license),
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  License
+    MIT License             
+    Copyright (c) Ollama    
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+}
--- a/cmd/interactive.go
+++ b/cmd/interactive.go
@ -18,7 +18,6 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/parser"
-	"github.com/ollama/ollama/progress"
 	"github.com/ollama/ollama/readline"
 	"github.com/ollama/ollama/types/errtypes"
 )
@ -31,26 +30,6 @@ const (
 	MultilineSystem
 )

-func loadModel(cmd *cobra.Command, opts *runOptions) error {
-	p := progress.NewProgress(os.Stderr)
-	defer p.StopAndClear()
-
-	spinner := progress.NewSpinner("")
-	p.Add("", spinner)
-
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
-	chatReq := &api.ChatRequest{
-		Model:     opts.Model,
-		KeepAlive: opts.KeepAlive,
-	}
-
-	return client.Chat(cmd.Context(), chatReq, func(api.ChatResponse) error { return nil })
-}
-
 func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 	usage := func() {
 		fmt.Fprintln(os.Stderr, "Available Commands:")
@ -217,7 +196,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			opts.Model = args[1]
 			opts.Messages = []api.Message{}
 			fmt.Printf("Loading model '%s'\n", opts.Model)
-			if err := loadModel(cmd, &opts); err != nil {
+			if err := loadOrUnloadModel(cmd, &opts); err != nil {
 				return err
 			}
 			continue
@ -371,7 +350,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {

 				switch args[1] {
 				case "info":
-					showInfo(resp)
+					_ = showInfo(resp, os.Stderr)
 				case "license":
 					if resp.License == "" {
 						fmt.Println("No license was specified for this model.")
--- a/convert/convert.go
+++ b/convert/convert.go
@ -208,14 +208,18 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		return err
 	}

-	if vocabSize := int(p.VocabSize); vocabSize > len(t.Vocabulary.Tokens) {
-		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", p.VocabSize, "actual", len(t.Vocabulary.Tokens))
+	vocabSize := int(p.VocabSize)
+	switch {
+	case vocabSize > len(t.Vocabulary.Tokens):
+		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
 		for i := range vocabSize - len(t.Vocabulary.Tokens) {
 			t.Vocabulary.Tokens = append(t.Vocabulary.Tokens, fmt.Sprintf("[PAD%d]", i))
 			t.Vocabulary.Scores = append(t.Vocabulary.Scores, -1)
 			t.Vocabulary.Types = append(t.Vocabulary.Types, tokenTypeUserDefined)
 		}
-	} else {
+	case vocabSize < len(t.Vocabulary.Tokens):
+		return fmt.Errorf("vocabulary is larger than expected '%d' instead of '%d'", len(t.Vocabulary.Tokens), vocabSize)
+	default:
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}

--- a/docs/api.md
+++ b/docs/api.md
@ -69,7 +69,7 @@ Enable JSON mode by setting the `format` parameter to `json`. This will structur

 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
  "prompt": "Why is the sky blue?"
 }'
 ```
@ -80,7 +80,7 @@ A stream of JSON objects is returned:

 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
  "created_at": "2023-08-04T08:52:19.385406455-07:00",
  "response": "The",
  "done": false
@ -102,7 +102,7 @@ To calculate how fast the response is generated in tokens per second (token/s),

 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
  "created_at": "2023-08-04T19:22:45.499127Z",
  "response": "",
  "done": true,
@ -124,7 +124,7 @@ A response can be received in one reply when streaming is off.

 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
  "prompt": "Why is the sky blue?",
  "stream": false
 }'
@ -136,7 +136,7 @@ If `stream` is set to `false`, the response will be a single JSON object:

 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
  "created_at": "2023-08-04T19:22:45.499127Z",
  "response": "The sky is blue because it is the color of the sky.",
  "done": true,
@ -194,7 +194,7 @@ curl http://localhost:11434/api/generate -d '{

 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
  "prompt": "What color is the sky at different times of the day? Respond using JSON",
  "format": "json",
  "stream": false
@ -205,7 +205,7 @@ curl http://localhost:11434/api/generate -d '{

 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
  "created_at": "2023-11-09T21:07:55.186497Z",
  "response": "{\n\"morning\": {\n\"color\": \"blue\"\n},\n\"noon\": {\n\"color\": \"blue-gray\"\n},\n\"afternoon\": {\n\"color\": \"warm gray\"\n},\n\"evening\": {\n\"color\": \"orange\"\n}\n}\n",
  "done": true,
@ -327,7 +327,7 @@ If you want to set custom options for the model at runtime rather than in the Mo

 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
  "prompt": "Why is the sky blue?",
  "stream": false,
  "options": {
@ -368,7 +368,7 @@ curl http://localhost:11434/api/generate -d '{

 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
  "created_at": "2023-08-04T19:22:45.499127Z",
  "response": "The sky is blue because it is the color of the sky.",
  "done": true,
@ -390,7 +390,7 @@ If an empty prompt is provided, the model will be loaded into memory.

 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3"
+  "model": "llama3.1"
 }'
 ```

@ -400,7 +400,7 @@ A single JSON object is returned:

 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
  "created_at": "2023-12-18T19:52:07.071755Z",
  "response": "",
  "done": true
@ -445,7 +445,7 @@ Send a chat message with a streaming response.

 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3",
+  "model": "llama3.1",
  "messages": [
    {
      "role": "user",
@ -461,7 +461,7 @@ A stream of JSON objects is returned:

 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
  "created_at": "2023-08-04T08:52:19.385406455-07:00",
  "message": {
    "role": "assistant",
@ -476,7 +476,7 @@ Final response:

 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
  "created_at": "2023-08-04T19:22:45.499127Z",
  "done": true,
  "total_duration": 4883583458,
@ -494,7 +494,7 @@ Final response:

 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3",
+  "model": "llama3.1",
  "messages": [
    {
      "role": "user",
@ -509,7 +509,7 @@ curl http://localhost:11434/api/chat -d '{

 ```json
 {
-  "model": "registry.ollama.ai/library/llama3:latest",
+  "model": "llama3.1",
  "created_at": "2023-12-12T14:13:43.416799Z",
  "message": {
    "role": "assistant",
@ -533,7 +533,7 @@ Send a chat message with a conversation history. You can use this same approach

 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3",
+  "model": "llama3.1",
  "messages": [
    {
      "role": "user",
@ -557,7 +557,7 @@ A stream of JSON objects is returned:

 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
  "created_at": "2023-08-04T08:52:19.385406455-07:00",
  "message": {
    "role": "assistant",
@ -571,7 +571,7 @@ Final response:

 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
  "created_at": "2023-08-04T19:22:45.499127Z",
  "done": true,
  "total_duration": 8113331500,
@ -629,7 +629,7 @@ curl http://localhost:11434/api/chat -d '{

 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3",
+  "model": "llama3.1",
  "messages": [
    {
      "role": "user",
@ -647,7 +647,7 @@ curl http://localhost:11434/api/chat -d '{

 ```json
 {
-  "model": "registry.ollama.ai/library/llama3:latest",
+  "model": "llama3.1",
  "created_at": "2023-12-12T14:13:43.416799Z",
  "message": {
    "role": "assistant",
@ -904,7 +904,7 @@ Show information about a model including details, modelfile, template, parameter

 ```shell
 curl http://localhost:11434/api/show -d '{
-  "name": "llama3"
+  "name": "llama3.1"
 }'
 ```

@ -965,7 +965,7 @@ Copy a model. Creates a model with another name from an existing model.

 ```shell
 curl http://localhost:11434/api/copy -d '{
-  "source": "llama3",
+  "source": "llama3.1",
  "destination": "llama3-backup"
 }'
 ```
@ -1020,7 +1020,7 @@ Download a model from the ollama library. Cancelled pulls are resumed from where

 ```shell
 curl http://localhost:11434/api/pull -d '{
-  "name": "llama3"
+  "name": "llama3.1"
 }'
 ```

--- a/docs/faq.md
+++ b/docs/faq.md
@ -32,7 +32,7 @@ When using the API, specify the `num_ctx` parameter:

 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
  "prompt": "Why is the sky blue?",
  "options": {
    "num_ctx": 4096
@ -247,12 +247,12 @@ The `keep_alive` parameter can be set to:

 For example, to preload a model and leave it in memory use:
 ```shell
-curl http://localhost:11434/api/generate -d '{"model": "llama3", "keep_alive": -1}'
+curl http://localhost:11434/api/generate -d '{"model": "llama3.1", "keep_alive": -1}'
 ```

 To unload the model and free up memory use:
 ```shell
-curl http://localhost:11434/api/generate -d '{"model": "llama3", "keep_alive": 0}'
+curl http://localhost:11434/api/generate -d '{"model": "llama3.1", "keep_alive": 0}'
 ```

 Alternatively, you can change the amount of time all models are loaded into memory by setting the `OLLAMA_KEEP_ALIVE` environment variable when starting the Ollama server. The `OLLAMA_KEEP_ALIVE` variable uses the same parameter types as the `keep_alive` parameter types mentioned above. Refer to section explaining [how to configure the Ollama server](#how-do-i-configure-ollama-server) to correctly set the environment variable.
--- a/docs/modelfile.md
+++ b/docs/modelfile.md
@ -11,7 +11,7 @@ A model file is the blueprint to create and share models with Ollama.
 - [Examples](#examples)
 - [Instructions](#instructions)
  - [FROM (Required)](#from-required)
-    - [Build from llama3.1](#build-from-llama31)
+    - [Build from existing model](#build-from-existing-model)
    - [Build from a Safetensors model](#build-from-a-safetensors-model)
    - [Build from a GGUF file](#build-from-a-gguf-file)
  - [PARAMETER](#parameter)
@ -50,7 +50,7 @@ INSTRUCTION arguments
 An example of a `Modelfile` creating a mario blueprint:

 ```modelfile
-FROM llama3
+FROM llama3.1
 # sets the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
 # sets the context window size to 4096, this controls how many tokens the LLM can use as context to generate the next token
@ -72,10 +72,10 @@ More examples are available in the [examples directory](../examples).
 To view the Modelfile of a given model, use the `ollama show --modelfile` command.

  ```bash
-  > ollama show --modelfile llama3
+  > ollama show --modelfile llama3.1
  # Modelfile generated by "ollama show"
  # To build a new Modelfile based on this one, replace the FROM line with:
-  # FROM llama3:latest
+  # FROM llama3.1:latest
  FROM /Users/pdevine/.ollama/models/blobs/sha256-00e1317cbf74d901080d7100f57580ba8dd8de57203072dc6f668324ba545f29
  TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>

@ -100,7 +100,7 @@ The `FROM` instruction defines the base model to use when creating a model.
 FROM <model name>:<tag>
 ```

-#### Build from llama3.1
+#### Build from existing model

 ```modelfile
 FROM llama3.1
--- a/docs/openai.md
+++ b/docs/openai.md
@ -25,7 +25,7 @@ chat_completion = client.chat.completions.create(
            'content': 'Say this is a test',
        }
    ],
-    model='llama3',
+    model='llama3.1',
 )

 response = client.chat.completions.create(
@ -46,13 +46,13 @@ response = client.chat.completions.create(
 )

 completion = client.completions.create(
-    model="llama3",
+    model="llama3.1",
    prompt="Say this is a test",
 )

 list_completion = client.models.list()

-model = client.models.retrieve("llama3")
+model = client.models.retrieve("llama3.1")

 embeddings = client.embeddings.create(
    model="all-minilm",
@ -74,7 +74,7 @@ const openai = new OpenAI({

 const chatCompletion = await openai.chat.completions.create({
    messages: [{ role: 'user', content: 'Say this is a test' }],
-    model: 'llama3',
+    model: 'llama3.1',
 })

 const response = await openai.chat.completions.create({
@ -94,13 +94,13 @@ const response = await openai.chat.completions.create({
 })

 const completion = await openai.completions.create({
-    model: "llama3",
+    model: "llama3.1",
    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("llama3.1")

 const embedding = await openai.embeddings.create({
  model: "all-minilm",
@ -114,7 +114,7 @@ const embedding = await openai.embeddings.create({
 curl http://localhost:11434/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
-        "model": "llama3",
+        "model": "llama3.1",
        "messages": [
            {
                "role": "system",
@ -154,13 +154,13 @@ curl http://localhost:11434/v1/chat/completions \
 curl http://localhost:11434/v1/completions \
    -H "Content-Type: application/json" \
    -d '{
-        "model": "llama3",
+        "model": "llama3.1",
        "prompt": "Say this is a test"
    }'

 curl http://localhost:11434/v1/models

-curl http://localhost:11434/v1/models/llama3
+curl http://localhost:11434/v1/models/llama3.1

 curl http://localhost:11434/v1/embeddings \
    -H "Content-Type: application/json" \
@ -274,7 +274,7 @@ curl http://localhost:11434/v1/embeddings \
 Before using a model, pull it locally `ollama pull`:

 ```shell
-ollama pull llama3
+ollama pull llama3.1
 ```

 ### Default model names
@ -282,7 +282,7 @@ ollama pull llama3
 For tooling that relies on default OpenAI model names such as `gpt-3.5-turbo`, use `ollama cp` to copy an existing model name to a temporary name:

 ```
-ollama cp llama3 gpt-3.5-turbo
+ollama cp llama3.1 gpt-3.5-turbo
 ```

 Afterwards, this new model name can be specified the `model` field:
--- a/docs/template.md
+++ b/docs/template.md
@ -33,7 +33,7 @@ Omitting a template in these models puts the responsibility of correctly templat
 To add templates in your model, you'll need to add a `TEMPLATE` command to the Modelfile. Here's an example using Meta's Llama 3.

 ```dockerfile
-FROM llama3
+FROM llama3.1

 TEMPLATE """{{- if .System }}<|start_header_id|>system<|end_header_id|>

--- a/docs/troubleshooting.md
+++ b/docs/troubleshooting.md
@ -91,6 +91,17 @@ If none of those resolve the problem, gather additional information and file an
 - Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`


+## AMD GPU Discovery
+
+On linux, AMD GPU access typically requires `video` and/or `render` group membership to access the `/dev/kfd` device.  If permissions are not set up correctly, Ollama will detect this and report an error in the server log.
+
+When running in a container, in some Linux distributions and container runtimes, the ollama process may be unable to access the GPU.  Use `ls -ld /dev/kfd /dev/dri /dev/dri/*` on the host system to determine the group assignments on your system, and pass additional `--group-add ...` arguments to the container so it can access the required devices.
+
+If you are experiencing problems getting Ollama to correctly discover or use your GPU for inference, the following may help isolate the failure.
+- `AMD_LOG_LEVEL=3` Enable info log levels in the AMD HIP/ROCm libraries.  This can help show more detailed error codes that can help troubleshoot problems
+- `OLLAMA_DEBUG=1` During GPU discovery additional information will be reported
+- Check dmesg for any errors from amdgpu or kfd drivers `sudo dmesg | grep -i amdgpu` and `sudo dmesg | grep -i kfd`
+
 ## Windows Terminal Errors

 Older versions of Windows 10 (e.g., 21H1) are known to have a bug where the standard terminal program does not display control characters correctly.  This can result in a long string of strings like `←[?25h←[?25l` being displayed, sometimes erroring with `The parameter is incorrect`  To resolve this problem, please update to Win 10 22H1 or newer.
--- a/docs/windows.md
+++ b/docs/windows.md
@ -29,7 +29,7 @@ Ollama uses unicode characters for progress indication, which may render as unkn

 Here's a quick example showing API access from `powershell`
 ```powershell
-(Invoke-WebRequest -method POST -Body '{"model":"llama3", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
+(Invoke-WebRequest -method POST -Body '{"model":"llama3.1", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
 ```

 ## Troubleshooting
--- a/envconfig/config.go
+++ b/envconfig/config.go
@ -179,53 +179,6 @@ var (
 	HsaOverrideGfxVersion = String("HSA_OVERRIDE_GFX_VERSION")
 )

-func RunnersDir() (p string) {
-	if p := Var("OLLAMA_RUNNERS_DIR"); p != "" {
-		return p
-	}
-
-	if runtime.GOOS != "windows" {
-		return
-	}
-
-	defer func() {
-		if p == "" {
-			slog.Error("unable to locate llm runner directory. Set OLLAMA_RUNNERS_DIR to the location of 'ollama/runners'")
-		}
-	}()
-
-	// On Windows we do not carry the payloads inside the main executable
-	exe, err := os.Executable()
-	if err != nil {
-		return
-	}
-
-	cwd, err := os.Getwd()
-	if err != nil {
-		return
-	}
-
-	var paths []string
-	for _, root := range []string{filepath.Dir(exe), filepath.Join(filepath.Dir(exe), LibRelativeToExe()), cwd} {
-		paths = append(paths,
-			root,
-			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, "lib", "ollama", "runners")
-		if _, err := os.Stat(candidate); err == nil {
-			p = candidate
-			break
-		}
-	}
-
-	return p
-}
-
 func Uint(key string, defaultValue uint) func() uint {
 	return func() uint {
 		if s := Var(key); s != "" {
@ -290,10 +243,22 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune(), "Do not prune model blobs on startup"},
 		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel(), "Maximum number of parallel requests"},
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", Origins(), "A comma separated list of allowed origins"},
-		"OLLAMA_RUNNERS_DIR":       {"OLLAMA_RUNNERS_DIR", RunnersDir(), "Location for runners"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_TMPDIR":            {"OLLAMA_TMPDIR", TmpDir(), "Location for temporary files"},
+
+		// Informational
+		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},
+		"HTTPS_PROXY": {"HTTPS_PROXY", String("HTTPS_PROXY")(), "HTTPS proxy"},
+		"NO_PROXY":    {"NO_PROXY", String("NO_PROXY")(), "No proxy"},
 	}
+
+	if runtime.GOOS != "windows" {
+		// Windows environment variables are case-insensitive so there's no need to duplicate them
+		ret["http_proxy"] = EnvVar{"http_proxy", String("http_proxy")(), "HTTP proxy"}
+		ret["https_proxy"] = EnvVar{"https_proxy", String("https_proxy")(), "HTTPS proxy"}
+		ret["no_proxy"] = EnvVar{"no_proxy", String("no_proxy")(), "No proxy"}
+	}
+
 	if runtime.GOOS != "darwin" {
 		ret["CUDA_VISIBLE_DEVICES"] = EnvVar{"CUDA_VISIBLE_DEVICES", CudaVisibleDevices(), "Set which NVIDIA devices are visible"}
 		ret["HIP_VISIBLE_DEVICES"] = EnvVar{"HIP_VISIBLE_DEVICES", HipVisibleDevices(), "Set which AMD devices are visible"}
@ -302,6 +267,7 @@ func AsMap() map[string]EnvVar {
 		ret["HSA_OVERRIDE_GFX_VERSION"] = EnvVar{"HSA_OVERRIDE_GFX_VERSION", HsaOverrideGfxVersion(), "Override the gfx used for all detected AMD GPUs"}
 		ret["OLLAMA_INTEL_GPU"] = EnvVar{"OLLAMA_INTEL_GPU", IntelGPU(), "Enable experimental Intel GPU detection"}
 	}
+
 	return ret
 }

--- a/examples/langchain-python-rag-privategpt/requirements.txt
+++ b/examples/langchain-python-rag-privategpt/requirements.txt
@ -1,6 +1,6 @@
 langchain==0.0.274
 gpt4all==1.0.8
-chromadb==0.4.7
+chromadb==0.5.0
 llama-cpp-python==0.1.81
 urllib3==2.0.4
 PyMuPDF==1.23.5
--- a/examples/python-loganalysis/Modelfile
+++ b/examples/python-loganalysis/Modelfile
@ -4,5 +4,5 @@ SYSTEM """
 You are a log file analyzer. You will receive a set of lines from a log file for some software application, find the errors and other interesting aspects of the logs, and explain them so a new user can understand what they mean. If there are any steps they can do to resolve them, list the steps in your answer.
 """

-PARAMETER TEMPERATURE 0.3
+PARAMETER temperature 0.3

--- a/examples/python-loganalysis/readme.md
+++ b/examples/python-loganalysis/readme.md
@ -21,6 +21,8 @@ You can try this with the `logtest.logfile` file included in this directory.
 2. Install the Python Requirements.

   ```bash
+   python3 -m venv .venv
+   source .venv/bin/activate
   pip install -r requirements.txt
   ```

--- a/examples/python-loganalysis/requirements.txt
+++ b/examples/python-loganalysis/requirements.txt
@ -1 +1 @@
-Requests==2.31.0
+Requests>=2.32.3
--- a/gpu/amd_linux.go
+++ b/gpu/amd_linux.go
@ -5,6 +5,7 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"io/fs"
 	"log/slog"
 	"os"
 	"path/filepath"
@ -359,6 +360,10 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 	if len(resp) == 0 {
 		slog.Info("no compatible amdgpu devices detected")
 	}
+	if err := verifyKFDDriverAccess(); err != nil {
+		slog.Error("amdgpu devices detected but permission problems block access", "error", err)
+		return nil
+	}
 	return resp
 }

@ -455,3 +460,19 @@ func getFreeMemory(usedFile string) (uint64, error) {
 	}
 	return usedMemory, nil
 }
+
+func verifyKFDDriverAccess() error {
+	// Verify we have permissions - either running as root, or we have group access to the driver
+	fd, err := os.OpenFile("/dev/kfd", os.O_RDWR, 0o666)
+	if err != nil {
+		if errors.Is(err, fs.ErrPermission) {
+			return fmt.Errorf("permissions not set up properly.  Either run ollama as root, or add you user account to the render group. %w", err)
+		} else if errors.Is(err, fs.ErrNotExist) {
+			// Container runtime failure?
+			return fmt.Errorf("kfd driver not loaded.  If running in a container, remember to include '--device /dev/kfd --device /dev/dri'")
+		}
+		return fmt.Errorf("failed to check permission on /dev/kfd: %w", err)
+	}
+	fd.Close()
+	return nil
+}
--- a/gpu/assets.go
+++ b/gpu/assets.go
@ -1,148 +0,0 @@
-package gpu
-
-import (
-	"errors"
-	"fmt"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"runtime"
-	"strconv"
-	"strings"
-	"sync"
-	"syscall"
-	"time"
-
-	"github.com/ollama/ollama/envconfig"
-)
-
-var (
-	lock        sync.Mutex
-	payloadsDir = ""
-)
-
-func PayloadsDir() (string, error) {
-	lock.Lock()
-	defer lock.Unlock()
-	var err error
-	if payloadsDir == "" {
-		runnersDir := envconfig.RunnersDir()
-
-		if runnersDir != "" {
-			payloadsDir = runnersDir
-			return payloadsDir, nil
-		}
-
-		// The remainder only applies on non-windows where we still carry payloads in the main executable
-		cleanupTmpDirs()
-		tmpDir := envconfig.TmpDir()
-		if tmpDir == "" {
-			tmpDir, err = os.MkdirTemp("", "ollama")
-			if err != nil {
-				return "", fmt.Errorf("failed to generate tmp dir: %w", err)
-			}
-		} else {
-			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
-		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
-		// This will typically look like /tmp/ollama3208993108/runners on linux
-		payloadsDir = filepath.Join(tmpDir, "runners")
-	}
-	return payloadsDir, nil
-}
-
-// Best effort to clean up prior tmpdirs
-func cleanupTmpDirs() {
-	matches, err := filepath.Glob(filepath.Join(os.TempDir(), "ollama*", "ollama.pid"))
-	if err != nil {
-		return
-	}
-
-	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
-		} 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("invalid pid, skipping", "path", match, "error", err)
-			continue
-		}
-
-		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(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)
-		}
-	}
-}
-
-func Cleanup() {
-	lock.Lock()
-	defer lock.Unlock()
-	runnersDir := envconfig.RunnersDir()
-	if payloadsDir != "" && runnersDir == "" && runtime.GOOS != "windows" {
-		// We want to fully clean up the tmpdir parent of the payloads dir
-		tmpDir := filepath.Clean(filepath.Join(payloadsDir, ".."))
-		slog.Debug("cleaning up", "dir", tmpDir)
-		err := os.RemoveAll(tmpDir)
-		if err != nil {
-			// On windows, if we remove too quickly the llama.dll may still be in-use and fail to remove
-			time.Sleep(1000 * time.Millisecond)
-			err = os.RemoveAll(tmpDir)
-			if err != nil {
-				slog.Warn("failed to clean up", "dir", tmpDir, "err", err)
-			}
-		}
-	}
-}
-
-func UpdatePath(dir string) {
-	if runtime.GOOS == "windows" {
-		tmpDir := filepath.Dir(dir)
-		pathComponents := strings.Split(os.Getenv("PATH"), ";")
-		i := 0
-		for _, comp := range pathComponents {
-			if strings.EqualFold(comp, dir) {
-				return
-			}
-			// Remove any other prior paths to our temp dir
-			if !strings.HasPrefix(strings.ToLower(comp), strings.ToLower(tmpDir)) {
-				pathComponents[i] = comp
-				i++
-			}
-		}
-		newPath := strings.Join(append([]string{dir}, pathComponents...), ";")
-		slog.Info("updating", "PATH", newPath)
-		os.Setenv("PATH", newPath)
-	}
-	// linux and darwin rely on rpath
-}
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@ -93,10 +93,9 @@ func initCudaHandles() *cudaHandles {
 		localAppData := os.Getenv("LOCALAPPDATA")
 		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", CudartMgmtName)}
 	}
-	tmpDir, _ := PayloadsDir()
-	if tmpDir != "" {
-		// TODO - add "payloads" for subprocess
-		cudartMgmtPatterns = []string{filepath.Join(tmpDir, "cuda*", CudartMgmtName)}
+	libDir := LibraryDir()
+	if libDir != "" {
+		cudartMgmtPatterns = []string{filepath.Join(libDir, CudartMgmtName)}
 	}
 	cudartMgmtPatterns = append(cudartMgmtPatterns, CudartGlobs...)

--- a/llm/ext_server/server.cpp
+++ b/llm/ext_server/server.cpp
@ -913,7 +913,9 @@ struct llama_server_context
        slot.sampled = result.tok;

        // search stop word and delete it
+        if (!llama_token_is_eog(model, result.tok))
            slot.generated_text += token_str;
+
        slot.has_next_token = true;

        if (slot.ctx_sampling->params.use_penalty_prompt_tokens && result.tok != -1)
@ -954,6 +956,8 @@ struct llama_server_context
        if (!incomplete)
        {
            size_t pos = std::min(slot.n_sent_text, slot.generated_text.size());
+
+            if (!llama_token_is_eog(model, result.tok)) {
                const std::string str_test = slot.generated_text.substr(pos);
                bool is_stop_full = false;
                size_t stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_FULL, slot);
@ -979,6 +983,10 @@ struct llama_server_context
                    slot.n_sent_text += result.text_to_send.size();
                    // add the token to slot queue and cache
                }
+            } else {
+                    result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
+                    slot.n_sent_text += result.text_to_send.size();
+            }

            if (slot.params.stream)
            {
@ -1117,9 +1125,7 @@ struct llama_server_context
            {"multimodal", multimodal}
        };

-        if (!llama_token_is_eog(model, tkn.tok)) {
        res.result_json["content"] = tkn.text_to_send;
-        }

        if (slot.sparams.n_probs > 0)
        {
--- a/llm/generate/gen_common.sh
+++ b/llm/generate/gen_common.sh
@ -31,6 +31,7 @@ init_vars() {
        NO_WHOLE_ARCHIVE=""
        GCC_ARCH="-arch ${ARCH}"
        DIST_BASE=../../dist/darwin-${GOARCH}/
+        PAYLOAD_BASE=../../build/darwin/${GOARCH}
        ;;
    "Linux")
        LIB_EXT="so"
@ -40,6 +41,7 @@ init_vars() {
        # Cross compiling not supported on linux - Use docker
        GCC_ARCH=""
        DIST_BASE=../../dist/linux-${GOARCH}/
+        PAYLOAD_BASE=../../build/linux/${GOARCH}
        ;;
    *)
        ;;
@ -47,7 +49,8 @@ init_vars() {
    if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then
        CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
    fi
-    GZIP=$(which pigz 2>/dev/null || echo "gzip")
+    GZIP=$(command -v pigz 2>/dev/null || echo "gzip")
+    RUNNER_BASE="${DIST_BASE}/lib/ollama/runners"
 }

 git_module_setup() {
@ -91,17 +94,34 @@ build() {
    rm -f ${BUILD_DIR}/bin/ggml-common.h ${BUILD_DIR}/bin/ggml-metal.metal
 }

-compress() {
-    echo "Compressing payloads to reduce overall binary size..."
-    rm -rf ${BUILD_DIR}/bin/*.gz
+dist() {
+    [ -z "${RUNNER}" ] && exit 1
+    mkdir -p ${RUNNER_BASE}/${RUNNER}/
    for f in ${BUILD_DIR}/bin/* ; do
-        ${GZIP} -n --best -f ${f} &
+        cp ${f} ${RUNNER_BASE}/${RUNNER}/
+    done
+    # check for lib directory
+    if [ -d ${BUILD_DIR}/lib ]; then
+        for f in ${BUILD_DIR}/lib/* ; do
+            cp ${f} ${RUNNER_BASE}/${RUNNER}/
+        done
+    fi
+}
+
+# Compress from the build $BUILD_DIR into the $PAYLOAD_BASE/$RUNNER dir
+compress() {
+    [ -z "${RUNNER}" ] && exit 1
+    echo "Compressing payloads with ${GZIP} to reduce overall binary size..."
+    rm -rf "${PAYLOAD_BASE}/${RUNNER}/"
+    mkdir -p "${PAYLOAD_BASE}/${RUNNER}/"
+    for f in ${BUILD_DIR}/bin/* ; do
+        ${GZIP} -c --best ${f} > "${PAYLOAD_BASE}/${RUNNER}/$(basename ${f}).gz" &
        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} &
+            ${GZIP} -c --best ${f} > "${PAYLOAD_BASE}/${RUNNER}/$(basename ${f}).gz" &
            compress_pids+=" $!"
        done
    fi
@ -117,7 +137,7 @@ wait_for_compress() {

 install() {
    echo "Installing libraries to bin dir ${BUILD_DIR}/bin/"
-    for lib in $(find ${BUILD_DIR} -name \*.${LIB_EXT}); do
+    for lib in $(find ${BUILD_DIR} -name \*.${LIB_EXT} | grep -v "${BUILD_DIR}/bin/" ); do
        rm -f "${BUILD_DIR}/bin/$(basename ${lib})"
        cp -af "${lib}" "${BUILD_DIR}/bin/"
    done
--- a/llm/generate/gen_darwin.sh
+++ b/llm/generate/gen_darwin.sh
@ -39,7 +39,8 @@ case "${GOARCH}" in
        #
        init_vars
        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/cpu"
+        RUNNER=cpu
+        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
        echo "Building LCD CPU"
        build
        sign ${BUILD_DIR}/bin/ollama_llama_server
@ -51,7 +52,8 @@ case "${GOARCH}" in
        #
        init_vars
        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
+        RUNNER=cpu_avx
+        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
        echo "Building AVX CPU"
        build
        sign ${BUILD_DIR}/bin/ollama_llama_server
@ -63,7 +65,8 @@ case "${GOARCH}" in
        #
        init_vars
        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=on -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
+        RUNNER=cpu_avx2
+        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
        echo "Building AVX2 CPU"
        EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
        build
@ -84,7 +87,8 @@ case "${GOARCH}" in
    if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
        init_vars
        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/metal"
+        RUNNER="metal"
+        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
        EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
        build
        sign ${BUILD_DIR}/bin/ollama_llama_server
--- a/llm/generate/gen_linux.sh
+++ b/llm/generate/gen_linux.sh
@ -79,10 +79,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
        init_vars
        echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
-        BUILD_DIR="../build/linux/${ARCH}/cpu"
+        RUNNER="cpu"
+        BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
        echo "Building custom CPU"
        build
        install
+        dist
        compress
    else
        # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
@ -102,10 +104,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
            #
            init_vars
            CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
-            BUILD_DIR="../build/linux/${ARCH}/cpu"
+            RUNNER=cpu
+            BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
            echo "Building LCD CPU"
            build
            install
+            dist
            compress
        fi

@ -120,10 +124,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                #
                init_vars
                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
-                BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
+                RUNNER=cpu_avx
+                BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
                echo "Building AVX CPU"
                build
                install
+                dist
                compress
            fi

@ -134,10 +140,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                #
                init_vars
                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
-                BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
+                RUNNER=cpu_avx2
+                BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
                echo "Building AVX2 CPU"
                build
                install
+                dist
                compress
            fi
        fi
@ -187,11 +195,13 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
    fi
    export CUDAFLAGS="-t8"
    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS} -DGGML_STATIC=off"
-    BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
+    RUNNER=cuda${CUDA_VARIANT}
+    BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
    export LLAMA_SERVER_LDFLAGS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
    CUDA_DIST_DIR="${CUDA_DIST_DIR:-${DIST_BASE}/lib/ollama}"
    build
    install
+    dist
    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* ${CUDA_LIB_DIR}/libcublasLt.so* ; do
@ -212,7 +222,8 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
    source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
    CC=icx
    CMAKE_DEFS="${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"
+    RUNNER=oneapi
+    BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
    ONEAPI_DIST_DIR="${DIST_BASE}/lib/ollama"
    export LLAMA_SERVER_LDFLAGS="-fsycl -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
    DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
@ -231,6 +242,7 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
    cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${ONEAPI_DIST_DIR}"
    cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${ONEAPI_DIST_DIR}"
    install
+    dist
    compress
 fi

@ -259,7 +271,8 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
        CMAKE_DEFS="${CMAKE_DEFS} ${OLLAMA_CUSTOM_ROCM_DEFS}"
        echo "Building custom ROCM GPU"
    fi
-    BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
+    RUNNER=rocm${ROCM_VARIANT}
+    BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
    # 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)
@ -269,13 +282,17 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then

    # 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
+    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -v "${GOARCH}/rocm${ROCM_VARIANT}" | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf ); do
        cp -a "${dep}"* "${ROCM_DIST_DIR}"
+        if [ $(readlink -f "${dep}") != "${dep}" ] ; then
+            cp $(readlink -f "${dep}") "${ROCM_DIST_DIR}"
+        fi
    done
    install
+    dist
    compress
 fi

 cleanup
 wait_for_compress
-echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
+echo "go generate completed.  LLM runners: $(cd ${PAYLOAD_BASE}; echo *)"
--- a/llm/llm_darwin_arm64.go
+++ b/llm/llm_darwin_arm64.go
@ -1,11 +1,7 @@
 package llm

 import (
-	"embed"
 	"syscall"
 )

-//go:embed build/darwin/arm64/*/bin/*
-var libEmbed embed.FS
-
 var LlamaServerSysProcAttr = &syscall.SysProcAttr{}
--- a/llm/llm_darwin_amd64.go
+++ b/llm/llm_darwin_amd64.go
@ -1,11 +0,0 @@
-package llm
-
-import (
-	"embed"
-	"syscall"
-)
-
-//go:embed build/darwin/x86_64/*/bin/*
-var libEmbed embed.FS
-
-var LlamaServerSysProcAttr = &syscall.SysProcAttr{}
--- a/llm/llm_linux.go
+++ b/llm/llm_linux.go
@ -1,11 +1,7 @@
 package llm

 import (
-	"embed"
 	"syscall"
 )

-//go:embed build/linux/*/*/bin/*
-var libEmbed embed.FS
-
 var LlamaServerSysProcAttr = &syscall.SysProcAttr{}
--- a/llm/llm_windows.go
+++ b/llm/llm_windows.go
@ -1,13 +1,9 @@
 package llm

 import (
-	"embed"
 	"syscall"
 )

-// unused on windows
-var libEmbed embed.FS
-
 const CREATE_DEFAULT_ERROR_MODE = 0x04000000

 var LlamaServerSysProcAttr = &syscall.SysProcAttr{
--- a/llm/payload.go
+++ b/llm/payload.go
@ -1,233 +0,0 @@
-package llm
-
-import (
-	"compress/gzip"
-	"errors"
-	"fmt"
-	"io"
-	"io/fs"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"runtime"
-	"slices"
-	"strings"
-
-	"golang.org/x/sync/errgroup"
-
-	"github.com/ollama/ollama/gpu"
-)
-
-var errPayloadMissing = errors.New("expected payloads not included in this build of ollama")
-
-func Init() error {
-	payloadsDir, err := gpu.PayloadsDir()
-	if err != nil {
-		return err
-	}
-
-	if runtime.GOOS != "windows" {
-		slog.Info("extracting embedded files", "dir", payloadsDir)
-		binGlob := "build/*/*/*/bin/*"
-
-		// extract server libraries
-		err = extractFiles(payloadsDir, binGlob)
-		if err != nil {
-			return fmt.Errorf("extract binaries: %v", err)
-		}
-	}
-
-	var variants []string
-	for v := range getAvailableServers() {
-		variants = append(variants, v)
-	}
-	slog.Info(fmt.Sprintf("Dynamic LLM libraries %v", variants))
-	slog.Debug("Override detection logic by setting OLLAMA_LLM_LIBRARY")
-
-	return nil
-}
-
-// binary names may contain an optional variant separated by '_'
-// For example, "ollama_rocm_v6" and "ollama_rocm_v5" or "ollama_cpu" and "ollama_cpu_avx2"
-// Any library without a variant is the lowest common denominator
-func getAvailableServers() map[string]string {
-	payloadsDir, err := gpu.PayloadsDir()
-	if err != nil {
-		slog.Error("payload lookup error", "error", err)
-		return nil
-	}
-
-	// glob payloadsDir for files that start with ollama_
-	pattern := filepath.Join(payloadsDir, "*", "ollama_*")
-
-	files, err := filepath.Glob(pattern)
-	if err != nil {
-		slog.Debug("could not glob", "pattern", pattern, "error", err)
-		return nil
-	}
-
-	servers := make(map[string]string)
-	for _, file := range files {
-		slog.Debug("availableServers : found", "file", file)
-		servers[filepath.Base(filepath.Dir(file))] = filepath.Dir(file)
-	}
-
-	return servers
-}
-
-// serversForGpu returns a list of compatible servers give the provided GPU
-// info, ordered by performance. assumes Init() has been called
-// TODO - switch to metadata based mapping
-func serversForGpu(info gpu.GpuInfo) []string {
-	// glob workDir for files that start with ollama_
-	availableServers := getAvailableServers()
-	requested := info.Library
-	if info.Variant != gpu.CPUCapabilityNone.String() {
-		requested += "_" + info.Variant
-	}
-
-	servers := []string{}
-
-	// exact match first
-	for a := range availableServers {
-		if a == requested {
-			servers = []string{a}
-
-			if a == "metal" {
-				return servers
-			}
-
-			break
-		}
-	}
-
-	alt := []string{}
-
-	// Then for GPUs load alternates and sort the list for consistent load ordering
-	if info.Library != "cpu" {
-		for a := range availableServers {
-			if info.Library == strings.Split(a, "_")[0] && a != requested {
-				alt = append(alt, a)
-			}
-		}
-
-		slices.Sort(alt)
-		servers = append(servers, alt...)
-	}
-
-	if !(runtime.GOOS == "darwin" && runtime.GOARCH == "arm64") {
-		// Load up the best CPU variant if not primary requested
-		if info.Library != "cpu" {
-			variant := gpu.GetCPUCapability()
-			// If no variant, then we fall back to default
-			// If we have a variant, try that if we find an exact match
-			// Attempting to run the wrong CPU instructions will panic the
-			// process
-			if variant != gpu.CPUCapabilityNone {
-				for cmp := range availableServers {
-					if cmp == "cpu_"+variant.String() {
-						servers = append(servers, cmp)
-						break
-					}
-				}
-			} else {
-				servers = append(servers, "cpu")
-			}
-		}
-
-		if len(servers) == 0 {
-			servers = []string{"cpu"}
-		}
-	}
-
-	return servers
-}
-
-// Return the optimal server for this CPU architecture
-func serverForCpu() string {
-	if runtime.GOOS == "darwin" && runtime.GOARCH == "arm64" {
-		return "metal"
-	}
-	variant := gpu.GetCPUCapability()
-	availableServers := getAvailableServers()
-	if variant != gpu.CPUCapabilityNone {
-		for cmp := range availableServers {
-			if cmp == "cpu_"+variant.String() {
-				return cmp
-			}
-		}
-	}
-	return "cpu"
-}
-
-// extract extracts the embedded files to the target directory
-func extractFiles(targetDir string, glob string) error {
-	files, err := fs.Glob(libEmbed, glob)
-	if err != nil || len(files) == 0 {
-		return errPayloadMissing
-	}
-
-	if err := os.MkdirAll(targetDir, 0o755); err != nil {
-		return fmt.Errorf("extractFiles could not mkdir %s: %v", targetDir, err)
-	}
-
-	g := new(errgroup.Group)
-
-	// build/$OS/$GOARCH/$VARIANT/{bin,lib}/$FILE
-	for _, file := range files {
-		filename := file
-
-		variant := filepath.Base(filepath.Dir(filepath.Dir(filename)))
-
-		slog.Debug("extracting", "variant", variant, "file", filename)
-
-		g.Go(func() error {
-			srcf, err := libEmbed.Open(filename)
-			if err != nil {
-				return err
-			}
-			defer srcf.Close()
-
-			src := io.Reader(srcf)
-			if strings.HasSuffix(filename, ".gz") {
-				src, err = gzip.NewReader(src)
-				if err != nil {
-					return fmt.Errorf("decompress payload %s: %v", filename, err)
-				}
-				filename = strings.TrimSuffix(filename, ".gz")
-			}
-
-			variantDir := filepath.Join(targetDir, variant)
-			if err := os.MkdirAll(variantDir, 0o755); err != nil {
-				return fmt.Errorf("extractFiles could not mkdir %s: %v", variantDir, err)
-			}
-
-			base := filepath.Base(filename)
-			destFilename := filepath.Join(variantDir, base)
-
-			_, err = os.Stat(destFilename)
-			switch {
-			case errors.Is(err, os.ErrNotExist):
-				destFile, err := os.OpenFile(destFilename, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
-				if err != nil {
-					return fmt.Errorf("write payload %s: %v", filename, err)
-				}
-				defer destFile.Close()
-				if _, err := io.Copy(destFile, src); err != nil {
-					return fmt.Errorf("copy payload %s: %v", filename, err)
-				}
-			case err != nil:
-				return fmt.Errorf("stat payload %s: %v", filename, err)
-			}
-			return nil
-		})
-	}
-
-	err = g.Wait()
-	if err != nil {
-		// If we fail to extract, the payload dir is unusable, so cleanup whatever we extracted
-		gpu.Cleanup()
-		return err
-	}
-	return nil
-}
--- a/llm/server.go
+++ b/llm/server.go
@ -24,9 +24,11 @@ import (
 	"golang.org/x/sync/semaphore"

 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/build"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
+	"github.com/ollama/ollama/runners"
 )

 type LlamaServer interface {
@ -106,7 +108,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		gpus = gpu.GetCPUInfo()
 	}
 	if len(gpus) == 1 && gpus[0].Library == "cpu" {
-		cpuRunner = serverForCpu()
+		cpuRunner = runners.ServerForCpu()
 		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
 	} else {
 		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
@ -118,7 +120,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			opts.NumGPU = 0
 		case gpus[0].Library != "metal" && estimate.Layers == 0:
 			// Don't bother loading into the GPU if no layers can fit
-			cpuRunner = serverForCpu()
+			cpuRunner = runners.ServerForCpu()
 			gpus = gpu.GetCPUInfo()
 		case opts.NumGPU < 0 && estimate.Layers > 0 && gpus[0].Library != "cpu":
 			opts.NumGPU = estimate.Layers
@ -145,25 +147,20 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
 	}

-	availableServers := getAvailableServers()
-	if len(availableServers) == 0 {
-		if runtime.GOOS != "windows" {
-			slog.Warn("llama server binary disappeared, reinitializing payloads")
-			err = Init()
+	rDir, err := runners.Refresh(build.EmbedFS)
 	if err != nil {
-				slog.Warn("failed to reinitialize payloads", "error", err)
 		return nil, err
 	}
-			availableServers = getAvailableServers()
-		} else {
+
+	availableServers := runners.GetAvailableServers(rDir)
+	if len(availableServers) == 0 {
 		return nil, finalErr
 	}
-	}
 	var servers []string
 	if cpuRunner != "" {
 		servers = []string{cpuRunner}
 	} else {
-		servers = serversForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
+		servers = runners.ServersForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
 	}
 	demandLib := envconfig.LLMLibrary()
 	if demandLib != "" {
@ -274,7 +271,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--tensor-split", estimate.TensorSplit)
 	}

-	for i := range len(servers) {
+	for i := range servers {
 		dir := availableServers[servers[i]]
 		if dir == "" {
 			// Shouldn't happen
@ -330,7 +327,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		_, err := os.Stat(server)
 		if errors.Is(err, os.ErrNotExist) {
 			slog.Warn("llama server disappeared, reinitializing payloads", "path", server, "error", err)
-			err = Init()
+			_, err = runners.Refresh(build.EmbedFS)
 			if err != nil {
 				slog.Warn("failed to reinitialize payloads", "error", err)
 				return nil, err
--- a/runners/common.go
+++ b/runners/common.go
@ -0,0 +1,384 @@
+package runners
+
+import (
+	"compress/gzip"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"runtime"
+	"slices"
+	"strconv"
+	"strings"
+	"sync"
+	"syscall"
+
+	"golang.org/x/sync/errgroup"
+
+	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/gpu"
+)
+
+const (
+	binGlob = "*/*/*/*"
+)
+
+var (
+	lock       sync.Mutex
+	runnersDir = ""
+)
+
+// Return the location where runners are stored
+// If runners are payloads, this will either extract them
+// or refresh them if any have disappeared due to tmp cleaners
+func Refresh(payloadFS fs.FS) (string, error) {
+	lock.Lock()
+	defer lock.Unlock()
+	var err error
+
+	// Wire up extra logging on our first load
+	if runnersDir == "" {
+		defer func() {
+			var runners []string
+			for v := range GetAvailableServers(runnersDir) {
+				runners = append(runners, v)
+			}
+			slog.Info("Dynamic LLM libraries", "runners", runners)
+			slog.Debug("Override detection logic by setting OLLAMA_LLM_LIBRARY")
+		}()
+	}
+
+	if hasPayloads(payloadFS) {
+		if runnersDir == "" {
+			runnersDir, err = extractRunners(payloadFS)
+		} else {
+			err = refreshRunners(payloadFS, runnersDir)
+		}
+	} else if runnersDir == "" {
+		runnersDir, err = locateRunners()
+	}
+
+	return runnersDir, err
+}
+
+func Cleanup(payloadFS fs.FS) {
+	lock.Lock()
+	defer lock.Unlock()
+	if hasPayloads(payloadFS) && runnersDir != "" {
+		// We want to fully clean up the tmpdir parent of the payloads dir
+		tmpDir := filepath.Clean(filepath.Join(runnersDir, ".."))
+		slog.Debug("cleaning up", "dir", tmpDir)
+		err := os.RemoveAll(tmpDir)
+		if err != nil {
+			slog.Warn("failed to clean up", "dir", tmpDir, "err", err)
+		}
+	}
+}
+
+func locateRunners() (string, error) {
+	exe, err := os.Executable()
+	if err != nil {
+		return "", err
+	}
+
+	cwd, err := os.Getwd()
+	if err != nil {
+		return "", err
+	}
+
+	var paths []string
+	for _, root := range []string{filepath.Dir(exe), filepath.Join(filepath.Dir(exe), envconfig.LibRelativeToExe()), cwd} {
+		paths = append(paths,
+			root,
+			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, "lib", "ollama", "runners")
+		if _, err := os.Stat(candidate); err == nil {
+			return candidate, nil
+		}
+	}
+	return "", fmt.Errorf("unable to locate runners in any search path %v", paths)
+}
+
+// Return true if we're carying nested payloads for the runners
+func hasPayloads(payloadFS fs.FS) bool {
+	files, err := fs.Glob(payloadFS, binGlob)
+	if err != nil || len(files) == 0 || (len(files) == 1 && strings.Contains(files[0], "placeholder")) {
+		return false
+	}
+	return true
+}
+
+func extractRunners(payloadFS fs.FS) (string, error) {
+	cleanupTmpDirs()
+	tmpDir, err := os.MkdirTemp(envconfig.TmpDir(), "ollama")
+	if err != nil {
+		return "", fmt.Errorf("failed to generate tmp dir: %w", err)
+	}
+	// Track our pid so we can clean up orphaned tmpdirs
+	n := filepath.Join(tmpDir, "ollama.pid")
+	if err := os.WriteFile(n, []byte(strconv.Itoa(os.Getpid())), 0o644); err != nil {
+		slog.Warn("failed to write pid file", "file", n, "error", err)
+	}
+	// We create a distinct subdirectory for payloads within the tmpdir
+	// This will typically look like /tmp/ollama3208993108/runners on linux
+	rDir := filepath.Join(tmpDir, "runners")
+
+	slog.Info("extracting embedded files", "dir", rDir)
+	return rDir, refreshRunners(payloadFS, rDir)
+}
+
+func refreshRunners(payloadFS fs.FS, rDir string) error {
+	// extract or refresh server libraries
+	err := extractFiles(payloadFS, rDir, binGlob)
+	if err != nil {
+		return fmt.Errorf("extract binaries: %v", err)
+	}
+	return nil
+}
+
+// extract extracts the embedded files to the target directory
+func extractFiles(payloadFS fs.FS, targetDir string, glob string) error {
+	files, err := fs.Glob(payloadFS, glob)
+	if err != nil || len(files) == 0 {
+		// Should not happen
+		return fmt.Errorf("extractFiles called without payload present")
+	}
+
+	if err := os.MkdirAll(targetDir, 0o755); err != nil {
+		return fmt.Errorf("extractFiles could not mkdir %s: %v", targetDir, err)
+	}
+
+	g := new(errgroup.Group)
+
+	// $OS/$GOARCH/$RUNNER/$FILE
+	for _, file := range files {
+		filename := file
+
+		runner := filepath.Base(filepath.Dir(filename))
+
+		slog.Debug("extracting", "runner", runner, "payload", filename)
+
+		g.Go(func() error {
+			srcf, err := payloadFS.Open(filename)
+			if err != nil {
+				return err
+			}
+			defer srcf.Close()
+
+			src := io.Reader(srcf)
+			if strings.HasSuffix(filename, ".gz") {
+				src, err = gzip.NewReader(src)
+				if err != nil {
+					return fmt.Errorf("decompress payload %s: %v", filename, err)
+				}
+				filename = strings.TrimSuffix(filename, ".gz")
+			}
+
+			runnerDir := filepath.Join(targetDir, runner)
+			if err := os.MkdirAll(runnerDir, 0o755); err != nil {
+				return fmt.Errorf("extractFiles could not mkdir %s: %v", runnerDir, err)
+			}
+
+			base := filepath.Base(filename)
+			destFilename := filepath.Join(runnerDir, base)
+
+			_, err = os.Stat(destFilename)
+			switch {
+			case errors.Is(err, os.ErrNotExist):
+				destFile, err := os.OpenFile(destFilename, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+				if err != nil {
+					return fmt.Errorf("write payload %s: %v", filename, err)
+				}
+				defer destFile.Close()
+				if _, err := io.Copy(destFile, src); err != nil {
+					return fmt.Errorf("copy payload %s: %v", filename, err)
+				}
+			case err != nil:
+				return fmt.Errorf("stat payload %s: %v", filename, err)
+			}
+			return nil
+		})
+	}
+
+	err = g.Wait()
+	if err != nil {
+		slog.Error("failed to extract files", "error", err)
+		// If we fail to extract, the payload dir is most likely unusable, so cleanup whatever we extracted
+		err := os.RemoveAll(targetDir)
+		if err != nil {
+			slog.Warn("failed to cleanup incomplete payload dir", "dir", targetDir, "error", err)
+		}
+		return err
+	}
+	return nil
+}
+
+// Best effort to clean up prior tmpdirs
+func cleanupTmpDirs() {
+	tmpDir := envconfig.TmpDir()
+	if tmpDir == "" {
+		tmpDir = os.TempDir()
+	}
+	matches, err := filepath.Glob(filepath.Join(tmpDir, "ollama*", "ollama.pid"))
+	if err != nil {
+		return
+	}
+
+	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
+		} 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("invalid pid, skipping", "path", match, "error", err)
+			continue
+		}
+
+		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(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)
+		}
+	}
+}
+
+// directory names are the name of the runner and may contain an optional
+// variant prefixed with '_' as the separator. For example, "cuda_v11" and
+// "cuda_v12" or "cpu" and "cpu_avx2". Any library without a variant is the
+// lowest common denominator
+func GetAvailableServers(payloadsDir string) map[string]string {
+	if payloadsDir == "" {
+		slog.Error("empty runner dir")
+		return nil
+	}
+
+	// glob payloadsDir for files that start with ollama_
+	pattern := filepath.Join(payloadsDir, "*", "ollama_*")
+
+	files, err := filepath.Glob(pattern)
+	if err != nil {
+		slog.Debug("could not glob", "pattern", pattern, "error", err)
+		return nil
+	}
+
+	servers := make(map[string]string)
+	for _, file := range files {
+		slog.Debug("availableServers : found", "file", file)
+		servers[filepath.Base(filepath.Dir(file))] = filepath.Dir(file)
+	}
+
+	return servers
+}
+
+// serversForGpu returns a list of compatible servers give the provided GPU
+// info, ordered by performance. assumes Init() has been called
+// TODO - switch to metadata based mapping
+func ServersForGpu(info gpu.GpuInfo) []string {
+	// glob workDir for files that start with ollama_
+	availableServers := GetAvailableServers(runnersDir)
+	requested := info.Library
+	if info.Variant != gpu.CPUCapabilityNone.String() {
+		requested += "_" + info.Variant
+	}
+
+	servers := []string{}
+
+	// exact match first
+	for a := range availableServers {
+		if a == requested {
+			servers = []string{a}
+
+			if a == "metal" {
+				return servers
+			}
+
+			break
+		}
+	}
+
+	alt := []string{}
+
+	// Then for GPUs load alternates and sort the list for consistent load ordering
+	if info.Library != "cpu" {
+		for a := range availableServers {
+			if info.Library == strings.Split(a, "_")[0] && a != requested {
+				alt = append(alt, a)
+			}
+		}
+
+		slices.Sort(alt)
+		servers = append(servers, alt...)
+	}
+
+	if !(runtime.GOOS == "darwin" && runtime.GOARCH == "arm64") {
+		// Load up the best CPU variant if not primary requested
+		if info.Library != "cpu" {
+			variant := gpu.GetCPUCapability()
+			// If no variant, then we fall back to default
+			// If we have a variant, try that if we find an exact match
+			// Attempting to run the wrong CPU instructions will panic the
+			// process
+			if variant != gpu.CPUCapabilityNone {
+				for cmp := range availableServers {
+					if cmp == "cpu_"+variant.String() {
+						servers = append(servers, cmp)
+						break
+					}
+				}
+			} else {
+				servers = append(servers, "cpu")
+			}
+		}
+
+		if len(servers) == 0 {
+			servers = []string{"cpu"}
+		}
+	}
+
+	return servers
+}
+
+// Return the optimal server for this CPU architecture
+func ServerForCpu() string {
+	if runtime.GOOS == "darwin" && runtime.GOARCH == "arm64" {
+		return "metal"
+	}
+	variant := gpu.GetCPUCapability()
+	availableServers := GetAvailableServers(runnersDir)
+	if variant != gpu.CPUCapabilityNone {
+		for cmp := range availableServers {
+			if cmp == "cpu_"+variant.String() {
+				return cmp
+			}
+		}
+	}
+	return "cpu"
+}
--- a/runners/runners_test.go
+++ b/runners/runners_test.go
@ -0,0 +1,50 @@
+package runners
+
+import (
+	"log/slog"
+	"os"
+	"path"
+	"runtime"
+	"strings"
+	"testing"
+	"testing/fstest"
+)
+
+func TestRefreshRunners(t *testing.T) {
+	slog.SetLogLoggerLevel(slog.LevelDebug)
+
+	payloadFS := fstest.MapFS{
+		path.Join(runtime.GOOS, runtime.GOARCH, "foo", "ollama_llama_server"): {Data: []byte("hello, world\n")},
+	}
+	tmpDir, err := os.MkdirTemp("", "testing")
+	if err != nil {
+		t.Fatalf("failed to make tmp dir %s", err)
+	}
+	t.Setenv("OLLAMA_TMPDIR", tmpDir)
+	rDir, err := Refresh(payloadFS)
+	if err != nil {
+		t.Fatalf("failed to extract to %s %s", tmpDir, err)
+	}
+	if !strings.Contains(rDir, tmpDir) {
+		t.Fatalf("runner dir %s was not in tmp dir %s", rDir, tmpDir)
+	}
+
+	// spot check results
+	servers := GetAvailableServers(rDir)
+	if len(servers) < 1 {
+		t.Fatalf("expected at least 1 server")
+	}
+
+	// Refresh contents
+	rDir, err = extractRunners(payloadFS)
+	if err != nil {
+		t.Fatalf("failed to extract to %s %s", tmpDir, err)
+	}
+	if !strings.Contains(rDir, tmpDir) {
+		t.Fatalf("runner dir %s was not in tmp dir %s", rDir, tmpDir)
+	}
+
+	cleanupTmpDirs()
+
+	Cleanup(payloadFS)
+}
--- a/scripts/build_darwin.sh
+++ b/scripts/build_darwin.sh
@ -2,8 +2,7 @@

 set -e

-export VERSION=${VERSION:-$(git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")}
-export GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$VERSION\" \"-X=github.com/ollama/ollama/server.mode=release\"'"
+. $(dirname $0)/env.sh

 mkdir -p dist

--- a/scripts/build_docker.sh
+++ b/scripts/build_docker.sh
@ -2,76 +2,34 @@

 set -eu

-export VERSION=${VERSION:-$(git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")}
-export GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$VERSION\" \"-X=github.com/ollama/ollama/server.mode=release\"'"
-
-# We use 2 different image repositories to handle combining architecture images into multiarch manifest
-# (The ROCm image is x86 only and is not a multiarch manifest)
-# For developers, you can override the DOCKER_ORG to generate multiarch manifests
-#  DOCKER_ORG=jdoe PUSH=1 ./scripts/build_docker.sh
-DOCKER_ORG=${DOCKER_ORG:-"ollama"}
-RELEASE_IMAGE_REPO=${RELEASE_IMAGE_REPO:-"${DOCKER_ORG}/release"}
-FINAL_IMAGE_REPO=${FINAL_IMAGE_REPO:-"${DOCKER_ORG}/ollama"}
-
-BUILD_ARCH=${BUILD_ARCH:-"amd64 arm64"}
+. $(dirname $0)/env.sh

 # Set PUSH to a non-empty string to trigger push instead of load
 PUSH=${PUSH:-""}

-# In CI mode, we break things down
-OLLAMA_SKIP_MANIFEST_CREATE=${OLLAMA_SKIP_MANIFEST_CREATE:-""}
-OLLAMA_SKIP_IMAGE_BUILD=${OLLAMA_SKIP_IMAGE_BUILD:-""}
-
 if [ -z "${PUSH}" ] ; then
+    echo "Building ${FINAL_IMAGE_REPO}:$VERSION locally.  set PUSH=1 to push"
    LOAD_OR_PUSH="--load"
 else
-    echo "Will be pushing ${RELEASE_IMAGE_REPO}:$VERSION for ${BUILD_ARCH}"
+    echo "Will be pushing ${FINAL_IMAGE_REPO}:$VERSION"
    LOAD_OR_PUSH="--push"
 fi

-if [ -z "${OLLAMA_SKIP_IMAGE_BUILD}" ]; then
-    for TARGETARCH in ${BUILD_ARCH}; do
-        docker build \
+docker buildx build \
    ${LOAD_OR_PUSH} \
-            --platform=linux/${TARGETARCH} \
-            --build-arg=VERSION \
-            --build-arg=GOFLAGS \
+    --platform=${PLATFORM} \
+    ${OLLAMA_COMMON_BUILD_ARGS} \
    -f Dockerfile \
-            -t ${RELEASE_IMAGE_REPO}:$VERSION-${TARGETARCH} \
+    -t ${FINAL_IMAGE_REPO}:$VERSION \
    .
-    done

-    if echo ${BUILD_ARCH} | grep "amd64" > /dev/null; then
-        docker build \
+if echo $PLATFORM | grep "amd64" > /dev/null; then
+    docker buildx build \
        ${LOAD_OR_PUSH} \
        --platform=linux/amd64 \
-            --build-arg=VERSION \
-            --build-arg=GOFLAGS \
+        ${OLLAMA_COMMON_BUILD_ARGS} \
        --target runtime-rocm \
        -f Dockerfile \
-            -t ${RELEASE_IMAGE_REPO}:$VERSION-rocm \
+        -t ${FINAL_IMAGE_REPO}:$VERSION-rocm \
        .
-    fi
-fi
-
-if [ -z "${OLLAMA_SKIP_MANIFEST_CREATE}" ]; then
-    if [ -n "${PUSH}" ]; then
-        docker manifest create ${FINAL_IMAGE_REPO}:$VERSION \
-            ${RELEASE_IMAGE_REPO}:$VERSION-amd64 \
-            ${RELEASE_IMAGE_REPO}:$VERSION-arm64
-        docker manifest push ${FINAL_IMAGE_REPO}:$VERSION
-
-        # For symmetry, tag/push the rocm image
-        if [ "${RELEASE_IMAGE_REPO}" != "${FINAL_IMAGE_REPO}" ]; then
-            echo "Tagging and pushing rocm image"
-            docker pull ${RELEASE_IMAGE_REPO}:$VERSION-rocm
-            docker tag ${RELEASE_IMAGE_REPO}:$VERSION-rocm ${FINAL_IMAGE_REPO}:$VERSION-rocm
-            docker push ${FINAL_IMAGE_REPO}:$VERSION-rocm
-        fi
-    else
-        echo "Skipping manifest generation when not pushing images are available locally as "
-        echo "  ${RELEASE_IMAGE_REPO}:$VERSION-amd64"
-        echo "  ${RELEASE_IMAGE_REPO}:$VERSION-arm64"
-        echo "  ${RELEASE_IMAGE_REPO}:$VERSION-rocm"
-    fi
 fi
--- a/scripts/build_linux.sh
+++ b/scripts/build_linux.sh
@ -1,37 +1,29 @@
 #!/bin/sh
+#
+# Mac ARM users, rosetta can be flaky, so to use a remote x86 builder
+#
+# docker context create amd64 --docker host=ssh://mybuildhost
+# docker buildx create --name mybuilder amd64 --platform linux/amd64
+# docker buildx create --name mybuilder --append desktop-linux --platform linux/arm64
+# docker buildx use mybuilder
+

 set -eu

-export VERSION=${VERSION:-$(git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")}
-export GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$VERSION\" \"-X=github.com/ollama/ollama/server.mode=release\"'"
-GZIP=$(which pigz 2>/dev/null || echo "gzip")
+. $(dirname $0)/env.sh

-BUILD_ARCH=${BUILD_ARCH:-"amd64 arm64"}
-export AMDGPU_TARGETS=${AMDGPU_TARGETS:=""}
 mkdir -p dist

-for TARGETARCH in ${BUILD_ARCH}; do
-    docker build \
-        --platform=linux/$TARGETARCH \
-        --build-arg=GOFLAGS \
-        --build-arg=CGO_CFLAGS \
-        --build-arg=OLLAMA_CUSTOM_CPU_DEFS \
-        --build-arg=AMDGPU_TARGETS \
-        --target build-$TARGETARCH \
+docker buildx build \
+        --output type=local,dest=./dist/ \
+        --platform=${PLATFORM} \
+        ${OLLAMA_COMMON_BUILD_ARGS} \
+        --target dist \
        -f Dockerfile \
-        -t builder:$TARGETARCH \
        .
-    docker create --platform linux/$TARGETARCH --name builder-$TARGETARCH builder:$TARGETARCH
-    rm -rf ./dist/linux-$TARGETARCH
-    docker cp builder-$TARGETARCH:/go/src/github.com/ollama/ollama/dist/linux-$TARGETARCH ./dist
-    if echo ${TARGETARCH} | grep "amd64" > /dev/null; then
-        docker cp builder-$TARGETARCH:/go/src/github.com/ollama/ollama/dist/linux-$TARGETARCH-rocm ./dist
-    fi
-    docker rm builder-$TARGETARCH
-    echo "Compressing final linux bundle..."
-    rm -f ./dist/ollama-linux-$TARGETARCH.tgz
-    (cd dist/linux-$TARGETARCH && tar cf - . | ${GZIP} --best > ../ollama-linux-$TARGETARCH.tgz )
-    if [ -d dist/linux-$TARGETARCH-rocm ]; then
-        (cd dist/linux-$TARGETARCH-rocm && tar cf - . | ${GZIP} --best > ../ollama-linux-$TARGETARCH-rocm.tgz )
-    fi
-done
+
+# buildx behavior changes for single vs. multiplatform
+if echo $PLATFORM | grep "," > /dev/null ; then 
+        mv -f ./dist/linux_*64/ollama* ./dist/
+        rmdir ./dist/linux_*64
+fi
--- a/scripts/env.sh
+++ b/scripts/env.sh
@ -0,0 +1,14 @@
+# Common environment setup across build*.sh scripts
+
+export VERSION=${VERSION:-$(git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")}
+export GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$VERSION\" \"-X=github.com/ollama/ollama/server.mode=release\"'"
+# TODO - consider `docker buildx ls --format=json` to autodiscover platform capability
+PLATFORM=${PLATFORM:-"linux/arm64,linux/amd64"}
+DOCKER_ORG=${DOCKER_ORG:-"ollama"}
+RELEASE_IMAGE_REPO=${RELEASE_IMAGE_REPO:-"${DOCKER_ORG}/release"}
+FINAL_IMAGE_REPO=${FINAL_IMAGE_REPO:-"${DOCKER_ORG}/ollama"}
+OLLAMA_COMMON_BUILD_ARGS="--build-arg=VERSION --build-arg=GOFLAGS --build-arg=OLLAMA_CUSTOM_CPU_DEFS --build-arg=AMDGPU_TARGETS"
+
+echo "Building Ollama"
+echo "VERSION=$VERSION"
+echo "PLATFORM=$PLATFORM"
--- a/server/routes.go
+++ b/server/routes.go
@ -26,11 +26,13 @@ import (
 	"golang.org/x/sync/errgroup"

 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/build"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/gpu"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/openai"
 	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/runners"
 	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
@ -117,6 +119,32 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}

+	// expire the runner
+	if req.Prompt == "" && req.KeepAlive != nil && int(req.KeepAlive.Seconds()) == 0 {
+		model, err := GetModel(req.Model)
+		if err != nil {
+			switch {
+			case os.IsNotExist(err):
+				c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Model)})
+			case err.Error() == "invalid model name":
+				c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+			default:
+				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			}
+			return
+		}
+		s.sched.expireRunner(model)
+
+		c.JSON(http.StatusOK, api.GenerateResponse{
+			Model:      req.Model,
+			CreatedAt:  time.Now().UTC(),
+			Response:   "",
+			Done:       true,
+			DoneReason: "unload",
+		})
+		return
+	}
+
 	if req.Format != "" && req.Format != "json" {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "format must be empty or \"json\""})
 		return
@ -1190,12 +1218,12 @@ func Serve(ln net.Listener) error {
 		srvr.Close()
 		schedDone()
 		sched.unloadAllRunners()
-		gpu.Cleanup()
+		runners.Cleanup(build.EmbedFS)
 		done()
 	}()

-	if err := llm.Init(); err != nil {
-		return fmt.Errorf("unable to initialize llm library %w", err)
+	if _, err := runners.Refresh(build.EmbedFS); err != nil {
+		return fmt.Errorf("unable to initialize llm runners %w", err)
 	}

 	s.sched.Run(schedCtx)
@ -1322,6 +1350,32 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}

+	// expire the runner
+	if len(req.Messages) == 0 && req.KeepAlive != nil && int(req.KeepAlive.Seconds()) == 0 {
+		model, err := GetModel(req.Model)
+		if err != nil {
+			switch {
+			case os.IsNotExist(err):
+				c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Model)})
+			case err.Error() == "invalid model name":
+				c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+			default:
+				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			}
+			return
+		}
+		s.sched.expireRunner(model)
+
+		c.JSON(http.StatusOK, api.ChatResponse{
+			Model:      req.Model,
+			CreatedAt:  time.Now().UTC(),
+			Message:    api.Message{Role: "assistant"},
+			Done:       true,
+			DoneReason: "unload",
+		})
+		return
+	}
+
 	caps := []Capability{CapabilityCompletion}
 	if len(req.Tools) > 0 {
 		caps = append(caps, CapabilityTools)
--- a/server/sched.go
+++ b/server/sched.go
@ -360,7 +360,6 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
 			slog.Debug("runner expired event received", "modelPath", runner.modelPath)
 			runner.refMu.Lock()
 			if runner.refCount > 0 {
-				// Shouldn't happen, but safeguard to ensure no leaked runners
 				slog.Debug("expired event with positive ref count, retrying", "modelPath", runner.modelPath, "refCount", runner.refCount)
 				go func(runner *runnerRef) {
 					// We can't unload yet, but want to as soon as the current request completes
@ -802,6 +801,25 @@ func (s *Scheduler) unloadAllRunners() {
 	}
 }

+func (s *Scheduler) expireRunner(model *Model) {
+	s.loadedMu.Lock()
+	defer s.loadedMu.Unlock()
+	runner, ok := s.loaded[model.ModelPath]
+	if ok {
+		runner.refMu.Lock()
+		runner.expiresAt = time.Now()
+		if runner.expireTimer != nil {
+			runner.expireTimer.Stop()
+			runner.expireTimer = nil
+		}
+		runner.sessionDuration = 0
+		if runner.refCount <= 0 {
+			s.expiredCh <- runner
+		}
+		runner.refMu.Unlock()
+	}
+}
+
 // If other runners are loaded, make sure the pending request will fit in system memory
 // If not, pick a runner to unload, else return nil and the request can be loaded
 func (s *Scheduler) maybeFindCPURunnerToUnload(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) *runnerRef {
--- a/server/sched_test.go
+++ b/server/sched_test.go
@ -406,6 +406,52 @@ func TestGetRunner(t *testing.T) {
 	b.ctxDone()
 }

+func TestExpireRunner(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 20*time.Millisecond)
+	defer done()
+	s := InitScheduler(ctx)
+	req := &LlmRequest{
+		ctx:             ctx,
+		model:           &Model{ModelPath: "foo"},
+		opts:            api.DefaultOptions(),
+		successCh:       make(chan *runnerRef, 1),
+		errCh:           make(chan error, 1),
+		sessionDuration: &api.Duration{Duration: 2 * time.Minute},
+	}
+
+	var ggml *llm.GGML
+	gpus := gpu.GpuInfoList{}
+	server := &mockLlm{estimatedVRAM: 10, estimatedVRAMByGPU: map[string]uint64{}}
+	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
+		return server, nil
+	}
+	s.load(req, ggml, gpus, 0)
+
+	select {
+	case err := <-req.errCh:
+		if err != nil {
+			t.Fatalf("expected no errors when loading, got '%s'", err.Error())
+		}
+	case resp := <-req.successCh:
+		s.loadedMu.Lock()
+		if resp.refCount != uint(1) || len(s.loaded) != 1 {
+			t.Fatalf("expected a model to be loaded")
+		}
+		s.loadedMu.Unlock()
+	}
+
+	s.expireRunner(&Model{ModelPath: "foo"})
+
+	s.finishedReqCh <- req
+	s.processCompleted(ctx)
+
+	s.loadedMu.Lock()
+	if len(s.loaded) != 0 {
+		t.Fatalf("expected model to be unloaded")
+	}
+	s.loadedMu.Unlock()
+}
+
 // TODO - add one scenario that triggers the bogus finished event with positive ref count
 func TestPrematureExpired(t *testing.T) {
 	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
				`@ -0,0 +1 @@`
				`This is here to make sure the build/ directory exists for the go:embed command`