thinking: force newer qwen3 models to always use thinking

openai: refactor to split compat layer and middleware

.github/workflows/release.yaml

@@ -94,7 +94,7 @@ jobs:
             install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
             rocm-version: '6.2'
             flags: '-DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
-            runner_dir: ''
+            runner_dir: 'rocm'
     runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
     environment: release
     env:
@@ -163,7 +163,7 @@ jobs:
           cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }} -DOLLAMA_RUNNER_DIR="${{ matrix.runner_dir }}"
           cmake --build --parallel --preset "${{ matrix.preset }}"
           cmake --install build --component "${{ startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || 'CPU' }}" --strip --parallel 8
-          rm -f dist\lib\ollama\rocm\rocblas\library\*gfx906*
+          Remove-Item -Path dist\lib\ollama\rocm\rocblas\library\*gfx906* -ErrorAction SilentlyContinue
         env:
           CMAKE_GENERATOR: Ninja
       - uses: actions/upload-artifact@v4
@@ -176,19 +176,19 @@ jobs:
       matrix:
         os: [windows]
         arch: [amd64, arm64]
+        include:
+        - os: windows
+          arch: amd64
+          llvmarch: x86_64
+        - os: windows
+          arch: arm64
+          llvmarch: aarch64
     runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
     environment: release
     needs: [setup-environment]
     env:
       GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
     steps:
-      - name: Install AMD64 system dependencies
-        if: matrix.arch == 'amd64'
-        run: |
-          $ErrorActionPreference = "Stop"
-          Start-Process "C:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
-          echo "C:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
       - name: Install ARM64 system dependencies
         if: matrix.arch == 'arm64'
         run: |
@@ -200,15 +200,29 @@ jobs:
 
           choco install -y --no-progress git gzip
           echo "C:\Program Files\Git\cmd" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-
-          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-aarch64.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt-aarch64.zip"
-          Expand-Archive -Path ${{ runner.temp }}\llvm-mingw-ucrt-aarch64.zip -DestinationPath "C:\Program Files\"
-          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt-aarch64").path
-          echo $installPath\bin | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install clang and gcc-compat
+        run: |
+          $ErrorActionPreference = "Stop"
+          Set-ExecutionPolicy Bypass -Scope Process -Force
+          Invoke-WebRequest -Uri "https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-${{ matrix.llvmarch }}.zip" -OutFile "${{ runner.temp }}\llvm-mingw-ucrt.zip"
+          Expand-Archive -Path ${{ runner.temp }}\llvm-mingw-ucrt.zip -DestinationPath "C:\Program Files\"
+          $installPath=(Resolve-Path -Path "C:\Program Files\llvm-mingw-*-ucrt*").path
+          echo "$installPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod
+      - name: Verify gcc is actually clang
+        run: |
+          $ErrorActionPreference='Continue'
+          $version=& gcc -v 2>&1
+          $version=$version -join "`n"
+          echo "gcc is $version"
+          if ($version -notmatch 'clang') {
+            echo "ERROR: GCC must be clang for proper utf16 handling"
+            exit 1
+          }
+          $ErrorActionPreference='Stop'
       - run: |
           go build -o dist/${{ matrix.os }}-${{ matrix.arch }}/ .
       - uses: actions/upload-artifact@v4

discover/runner.go

@@ -330,6 +330,9 @@ func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.Dev
 		}
 	}
 
+	// Apply any iGPU workarounds
+	iGPUWorkarounds(devices)
+
 	return devices
 }
 
@@ -439,15 +442,18 @@ func bootstrapDevices(ctx context.Context, ollamaLibDirs []string, extraEnvs []s
 		cmd.Stderr = os.Stderr
 	}
 	// cmd.SysProcAttr = llm.LlamaServerSysProcAttr // circular dependency - bring back once refactored
-	cmd.Env = append(cmd.Env, "OLLAMA_LIBRARY_PATH="+strings.Join(ollamaLibDirs, string(filepath.ListSeparator)))
 	pathEnvVal := strings.Join(libraryPaths, string(filepath.ListSeparator))
 	pathNeeded := true
+	ollamaPathNeeded := true
 	extraDone := make([]bool, len(extraEnvs))
 	for i := range cmd.Env {
 		cmp := strings.SplitN(cmd.Env[i], "=", 2)
 		if strings.EqualFold(cmp[0], pathEnv) {
 			cmd.Env[i] = pathEnv + "=" + pathEnvVal
 			pathNeeded = false
+		} else if strings.EqualFold(cmp[0], "OLLAMA_LIBRARY_PATH") {
+			cmd.Env[i] = "OLLAMA_LIBRARY_PATH=" + strings.Join(ollamaLibDirs, string(filepath.ListSeparator))
+			ollamaPathNeeded = false
 		} else {
 			for j := range extraEnvs {
 				if extraDone[j] {
@@ -464,6 +470,9 @@ func bootstrapDevices(ctx context.Context, ollamaLibDirs []string, extraEnvs []s
 	if pathNeeded {
 		cmd.Env = append(cmd.Env, pathEnv+"="+pathEnvVal)
 	}
+	if ollamaPathNeeded {
+		cmd.Env = append(cmd.Env, "OLLAMA_LIBRARY_PATH="+strings.Join(ollamaLibDirs, string(filepath.ListSeparator)))
+	}
 	for i := range extraDone {
 		if !extraDone[i] {
 			cmd.Env = append(cmd.Env, extraEnvs[i])
@@ -540,3 +549,32 @@ func GetDevicesFromRunner(ctx context.Context, runner BaseRunner) ([]ml.DeviceIn
 		}
 	}
 }
+
+func iGPUWorkarounds(devices []ml.DeviceInfo) {
+	// short circuit if we have no iGPUs
+	anyiGPU := false
+	for i := range devices {
+		if devices[i].Integrated {
+			anyiGPU = true
+			break
+		}
+	}
+	if !anyiGPU {
+		return
+	}
+
+	memInfo, err := GetCPUMem()
+	if err != nil {
+		slog.Debug("failed to fetch system memory information for iGPU", "error", err)
+		return
+	}
+	for i := range devices {
+		if !devices[i].Integrated {
+			continue
+		}
+		// NVIDIA iGPUs return useless free VRAM data which ignores system buff/cache
+		if devices[i].Library == "CUDA" {
+			devices[i].FreeMemory = memInfo.FreeMemory
+		}
+	}
+}

fs/ggml/ggml.go

@@ -870,11 +870,6 @@ func (f GGML) SupportsKVCacheType(cacheType string) bool {
 		return true
 	}
 
-	if arch := f.KV().Architecture(); slices.Contains([]string{"gptoss", "gpt-oss"}, arch) {
-		// gpt-oss uses attention with sinks which does not support quantized cache types
-		slog.Warn("model only supports non-quantized cache types", "model", arch)
-		return false
-	}
 	return slices.Contains([]string{"q8_0", "q4_0"}, cacheType)
 }
 

llm/server.go

@@ -359,20 +359,22 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		s.cmd.Stderr = s.status
 		s.cmd.SysProcAttr = LlamaServerSysProcAttr
 
-		s.cmd.Env = append(s.cmd.Env, "OLLAMA_LIBRARY_PATH="+strings.Join(ggmlPaths, string(filepath.ListSeparator)))
-
 		// Always filter down the set of GPUs in case there are any unsupported devices that might crash
 		envWorkarounds := gpus.GetVisibleDevicesEnv()
 		pathEnvVal := strings.Join(libraryPaths, string(filepath.ListSeparator))
 
 		// Update or add the path variable with our adjusted version
 		pathNeeded := true
+		ollamaPathNeeded := true
 		envWorkaroundDone := make([]bool, len(envWorkarounds))
 		for i := range s.cmd.Env {
 			cmp := strings.SplitN(s.cmd.Env[i], "=", 2)
 			if strings.EqualFold(cmp[0], pathEnv) {
 				s.cmd.Env[i] = pathEnv + "=" + pathEnvVal
 				pathNeeded = false
+			} else if strings.EqualFold(cmp[0], "OLLAMA_LIBRARY_PATH") {
+				s.cmd.Env[i] = "OLLAMA_LIBRARY_PATH=" + strings.Join(ggmlPaths, string(filepath.ListSeparator))
+				ollamaPathNeeded = false
 			} else if len(envWorkarounds) != 0 {
 				for j, kv := range envWorkarounds {
 					tmp := strings.SplitN(kv, "=", 2)
@@ -386,6 +388,9 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		if pathNeeded {
 			s.cmd.Env = append(s.cmd.Env, pathEnv+"="+pathEnvVal)
 		}
+		if ollamaPathNeeded {
+			s.cmd.Env = append(s.cmd.Env, "OLLAMA_LIBRARY_PATH="+strings.Join(ggmlPaths, string(filepath.ListSeparator)))
+		}
 		for i, done := range envWorkaroundDone {
 			if !done {
 				s.cmd.Env = append(s.cmd.Env, envWorkarounds[i])

middleware/openai.go

@@ -0,0 +1,424 @@
+package middleware
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"math/rand"
+	"net/http"
+
+	"github.com/gin-gonic/gin"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/openai"
+)
+
+type BaseWriter struct {
+	gin.ResponseWriter
+}
+
+type ChatWriter struct {
+	stream        bool
+	streamOptions *openai.StreamOptions
+	id            string
+	toolCallSent  bool
+	BaseWriter
+}
+
+type CompleteWriter struct {
+	stream        bool
+	streamOptions *openai.StreamOptions
+	id            string
+	BaseWriter
+}
+
+type ListWriter struct {
+	BaseWriter
+}
+
+type RetrieveWriter struct {
+	BaseWriter
+	model string
+}
+
+type EmbedWriter struct {
+	BaseWriter
+	model string
+}
+
+func (w *BaseWriter) writeError(data []byte) (int, error) {
+	var serr api.StatusError
+	err := json.Unmarshal(data, &serr)
+	if err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.NewError(http.StatusInternalServerError, serr.Error()))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *ChatWriter) writeResponse(data []byte) (int, error) {
+	var chatResponse api.ChatResponse
+	err := json.Unmarshal(data, &chatResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// chat chunk
+	if w.stream {
+		c := openai.ToChunk(w.id, chatResponse, w.toolCallSent)
+		d, err := json.Marshal(c)
+		if err != nil {
+			return 0, err
+		}
+		if !w.toolCallSent && len(c.Choices) > 0 && len(c.Choices[0].Delta.ToolCalls) > 0 {
+			w.toolCallSent = true
+		}
+
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+		_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
+		if err != nil {
+			return 0, err
+		}
+
+		if chatResponse.Done {
+			if w.streamOptions != nil && w.streamOptions.IncludeUsage {
+				u := openai.ToUsage(chatResponse)
+				c.Usage = &u
+				c.Choices = []openai.ChunkChoice{}
+				d, err := json.Marshal(c)
+				if err != nil {
+					return 0, err
+				}
+				_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
+				if err != nil {
+					return 0, err
+				}
+			}
+			_, err = w.ResponseWriter.Write([]byte("data: [DONE]\n\n"))
+			if err != nil {
+				return 0, err
+			}
+		}
+
+		return len(data), nil
+	}
+
+	// chat completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.ToChatCompletion(w.id, chatResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *ChatWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *CompleteWriter) writeResponse(data []byte) (int, error) {
+	var generateResponse api.GenerateResponse
+	err := json.Unmarshal(data, &generateResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// completion chunk
+	if w.stream {
+		c := openai.ToCompleteChunk(w.id, generateResponse)
+		if w.streamOptions != nil && w.streamOptions.IncludeUsage {
+			c.Usage = &openai.Usage{}
+		}
+		d, err := json.Marshal(c)
+		if err != nil {
+			return 0, err
+		}
+
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+		_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
+		if err != nil {
+			return 0, err
+		}
+
+		if generateResponse.Done {
+			if w.streamOptions != nil && w.streamOptions.IncludeUsage {
+				u := openai.ToUsageGenerate(generateResponse)
+				c.Usage = &u
+				c.Choices = []openai.CompleteChunkChoice{}
+				d, err := json.Marshal(c)
+				if err != nil {
+					return 0, err
+				}
+				_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
+				if err != nil {
+					return 0, err
+				}
+			}
+			_, err = w.ResponseWriter.Write([]byte("data: [DONE]\n\n"))
+			if err != nil {
+				return 0, err
+			}
+		}
+
+		return len(data), nil
+	}
+
+	// completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.ToCompletion(w.id, generateResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *CompleteWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *ListWriter) writeResponse(data []byte) (int, error) {
+	var listResponse api.ListResponse
+	err := json.Unmarshal(data, &listResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.ToListCompletion(listResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *ListWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *RetrieveWriter) writeResponse(data []byte) (int, error) {
+	var showResponse api.ShowResponse
+	err := json.Unmarshal(data, &showResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// retrieve completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.ToModel(showResponse, w.model))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *RetrieveWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *EmbedWriter) writeResponse(data []byte) (int, error) {
+	var embedResponse api.EmbedResponse
+	err := json.Unmarshal(data, &embedResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(openai.ToEmbeddingList(w.model, embedResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *EmbedWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func ListMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		w := &ListWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func RetrieveMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(api.ShowRequest{Name: c.Param("model")}); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &RetrieveWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			model:      c.Param("model"),
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func CompletionsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.CompletionRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		var b bytes.Buffer
+		genReq, err := openai.FromCompleteRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if err := json.NewEncoder(&b).Encode(genReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &CompleteWriter{
+			BaseWriter:    BaseWriter{ResponseWriter: c.Writer},
+			stream:        req.Stream,
+			id:            fmt.Sprintf("cmpl-%d", rand.Intn(999)),
+			streamOptions: req.StreamOptions,
+		}
+
+		c.Writer = w
+		c.Next()
+	}
+}
+
+func EmbeddingsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.EmbedRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Input == "" {
+			req.Input = []string{""}
+		}
+
+		if req.Input == nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "invalid input"))
+			return
+		}
+
+		if v, ok := req.Input.([]any); ok && len(v) == 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "invalid input"))
+			return
+		}
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(api.EmbedRequest{Model: req.Model, Input: req.Input, Dimensions: req.Dimensions}); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &EmbedWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			model:      req.Model,
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func ChatMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.ChatCompletionRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if len(req.Messages) == 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, "[] is too short - 'messages'"))
+			return
+		}
+
+		var b bytes.Buffer
+
+		chatReq, err := openai.FromChatRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &ChatWriter{
+			BaseWriter:    BaseWriter{ResponseWriter: c.Writer},
+			stream:        req.Stream,
+			id:            fmt.Sprintf("chatcmpl-%d", rand.Intn(999)),
+			streamOptions: req.StreamOptions,
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}

middleware/openai_test.go

@@ -0,0 +1,928 @@
+package middleware
+
+import (
+	"bytes"
+	"encoding/base64"
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/openai"
+)
+
+const (
+	prefix = `data:image/jpeg;base64,`
+	image  = `iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=`
+)
+
+var (
+	False = false
+	True  = true
+)
+
+func captureRequestMiddleware(capturedRequest any) gin.HandlerFunc {
+	return func(c *gin.Context) {
+		bodyBytes, _ := io.ReadAll(c.Request.Body)
+		c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+		err := json.Unmarshal(bodyBytes, capturedRequest)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, "failed to unmarshal request")
+		}
+		c.Next()
+	}
+}
+
+func TestChatMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.ChatRequest
+		err  openai.ErrorResponse
+	}
+
+	var capturedRequest *api.ChatRequest
+
+	testCases := []testCase{
+		{
+			name: "chat handler",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Hello",
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "chat handler with options",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				],
+				"stream":            true,
+				"max_tokens":        999,
+				"seed":              123,
+				"stop":              ["\n", "stop"],
+				"temperature":       3.0,
+				"frequency_penalty": 4.0,
+				"presence_penalty":  5.0,
+				"top_p":             6.0,
+				"response_format":   {"type": "json_object"}
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Hello",
+					},
+				},
+				Options: map[string]any{
+					"num_predict":       999.0, // float because JSON doesn't distinguish between float and int
+					"seed":              123.0,
+					"stop":              []any{"\n", "stop"},
+					"temperature":       3.0,
+					"frequency_penalty": 4.0,
+					"presence_penalty":  5.0,
+					"top_p":             6.0,
+				},
+				Format: json.RawMessage(`"json"`),
+				Stream: &True,
+			},
+		},
+		{
+			name: "chat handler with streaming usage",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				],
+				"stream":            true,
+				"stream_options":    {"include_usage": true},
+				"max_tokens":        999,
+				"seed":              123,
+				"stop":              ["\n", "stop"],
+				"temperature":       3.0,
+				"frequency_penalty": 4.0,
+				"presence_penalty":  5.0,
+				"top_p":             6.0,
+				"response_format":   {"type": "json_object"}
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Hello",
+					},
+				},
+				Options: map[string]any{
+					"num_predict":       999.0, // float because JSON doesn't distinguish between float and int
+					"seed":              123.0,
+					"stop":              []any{"\n", "stop"},
+					"temperature":       3.0,
+					"frequency_penalty": 4.0,
+					"presence_penalty":  5.0,
+					"top_p":             6.0,
+				},
+				Format: json.RawMessage(`"json"`),
+				Stream: &True,
+			},
+		},
+		{
+			name: "chat handler with image content",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{
+						"role": "user",
+						"content": [
+							{
+								"type": "text",
+								"text": "Hello"
+							},
+							{
+								"type": "image_url",
+								"image_url": {
+									"url": "` + prefix + image + `"
+								}
+							}
+						]
+					}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Hello",
+					},
+					{
+						Role: "user",
+						Images: []api.ImageData{
+							func() []byte {
+								img, _ := base64.StdEncoding.DecodeString(image)
+								return img
+							}(),
+						},
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "chat handler with tools",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "What's the weather like in Paris Today?"},
+					{"role": "assistant", "tool_calls": [{"id": "id", "type": "function", "function": {"name": "get_current_weather", "arguments": "{\"location\": \"Paris, France\", \"format\": \"celsius\"}"}}]}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "What's the weather like in Paris Today?",
+					},
+					{
+						Role: "assistant",
+						ToolCalls: []api.ToolCall{
+							{
+								Function: api.ToolCallFunction{
+									Name: "get_current_weather",
+									Arguments: map[string]any{
+										"location": "Paris, France",
+										"format":   "celsius",
+									},
+								},
+							},
+						},
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "chat handler with tools and content",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "What's the weather like in Paris Today?"},
+					{"role": "assistant", "content": "Let's see what the weather is like in Paris", "tool_calls": [{"id": "id", "type": "function", "function": {"name": "get_current_weather", "arguments": "{\"location\": \"Paris, France\", \"format\": \"celsius\"}"}}]}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "What's the weather like in Paris Today?",
+					},
+					{
+						Role:    "assistant",
+						Content: "Let's see what the weather is like in Paris",
+						ToolCalls: []api.ToolCall{
+							{
+								Function: api.ToolCallFunction{
+									Name: "get_current_weather",
+									Arguments: map[string]any{
+										"location": "Paris, France",
+										"format":   "celsius",
+									},
+								},
+							},
+						},
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "chat handler with tools and empty content",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "What's the weather like in Paris Today?"},
+					{"role": "assistant", "content": "", "tool_calls": [{"id": "id", "type": "function", "function": {"name": "get_current_weather", "arguments": "{\"location\": \"Paris, France\", \"format\": \"celsius\"}"}}]}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "What's the weather like in Paris Today?",
+					},
+					{
+						Role: "assistant",
+						ToolCalls: []api.ToolCall{
+							{
+								Function: api.ToolCallFunction{
+									Name: "get_current_weather",
+									Arguments: map[string]any{
+										"location": "Paris, France",
+										"format":   "celsius",
+									},
+								},
+							},
+						},
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "chat handler with tools and thinking content",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "What's the weather like in Paris Today?"},
+					{"role": "assistant", "reasoning": "Let's see what the weather is like in Paris", "tool_calls": [{"id": "id", "type": "function", "function": {"name": "get_current_weather", "arguments": "{\"location\": \"Paris, France\", \"format\": \"celsius\"}"}}]}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "What's the weather like in Paris Today?",
+					},
+					{
+						Role:     "assistant",
+						Thinking: "Let's see what the weather is like in Paris",
+						ToolCalls: []api.ToolCall{
+							{
+								Function: api.ToolCallFunction{
+									Name: "get_current_weather",
+									Arguments: map[string]any{
+										"location": "Paris, France",
+										"format":   "celsius",
+									},
+								},
+							},
+						},
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "tool response with call ID",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "What's the weather like in Paris Today?"},
+					{"role": "assistant", "tool_calls": [{"id": "id_abc", "type": "function", "function": {"name": "get_current_weather", "arguments": "{\"location\": \"Paris, France\", \"format\": \"celsius\"}"}}]},
+					{"role": "tool", "tool_call_id": "id_abc", "content": "The weather in Paris is 20 degrees Celsius"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "What's the weather like in Paris Today?",
+					},
+					{
+						Role: "assistant",
+						ToolCalls: []api.ToolCall{
+							{
+								Function: api.ToolCallFunction{
+									Name: "get_current_weather",
+									Arguments: map[string]any{
+										"location": "Paris, France",
+										"format":   "celsius",
+									},
+								},
+							},
+						},
+					},
+					{
+						Role:     "tool",
+						Content:  "The weather in Paris is 20 degrees Celsius",
+						ToolName: "get_current_weather",
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "tool response with name",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "What's the weather like in Paris Today?"},
+					{"role": "assistant", "tool_calls": [{"id": "id", "type": "function", "function": {"name": "get_current_weather", "arguments": "{\"location\": \"Paris, France\", \"format\": \"celsius\"}"}}]},
+					{"role": "tool", "name": "get_current_weather", "content": "The weather in Paris is 20 degrees Celsius"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "What's the weather like in Paris Today?",
+					},
+					{
+						Role: "assistant",
+						ToolCalls: []api.ToolCall{
+							{
+								Function: api.ToolCallFunction{
+									Name: "get_current_weather",
+									Arguments: map[string]any{
+										"location": "Paris, France",
+										"format":   "celsius",
+									},
+								},
+							},
+						},
+					},
+					{
+						Role:     "tool",
+						Content:  "The weather in Paris is 20 degrees Celsius",
+						ToolName: "get_current_weather",
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "chat handler with streaming tools",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "What's the weather like in Paris?"}
+				],
+				"stream": true,
+				"tools": [{
+					"type": "function",
+					"function": {
+						"name": "get_weather",
+						"description": "Get the current weather",
+						"parameters": {
+							"type": "object",
+							"required": ["location"],
+							"properties": {
+								"location": {
+									"type": "string",
+									"description": "The city and state"
+								},
+								"unit": {
+									"type": "string",
+									"enum": ["celsius", "fahrenheit"]
+								}
+							}
+						}
+					}
+				}]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "What's the weather like in Paris?",
+					},
+				},
+				Tools: []api.Tool{
+					{
+						Type: "function",
+						Function: api.ToolFunction{
+							Name:        "get_weather",
+							Description: "Get the current weather",
+							Parameters: struct {
+								Type       string                      `json:"type"`
+								Defs       any                         `json:"$defs,omitempty"`
+								Items      any                         `json:"items,omitempty"`
+								Required   []string                    `json:"required"`
+								Properties map[string]api.ToolProperty `json:"properties"`
+							}{
+								Type:     "object",
+								Required: []string{"location"},
+								Properties: map[string]api.ToolProperty{
+									"location": {
+										Type:        api.PropertyType{"string"},
+										Description: "The city and state",
+									},
+									"unit": {
+										Type: api.PropertyType{"string"},
+										Enum: []any{"celsius", "fahrenheit"},
+									},
+								},
+							},
+						},
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &True,
+			},
+		},
+		{
+			name: "chat handler error forwarding",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": 2}
+				]
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "invalid message content type: float64",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(ChatMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/chat", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/chat", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			defer func() { capturedRequest = nil }()
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			var errResp openai.ErrorResponse
+			if resp.Code != http.StatusOK {
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatal(err)
+				}
+				return
+			}
+			if diff := cmp.Diff(&tc.req, capturedRequest); diff != "" {
+				t.Fatalf("requests did not match: %+v", diff)
+			}
+			if diff := cmp.Diff(tc.err, errResp); diff != "" {
+				t.Fatalf("errors did not match for %s:\n%s", tc.name, diff)
+			}
+		})
+	}
+}
+
+func TestCompletionsMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.GenerateRequest
+		err  openai.ErrorResponse
+	}
+
+	var capturedRequest *api.GenerateRequest
+
+	testCases := []testCase{
+		{
+			name: "completions handler",
+			body: `{
+				"model": "test-model",
+				"prompt": "Hello",
+				"temperature": 0.8,
+				"stop": ["\n", "stop"],
+				"suffix": "suffix"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "Hello",
+				Options: map[string]any{
+					"frequency_penalty": 0.0,
+					"presence_penalty":  0.0,
+					"temperature":       0.8,
+					"top_p":             1.0,
+					"stop":              []any{"\n", "stop"},
+				},
+				Suffix: "suffix",
+				Stream: &False,
+			},
+		},
+		{
+			name: "completions handler stream",
+			body: `{
+				"model": "test-model",
+				"prompt": "Hello",
+				"stream": true,
+				"temperature": 0.8,
+				"stop": ["\n", "stop"],
+				"suffix": "suffix"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "Hello",
+				Options: map[string]any{
+					"frequency_penalty": 0.0,
+					"presence_penalty":  0.0,
+					"temperature":       0.8,
+					"top_p":             1.0,
+					"stop":              []any{"\n", "stop"},
+				},
+				Suffix: "suffix",
+				Stream: &True,
+			},
+		},
+		{
+			name: "completions handler stream with usage",
+			body: `{
+				"model": "test-model",
+				"prompt": "Hello",
+				"stream": true,
+				"stream_options": {"include_usage": true},
+				"temperature": 0.8,
+				"stop": ["\n", "stop"],
+				"suffix": "suffix"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "Hello",
+				Options: map[string]any{
+					"frequency_penalty": 0.0,
+					"presence_penalty":  0.0,
+					"temperature":       0.8,
+					"top_p":             1.0,
+					"stop":              []any{"\n", "stop"},
+				},
+				Suffix: "suffix",
+				Stream: &True,
+			},
+		},
+		{
+			name: "completions handler error forwarding",
+			body: `{
+				"model": "test-model",
+				"prompt": "Hello",
+				"temperature": null,
+				"stop": [1, 2],
+				"suffix": "suffix"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "invalid type for 'stop' field: float64",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(CompletionsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/generate", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/generate", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			var errResp openai.ErrorResponse
+			if resp.Code != http.StatusOK {
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatal(err)
+				}
+			}
+
+			if capturedRequest != nil && !reflect.DeepEqual(tc.req, *capturedRequest) {
+				t.Fatal("requests did not match")
+			}
+
+			if !reflect.DeepEqual(tc.err, errResp) {
+				t.Fatal("errors did not match")
+			}
+
+			capturedRequest = nil
+		})
+	}
+}
+
+func TestEmbeddingsMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.EmbedRequest
+		err  openai.ErrorResponse
+	}
+
+	var capturedRequest *api.EmbedRequest
+
+	testCases := []testCase{
+		{
+			name: "embed handler single input",
+			body: `{
+				"input": "Hello",
+				"model": "test-model"
+			}`,
+			req: api.EmbedRequest{
+				Input: "Hello",
+				Model: "test-model",
+			},
+		},
+		{
+			name: "embed handler batch input",
+			body: `{
+				"input": ["Hello", "World"],
+				"model": "test-model"
+			}`,
+			req: api.EmbedRequest{
+				Input: []any{"Hello", "World"},
+				Model: "test-model",
+			},
+		},
+		{
+			name: "embed handler error forwarding",
+			body: `{
+				"model": "test-model"
+			}`,
+			err: openai.ErrorResponse{
+				Error: openai.Error{
+					Message: "invalid input",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(EmbeddingsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/embed", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/embed", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			var errResp openai.ErrorResponse
+			if resp.Code != http.StatusOK {
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatal(err)
+				}
+			}
+
+			if capturedRequest != nil && !reflect.DeepEqual(tc.req, *capturedRequest) {
+				t.Fatal("requests did not match")
+			}
+
+			if !reflect.DeepEqual(tc.err, errResp) {
+				t.Fatal("errors did not match")
+			}
+
+			capturedRequest = nil
+		})
+	}
+}
+
+func TestListMiddleware(t *testing.T) {
+	type testCase struct {
+		name     string
+		endpoint func(c *gin.Context)
+		resp     string
+	}
+
+	testCases := []testCase{
+		{
+			name: "list handler",
+			endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusOK, api.ListResponse{
+					Models: []api.ListModelResponse{
+						{
+							Name:       "test-model",
+							ModifiedAt: time.Unix(int64(1686935002), 0).UTC(),
+						},
+					},
+				})
+			},
+			resp: `{
+				"object": "list",
+				"data": [
+					{
+						"id": "test-model",
+						"object": "model",
+						"created": 1686935002,
+						"owned_by": "library"
+					}
+				]
+			}`,
+		},
+		{
+			name: "list handler empty output",
+			endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusOK, api.ListResponse{})
+			},
+			resp: `{
+				"object": "list",
+				"data": null
+			}`,
+		},
+	}
+
+	gin.SetMode(gin.TestMode)
+
+	for _, tc := range testCases {
+		router := gin.New()
+		router.Use(ListMiddleware())
+		router.Handle(http.MethodGet, "/api/tags", tc.endpoint)
+		req, _ := http.NewRequest(http.MethodGet, "/api/tags", nil)
+
+		resp := httptest.NewRecorder()
+		router.ServeHTTP(resp, req)
+
+		var expected, actual map[string]any
+		err := json.Unmarshal([]byte(tc.resp), &expected)
+		if err != nil {
+			t.Fatalf("failed to unmarshal expected response: %v", err)
+		}
+
+		err = json.Unmarshal(resp.Body.Bytes(), &actual)
+		if err != nil {
+			t.Fatalf("failed to unmarshal actual response: %v", err)
+		}
+
+		if !reflect.DeepEqual(expected, actual) {
+			t.Errorf("responses did not match\nExpected: %+v\nActual: %+v", expected, actual)
+		}
+	}
+}
+
+func TestRetrieveMiddleware(t *testing.T) {
+	type testCase struct {
+		name     string
+		endpoint func(c *gin.Context)
+		resp     string
+	}
+
+	testCases := []testCase{
+		{
+			name: "retrieve handler",
+			endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusOK, api.ShowResponse{
+					ModifiedAt: time.Unix(int64(1686935002), 0).UTC(),
+				})
+			},
+			resp: `{
+				"id":"test-model",
+				"object":"model",
+				"created":1686935002,
+				"owned_by":"library"}
+			`,
+		},
+		{
+			name: "retrieve handler error forwarding",
+			endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusBadRequest, gin.H{"error": "model not found"})
+			},
+			resp: `{
+				"error": {
+				  "code": null,
+				  "message": "model not found",
+				  "param": null,
+				  "type": "api_error"
+				}
+			}`,
+		},
+	}
+
+	gin.SetMode(gin.TestMode)
+
+	for _, tc := range testCases {
+		router := gin.New()
+		router.Use(RetrieveMiddleware())
+		router.Handle(http.MethodGet, "/api/show/:model", tc.endpoint)
+		req, _ := http.NewRequest(http.MethodGet, "/api/show/test-model", nil)
+
+		resp := httptest.NewRecorder()
+		router.ServeHTTP(resp, req)
+
+		var expected, actual map[string]any
+		err := json.Unmarshal([]byte(tc.resp), &expected)
+		if err != nil {
+			t.Fatalf("failed to unmarshal expected response: %v", err)
+		}
+
+		err = json.Unmarshal(resp.Body.Bytes(), &actual)
+		if err != nil {
+			t.Fatalf("failed to unmarshal actual response: %v", err)
+		}
+
+		if !reflect.DeepEqual(expected, actual) {
+			t.Errorf("responses did not match\nExpected: %+v\nActual: %+v", expected, actual)
+		}
+	}
+}

model/models/deepseek2/model.go

@@ -150,7 +150,9 @@ func (moe *sparse) Moe(ctx ml.Context, hiddenStates, topKIndices, topKWeights ml
 }
 
 func (moe *sparse) topKIndices(ctx ml.Context, scores ml.Tensor, opts *Options) ml.Tensor {
-	scores = scores.Add(ctx, moe.ExpProbsBias)
+	if moe.ExpProbsBias != nil {
+		scores = scores.Add(ctx, moe.ExpProbsBias)
+	}
 	topKIndices := scores.TopK(ctx, opts.numExpertsUsed)
 	return topKIndices
 }

openai/openai.go

@@ -1,21 +1,17 @@
-// openai package provides middleware for partial compatibility with the OpenAI REST API
+// openai package provides core transformation logic for partial compatibility with the OpenAI REST API
 package openai
 
 import (
-	"bytes"
 	"encoding/base64"
 	"encoding/json"
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"math/rand"
 	"net/http"
 	"strings"
 	"time"
 
-	"github.com/gin-gonic/gin"
-
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/types/model"
 )
@@ -220,11 +216,12 @@ func NewError(code int, message string) ErrorResponse {
 	return ErrorResponse{Error{Type: etype, Message: message}}
 }
 
-func toUsage(r api.ChatResponse) Usage {
+// ToUsage converts an api.ChatResponse to Usage
+func ToUsage(r api.ChatResponse) Usage {
 	return Usage{
-		PromptTokens:     r.PromptEvalCount,
-		CompletionTokens: r.EvalCount,
-		TotalTokens:      r.PromptEvalCount + r.EvalCount,
+		PromptTokens:     r.Metrics.PromptEvalCount,
+		CompletionTokens: r.Metrics.EvalCount,
+		TotalTokens:      r.Metrics.PromptEvalCount + r.Metrics.EvalCount,
 	}
 }
 
@@ -256,7 +253,8 @@ func toToolCalls(tc []api.ToolCall) []ToolCall {
 	return toolCalls
 }
 
-func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
+// ToChatCompletion converts an api.ChatResponse to ChatCompletion
+func ToChatCompletion(id string, r api.ChatResponse) ChatCompletion {
 	toolCalls := toToolCalls(r.Message.ToolCalls)
 	return ChatCompletion{
 		Id:                id,
@@ -276,12 +274,13 @@ func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
 				}
 				return nil
 			}(r.DoneReason),
-		}}, Usage: toUsage(r),
+		}}, Usage: ToUsage(r),
 		DebugInfo: r.DebugInfo,
 	}
 }
 
-func toChunk(id string, r api.ChatResponse, toolCallSent bool) ChatCompletionChunk {
+// ToChunk converts an api.ChatResponse to ChatCompletionChunk
+func ToChunk(id string, r api.ChatResponse, toolCallSent bool) ChatCompletionChunk {
 	toolCalls := toToolCalls(r.Message.ToolCalls)
 	return ChatCompletionChunk{
 		Id:                id,
@@ -305,15 +304,17 @@ func toChunk(id string, r api.ChatResponse, toolCallSent bool) ChatCompletionChu
 	}
 }
 
-func toUsageGenerate(r api.GenerateResponse) Usage {
+// ToUsageGenerate converts an api.GenerateResponse to Usage
+func ToUsageGenerate(r api.GenerateResponse) Usage {
 	return Usage{
-		PromptTokens:     r.PromptEvalCount,
-		CompletionTokens: r.EvalCount,
-		TotalTokens:      r.PromptEvalCount + r.EvalCount,
+		PromptTokens:     r.Metrics.PromptEvalCount,
+		CompletionTokens: r.Metrics.EvalCount,
+		TotalTokens:      r.Metrics.PromptEvalCount + r.Metrics.EvalCount,
 	}
 }
 
-func toCompletion(id string, r api.GenerateResponse) Completion {
+// ToCompletion converts an api.GenerateResponse to Completion
+func ToCompletion(id string, r api.GenerateResponse) Completion {
 	return Completion{
 		Id:                id,
 		Object:            "text_completion",
@@ -330,11 +331,12 @@ func toCompletion(id string, r api.GenerateResponse) Completion {
 				return nil
 			}(r.DoneReason),
 		}},
-		Usage: toUsageGenerate(r),
+		Usage: ToUsageGenerate(r),
 	}
 }
 
-func toCompleteChunk(id string, r api.GenerateResponse) CompletionChunk {
+// ToCompleteChunk converts an api.GenerateResponse to CompletionChunk
+func ToCompleteChunk(id string, r api.GenerateResponse) CompletionChunk {
 	return CompletionChunk{
 		Id:                id,
 		Object:            "text_completion",
@@ -354,7 +356,8 @@ func toCompleteChunk(id string, r api.GenerateResponse) CompletionChunk {
 	}
 }
 
-func toListCompletion(r api.ListResponse) ListCompletion {
+// ToListCompletion converts an api.ListResponse to ListCompletion
+func ToListCompletion(r api.ListResponse) ListCompletion {
 	var data []Model
 	for _, m := range r.Models {
 		data = append(data, Model{
@@ -371,7 +374,8 @@ func toListCompletion(r api.ListResponse) ListCompletion {
 	}
 }
 
-func toEmbeddingList(model string, r api.EmbedResponse) EmbeddingList {
+// ToEmbeddingList converts an api.EmbedResponse to EmbeddingList
+func ToEmbeddingList(model string, r api.EmbedResponse) EmbeddingList {
 	if r.Embeddings != nil {
 		var data []Embedding
 		for i, e := range r.Embeddings {
@@ -396,7 +400,8 @@ func toEmbeddingList(model string, r api.EmbedResponse) EmbeddingList {
 	return EmbeddingList{}
 }
 
-func toModel(r api.ShowResponse, m string) Model {
+// ToModel converts an api.ShowResponse to Model
+func ToModel(r api.ShowResponse, m string) Model {
 	return Model{
 		Id:      m,
 		Object:  "model",
@@ -405,7 +410,8 @@ func toModel(r api.ShowResponse, m string) Model {
 	}
 }
 
-func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
+// FromChatRequest converts a ChatCompletionRequest to api.ChatRequest
+func FromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 	var messages []api.Message
 	for _, msg := range r.Messages {
 		toolName := ""
@@ -609,7 +615,8 @@ func fromCompletionToolCall(toolCalls []ToolCall) ([]api.ToolCall, error) {
 	return apiToolCalls, nil
 }
 
-func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
+// FromCompleteRequest converts a CompletionRequest to api.GenerateRequest
+func FromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
 	options := make(map[string]any)
 
 	switch stop := r.Stop.(type) {
@@ -660,413 +667,3 @@ func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
 		DebugRenderOnly: r.DebugRenderOnly,
 	}, nil
 }
-
-type BaseWriter struct {
-	gin.ResponseWriter
-}
-
-type ChatWriter struct {
-	stream        bool
-	streamOptions *StreamOptions
-	id            string
-	toolCallSent  bool
-	BaseWriter
-}
-
-type CompleteWriter struct {
-	stream        bool
-	streamOptions *StreamOptions
-	id            string
-	BaseWriter
-}
-
-type ListWriter struct {
-	BaseWriter
-}
-
-type RetrieveWriter struct {
-	BaseWriter
-	model string
-}
-
-type EmbedWriter struct {
-	BaseWriter
-	model string
-}
-
-func (w *BaseWriter) writeError(data []byte) (int, error) {
-	var serr api.StatusError
-	err := json.Unmarshal(data, &serr)
-	if err != nil {
-		return 0, err
-	}
-
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(NewError(http.StatusInternalServerError, serr.Error()))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *ChatWriter) writeResponse(data []byte) (int, error) {
-	var chatResponse api.ChatResponse
-	err := json.Unmarshal(data, &chatResponse)
-	if err != nil {
-		return 0, err
-	}
-
-	// chat chunk
-	if w.stream {
-		c := toChunk(w.id, chatResponse, w.toolCallSent)
-		d, err := json.Marshal(c)
-		if err != nil {
-			return 0, err
-		}
-		if !w.toolCallSent && len(c.Choices) > 0 && len(c.Choices[0].Delta.ToolCalls) > 0 {
-			w.toolCallSent = true
-		}
-
-		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
-		_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
-		if err != nil {
-			return 0, err
-		}
-
-		if chatResponse.Done {
-			if w.streamOptions != nil && w.streamOptions.IncludeUsage {
-				u := toUsage(chatResponse)
-				c.Usage = &u
-				c.Choices = []ChunkChoice{}
-				d, err := json.Marshal(c)
-				if err != nil {
-					return 0, err
-				}
-				_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
-				if err != nil {
-					return 0, err
-				}
-			}
-			_, err = w.ResponseWriter.Write([]byte("data: [DONE]\n\n"))
-			if err != nil {
-				return 0, err
-			}
-		}
-
-		return len(data), nil
-	}
-
-	// chat completion
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(toChatCompletion(w.id, chatResponse))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *ChatWriter) Write(data []byte) (int, error) {
-	code := w.ResponseWriter.Status()
-	if code != http.StatusOK {
-		return w.writeError(data)
-	}
-
-	return w.writeResponse(data)
-}
-
-func (w *CompleteWriter) writeResponse(data []byte) (int, error) {
-	var generateResponse api.GenerateResponse
-	err := json.Unmarshal(data, &generateResponse)
-	if err != nil {
-		return 0, err
-	}
-
-	// completion chunk
-	if w.stream {
-		c := toCompleteChunk(w.id, generateResponse)
-		if w.streamOptions != nil && w.streamOptions.IncludeUsage {
-			c.Usage = &Usage{}
-		}
-		d, err := json.Marshal(c)
-		if err != nil {
-			return 0, err
-		}
-
-		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
-		_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
-		if err != nil {
-			return 0, err
-		}
-
-		if generateResponse.Done {
-			if w.streamOptions != nil && w.streamOptions.IncludeUsage {
-				u := toUsageGenerate(generateResponse)
-				c.Usage = &u
-				c.Choices = []CompleteChunkChoice{}
-				d, err := json.Marshal(c)
-				if err != nil {
-					return 0, err
-				}
-				_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
-				if err != nil {
-					return 0, err
-				}
-			}
-			_, err = w.ResponseWriter.Write([]byte("data: [DONE]\n\n"))
-			if err != nil {
-				return 0, err
-			}
-		}
-
-		return len(data), nil
-	}
-
-	// completion
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(toCompletion(w.id, generateResponse))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *CompleteWriter) Write(data []byte) (int, error) {
-	code := w.ResponseWriter.Status()
-	if code != http.StatusOK {
-		return w.writeError(data)
-	}
-
-	return w.writeResponse(data)
-}
-
-func (w *ListWriter) writeResponse(data []byte) (int, error) {
-	var listResponse api.ListResponse
-	err := json.Unmarshal(data, &listResponse)
-	if err != nil {
-		return 0, err
-	}
-
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(toListCompletion(listResponse))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *ListWriter) Write(data []byte) (int, error) {
-	code := w.ResponseWriter.Status()
-	if code != http.StatusOK {
-		return w.writeError(data)
-	}
-
-	return w.writeResponse(data)
-}
-
-func (w *RetrieveWriter) writeResponse(data []byte) (int, error) {
-	var showResponse api.ShowResponse
-	err := json.Unmarshal(data, &showResponse)
-	if err != nil {
-		return 0, err
-	}
-
-	// retrieve completion
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(toModel(showResponse, w.model))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *RetrieveWriter) Write(data []byte) (int, error) {
-	code := w.ResponseWriter.Status()
-	if code != http.StatusOK {
-		return w.writeError(data)
-	}
-
-	return w.writeResponse(data)
-}
-
-func (w *EmbedWriter) writeResponse(data []byte) (int, error) {
-	var embedResponse api.EmbedResponse
-	err := json.Unmarshal(data, &embedResponse)
-	if err != nil {
-		return 0, err
-	}
-
-	w.ResponseWriter.Header().Set("Content-Type", "application/json")
-	err = json.NewEncoder(w.ResponseWriter).Encode(toEmbeddingList(w.model, embedResponse))
-	if err != nil {
-		return 0, err
-	}
-
-	return len(data), nil
-}
-
-func (w *EmbedWriter) Write(data []byte) (int, error) {
-	code := w.ResponseWriter.Status()
-	if code != http.StatusOK {
-		return w.writeError(data)
-	}
-
-	return w.writeResponse(data)
-}
-
-func ListMiddleware() gin.HandlerFunc {
-	return func(c *gin.Context) {
-		w := &ListWriter{
-			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
-		}
-
-		c.Writer = w
-
-		c.Next()
-	}
-}
-
-func RetrieveMiddleware() gin.HandlerFunc {
-	return func(c *gin.Context) {
-		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(api.ShowRequest{Name: c.Param("model")}); err != nil {
-			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
-			return
-		}
-
-		c.Request.Body = io.NopCloser(&b)
-
-		// response writer
-		w := &RetrieveWriter{
-			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
-			model:      c.Param("model"),
-		}
-
-		c.Writer = w
-
-		c.Next()
-	}
-}
-
-func CompletionsMiddleware() gin.HandlerFunc {
-	return func(c *gin.Context) {
-		var req CompletionRequest
-		err := c.ShouldBindJSON(&req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		var b bytes.Buffer
-		genReq, err := fromCompleteRequest(req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		if err := json.NewEncoder(&b).Encode(genReq); err != nil {
-			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
-			return
-		}
-
-		c.Request.Body = io.NopCloser(&b)
-
-		w := &CompleteWriter{
-			BaseWriter:    BaseWriter{ResponseWriter: c.Writer},
-			stream:        req.Stream,
-			id:            fmt.Sprintf("cmpl-%d", rand.Intn(999)),
-			streamOptions: req.StreamOptions,
-		}
-
-		c.Writer = w
-		c.Next()
-	}
-}
-
-func EmbeddingsMiddleware() gin.HandlerFunc {
-	return func(c *gin.Context) {
-		var req EmbedRequest
-		err := c.ShouldBindJSON(&req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		if req.Input == "" {
-			req.Input = []string{""}
-		}
-
-		if req.Input == nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, "invalid input"))
-			return
-		}
-
-		if v, ok := req.Input.([]any); ok && len(v) == 0 {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, "invalid input"))
-			return
-		}
-
-		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(api.EmbedRequest{Model: req.Model, Input: req.Input, Dimensions: req.Dimensions}); err != nil {
-			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
-			return
-		}
-
-		c.Request.Body = io.NopCloser(&b)
-
-		w := &EmbedWriter{
-			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
-			model:      req.Model,
-		}
-
-		c.Writer = w
-
-		c.Next()
-	}
-}
-
-func ChatMiddleware() gin.HandlerFunc {
-	return func(c *gin.Context) {
-		var req ChatCompletionRequest
-		err := c.ShouldBindJSON(&req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		if len(req.Messages) == 0 {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, "[] is too short - 'messages'"))
-			return
-		}
-
-		var b bytes.Buffer
-
-		chatReq, err := fromChatRequest(req)
-		if err != nil {
-			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
-			return
-		}
-
-		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
-			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
-			return
-		}
-
-		c.Request.Body = io.NopCloser(&b)
-
-		w := &ChatWriter{
-			BaseWriter:    BaseWriter{ResponseWriter: c.Writer},
-			stream:        req.Stream,
-			id:            fmt.Sprintf("chatcmpl-%d", rand.Intn(999)),
-			streamOptions: req.StreamOptions,
-		}
-
-		c.Writer = w
-
-		c.Next()
-	}
-}

scripts/build_windows.ps1

@@ -179,7 +179,7 @@ function buildROCm() {
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
             & cmake --install build --component "HIP" --strip
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-            rm -f $script:DIST_DIR\lib\ollama\rocm\rocblas\library\*gfx906*
+            Remove-Item -Path $script:DIST_DIR\lib\ollama\rocm\rocblas\library\*gfx906* -ErrorAction SilentlyContinue
         }
     }
 }

server/images.go

@@ -105,12 +105,16 @@ func (m *Model) Capabilities() []model.Capability {
 
 	builtinParser := parsers.ParserForName(m.Config.Parser)
 	// Check for tools capability
-	if slices.Contains(m.Template.Vars(), "tools") || (builtinParser != nil && builtinParser.HasToolSupport()) {
+	v, err := m.Template.Vars()
+	if err != nil {
+		slog.Warn("model template contains errors", "error", err)
+	}
+	if slices.Contains(v, "tools") || (builtinParser != nil && builtinParser.HasToolSupport()) {
 		capabilities = append(capabilities, model.CapabilityTools)
 	}
 
 	// Check for insert capability
-	if slices.Contains(m.Template.Vars(), "suffix") {
+	if slices.Contains(v, "suffix") {
 		capabilities = append(capabilities, model.CapabilityInsert)
 	}
 

server/routes.go

@@ -37,8 +37,8 @@ import (
 	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/logutil"
+	"github.com/ollama/ollama/middleware"
 	"github.com/ollama/ollama/model/parsers"
-	"github.com/ollama/ollama/openai"
 	"github.com/ollama/ollama/server/internal/client/ollama"
 	"github.com/ollama/ollama/server/internal/registry"
 	"github.com/ollama/ollama/template"
@@ -63,6 +63,15 @@ func shouldUseHarmony(model *Model) bool {
 	return false
 }
 
+func shouldForceThinking(model *Model) bool {
+	if model.Config.ModelFamily == "qwen3" && strings.Contains(model.Name, "thinking-2507") {
+		// Force thinking mode for the newer qwen3 thinking models otherwise they end up
+		// leaving the thinking trace in the return content w/ unmatched <think></think> tags
+		return true
+	}
+	return false
+}
+
 func experimentEnabled(name string) bool {
 	return slices.Contains(strings.Split(os.Getenv("OLLAMA_EXPERIMENT"), ","), name)
 }
@@ -330,7 +339,10 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 	if req.Suffix != "" {
 		caps = append(caps, model.CapabilityInsert)
 	}
-	if req.Think != nil && req.Think.Bool() {
+	if req.Think == nil && shouldForceThinking(m) {
+		req.Think = &api.ThinkValue{Value: true}
+		caps = append(caps, model.CapabilityThinking)
+	} else if req.Think != nil && req.Think.Bool() {
 		caps = append(caps, model.CapabilityThinking)
 		// TODO(drifkin): consider adding a warning if it's false and the model
 		// doesn't support thinking. It's not strictly required, but it can be a
@@ -1449,11 +1461,11 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.POST("/api/embeddings", s.EmbeddingsHandler)
 
 	// Inference (OpenAI compatibility)
-	r.POST("/v1/chat/completions", openai.ChatMiddleware(), s.ChatHandler)
-	r.POST("/v1/completions", openai.CompletionsMiddleware(), s.GenerateHandler)
-	r.POST("/v1/embeddings", openai.EmbeddingsMiddleware(), s.EmbedHandler)
-	r.GET("/v1/models", openai.ListMiddleware(), s.ListHandler)
-	r.GET("/v1/models/:model", openai.RetrieveMiddleware(), s.ShowHandler)
+	r.POST("/v1/chat/completions", middleware.ChatMiddleware(), s.ChatHandler)
+	r.POST("/v1/completions", middleware.CompletionsMiddleware(), s.GenerateHandler)
+	r.POST("/v1/embeddings", middleware.EmbeddingsMiddleware(), s.EmbedHandler)
+	r.GET("/v1/models", middleware.ListMiddleware(), s.ListHandler)
+	r.GET("/v1/models/:model", middleware.RetrieveMiddleware(), s.ShowHandler)
 
 	if rc != nil {
 		// wrap old with new
@@ -1871,7 +1883,10 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	if len(req.Tools) > 0 {
 		caps = append(caps, model.CapabilityTools)
 	}
-	if req.Think != nil && req.Think.Bool() {
+	if req.Think == nil && shouldForceThinking(m) {
+		req.Think = &api.ThinkValue{Value: true}
+		caps = append(caps, model.CapabilityThinking)
+	} else if req.Think != nil && req.Think.Bool() {
 		caps = append(caps, model.CapabilityThinking)
 	}
 

template/template.go

@@ -148,7 +148,12 @@ func Parse(s string) (*Template, error) {
 	}
 
 	t := Template{Template: tmpl, raw: s}
-	if vars := t.Vars(); !slices.Contains(vars, "messages") && !slices.Contains(vars, "response") {
+	vars, err := t.Vars()
+	if err != nil {
+		return nil, err
+	}
+
+	if !slices.Contains(vars, "messages") && !slices.Contains(vars, "response") {
 		// touch up the template and append {{ .Response }}
 		tmpl.Tree.Root.Nodes = append(tmpl.Tree.Root.Nodes, &response)
 	}
@@ -160,11 +165,15 @@ func (t *Template) String() string {
 	return t.raw
 }
 
-func (t *Template) Vars() []string {
+func (t *Template) Vars() ([]string, error) {
 	var vars []string
 	for _, tt := range t.Templates() {
 		for _, n := range tt.Root.Nodes {
-			vars = append(vars, Identifiers(n)...)
+			v, err := Identifiers(n)
+			if err != nil {
+				return vars, err
+			}
+			vars = append(vars, v...)
 		}
 	}
 
@@ -173,7 +182,7 @@ func (t *Template) Vars() []string {
 		set[strings.ToLower(n)] = struct{}{}
 	}
 
-	return slices.Sorted(maps.Keys(set))
+	return slices.Sorted(maps.Keys(set)), nil
 }
 
 func (t *Template) Contains(s string) bool {
@@ -244,6 +253,10 @@ func (t *Template) Subtree(fn func(parse.Node) bool) *template.Template {
 
 func (t *Template) Execute(w io.Writer, v Values) error {
 	system, messages := collate(v.Messages)
+	vars, err := t.Vars()
+	if err != nil {
+		return err
+	}
 	if v.Prompt != "" && v.Suffix != "" {
 		return t.Template.Execute(w, map[string]any{
 			"Prompt":     v.Prompt,
@@ -253,7 +266,7 @@ func (t *Template) Execute(w io.Writer, v Values) error {
 			"ThinkLevel": v.ThinkLevel,
 			"IsThinkSet": v.IsThinkSet,
 		})
-	} else if !v.forceLegacy && slices.Contains(t.Vars(), "messages") {
+	} else if !v.forceLegacy && slices.Contains(vars, "messages") {
 		return t.Template.Execute(w, map[string]any{
 			"System":     system,
 			"Messages":   messages,
@@ -329,7 +342,7 @@ func (t *Template) Execute(w io.Writer, v Values) error {
 		return err
 	}
 
-	_, err := io.Copy(w, &b)
+	_, err = io.Copy(w, &b)
 	return err
 }
 
@@ -358,27 +371,47 @@ func collate(msgs []api.Message) (string, []*api.Message) {
 }
 
 // Identifiers walks the node tree returning any identifiers it finds along the way
-func Identifiers(n parse.Node) []string {
+func Identifiers(n parse.Node) ([]string, error) {
 	switch n := n.(type) {
 	case *parse.ListNode:
 		var names []string
 		for _, n := range n.Nodes {
-			names = append(names, Identifiers(n)...)
+			i, err := Identifiers(n)
+			if err != nil {
+				return names, err
+			}
+			names = append(names, i...)
 		}
 
-		return names
+		return names, nil
 	case *parse.TemplateNode:
+		if n.Pipe == nil {
+			return nil, errors.New("undefined template specified")
+		}
 		return Identifiers(n.Pipe)
 	case *parse.ActionNode:
+		if n.Pipe == nil {
+			return nil, errors.New("undefined action in template")
+		}
 		return Identifiers(n.Pipe)
 	case *parse.BranchNode:
-		names := Identifiers(n.Pipe)
+		if n.Pipe == nil {
+			return nil, errors.New("undefined branch")
+		}
+		names, err := Identifiers(n.Pipe)
+		if err != nil {
+			return names, err
+		}
 		for _, n := range []*parse.ListNode{n.List, n.ElseList} {
 			if n != nil {
-				names = append(names, Identifiers(n)...)
+				i, err := Identifiers(n)
+				if err != nil {
+					return names, err
+				}
+				names = append(names, i...)
 			}
 		}
-		return names
+		return names, nil
 	case *parse.IfNode:
 		return Identifiers(&n.BranchNode)
 	case *parse.RangeNode:
@@ -389,17 +422,21 @@ func Identifiers(n parse.Node) []string {
 		var names []string
 		for _, c := range n.Cmds {
 			for _, a := range c.Args {
-				names = append(names, Identifiers(a)...)
+				i, err := Identifiers(a)
+				if err != nil {
+					return names, err
+				}
+				names = append(names, i...)
 			}
 		}
-		return names
+		return names, nil
 	case *parse.FieldNode:
-		return n.Ident
+		return n.Ident, nil
 	case *parse.VariableNode:
-		return n.Ident
+		return n.Ident, nil
 	}
 
-	return nil
+	return nil, nil
 }
 
 // deleteNode walks the node list and deletes nodes that match the predicate

template/template_test.go

@@ -154,24 +154,55 @@ func TestTemplate(t *testing.T) {
 }
 
 func TestParse(t *testing.T) {
-	cases := []struct {
+	validCases := []struct {
+		name     string
 		template string
 		vars     []string
 	}{
-		{"{{ .Prompt }}", []string{"prompt", "response"}},
-		{"{{ .System }} {{ .Prompt }}", []string{"prompt", "response", "system"}},
-		{"{{ .System }} {{ .Prompt }} {{ .Response }}", []string{"prompt", "response", "system"}},
-		{"{{ with .Tools }}{{ . }}{{ end }} {{ .System }} {{ .Prompt }}", []string{"prompt", "response", "system", "tools"}},
-		{"{{ range .Messages }}{{ .Role }} {{ .Content }}{{ end }}", []string{"content", "messages", "role"}},
-		{"{{ range .Messages }}{{ if eq .Role \"tool\" }}Tool Result: {{ .ToolName }} {{ .Content }}{{ end }}{{ end }}", []string{"content", "messages", "role", "toolname"}},
-		{`{{- range .Messages }}
+		{
+			name:     "PromptOnly",
+			template: "{{ .Prompt }}",
+			vars:     []string{"prompt", "response"},
+		},
+		{
+			name:     "SystemAndPrompt",
+			template: "{{ .System }} {{ .Prompt }}",
+			vars:     []string{"prompt", "response", "system"},
+		},
+		{
+			name:     "PromptResponseSystem",
+			template: "{{ .System }} {{ .Prompt }} {{ .Response }}",
+			vars:     []string{"prompt", "response", "system"},
+		},
+		{
+			name:     "ToolsBlock",
+			template: "{{ with .Tools }}{{ . }}{{ end }} {{ .System }} {{ .Prompt }}",
+			vars:     []string{"prompt", "response", "system", "tools"},
+		},
+		{
+			name:     "MessagesRange",
+			template: "{{ range .Messages }}{{ .Role }} {{ .Content }}{{ end }}",
+			vars:     []string{"content", "messages", "role"},
+		},
+		{
+			name:     "ToolResultConditional",
+			template: "{{ range .Messages }}{{ if eq .Role \"tool\" }}Tool Result: {{ .ToolName }} {{ .Content }}{{ end }}{{ end }}",
+			vars:     []string{"content", "messages", "role", "toolname"},
+		},
+		{
+			name: "MultilineSystemUserAssistant",
+			template: `{{- range .Messages }}
 {{- if eq .Role "system" }}SYSTEM:
 {{- else if eq .Role "user" }}USER:
 {{- else if eq .Role "assistant" }}ASSISTANT:
-{{- else if eq .Role "tool" }}TOOL: 
+{{- else if eq .Role "tool" }}TOOL:
 {{- end }} {{ .Content }}
-{{- end }}`, []string{"content", "messages", "role"}},
-		{`{{- if .Messages }}
+{{- end }}`,
+			vars: []string{"content", "messages", "role"},
+		},
+		{
+			name: "ChatMLLike",
+			template: `{{- if .Messages }}
 {{- range .Messages }}<|im_start|>{{ .Role }}
 {{ .Content }}<|im_end|>
 {{ end }}<|im_start|>assistant
@@ -182,18 +213,60 @@ func TestParse(t *testing.T) {
 {{ .Prompt }}<|im_end|>
 {{ end }}<|im_start|>assistant
 {{ .Response }}<|im_end|>
-{{- end -}}`, []string{"content", "messages", "prompt", "response", "role", "system"}},
+{{- end -}}`,
+			vars: []string{"content", "messages", "prompt", "response", "role", "system"},
+		},
 	}
 
-	for _, tt := range cases {
-		t.Run("", func(t *testing.T) {
+	for _, tt := range validCases {
+		tt := tt
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
 			tmpl, err := Parse(tt.template)
 			if err != nil {
-				t.Fatal(err)
+				t.Fatalf("Parse returned unexpected error: %v", err)
 			}
 
-			if diff := cmp.Diff(tmpl.Vars(), tt.vars); diff != "" {
-				t.Errorf("mismatch (-got +want):\n%s", diff)
+			gotVars, err := tmpl.Vars()
+			if err != nil {
+				t.Fatalf("Vars returned unexpected error: %v", err)
+			}
+
+			if diff := cmp.Diff(gotVars, tt.vars); diff != "" {
+				t.Errorf("Vars mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestParseError(t *testing.T) {
+	invalidCases := []struct {
+		name     string
+		template string
+		errorStr string
+	}{
+		{
+			"TemplateNotClosed",
+			"{{ .Prompt ",
+			"unclosed action",
+		},
+		{
+			"Template",
+			`{{define "x"}}{{template "x"}}{{end}}{{template "x"}}`,
+			"undefined template specified",
+		},
+	}
+
+	for _, tt := range invalidCases {
+		t.Run(tt.name, func(t *testing.T) {
+			_, err := Parse(tt.template)
+			if err == nil {
+				t.Fatalf("expected Parse to return an error for an invalid template, got nil")
+			}
+
+			if !strings.Contains(strings.ToLower(err.Error()), strings.ToLower(tt.errorStr)) {
+				t.Errorf("unexpected error message.\n got: %q\n want substring (case‑insensitive): %q", err.Error(), tt.errorStr)
 			}
 		})
 	}