Revert "runner: move harmony to runner (#12052)"

This reverts commit 1a558f98e2.

.github/workflows/release.yaml

@@ -65,14 +65,36 @@ jobs:
             arch: amd64
             preset: 'CUDA 12'
             install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
+            cuda-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
             cuda-version: '12.8'
             flags: ''
+            runner_dir: 'cuda_v12'
+          - os: windows
+            arch: amd64
+            preset: 'CUDA 13'
+            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
+            cuda-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
+              - '"crt"'
+              - '"nvvm"'
+              - '"nvptxcompiler"'
+            cuda-version: '13.0'
+            flags: ''
+            runner_dir: 'cuda_v13'
           - os: windows
             arch: amd64
             preset: 'ROCm 6'
             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: ''
     runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
     environment: release
     env:
@@ -96,7 +118,7 @@ jobs:
           $ErrorActionPreference = "Stop"
           if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
             Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            $subpackages = @("cudart", "nvcc", "cublas", "cublas_dev") | Foreach-Object {"${_}_${{ matrix.cuda-version }}"}
+            $subpackages = @(${{ join(matrix.cuda-components, ', ') }}) | Foreach-Object {"${_}_${{ matrix.cuda-version }}"}
             Start-Process -FilePath .\install.exe -ArgumentList (@("-s") + $subpackages) -NoNewWindow -Wait
           }
 
@@ -138,7 +160,7 @@ jobs:
         run: |
           Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
           Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
+          cmake --preset "${{ 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
         env:
@@ -232,7 +254,7 @@ jobs:
             case "$COMPONENT" in
               bin/ollama)                echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/cuda_sbsa)      echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/cuda_jetpack5)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
               lib/ollama/cuda_jetpack6)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
               lib/ollama/rocm)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;

.github/workflows/test.yaml

@@ -46,7 +46,7 @@ jobs:
         include:
           - preset: CPU
           - preset: CUDA
-            container: nvidia/cuda:12.8.1-devel-ubuntu22.04
+            container: nvidia/cuda:13.0.0-devel-ubuntu22.04
             flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
           - preset: ROCm
             container: rocm/dev-ubuntu-22.04:6.1.2
@@ -78,8 +78,17 @@ jobs:
         include:
           - preset: CPU
           - preset: CUDA
-            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
+            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
             flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
+            cuda-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
+              - '"crt"'
+              - '"nvvm"'
+              - '"nvptxcompiler"'
+            cuda-version: '13.0'
           - preset: ROCm
             install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
             flags: '-DAMDGPU_TARGETS=gfx1010 -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"'
@@ -102,7 +111,8 @@ jobs:
           $ErrorActionPreference = "Stop"
           if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
             Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_12.8", "nvcc_12.8", "cublas_12.8", "cublas_dev_12.8")) -NoNewWindow -Wait
+            $subpackages = @(${{ join(matrix.cuda-components, ', ') }}) | Foreach-Object {"${_}_${{ matrix.cuda-version }}"}
+            Start-Process -FilePath .\install.exe -ArgumentList (@("-s") + $subpackages) -NoNewWindow -Wait
           }
 
           $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path

CMakeLists.txt

@@ -38,7 +38,7 @@ if (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
 endif()
 
 set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
-set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama)
+set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama/${OLLAMA_RUNNER_DIR})
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
@@ -81,7 +81,7 @@ if(CMAKE_CUDA_COMPILER)
     add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cuda)
     install(TARGETS ggml-cuda
         RUNTIME_DEPENDENCIES
-            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_LIBRARY_DIR}
+            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
             PRE_INCLUDE_REGEXES cublas cublasLt cudart
             PRE_EXCLUDE_REGEXES ".*"
         RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA

CMakePresets.json

@@ -18,6 +18,14 @@
       "name": "CUDA",
       "inherits": [ "Default" ]
     },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "cacheVariables": {
+        "CMAKE_CUDA_ARCHITECTURES": "50-virtual;60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual",
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
+      }
+    },
     {
       "name": "CUDA 12",
       "inherits": [ "CUDA" ],
@@ -26,6 +34,14 @@
         "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
       }
     },
+    {
+      "name": "CUDA 13",
+      "inherits": [ "CUDA" ],
+      "cacheVariables": {
+        "CMAKE_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;110-virtual;120-virtual;121-virtual",
+        "CMAKE_CUDA_FLAGS": "-t 2"
+      }
+    },
     {
       "name": "JetPack 5",
       "inherits": [ "CUDA" ],
@@ -72,11 +88,21 @@
       "configurePreset": "CUDA",
       "targets": [ "ggml-cuda" ]
     },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "configurePreset": "CUDA 11"
+    },
     {
       "name": "CUDA 12",
       "inherits": [ "CUDA" ],
       "configurePreset": "CUDA 12"
     },
+    {
+      "name": "CUDA 13",
+      "inherits": [ "CUDA" ],
+      "configurePreset": "CUDA 13"
+    },
     {
       "name": "JetPack 5",
       "inherits": [ "CUDA" ],

Dockerfile

@@ -39,15 +39,35 @@ RUN --mount=type=cache,target=/root/.ccache \
         && cmake --build --parallel --preset 'CPU' \
         && cmake --install build --component CPU --strip --parallel 8
 
+FROM base AS cuda-11
+ARG CUDA11VERSION=11.8
+RUN dnf install -y cuda-toolkit-${CUDA11VERSION//./-}
+ENV PATH=/usr/local/cuda-11/bin:$PATH
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'CUDA 11' -DOLLAMA_RUNNER_DIR="cuda_v11" \
+        && cmake --build --parallel --preset 'CUDA 11' \
+        && cmake --install build --component CUDA --strip --parallel 8
+
 FROM base AS cuda-12
 ARG CUDA12VERSION=12.8
 RUN dnf install -y cuda-toolkit-${CUDA12VERSION//./-}
 ENV PATH=/usr/local/cuda-12/bin:$PATH
 RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 12' \
+    cmake --preset 'CUDA 12' -DOLLAMA_RUNNER_DIR="cuda_v12"\
         && cmake --build --parallel --preset 'CUDA 12' \
         && cmake --install build --component CUDA --strip --parallel 8
 
+
+FROM base AS cuda-13
+ARG CUDA13VERSION=13.0
+RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-}
+ENV PATH=/usr/local/cuda-13/bin:$PATH
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'CUDA 13' -DOLLAMA_RUNNER_DIR="cuda_v13" \
+        && cmake --build --parallel --preset 'CUDA 13' \
+        && cmake --install build --component CUDA --strip --parallel 8
+
+
 FROM base AS rocm-6
 ENV PATH=/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:/opt/rocm/hcc/bin:$PATH
 RUN --mount=type=cache,target=/root/.ccache \
@@ -92,10 +112,14 @@ RUN --mount=type=cache,target=/root/.cache/go-build \
     go build -trimpath -buildmode=pie -o /bin/ollama .
 
 FROM --platform=linux/amd64 scratch AS amd64
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama
+# COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
+COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
+COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/
 
 FROM --platform=linux/arm64 scratch AS arm64
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama/cuda_sbsa
+# COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
+COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
+COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/
 COPY --from=jetpack-5 dist/lib/ollama /lib/ollama/cuda_jetpack5
 COPY --from=jetpack-6 dist/lib/ollama /lib/ollama/cuda_jetpack6
 

README.md

@@ -414,6 +414,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Serene Pub](https://github.com/doolijb/serene-pub) (Beginner friendly, open source AI Roleplaying App for Windows, Mac OS and Linux. Search, download and use models with Ollama all inside the app.)
 - [Andes](https://github.com/aqerd/andes) (A Visual Studio Code extension that provides a local UI interface for Ollama models)
 - [Clueless](https://github.com/KashyapTan/clueless) (Open Source & Local Cluely: A desktop application LLM assistant to help you talk to anything on your screen using locally served Ollama models. Also undetectable to screenshare)
+- [ollama-co2](https://github.com/carbonatedWaterOrg/ollama-co2) (FastAPI web interface for monitoring and managing local and remote Ollama servers with real-time model monitoring and concurrent downloads)
 
 ### Cloud
 

api/types.go

@@ -388,8 +388,12 @@ type EmbedRequest struct {
 	// this request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
+	// Truncate truncates the input to fit the model's max sequence length.
 	Truncate *bool `json:"truncate,omitempty"`
 
+	// Dimensions truncates the output embedding to the specified dimension.
+	Dimensions int `json:"dimensions,omitempty"`
+
 	// Options lists model-specific options.
 	Options map[string]any `json:"options"`
 }

cmd/cmd.go

@@ -56,10 +56,8 @@ func ensureThinkingSupport(ctx context.Context, client *api.Client, name string)
 	if err != nil {
 		return
 	}
-	for _, cap := range resp.Capabilities {
-		if cap == model.CapabilityThinking {
-			return
-		}
+	if slices.Contains(resp.Capabilities, model.CapabilityThinking) {
+		return
 	}
 	fmt.Fprintf(os.Stderr, "warning: model %q does not support thinking output\n", name)
 }

discover/cuda_common.go

@@ -43,14 +43,15 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 				}
 			}
 		}
-		return "sbsa"
 	}
 
-	// driver 12.0 has problems with the cuda v12 library, so run v11 on those older drivers
-	if gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
-		// The detected driver is older than Feb 2023
-		slog.Warn("old CUDA driver detected - please upgrade to a newer driver", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
-		return "v11"
+	if gpuInfo.DriverMajor < 13 {
+		// The detected driver is older than 580 (Aug 2025)
+		// Warn if their CC is compatible with v13 and they should upgrade their driver to get better performance
+		if gpuInfo.computeMajor > 7 || (gpuInfo.computeMajor == 7 && gpuInfo.computeMinor >= 5) {
+			slog.Warn("old CUDA driver detected - please upgrade to a newer driver for best performance", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
+		}
+		return "v12"
 	}
-	return "v12"
+	return "v13"
 }

docs/api.md

@@ -1708,6 +1708,7 @@ Advanced parameters:
 - `truncate`: truncates the end of each input to fit within context length. Returns error if `false` and context length is exceeded. Defaults to `true`
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
+- `dimensions`: number of dimensions for the embedding
 
 ### Examples
 

docs/linux.md

@@ -11,12 +11,13 @@ curl -fsSL https://ollama.com/install.sh | sh
 ## Manual install
 
 > [!NOTE]
-> If you are upgrading from a prior version, you should remove the old libraries with `sudo rm -rf /usr/lib/ollama` first.
+> If you are upgrading from a prior version, you **MUST** remove the old libraries with `sudo rm -rf /usr/lib/ollama` first.
 
 Download and extract the package:
 
 ```shell
 curl -LO https://ollama.com/download/ollama-linux-amd64.tgz
+sudo rm -rf /usr/lib/ollama
 sudo tar -C /usr -xzf ollama-linux-amd64.tgz
 ```
 

envconfig/config.go

@@ -185,8 +185,6 @@ var (
 	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
 	// Auth enables authentication between the Ollama client and server
 	UseAuth = Bool("OLLAMA_AUTH")
-	// Enable the new memory estimation logic
-	NewMemoryEstimates = Bool("OLLAMA_NEW_ESTIMATES")
 )
 
 func String(s string) func() string {
@@ -272,7 +270,6 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
 		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4096)"},
 		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
-		"OLLAMA_NEW_ESTIMATES":     {"OLLAMA_NEW_ESTIMATES", NewMemoryEstimates(), "Enable the new memory estimation logic"},
 
 		// Informational
 		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},

fs/ggml/ggml.go

@@ -864,12 +864,16 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
 
 // SupportsKVCacheType checks if the requested cache type is supported
 func (f GGML) SupportsKVCacheType(cacheType string) bool {
+	if cacheType == "" || cacheType == "f16" {
+		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 ", "mode", arch)
-		return cacheType == "f16"
+		slog.Warn("model only supports non-quantized cache types", "model", arch)
+		return false
 	}
-	return slices.Contains([]string{"f16", "q8_0", "q4_0"}, cacheType)
+	return slices.Contains([]string{"q8_0", "q4_0"}, cacheType)
 }
 
 // SupportsFlashAttention checks if the model supports flash attention
@@ -879,6 +883,10 @@ func (f GGML) SupportsFlashAttention() bool {
 		return false
 	}
 
+	if arch := f.KV().Architecture(); slices.Contains([]string{"gemma2"}, arch) {
+		return false
+	}
+
 	// Check head counts match and are non-zero
 	headCountK := f.KV().EmbeddingHeadCountK()
 	headCountV := f.KV().EmbeddingHeadCountV()

harmony/harmonyparser.go

@@ -3,29 +3,15 @@ package harmony
 import (
 	"fmt"
 	"log/slog"
-	"slices"
 	"strings"
 	"unicode"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/logutil"
-	"github.com/ollama/ollama/template"
 )
 
 type harmonyParserState int
 
-func ShouldUseHarmony(modelFamily string, template *template.Template) bool {
-	if slices.Contains([]string{"gptoss", "gpt-oss"}, modelFamily) {
-		// heuristic to check whether the template expects to be parsed via harmony:
-		// search for harmony tags that are nearly always used
-		if template.Contains("<|start|>") && template.Contains("<|end|>") {
-			return true
-		}
-	}
-
-	return false
-}
-
 const (
 	harmonyParserState_LookingForMessageStart harmonyParserState = iota
 	harmonyParserState_ParsingHeader
@@ -89,28 +75,18 @@ func (s *HarmonyParser) AddImplicitStart() {
 	s.acc.WriteString("<|start|>assistant")
 }
 
-func Prefill(lastMessage api.Message) string {
-	if lastMessage.Role != "assistant" {
-		return ""
-	}
-
-	switch {
-	case strings.TrimSpace(lastMessage.Content) != "":
-		return "<|start|>assistant<|channel|>final<|message|>"
-	case strings.TrimSpace(lastMessage.Thinking) != "":
-		return "<|start|>assistant<|channel|>analysis<|message|>"
-	default:
-		return ""
-	}
-}
-
-// AddImplicitStartOrPrefill adds an implicit start tag or prefill string if provided
-func (s *HarmonyParser) AddImplicitStartOrPrefill(prefillString string) {
-	if strings.TrimSpace(prefillString) != "" {
-		s.acc.WriteString(prefillString)
-	} else {
-		s.AddImplicitStart()
+func (s *HarmonyParser) AddImplicitStartOrPrefill(lastMessage *api.Message) {
+	if lastMessage != nil && lastMessage.Role == "assistant" {
+		// handle prefilling conditions
+		if lastMessage.Content != "" {
+			s.acc.WriteString("<|start|>assistant<|channel|>final<|message|>")
+			return
+		} else if lastMessage.Thinking != "" {
+			s.acc.WriteString("<|start|>assistant<|channel|>analysis<|message|>")
+			return
+		}
 	}
+	s.AddImplicitStart()
 }
 
 func (s *HarmonyParser) AddContent(content string) []HarmonyEvent {

harmony/harmonyparser_test.go

@@ -3,7 +3,6 @@ package harmony
 import (
 	"fmt"
 	"reflect"
-	"strings"
 	"testing"
 )
 
@@ -536,202 +535,3 @@ func TestFunctionConvertAndAdd(t *testing.T) {
 		})
 	}
 }
-
-func TestHarmonyMessageHandlerStreamingScenarios(t *testing.T) {
-	t.Run("thinking_then_content_streams", func(t *testing.T) {
-		handler := NewHarmonyMessageHandler()
-		handler.HarmonyParser.AddImplicitStart()
-		tp := handler.CreateToolParser()
-		type step struct {
-			in           string
-			wantContent  string
-			wantThinking string
-		}
-		steps := []step{
-			{in: "<|channel|>analysis<|message|>Thinking...", wantThinking: "Thinking..."},
-			{in: "<|end|>", wantThinking: ""},
-			{in: "<|start|>assistant<|message|>Answer", wantContent: "Answer"},
-			{in: "<|end|>", wantContent: ""},
-		}
-		for i, s := range steps {
-			content, thinking, tool := handler.AddContent(s.in, tp)
-			if tool != "" {
-				tp.Add(tool)
-			}
-			if content != s.wantContent || thinking != s.wantThinking {
-				t.Fatalf("step %d: got (content=%q thinking=%q), want (content=%q thinking=%q)", i, content, thinking, s.wantContent, s.wantThinking)
-			}
-		}
-	})
-
-	t.Run("content_streams_as_it_arrives", func(t *testing.T) {
-		handler := NewHarmonyMessageHandler()
-		handler.HarmonyParser.AddImplicitStart()
-		tp := handler.CreateToolParser()
-		inputs := []string{
-			"<|start|>assistant<|message|>Hello",
-			", world",
-			"!<|end|>",
-		}
-		var got []string
-		for _, in := range inputs {
-			content, thinking, tool := handler.AddContent(in, tp)
-			if tool != "" {
-				tp.Add(tool)
-			}
-			if thinking != "" {
-				t.Fatalf("unexpected thinking %q", thinking)
-			}
-			if content != "" {
-				got = append(got, content)
-			}
-		}
-		want := []string{"Hello", ", world", "!"}
-		if !reflect.DeepEqual(got, want) {
-			t.Fatalf("content pieces mismatch: got %v want %v", got, want)
-		}
-	})
-
-	t.Run("thinking_streams_separately_from_content", func(t *testing.T) {
-		handler := NewHarmonyMessageHandler()
-		handler.HarmonyParser.AddImplicitStart()
-		tp := handler.CreateToolParser()
-		inputs := []string{
-			"<|channel|>analysis<|message|>Thinking...",
-			"<|end|>",
-			"<|start|>assistant<|message|>Answer",
-			"<|end|>",
-		}
-		var got []string
-		for _, in := range inputs {
-			content, thinking, tool := handler.AddContent(in, tp)
-			if tool != "" {
-				tp.Add(tool)
-			}
-			if thinking != "" {
-				got = append(got, thinking)
-			}
-			if content != "" {
-				got = append(got, content)
-			}
-		}
-		want := []string{"Thinking...", "Answer"}
-		if !reflect.DeepEqual(got, want) {
-			t.Fatalf("content pieces mismatch: got %v want %v", got, want)
-		}
-	})
-
-	t.Run("partial_tags_buffer_until_complete", func(t *testing.T) {
-		handler := NewHarmonyMessageHandler()
-		handler.HarmonyParser.AddImplicitStart()
-		tp := handler.CreateToolParser()
-		inputs := []string{
-			"<|chan",
-			"nel|>analysis<|mess",
-			"age|>Deep ",
-			"thought",
-			"<|end|>",
-			"<|start|>assistant<|message|>Done",
-			"<|end|>",
-		}
-		var thinkingPieces []string
-		var contentPieces []string
-		for _, in := range inputs {
-			content, thinking, tool := handler.AddContent(in, tp)
-			if tool != "" {
-				tp.Add(tool)
-			}
-			if thinking != "" {
-				thinkingPieces = append(thinkingPieces, thinking)
-			}
-			if content != "" {
-				contentPieces = append(contentPieces, content)
-			}
-		}
-		if want := []string{"Deep ", "thought"}; !reflect.DeepEqual(thinkingPieces, want) {
-			t.Fatalf("thinking pieces mismatch: got %v want %v", thinkingPieces, want)
-		}
-		if want := []string{"Done"}; !reflect.DeepEqual(contentPieces, want) {
-			t.Fatalf("content pieces mismatch: got %v want %v", contentPieces, want)
-		}
-	})
-
-	t.Run("simple_assistant_after_analysis", func(t *testing.T) {
-		handler := NewHarmonyMessageHandler()
-		handler.HarmonyParser.AddImplicitStart()
-		tp := handler.CreateToolParser()
-		inputs := []string{
-			"<|channel|>analysis<|message|>Think",
-			"<|end|>",
-			"<|start|>assistant<|message|>Answer",
-			"<|end|>",
-		}
-		var contentSb, thinkingSb strings.Builder
-		for _, in := range inputs {
-			content, thinking, tool := handler.AddContent(in, tp)
-			if tool != "" {
-				tp.Add(tool)
-			}
-			contentSb.WriteString(content)
-			thinkingSb.WriteString(thinking)
-		}
-		if contentSb.String() != "Answer" {
-			t.Fatalf("content mismatch: got %q want %q", contentSb.String(), "Answer")
-		}
-		if thinkingSb.String() != "Think" {
-			t.Fatalf("thinking mismatch: got %q want %q", thinkingSb.String(), "Think")
-		}
-	})
-
-	t.Run("tool_call_parsed_and_returned_correctly", func(t *testing.T) {
-		handler := NewHarmonyMessageHandler()
-		handler.HarmonyParser.AddImplicitStart()
-		tp := handler.CreateToolParser()
-		inputs := []string{
-			"<|channel|>commentary to=functions.calculate<|message|>{\"expression\":\"2+2\"}<|end|>",
-		}
-		for _, in := range inputs {
-			content, thinking, tool := handler.AddContent(in, tp)
-			if content != "" || thinking != "" {
-				continue
-			}
-			if tool != "" {
-				tp.Add(tool)
-			}
-		}
-		name, args := tp.Drain()
-		if name == nil || *name != "functions.calculate" {
-			t.Fatalf("unexpected tool name: %v", name)
-		}
-		if got, want := args, "{\"expression\":\"2+2\"}"; got != want {
-			t.Fatalf("unexpected tool args: got %s want %s", got, want)
-		}
-	})
-
-	t.Run("tool_call_across_chunks", func(t *testing.T) {
-		handler := NewHarmonyMessageHandler()
-		handler.HarmonyParser.AddImplicitStart()
-		tp := handler.CreateToolParser()
-		inputs := []string{
-			"<|channel|>commentary to=functions.calculate<|message|>{\"expression\":\"2+",
-			"2\"}",
-			"<|end|>",
-		}
-		for _, in := range inputs {
-			content, thinking, tool := handler.AddContent(in, tp)
-			if content != "" || thinking != "" {
-				continue
-			}
-			if tool != "" {
-				tp.Add(tool)
-			}
-		}
-		name, args := tp.Drain()
-		if name == nil || *name != "functions.calculate" {
-			t.Fatalf("unexpected tool name: %v", name)
-		}
-		if got, want := args, "{\"expression\":\"2+2\"}"; got != want {
-			t.Fatalf("unexpected tool args: got %s want %s", got, want)
-		}
-	})
-}

integration/context_test.go

@@ -50,7 +50,7 @@ func TestContextExhaustion(t *testing.T) {
 	// Set up the test data
 	req := api.GenerateRequest{
 		Model:  smol,
-		Prompt: "Write me a story with a ton of emojis?",
+		Prompt: "Write me a story in english with a lot of emojis",
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,

integration/utils_test.go

@@ -561,7 +561,7 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 			}, {
 				Model:     smol,
-				Prompt:    "what is the origin of the US thanksgiving holiday? Be brief but factual in your reply",
+				Prompt:    "how do rainbows form? Be brief but factual in your reply",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 			}, {
@@ -579,9 +579,9 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 		[][]string{
 			{"sunlight", "scattering", "interact", "color", "surface", "depth", "red", "orange", "yellow", "absorbs", "wavelength"},
 			{"soil", "organic", "earth", "black", "tan", "chemical", "processes", "pigments", "particles", "iron oxide", "rust", "air", "water", "mixture", "mixing"},
-			{"england", "english", "massachusetts", "pilgrims", "colonists", "independence", "british", "feast", "family", "gatherings", "traditions", "turkey", "colonial", "period", "harvest", "agricultural", "european settlers", "american revolution", "civil war", "16th century", "17th century", "native american", "united states", "cultural", "hardship", "autumn", "festival"},
+			{"water", "droplet", "refracted", "reflect", "color", "spectrum"},
 			{"fourth", "july", "declaration", "independence"},
-			{"nitrogen", "oxygen", "carbon", "dioxide"},
+			{"nitrogen", "oxygen", "carbon", "dioxide", "water", "vapor"},
 		}
 }
 

llama/llama.go

@@ -515,33 +515,34 @@ func (c *MtmdContext) NewEmbed(llamaContext *Context, data []byte) ([][]float32,
 	}
 	nChunks := C.mtmd_input_chunks_size(ic)
 	numEmbed := llamaContext.Model().NEmbd()
-	lastChunkSize := 0
+	embed := make([][]float32, 0)
 	for i := range int(nChunks) {
 		chunk := C.mtmd_input_chunks_get(ic, C.size_t(i))
 		numTokens := int(C.mtmd_input_chunk_get_n_tokens(chunk))
-		lastChunkSize = numTokens
+		slog.Debug("chunk tokens", "index", i, "numTokens", numTokens)
 
 		// Encode the chunk
 		if C.int32_t(0) != C.mtmd_encode_chunk(c.c, chunk) {
 			return nil, errors.New("unable to encode mtmd image chunk")
 		}
-	}
 
-	// Get the embeddings
-	embed := make([][]float32, lastChunkSize)
-	embd := C.mtmd_get_output_embd(c.c)
-	if nil == embd {
-		return nil, errors.New("failed to get image embedding")
-	}
+		// Get the embeddings for this chunk
+		chunkEmbed := make([][]float32, numTokens)
+		chunkEmbd := C.mtmd_get_output_embd(c.c)
+		if nil == chunkEmbd {
+			continue
+		}
 
-	// Extend the embedding array for each token
-	s := unsafe.Slice((*float32)(embd), numEmbed*lastChunkSize)
-	rows := make([]float32, len(s))
-	copy(rows, s)
-	for i := range lastChunkSize {
-		embed[i] = rows[i*numEmbed : (i+1)*numEmbed]
+		// Extend the embedding array for each token
+		s := unsafe.Slice((*float32)(chunkEmbd), numTokens*numEmbed)
+		rows := make([]float32, len(s))
+		copy(rows, s)
+		for i := range numTokens {
+			chunkEmbed[i] = rows[i*numEmbed : (i+1)*numEmbed]
+		}
+		embed = append(embed, chunkEmbed...)
 	}
-
+	slog.Debug("image embeddings", "totalEmbeddings", len(embed))
 	return embed, nil
 }
 

llm/memory.go

@@ -202,7 +202,7 @@ func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 	var kvct string
 	if useFlashAttention {
 		requested := strings.ToLower(envconfig.KvCacheType())
-		if requested != "" && f.SupportsKVCacheType(requested) {
+		if f.SupportsKVCacheType(requested) {
 			kvct = requested
 		}
 	}

llm/server.go

@@ -148,7 +148,11 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 	var textProcessor model.TextProcessor
 	var err error
 	if envconfig.NewEngine() || f.KV().OllamaEngineRequired() {
-		textProcessor, err = model.NewTextProcessor(modelPath)
+		if len(projectors) == 0 {
+			textProcessor, err = model.NewTextProcessor(modelPath)
+		} else {
+			err = errors.New("split vision models aren't supported")
+		}
 		if err != nil {
 			// To prepare for opt-out mode, instead of treating this as an error, we fallback to the old runner
 			slog.Debug("model not yet supported by Ollama engine, switching to compatibility mode", "model", modelPath, "error", err)
@@ -161,11 +165,6 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		}
 	}
 
-	newEstimates := textProcessor != nil && envconfig.NewMemoryEstimates()
-	if newEstimates {
-		slog.Info("enabling new memory estimates")
-	}
-
 	// Verify the requested context size is <= the model training size
 	trainCtx := f.KV().ContextLength()
 	if opts.NumCtx > int(trainCtx) && trainCtx > 0 {
@@ -220,7 +219,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 
 		// Flash Attention also supports kv cache quantization
 		// Enable if the requested and kv cache type is supported by the model
-		if kvct != "" && f.SupportsKVCacheType(kvct) {
+		if f.SupportsKVCacheType(kvct) {
 			loadRequest.KvCacheType = kvct
 		} else {
 			slog.Warn("kv cache type not supported by model", "type", kvct)
@@ -433,7 +432,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 			}
 		}()
 
-		if newEstimates {
+		if textProcessor != nil {
 			return &ollamaServer{llmServer: s}, nil
 		} else {
 			return &llamaServer{llmServer: s, ggml: f}, nil
@@ -1349,9 +1348,7 @@ type CompletionRequest struct {
 	Images  []ImageData
 	Options *api.Options
 
-	Grammar       string // set before sending the request to the subprocess
-	UseHarmony    bool
-	PrefillString string
+	Grammar string // set before sending the request to the subprocess
 }
 
 // DoneReason represents the reason why a completion response is done
@@ -1364,8 +1361,6 @@ const (
 	DoneReasonLength
 	// DoneReasonConnectionClosed indicates the completion stopped due to the connection being closed
 	DoneReasonConnectionClosed
-	// DoneReasonTokenRepeatLimit indicates the completion stopped due to a token repeat limit
-	DoneReasonTokenRepeatLimit
 )
 
 func (d DoneReason) String() string {
@@ -1374,23 +1369,19 @@ func (d DoneReason) String() string {
 		return "length"
 	case DoneReasonStop:
 		return "stop"
-	case DoneReasonTokenRepeatLimit:
-		return "token_repeat_limit"
 	default:
 		return "" // closed
 	}
 }
 
 type CompletionResponse struct {
-	Content            string         `json:"content"`
-	Thinking           string         `json:"thinking"`
-	ToolCalls          []api.ToolCall `json:"tool_calls"`
-	DoneReason         DoneReason     `json:"done_reason"`
-	Done               bool           `json:"done"`
-	PromptEvalCount    int            `json:"prompt_eval_count"`
-	PromptEvalDuration time.Duration  `json:"prompt_eval_duration"`
-	EvalCount          int            `json:"eval_count"`
-	EvalDuration       time.Duration  `json:"eval_duration"`
+	Content            string        `json:"content"`
+	DoneReason         DoneReason    `json:"done_reason"`
+	Done               bool          `json:"done"`
+	PromptEvalCount    int           `json:"prompt_eval_count"`
+	PromptEvalDuration time.Duration `json:"prompt_eval_duration"`
+	EvalCount          int           `json:"eval_count"`
+	EvalDuration       time.Duration `json:"eval_duration"`
 }
 
 func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error {
@@ -1508,8 +1499,7 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 				return fmt.Errorf("error unmarshalling llm prediction response: %v", err)
 			}
 			switch {
-			// TODO(parthsareen): token repeat limit is now handled in the runner, this currently support legacy model and can be removed in the future
-			case strings.TrimSpace(c.Content) == lastToken && c.Content != "":
+			case strings.TrimSpace(c.Content) == lastToken:
 				tokenRepeat++
 			default:
 				lastToken = strings.TrimSpace(c.Content)
@@ -1522,14 +1512,16 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 				return ctx.Err()
 			}
 
+			if c.Content != "" {
+				fn(CompletionResponse{
+					Content: c.Content,
+				})
+			}
+
 			if c.Done {
 				fn(c)
 				return nil
 			}
-
-			if c.Content != "" || c.Thinking != "" || len(c.ToolCalls) > 0 {
-				fn(c)
-			}
 		}
 	}
 

model/models/models.go

@@ -12,5 +12,4 @@ import (
 	_ "github.com/ollama/ollama/model/models/qwen2"
 	_ "github.com/ollama/ollama/model/models/qwen25vl"
 	_ "github.com/ollama/ollama/model/models/qwen3"
-	_ "github.com/ollama/ollama/model/models/qwen3vl"
 )

model/models/qwen25vl/model.go

@@ -44,8 +44,8 @@ func New(c fs.Config) (model.Model, error) {
 			},
 		),
 		TextModel:      NewTextModel(c),
-		VisionModel:    NewVisionModel(c),
-		ImageProcessor: NewImageProcessor(c),
+		VisionModel:    newVisionModel(c),
+		ImageProcessor: newImageProcessor(c),
 	}
 
 	m.Cache = kvcache.NewCausalCache(m.TextModel.Shift)
@@ -65,8 +65,8 @@ func (m *Model) PixelValues(ctx ml.Context, multimodalData []byte) (ml.Tensor, *
 	}
 
 	// Calculate tensor dimensions
-	patchDim := m.ImageProcessor.NumChannels * m.ImageProcessor.TemporalPatchSize *
-		m.ImageProcessor.PatchSize * m.ImageProcessor.PatchSize
+	patchDim := m.ImageProcessor.numChannels * m.ImageProcessor.temporalPatchSize *
+		m.ImageProcessor.patchSize * m.ImageProcessor.patchSize
 	numPatches := grid.Temporal * grid.Height * grid.Width
 
 	pixelValues := ctx.Input().FromFloatSlice(f32s, patchDim, numPatches)

model/models/qwen25vl/model_vision.go

@@ -345,8 +345,8 @@ func (m *VisionModel) PositionalEmbedding(ctx ml.Context, grid *Grid) ml.Tensor
 	return positionalEmbedding
 }
 
-// NewVisionModel creates a new instance of the Qwen vision model
-func NewVisionModel(c fs.Config) *VisionModel {
+// newVisionModel creates a new instance of the Qwen vision model
+func newVisionModel(c fs.Config) *VisionModel {
 	patchSize := int(c.Uint("vision.patch_size", 14))
 	hiddenSize := int(c.Uint("vision.embedding_length", 1280))
 	numHeads := int(c.Uint("vision.attention.head_count", 16))

model/models/qwen25vl/process_image.go

@@ -11,40 +11,40 @@ import (
 
 // ImageProcessor contains configuration for the Qwen 2.5 VL image processing
 type ImageProcessor struct {
-	NumChannels       int
-	PatchSize         int
-	TemporalPatchSize int
-	MergeSize         int
-	MinPixels         int
-	MaxPixels         int
-	Factor            int
-	RescaleFactor     float32
-	ImageMean         []float32
-	ImageStd          []float32
+	numChannels       int
+	patchSize         int
+	temporalPatchSize int
+	mergeSize         int
+	minPixels         int
+	maxPixels         int
+	factor            int
+	rescaleFactor     float32
+	imageMean         []float32
+	imageStd          []float32
 }
 
 // newImageProcessor creates a new image processor with default values
-func NewImageProcessor(c fs.Config) ImageProcessor {
+func newImageProcessor(c fs.Config) ImageProcessor {
 	patchSize := int(c.Uint("vision.patch_size", 14))
 	mergeSize := int(c.Uint("vision.spatial_merge_size", 2))
 
 	return ImageProcessor{
-		NumChannels:       int(c.Uint("vision.num_channels", 3)), // not set
-		PatchSize:         patchSize,
-		TemporalPatchSize: 2,
-		MergeSize:         mergeSize,
-		MinPixels:         56 * 56,
-		MaxPixels:         int(c.Uint("vision.max_pixels", 28*28*1280)), // 1MP limit
-		Factor:            patchSize * mergeSize,
-		RescaleFactor:     1.0 / 255.0,
-		ImageMean:         imageproc.ClipDefaultMean[:],
-		ImageStd:          imageproc.ClipDefaultSTD[:],
+		numChannels:       int(c.Uint("vision.num_channels", 3)), // not set
+		patchSize:         patchSize,
+		temporalPatchSize: 2,
+		mergeSize:         mergeSize,
+		minPixels:         56 * 56,
+		maxPixels:         int(c.Uint("vision.max_pixels", 28*28*1280)), // 1MP limit
+		factor:            patchSize * mergeSize,
+		rescaleFactor:     1.0 / 255.0,
+		imageMean:         imageproc.ClipDefaultMean[:],
+		imageStd:          imageproc.ClipDefaultSTD[:],
 	}
 }
 
 // SmartResize implements the smart resize algorithm
 func (p *ImageProcessor) SmartResize(height, width int) (int, int) {
-	factor := p.Factor
+	factor := p.factor
 
 	if height < factor || width < factor {
 		panic(fmt.Sprintf("height:%d or width:%d must be larger than factor:%d", height, width, factor))
@@ -57,13 +57,13 @@ func (p *ImageProcessor) SmartResize(height, width int) (int, int) {
 	hBar := round(float64(height)/float64(factor)) * factor
 	wBar := round(float64(width)/float64(factor)) * factor
 
-	if hBar*wBar > p.MaxPixels {
-		beta := math.Sqrt(float64(height*width) / float64(p.MaxPixels))
+	if hBar*wBar > p.maxPixels {
+		beta := math.Sqrt(float64(height*width) / float64(p.maxPixels))
 
 		hBar = int(math.Floor(float64(height)/beta/float64(factor))) * factor
 		wBar = int(math.Floor(float64(width)/beta/float64(factor))) * factor
-	} else if hBar*wBar < p.MinPixels {
-		beta := math.Sqrt(float64(p.MinPixels) / float64(height*width))
+	} else if hBar*wBar < p.minPixels {
+		beta := math.Sqrt(float64(p.minPixels) / float64(height*width))
 
 		hBar = int(math.Ceil(float64(height)*beta/float64(factor))) * factor
 		wBar = int(math.Ceil(float64(width)*beta/float64(factor))) * factor
@@ -90,16 +90,16 @@ func (p *ImageProcessor) ProcessImage(img image.Image) ([]float32, *Grid, error)
 
 	normalizedPixels := imageproc.Normalize(
 		resizedImg,
-		[3]float32{p.ImageMean[0], p.ImageMean[1], p.ImageMean[2]},
-		[3]float32{p.ImageStd[0], p.ImageStd[1], p.ImageStd[2]},
+		[3]float32{p.imageMean[0], p.imageMean[1], p.imageMean[2]},
+		[3]float32{p.imageStd[0], p.imageStd[1], p.imageStd[2]},
 		true, // rescale
 		true, // channelFirst
 	)
 
 	// Calculate grid dimensions
 	grid := &Grid{
-		Height:   resizedHeight / p.PatchSize,
-		Width:    resizedWidth / p.PatchSize,
+		Height:   resizedHeight / p.patchSize,
+		Width:    resizedWidth / p.patchSize,
 		Temporal: 1, // For single images, temporal dimension is 1
 	}
 
@@ -113,10 +113,10 @@ func (p *ImageProcessor) ProcessImage(img image.Image) ([]float32, *Grid, error)
 }
 
 func (p *ImageProcessor) createPatches(pixels []float32, height, width int, grid *Grid) ([]float32, error) {
-	channels := p.NumChannels
-	patchSize := p.PatchSize
-	mergeSize := p.MergeSize
-	temporalPatchSize := p.TemporalPatchSize
+	channels := p.numChannels
+	patchSize := p.patchSize
+	mergeSize := p.mergeSize
+	temporalPatchSize := p.temporalPatchSize
 
 	// Calculate output dimensions
 	numPatches := grid.Temporal * grid.Height * grid.Width

model/models/qwen3vl/model.go

@@ -1,153 +0,0 @@
-package qwen3vl
-
-import (
-	"bytes"
-	"fmt"
-	"image"
-	"slices"
-
-	"github.com/ollama/ollama/fs"
-	"github.com/ollama/ollama/kvcache"
-	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/model"
-	"github.com/ollama/ollama/model/input"
-	"github.com/ollama/ollama/model/models/qwen25vl"
-	"github.com/ollama/ollama/model/models/qwen3"
-)
-
-type Model struct {
-	model.Base
-	model.BytePairEncoding
-
-	TextModel *qwen3.Model
-	*qwen25vl.VisionModel
-
-	qwen25vl.ImageProcessor
-}
-
-var _ model.MultimodalProcessor = (*Model)(nil)
-
-func New(c fs.Config) (model.Model, error) {
-	textModel, err := qwen3.New(c)
-	if err != nil {
-		return nil, err
-	}
-
-	m := &Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
-			&model.Vocabulary{
-				Values: c.Strings("tokenizer.ggml.tokens"),
-				Types:  c.Ints("tokenizer.ggml.token_type"),
-				Merges: c.Strings("tokenizer.ggml.merges"),
-				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
-				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
-				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
-				EOS: append(
-					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
-					c.Ints("tokenizer.ggml.eos_token_ids")...,
-				),
-			},
-		),
-		TextModel:      textModel.(*qwen3.Model),
-		VisionModel:    qwen25vl.NewVisionModel(c),
-		ImageProcessor: qwen25vl.NewImageProcessor(c),
-	}
-
-	m.Cache = kvcache.NewCausalCache(m.TextModel.Shift)
-
-	return m, nil
-}
-
-func (m *Model) PixelValues(ctx ml.Context, multimodalData []byte) (ml.Tensor, *qwen25vl.Grid, error) {
-	image, _, err := image.Decode(bytes.NewReader(multimodalData))
-	if err != nil {
-		return nil, nil, err
-	}
-
-	f32s, grid, err := m.ImageProcessor.ProcessImage(image)
-	if err != nil {
-		return nil, nil, err
-	}
-
-	// Calculate tensor dimensions
-	patchDim := m.ImageProcessor.NumChannels * m.ImageProcessor.TemporalPatchSize *
-		m.ImageProcessor.PatchSize * m.ImageProcessor.PatchSize
-	numPatches := grid.Temporal * grid.Height * grid.Width
-
-	pixelValues := ctx.Input().FromFloatSlice(f32s, patchDim, numPatches)
-
-	return pixelValues, grid, nil
-}
-
-func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input.Multimodal, error) {
-	if len(m.VisionModel.Layers) == 0 {
-		return nil, model.ErrNoVisionModel
-	}
-
-	pixels, grid, err := m.PixelValues(ctx, multimodalData)
-	if err != nil {
-		return nil, err
-	}
-
-	visionOutputs := m.VisionModel.Forward(ctx, pixels, grid)
-	return []input.Multimodal{{Tensor: visionOutputs}}, nil
-}
-
-// PostTokenize arranges Qwen-3-VL's inputs for the forward pass
-func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
-	var result []*input.Input
-
-	var (
-		imageToken       int32 = 151655
-		visionStartToken int32 = 151652
-		visionEndToken   int32 = 151653
-	)
-
-	nImg := 0
-	for _, inp := range inputs {
-		if inp.Multimodal == nil {
-			// If not a multimodal input, add it to the result unchanged
-			result = append(result, inp)
-		} else {
-			// Adding the 'Picture' prefix is a hack, at the time of writing there is no way to prefix
-			// the image tokens with a prompt, so we add a prefix here
-			nImg++
-			pre, err := m.Encode(fmt.Sprintf(" Picture %d: ", nImg), true)
-			if err != nil {
-				return nil, fmt.Errorf("failed to encode image prompt: %w", err)
-			}
-			for i := range pre {
-				result = append(result, &input.Input{Token: pre[i]})
-			}
-
-			patchesPerChunk := inp.Multimodal[0].Tensor.Dim(1)
-
-			// First add the vision start token
-			result = append(result, &input.Input{Token: visionStartToken})
-
-			// Add the image token with the multimodal tensor data at the first position
-			result = append(result, &input.Input{
-				Token:          imageToken,
-				Multimodal:     inp.Multimodal,
-				MultimodalHash: inp.MultimodalHash,
-				SameBatch:      patchesPerChunk,
-			})
-
-			// Add the placeholder tokens for the remaining positions (tokensPerGrid-1)
-			result = append(result, slices.Repeat([]*input.Input{{Token: imageToken}}, patchesPerChunk-1)...)
-
-			result = append(result, &input.Input{Token: visionEndToken})
-		}
-	}
-
-	return result, nil
-}
-
-func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	return m.TextModel.Forward(ctx, batch)
-}
-
-func init() {
-	model.Register("qwen3vl", New)
-}

openai/openai.go

@@ -76,8 +76,9 @@ type JsonSchema struct {
 }
 
 type EmbedRequest struct {
-	Input any    `json:"input"`
-	Model string `json:"model"`
+	Input      any    `json:"input"`
+	Model      string `json:"model"`
+	Dimensions int    `json:"dimensions,omitempty"`
 }
 
 type StreamOptions struct {
@@ -1005,7 +1006,7 @@ func EmbeddingsMiddleware() gin.HandlerFunc {
 		}
 
 		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(api.EmbedRequest{Model: req.Model, Input: req.Input}); err != nil {
+		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
 		}

runner/ollamarunner/runner.go

@@ -18,7 +18,6 @@ import (
 	"reflect"
 	"regexp"
 	"runtime"
-	"runtime/debug"
 	"strconv"
 	"strings"
 	"sync"
@@ -30,7 +29,6 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/harmony"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/ml"
@@ -782,14 +780,6 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	var harmonyMessageHandler *harmony.HarmonyMessageHandler
-	var harmonyToolParser *harmony.HarmonyToolCallAccumulator
-	if req.UseHarmony {
-		harmonyMessageHandler = harmony.NewHarmonyMessageHandler()
-		harmonyMessageHandler.HarmonyParser.AddImplicitStartOrPrefill(req.PrefillString)
-		harmonyToolParser = harmonyMessageHandler.CreateToolParser()
-	}
-
 	if req.Options == nil {
 		opts := api.DefaultOptions()
 		req.Options = &opts
@@ -872,9 +862,6 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 		http.Error(w, "could not find an available sequence", http.StatusInternalServerError)
 		return
 	}
-	var lastToken string
-	tokenRepeat := 0
-	const tokenRepeatLimit = 30
 
 	for {
 		select {
@@ -883,27 +870,8 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 			return
 		case content, ok := <-seq.responses:
 			if ok {
-				if strings.TrimSpace(content) == lastToken {
-					tokenRepeat++
-				}
-				if tokenRepeat == tokenRepeatLimit {
-					http.Error(w, "token repeat limit reached", http.StatusInternalServerError)
-					seq.doneReason = llm.DoneReasonTokenRepeatLimit
-					close(seq.quit)
-					return
-				}
-				lastToken = strings.TrimSpace(content)
-
-				var thinking string
-				if harmonyMessageHandler != nil {
-					var toolContent string
-					content, thinking, toolContent = harmonyMessageHandler.AddContent(content, harmonyToolParser)
-					harmonyToolParser.Add(toolContent)
-				}
-
 				if err := json.NewEncoder(w).Encode(&llm.CompletionResponse{
-					Content:  content,
-					Thinking: thinking,
+					Content: content,
 				}); err != nil {
 					http.Error(w, fmt.Sprintf("failed to encode response: %v", err), http.StatusInternalServerError)
 					close(seq.quit)
@@ -912,29 +880,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 
 				flusher.Flush()
 			} else {
-				var toolCalls []api.ToolCall
-				if harmonyMessageHandler != nil {
-					// these tools still need to be transformed to the original function name
-					toolName, toolContent := harmonyToolParser.Drain()
-					if toolName != nil {
-						*toolName = strings.TrimPrefix(*toolName, "functions.")
-						var args api.ToolCallFunctionArguments
-						if err := json.Unmarshal([]byte(toolContent), &args); err != nil {
-							http.Error(w, fmt.Sprintf("failed to unmarshal tool call function arguments: %v", err), http.StatusInternalServerError)
-							close(seq.quit)
-							return
-						}
-						toolCalls = append(toolCalls, api.ToolCall{
-							Function: api.ToolCallFunction{
-								Name:      *toolName,
-								Arguments: args,
-							},
-						})
-					}
-				}
-
 				if err := json.NewEncoder(w).Encode(&llm.CompletionResponse{
-					ToolCalls:          toolCalls,
 					Done:               true,
 					DoneReason:         seq.doneReason,
 					PromptEvalCount:    seq.numPromptInputs,
@@ -1139,9 +1085,13 @@ func (s *Server) allocModel(
 	// Convert memory allocation panics to errors
 	defer func() {
 		if r := recover(); r != nil {
-			debug.PrintStack()
 			if err, ok := r.(error); ok {
-				panicErr = err
+				var noMem ml.ErrNoMem
+				if errors.As(err, &noMem) {
+					panicErr = noMem
+				} else {
+					panic(r)
+				}
 			} else {
 				panic(r)
 			}

scripts/build_windows.ps1

@@ -78,7 +78,7 @@ function checkEnv() {
 }
 
 
-function buildOllama() {
+function buildCPU() {
     mkdir -Force -path "${script:DIST_DIR}\"
     if ($script:ARCH -ne "arm64") {
         Remove-Item -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}"
@@ -90,20 +90,72 @@ function buildOllama() {
         if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
         & cmake --install build --component CPU --strip
         if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+    }
+}
 
+function buildCUDA11() {
+    # CUDA v11 claims to be compatible with MSVC 2022, but the latest updates are no longer compatible
+    # 19.40 is the last compiler version that works, but recent udpates are 19.43
+    # So this pins to MSVC 2019 for best compatibility
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
         $hashEnv = @{}
         Get-ChildItem env: | foreach { $hashEnv[$_.Name] = $_.Value }
-        if ("$script:CUDA_DIRS".Contains("v12")) {
-            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V12")) { $v12="$_" }}
-            $env:CUDAToolkit_ROOT=$hashEnv[$v12]
-            write-host "Building CUDA v12 backend libraries"
-            & cmake --fresh --preset "CUDA 12" --install-prefix $script:DIST_DIR
+        if ("$script:CUDA_DIRS".Contains("v11")) {
+            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V11")) { $x=$hashEnv[$_]; if (test-path -literalpath "$x\bin\nvcc.exe" ) { $cuda=$x}  }}
+            write-host "Building CUDA v11 backend libraries $cuda"
+            $env:CUDAToolkit_ROOT=$cuda
+            & cmake --fresh --preset "CUDA 11" -T cuda="$cuda" -DCMAKE_CUDA_COMPILER="$cuda\bin\nvcc.exe" -G "Visual Studio 16 2019" --install-prefix $script:DIST_DIR -DOLLAMA_RUNNER_DIR="cuda_v11"
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --build --preset "CUDA 11"  --config Release --parallel $script:JOBS
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --install build --component "CUDA" --strip
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+        }
+    }
+}
+
+function buildCUDA12() {
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
+        $hashEnv = @{}
+        Get-ChildItem env: | foreach { $hashEnv[$_.Name] = $_.Value }
+        if ("$script:CUDA_DIRS".Contains("v12.8")) {
+            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V12_8")) { $x=$hashEnv[$_]; if (test-path -literalpath "$x\bin\nvcc.exe" ) { $cuda=$x}  }}
+            write-host "Building CUDA v12 backend libraries $cuda"
+            $env:CUDAToolkit_ROOT=$cuda
+            & cmake --fresh --preset "CUDA 12" -T cuda="$cuda" --install-prefix $script:DIST_DIR -DOLLAMA_RUNNER_DIR="cuda_v12"
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
             & cmake --build --preset "CUDA 12"  --config Release --parallel $script:JOBS
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
             & cmake --install build --component "CUDA" --strip
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
         }
+    }
+}
+
+function buildCUDA13() {
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
+        $hashEnv = @{}
+        Get-ChildItem env: | foreach { $hashEnv[$_.Name] = $_.Value }
+        if ("$script:CUDA_DIRS".Contains("v13")) {
+            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V13")) { $x=$hashEnv[$_]; if (test-path -literalpath "$x\bin\nvcc.exe" ) { $cuda=$x}  }}
+            $env:CUDAToolkit_ROOT=$cuda
+            write-host "Building CUDA v13 backend libraries $cuda"
+            & cmake --fresh --preset "CUDA 13" -T cuda="$cuda" --install-prefix $script:DIST_DIR -DOLLAMA_RUNNER_DIR="cuda_v13"
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --build --preset "CUDA 13"  --config Release --parallel $script:JOBS
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --install build --component "CUDA" --strip
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+        }
+    }
+}
+
+function buildROCm() {
+    mkdir -Force -path "${script:DIST_DIR}\"
+    if ($script:ARCH -ne "arm64") {
         if ($env:HIP_PATH) {
             write-host "Building ROCm backend libraries"
             if (-Not (get-command -ErrorAction silent ninja)) {
@@ -129,6 +181,10 @@ function buildOllama() {
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
         }
     }
+}
+
+function buildOllama() {
+    mkdir -Force -path "${script:DIST_DIR}\"
     write-host "Building ollama CLI"
     & go build -trimpath -ldflags "-s -w -X=github.com/ollama/ollama/version.Version=$script:VERSION -X=github.com/ollama/ollama/server.mode=release" .
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
@@ -236,6 +292,10 @@ function distZip() {
 checkEnv
 try {
     if ($($args.count) -eq 0) {
+        buildCPU
+        buildCUDA12
+        buildCUDA13
+        buildROCm
         buildOllama
         buildApp
         gatherDependencies

server/routes.go

@@ -46,6 +46,18 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
+func shouldUseHarmony(model *Model) bool {
+	if slices.Contains([]string{"gptoss", "gpt-oss"}, model.Config.ModelFamily) {
+		// heuristic to check whether the template expects to be parsed via harmony:
+		// search for harmony tags that are nearly always used
+		if model.Template.Contains("<|start|>") && model.Template.Contains("<|end|>") {
+			return true
+		}
+	}
+
+	return false
+}
+
 func experimentEnabled(name string) bool {
 	return slices.Contains(strings.Split(os.Getenv("OLLAMA_EXPERIMENT"), ","), name)
 }
@@ -195,11 +207,13 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	useHarmony := harmony.ShouldUseHarmony(m.Config.ModelFamily, m.Template) && !req.Raw
-	var functionNameMap *harmony.FunctionNameMap
-
+	useHarmony := shouldUseHarmony(m) && !req.Raw
+	var harmonyMessageHandler *harmony.HarmonyMessageHandler
+	var harmonyToolParser *harmony.HarmonyToolCallAccumulator
 	if useHarmony {
-		functionNameMap = harmony.NewFunctionNameMap()
+		harmonyMessageHandler = harmony.NewHarmonyMessageHandler()
+		harmonyMessageHandler.HarmonyParser.AddImplicitStart()
+		harmonyToolParser = harmonyMessageHandler.CreateToolParser()
 	}
 
 	// Validate Think value: string values currently only allowed for gptoss models
@@ -343,19 +357,16 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		var sb strings.Builder
 		defer close(ch)
 		if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
-			Prompt:     prompt,
-			Images:     images,
-			Format:     req.Format,
-			Options:    opts,
-			UseHarmony: useHarmony,
+			Prompt:  prompt,
+			Images:  images,
+			Format:  req.Format,
+			Options: opts,
 		}, func(cr llm.CompletionResponse) {
 			res := api.GenerateResponse{
 				Model:     req.Model,
 				CreatedAt: time.Now().UTC(),
 				Response:  cr.Content,
 				Done:      cr.Done,
-				Thinking:  cr.Thinking,
-				ToolCalls: cr.ToolCalls,
 				Metrics: api.Metrics{
 					PromptEvalCount:    cr.PromptEvalCount,
 					PromptEvalDuration: cr.PromptEvalDuration,
@@ -364,22 +375,12 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 				},
 			}
 
-			if res.Done {
-				res.DoneReason = cr.DoneReason.String()
-				res.TotalDuration = time.Since(checkpointStart)
-				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
-			}
-
 			if useHarmony {
-				for i, tool := range res.ToolCalls {
-					res.ToolCalls[i].Function.Name = functionNameMap.OriginalFromConverted(tool.Function.Name)
-				}
-				if res.Response != "" || res.Thinking != "" || len(res.ToolCalls) > 0 || res.Done {
-					ch <- res
-				}
-				return
-			}
-			if thinkingState != nil {
+				content, thinking, toolContent := harmonyMessageHandler.AddContent(cr.Content, harmonyToolParser)
+				res.Response = content
+				res.Thinking = thinking
+				harmonyToolParser.Add(toolContent)
+			} else if thinkingState != nil {
 				thinking, content := thinkingState.AddContent(cr.Content)
 				res.Thinking = thinking
 				res.Response = content
@@ -390,6 +391,30 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 			}
 
 			if cr.Done {
+				if useHarmony {
+					toolName, toolContent := harmonyToolParser.Drain()
+					if toolName != nil {
+						*toolName = strings.TrimPrefix(*toolName, "functions.")
+						var args api.ToolCallFunctionArguments
+						if err := json.Unmarshal([]byte(toolContent), &args); err != nil {
+							errStr := fmt.Sprintf("error parsing tool call: raw='%s', err=%s", toolContent, err.Error())
+							ch <- gin.H{"error": errStr}
+							return
+						}
+
+						res.ToolCalls = append(res.ToolCalls, api.ToolCall{
+							Function: api.ToolCallFunction{
+								Name:      *toolName,
+								Arguments: args,
+							},
+						})
+					}
+				}
+
+				res.DoneReason = cr.DoneReason.String()
+				res.TotalDuration = time.Since(checkpointStart)
+				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+
 				if !req.Raw {
 					tokens, err := r.Tokenize(c.Request.Context(), prompt+sb.String())
 					if err != nil {
@@ -551,7 +576,12 @@ func (s *Server) EmbedHandler(c *gin.Context) {
 			if err != nil {
 				return err
 			}
-			embeddings[i] = normalize(embedding)
+			// TODO: this first normalization should be done by the model
+			embedding = normalize(embedding)
+			if req.Dimensions > 0 && req.Dimensions < len(embedding) {
+				embedding = normalize(embedding[:req.Dimensions])
+			}
+			embeddings[i] = embedding
 			return nil
 		})
 	}
@@ -577,11 +607,7 @@ func normalize(vec []float32) []float32 {
 		sum += v * v
 	}
 
-	norm := float32(0.0)
-	if sum > 0 {
-		norm = float32(1.0 / math.Sqrt(float64(sum)))
-	}
-
+	norm := float32(1.0 / max(math.Sqrt(float64(sum)), 1e-12))
 	for i := range vec {
 		vec[i] *= norm
 	}
@@ -1591,21 +1617,27 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	}
 	msgs = filterThinkTags(msgs, m)
 
-	useHarmony := harmony.ShouldUseHarmony(m.Config.ModelFamily, m.Template)
+	var harmonyMessageHandler *harmony.HarmonyMessageHandler
+	var harmonyToolParser *harmony.HarmonyToolCallAccumulator
+
+	useHarmony := shouldUseHarmony(m)
 
 	processedTools := req.Tools
-	var functionNameMap *harmony.FunctionNameMap
-	var prefillString string
-	// TODO(parthsareen): this can be abstracted to not be model specific and potentially moved to the runner
 	if useHarmony {
-		prefillString = harmony.Prefill(msgs[len(msgs)-1])
-		functionNameMap = harmony.NewFunctionNameMap()
+		harmonyMessageHandler = harmony.NewHarmonyMessageHandler()
+		var lastMessage *api.Message
+		if len(msgs) > 0 {
+			lastMessage = &msgs[len(msgs)-1]
+		}
+		harmonyMessageHandler.HarmonyParser.AddImplicitStartOrPrefill(lastMessage)
+		harmonyToolParser = harmonyMessageHandler.CreateToolParser()
+
 		// make a copy of tools to pass to the chat prompt. Function names may be
 		// renamed to be valid Harmony function names.
 		processedTools = make([]api.Tool, len(req.Tools))
 		copy(processedTools, req.Tools)
 		for i, tool := range processedTools {
-			processedTools[i].Function.Name = functionNameMap.ConvertAndAdd(tool.Function.Name)
+			processedTools[i].Function.Name = harmonyMessageHandler.FunctionNameMap.ConvertAndAdd(tool.Function.Name)
 		}
 	}
 
@@ -1658,17 +1690,15 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		defer close(ch)
 
 		if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
-			Prompt:        prompt,
-			Images:        images,
-			Format:        req.Format,
-			Options:       opts,
-			UseHarmony:    useHarmony,
-			PrefillString: prefillString,
+			Prompt:  prompt,
+			Images:  images,
+			Format:  req.Format,
+			Options: opts,
 		}, func(r llm.CompletionResponse) {
 			res := api.ChatResponse{
 				Model:     req.Model,
 				CreatedAt: time.Now().UTC(),
-				Message:   api.Message{Role: "assistant", Content: r.Content, Thinking: r.Thinking, ToolCalls: r.ToolCalls},
+				Message:   api.Message{Role: "assistant", Content: r.Content},
 				Done:      r.Done,
 				Metrics: api.Metrics{
 					PromptEvalCount:    r.PromptEvalCount,
@@ -1684,13 +1714,31 @@ func (s *Server) ChatHandler(c *gin.Context) {
 			}
 
 			if useHarmony {
-				for i, tool := range res.Message.ToolCalls {
-					res.Message.ToolCalls[i].Function.Name = functionNameMap.OriginalFromConverted(tool.Function.Name)
+				content, thinking, toolContent := harmonyMessageHandler.AddContent(r.Content, harmonyToolParser)
+				res.Message.Content = content
+				res.Message.Thinking = thinking
+				harmonyToolParser.Add(toolContent)
+
+				if r.Done {
+					toolName, toolContent := harmonyToolParser.Drain()
+					if toolName != nil {
+						*toolName = strings.TrimPrefix(*toolName, "functions.")
+						*toolName = harmonyMessageHandler.FunctionNameMap.OriginalFromConverted(*toolName)
+						var args api.ToolCallFunctionArguments
+						if err := json.Unmarshal([]byte(toolContent), &args); err != nil {
+							errStr := fmt.Sprintf("error parsing tool call: raw='%s', err=%s", toolContent, err.Error())
+							ch <- gin.H{"error": errStr}
+							return
+						}
+						res.Message.ToolCalls = []api.ToolCall{{Function: api.ToolCallFunction{Name: *toolName, Arguments: args}}}
+					}
 				}
+
 				// only send messages with meaningful content (empty messages confuse clients)
 				if res.Message.Content != "" || res.Message.Thinking != "" || len(res.Message.ToolCalls) > 0 || res.Done {
 					ch <- res
 				}
+
 				return
 			}
 

server/routes_harmony_streaming_test.go

@@ -7,6 +7,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"net/http"
 	"strings"
 	"testing"
 	"time"
@@ -117,7 +118,7 @@ func TestChatHarmonyParserStreamingRealtime(t *testing.T) {
 			name: "content streams as it arrives",
 			steps: []step{
 				{
-					input:       llm.CompletionResponse{Content: "Hello", Done: false},
+					input:       llm.CompletionResponse{Content: "<|message|>Hello", Done: false},
 					wantContent: "Hello",
 				},
 				{
@@ -125,7 +126,7 @@ func TestChatHarmonyParserStreamingRealtime(t *testing.T) {
 					wantContent: ", world",
 				},
 				{
-					input:       llm.CompletionResponse{Content: "!", Done: true, DoneReason: llm.DoneReasonStop},
+					input:       llm.CompletionResponse{Content: "!<|end|>", Done: true, DoneReason: llm.DoneReasonStop},
 					wantContent: "!",
 				},
 			},
@@ -134,15 +135,20 @@ func TestChatHarmonyParserStreamingRealtime(t *testing.T) {
 			name: "thinking streams separately from content",
 			steps: []step{
 				{
-					input:        llm.CompletionResponse{Thinking: "Thinking...", Done: false},
+					input:        llm.CompletionResponse{Content: "<|channel|>analysis<|message|>Thinking...", Done: false},
 					wantThinking: "Thinking...",
 				},
 				{
-					input:       llm.CompletionResponse{Content: "Answer", Done: false},
-					wantContent: "Answer",
+					input: llm.CompletionResponse{Content: "<|end|>", Done: false},
+					// No output expected - just closes the analysis message and resets state to normal
 				},
 				{
-					input: llm.CompletionResponse{Done: true, DoneReason: llm.DoneReasonStop},
+					input:       llm.CompletionResponse{Content: "<|start|>assistant<|message|>Answer", Done: false},
+					wantContent: "Answer", // After message end, state is reset to normal
+				},
+				{
+					input: llm.CompletionResponse{Content: "<|end|>", Done: true, DoneReason: llm.DoneReasonStop},
+					// No output expected - just closes the assistant message
 				},
 			},
 		},
@@ -150,16 +156,24 @@ func TestChatHarmonyParserStreamingRealtime(t *testing.T) {
 			name: "partial tags buffer until complete",
 			steps: []step{
 				{
-					input:        llm.CompletionResponse{Thinking: "Deep ", Done: false},
+					input: llm.CompletionResponse{Content: "<|chan", Done: false},
+					// No output - partial tag
+				},
+				{
+					input: llm.CompletionResponse{Content: "nel|>analysis<|mess", Done: false},
+					// No output - still building tags
+				},
+				{
+					input:        llm.CompletionResponse{Content: "age|>Deep ", Done: false},
 					wantThinking: "Deep ",
 				},
 				{
-					input:        llm.CompletionResponse{Thinking: "thought", Done: false},
+					input:        llm.CompletionResponse{Content: "thought<|end|>", Done: false},
 					wantThinking: "thought",
 				},
 				{
-					input:       llm.CompletionResponse{Content: "Done", Done: true, DoneReason: llm.DoneReasonStop},
-					wantContent: "Done",
+					input:       llm.CompletionResponse{Content: "<|start|>assistant<|message|>Done<|end|>", Done: true, DoneReason: llm.DoneReasonStop},
+					wantContent: "Done", // After message end, state is reset to normal
 				},
 			},
 		},
@@ -167,7 +181,7 @@ func TestChatHarmonyParserStreamingRealtime(t *testing.T) {
 			name: "simple assistant after analysis",
 			steps: []step{
 				{
-					input:        llm.CompletionResponse{Thinking: "Think", Content: "Answer", Done: true, DoneReason: llm.DoneReasonStop},
+					input:        llm.CompletionResponse{Content: "<|channel|>analysis<|message|>Think<|end|><|start|>assistant<|message|>Answer<|end|>", Done: true, DoneReason: llm.DoneReasonStop},
 					wantContent:  "Answer",
 					wantThinking: "Think",
 				},
@@ -177,7 +191,7 @@ func TestChatHarmonyParserStreamingRealtime(t *testing.T) {
 			name: "tool call parsed and returned correctly",
 			steps: []step{
 				{
-					input:       llm.CompletionResponse{Content: "The weather is sunny", ToolCalls: []api.ToolCall{{Function: api.ToolCallFunction{Name: "get_weather", Arguments: api.ToolCallFunctionArguments{"location": "San Francisco"}}}}, Done: true, DoneReason: llm.DoneReasonStop},
+					input:       llm.CompletionResponse{Content: "<|channel|>commentary to=functions.get_weather<|message|>{\"location\":\"San Francisco\"}<|end|><|start|>assistant<|message|>The weather is sunny<|end|>", Done: true, DoneReason: llm.DoneReasonStop},
 					wantContent: "The weather is sunny",
 					wantToolCalls: []api.ToolCall{
 						{
@@ -196,10 +210,15 @@ func TestChatHarmonyParserStreamingRealtime(t *testing.T) {
 			name: "tool call with streaming JSON across chunks",
 			steps: []step{
 				{
-					input: llm.CompletionResponse{Done: false},
+					input: llm.CompletionResponse{Content: "<|channel|>commentary to=functions.calculate<|message|>{\"expr", Done: false},
+					// No output yet - incomplete JSON
 				},
 				{
-					input: llm.CompletionResponse{ToolCalls: []api.ToolCall{{Function: api.ToolCallFunction{Name: "calculate", Arguments: api.ToolCallFunctionArguments{"expression": "2+2"}}}}, Done: true},
+					input: llm.CompletionResponse{Content: "ession\":\"2+", Done: false},
+					// Still no output - incomplete JSON
+				},
+				{
+					input: llm.CompletionResponse{Content: "2\"}", Done: true},
 					wantToolCalls: []api.ToolCall{
 						{
 							Function: api.ToolCallFunction{
@@ -381,9 +400,9 @@ func TestChatHarmonyParserStreamingSimple(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 
 	mockResponses := []llm.CompletionResponse{
-		{Content: "First ", Done: false},
+		{Content: "<|message|>First ", Done: false},
 		{Content: "chunk ", Done: false},
-		{Content: "here", Done: true, DoneReason: llm.DoneReasonStop},
+		{Content: "here<|end|>", Done: true, DoneReason: llm.DoneReasonStop},
 	}
 
 	mock := mockRunner{
@@ -488,3 +507,189 @@ func TestChatHarmonyParserStreamingSimple(t *testing.T) {
 		t.Errorf("expected at least 2 content chunks for streaming, got %d", contentChunks)
 	}
 }
+
+func TestChatHarmonyParserStreaming(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	type expectedChunk struct {
+		afterResponse int    // Which mock response this chunk should appear after
+		content       string // Expected content in this chunk
+		thinking      string // Expected thinking in this chunk
+	}
+
+	testCases := []struct {
+		name           string
+		mockResponses  []llm.CompletionResponse
+		expectedChunks []expectedChunk
+		wantContent    string
+		wantThinking   string
+	}{
+		{
+			name: "simple message without thinking",
+			mockResponses: []llm.CompletionResponse{
+				{Content: "<|start|>assistant<|message|>Hello, ", Done: false},
+				{Content: "how can I help?", Done: false},
+				{Content: "<|end|>", Done: true, DoneReason: llm.DoneReasonStop},
+			},
+			expectedChunks: []expectedChunk{
+				{afterResponse: 1, content: "Hello, "},
+				{afterResponse: 2, content: "how can I help?"},
+			},
+			wantContent: "Hello, how can I help?",
+		},
+		{
+			name: "message with analysis channel for thinking",
+			mockResponses: []llm.CompletionResponse{
+				{Content: "<|channel|>analysis<|message|>", Done: false},
+				{Content: "Let me think ", Done: false},
+				{Content: "about this problem...", Done: false},
+				{Content: "<|end|>", Done: false},
+				{Content: "<|start|>assistant<|message|>", Done: false},
+				{Content: "The answer ", Done: false},
+				{Content: "is 42", Done: false},
+				{Content: "<|end|>", Done: true, DoneReason: llm.DoneReasonStop},
+			},
+			expectedChunks: []expectedChunk{
+				{afterResponse: 2, thinking: "Let me think "},
+				{afterResponse: 3, thinking: "about this problem..."},
+				{afterResponse: 6, content: "The answer "},
+				{afterResponse: 7, content: "is 42"},
+			},
+			wantContent:  "The answer is 42",
+			wantThinking: "Let me think about this problem...",
+		},
+		{
+			name: "streaming with partial tags across boundaries",
+			mockResponses: []llm.CompletionResponse{
+				{Content: "<|chan", Done: false},
+				{Content: "nel|>analy", Done: false},
+				{Content: "sis<|mess", Done: false},
+				{Content: "age|>Think", Done: false},
+				{Content: "ing deeply...<|end|>", Done: false},
+				{Content: "<|start|>assi", Done: false},
+				{Content: "stant<|message|>Result ", Done: false},
+				{Content: "computed<|e", Done: false},
+				{Content: "nd|>", Done: true, DoneReason: llm.DoneReasonStop},
+			},
+			expectedChunks: []expectedChunk{
+				{afterResponse: 4, thinking: "Think"},
+				{afterResponse: 5, thinking: "ing deeply..."},
+				{afterResponse: 7, content: "Result "},
+				{afterResponse: 8, content: "computed"},
+			},
+			wantContent:  "Result computed",
+			wantThinking: "Thinking deeply...",
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			// Channel to synchronize mock responses with chunk verification
+			responsesSent := make(chan int, len(tc.mockResponses))
+
+			mock := mockRunner{
+				CompletionFn: func(ctx context.Context, r llm.CompletionRequest, fn func(llm.CompletionResponse)) error {
+					// Send mock responses one at a time, notifying when each is sent
+					for i, resp := range tc.mockResponses {
+						fn(resp)
+						responsesSent <- i + 1
+					}
+					close(responsesSent)
+					return nil
+				},
+			}
+
+			s := Server{
+				sched: &Scheduler{
+					pendingReqCh:  make(chan *LlmRequest, 1),
+					finishedReqCh: make(chan *LlmRequest, 1),
+					expiredCh:     make(chan *runnerRef, 1),
+					unloadedCh:    make(chan any, 1),
+					loaded:        make(map[string]*runnerRef),
+					newServerFn:   newMockServer(&mock),
+					getGpuFn:      discover.GetGPUInfo,
+					getCpuFn:      discover.GetCPUInfo,
+					reschedDelay:  250 * time.Millisecond,
+					loadFn: func(req *LlmRequest, _ *ggml.GGML, _ discover.GpuInfoList, _ bool) bool {
+						req.successCh <- &runnerRef{
+							llama: &mock,
+						}
+						return false
+					},
+				},
+			}
+
+			go s.sched.Run(t.Context())
+
+			// Create a minimal model
+			_, digest := createHarmonyTestModel(t)
+
+			// Create model with passthrough template
+			stream := false
+			w := createRequest(t, s.CreateHandler, api.CreateRequest{
+				Model:    "harmony-test",
+				Files:    map[string]string{"file.gguf": digest},
+				Template: `<|start|><|end|>{{ with .Tools }}{{ end }}{{ .Prompt }}`,
+				Stream:   &stream,
+			})
+
+			if w.Code != http.StatusOK {
+				t.Fatalf("failed to create model: %d", w.Code)
+			}
+
+			// Test chat endpoint with streaming
+			streamTrue := true
+			w = createRequest(t, s.ChatHandler, api.ChatRequest{
+				Model:    "harmony-test",
+				Messages: []api.Message{{Role: "user", Content: "Hello"}},
+				Stream:   &streamTrue,
+				Tools:    getTestTools(),
+			})
+
+			if w.Code != http.StatusOK {
+				t.Fatalf("chat request failed: %d - %s", w.Code, w.Body.String())
+			}
+
+			// Parse streaming response
+			var chunks []api.ChatResponse
+			var content, thinking strings.Builder
+
+			decoder := json.NewDecoder(w.Body)
+			for decoder.More() {
+				var chunk api.ChatResponse
+				if err := decoder.Decode(&chunk); err != nil {
+					t.Fatalf("failed to decode chunk: %v", err)
+				}
+				chunks = append(chunks, chunk)
+
+				// Accumulate content and thinking from each chunk
+				content.WriteString(chunk.Message.Content)
+				thinking.WriteString(chunk.Message.Thinking)
+
+				// Debug output
+				t.Logf("Chunk %d: content=%q thinking=%q done=%v", len(chunks), chunk.Message.Content, chunk.Message.Thinking, chunk.Done)
+			}
+
+			// Verify we got streaming chunks
+			if len(chunks) == 0 {
+				t.Fatal("expected streaming chunks, got none")
+			}
+
+			gotContent := content.String()
+			gotThinking := thinking.String()
+
+			if gotContent != tc.wantContent {
+				t.Errorf("content mismatch: got %q, want %q", gotContent, tc.wantContent)
+			}
+			if gotThinking != tc.wantThinking {
+				t.Errorf("thinking mismatch: got %q, want %q", gotThinking, tc.wantThinking)
+			}
+
+			// Verify last chunk has done=true
+			lastChunk := chunks[len(chunks)-1]
+			if !lastChunk.Done {
+				t.Error("expected last chunk to have done=true")
+			}
+		})
+	}
+}