runner: enable returning more info from runner processing

Currently we return only the text predicted from the LLM. This was nice in
that it was simple, but there may be other info we want to know from the
processing. This change adds the ability to return more information from the
runner than just the text predicted.

.github/workflows/release.yaml

@@ -103,6 +103,11 @@ jobs:
         arch: [amd64]
         preset: ['CPU']
         include:
+          - os: windows
+            arch: amd64
+            preset: 'CUDA 11'
+            install: https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe
+            cuda-version: '11.3'
           - os: windows
             arch: amd64
             preset: 'CUDA 12'
@@ -155,9 +160,6 @@ jobs:
           echo "$hipPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "CC=$hipPath\bin\clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
           echo "CXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "HIPCXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "HIP_PLATFORM=amd" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "CMAKE_PREFIX_PATH=$hipPath" | Out-File -FilePath $env:GITHUB_ENV -Append
       - if: matrix.preset == 'CPU'
         run: |
           echo "CC=clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
@@ -176,8 +178,8 @@ jobs:
           key: ccache-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}
       - name: Build target "${{ matrix.preset }}"
         run: |
-          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
+          Import-Module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -VsInstallPath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
           cmake --preset "${{ matrix.preset }}"
           cmake --build --parallel --preset "${{ matrix.preset }}"
           cmake --install build --component "${{ startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || 'CPU' }}" --strip --parallel 8
@@ -244,7 +246,7 @@ jobs:
             dist\${{ matrix.os }}-${{ matrix.arch }}-app.exe
 
   windows-sign:
-    runs-on: windows
+    runs-on: windows-2022
     environment: release
     needs: [windows-depends, windows-build]
     steps:
@@ -322,6 +324,8 @@ 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_v11)      echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_v12)      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

@@ -36,7 +36,7 @@ jobs:
               | xargs python3 -c "import sys; from pathlib import Path; print(any(Path(x).match(glob) for x in sys.argv[1:] for glob in '$*'.split(' ')))"
           }
 
-          echo changed=$(changed 'llama/llama.cpp/**/*' 'ml/backend/ggml/ggml/**/*') | tee -a $GITHUB_OUTPUT
+          echo changed=$(changed 'llama/llama.cpp/**' 'ml/backend/ggml/ggml/**') | tee -a $GITHUB_OUTPUT
 
   linux:
     needs: [changes]
@@ -46,7 +46,7 @@ jobs:
         include:
           - preset: CPU
           - preset: CUDA
-            container: nvidia/cuda:12.8.1-devel-ubuntu22.04
+            container: nvidia/cuda:11.8.0-devel-ubuntu22.04
             flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
           - preset: ROCm
             container: rocm/dev-ubuntu-22.04:6.1.2
@@ -78,7 +78,7 @@ 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/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe
             flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
           - preset: ROCm
             install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
@@ -102,7 +102,7 @@ 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
+            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_11.3", "nvcc_11.3", "cublas_11.3", "cublas_dev_11.3")) -NoNewWindow -Wait
           }
 
           $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path
@@ -120,9 +120,6 @@ jobs:
           echo "$hipPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "CC=$hipPath\bin\clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
           echo "CXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "HIPCXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "HIP_PLATFORM=amd" | Out-File -FilePath $env:GITHUB_ENV -Append
-          echo "CMAKE_PREFIX_PATH=$hipPath" | Out-File -FilePath $env:GITHUB_ENV -Append
       - if: ${{ !cancelled() && steps.cache-install.outputs.cache-hit != 'true' }}
         uses: actions/cache/save@v4
         with:
@@ -136,8 +133,8 @@ jobs:
           path: ${{ github.workspace }}\.ccache
           key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}
       - 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'
+          Import-Module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -VsInstallPath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
           cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
           cmake --build --parallel --preset "${{ matrix.preset }}"
         env:

CMakeLists.txt

@@ -78,13 +78,14 @@ if(CMAKE_CUDA_COMPILER)
 
     find_package(CUDAToolkit)
     add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cuda)
+    set(OLLAMA_CUDA_INSTALL_DIR ${OLLAMA_INSTALL_DIR}/cuda_v${CUDAToolkit_VERSION_MAJOR})
     install(TARGETS ggml-cuda
         RUNTIME_DEPENDENCIES
             DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_LIBRARY_DIR}
             PRE_INCLUDE_REGEXES cublas cublasLt cudart
             PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
-        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
+        RUNTIME DESTINATION ${OLLAMA_CUDA_INSTALL_DIR} COMPONENT CUDA
+        LIBRARY DESTINATION ${OLLAMA_CUDA_INSTALL_DIR} COMPONENT CUDA
     )
 endif()
 
@@ -115,11 +116,7 @@ if(CMAKE_HIP_COMPILER)
 
         set(OLLAMA_HIP_INSTALL_DIR ${OLLAMA_INSTALL_DIR}/rocm)
         install(TARGETS ggml-hip
-            RUNTIME_DEPENDENCY_SET rocm
-            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
-        )
-        install(RUNTIME_DEPENDENCY_SET rocm
+            RUNTIME_DEPENDENCIES
                 DIRECTORIES ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR}
                 PRE_INCLUDE_REGEXES hipblas rocblas amdhip64 rocsolver amd_comgr hsa-runtime64 rocsparse tinfo rocprofiler-register drm drm_amdgpu numa elf
                 PRE_EXCLUDE_REGEXES ".*"

CMakePresets.json

@@ -17,12 +17,20 @@
       "name": "CUDA",
       "inherits": [ "Default" ]
     },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "cacheVariables": {
+        "CMAKE_CUDA_ARCHITECTURES": "50;52;53;60;61;70;75;80;86",
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets"
+      }
+    },
     {
       "name": "CUDA 12",
       "inherits": [ "CUDA" ],
       "cacheVariables": {
         "CMAKE_CUDA_ARCHITECTURES": "50;60;61;70;75;80;86;87;89;90;90a;120",
-        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets"
       }
     },
     {
@@ -50,7 +58,6 @@
       "name": "ROCm 6",
       "inherits": [ "ROCm" ],
       "cacheVariables": {
-        "CMAKE_HIP_FLAGS": "-parallel-jobs=4",
         "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
       }
     }
@@ -71,6 +78,11 @@
       "configurePreset": "CUDA",
       "targets": [ "ggml-cuda" ]
     },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "configurePreset": "CUDA 11"
+    },
     {
       "name": "CUDA 12",
       "inherits": [ "CUDA" ],

Dockerfile

@@ -7,13 +7,12 @@ ARG JETPACK5VERSION=r35.4.1
 ARG JETPACK6VERSION=r36.4.0
 ARG CMAKEVERSION=3.31.2
 
-# We require gcc v10 minimum.  v10.3 has regressions, so the rockylinux 8.5 AppStream has the latest compatible version
+# CUDA v11 requires gcc v10.  v10.3 has regressions, so the rockylinux 8.5 AppStream has the latest compatible version
 FROM --platform=linux/amd64 rocm/dev-almalinux-8:${ROCMVERSION}-complete AS base-amd64
 RUN yum install -y yum-utils \
     && yum-config-manager --add-repo https://dl.rockylinux.org/vault/rocky/8.5/AppStream/\$basearch/os/ \
     && rpm --import https://dl.rockylinux.org/pub/rocky/RPM-GPG-KEY-Rocky-8 \
     && dnf install -y yum-utils ccache gcc-toolset-10-gcc-10.2.1-8.2.el8 gcc-toolset-10-gcc-c++-10.2.1-8.2.el8 gcc-toolset-10-binutils-2.35-11.el8 \
-    && dnf install -y ccache \
     && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
 ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
 
@@ -39,6 +38,15 @@ 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=11.3
+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' \
+        && 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//./-}
@@ -90,15 +98,17 @@ 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/cuda_v11 /lib/ollama/cuda_v11
+COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
 
 FROM --platform=linux/arm64 scratch AS arm64
-COPY --from=cuda-12 dist/lib/ollama /lib/ollama/cuda_sbsa
-COPY --from=jetpack-5 dist/lib/ollama /lib/ollama/cuda_jetpack5
-COPY --from=jetpack-6 dist/lib/ollama /lib/ollama/cuda_jetpack6
+COPY --from=cuda-11 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_v11
+COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
+COPY --from=jetpack-5 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_jetpack5
+COPY --from=jetpack-6 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_jetpack6
 
 FROM scratch AS rocm
-COPY --from=rocm-6 dist/lib/ollama /lib/ollama
+COPY --from=rocm-6 dist/lib/ollama/rocm /lib/ollama/rocm
 
 FROM ${FLAVOR} AS archive
 COPY --from=cpu dist/lib/ollama /lib/ollama

README.md

@@ -407,9 +407,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Lumina](https://github.com/cushydigit/lumina.git) (A lightweight, minimal React.js frontend for interacting with Ollama servers)
 - [Tiny Notepad](https://pypi.org/project/tiny-notepad) (A lightweight, notepad-like interface to chat with ollama available on PyPI)
 - [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.) 
-- [GPTranslate](https://github.com/philberndt/GPTranslate) (A fast and lightweight, AI powered desktop translation application written with Rust and Tauri. Features real-time translation with OpenAI/Azure/Ollama.)
-- [ollama launcher](https://github.com/NGC13009/ollama-launcher) (A launcher for Ollama, aiming to provide users with convenient functions such as ollama server launching, management, or configuration.)
-- [ai-hub](https://github.com/Aj-Seven/ai-hub) (AI Hub supports multiple models via API keys and Chat support via Ollama API.)
 
 ### Cloud
 

benchmark/server_benchmark_test.go

@@ -0,0 +1,178 @@
+package benchmark
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// Command line flags
+var modelFlag string
+
+func init() {
+	flag.StringVar(&modelFlag, "m", "", "Name of the model to benchmark")
+	flag.Lookup("m").DefValue = "model"
+}
+
+// modelName returns the model name from flags, failing the test if not set
+func modelName(b *testing.B) string {
+	if modelFlag == "" {
+		b.Fatal("Error: -m flag is required for benchmark tests")
+	}
+	return modelFlag
+}
+
+type TestCase struct {
+	name      string
+	prompt    string
+	maxTokens int
+}
+
+// runGenerateBenchmark contains the common generate and metrics logic
+func runGenerateBenchmark(b *testing.B, ctx context.Context, client *api.Client, req *api.GenerateRequest) {
+	start := time.Now()
+	var ttft time.Duration
+	var metrics api.Metrics
+
+	err := client.Generate(ctx, req, func(resp api.GenerateResponse) error {
+		if ttft == 0 && resp.Response != "" {
+			ttft = time.Since(start)
+		}
+		if resp.Done {
+			metrics = resp.Metrics
+		}
+		return nil
+	})
+
+	// Report custom metrics as part of the benchmark results
+	b.ReportMetric(float64(ttft.Milliseconds()), "ttft_ms")
+	b.ReportMetric(float64(metrics.LoadDuration.Milliseconds()), "load_ms")
+
+	// Token throughput metrics
+	promptThroughput := float64(metrics.PromptEvalCount) / metrics.PromptEvalDuration.Seconds()
+	genThroughput := float64(metrics.EvalCount) / metrics.EvalDuration.Seconds()
+	b.ReportMetric(promptThroughput, "prompt_tok/s")
+	b.ReportMetric(genThroughput, "gen_tok/s")
+
+	// Token counts
+	b.ReportMetric(float64(metrics.PromptEvalCount), "prompt_tokens")
+	b.ReportMetric(float64(metrics.EvalCount), "gen_tokens")
+	if err != nil {
+		b.Fatal(err)
+	}
+}
+
+// BenchmarkColdStart runs benchmarks with model loading from cold state
+func BenchmarkColdStart(b *testing.B) {
+	client := setup(b)
+	tests := []TestCase{
+		{"short_prompt", "Write a long story", 100},
+		{"medium_prompt", "Write a detailed economic analysis", 500},
+		{"long_prompt", "Write a comprehensive AI research paper", 1000},
+	}
+	m := modelName(b)
+
+	for _, tt := range tests {
+		b.Run(fmt.Sprintf("%s/cold/%s", m, tt.name), func(b *testing.B) {
+			ctx := b.Context()
+
+			// Set number of tokens as our throughput metric
+			b.SetBytes(int64(tt.maxTokens))
+
+			for b.Loop() {
+				b.StopTimer()
+				// Ensure model is unloaded before each iteration
+				unload(client, m, b)
+				b.StartTimer()
+
+				req := &api.GenerateRequest{
+					Model:   m,
+					Prompt:  tt.prompt,
+					Options: map[string]any{"num_predict": tt.maxTokens, "temperature": 0.1},
+				}
+
+				runGenerateBenchmark(b, ctx, client, req)
+			}
+		})
+	}
+}
+
+// BenchmarkWarmStart runs benchmarks with pre-loaded model
+func BenchmarkWarmStart(b *testing.B) {
+	client := setup(b)
+	tests := []TestCase{
+		{"short_prompt", "Write a long story", 100},
+		{"medium_prompt", "Write a detailed economic analysis", 500},
+		{"long_prompt", "Write a comprehensive AI research paper", 1000},
+	}
+	m := modelName(b)
+
+	for _, tt := range tests {
+		b.Run(fmt.Sprintf("%s/warm/%s", m, tt.name), func(b *testing.B) {
+			ctx := b.Context()
+
+			// Pre-warm the model
+			warmup(client, m, tt.prompt, b)
+
+			// Set number of tokens as our throughput metric
+			b.SetBytes(int64(tt.maxTokens))
+
+			for b.Loop() {
+				req := &api.GenerateRequest{
+					Model:   m,
+					Prompt:  tt.prompt,
+					Options: map[string]any{"num_predict": tt.maxTokens, "temperature": 0.1},
+				}
+
+				runGenerateBenchmark(b, ctx, client, req)
+			}
+		})
+	}
+}
+
+// setup verifies server and model availability
+func setup(b *testing.B) *api.Client {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		b.Fatal(err)
+	}
+	if _, err := client.Show(b.Context(), &api.ShowRequest{Model: modelName(b)}); err != nil {
+		b.Fatalf("Model unavailable: %v", err)
+	}
+
+	return client
+}
+
+// warmup ensures the model is loaded and warmed up
+func warmup(client *api.Client, model string, prompt string, b *testing.B) {
+	for range 3 {
+		err := client.Generate(
+			context.Background(),
+			&api.GenerateRequest{
+				Model:   model,
+				Prompt:  prompt,
+				Options: map[string]any{"num_predict": 50, "temperature": 0.1},
+			},
+			func(api.GenerateResponse) error { return nil },
+		)
+		if err != nil {
+			b.Logf("Error during model warm-up: %v", err)
+		}
+	}
+}
+
+// unload forces model unloading using KeepAlive: 0 parameter
+func unload(client *api.Client, model string, b *testing.B) {
+	req := &api.GenerateRequest{
+		Model:     model,
+		KeepAlive: &api.Duration{Duration: 0},
+	}
+	if err := client.Generate(context.Background(), req, func(api.GenerateResponse) error { return nil }); err != nil {
+		b.Logf("Unload error: %v", err)
+	}
+	time.Sleep(1 * time.Second)
+}

convert/convert_mixtral.go

@@ -2,6 +2,9 @@ package convert
 
 import (
 	"fmt"
+	"io"
+	"slices"
+	"strings"
 
 	"github.com/ollama/ollama/fs/ggml"
 )
@@ -27,38 +30,65 @@ func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
 }
 
 func (p *mixtralModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	merges := make([]merge, 0, p.NumHiddenLayers*6)
-	for i := range p.NumHiddenLayers {
-		merges = append(merges, merge{
-			fmt.Sprintf("blk.%d.*.w1.weight", i),
-			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
-		}, merge{
-			fmt.Sprintf("blk.%d.*.w1.bias", i),
-			fmt.Sprintf("blk.%d.ffn_gate_exps.bias", i),
-		}, merge{
-			fmt.Sprintf("blk.%d.*.w2.weight", i),
-			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
-		}, merge{
-			fmt.Sprintf("blk.%d.*.w2.bias", i),
-			fmt.Sprintf("blk.%d.ffn_up_exps.bias", i),
-		}, merge{
-			fmt.Sprintf("blk.%d.*.w3.weight", i),
-			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
-		}, merge{
-			fmt.Sprintf("blk.%d.*.w3.bias", i),
-			fmt.Sprintf("blk.%d.ffn_down_exps.bias", i),
+	oldnew := []string{
+		"model.layers", "blk",
+		"w1", "ffn_gate_exps",
+		"w2", "ffn_down_exps",
+		"w3", "ffn_up_exps",
+	}
+
+	for i := range p.NumLocalExperts {
+		oldnew = append(oldnew, fmt.Sprintf(".block_sparse_moe.experts.%d.", i), ".")
+	}
+
+	// group experts of the same layer (model.layers.%d) and type (w[123]) into a single tensor
+	namer := strings.NewReplacer(oldnew...)
+	experts := make(map[string]experts)
+
+	// merge experts into a single tensor while removing them from ts
+	ts = slices.DeleteFunc(ts, func(t Tensor) bool {
+		if !strings.Contains(t.Name(), ".block_sparse_moe.experts.") {
+			return false
+		}
+
+		name := namer.Replace(t.Name())
+		experts[name] = append(experts[name], t)
+		return true
+	})
+
+	var out []*ggml.Tensor
+	for n, e := range experts {
+		// TODO(mxyng): sanity check experts
+		out = append(out, &ggml.Tensor{
+			Name:     n,
+			Kind:     e[0].Kind(),
+			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),
+			WriterTo: e,
 		})
 	}
 
-	out, ts := mergeTensors(ts, merges...)
 	return append(out, p.llamaModel.Tensors(ts)...)
 }
 
 func (p *mixtralModel) Replacements() []string {
 	return append(
 		p.llamaModel.Replacements(),
-		"model.layers", "blk",
 		"block_sparse_moe.gate", "ffn_gate_inp",
-		"block_sparse_moe.experts.", ".",
 	)
 }
+
+type experts []Tensor
+
+func (e experts) WriteTo(w io.Writer) (int64, error) {
+	// TODO(mxyng): experts _should_ be numerically sorted by expert but this should check
+	for _, t := range e {
+		// the canonical merged experts tensor stacks all experts along a new, 0 axis,
+		// e.g. `tensor.Stack(0, e[0], e[1:]...)`, which requires allocating temporary buffers
+		// this accomplishes the same thing by writing each expert tensor in sequence
+		if _, err := t.WriteTo(w); err != nil {
+			return 0, err
+		}
+	}
+
+	return 0, nil
+}

convert/tensor.go

@@ -2,9 +2,7 @@ package convert
 
 import (
 	"cmp"
-	"io"
 	"iter"
-	"path"
 	"slices"
 	"strings"
 
@@ -76,54 +74,3 @@ func splitDim(t Tensor, dim int, splits ...split) iter.Seq[*ggml.Tensor] {
 		}
 	}
 }
-
-type merge struct {
-	pattern, name string
-}
-
-// mergeTensors merges tensors that match a given pattern into a single tensor.
-func mergeTensors(unmatched []Tensor, merges ...merge) (out []*ggml.Tensor, _ []Tensor) {
-	var matched []Tensor
-	for i := range merges {
-		matched, unmatched = slicesSplitFunc(unmatched, func(t Tensor) bool {
-			matched, _ := path.Match(merges[i].pattern, t.Name())
-			return matched
-		})
-
-		if len(matched) > 0 {
-			out = append(out, &ggml.Tensor{
-				Name:     merges[i].name,
-				Kind:     matched[0].Kind(),
-				Shape:    append([]uint64{uint64(len(matched))}, matched[0].Shape()...),
-				WriterTo: mergeGroup(matched),
-			})
-		}
-	}
-
-	return out, unmatched
-}
-
-// slicesSplitFunc splits a slice into two slices based on a predicate function.
-func slicesSplitFunc[S ~[]E, E comparable](s S, fn func(e E) bool) (matched, unmatched S) {
-	for _, e := range s {
-		if fn(e) {
-			matched = append(matched, e)
-		} else {
-			unmatched = append(unmatched, e)
-		}
-	}
-
-	return matched, unmatched
-}
-
-type mergeGroup []Tensor
-
-func (g mergeGroup) WriteTo(w io.Writer) (int64, error) {
-	for _, t := range g {
-		if _, err := t.WriteTo(w); err != nil {
-			return 0, err
-		}
-	}
-
-	return 0, nil
-}

convert/tensor_test.go

@@ -9,8 +9,6 @@ import (
 	"strings"
 	"testing"
 
-	"github.com/google/go-cmp/cmp"
-	"github.com/ollama/ollama/fs/ggml"
 	"github.com/pdevine/tensor"
 )
 
@@ -304,99 +302,3 @@ func TestSplitDim(t *testing.T) {
 		}
 	})
 }
-
-func TestMerge(t *testing.T) {
-	unmatched := []Tensor{
-		&fakeTensor{
-			name:  "a.0.b",
-			shape: []uint64{5, 2},
-			data:  []float32{10, 11, 12, 13, 14, 15, 16, 17, 18, 19},
-		},
-		&fakeTensor{
-			name:  "a.1.b",
-			shape: []uint64{5, 2},
-			data:  []float32{20, 21, 22, 23, 24, 25, 26, 27, 28, 29},
-		},
-		&fakeTensor{
-			name:  "c.0.d",
-			shape: []uint64{5, 2},
-			data:  []float32{30, 31, 32, 33, 34, 35, 36, 37, 38, 39},
-		},
-		&fakeTensor{
-			name:  "c.1.d",
-			shape: []uint64{5, 2},
-			data:  []float32{40, 41, 42, 43, 44, 45, 46, 47, 48, 49},
-		},
-		&fakeTensor{
-			name:  "e.0.f",
-			shape: []uint64{5, 2},
-			data:  []float32{50, 51, 52, 53, 54, 55, 56, 57, 58, 59},
-		},
-	}
-
-	checkMatched := func(t *testing.T, n int, matched []*ggml.Tensor) {
-		for i := range n {
-			got := matched[i]
-			if diff := cmp.Diff([]uint64{2, 5, 2}, got.Shape); diff != "" {
-				t.Errorf("unexpected (-want +got):\n%s", diff)
-			}
-
-			var b bytes.Buffer
-			if _, err := got.WriteTo(&b); err != nil {
-				t.Fatal(err)
-			}
-
-			f32s := make([]float32, 20)
-			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
-				t.Fatal(err)
-			}
-
-			offset := 10 + (i * 20)
-			want := make([]float32, 20)
-			for j := range 20 {
-				want[j] = float32(offset + j)
-			}
-
-			if diff := cmp.Diff(want, f32s); diff != "" {
-				t.Errorf("unexpected data (-want +got):\n%s", diff)
-			}
-		}
-	}
-
-	t.Run("single merge", func(t *testing.T) {
-		matched, unmatched := mergeTensors(unmatched, merge{"a.*.b", "a.b"})
-		if len(unmatched) != 3 {
-			t.Error("expected 3 remaining tensors, got", len(unmatched))
-		}
-
-		if len(matched) != 1 {
-			t.Error("expected 1 merged tensor, got", len(matched))
-		}
-
-		checkMatched(t, 1, matched)
-	})
-
-	t.Run("multiple merges", func(t *testing.T) {
-		matched, unmatched := mergeTensors(unmatched, merge{"a.*.b", "a.b"}, merge{"c.*.d", "c.d"})
-		if len(unmatched) != 1 {
-			t.Error("expected 1 remaining tensors, got", len(unmatched))
-		}
-
-		if len(matched) != 2 {
-			t.Error("expected 2 merged tensor, got", len(matched))
-		}
-
-		checkMatched(t, 2, matched)
-	})
-
-	t.Run("no match", func(t *testing.T) {
-		matched, unmatched := mergeTensors(unmatched, merge{"x.*.y", "x.y"})
-		if len(unmatched) != 5 {
-			t.Error("expected 5 remaining tensors, got", len(unmatched))
-		}
-
-		if len(matched) != 0 {
-			t.Error("expected no merged tensors, got", len(matched))
-		}
-	})
-}

discover/cuda_common.go

@@ -3,7 +3,6 @@
 package discover
 
 import (
-	"fmt"
 	"log/slog"
 	"os"
 	"regexp"
@@ -56,13 +55,10 @@ 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"
 	}
 	return "v12"

discover/path.go

@@ -12,7 +12,7 @@ import (
 // '../lib/ollama' on Linux and the executable's directory on macOS
 // note: distribution builds, additional GPU-specific libraries are
 // found in subdirectories of the returned path, such as
-// 'cuda_v12', 'rocm', etc.
+// 'cuda_v11', 'cuda_v12', 'rocm', etc.
 var LibOllamaPath string = func() string {
 	exe, err := os.Executable()
 	if err != nil {

docs/benchmark.md

@@ -0,0 +1,59 @@
+# Benchmark
+
+Go benchmark tests that measure end-to-end performance of a running Ollama server. Run these tests to evaluate model inference performance on your hardware and measure the impact of code changes.
+
+## When to use
+
+Run these benchmarks when:
+- Making changes to the model inference engine
+- Modifying model loading/unloading logic
+- Changing prompt processing or token generation code
+- Implementing a new model architecture
+- Testing performance across different hardware setups
+
+## Prerequisites
+- Ollama server running locally with `ollama serve` on `127.0.0.1:11434`
+## Usage and Examples
+
+>[!NOTE]
+>All commands must be run from the root directory of the Ollama project.
+
+Basic syntax:
+```bash
+go test -bench=. ./benchmark/... -m $MODEL_NAME
+```
+
+Required flags:
+- `-bench=.`: Run all benchmarks
+- `-m`: Model name to benchmark
+
+Optional flags:
+- `-count N`: Number of times to run the benchmark (useful for statistical analysis)
+- `-timeout T`: Maximum time for the benchmark to run (e.g. "10m" for 10 minutes)
+
+Common usage patterns:
+
+Single benchmark run with a model specified:
+```bash
+go test -bench=. ./benchmark/... -m llama3.3
+```
+
+## Output metrics
+
+The benchmark reports several key metrics:
+
+- `gen_tok/s`: Generated tokens per second
+- `prompt_tok/s`: Prompt processing tokens per second
+- `ttft_ms`: Time to first token in milliseconds
+- `load_ms`: Model load time in milliseconds
+- `gen_tokens`: Total tokens generated
+- `prompt_tokens`: Total prompt tokens processed
+
+Each benchmark runs two scenarios:
+- Cold start: Model is loaded from disk for each test
+- Warm start: Model is pre-loaded in memory
+
+Three prompt lengths are tested for each scenario:
+- Short prompt (100 tokens)
+- Medium prompt (500 tokens)
+- Long prompt (1000 tokens)

docs/gpu.md

@@ -1,6 +1,6 @@
 # GPU
 ## Nvidia
-Ollama supports Nvidia GPUs with compute capability 5.0+ and driver version 531 and newer.
+Ollama supports Nvidia GPUs with compute capability 5.0+.
 
 Check your compute compatibility to see if your card is supported:
 [https://developer.nvidia.com/cuda-gpus](https://developer.nvidia.com/cuda-gpus)

docs/troubleshooting.md

@@ -43,7 +43,7 @@ Ollama includes multiple LLM libraries compiled for different GPUs and CPU vecto
 In the server log, you will see a message that looks something like this (varies from release to release):
 
 ```
-Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v12 rocm_v5]
+Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v11 rocm_v5]
 ```
 
 **Experimental LLM Library Override**

fs/ggml/gguf.go

@@ -527,17 +527,23 @@ func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 		return err
 	}
 
-	for _, key := range slices.Sorted(maps.Keys(kv)) {
+	keys := slices.Collect(maps.Keys(kv))
+	slices.Sort(keys)
+
+	for _, key := range keys {
 		if err := ggufWriteKV(f, key, kv[key]); err != nil {
 			return err
 		}
 	}
 
 	slices.SortStableFunc(ts, func(a, b *Tensor) int {
-		if i, j := a.block(), b.block(); i > 0 && j > 0 {
+		if i, j := a.block(), b.block(); i < 0 && j > 0 {
+			return 1
+		} else if i > 0 && j < 0 {
+			return -1
+		} else {
 			return cmp.Compare(i, j)
 		}
-		return cmp.Compare(a.Name, b.Name)
 	})
 
 	var s uint64

fs/ggml/gguf_test.go

@@ -2,82 +2,62 @@ package ggml
 
 import (
 	"bytes"
-	"math/rand/v2"
 	"os"
-	"strings"
+	"slices"
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
 )
 
 func TestWriteGGUF(t *testing.T) {
-	r := rand.New(rand.NewPCG(0, 0))
-	for range 8 {
-		t.Run("shuffle", func(t *testing.T) {
-			t.Parallel()
+	w, err := os.CreateTemp(t.TempDir(), "*.bin")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer w.Close()
 
-			ts := []*Tensor{
-				{Name: "token_embd.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.0.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.1.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.2.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.3.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.4.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "blk.5.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-				{Name: "output_norm.weight", Shape: []uint64{3, 2}, WriterTo: bytes.NewBuffer(make([]byte, 3*2))},
-				{Name: "output.weight", Shape: []uint64{3, 2}, WriterTo: bytes.NewBuffer(make([]byte, 3*2))},
-			}
+	if err := WriteGGUF(w, KV{
+		"general.alignment": uint32(16),
+	}, []*Tensor{
+		{Name: "test.0", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.1", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.2", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.3", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.4", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.5", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+	}); err != nil {
+		t.Fatal(err)
+	}
 
-			r.Shuffle(len(ts), func(i, j int) {
-				ts[i], ts[j] = ts[j], ts[i]
-			})
+	r, err := os.Open(w.Name())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer r.Close()
 
-			w, err := os.CreateTemp(t.TempDir(), strings.ReplaceAll(t.Name(), "/", "_")+"*.bin")
-			if err != nil {
-				t.Fatal(err)
-			}
-			defer w.Close()
+	ff, err := Decode(r, 0)
+	if err != nil {
+		t.Fatal(err)
+	}
 
-			if err := WriteGGUF(w, KV{
-				"general.alignment": uint32(16),
-			}, ts); err != nil {
-				t.Fatal(err)
-			}
+	if diff := cmp.Diff(ff.KV(), KV{
+		"general.alignment":       uint32(16),
+		"general.parameter_count": uint64(36),
+	}); diff != "" {
+		t.Errorf("Mismatch (-want +got):\n%s", diff)
+	}
 
-			r, err := os.Open(w.Name())
-			if err != nil {
-				t.Fatal(err)
-			}
-			defer r.Close()
-
-			ff, err := Decode(r, 0)
-			if err != nil {
-				t.Fatal(err)
-			}
-
-			if diff := cmp.Diff(KV{
-				"general.alignment":       uint32(16),
-				"general.parameter_count": uint64(54),
-			}, ff.KV()); diff != "" {
-				t.Errorf("Mismatch (-want +got):\n%s", diff)
-			}
-
-			if diff := cmp.Diff(Tensors{
-				Offset: 608,
-				items: []*Tensor{
-					{Name: "blk.0.attn_norm.weight", Offset: 0, Shape: []uint64{2, 3}},
-					{Name: "blk.1.attn_norm.weight", Offset: 32, Shape: []uint64{2, 3}},
-					{Name: "blk.2.attn_norm.weight", Offset: 64, Shape: []uint64{2, 3}},
-					{Name: "blk.3.attn_norm.weight", Offset: 96, Shape: []uint64{2, 3}},
-					{Name: "blk.4.attn_norm.weight", Offset: 128, Shape: []uint64{2, 3}},
-					{Name: "blk.5.attn_norm.weight", Offset: 160, Shape: []uint64{2, 3}},
-					{Name: "output.weight", Offset: 192, Shape: []uint64{3, 2}},
-					{Name: "output_norm.weight", Offset: 224, Shape: []uint64{3, 2}},
-					{Name: "token_embd.weight", Offset: 256, Shape: []uint64{2, 3}},
-				},
-			}, ff.Tensors(), cmp.AllowUnexported(Tensors{})); diff != "" {
-				t.Errorf("Mismatch (-want +got):\n%s", diff)
-			}
-		})
+	if diff := cmp.Diff(ff.Tensors(), Tensors{
+		Offset: 336,
+		items: []*Tensor{
+			{Name: "test.0", Offset: 0, Shape: []uint64{2, 3}},
+			{Name: "test.1", Offset: 32, Shape: []uint64{2, 3}},
+			{Name: "test.2", Offset: 64, Shape: []uint64{2, 3}},
+			{Name: "test.3", Offset: 96, Shape: []uint64{2, 3}},
+			{Name: "test.4", Offset: 128, Shape: []uint64{2, 3}},
+			{Name: "test.5", Offset: 160, Shape: []uint64{2, 3}},
+		},
+	}, cmp.AllowUnexported(Tensors{})); diff != "" {
+		t.Errorf("Mismatch (-want +got):\n%s", diff)
 	}
 }

fs/gguf/gguf.go

@@ -65,7 +65,7 @@ func Open(path string) (f *File, err error) {
 		return nil, err
 	}
 
-	if f.Version < 2 {
+	if f.Version != 3 {
 		return nil, fmt.Errorf("%w version %v", ErrUnsupported, f.Version)
 	}
 

integration/model_arch_test.go

@@ -45,8 +45,6 @@ var (
 		"qwen2.5-coder:latest",
 		"qwen:latest",
 		"solar-pro:latest",
-		"codellama:latest",
-		"nous-hermes:latest",
 	}
 )
 

llama/patches/0018-temporary-prevent-rocm-cuda-mixed-loading.patch

@@ -1,32 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Daniel Hiltgen <daniel@ollama.com>
-Date: Sun, 22 Jun 2025 09:22:05 -0700
-Subject: [PATCH] temporary prevent rocm+cuda mixed loading
-
----
- ggml/src/ggml-backend-reg.cpp | 12 ++++++++++--
- 1 file changed, 10 insertions(+), 2 deletions(-)
-
-diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
-index 4e67d243..8f49f084 100644
---- a/ggml/src/ggml-backend-reg.cpp
-+++ b/ggml/src/ggml-backend-reg.cpp
-@@ -573,8 +573,16 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
- 
-     ggml_backend_load_best("blas", silent, dir_path);
-     ggml_backend_load_best("cann", silent, dir_path);
--    ggml_backend_load_best("cuda", silent, dir_path);
--    ggml_backend_load_best("hip", silent, dir_path);
-+
-+    // Avoid mixed hip+cuda configurations
-+    const char * hip_devices = std::getenv("HIP_VISIBLE_DEVICES");
-+    const char * rocr_devices = std::getenv("ROCR_VISIBLE_DEVICES"); 
-+    if (!hip_devices && !rocr_devices) {
-+        ggml_backend_load_best("cuda", silent, dir_path);
-+    } else {
-+        ggml_backend_load_best("hip", silent, dir_path);
-+    }
-+    
-     ggml_backend_load_best("kompute", silent, dir_path);
-     ggml_backend_load_best("metal", silent, dir_path);
-     ggml_backend_load_best("rpc", silent, dir_path);

llm/server.go

@@ -22,6 +22,7 @@ import (
 	"strings"
 	"sync"
 	"time"
+	"unicode/utf8"
 
 	"golang.org/x/sync/semaphore"
 
@@ -139,13 +140,6 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		gpus = discover.GetCPUInfo()
 	}
 
-	// Verify the requested context size is <= the model training size
-	trainCtx := f.KV().ContextLength()
-	if opts.NumCtx/numParallel > int(trainCtx) && trainCtx > 0 {
-		slog.Warn("requested context size too large for model", "num_ctx", opts.NumCtx, "num_parallel", numParallel, "n_ctx_train", trainCtx)
-		opts.NumCtx = int(trainCtx) * numParallel
-	}
-
 	estimate := EstimateGPULayers(gpus, f, projectors, opts, numParallel)
 	if len(gpus) > 1 || gpus[0].Library != "cpu" {
 		switch {
@@ -318,7 +312,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		params = append(params, "--mmproj", projectors[0])
 	}
 
-	// iterate through compatible GPU libraries such as 'cuda_v12', 'rocm', etc.
+	// iterate through compatible GPU libraries such as 'cuda_v12', 'cuda_v11', 'rocm', etc.
 	// adding each library's respective path to the LD_LIBRARY_PATH, until finally running
 	// without any LD_LIBRARY_PATH flags
 	for {
@@ -732,10 +726,68 @@ type CompletionResponse struct {
 	EvalDuration       time.Duration `json:"eval_duration"`
 }
 
+// unicodeBufferHandler wraps a completion response callback to handle partial UTF-8 sequences.
+// This function creates a stateful closure that is NOT safe for concurrent use.
+// Each completion request should create its own handler instance.
+func unicodeBufferHandler(fn func(CompletionResponse)) func(CompletionResponse) {
+	var pendingUTF8 string
+
+	return func(resp CompletionResponse) {
+		if resp.Content == "" && !resp.Done {
+			// No content to process, just pass through
+			fn(resp)
+			return
+		}
+
+		// Combine any pending UTF-8 with current content
+		combinedContent := pendingUTF8 + resp.Content
+		pendingUTF8 = ""
+
+		// Check if combined content is valid UTF-8
+		if utf8.ValidString(combinedContent) {
+			// Valid UTF-8, send it
+			resp.Content = combinedContent
+			fn(resp)
+		} else {
+			// Invalid UTF-8
+			if resp.Done {
+				// This is the final response, trim incomplete UTF-8
+				trimmedContent := combinedContent
+				for !utf8.ValidString(trimmedContent) && len(trimmedContent) > 0 {
+					trimmedContent = trimmedContent[:len(trimmedContent)-1]
+				}
+				resp.Content = trimmedContent
+				fn(resp)
+			} else {
+				// Not final response, split valid and invalid parts
+				validPrefix := combinedContent
+				for !utf8.ValidString(validPrefix) && len(validPrefix) > 0 {
+					validPrefix = validPrefix[:len(validPrefix)-1]
+				}
+
+				if len(validPrefix) > 0 {
+					// Send valid prefix
+					resp.Content = validPrefix
+					fn(resp)
+					// Buffer the remainder
+					pendingUTF8 = combinedContent[len(validPrefix):]
+				} else {
+					// No valid prefix, buffer everything
+					pendingUTF8 = combinedContent
+					// Don't send this response
+				}
+			}
+		}
+	}
+}
+
 func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error {
 	slog.Debug("completion request", "images", len(req.Images), "prompt", len(req.Prompt), "format", string(req.Format))
 	slog.Log(ctx, logutil.LevelTrace, "completion request", "prompt", req.Prompt)
 
+	// Wrap the callback with unicode buffer handling
+	unicodeFn := unicodeBufferHandler(fn)
+
 	if len(req.Format) > 0 {
 		switch string(req.Format) {
 		case `null`, `""`:
@@ -861,13 +913,13 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 			}
 
 			if c.Content != "" {
-				fn(CompletionResponse{
+				unicodeFn(CompletionResponse{
 					Content: c.Content,
 				})
 			}
 
 			if c.Done {
-				fn(c)
+				unicodeFn(c)
 				return nil
 			}
 		}

llm/server_test.go

@@ -70,3 +70,152 @@ func TestLLMServerCompletionFormat(t *testing.T) {
 	}, nil)
 	checkValid(err)
 }
+
+func TestUnicodeBufferHandler(t *testing.T) {
+	tests := []struct {
+		name              string
+		inputResponses    []CompletionResponse
+		expectedResponses []CompletionResponse
+		description       string
+	}{
+		{
+			name: "complete_unicode",
+			inputResponses: []CompletionResponse{
+				{Content: "Hello", Done: false},
+				{Content: " world", Done: false},
+				{Content: "!", Done: true},
+			},
+			expectedResponses: []CompletionResponse{
+				{Content: "Hello", Done: false},
+				{Content: " world", Done: false},
+				{Content: "!", Done: true},
+			},
+			description: "All responses with valid unicode should pass through unchanged",
+		},
+		{
+			name: "incomplete_unicode_at_end_with_done",
+			inputResponses: []CompletionResponse{
+				{Content: "Hello", Done: false},
+				{Content: string([]byte{0xF0, 0x9F}), Done: true}, // Incomplete emoji with Done=true
+			},
+			expectedResponses: []CompletionResponse{
+				{Content: "Hello", Done: false},
+				{Content: "", Done: true}, // Content is trimmed but response is still sent with Done=true
+			},
+			description: "When Done=true, incomplete Unicode at the end should be trimmed",
+		},
+		{
+			name: "split_unicode_across_responses",
+			inputResponses: []CompletionResponse{
+				{Content: "Hello " + string([]byte{0xF0, 0x9F}), Done: false}, // First part of 😀
+				{Content: string([]byte{0x98, 0x80}) + " world!", Done: true}, // Second part of 😀 and more text
+			},
+			expectedResponses: []CompletionResponse{
+				{Content: "Hello ", Done: false},  // Incomplete Unicode trimmed
+				{Content: "😀 world!", Done: true}, // Complete emoji in second response
+			},
+			description: "Unicode split across responses should be handled correctly",
+		},
+		{
+			name: "incomplete_unicode_buffered",
+			inputResponses: []CompletionResponse{
+				{Content: "Test " + string([]byte{0xF0, 0x9F}), Done: false}, // Incomplete emoji
+				{Content: string([]byte{0x98, 0x80}), Done: false},           // Complete the emoji
+				{Content: " done", Done: true},
+			},
+			expectedResponses: []CompletionResponse{
+				{Content: "Test ", Done: false}, // First part without incomplete unicode
+				{Content: "😀", Done: false},     // Complete emoji
+				{Content: " done", Done: true},
+			},
+			description: "Incomplete unicode should be buffered and combined with next response",
+		},
+		{
+			name: "empty_response_with_done",
+			inputResponses: []CompletionResponse{
+				{Content: "Complete response", Done: false},
+				{Content: "", Done: true}, // Empty response with Done=true
+			},
+			expectedResponses: []CompletionResponse{
+				{Content: "Complete response", Done: false},
+				{Content: "", Done: true}, // Should still be sent because Done=true
+			},
+			description: "Empty final response with Done=true should still be sent",
+		},
+		{
+			name: "done_reason_preserved",
+			inputResponses: []CompletionResponse{
+				{Content: "Response", Done: false},
+				{Content: " complete", Done: true, DoneReason: DoneReasonStop},
+			},
+			expectedResponses: []CompletionResponse{
+				{Content: "Response", Done: false},
+				{Content: " complete", Done: true, DoneReason: DoneReasonStop},
+			},
+			description: "DoneReason should be preserved in the final response",
+		},
+		{
+			name: "only_incomplete_unicode_not_done",
+			inputResponses: []CompletionResponse{
+				{Content: string([]byte{0xF0, 0x9F}), Done: false}, // Only incomplete unicode
+			},
+			expectedResponses: []CompletionResponse{
+				// No response expected - should be buffered
+			},
+			description: "Response with only incomplete unicode should be buffered if not done",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var actualResponses []CompletionResponse
+
+			// Create a callback that collects responses
+			callback := func(resp CompletionResponse) {
+				actualResponses = append(actualResponses, resp)
+			}
+
+			// Create the unicode buffer handler
+			handler := unicodeBufferHandler(callback)
+
+			// Send all input responses through the handler
+			for _, resp := range tt.inputResponses {
+				handler(resp)
+			}
+
+			// Verify the number of responses
+			if len(actualResponses) != len(tt.expectedResponses) {
+				t.Fatalf("%s: got %d responses, want %d responses",
+					tt.description, len(actualResponses), len(tt.expectedResponses))
+			}
+
+			// Verify each response matches the expected one
+			for i, expected := range tt.expectedResponses {
+				if i >= len(actualResponses) {
+					t.Fatalf("%s: missing response at index %d", tt.description, i)
+					continue
+				}
+
+				actual := actualResponses[i]
+
+				// Verify content
+				if actual.Content != expected.Content {
+					t.Errorf("%s: response[%d].Content = %q, want %q",
+						tt.description, i, actual.Content, expected.Content)
+				}
+
+				// Verify Done flag
+				if actual.Done != expected.Done {
+					t.Errorf("%s: response[%d].Done = %v, want %v",
+						tt.description, i, actual.Done, expected.Done)
+				}
+
+				// Verify DoneReason if specified
+				if actual.DoneReason != expected.DoneReason {
+					t.Errorf("%s: response[%d].DoneReason = %v, want %v",
+						tt.description, i, actual.DoneReason, expected.DoneReason)
+				}
+			}
+		})
+	}
+}

ml/backend/ggml/ggml.go

@@ -602,9 +602,7 @@ func (c *Context) Forward(tensors ...ml.Tensor) ml.Context {
 }
 
 func (c *Context) Compute(tensors ...ml.Tensor) {
-	if status := C.ggml_backend_sched_graph_compute_async(c.b.sched, c.graph); status != C.GGML_STATUS_SUCCESS {
-		panic(fmt.Errorf("error computing ggml graph: %v", status))
-	}
+	C.ggml_backend_sched_graph_compute_async(c.b.sched, c.graph)
 	C.ggml_backend_sched_reset(c.b.sched)
 
 	needSync := true

ml/backend/ggml/ggml/src/ggml-backend-reg.cpp

@@ -573,16 +573,8 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
 
     ggml_backend_load_best("blas", silent, dir_path);
     ggml_backend_load_best("cann", silent, dir_path);
-
-    // Avoid mixed hip+cuda configurations
-    const char * hip_devices = std::getenv("HIP_VISIBLE_DEVICES");
-    const char * rocr_devices = std::getenv("ROCR_VISIBLE_DEVICES"); 
-    if (!hip_devices && !rocr_devices) {
-        ggml_backend_load_best("cuda", silent, dir_path);
-    } else {
-        ggml_backend_load_best("hip", silent, dir_path);
-    }
-    
+    ggml_backend_load_best("cuda", silent, dir_path);
+    ggml_backend_load_best("hip", silent, dir_path);
     ggml_backend_load_best("kompute", silent, dir_path);
     ggml_backend_load_best("metal", silent, dir_path);
     ggml_backend_load_best("rpc", silent, dir_path);

model/vocabulary.go

@@ -87,7 +87,7 @@ func (v *Vocabulary) Decode(id int32) string {
 func (v *Vocabulary) SpecialVocabulary() []string {
 	v.specialOnce.Do(func() {
 		for i := range v.Values {
-			if v.Types[i] == TOKEN_TYPE_CONTROL || v.Types[i] == TOKEN_TYPE_USER_DEFINED {
+			if v.Types[i] == TOKEN_TYPE_CONTROL {
 				v.special = append(v.special, v.Values[i])
 			}
 		}

model/vocabulary_test.go

@@ -1,16 +0,0 @@
-package model
-
-import "testing"
-
-func TestVocabulary_SpecialVocabulary(t *testing.T) {
-	vocab := &Vocabulary{
-		Values: []string{"<|startoftext|>", "<|endoftext|>", "<|tool_call_start|>", "<|tool_call_end|>", "hi"},
-		Types:  []int32{TOKEN_TYPE_CONTROL, TOKEN_TYPE_CONTROL, TOKEN_TYPE_USER_DEFINED, TOKEN_TYPE_USER_DEFINED, TOKEN_TYPE_NORMAL},
-	}
-
-	specialVocab := vocab.SpecialVocabulary()
-
-	if len(specialVocab) != 4 {
-		t.Errorf("expected 4 special tokens, got %d", len(specialVocab))
-	}
-}

runner/common/stop.go

@@ -2,6 +2,8 @@ package common
 
 import (
 	"strings"
+
+	"github.com/ollama/ollama/llm"
 )
 
 func FindStop(sequence string, stops []string) (bool, string) {
@@ -29,68 +31,41 @@ func ContainsStopSuffix(sequence string, stops []string) bool {
 // truncateStop removes the provided stop string from pieces,
 // returning the partial pieces with stop removed, including truncating
 // the last piece if required (and signalling if this was the case)
-func TruncateStop(pieces []string, stop string) ([]string, bool) {
-	joined := strings.Join(pieces, "")
-
-	index := strings.Index(joined, stop)
-	if index == -1 {
-		return pieces, false
+func TruncateStop(resps []llm.CompletionResponse, stop string) ([]llm.CompletionResponse, bool) {
+	var sequence string
+	for _, resp := range resps {
+		sequence += resp.Content
 	}
 
-	joined = joined[:index]
-
-	// Split truncated string back into pieces of original lengths
-	lengths := make([]int, len(pieces))
-	for i, piece := range pieces {
-		lengths[i] = len(piece)
+	idx := strings.Index(sequence, stop)
+	if idx < 0 {
+		return resps, false
 	}
 
-	var result []string
-	tokenTruncated := false
-	start := 0
-	for _, length := range lengths {
-		if start >= len(joined) {
+	truncated := sequence[:idx]
+	if len(truncated) == 0 {
+		return nil, true
+	}
+
+	result := make([]llm.CompletionResponse, 0, len(resps))
+
+	// Track position in truncated sequence
+	pos := 0
+	truncationHappened := false
+	for _, resp := range resps {
+		if pos >= len(truncated) {
 			break
 		}
 
-		end := start + length
-		if end > len(joined) {
-			end = len(joined)
-			tokenTruncated = true
+		chunk := truncated[pos:min(pos+len(resp.Content), len(truncated))]
+		if len(chunk) < len(resp.Content) {
+			truncationHappened = true
 		}
-		result = append(result, joined[start:end])
-		start = end
+		if len(chunk) > 0 {
+			result = append(result, llm.CompletionResponse{Content: chunk})
+		}
+		pos += len(resp.Content)
 	}
 
-	return result, tokenTruncated
-}
-
-func IncompleteUnicode(token string) bool {
-	incomplete := false
-
-	// check if there is incomplete UTF-8 character at the end
-	for i := 1; i < 5 && i <= len(token); i++ {
-		c := token[len(token)-i]
-
-		if (c & 0xc0) == 0x80 {
-			// continuation byte: 10xxxxxx
-			continue
-		}
-
-		if (c & 0xe0) == 0xc0 {
-			// 2-byte character: 110xxxxx ...
-			incomplete = i < 2
-		} else if (c & 0xf0) == 0xe0 {
-			// 3-byte character: 1110xxxx ...
-			incomplete = i < 3
-		} else if (c & 0xf8) == 0xf0 {
-			// 4-byte character: 11110xxx ...
-			incomplete = i < 4
-		}
-
-		// else 1-byte character or invalid byte
-		break
-	}
-
-	return incomplete
+	return result, truncationHappened
 }

runner/common/stop_test.go

@@ -1,51 +1,84 @@
 package common
 
 import (
+	"fmt"
 	"reflect"
 	"testing"
+
+	"github.com/ollama/ollama/llm"
 )
 
 func TestTruncateStop(t *testing.T) {
 	tests := []struct {
 		name          string
-		pieces        []string
+		pieces        []llm.CompletionResponse
 		stop          string
-		expected      []string
+		expected      []llm.CompletionResponse
 		expectedTrunc bool
 	}{
 		{
-			name:          "Single word",
-			pieces:        []string{"hello", "world"},
-			stop:          "world",
-			expected:      []string{"hello"},
+			name: "Single word",
+			pieces: []llm.CompletionResponse{
+				{Content: "Hello"},
+				{Content: "world"},
+			},
+			stop: "world",
+			expected: []llm.CompletionResponse{
+				{Content: "Hello"},
+			},
 			expectedTrunc: false,
 		},
 		{
-			name:          "Partial",
-			pieces:        []string{"hello", "wor"},
-			stop:          "or",
-			expected:      []string{"hello", "w"},
+			name: "Partial",
+			pieces: []llm.CompletionResponse{
+				{Content: "Hello"},
+				{Content: " wor"},
+			},
+			stop: "or",
+			expected: []llm.CompletionResponse{
+				{Content: "Hello"},
+				{Content: " w"},
+			},
 			expectedTrunc: true,
 		},
 		{
-			name:          "Suffix",
-			pieces:        []string{"Hello", " there", "!"},
-			stop:          "!",
-			expected:      []string{"Hello", " there"},
+			name: "Suffix",
+			pieces: []llm.CompletionResponse{
+				{Content: "Hello"},
+				{Content: " there"},
+				{Content: "!"},
+			},
+			stop: "!",
+			expected: []llm.CompletionResponse{
+				{Content: "Hello"},
+				{Content: " there"},
+			},
 			expectedTrunc: false,
 		},
 		{
-			name:          "Suffix partial",
-			pieces:        []string{"Hello", " the", "re!"},
-			stop:          "there!",
-			expected:      []string{"Hello", " "},
+			name: "Suffix partial",
+			pieces: []llm.CompletionResponse{
+				{Content: "Hello"},
+				{Content: " the"},
+				{Content: "re!"},
+			},
+			stop: "there!",
+			expected: []llm.CompletionResponse{
+				{Content: "Hello"},
+				{Content: " "},
+			},
 			expectedTrunc: true,
 		},
 		{
-			name:          "Middle",
-			pieces:        []string{"hello", " wor"},
-			stop:          "llo w",
-			expected:      []string{"he"},
+			name: "Middle",
+			pieces: []llm.CompletionResponse{
+				{Content: "Hello"},
+				{Content: " wo"},
+			},
+			stop: "llo w",
+			expected: []llm.CompletionResponse{
+				{Content: "He"},
+			},
 			expectedTrunc: true,
 		},
 	}
@@ -54,76 +87,23 @@ func TestTruncateStop(t *testing.T) {
 		t.Run(tt.name, func(t *testing.T) {
 			result, resultTrunc := TruncateStop(tt.pieces, tt.stop)
 			if !reflect.DeepEqual(result, tt.expected) || resultTrunc != tt.expectedTrunc {
-				t.Errorf("truncateStop(%v, %s): have %v (%v); want %v (%v)", tt.pieces, tt.stop, result, resultTrunc, tt.expected, tt.expectedTrunc)
+				t.Errorf("truncateStop(%v, %v):\n%shave truncated %v\nwant truncated %v",
+					tt.pieces, tt.stop, formatContentDiff(result, tt.expected), resultTrunc, tt.expectedTrunc)
 			}
 		})
 	}
 }
 
-func TestIncompleteUnicode(t *testing.T) {
-	tests := []struct {
-		name     string
-		input    string
-		expected bool
-	}{
-		{
-			name:     "Basic",
-			input:    "hi",
-			expected: false,
-		},
-		{
-			name:     "Two byte",
-			input:    "hi" + string([]byte{0xc2, 0xa3}),
-			expected: false,
-		},
-		{
-			name:     "Two byte - missing last",
-			input:    "hi" + string([]byte{0xc2}),
-			expected: true,
-		},
-		{
-			name:     "Three byte",
-			input:    "hi" + string([]byte{0xe0, 0xA0, 0x80}),
-			expected: false,
-		},
-		{
-			name:     "Three byte - missing last",
-			input:    "hi" + string([]byte{0xe0, 0xA0}),
-			expected: true,
-		},
-		{
-			name:     "Three byte - missing last 2",
-			input:    "hi" + string([]byte{0xe0}),
-			expected: true,
-		},
-		{
-			name:     "Four byte",
-			input:    "hi" + string([]byte{0xf0, 0x92, 0x8a, 0xb7}),
-			expected: false,
-		},
-		{
-			name:     "Four byte - missing last",
-			input:    "hi" + string([]byte{0xf0, 0x92, 0x8a}),
-			expected: true,
-		},
-		{
-			name:     "Four byte - missing last 2",
-			input:    "hi" + string([]byte{0xf0, 0x92}),
-			expected: true,
-		},
-		{
-			name:     "Four byte - missing last 3",
-			input:    "hi" + string([]byte{0xf0}),
-			expected: true,
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			result := IncompleteUnicode(tt.input)
-			if result != tt.expected {
-				t.Errorf("incompleteUnicode(%s): have %v; want %v", tt.input, result, tt.expected)
-			}
-		})
+func formatContentDiff(result, expected []llm.CompletionResponse) string {
+	var s string
+	for i := 0; i < len(result) || i < len(expected); i++ {
+		if i < len(result) && i < len(expected) && result[i].Content != expected[i].Content {
+			s += fmt.Sprintf("[%d] %q vs %q\n", i, result[i].Content, expected[i].Content)
+		} else if i < len(result) && i >= len(expected) {
+			s += fmt.Sprintf("[%d] extra %q\n", i, result[i].Content)
+		} else if i >= len(result) && i < len(expected) {
+			s += fmt.Sprintf("[%d] missing %q\n", i, expected[i].Content)
+		}
 	}
+	return s
 }

runner/llamarunner/runner.go

@@ -17,7 +17,6 @@ import (
 	"strings"
 	"sync"
 	"time"
-	"unicode/utf8"
 
 	"golang.org/x/sync/semaphore"
 
@@ -52,13 +51,13 @@ type Sequence struct {
 	pendingInputs []input
 
 	// tokens that have been generated but not returned yet (e.g. for stop sequences)
-	pendingResponses []string
+	pendingResponses []llm.CompletionResponse
 
 	// input cache being used by this sequence
 	cache *InputCacheSlot
 
 	// channel to send responses over
-	responses chan string
+	responses chan llm.CompletionResponse
 
 	// channel to stop decoding (such as if the remote connection is closed)
 	quit chan bool
@@ -89,6 +88,19 @@ type Sequence struct {
 	numPromptInputs     int
 }
 
+func (seq *Sequence) send(resp llm.CompletionResponse) bool {
+	if len(resp.Content) > 0 || resp.Done {
+		select {
+		case seq.responses <- resp:
+			// Successfully sent
+			return true
+		case <-seq.quit:
+			return false
+		}
+	}
+	return true
+}
+
 type NewSequenceParams struct {
 	numPredict     int
 	stop           []string
@@ -147,8 +159,8 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 		numPromptInputs:     len(inputs),
 		startProcessingTime: startTime,
 		numPredict:          params.numPredict,
-		pendingResponses:    make([]string, 0),
-		responses:           make(chan string, 100),
+		pendingResponses:    make([]llm.CompletionResponse, 0),
+		responses:           make(chan llm.CompletionResponse, 100),
 		quit:                make(chan bool, 1),
 		embedding:           make(chan []float32, 1),
 		samplingCtx:         sc,
@@ -272,36 +284,15 @@ func (s *Server) allNil() bool {
 	return true
 }
 
-func flushPending(seq *Sequence) bool {
-	joined := strings.Join(seq.pendingResponses, "")
-	seq.pendingResponses = []string{}
-
-	// Check if there are any partial UTF-8 characters remaining.
-	// We already check and queue as we are generating but some may
-	// still make it here:
-	// - Sequence is ending, e.g. generation limit has been hit
-	// - Invalid characters in the middle of a string
-	// This is a stricter check to ensure we never output invalid Unicode.
-	for !utf8.ValidString(joined) {
-		joined = joined[:len(joined)-1]
-	}
-
-	if len(joined) == 0 {
-		return true
-	}
-
-	select {
-	case seq.responses <- joined:
-		return true
-	case <-seq.quit:
-		return false
-	}
-}
-
 func (s *Server) removeSequence(seqIndex int, reason llm.DoneReason) {
 	seq := s.seqs[seqIndex]
 
-	flushPending(seq)
+	// Send any remaining pending responses
+	for _, resp := range seq.pendingResponses {
+		seq.send(resp)
+	}
+	seq.pendingResponses = []llm.CompletionResponse{}
+
 	seq.doneReason = reason
 	close(seq.responses)
 	close(seq.embedding)
@@ -490,8 +481,11 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 
 		seq.inputs = []input{{token: token}}
 
-		seq.pendingResponses = append(seq.pendingResponses, piece)
-		sequence := strings.Join(seq.pendingResponses, "")
+		seq.pendingResponses = append(seq.pendingResponses, llm.CompletionResponse{Content: piece})
+		sequence := ""
+		for _, r := range seq.pendingResponses {
+			sequence += r.Content
+		}
 
 		if ok, stop := common.FindStop(sequence, seq.stop); ok {
 			slog.Debug("hit stop token", "pending", seq.pendingResponses, "stop", stop)
@@ -523,13 +517,13 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 			continue
 		}
 
-		if common.IncompleteUnicode(sequence) {
-			continue
-		}
-
-		if !flushPending(seq) {
-			s.removeSequence(i, llm.DoneReasonConnectionClosed)
+		for _, resp := range seq.pendingResponses {
+			if !seq.send(resp) {
+				s.removeSequence(i, llm.DoneReasonConnectionClosed)
+				break
+			}
 		}
+		seq.pendingResponses = []llm.CompletionResponse{}
 	}
 
 	return nil
@@ -627,9 +621,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 			return
 		case content, ok := <-seq.responses:
 			if ok {
-				if err := json.NewEncoder(w).Encode(&llm.CompletionResponse{
-					Content: content,
-				}); err != nil {
+				if err := json.NewEncoder(w).Encode(&content); err != nil {
 					http.Error(w, fmt.Sprintf("failed to encode response: %v", err), http.StatusInternalServerError)
 					close(seq.quit)
 					return

runner/ollamarunner/runner.go

@@ -20,7 +20,6 @@ import (
 	"strings"
 	"sync"
 	"time"
-	"unicode/utf8"
 
 	"golang.org/x/image/bmp"
 	"golang.org/x/sync/semaphore"
@@ -56,13 +55,13 @@ type Sequence struct {
 	pendingInputs []input.Input
 
 	// tokens that have been generated but not returned yet (e.g. for stop sequences)
-	pendingResponses []string
+	pendingResponses []llm.CompletionResponse
 
 	// input cache being used by this sequence
 	cache *InputCacheSlot
 
 	// channel to send responses over
-	responses chan string
+	responses chan llm.CompletionResponse
 
 	// channel to stop decoding (such as if the remote connection is closed)
 	quit chan bool
@@ -94,6 +93,19 @@ type Sequence struct {
 	numPromptInputs     int
 }
 
+func (seq *Sequence) send(resp llm.CompletionResponse) bool {
+	if len(resp.Content) > 0 || resp.Done {
+		select {
+		case seq.responses <- resp:
+			// Successfully sent
+			return true
+		case <-seq.quit:
+			return false
+		}
+	}
+	return true
+}
+
 type NewSequenceParams struct {
 	numPredict int
 	stop       []string
@@ -167,8 +179,8 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 		numPromptInputs:     len(inputs),
 		startProcessingTime: startTime,
 		numPredict:          params.numPredict,
-		pendingResponses:    make([]string, 0),
-		responses:           make(chan string, 100),
+		pendingResponses:    make([]llm.CompletionResponse, 0),
+		responses:           make(chan llm.CompletionResponse, 100),
 		quit:                make(chan bool, 1),
 		embedding:           make(chan []float32, 1),
 		sampler:             params.sampler,
@@ -313,36 +325,15 @@ func (s *Server) allNil() bool {
 	return true
 }
 
-func flushPending(seq *Sequence) bool {
-	joined := strings.Join(seq.pendingResponses, "")
-	seq.pendingResponses = []string{}
-
-	// Check if there are any partial UTF-8 characters remaining.
-	// We already check and queue as we are generating but some may
-	// still make it here:
-	// - Sequence is ending, e.g. generation limit has been hit
-	// - Invalid characters in the middle of a string
-	// This is a stricter check to ensure we never output invalid Unicode.
-	for !utf8.ValidString(joined) {
-		joined = joined[:len(joined)-1]
-	}
-
-	if len(joined) == 0 {
-		return true
-	}
-
-	select {
-	case seq.responses <- joined:
-		return true
-	case <-seq.quit:
-		return false
-	}
-}
-
 func (s *Server) removeSequence(seqIndex int, reason llm.DoneReason) {
 	seq := s.seqs[seqIndex]
 
-	flushPending(seq)
+	// Send any remaining pending responses
+	for _, resp := range seq.pendingResponses {
+		seq.send(resp)
+	}
+	seq.pendingResponses = []llm.CompletionResponse{}
+
 	seq.doneReason = reason
 	close(seq.responses)
 	close(seq.embedding)
@@ -541,8 +532,11 @@ func (s *Server) processBatch() error {
 
 		seq.inputs = []input.Input{{Token: token}}
 
-		seq.pendingResponses = append(seq.pendingResponses, piece)
-		sequence := strings.Join(seq.pendingResponses, "")
+		seq.pendingResponses = append(seq.pendingResponses, llm.CompletionResponse{Content: piece})
+		sequence := ""
+		for _, r := range seq.pendingResponses {
+			sequence += r.Content
+		}
 
 		if ok, stop := common.FindStop(sequence, seq.stop); ok {
 			slog.Debug("hit stop token", "pending", seq.pendingResponses, "stop", stop)
@@ -574,13 +568,14 @@ func (s *Server) processBatch() error {
 			continue
 		}
 
-		if common.IncompleteUnicode(sequence) {
-			continue
-		}
-
-		if !flushPending(seq) {
-			s.removeSequence(i, llm.DoneReasonConnectionClosed)
+		// Send all pending responses directly without unicode checking
+		for _, resp := range seq.pendingResponses {
+			if !seq.send(resp) {
+				s.removeSequence(i, llm.DoneReasonConnectionClosed)
+				break
+			}
 		}
+		seq.pendingResponses = []llm.CompletionResponse{}
 	}
 
 	return nil
@@ -683,9 +678,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 			return
 		case content, ok := <-seq.responses:
 			if ok {
-				if err := json.NewEncoder(w).Encode(&llm.CompletionResponse{
-					Content: content,
-				}); err != nil {
+				if err := json.NewEncoder(w).Encode(&content); err != nil {
 					http.Error(w, fmt.Sprintf("failed to encode response: %v", err), http.StatusInternalServerError)
 					close(seq.quit)
 					return

scripts/build_windows.ps1

@@ -27,6 +27,7 @@ function checkEnv() {
         $env:VCToolsRedistDir=(get-item "${MSVC_INSTALL}\VC\Redist\MSVC\*")[0]
     }
     # Locate CUDA versions
+    # Note: this assumes every version found will be built
     $cudaList=(get-item "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v*\bin\" -ea 'silentlycontinue')
     if ($cudaList.length -eq 0) {
         $d=(get-command -ea 'silentlycontinue' nvcc).path
@@ -93,6 +94,19 @@ function buildOllama() {
 
         $hashEnv = @{}
         Get-ChildItem env: | foreach { $hashEnv[$_.Name] = $_.Value }
+        if ("$script:CUDA_DIRS".Contains("v11")) {
+            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V11")) { $v11="$_" }}
+            $env:CUDAToolkit_ROOT=$hashEnv[$v11]
+            write-host "Building CUDA v11 backend libraries"
+            # Note: cuda v11 requires msvc 2019 so force the older generator
+            # to avoid 2022 (or newer) from being used as the default
+            & cmake --fresh --preset "CUDA 11" -G "Visual Studio 16 2019" --install-prefix $script:DIST_DIR
+            if ($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($LASTEXITCODE)}
+        }
         if ("$script:CUDA_DIRS".Contains("v12")) {
             $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V12")) { $v12="$_" }}
             $env:CUDAToolkit_ROOT=$hashEnv[$v12]

scripts/env.sh

@@ -10,7 +10,9 @@ OLLAMA_COMMON_BUILD_ARGS="--build-arg=VERSION \
     --build-arg=GOFLAGS \
     --build-arg=OLLAMA_CUSTOM_CPU_DEFS \
     --build-arg=OLLAMA_SKIP_CUDA_GENERATE \
+    --build-arg=OLLAMA_SKIP_CUDA_11_GENERATE \
     --build-arg=OLLAMA_SKIP_CUDA_12_GENERATE \
+    --build-arg=CUDA_V11_ARCHITECTURES \
     --build-arg=CUDA_V12_ARCHITECTURES \
     --build-arg=OLLAMA_SKIP_ROCM_GENERATE \
     --build-arg=OLLAMA_FAST_BUILD \

server/internal/cache/blob/cache.go

@@ -59,7 +59,7 @@ type DiskCache struct {
 	testHookBeforeFinalWrite func(f *os.File)
 }
 
-// PutBytes is a convenience function for c.Put(d, strings.NewReader(s), int64(len(s))).
+// PutString is a convenience function for c.Put(d, strings.NewReader(s), int64(len(s))).
 func PutBytes[S string | []byte](c *DiskCache, d Digest, data S) error {
 	return c.Put(d, bytes.NewReader([]byte(data)), int64(len(data)))
 }

tools/tools.go

@@ -18,8 +18,9 @@ const (
 )
 
 type Parser struct {
-	tag   string
-	tools []api.Tool
+	tag        string
+	names      []string
+	properties []string
 
 	state  toolsState
 	buffer []byte
@@ -33,10 +34,15 @@ func NewParser(tmpl *template.Template, tools []api.Tool) *Parser {
 }
 
 func NewParserWithTag(tools []api.Tool, tag string) *Parser {
-	return &Parser{
-		tag:   tag,
-		tools: tools,
+	var p Parser
+	for _, t := range tools {
+		p.names = append(p.names, t.Function.Name)
+		for r := range t.Function.Parameters.Properties {
+			p.properties = append(p.properties, r)
+		}
 	}
+	p.tag = tag
+	return &p
 }
 
 // Add processes a string input to parse tool calls and content that
@@ -115,40 +121,36 @@ func (p *Parser) findTag() (int, bool) {
 // parseToolCall finds the next complete tool call in the buffer
 // incrementing n and advancing the buffer.
 func (p *Parser) parseToolCall() *api.ToolCall {
-	var tool *api.Tool
+	var name string
+	var args map[string]any
 	var end int = len(p.buffer)
-	var i int
 
 	// find tool name
-	for _, t := range p.tools {
-		n := t.Function.Name
+	var i int
+	for _, n := range p.names {
 		if i = bytes.Index(p.buffer, []byte(n)); i != -1 {
 			if i+len(n) < end {
-				tool = &t
+				name = n
 				end = i + len(n)
 			}
 		}
 	}
 
-	if tool == nil {
+	if name == "" {
 		return nil
 	}
 
-	// only look for arguments if the tool has parameters
-	args := map[string]any{}
-	if len(tool.Function.Parameters.Properties) > 0 {
-		if args, i = p.findArguments(*tool); args == nil {
-			return nil
-		}
+	if args, i = p.findArguments(); args == nil {
+		return nil
+	}
 
-		if i > end {
-			end = i
-		}
+	if i > end {
+		end = i
 	}
 
 	tc := &api.ToolCall{
 		Function: api.ToolCallFunction{
-			Name:      tool.Function.Name,
+			Name:      name,
 			Arguments: args,
 			Index:     p.n,
 		},
@@ -160,17 +162,13 @@ func (p *Parser) parseToolCall() *api.ToolCall {
 }
 
 // findArguments returns the first object that appears to be
-// arguments for the provided tool, returning nil
-func (p *Parser) findArguments(tool api.Tool) (map[string]any, int) {
+// arguments and the position where the arguments end, returning nil and 0 if
+// an invalid JSON object or non-arguments object is found first
+func (p *Parser) findArguments() (map[string]any, int) {
 	if len(p.buffer) == 0 {
 		return nil, 0
 	}
 
-	// no arguments to parse
-	if len(tool.Function.Parameters.Properties) == 0 {
-		return nil, 0
-	}
-
 	var braces int
 	var start int = -1
 	var end int
@@ -186,13 +184,11 @@ func (p *Parser) findArguments(tool api.Tool) (map[string]any, int) {
 		}
 
 		if c == '}' {
-			if start != -1 {
-				braces--
-				if braces == 0 {
-					end = i + 1
-					object = p.buffer[start:end]
-					break
-				}
+			braces--
+			if braces == 0 && start != -1 {
+				end = i + 1
+				object = p.buffer[start:end]
+				break
 			}
 		}
 	}
@@ -210,27 +206,24 @@ func (p *Parser) findArguments(tool api.Tool) (map[string]any, int) {
 
 	var find func(obj any) map[string]any
 	find = func(obj any) map[string]any {
-		switch obj := obj.(type) {
+		switch v := obj.(type) {
 		case map[string]any:
-			found := true
-			for key := range obj {
-				if _, exists := tool.Function.Parameters.Properties[key]; !exists {
-					found = false
-					break
+			// check if the object keys are valid tool properties
+			// TODO (jmorganca): check only sets of properties that
+			// go together instead of the entire set
+			for _, prop := range p.properties {
+				if _, exists := v[prop]; exists {
+					return v
 				}
 			}
 
-			if found {
-				return obj
-			}
-
-			for _, value := range obj {
+			for _, value := range v {
 				if result := find(value); result != nil {
 					return result
 				}
 			}
 		case []any:
-			for _, item := range obj {
+			for _, item := range v {
 				if result := find(item); result != nil {
 					return result
 				}

tools/tools_test.go

@@ -104,13 +104,6 @@ func TestParser(t *testing.T) {
 				},
 			},
 		},
-		{
-			Type: "function",
-			Function: api.ToolFunction{
-				Name:        "say_hello",
-				Description: "Say hello",
-			},
-		},
 	}
 
 	tests := []struct {
@@ -151,20 +144,6 @@ func TestParser(t *testing.T) {
 				},
 			},
 		},
-		{
-			name:    "invalid arguments",
-			inputs:  []string{`<tool_call>{"name": "get_conditions", "arguments": {"city": "San Francisco"}}</tool_call>`},
-			content: "",
-			tmpl:    qwen,
-			calls:   nil,
-		},
-		{
-			name:    "missing args",
-			inputs:  []string{`<tool_call>{"name": "get_conditions"}</tool_call>`},
-			content: "",
-			tmpl:    qwen,
-			calls:   nil,
-		},
 		{
 			name:    "text before tool call",
 			inputs:  []string{`Let me check the weather. <tool_call>{"name": "get_temperature", "arguments": {"city": "New York"}}</tool_call>`},
@@ -182,28 +161,6 @@ func TestParser(t *testing.T) {
 				},
 			},
 		},
-		{
-			name:    "qwen no args tool call",
-			inputs:  []string{`Let me say hello to the user. I'll use the say_hello tool <tool_call>{"name": "say_hello"}</tool_call>`},
-			content: "Let me say hello to the user. I'll use the say_hello tool ",
-			tmpl:    qwen,
-			calls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Index:     0,
-						Name:      "say_hello",
-						Arguments: api.ToolCallFunctionArguments{},
-					},
-				},
-			},
-		},
-		{
-			name:    "qwen no args with text",
-			inputs:  []string{"Let me say hello to the user. I'll use the say_hello tool. "},
-			content: "Let me say hello to the user. I'll use the say_hello tool. ",
-			tmpl:    qwen,
-			calls:   nil,
-		},
 		{
 			name:    "two tool calls in a list",
 			inputs:  []string{`[TOOL_CALLS] [{"name": "get_temperature", "arguments": {"city": "London", "format": "fahrenheit"}}, {"name": "get_conditions", "arguments": {"location": "Tokyo"}}][/TOOL_CALLS]`},
@@ -232,7 +189,7 @@ func TestParser(t *testing.T) {
 			},
 		},
 		{
-			name:    "qwen two tool calls",
+			name:    "two tool calls",
 			inputs:  []string{`Okay, let's call both tools! <tool_call>{"name": "get_temperature", "arguments": {"city": "London", "format": "fahrenheit"}}</tool_call><tool_call>{"name": "get_conditions", "arguments": {"location": "Tokyo"}}</tool_call>`},
 			content: "Okay, let's call both tools! ",
 			tmpl:    qwen,
@@ -258,30 +215,6 @@ func TestParser(t *testing.T) {
 				},
 			},
 		},
-		{
-			name:    "qwen two tool calls one with no args",
-			inputs:  []string{`Let me check the weather. <tool_call>{"name": "say_hello"}</tool_call><tool_call>{"name": "get_conditions", "arguments": {"location": "Tokyo"}}`},
-			content: "Let me check the weather. ",
-			tmpl:    qwen,
-			calls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Index:     0,
-						Name:      "say_hello",
-						Arguments: api.ToolCallFunctionArguments{},
-					},
-				},
-				{
-					Function: api.ToolCallFunction{
-						Index: 1,
-						Name:  "get_conditions",
-						Arguments: api.ToolCallFunctionArguments{
-							"location": "Tokyo",
-						},
-					},
-				},
-			},
-		},
 		{
 			name:    "deepseek",
 			inputs:  []string{"<think>Wait, I need to call a tool</think><|tool▁calls▁begin|><|tool▁call▁begin|>function<|tool▁sep|>get_temperature\n```json\n{\"city\": \"Tokyo\"}\n```<|tool▁call▁end|><|tool▁calls▁end|><|end▁of▁sentence|>"},
@@ -405,52 +338,6 @@ func TestParser(t *testing.T) {
 			content: "for { fmt.Println(\"hello\") }",
 			tmpl:    json,
 		},
-		{
-			name: "json no args tool call",
-			inputs: []string{
-				"{\"name\": \"say_hello\"}",
-			},
-			content: "",
-			tmpl:    json,
-			calls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Index:     0,
-						Name:      "say_hello",
-						Arguments: api.ToolCallFunctionArguments{},
-					},
-				},
-			},
-		},
-		{
-			name: "json no args no tool call",
-			inputs: []string{
-				"I'll use the say_hello tool to say hello to the user.",
-			},
-			content: "I'll use the say_hello tool to say hello to the user.",
-			tmpl:    json,
-			calls:   nil,
-		},
-
-		// TODO (jmorganca): this is a false positive, we should
-		// not be parsing this as a tool call
-		{
-			name: "json no args false positive",
-			inputs: []string{
-				`{say_hello!!!}`,
-			},
-			content: "",
-			tmpl:    json,
-			calls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Index:     0,
-						Name:      "say_hello",
-						Arguments: api.ToolCallFunctionArguments{},
-					},
-				},
-			},
-		},
 		{
 			name: "list multiple",
 			inputs: []string{
@@ -493,30 +380,6 @@ func TestParser(t *testing.T) {
 		},
 		{
 			name: "list partial",
-			inputs: []string{
-				"[{",
-				"\"name\": \"get_conditions\", ",
-				"\"arguments\": {",
-				"\"location\": \"Tokyo\"",
-				"}",
-				"}",
-			},
-			content: "",
-			tmpl:    list,
-			calls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Index: 0,
-						Name:  "get_conditions",
-						Arguments: api.ToolCallFunctionArguments{
-							"location": "Tokyo",
-						},
-					},
-				},
-			},
-		},
-		{
-			name: "list invalid",
 			inputs: []string{
 				"[",
 				"{",
@@ -530,33 +393,6 @@ func TestParser(t *testing.T) {
 			tmpl:    list,
 			calls:   nil,
 		},
-		{
-			name: "list trailing ]",
-			inputs: []string{
-				"[",
-				"{",
-				"\"name\": \"get_conditions\", ",
-				"\"arguments\": {",
-				"\"location\": \"Tokyo\"",
-				"}",
-				"}",
-				"]",
-				"]",
-			},
-			content: "",
-			tmpl:    list,
-			calls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Index: 0,
-						Name:  "get_conditions",
-						Arguments: api.ToolCallFunctionArguments{
-							"location": "Tokyo",
-						},
-					},
-				},
-			},
-		},
 		{
 			name: "list not a tool call",
 			inputs: []string{
@@ -568,26 +404,6 @@ func TestParser(t *testing.T) {
 			tmpl:    list,
 			calls:   nil,
 		},
-		{
-			name: "list with no arguments",
-			inputs: []string{
-				"[",
-				"{",
-				"\"name\": \"say_hello\"",
-				"}",
-			},
-			content: "",
-			tmpl:    list,
-			calls: []api.ToolCall{
-				{
-					Function: api.ToolCallFunction{
-						Index:     0,
-						Name:      "say_hello",
-						Arguments: api.ToolCallFunctionArguments{},
-					},
-				},
-			},
-		},
 	}
 
 	for _, tt := range tests {
@@ -884,75 +700,25 @@ func TestFindTag(t *testing.T) {
 }
 
 func TestFindArguments(t *testing.T) {
-	tool := api.Tool{
-		Type: "function",
-		Function: api.ToolFunction{
-			Name:        "get_temperature",
-			Description: "Retrieve the temperature for a given location",
-			Parameters: struct {
-				Type       string   `json:"type"`
-				Defs       any      `json:"$defs,omitempty"`
-				Items      any      `json:"items,omitempty"`
-				Required   []string `json:"required"`
-				Properties map[string]struct {
-					Type        api.PropertyType `json:"type"`
-					Items       any              `json:"items,omitempty"`
-					Description string           `json:"description"`
-					Enum        []any            `json:"enum,omitempty"`
-				} `json:"properties"`
-			}{
-				Type: "object",
-				Properties: map[string]struct {
-					Type        api.PropertyType `json:"type"`
-					Items       any              `json:"items,omitempty"`
-					Description string           `json:"description"`
-					Enum        []any            `json:"enum,omitempty"`
-				}{
-					"format": {
-						Type:        api.PropertyType{"string"},
-						Description: "The format to return the temperature in",
-						Enum:        []any{"fahrenheit", "celsius"},
-					},
-					"location": {
-						Type:        api.PropertyType{"string"},
-						Description: "The location to get the temperature for",
-					},
-				},
-			},
-		},
-	}
-
-	tool2 := api.Tool{
-		Type: "function",
-		Function: api.ToolFunction{
-			Name:        "say_hello",
-			Description: "Say hello to the user",
-		},
-	}
-
 	tests := []struct {
 		name   string
 		buffer []byte
 		want   map[string]any
-		tool   api.Tool
 	}{
 		{
 			name:   "empty string",
 			buffer: []byte{},
 			want:   nil,
-			tool:   tool,
 		},
 		{
 			name:   "whitespace only",
 			buffer: []byte("   \n\t  "),
 			want:   nil,
-			tool:   tool,
 		},
 		{
 			name:   "unbalanced braces - missing closing",
 			buffer: []byte(`{"format": "fahrenheit", "location": "San Francisco"`),
 			want:   nil,
-			tool:   tool,
 		},
 		{
 			name:   "unbalanced braces - extra closing",
@@ -960,13 +726,11 @@ func TestFindArguments(t *testing.T) {
 			want: map[string]any{
 				"format": "fahrenheit",
 			},
-			tool: tool,
 		},
 		{
 			name:   "invalid JSON",
 			buffer: []byte(`{format: fahrenheit, location: "San Francisco"}`),
 			want:   nil,
-			tool:   tool,
 		},
 		{
 			name:   "valid json",
@@ -975,7 +739,6 @@ func TestFindArguments(t *testing.T) {
 				"format":   "fahrenheit",
 				"location": "San Francisco, CA",
 			},
-			tool: tool,
 		},
 		{
 			name:   "valid arguments with special tokens",
@@ -984,7 +747,6 @@ func TestFindArguments(t *testing.T) {
 				"format":   "fahrenheit",
 				"location": "San Francisco, CA",
 			},
-			tool: tool,
 		},
 		{
 			name:   "valid arguments in array",
@@ -993,7 +755,6 @@ func TestFindArguments(t *testing.T) {
 				"format":   "fahrenheit",
 				"location": "San Francisco, CA",
 			},
-			tool: tool,
 		},
 		{
 			name:   "nested deep",
@@ -1002,49 +763,39 @@ func TestFindArguments(t *testing.T) {
 				"format":   "fahrenheit",
 				"location": "San Francisco, CA",
 			},
-			tool: tool,
 		},
 		{
 			name:   "one arg",
-			buffer: []byte(`get_temperature({"location": "San Francisco, CA"})`),
+			buffer: []byte(`get_weather({"location": "San Francisco, CA"})`),
 			want: map[string]any{
 				"location": "San Francisco, CA",
 			},
-			tool: tool,
 		},
 		{
 			name:   "two args",
-			buffer: []byte(`[{"name": "get_temperature", "arguments": {"location": "San Francisco, CA", "format": "fahrenheit"}}, {"name": "get_weather", "arguments": {"location": "San Francisco, CA", "format": "fahrenheit"}}]`),
+			buffer: []byte(`[{"name": "get_weather", "arguments": {"location": "San Francisco, CA", "format": "fahrenheit"}}, {"name": "get_weather", "arguments": {"location": "San Francisco, CA", "format": "fahrenheit"}}]`),
 			want: map[string]any{
 				"location": "San Francisco, CA",
 				"format":   "fahrenheit",
 			},
-			tool: tool,
-		},
-		{
-			name:   "no args",
-			buffer: []byte(`{"name": "say_hello"}`),
-			want:   nil,
-			tool:   tool2,
 		},
 		{
 			name:   "deepseek",
-			buffer: []byte("<|tool▁calls▁begin|><|tool▁call▁begin|>function<|tool▁sep|>get_temperature\n```json\n{\"location\": \"Tokyo\"}\n```<|tool▁call▁end|><|tool▁calls▁end|><|end▁of▁sentence|>"),
+			buffer: []byte("<|tool▁calls▁begin|><|tool▁call▁begin|>function<|tool▁sep|>get_current_weather\n```json\n{\"location\": \"Tokyo\"}\n```<|tool▁call▁end|><|tool▁calls▁end|><|end▁of▁sentence|>"),
 			want: map[string]any{
 				"location": "Tokyo",
 			},
-			tool: tool,
 		},
 	}
 
 	for _, tt := range tests {
 		parser := &Parser{
-			buffer: tt.buffer,
-			tools:  []api.Tool{tool, tool2},
+			buffer:     tt.buffer,
+			properties: []string{"format", "location"},
 		}
 
 		t.Run(tt.name, func(t *testing.T) {
-			got, _ := parser.findArguments(tool)
+			got, _ := parser.findArguments()
 
 			if diff := cmp.Diff(got, tt.want); diff != "" {
 				t.Errorf("scanArguments() args mismatch (-got +want):\n%s", diff)