tools: fix parsing tool calls with empty arguments, missing required fields (#11233)

This is causing segfaults, so disable it. Currently UUIDs are only
used for debugging purposes, although they planned to be used in
additional ways in the future.

Bug #11211

.github/workflows/release.yaml

@@ -54,48 +54,6 @@ jobs:
           name: build-${{ matrix.os }}-${{ matrix.arch }}
           path: dist/*
 
-  darwin-sign:
-    runs-on: macos-13
-    environment: release
-    needs: darwin-build
-    steps:
-      - uses: actions/checkout@v4
-      - run: |
-          echo $MACOS_SIGNING_KEY | base64 --decode > certificate.p12
-          security create-keychain -p password build.keychain
-          security default-keychain -s build.keychain
-          security unlock-keychain -p password build.keychain
-          security import certificate.p12 -k build.keychain -P $MACOS_SIGNING_KEY_PASSWORD -T /usr/bin/codesign
-          security set-key-partition-list -S apple-tool:,apple:,codesign: -s -k password build.keychain
-          security set-keychain-settings -lut 3600 build.keychain
-        env:
-          MACOS_SIGNING_KEY: ${{ secrets.MACOS_SIGNING_KEY }}
-          MACOS_SIGNING_KEY_PASSWORD: ${{ secrets.MACOS_SIGNING_KEY_PASSWORD }}
-      - uses: actions/download-artifact@v4
-        with:
-          name: build-darwin-amd64
-          path: dist/darwin-amd64
-      - uses: actions/download-artifact@v4
-        with:
-          name: build-darwin-arm64
-          path: dist/darwin-arm64
-      - run: |
-          export VERSION=${GITHUB_REF_NAME#v}
-          ./scripts/build_darwin.sh sign macapp
-        env:
-          APPLE_IDENTITY: ${{ secrets.APPLE_IDENTITY }}
-          APPLE_PASSWORD: ${{ secrets.APPLE_PASSWORD }}
-          APPLE_TEAM_ID: ${{ vars.APPLE_TEAM_ID }}
-          APPLE_ID: ${{ vars.APPLE_ID }}
-          SDKROOT: /Applications/Xcode_14.1.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk
-          DEVELOPER_DIR: /Applications/Xcode_14.1.0.app/Contents/Developer
-      - uses: actions/upload-artifact@v4
-        with:
-          name: dist-darwin
-          path: |
-            dist/Ollama-darwin.zip
-            dist/ollama-darwin.tgz
-
   windows-depends:
     strategy:
       matrix:
@@ -230,61 +188,11 @@ jobs:
           go-version-file: go.mod
       - run: |
           go build -o dist/${{ matrix.os }}-${{ matrix.arch }}/ .
-      - if: matrix.arch == 'arm64'
-        run: |
-          Invoke-WebRequest -Uri "https://aka.ms/vs/17/release/vc_redist.arm64.exe" -OutFile "dist\windows-arm64\vc_redist.arm64.exe"
-      - run: |
-          $env:VERSION='${{ github.ref_name }}' -Replace "v(.*)", '$1'
-          & .\scripts\build_windows.ps1 buildApp
-        env:
-          VCToolsRedistDir: stub
       - uses: actions/upload-artifact@v4
         with:
           name: build-${{ matrix.os }}-${{ matrix.arch }}
           path: |
             dist\${{ matrix.os }}-${{ matrix.arch }}\*.exe
-            dist\${{ matrix.os }}-${{ matrix.arch }}-app.exe
-
-  windows-sign:
-    runs-on: windows
-    environment: release
-    needs: [windows-depends, windows-build]
-    steps:
-      - uses: actions/checkout@v4
-      - uses: google-github-actions/auth@v2
-        with:
-          project_id: ollama
-          credentials_json: ${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}
-      - run: |
-          $ErrorActionPreference = "Stop"
-          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${{ runner.temp }}\sdksetup.exe"
-          Start-Process "${{ runner.temp }}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
-
-          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${{ runner.temp }}\plugin.zip"
-          Expand-Archive -Path "${{ runner.temp }}\plugin.zip" -DestinationPath "${{ runner.temp }}\plugin\"
-          & "${{ runner.temp }}\plugin\*\kmscng.msi" /quiet
-
-          echo "${{ vars.OLLAMA_CERT }}" >ollama_inc.crt
-      - uses: actions/download-artifact@v4
-        with:
-          pattern: build-windows-*
-          path: dist\
-          merge-multiple: true
-      - uses: actions/download-artifact@v4
-        with:
-          pattern: depends-windows-amd64-*
-          path: dist\windows-amd64\
-          merge-multiple: true
-      - run: |
-          & .\scripts\build_windows.ps1 gatherDependencies sign buildInstaller distZip
-        env:
-          KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
-      - uses: actions/upload-artifact@v4
-        with:
-          name: dist-windows
-          path: |
-            dist\OllamaSetup.exe
-            dist\ollama-windows-*.zip
 
   linux-build:
     strategy:
@@ -441,58 +349,16 @@ jobs:
   trigger:
     runs-on: ubuntu-latest
     environment: release
-    needs: [darwin-build, windows-build, windows-depends]
-    steps:
-      - name: Trigger downstream release process
-        run: |
-          curl -L \
-            -X POST \
-            -H "Accept: application/vnd.github+json" \
-            -H "Authorization: Bearer ${{ secrets.RELEASE_TOKEN }}" \
-            -H "X-GitHub-Api-Version: 2022-11-28" \
-            https://api.github.com/repos/ollama/${{ vars.RELEASE_REPO }}/dispatches \
-            -d "{\"event_type\": \"trigger-workflow\", \"client_payload\": {\"run_id\": \"${GITHUB_RUN_ID}\", \"version\": \"${GITHUB_REF_NAME#v}\"}}"
-
-  # Aggregate all the assets and ship a release
-  release:
-    needs: [darwin-sign, windows-sign, linux-build]
-    runs-on: linux
-    environment: release
+    needs: [darwin-build, windows-build, windows-depends, linux-build]
     permissions:
       contents: write
     env:
       GH_TOKEN: ${{ github.token }}
     steps:
       - uses: actions/checkout@v4
-      - uses: actions/download-artifact@v4
-        with:
-          name: dist-darwin
-          path: dist
-      - uses: actions/download-artifact@v4
-        with:
-          name: dist-windows
-          path: dist
-      - uses: actions/download-artifact@v4
-        with:
-          pattern: dist-linux-*
-          path: stage
-          merge-multiple: false
-      - name: Merge linux amd64 payload
-        working-directory: stage/dist-linux-amd64-archive
-        run: |
-          tar zxf ollama-linux-amd64.tgz
-          tar zxf ../dist-linux-amd64-rocm/ollama-linux-amd64.tgz
-          rm -f ollama-linux-amd64.tgz ../dist-linux-amd64-rocm/ollama-linux-amd64.tgz
-          tar -c -f- --owner 0 --group 0 . | pigz -9vc > ../ollama-linux-amd64.tgz
-      - name: Cleanup linux payloads
-        run: |
-          find stage -name ollama-linux\*.tgz -exec mv {} dist/ \;
-      - run: find . -type f -not -name 'sha256sum.txt' | xargs sha256sum | tee sha256sum.txt
-        working-directory: dist
-      - name: Create or update Release
+      - name: Create or update Release for tag
         run: |
           RELEASE_VERSION="$(echo ${GITHUB_REF_NAME} | cut -f1 -d-)"
-
           echo "Looking for existing release for ${RELEASE_VERSION}"
           OLD_TAG=$(gh release ls --json name,tagName | jq -r ".[] | select(.name == \"${RELEASE_VERSION}\") | .tagName")
           if [ -n "$OLD_TAG" ]; then
@@ -506,5 +372,12 @@ jobs:
               --generate-notes \
               --prerelease
           fi
-          echo "Uploading artifacts for tag ${GITHUB_REF_NAME}"
-          gh release upload ${GITHUB_REF_NAME} dist/* --clobber
+      - name: Trigger downstream release process
+        run: |
+          curl -L \
+            -X POST \
+            -H "Accept: application/vnd.github+json" \
+            -H "Authorization: Bearer ${{ secrets.RELEASE_TOKEN }}" \
+            -H "X-GitHub-Api-Version: 2022-11-28" \
+            https://api.github.com/repos/ollama/${{ vars.RELEASE_REPO }}/dispatches \
+            -d "{\"event_type\": \"trigger-workflow\", \"client_payload\": {\"run_id\": \"${GITHUB_RUN_ID}\", \"version\": \"${GITHUB_REF_NAME#v}\", \"publish\": \"1\"}}"

README.md

@@ -455,6 +455,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [GGUF-to-Ollama](https://github.com/jonathanhecl/gguf-to-ollama) - Importing GGUF to Ollama made easy (multiplatform)
 - [AWS-Strands-With-Ollama](https://github.com/rapidarchitect/ollama_strands) - AWS Strands Agents with Ollama Examples
 - [ollama-multirun](https://github.com/attogram/ollama-multirun) - A bash shell script to run a single prompt against any or all of your locally installed ollama models, saving the output and performance statistics as easily navigable web pages. ([Demo](https://attogram.github.io/ai_test_zone/))
+- [ollama-bash-toolshed](https://github.com/attogram/ollama-bash-toolshed) - Bash scripts to chat with tool using models. Add new tools to your shed with ease. Runs on Ollama.
 
 ### Apple Vision Pro
 

convert/convert_gemma3n.go

@@ -24,7 +24,6 @@ type gemma3nModel struct {
 		HiddenSize                uint32    `json:"hidden_size"`
 		HiddenSizePerLayerInput   uint32    `json:"hidden_size_per_layer_input"`
 		IntermediateSize          uint32    `json:"intermediate_size"`
-		LaurelRank                uint32    `json:"laurel_rank"`
 		MaxPositionEmbeddings     uint32    `json:"max_position_embeddings"`
 		NumAttentionHeads         uint32    `json:"num_attention_heads"`
 		NumHiddenLayers           uint32    `json:"num_hidden_layers"`
@@ -72,8 +71,6 @@ func (m *gemma3nModel) KV(t *Tokenizer) ggml.KV {
 	kv["gemma3n.embedding_length"] = m.TextModel.HiddenSize
 	kv["gemma3n.feed_forward_length"] = m.TextModel.IntermediateSize
 	kv["gemma3n.head_dim"] = m.TextModel.HeadDim
-	kv["gemma3n.laurel_rank"] = m.TextModel.LaurelRank
-	kv["gemma3n.num_kv_shared_layers"] = m.TextModel.NumKVSharedLayers
 	kv["gemma3n.rope.freq_base_local"] = m.TextModel.RopeLocalBaseFreq
 	kv["gemma3n.rope.freq_base"] = m.TextModel.RopeTheta
 	return kv

ml/backend/ggml/ggml.go

@@ -138,7 +138,10 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 	requiredMemory.CPU.Name = C.GoString(C.ggml_backend_dev_name(cpuDeviceBufferType.d))
 	var props C.struct_ggml_backend_dev_props
 	C.ggml_backend_dev_get_props(cpuDeviceBufferType.d, &props)
-	requiredMemory.CPU.UUID = C.GoString(props.uuid)
+
+	// Bug #11211: Reporting of UUIDs is temporarily disabled due to causing segfaults
+	// This only affects debug information until the new memory management code is in place
+	// requiredMemory.CPU.UUID = C.GoString(props.uuid)
 	requiredMemory.CPU.Weights = make([]ml.Memory, blocks+1)
 	requiredMemory.CPU.Cache = make([]ml.Memory, blocks+1)
 
@@ -155,7 +158,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 		requiredMemory.GPUs[i].Name = C.GoString(C.ggml_backend_dev_name(d))
 		var props C.struct_ggml_backend_dev_props
 		C.ggml_backend_dev_get_props(d, &props)
-		requiredMemory.GPUs[i].UUID = C.GoString(props.uuid)
+		// requiredMemory.GPUs[i].UUID = C.GoString(props.uuid)
 		requiredMemory.GPUs[i].Weights = make([]ml.Memory, blocks+1)
 		requiredMemory.GPUs[i].Cache = make([]ml.Memory, blocks+1)
 	}

model/models/gemma3n/model.go

@@ -39,7 +39,6 @@ func New(c fs.Config) (model.Model, error) {
 		),
 	}
 
-	// TODO: setup hybrid (local sliding window + global) cache
 	m.Cache = kvcache.NewWrapperCache(
 		kvcache.NewCausalCache(m.Shift),
 		kvcache.NewSWACache(int32(c.Uint("attention.sliding_window")), m.Shift),

server/quantization.go

@@ -231,6 +231,8 @@ func newType(t *fsggml.Tensor, kv fsggml.KV, qs *quantizeState, ftype fsggml.Fil
 	// do not quantize relative position bias (T5)
 	quantize = quantize && !strings.Contains(name, "attn_rel_b.weight")
 
+	quantize = quantize && !strings.Contains(name, "per_layer_token_embd.weight")
+
 	newType := fsggml.TensorType(t.Kind)
 	if quantize {
 		// get more optimal quantization type based on the tensor shape, layer, etc.

tools/tools.go

@@ -134,16 +134,16 @@ func (p *Parser) parseToolCall() *api.ToolCall {
 		return nil
 	}
 
-	// only look for arguments if the tool has parameters
+	// only look for arguments after the tool name if the tool has parameters
+	// TODO (jmorganca): while probably uncommon, this doesn't support
+	// parsing arguments before the tool name, which may be needed in the future
 	args := map[string]any{}
 	if len(tool.Function.Parameters.Properties) > 0 {
-		if args, i = p.findArguments(*tool); args == nil {
+		if args, i = findArguments(*tool, p.buffer[end:]); args == nil {
 			return nil
 		}
 
-		if i > end {
-			end = i
-		}
+		end += i
 	}
 
 	tc := &api.ToolCall{
@@ -160,14 +160,14 @@ 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) {
-	if len(p.buffer) == 0 {
-		return nil, 0
-	}
-
-	// no arguments to parse
-	if len(tool.Function.Parameters.Properties) == 0 {
+// arguments for the provided tool in the provided buffer,
+// returning nil if no arguments are found.
+// TODO (jmorganca): this does not support parsing omitted arguments
+// objects for functions that have all-optional parameters
+// e.g. `{"name": "get_conditions", "arguments": {}}` will work but
+// `{"name": "get_conditions"}` will not currently work
+func findArguments(tool api.Tool, buffer []byte) (map[string]any, int) {
+	if len(buffer) == 0 {
 		return nil, 0
 	}
 
@@ -177,7 +177,7 @@ func (p *Parser) findArguments(tool api.Tool) (map[string]any, int) {
 	var object []byte
 
 	// find any outer json object
-	for i, c := range p.buffer {
+	for i, c := range buffer {
 		if c == '{' {
 			braces++
 			if start == -1 {
@@ -190,7 +190,7 @@ func (p *Parser) findArguments(tool api.Tool) (map[string]any, int) {
 				braces--
 				if braces == 0 {
 					end = i + 1
-					object = p.buffer[start:end]
+					object = buffer[start:end]
 					break
 				}
 			}
@@ -202,8 +202,6 @@ func (p *Parser) findArguments(tool api.Tool) (map[string]any, int) {
 	}
 
 	var data map[string]any
-
-	// not valid json
 	if err := json.Unmarshal(object, &data); err != nil {
 		return nil, 0
 	}
@@ -212,15 +210,27 @@ func (p *Parser) findArguments(tool api.Tool) (map[string]any, int) {
 	find = func(obj any) map[string]any {
 		switch obj := obj.(type) {
 		case map[string]any:
-			found := true
+			valid := true
+			// check if all keys in the object exist in the tool's parameters
 			for key := range obj {
 				if _, exists := tool.Function.Parameters.Properties[key]; !exists {
-					found = false
+					valid = false
 					break
 				}
 			}
 
-			if found {
+			// check for required parameters
+			// TODO (jmorganca): this should error instead of silently failing
+			if valid {
+				for _, required := range tool.Function.Parameters.Required {
+					if _, exists := obj[required]; !exists {
+						valid = false
+						break
+					}
+				}
+			}
+
+			if valid {
 				return obj
 			}
 

tools/tools_test.go

@@ -52,7 +52,8 @@ func TestParser(t *testing.T) {
 						Enum        []any            `json:"enum,omitempty"`
 					} `json:"properties"`
 				}{
-					Type: "object",
+					Type:     "object",
+					Required: []string{"city"},
 					Properties: map[string]struct {
 						Type        api.PropertyType `json:"type"`
 						Items       any              `json:"items,omitempty"`
@@ -159,8 +160,23 @@ func TestParser(t *testing.T) {
 			calls:   nil,
 		},
 		{
-			name:    "missing args",
-			inputs:  []string{`<tool_call>{"name": "get_conditions"}</tool_call>`},
+			name:    "empty args",
+			inputs:  []string{`<tool_call>{"name": "get_conditions", "arguments": {}}</tool_call>`},
+			content: "",
+			tmpl:    qwen,
+			calls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Index:     0,
+						Name:      "get_conditions",
+						Arguments: api.ToolCallFunctionArguments{},
+					},
+				},
+			},
+		},
+		{
+			name:    "missing required args",
+			inputs:  []string{`<tool_call>{"name": "get_temperature", "arguments": {}}</tool_call>`},
 			content: "",
 			tmpl:    qwen,
 			calls:   nil,
@@ -259,9 +275,9 @@ 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. ",
+			name:    "empty args followed by args",
+			inputs:  []string{`Let me say hello and check the weather. <tool_call>{"name": "say_hello", "arguments": {}}</tool_call><tool_call>{"name": "get_temperature", "arguments": {"city": "London", "format": "fahrenheit"}}</tool_call>`},
+			content: "Let me say hello and check the weather. ",
 			tmpl:    qwen,
 			calls: []api.ToolCall{
 				{
@@ -271,6 +287,31 @@ func TestParser(t *testing.T) {
 						Arguments: api.ToolCallFunctionArguments{},
 					},
 				},
+				{
+					Function: api.ToolCallFunction{
+						Index: 1,
+						Name:  "get_temperature",
+						Arguments: api.ToolCallFunctionArguments{
+							"city":   "London",
+							"format": "fahrenheit",
+						},
+					},
+				},
+			},
+		},
+		{
+			name:    "qwen empty followed by args",
+			inputs:  []string{`Let me check the weather. <tool_call>{"name": "get_conditions", "arguments": {}}</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:      "get_conditions",
+						Arguments: api.ToolCallFunctionArguments{},
+					},
+				},
 				{
 					Function: api.ToolCallFunction{
 						Index: 1,
@@ -1035,16 +1076,19 @@ func TestFindArguments(t *testing.T) {
 			},
 			tool: tool,
 		},
+		{
+			name:   "deepseek",
+			buffer: []byte(`", "arguments": {"location": "Tokyo"}}</tool_call>`),
+			want: map[string]any{
+				"location": "Tokyo",
+			},
+			tool: tool,
+		},
 	}
 
 	for _, tt := range tests {
-		parser := &Parser{
-			buffer: tt.buffer,
-			tools:  []api.Tool{tool, tool2},
-		}
-
 		t.Run(tt.name, func(t *testing.T) {
-			got, _ := parser.findArguments(tool)
+			got, _ := findArguments(tt.tool, tt.buffer)
 
 			if diff := cmp.Diff(got, tt.want); diff != "" {
 				t.Errorf("scanArguments() args mismatch (-got +want):\n%s", diff)