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

This reverts commit 1a558f98e2.

server/routes.go

@@ -46,6 +46,18 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
+func shouldUseHarmony(model *Model) bool {
+	if slices.Contains([]string{"gptoss", "gpt-oss"}, model.Config.ModelFamily) {
+		// heuristic to check whether the template expects to be parsed via harmony:
+		// search for harmony tags that are nearly always used
+		if model.Template.Contains("<|start|>") && model.Template.Contains("<|end|>") {
+			return true
+		}
+	}
+
+	return false
+}
+
 func experimentEnabled(name string) bool {
 	return slices.Contains(strings.Split(os.Getenv("OLLAMA_EXPERIMENT"), ","), name)
 }
@@ -195,11 +207,13 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	useHarmony := harmony.ShouldUseHarmony(m.Config.ModelFamily, m.Template) && !req.Raw
-	var functionNameMap *harmony.FunctionNameMap
-
+	useHarmony := shouldUseHarmony(m) && !req.Raw
+	var harmonyMessageHandler *harmony.HarmonyMessageHandler
+	var harmonyToolParser *harmony.HarmonyToolCallAccumulator
 	if useHarmony {
-		functionNameMap = harmony.NewFunctionNameMap()
+		harmonyMessageHandler = harmony.NewHarmonyMessageHandler()
+		harmonyMessageHandler.HarmonyParser.AddImplicitStart()
+		harmonyToolParser = harmonyMessageHandler.CreateToolParser()
 	}
 
 	// Validate Think value: string values currently only allowed for gptoss models
@@ -343,19 +357,16 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		var sb strings.Builder
 		defer close(ch)
 		if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
-			Prompt:     prompt,
-			Images:     images,
-			Format:     req.Format,
-			Options:    opts,
-			UseHarmony: useHarmony,
+			Prompt:  prompt,
+			Images:  images,
+			Format:  req.Format,
+			Options: opts,
 		}, func(cr llm.CompletionResponse) {
 			res := api.GenerateResponse{
 				Model:     req.Model,
 				CreatedAt: time.Now().UTC(),
 				Response:  cr.Content,
 				Done:      cr.Done,
-				Thinking:  cr.Thinking,
-				ToolCalls: cr.ToolCalls,
 				Metrics: api.Metrics{
 					PromptEvalCount:    cr.PromptEvalCount,
 					PromptEvalDuration: cr.PromptEvalDuration,
@@ -364,22 +375,12 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 				},
 			}
 
-			if res.Done {
-				res.DoneReason = cr.DoneReason.String()
-				res.TotalDuration = time.Since(checkpointStart)
-				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
-			}
-
 			if useHarmony {
-				for i, tool := range res.ToolCalls {
-					res.ToolCalls[i].Function.Name = functionNameMap.OriginalFromConverted(tool.Function.Name)
-				}
-				if res.Response != "" || res.Thinking != "" || len(res.ToolCalls) > 0 || res.Done {
-					ch <- res
-				}
-				return
-			}
-			if thinkingState != nil {
+				content, thinking, toolContent := harmonyMessageHandler.AddContent(cr.Content, harmonyToolParser)
+				res.Response = content
+				res.Thinking = thinking
+				harmonyToolParser.Add(toolContent)
+			} else if thinkingState != nil {
 				thinking, content := thinkingState.AddContent(cr.Content)
 				res.Thinking = thinking
 				res.Response = content
@@ -390,6 +391,30 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 			}
 
 			if cr.Done {
+				if useHarmony {
+					toolName, toolContent := harmonyToolParser.Drain()
+					if toolName != nil {
+						*toolName = strings.TrimPrefix(*toolName, "functions.")
+						var args api.ToolCallFunctionArguments
+						if err := json.Unmarshal([]byte(toolContent), &args); err != nil {
+							errStr := fmt.Sprintf("error parsing tool call: raw='%s', err=%s", toolContent, err.Error())
+							ch <- gin.H{"error": errStr}
+							return
+						}
+
+						res.ToolCalls = append(res.ToolCalls, api.ToolCall{
+							Function: api.ToolCallFunction{
+								Name:      *toolName,
+								Arguments: args,
+							},
+						})
+					}
+				}
+
+				res.DoneReason = cr.DoneReason.String()
+				res.TotalDuration = time.Since(checkpointStart)
+				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+
 				if !req.Raw {
 					tokens, err := r.Tokenize(c.Request.Context(), prompt+sb.String())
 					if err != nil {
@@ -1592,21 +1617,27 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	}
 	msgs = filterThinkTags(msgs, m)
 
-	useHarmony := harmony.ShouldUseHarmony(m.Config.ModelFamily, m.Template)
+	var harmonyMessageHandler *harmony.HarmonyMessageHandler
+	var harmonyToolParser *harmony.HarmonyToolCallAccumulator
+
+	useHarmony := shouldUseHarmony(m)
 
 	processedTools := req.Tools
-	var functionNameMap *harmony.FunctionNameMap
-	var prefillString string
-	// TODO(parthsareen): this can be abstracted to not be model specific and potentially moved to the runner
 	if useHarmony {
-		prefillString = harmony.Prefill(msgs[len(msgs)-1])
-		functionNameMap = harmony.NewFunctionNameMap()
+		harmonyMessageHandler = harmony.NewHarmonyMessageHandler()
+		var lastMessage *api.Message
+		if len(msgs) > 0 {
+			lastMessage = &msgs[len(msgs)-1]
+		}
+		harmonyMessageHandler.HarmonyParser.AddImplicitStartOrPrefill(lastMessage)
+		harmonyToolParser = harmonyMessageHandler.CreateToolParser()
+
 		// make a copy of tools to pass to the chat prompt. Function names may be
 		// renamed to be valid Harmony function names.
 		processedTools = make([]api.Tool, len(req.Tools))
 		copy(processedTools, req.Tools)
 		for i, tool := range processedTools {
-			processedTools[i].Function.Name = functionNameMap.ConvertAndAdd(tool.Function.Name)
+			processedTools[i].Function.Name = harmonyMessageHandler.FunctionNameMap.ConvertAndAdd(tool.Function.Name)
 		}
 	}
 
@@ -1659,17 +1690,15 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		defer close(ch)
 
 		if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
-			Prompt:        prompt,
-			Images:        images,
-			Format:        req.Format,
-			Options:       opts,
-			UseHarmony:    useHarmony,
-			PrefillString: prefillString,
+			Prompt:  prompt,
+			Images:  images,
+			Format:  req.Format,
+			Options: opts,
 		}, func(r llm.CompletionResponse) {
 			res := api.ChatResponse{
 				Model:     req.Model,
 				CreatedAt: time.Now().UTC(),
-				Message:   api.Message{Role: "assistant", Content: r.Content, Thinking: r.Thinking, ToolCalls: r.ToolCalls},
+				Message:   api.Message{Role: "assistant", Content: r.Content},
 				Done:      r.Done,
 				Metrics: api.Metrics{
 					PromptEvalCount:    r.PromptEvalCount,
@@ -1685,13 +1714,31 @@ func (s *Server) ChatHandler(c *gin.Context) {
 			}
 
 			if useHarmony {
-				for i, tool := range res.Message.ToolCalls {
-					res.Message.ToolCalls[i].Function.Name = functionNameMap.OriginalFromConverted(tool.Function.Name)
+				content, thinking, toolContent := harmonyMessageHandler.AddContent(r.Content, harmonyToolParser)
+				res.Message.Content = content
+				res.Message.Thinking = thinking
+				harmonyToolParser.Add(toolContent)
+
+				if r.Done {
+					toolName, toolContent := harmonyToolParser.Drain()
+					if toolName != nil {
+						*toolName = strings.TrimPrefix(*toolName, "functions.")
+						*toolName = harmonyMessageHandler.FunctionNameMap.OriginalFromConverted(*toolName)
+						var args api.ToolCallFunctionArguments
+						if err := json.Unmarshal([]byte(toolContent), &args); err != nil {
+							errStr := fmt.Sprintf("error parsing tool call: raw='%s', err=%s", toolContent, err.Error())
+							ch <- gin.H{"error": errStr}
+							return
+						}
+						res.Message.ToolCalls = []api.ToolCall{{Function: api.ToolCallFunction{Name: *toolName, Arguments: args}}}
+					}
 				}
+
 				// only send messages with meaningful content (empty messages confuse clients)
 				if res.Message.Content != "" || res.Message.Thinking != "" || len(res.Message.ToolCalls) > 0 || res.Done {
 					ch <- res
 				}
+
 				return
 			}
 

server/routes_harmony_streaming_test.go

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