add ability to turn on debug request logging

If `OLLAMA_DEBUG_LOG_REQUESTS` is set, then on server startup a temp
folder will be created. Upon any inference request, the body will be
logged to a file in this folder, as well as a small shell script to
"replay" the request using cURL.

This is just intended for debugging scenarios, not as something to turn
on normally.

anthropic/anthropic.go

@@ -518,26 +518,24 @@ func mapStopReason(reason string, hasToolCalls bool) string {
 
 // StreamConverter manages state for converting Ollama streaming responses to Anthropic format
 type StreamConverter struct {
-	ID                   string
-	Model                string
-	firstWrite           bool
-	contentIndex         int
-	inputTokens          int
-	outputTokens         int
-	estimatedInputTokens int // Estimated tokens from request (used when actual metrics are 0)
-	thinkingStarted      bool
-	thinkingDone         bool
-	textStarted          bool
-	toolCallsSent        map[string]bool
+	ID              string
+	Model           string
+	firstWrite      bool
+	contentIndex    int
+	inputTokens     int
+	outputTokens    int
+	thinkingStarted bool
+	thinkingDone    bool
+	textStarted     bool
+	toolCallsSent   map[string]bool
 }
 
-func NewStreamConverter(id, model string, estimatedInputTokens int) *StreamConverter {
+func NewStreamConverter(id, model string) *StreamConverter {
 	return &StreamConverter{
-		ID:                   id,
-		Model:                model,
-		firstWrite:           true,
-		estimatedInputTokens: estimatedInputTokens,
-		toolCallsSent:        make(map[string]bool),
+		ID:            id,
+		Model:         model,
+		firstWrite:    true,
+		toolCallsSent: make(map[string]bool),
 	}
 }
 
@@ -553,11 +551,7 @@ func (c *StreamConverter) Process(r api.ChatResponse) []StreamEvent {
 
 	if c.firstWrite {
 		c.firstWrite = false
-		// Use actual metrics if available, otherwise use estimate
 		c.inputTokens = r.Metrics.PromptEvalCount
-		if c.inputTokens == 0 && c.estimatedInputTokens > 0 {
-			c.inputTokens = c.estimatedInputTokens
-		}
 
 		events = append(events, StreamEvent{
 			Event: "message_start",
@@ -785,121 +779,3 @@ func mapToArgs(m map[string]any) api.ToolCallFunctionArguments {
 	}
 	return args
 }
-
-// CountTokensRequest represents an Anthropic count_tokens request
-type CountTokensRequest struct {
-	Model    string          `json:"model"`
-	Messages []MessageParam  `json:"messages"`
-	System   any             `json:"system,omitempty"`
-	Tools    []Tool          `json:"tools,omitempty"`
-	Thinking *ThinkingConfig `json:"thinking,omitempty"`
-}
-
-// EstimateInputTokens estimates input tokens from a MessagesRequest (reuses CountTokensRequest logic)
-func EstimateInputTokens(req MessagesRequest) int {
-	return estimateTokens(CountTokensRequest{
-		Model:    req.Model,
-		Messages: req.Messages,
-		System:   req.System,
-		Tools:    req.Tools,
-		Thinking: req.Thinking,
-	})
-}
-
-// CountTokensResponse represents an Anthropic count_tokens response
-type CountTokensResponse struct {
-	InputTokens int `json:"input_tokens"`
-}
-
-// estimateTokens returns a rough estimate of tokens (len/4)
-func estimateTokens(req CountTokensRequest) int {
-	var totalLen int
-
-	// Count system prompt
-	if req.System != nil {
-		totalLen += countAnyContent(req.System)
-	}
-
-	// Count messages
-	for _, msg := range req.Messages {
-		// Count role (always present)
-		totalLen += len(msg.Role)
-		// Count content
-		contentLen := countAnyContent(msg.Content)
-		totalLen += contentLen
-	}
-
-	for _, tool := range req.Tools {
-		totalLen += len(tool.Name) + len(tool.Description) + len(tool.InputSchema)
-	}
-
-	// Return len/4 as rough token estimate, minimum 1 if there's any content
-	tokens := totalLen / 4
-	if tokens == 0 && (len(req.Messages) > 0 || req.System != nil) {
-		tokens = 1
-	}
-	return tokens
-}
-
-func countAnyContent(content any) int {
-	if content == nil {
-		return 0
-	}
-
-	switch c := content.(type) {
-	case string:
-		return len(c)
-	case []any:
-		total := 0
-		for _, block := range c {
-			total += countContentBlock(block)
-		}
-		return total
-	default:
-		if data, err := json.Marshal(content); err == nil {
-			return len(data)
-		}
-		return 0
-	}
-}
-
-func countContentBlock(block any) int {
-	blockMap, ok := block.(map[string]any)
-	if !ok {
-		if s, ok := block.(string); ok {
-			return len(s)
-		}
-		return 0
-	}
-
-	total := 0
-	blockType, _ := blockMap["type"].(string)
-
-	if text, ok := blockMap["text"].(string); ok {
-		total += len(text)
-	}
-
-	if thinking, ok := blockMap["thinking"].(string); ok {
-		total += len(thinking)
-	}
-
-	if blockType == "tool_use" {
-		if data, err := json.Marshal(blockMap); err == nil {
-			total += len(data)
-		}
-	}
-
-	if blockType == "tool_result" {
-		if data, err := json.Marshal(blockMap); err == nil {
-			total += len(data)
-		}
-	}
-
-	if source, ok := blockMap["source"].(map[string]any); ok {
-		if data, ok := source["data"].(string); ok {
-			total += len(data)
-		}
-	}
-
-	return total
-}

anthropic/anthropic_test.go

@@ -605,7 +605,7 @@ func TestGenerateMessageID(t *testing.T) {
 }
 
 func TestStreamConverter_Basic(t *testing.T) {
-	conv := NewStreamConverter("msg_123", "test-model", 0)
+	conv := NewStreamConverter("msg_123", "test-model")
 
 	// First chunk
 	resp1 := api.ChatResponse{
@@ -678,7 +678,7 @@ func TestStreamConverter_Basic(t *testing.T) {
 }
 
 func TestStreamConverter_WithToolCalls(t *testing.T) {
-	conv := NewStreamConverter("msg_123", "test-model", 0)
+	conv := NewStreamConverter("msg_123", "test-model")
 
 	resp := api.ChatResponse{
 		Model: "test-model",
@@ -731,7 +731,7 @@ func TestStreamConverter_WithToolCalls(t *testing.T) {
 func TestStreamConverter_ToolCallWithUnmarshalableArgs(t *testing.T) {
 	// Test that unmarshalable arguments (like channels) are handled gracefully
 	// and don't cause a panic or corrupt stream
-	conv := NewStreamConverter("msg_123", "test-model", 0)
+	conv := NewStreamConverter("msg_123", "test-model")
 
 	// Create a channel which cannot be JSON marshaled
 	unmarshalable := make(chan int)
@@ -778,7 +778,7 @@ func TestStreamConverter_ToolCallWithUnmarshalableArgs(t *testing.T) {
 
 func TestStreamConverter_MultipleToolCallsWithMixedValidity(t *testing.T) {
 	// Test that valid tool calls still work when mixed with invalid ones
-	conv := NewStreamConverter("msg_123", "test-model", 0)
+	conv := NewStreamConverter("msg_123", "test-model")
 
 	unmarshalable := make(chan int)
 	badArgs := api.NewToolCallFunctionArguments()
@@ -903,7 +903,7 @@ func TestContentBlockJSON_EmptyFieldsPresent(t *testing.T) {
 // events include the required empty fields for SDK compatibility.
 func TestStreamConverter_ContentBlockStartIncludesEmptyFields(t *testing.T) {
 	t.Run("text block start includes empty text", func(t *testing.T) {
-		conv := NewStreamConverter("msg_123", "test-model", 0)
+		conv := NewStreamConverter("msg_123", "test-model")
 
 		resp := api.ChatResponse{
 			Model:   "test-model",
@@ -937,7 +937,7 @@ func TestStreamConverter_ContentBlockStartIncludesEmptyFields(t *testing.T) {
 	})
 
 	t.Run("thinking block start includes empty thinking", func(t *testing.T) {
-		conv := NewStreamConverter("msg_123", "test-model", 0)
+		conv := NewStreamConverter("msg_123", "test-model")
 
 		resp := api.ChatResponse{
 			Model:   "test-model",

api/client.go

@@ -466,15 +466,3 @@ func (c *Client) Whoami(ctx context.Context) (*UserResponse, error) {
 	}
 	return &resp, nil
 }
-
-// AliasRequest is the request body for creating or updating a model alias.
-type AliasRequest struct {
-	Alias          string `json:"alias"`
-	Target         string `json:"target"`
-	PrefixMatching bool   `json:"prefix_matching,omitempty"`
-}
-
-// SetAliasExperimental creates or updates a model alias via the experimental aliases API.
-func (c *Client) SetAliasExperimental(ctx context.Context, req *AliasRequest) error {
-	return c.do(ctx, http.MethodPost, "/api/experimental/aliases", req, nil)
-}

cmd/cmd.go

@@ -367,14 +367,25 @@ func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
 		return err
 	} else if info.RemoteHost != "" {
 		// Cloud model, no need to load/unload
+
+		isCloud := strings.HasPrefix(info.RemoteHost, "https://ollama.com")
+
+		// Check if user is signed in for ollama.com cloud models
+		if isCloud {
+			if _, err := client.Whoami(cmd.Context()); err != nil {
+				return err
+			}
+		}
+
 		if opts.ShowConnect {
 			p.StopAndClear()
-			if strings.HasPrefix(info.RemoteHost, "https://ollama.com") {
+			if isCloud {
 				fmt.Fprintf(os.Stderr, "Connecting to '%s' on 'ollama.com' ⚡\n", info.RemoteModel)
 			} else {
 				fmt.Fprintf(os.Stderr, "Connecting to '%s' on '%s'\n", info.RemoteModel, info.RemoteHost)
 			}
 		}
+
 		return nil
 	}
 

cmd/cmd_test.go

@@ -3,6 +3,7 @@ package cmd
 import (
 	"bytes"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"net/http"
@@ -1659,3 +1660,103 @@ func TestRunOptions_Copy_Independence(t *testing.T) {
 		t.Error("Copy Think should not be affected by original modification")
 	}
 }
+
+func TestLoadOrUnloadModel_CloudModelAuth(t *testing.T) {
+	tests := []struct {
+		name          string
+		remoteHost    string
+		whoamiStatus  int
+		whoamiResp    any
+		expectedError string
+	}{
+		{
+			name:         "ollama.com cloud model - user signed in",
+			remoteHost:   "https://ollama.com",
+			whoamiStatus: http.StatusOK,
+			whoamiResp:   api.UserResponse{Name: "testuser"},
+		},
+		{
+			name:         "ollama.com cloud model - user not signed in",
+			remoteHost:   "https://ollama.com",
+			whoamiStatus: http.StatusUnauthorized,
+			whoamiResp: map[string]string{
+				"error":      "unauthorized",
+				"signin_url": "https://ollama.com/signin",
+			},
+			expectedError: "unauthorized",
+		},
+		{
+			name:         "non-ollama.com remote - no auth check",
+			remoteHost:   "https://other-remote.com",
+			whoamiStatus: http.StatusUnauthorized, // should not be called
+			whoamiResp:   nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			whoamiCalled := false
+			mockServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				switch r.URL.Path {
+				case "/api/show":
+					w.Header().Set("Content-Type", "application/json")
+					if err := json.NewEncoder(w).Encode(api.ShowResponse{
+						RemoteHost:  tt.remoteHost,
+						RemoteModel: "test-model",
+					}); err != nil {
+						http.Error(w, err.Error(), http.StatusInternalServerError)
+					}
+				case "/api/me":
+					whoamiCalled = true
+					w.Header().Set("Content-Type", "application/json")
+					w.WriteHeader(tt.whoamiStatus)
+					if tt.whoamiResp != nil {
+						if err := json.NewEncoder(w).Encode(tt.whoamiResp); err != nil {
+							http.Error(w, err.Error(), http.StatusInternalServerError)
+						}
+					}
+				default:
+					http.NotFound(w, r)
+				}
+			}))
+			defer mockServer.Close()
+
+			t.Setenv("OLLAMA_HOST", mockServer.URL)
+
+			cmd := &cobra.Command{}
+			cmd.SetContext(t.Context())
+
+			opts := &runOptions{
+				Model:       "test-cloud-model",
+				ShowConnect: false,
+			}
+
+			err := loadOrUnloadModel(cmd, opts)
+
+			if strings.HasPrefix(tt.remoteHost, "https://ollama.com") {
+				if !whoamiCalled {
+					t.Error("expected whoami to be called for ollama.com cloud model")
+				}
+			} else {
+				if whoamiCalled {
+					t.Error("whoami should not be called for non-ollama.com remote")
+				}
+			}
+
+			if tt.expectedError != "" {
+				if err == nil {
+					t.Errorf("expected error containing %q, got nil", tt.expectedError)
+				} else {
+					var authErr api.AuthorizationError
+					if !errors.As(err, &authErr) {
+						t.Errorf("expected AuthorizationError, got %T: %v", err, err)
+					}
+				}
+			} else {
+				if err != nil {
+					t.Errorf("expected no error, got %v", err)
+				}
+			}
+		})
+	}
+}

cmd/config/claude.go

@@ -1,23 +1,18 @@
 package config
 
 import (
-	"context"
 	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"runtime"
 
-	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 )
 
-// Claude implements Runner and AliasConfigurer for Claude Code integration
+// Claude implements Runner for Claude Code integration
 type Claude struct{}
 
-// Compile-time check that Claude implements AliasConfigurer
-var _ AliasConfigurer = (*Claude)(nil)
-
 func (c *Claude) String() string { return "Claude Code" }
 
 func (c *Claude) args(model string, extra []string) []string {
@@ -65,96 +60,3 @@ func (c *Claude) Run(model string, args []string) error {
 	)
 	return cmd.Run()
 }
-
-// ConfigureAliases sets up Primary and Fast model aliases for Claude Code.
-func (c *Claude) ConfigureAliases(ctx context.Context, primaryModel string, existing map[string]string, force bool) (map[string]string, bool, error) {
-	aliases := make(map[string]string)
-	for k, v := range existing {
-		aliases[k] = v
-	}
-
-	if primaryModel != "" {
-		aliases["primary"] = primaryModel
-	}
-
-	if !force && aliases["primary"] != "" && aliases["fast"] != "" {
-		return aliases, false, nil
-	}
-
-	items, existingModels, cloudModels, client, err := listModels(ctx)
-	if err != nil {
-		return nil, false, err
-	}
-
-	fmt.Fprintf(os.Stderr, "\n%sModel Configuration%s\n", ansiBold, ansiReset)
-	fmt.Fprintf(os.Stderr, "%sClaude Code uses multiple models for various tasks%s\n\n", ansiGray, ansiReset)
-
-	fmt.Fprintf(os.Stderr, "%sPrimary%s\n", ansiBold, ansiReset)
-	fmt.Fprintf(os.Stderr, "%sHandles complex reasoning: planning, code generation, debugging.%s\n\n", ansiGray, ansiReset)
-
-	if aliases["primary"] == "" || force {
-		primary, err := selectPrompt("Select Primary model:", items)
-		if err != nil {
-			return nil, false, err
-		}
-		if err := pullIfNeeded(ctx, client, existingModels, primary); err != nil {
-			return nil, false, err
-		}
-		if err := ensureAuth(ctx, client, cloudModels, []string{primary}); err != nil {
-			return nil, false, err
-		}
-		aliases["primary"] = primary
-	} else {
-		fmt.Fprintf(os.Stderr, "  %s\n\n", aliases["primary"])
-	}
-
-	fmt.Fprintf(os.Stderr, "%sFast%s\n", ansiBold, ansiReset)
-	fmt.Fprintf(os.Stderr, "%sHandles quick operations: file searches, simple edits, status checks.%s\n", ansiGray, ansiReset)
-	fmt.Fprintf(os.Stderr, "%sSmaller models work well and respond faster.%s\n\n", ansiGray, ansiReset)
-
-	if aliases["fast"] == "" || force {
-		fast, err := selectPrompt("Select Fast model:", items)
-		if err != nil {
-			return nil, false, err
-		}
-		if err := pullIfNeeded(ctx, client, existingModels, fast); err != nil {
-			return nil, false, err
-		}
-		if err := ensureAuth(ctx, client, cloudModels, []string{fast}); err != nil {
-			return nil, false, err
-		}
-		aliases["fast"] = fast
-	}
-
-	return aliases, true, nil
-}
-
-// SetAliases syncs the configured aliases to the Ollama server using prefix matching.
-func (c *Claude) SetAliases(ctx context.Context, aliases map[string]string) error {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
-	prefixAliases := map[string]string{
-		"claude-sonnet-": aliases["primary"],
-		"claude-haiku-":  aliases["fast"],
-	}
-
-	var errs []string
-	for prefix, target := range prefixAliases {
-		req := &api.AliasRequest{
-			Alias:          prefix,
-			Target:         target,
-			PrefixMatching: true,
-		}
-		if err := client.SetAliasExperimental(ctx, req); err != nil {
-			errs = append(errs, prefix)
-		}
-	}
-
-	if len(errs) > 0 {
-		return fmt.Errorf("failed to set aliases: %v", errs)
-	}
-	return nil
-}

cmd/config/config.go

@@ -13,8 +13,7 @@ import (
 )
 
 type integration struct {
-	Models  []string          `json:"models"`
-	Aliases map[string]string `json:"aliases,omitempty"`
+	Models []string `json:"models"`
 }
 
 type config struct {
@@ -134,16 +133,8 @@ func saveIntegration(appName string, models []string) error {
 		return err
 	}
 
-	key := strings.ToLower(appName)
-	existing := cfg.Integrations[key]
-	var aliases map[string]string
-	if existing != nil && existing.Aliases != nil {
-		aliases = existing.Aliases
-	}
-
-	cfg.Integrations[key] = &integration{
-		Models:  models,
-		Aliases: aliases,
+	cfg.Integrations[strings.ToLower(appName)] = &integration{
+		Models: models,
 	}
 
 	return save(cfg)
@@ -163,33 +154,6 @@ func loadIntegration(appName string) (*integration, error) {
 	return ic, nil
 }
 
-func saveAliases(appName string, aliases map[string]string) error {
-	if appName == "" {
-		return errors.New("app name cannot be empty")
-	}
-
-	cfg, err := load()
-	if err != nil {
-		return err
-	}
-
-	key := strings.ToLower(appName)
-	existing := cfg.Integrations[key]
-	if existing == nil {
-		existing = &integration{}
-	}
-
-	if existing.Aliases == nil {
-		existing.Aliases = make(map[string]string)
-	}
-	for k, v := range aliases {
-		existing.Aliases[k] = v
-	}
-
-	cfg.Integrations[key] = existing
-	return save(cfg)
-}
-
 func listIntegrations() ([]integration, error) {
 	cfg, err := load()
 	if err != nil {

cmd/config/config_test.go

@@ -46,53 +46,6 @@ func TestIntegrationConfig(t *testing.T) {
 		}
 	})
 
-	t.Run("save and load aliases", func(t *testing.T) {
-		models := []string{"llama3.2"}
-		if err := saveIntegration("claude", models); err != nil {
-			t.Fatal(err)
-		}
-		aliases := map[string]string{
-			"primary": "llama3.2:70b",
-			"fast":    "llama3.2:8b",
-		}
-		if err := saveAliases("claude", aliases); err != nil {
-			t.Fatal(err)
-		}
-
-		config, err := loadIntegration("claude")
-		if err != nil {
-			t.Fatal(err)
-		}
-		if config.Aliases == nil {
-			t.Fatal("expected aliases to be saved")
-		}
-		for k, v := range aliases {
-			if config.Aliases[k] != v {
-				t.Errorf("alias %s: expected %s, got %s", k, v, config.Aliases[k])
-			}
-		}
-	})
-
-	t.Run("saveIntegration preserves aliases", func(t *testing.T) {
-		if err := saveIntegration("claude", []string{"model-a"}); err != nil {
-			t.Fatal(err)
-		}
-		if err := saveAliases("claude", map[string]string{"primary": "model-a", "fast": "model-small"}); err != nil {
-			t.Fatal(err)
-		}
-
-		if err := saveIntegration("claude", []string{"model-b"}); err != nil {
-			t.Fatal(err)
-		}
-		config, err := loadIntegration("claude")
-		if err != nil {
-			t.Fatal(err)
-		}
-		if config.Aliases["primary"] != "model-a" {
-			t.Errorf("expected aliases to be preserved, got %v", config.Aliases)
-		}
-	})
-
 	t.Run("defaultModel returns first model", func(t *testing.T) {
 		saveIntegration("codex", []string{"model-a", "model-b"})
 

cmd/config/integrations.go

@@ -39,15 +39,6 @@ type Editor interface {
 	Models() []string
 }
 
-// AliasConfigurer can configure model aliases (e.g., for subagent routing).
-// Integrations like Claude and Codex use this to route model requests to local models.
-type AliasConfigurer interface {
-	// ConfigureAliases prompts the user to configure aliases and returns the updated map.
-	ConfigureAliases(ctx context.Context, primaryModel string, existing map[string]string, force bool) (map[string]string, bool, error)
-	// SetAliases syncs the configured aliases to the server
-	SetAliases(ctx context.Context, aliases map[string]string) error
-}
-
 // integrations is the registry of available integrations.
 var integrations = map[string]Runner{
 	"claude":   &Claude{},
@@ -138,11 +129,7 @@ func selectModels(ctx context.Context, name, current string) ([]string, error) {
 			return nil, err
 		}
 	} else {
-		prompt := fmt.Sprintf("Select model for %s:", r)
-		if _, ok := r.(AliasConfigurer); ok {
-			prompt = fmt.Sprintf("Select Primary model for %s:", r)
-		}
-		model, err := selectPrompt(prompt, items)
+		model, err := selectPrompt(fmt.Sprintf("Select model for %s:", r), items)
 		if err != nil {
 			return nil, err
 		}
@@ -170,146 +157,73 @@ func selectModels(ctx context.Context, name, current string) ([]string, error) {
 		}
 	}
 
-	if err := ensureAuth(ctx, client, cloudModels, selected); err != nil {
-		return nil, err
-	}
-
-	return selected, nil
-}
-
-func pullIfNeeded(ctx context.Context, client *api.Client, existingModels map[string]bool, model string) error {
-	if existingModels[model] {
-		return nil
-	}
-	msg := fmt.Sprintf("Download %s?", model)
-	if ok, err := confirmPrompt(msg); err != nil {
-		return err
-	} else if !ok {
-		return errCancelled
-	}
-	fmt.Fprintf(os.Stderr, "\n")
-	if err := pullModel(ctx, client, model); err != nil {
-		return fmt.Errorf("failed to pull %s: %w", model, err)
-	}
-	return nil
-}
-
-func listModels(ctx context.Context) ([]selectItem, map[string]bool, map[string]bool, *api.Client, error) {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return nil, nil, nil, nil, err
-	}
-
-	models, err := client.List(ctx)
-	if err != nil {
-		return nil, nil, nil, nil, err
-	}
-
-	var existing []modelInfo
-	for _, m := range models.Models {
-		existing = append(existing, modelInfo{Name: m.Name, Remote: m.RemoteModel != ""})
-	}
-
-	items, _, existingModels, cloudModels := buildModelList(existing, nil, "")
-
-	if len(items) == 0 {
-		return nil, nil, nil, nil, fmt.Errorf("no models available, run 'ollama pull <model>' first")
-	}
-
-	return items, existingModels, cloudModels, client, nil
-}
-
-func ensureAuth(ctx context.Context, client *api.Client, cloudModels map[string]bool, selected []string) error {
 	var selectedCloudModels []string
 	for _, m := range selected {
 		if cloudModels[m] {
 			selectedCloudModels = append(selectedCloudModels, m)
 		}
 	}
-	if len(selectedCloudModels) == 0 {
-		return nil
-	}
+	if len(selectedCloudModels) > 0 {
+		// ensure user is signed in
+		user, err := client.Whoami(ctx)
+		if err == nil && user != nil && user.Name != "" {
+			return selected, nil
+		}
 
-	user, err := client.Whoami(ctx)
-	if err == nil && user != nil && user.Name != "" {
-		return nil
-	}
+		var aErr api.AuthorizationError
+		if !errors.As(err, &aErr) || aErr.SigninURL == "" {
+			return nil, err
+		}
 
-	var aErr api.AuthorizationError
-	if !errors.As(err, &aErr) || aErr.SigninURL == "" {
-		return err
-	}
+		modelList := strings.Join(selectedCloudModels, ", ")
+		yes, err := confirmPrompt(fmt.Sprintf("sign in to use %s?", modelList))
+		if err != nil || !yes {
+			return nil, fmt.Errorf("%s requires sign in", modelList)
+		}
 
-	modelList := strings.Join(selectedCloudModels, ", ")
-	yes, err := confirmPrompt(fmt.Sprintf("sign in to use %s?", modelList))
-	if err != nil || !yes {
-		return fmt.Errorf("%s requires sign in", modelList)
-	}
+		fmt.Fprintf(os.Stderr, "\nTo sign in, navigate to:\n    %s\n\n", aErr.SigninURL)
 
-	fmt.Fprintf(os.Stderr, "\nTo sign in, navigate to:\n    %s\n\n", aErr.SigninURL)
+		// TODO(parthsareen): extract into auth package for cmd
+		// Auto-open browser (best effort, fail silently)
+		switch runtime.GOOS {
+		case "darwin":
+			_ = exec.Command("open", aErr.SigninURL).Start()
+		case "linux":
+			_ = exec.Command("xdg-open", aErr.SigninURL).Start()
+		case "windows":
+			_ = exec.Command("rundll32", "url.dll,FileProtocolHandler", aErr.SigninURL).Start()
+		}
 
-	switch runtime.GOOS {
-	case "darwin":
-		_ = exec.Command("open", aErr.SigninURL).Start()
-	case "linux":
-		_ = exec.Command("xdg-open", aErr.SigninURL).Start()
-	case "windows":
-		_ = exec.Command("rundll32", "url.dll,FileProtocolHandler", aErr.SigninURL).Start()
-	}
+		spinnerFrames := []string{"|", "/", "-", "\\"}
+		frame := 0
 
-	spinnerFrames := []string{"|", "/", "-", "\\"}
-	frame := 0
+		fmt.Fprintf(os.Stderr, "\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[0])
 
-	fmt.Fprintf(os.Stderr, "\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[0])
+		ticker := time.NewTicker(200 * time.Millisecond)
+		defer ticker.Stop()
 
-	ticker := time.NewTicker(200 * time.Millisecond)
-	defer ticker.Stop()
+		for {
+			select {
+			case <-ctx.Done():
+				fmt.Fprintf(os.Stderr, "\r\033[K")
+				return nil, ctx.Err()
+			case <-ticker.C:
+				frame++
+				fmt.Fprintf(os.Stderr, "\r\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[frame%len(spinnerFrames)])
 
-	for {
-		select {
-		case <-ctx.Done():
-			fmt.Fprintf(os.Stderr, "\r\033[K")
-			return ctx.Err()
-		case <-ticker.C:
-			frame++
-			fmt.Fprintf(os.Stderr, "\r\033[90mwaiting for sign in to complete... %s\033[0m", spinnerFrames[frame%len(spinnerFrames)])
-
-			// poll every 10th frame (~2 seconds)
-			if frame%10 == 0 {
-				u, err := client.Whoami(ctx)
-				if err == nil && u != nil && u.Name != "" {
-					fmt.Fprintf(os.Stderr, "\r\033[K\033[A\r\033[K\033[1msigned in:\033[0m %s\n", u.Name)
-					return nil
+				// poll every 10th frame (~2 seconds)
+				if frame%10 == 0 {
+					u, err := client.Whoami(ctx)
+					if err == nil && u != nil && u.Name != "" {
+						fmt.Fprintf(os.Stderr, "\r\033[K\033[A\r\033[K\033[1msigned in:\033[0m %s\n", u.Name)
+						return selected, nil
+					}
 				}
 			}
 		}
 	}
-}
 
-func ensureAliases(ctx context.Context, r Runner, name string, primaryModel string, existing map[string]string, force bool) (bool, error) {
-	ac, ok := r.(AliasConfigurer)
-	if !ok {
-		return false, nil
-	}
-
-	aliases, updated, err := ac.ConfigureAliases(ctx, primaryModel, existing, force)
-	if err != nil {
-		return false, err
-	}
-	if !updated {
-		return false, nil
-	}
-
-	if err := saveAliases(name, aliases); err != nil {
-		return false, err
-	}
-
-	if err := ac.SetAliases(ctx, aliases); err != nil {
-		fmt.Fprintf(os.Stderr, "%sWarning: Could not sync aliases to server: %v%s\n", ansiGray, err, ansiReset)
-		fmt.Fprintf(os.Stderr, "%sAliases saved locally. Server sync will retry on next launch.%s\n\n", ansiGray, ansiReset)
-	}
-
-	return true, nil
+	return selected, nil
 }
 
 func runIntegration(name, modelName string, args []string) error {
@@ -317,17 +231,6 @@ func runIntegration(name, modelName string, args []string) error {
 	if !ok {
 		return fmt.Errorf("unknown integration: %s", name)
 	}
-
-	if _, ok := r.(AliasConfigurer); ok {
-		if config, err := loadIntegration(name); err == nil && config.Aliases != nil {
-			primary, fast := config.Aliases["primary"], config.Aliases["fast"]
-			if primary != "" && fast != "" {
-				fmt.Fprintf(os.Stderr, "\nLaunching %s with Primary: %s, Fast: %s...\n", r, primary, fast)
-				return r.Run(modelName, args)
-			}
-		}
-	}
-
 	fmt.Fprintf(os.Stderr, "\nLaunching %s with %s...\n", r, modelName)
 	return r.Run(modelName, args)
 }
@@ -401,50 +304,10 @@ Examples:
 
 			if !configFlag && modelFlag == "" {
 				if config, err := loadIntegration(name); err == nil && len(config.Models) > 0 {
-					if _, err := ensureAliases(cmd.Context(), r, name, config.Models[0], config.Aliases, false); errors.Is(err, errCancelled) {
-						return nil
-					} else if err != nil {
-						return err
-					}
 					return runIntegration(name, config.Models[0], passArgs)
 				}
 			}
 
-			if ac, ok := r.(AliasConfigurer); ok {
-				var existingAliases map[string]string
-				if existing, err := loadIntegration(name); err == nil {
-					existingAliases = existing.Aliases
-				}
-				aliases, updated, err := ac.ConfigureAliases(cmd.Context(), "", existingAliases, configFlag)
-				if errors.Is(err, errCancelled) {
-					return nil
-				}
-				if err != nil {
-					return err
-				}
-				if updated {
-					if err := saveAliases(name, aliases); err != nil {
-						return err
-					}
-					if err := ac.SetAliases(cmd.Context(), aliases); err != nil {
-						fmt.Fprintf(os.Stderr, "%sWarning: Could not sync aliases to server: %v%s\n", ansiGray, err, ansiReset)
-					}
-					fmt.Fprintf(os.Stderr, "\n%sConfiguration Complete%s\n", ansiBold, ansiReset)
-					fmt.Fprintf(os.Stderr, "Primary: %s\n", aliases["primary"])
-					fmt.Fprintf(os.Stderr, "Fast: %s\n\n", aliases["fast"])
-				}
-				if err := saveIntegration(name, []string{aliases["primary"]}); err != nil {
-					return fmt.Errorf("failed to save: %w", err)
-				}
-				if configFlag {
-					if launch, _ := confirmPrompt(fmt.Sprintf("Launch %s now?", r)); launch {
-						return runIntegration(name, aliases["primary"], passArgs)
-					}
-					return nil
-				}
-				return runIntegration(name, aliases["primary"], passArgs)
-			}
-
 			var models []string
 			if modelFlag != "" {
 				models = []string{modelFlag}

cmd/config/integrations_test.go

@@ -509,19 +509,3 @@ func TestBuildModelList_ReturnsExistingAndCloudMaps(t *testing.T) {
 		t.Error("llama3.2 should not be in cloudModels")
 	}
 }
-
-func TestAliasConfigurerInterface(t *testing.T) {
-	t.Run("claude implements AliasConfigurer", func(t *testing.T) {
-		claude := &Claude{}
-		if _, ok := interface{}(claude).(AliasConfigurer); !ok {
-			t.Error("Claude should implement AliasConfigurer")
-		}
-	})
-
-	t.Run("codex does not implement AliasConfigurer", func(t *testing.T) {
-		codex := &Codex{}
-		if _, ok := interface{}(codex).(AliasConfigurer); ok {
-			t.Error("Codex should not implement AliasConfigurer")
-		}
-	})
-}

cmd/config/selector.go

@@ -65,10 +65,6 @@ func (s *selectState) handleInput(event inputEvent, char byte) (done bool, resul
 		if len(filtered) > 0 && s.selected < len(filtered) {
 			return true, filtered[s.selected].Name, nil
 		}
-		// No matches but user typed something - return filter for pull prompt
-		if len(filtered) == 0 && s.filter != "" {
-			return true, s.filter, nil
-		}
 	case eventEscape:
 		return true, "", errCancelled
 	case eventBackspace:
@@ -287,11 +283,7 @@ func renderSelect(w io.Writer, prompt string, s *selectState) int {
 	lineCount := 1
 
 	if len(filtered) == 0 {
-		if s.filter != "" {
-			fmt.Fprintf(w, "  %s→ Download model: '%s'? Press Enter%s\r\n", ansiGray, s.filter, ansiReset)
-		} else {
-			fmt.Fprintf(w, "  %s(no matches)%s\r\n", ansiGray, ansiReset)
-		}
+		fmt.Fprintf(w, "  %s(no matches)%s\r\n", ansiGray, ansiReset)
 		lineCount++
 	} else {
 		displayCount := min(len(filtered), maxDisplayedItems)

cmd/config/selector_test.go

@@ -87,18 +87,10 @@ func TestSelectState(t *testing.T) {
 		}
 	})
 
-	t.Run("Enter_EmptyFilteredList_ReturnsFilter", func(t *testing.T) {
+	t.Run("Enter_EmptyFilteredList_DoesNothing", func(t *testing.T) {
 		s := newSelectState(items)
 		s.filter = "nonexistent"
 		done, result, err := s.handleInput(eventEnter, 0)
-		if !done || result != "nonexistent" || err != nil {
-			t.Errorf("expected (true, 'nonexistent', nil), got (%v, %v, %v)", done, result, err)
-		}
-	})
-
-	t.Run("Enter_EmptyFilteredList_EmptyFilter_DoesNothing", func(t *testing.T) {
-		s := newSelectState([]selectItem{})
-		done, result, err := s.handleInput(eventEnter, 0)
 		if done || result != "" || err != nil {
 			t.Errorf("expected (false, '', nil), got (%v, %v, %v)", done, result, err)
 		}
@@ -576,25 +568,14 @@ func TestRenderSelect(t *testing.T) {
 		}
 	})
 
-	t.Run("EmptyFilteredList_ShowsPullPrompt", func(t *testing.T) {
+	t.Run("EmptyFilteredList_ShowsNoMatches", func(t *testing.T) {
 		s := newSelectState(items)
 		s.filter = "xyz"
 		var buf bytes.Buffer
 		renderSelect(&buf, "Select:", s)
 
-		output := buf.String()
-		if !strings.Contains(output, "Download model: 'xyz'?") {
-			t.Errorf("expected 'Download model: xyz?' message, got: %s", output)
-		}
-	})
-
-	t.Run("EmptyFilteredList_EmptyFilter_ShowsNoMatches", func(t *testing.T) {
-		s := newSelectState([]selectItem{})
-		var buf bytes.Buffer
-		renderSelect(&buf, "Select:", s)
-
 		if !strings.Contains(buf.String(), "no matches") {
-			t.Error("expected 'no matches' message for empty list with no filter")
+			t.Error("expected 'no matches' message")
 		}
 	})
 

envconfig/config.go

@@ -188,6 +188,8 @@ func LogLevel() slog.Level {
 var (
 	// FlashAttention enables the experimental flash attention feature.
 	FlashAttention = BoolWithDefault("OLLAMA_FLASH_ATTENTION")
+	// DebugLogRequests logs inference requests to disk for replay/debugging.
+	DebugLogRequests = Bool("OLLAMA_DEBUG_LOG_REQUESTS")
 	// KvCacheType is the quantization type for the K/V cache.
 	KvCacheType = String("OLLAMA_KV_CACHE_TYPE")
 	// NoHistory disables readline history.
@@ -273,26 +275,27 @@ type EnvVar struct {
 
 func AsMap() map[string]EnvVar {
 	ret := map[string]EnvVar{
-		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", LogLevel(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
-		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(false), "Enabled flash attention"},
-		"OLLAMA_KV_CACHE_TYPE":     {"OLLAMA_KV_CACHE_TYPE", KvCacheType(), "Quantization type for the K/V cache (default: f16)"},
-		"OLLAMA_GPU_OVERHEAD":      {"OLLAMA_GPU_OVERHEAD", GpuOverhead(), "Reserve a portion of VRAM per GPU (bytes)"},
-		"OLLAMA_HOST":              {"OLLAMA_HOST", Host(), "IP Address for the ollama server (default 127.0.0.1:11434)"},
-		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive(), "The duration that models stay loaded in memory (default \"5m\")"},
-		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_LLM_LIBRARY", LLMLibrary(), "Set LLM library to bypass autodetection"},
-		"OLLAMA_LOAD_TIMEOUT":      {"OLLAMA_LOAD_TIMEOUT", LoadTimeout(), "How long to allow model loads to stall before giving up (default \"5m\")"},
-		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners(), "Maximum number of loaded models per GPU"},
-		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueue(), "Maximum number of queued requests"},
-		"OLLAMA_MODELS":            {"OLLAMA_MODELS", Models(), "The path to the models directory"},
-		"OLLAMA_NOHISTORY":         {"OLLAMA_NOHISTORY", NoHistory(), "Do not preserve readline history"},
-		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune(), "Do not prune model blobs on startup"},
-		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel(), "Maximum number of parallel requests"},
-		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowedOrigins(), "A comma separated list of allowed origins"},
-		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
-		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
-		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4k/32k/256k based on VRAM)"},
-		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
-		"OLLAMA_REMOTES":           {"OLLAMA_REMOTES", Remotes(), "Allowed hosts for remote models (default \"ollama.com\")"},
+		"OLLAMA_DEBUG":              {"OLLAMA_DEBUG", LogLevel(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
+		"OLLAMA_DEBUG_LOG_REQUESTS": {"OLLAMA_DEBUG_LOG_REQUESTS", DebugLogRequests(), "Log inference request bodies and replay curl commands to a temp directory"},
+		"OLLAMA_FLASH_ATTENTION":    {"OLLAMA_FLASH_ATTENTION", FlashAttention(false), "Enabled flash attention"},
+		"OLLAMA_KV_CACHE_TYPE":      {"OLLAMA_KV_CACHE_TYPE", KvCacheType(), "Quantization type for the K/V cache (default: f16)"},
+		"OLLAMA_GPU_OVERHEAD":       {"OLLAMA_GPU_OVERHEAD", GpuOverhead(), "Reserve a portion of VRAM per GPU (bytes)"},
+		"OLLAMA_HOST":               {"OLLAMA_HOST", Host(), "IP Address for the ollama server (default 127.0.0.1:11434)"},
+		"OLLAMA_KEEP_ALIVE":         {"OLLAMA_KEEP_ALIVE", KeepAlive(), "The duration that models stay loaded in memory (default \"5m\")"},
+		"OLLAMA_LLM_LIBRARY":        {"OLLAMA_LLM_LIBRARY", LLMLibrary(), "Set LLM library to bypass autodetection"},
+		"OLLAMA_LOAD_TIMEOUT":       {"OLLAMA_LOAD_TIMEOUT", LoadTimeout(), "How long to allow model loads to stall before giving up (default \"5m\")"},
+		"OLLAMA_MAX_LOADED_MODELS":  {"OLLAMA_MAX_LOADED_MODELS", MaxRunners(), "Maximum number of loaded models per GPU"},
+		"OLLAMA_MAX_QUEUE":          {"OLLAMA_MAX_QUEUE", MaxQueue(), "Maximum number of queued requests"},
+		"OLLAMA_MODELS":             {"OLLAMA_MODELS", Models(), "The path to the models directory"},
+		"OLLAMA_NOHISTORY":          {"OLLAMA_NOHISTORY", NoHistory(), "Do not preserve readline history"},
+		"OLLAMA_NOPRUNE":            {"OLLAMA_NOPRUNE", NoPrune(), "Do not prune model blobs on startup"},
+		"OLLAMA_NUM_PARALLEL":       {"OLLAMA_NUM_PARALLEL", NumParallel(), "Maximum number of parallel requests"},
+		"OLLAMA_ORIGINS":            {"OLLAMA_ORIGINS", AllowedOrigins(), "A comma separated list of allowed origins"},
+		"OLLAMA_SCHED_SPREAD":       {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
+		"OLLAMA_MULTIUSER_CACHE":    {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
+		"OLLAMA_CONTEXT_LENGTH":     {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4k/32k/256k based on VRAM)"},
+		"OLLAMA_NEW_ENGINE":         {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
+		"OLLAMA_REMOTES":            {"OLLAMA_REMOTES", Remotes(), "Allowed hosts for remote models (default \"ollama.com\")"},
 
 		// Informational
 		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},

integration/basic_test.go

@@ -144,3 +144,47 @@ func TestUnicodeModelDir(t *testing.T) {
 	}
 	ChatTestHelper(ctx, t, req, blueSkyExpected)
 }
+
+// TestNumPredict verifies that when num_predict is set, the model generates
+// exactly that many tokens. It uses logprobs to count the actual tokens output.
+func TestNumPredict(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	if err := PullIfMissing(ctx, client, "qwen3:0.6b"); err != nil {
+		t.Fatalf("failed to pull model: %v", err)
+	}
+
+	req := api.GenerateRequest{
+		Model:    "qwen3:0.6b",
+		Prompt:   "Write a long story.",
+		Stream:   &stream,
+		Logprobs: true,
+		Options: map[string]any{
+			"num_predict": 10,
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+
+	logprobCount := 0
+	var finalResponse api.GenerateResponse
+	err := client.Generate(ctx, &req, func(resp api.GenerateResponse) error {
+		logprobCount += len(resp.Logprobs)
+		if resp.Done {
+			finalResponse = resp
+		}
+		return nil
+	})
+	if err != nil {
+		t.Fatalf("generate failed: %v", err)
+	}
+
+	if logprobCount != 10 {
+		t.Errorf("expected 10 tokens (logprobs), got %d (EvalCount=%d, DoneReason=%s)",
+			logprobCount, finalResponse.EvalCount, finalResponse.DoneReason)
+	}
+}

middleware/anthropic.go

@@ -131,15 +131,12 @@ func AnthropicMessagesMiddleware() gin.HandlerFunc {
 
 		messageID := anthropic.GenerateMessageID()
 
-		// Estimate input tokens for streaming (actual count not available until generation completes)
-		estimatedTokens := anthropic.EstimateInputTokens(req)
-
 		w := &AnthropicWriter{
 			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
 			stream:     req.Stream,
 			id:         messageID,
 			model:      req.Model,
-			converter:  anthropic.NewStreamConverter(messageID, req.Model, estimatedTokens),
+			converter:  anthropic.NewStreamConverter(messageID, req.Model),
 		}
 
 		if req.Stream {

ml/backend.go

@@ -175,6 +175,7 @@ type Tensor interface {
 	SILU(ctx Context, up ...Tensor) Tensor
 	RELU(ctx Context, up ...Tensor) Tensor
 	Sigmoid(ctx Context) Tensor
+	SigmoidOut(ctx Context) Tensor
 
 	// AlphaLimitSILU is a variant of SILU that clamps the input to the range [-limit, limit]
 	SILUAlphaLimit(ctx Context, up Tensor, alpha, limit float32) Tensor

ml/backend/ggml/ggml.go

@@ -1468,6 +1468,13 @@ func (t *Tensor) Sigmoid(ctx ml.Context) ml.Tensor {
 	}
 }
 
+func (t *Tensor) SigmoidOut(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_sigmoid(ctx.(*Context).ctx, t.t),
+	}
+}
+
 func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	switch len(shape) {
 	case 1:

model/models/qwen3next/deltanet.go

@@ -406,8 +406,10 @@ func (gdn *GatedDeltaNet) deltaNetChunked(
 	gDiff := gCumsum.Neg(ctx).Add(ctx, gLast)
 	gDiffExp := gDiff.Exp(ctx)
 
-	// key_gdiff = k * exp(g_diff)
-	keyGDiff := k.Mul(ctx, gDiffExp)
+	// Reshapes g_diff_exp to [1, chunkSize, nChunks, ...]
+	gDiffExpReshaped := gDiffExp.Reshape(ctx, 1, chunkSize, nChunks, numVHeads*nSeqs)
+	keyGDiff := k.Mul(ctx, gDiffExpReshaped)
+	keyGDiffT := keyGDiff.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
 
 	// Process chunks and update state
 	var coreAttnOut ml.Tensor
@@ -444,12 +446,9 @@ func (gdn *GatedDeltaNet) deltaNetChunked(
 			coreAttnOut = coreAttnOut.Concat(ctx, coreAttnOutChunk, 1)
 		}
 
-		// Update state for next chunk using pre-computed values
+		// Update state for next chunk
 		gExpLastChunk := gLastExp.Slice(ctx, 2, chunk, chunk+1, 1)
-		kGDiffChunk := keyGDiff.Slice(ctx, 2, chunk, chunk+1, 1)
-
-		// kgdmulvnew = key_gdiff^T @ v_new
-		kGDiffChunkT := kGDiffChunk.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
+		kGDiffChunkT := keyGDiffT.Slice(ctx, 2, chunk, chunk+1, 1)
 		kgdMulVNew := vNewT.Mulmat(ctx, kGDiffChunkT)
 
 		// state = state * g_last + kgdmulvnew

model/models/qwen3next/model.go

@@ -135,7 +135,7 @@ func (mlp *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options
 		// Apply shared expert gating
 		if mlp.SharedGateInp != nil {
 			sharedGateVal := mlp.SharedGateInp.Forward(ctx, hiddenStates2D)
-			sharedGateVal = sharedGateVal.Sigmoid(ctx)
+			sharedGateVal = sharedGateVal.SigmoidOut(ctx)
 			// Broadcast gate to match dimensions
 			sharedGateVal = sharedGateVal.Repeat(ctx, 0, sharedOut.Dim(0))
 			sharedOut = sharedOut.Mul(ctx, sharedGateVal)

runner/ollamarunner/runner.go

@@ -514,13 +514,6 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
 			continue
 		}
 
-		// if past the num predict limit
-		if seq.numPredict > 0 && seq.numPredicted >= seq.numPredict {
-			s.removeSequence(seqIdx, llm.DoneReasonLength)
-			nextBatch.seqs[seqIdx] = nil
-			continue
-		}
-
 		if !s.cache.enabled {
 			seq.inputs = append(seq.cache.Inputs, seq.inputs...)
 			seq.cache.Inputs = []*input.Input{}
@@ -709,7 +702,6 @@ func (s *Server) computeBatch(activeBatch batchState) {
 			continue
 		}
 
-		seq.numPredicted++
 		nextToken := &input.Input{Token: 0} // placeholder we'll fill in after Compute/Floats
 		seq.inputs = []*input.Input{nextToken}
 		nextBatchTokens[i] = nextToken
@@ -740,8 +732,14 @@ func (s *Server) computeBatch(activeBatch batchState) {
 		if seq == nil || nextBatchTokens[i] == nil {
 			continue
 		}
+		// If the sequence was replaced while this batch was computing, discard results.
+		if activeBatch.seqs[i] != seq {
+			logutil.Trace("computeBatch: sequence replaced, discarding its results", "batchID", activeBatch.id, "seqIdx", i)
+			continue
+		}
 
 		seq.lastUpdatedAt = t
+		seq.numPredicted++
 		if seq.numPredicted == 1 {
 			seq.processingDuration = seq.lastUpdatedAt.Sub(seq.startedAt)
 			seq.startedAt = seq.lastUpdatedAt
@@ -787,6 +785,13 @@ func (s *Server) computeBatch(activeBatch batchState) {
 		}
 
 		seq.pendingResponses = append(seq.pendingResponses, piece)
+
+		// if past the num predict limit
+		if seq.numPredict > 0 && seq.numPredicted >= seq.numPredict {
+			s.removeSequence(i, llm.DoneReasonLength)
+			continue
+		}
+
 		sequence := strings.Join(seq.pendingResponses, "")
 
 		if ok, stop := common.FindStop(sequence, seq.stop); ok {

server/aliases.go

@@ -1,422 +0,0 @@
-package server
-
-import (
-	"encoding/json"
-	"errors"
-	"fmt"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"sort"
-	"strings"
-	"sync"
-
-	"github.com/ollama/ollama/manifest"
-	"github.com/ollama/ollama/types/model"
-)
-
-const (
-	routerConfigFilename = "server.json"
-	routerConfigVersion  = 1
-)
-
-var errAliasCycle = errors.New("alias cycle detected")
-
-type aliasEntry struct {
-	Alias          string `json:"alias"`
-	Target         string `json:"target"`
-	PrefixMatching bool   `json:"prefix_matching,omitempty"`
-}
-
-type routerConfig struct {
-	Version int          `json:"version"`
-	Aliases []aliasEntry `json:"aliases"`
-}
-
-type aliasStore struct {
-	mu            sync.RWMutex
-	path          string
-	entries       map[string]aliasEntry // normalized alias -> entry (exact matches)
-	prefixEntries []aliasEntry          // prefix matches, sorted longest-first
-}
-
-func newAliasStore(path string) (*aliasStore, error) {
-	store := &aliasStore{
-		path:    path,
-		entries: make(map[string]aliasEntry),
-	}
-	if err := store.load(); err != nil {
-		return nil, err
-	}
-	return store, nil
-}
-
-func (s *aliasStore) load() error {
-	data, err := os.ReadFile(s.path)
-	if err != nil {
-		if errors.Is(err, os.ErrNotExist) {
-			return nil
-		}
-		return err
-	}
-
-	var cfg routerConfig
-	if err := json.Unmarshal(data, &cfg); err != nil {
-		return err
-	}
-
-	if cfg.Version != 0 && cfg.Version != routerConfigVersion {
-		return fmt.Errorf("unsupported router config version %d", cfg.Version)
-	}
-
-	for _, entry := range cfg.Aliases {
-		targetName := model.ParseName(entry.Target)
-		if !targetName.IsValid() {
-			slog.Warn("invalid alias target in router config", "target", entry.Target)
-			continue
-		}
-		canonicalTarget := displayAliasName(targetName)
-
-		if entry.PrefixMatching {
-			// Prefix aliases don't need to be valid model names
-			alias := strings.TrimSpace(entry.Alias)
-			if alias == "" {
-				slog.Warn("empty prefix alias in router config")
-				continue
-			}
-			s.prefixEntries = append(s.prefixEntries, aliasEntry{
-				Alias:          alias,
-				Target:         canonicalTarget,
-				PrefixMatching: true,
-			})
-		} else {
-			aliasName := model.ParseName(entry.Alias)
-			if !aliasName.IsValid() {
-				slog.Warn("invalid alias name in router config", "alias", entry.Alias)
-				continue
-			}
-			canonicalAlias := displayAliasName(aliasName)
-			s.entries[normalizeAliasKey(aliasName)] = aliasEntry{
-				Alias:  canonicalAlias,
-				Target: canonicalTarget,
-			}
-		}
-	}
-
-	// Sort prefix entries by alias length descending (longest prefix wins)
-	s.sortPrefixEntriesLocked()
-
-	return nil
-}
-
-func (s *aliasStore) saveLocked() error {
-	dir := filepath.Dir(s.path)
-	if err := os.MkdirAll(dir, 0o755); err != nil {
-		return err
-	}
-
-	// Combine exact and prefix entries
-	entries := make([]aliasEntry, 0, len(s.entries)+len(s.prefixEntries))
-	for _, entry := range s.entries {
-		entries = append(entries, entry)
-	}
-	entries = append(entries, s.prefixEntries...)
-
-	sort.Slice(entries, func(i, j int) bool {
-		return strings.Compare(entries[i].Alias, entries[j].Alias) < 0
-	})
-
-	cfg := routerConfig{
-		Version: routerConfigVersion,
-		Aliases: entries,
-	}
-
-	f, err := os.CreateTemp(dir, "router-*.json")
-	if err != nil {
-		return err
-	}
-
-	enc := json.NewEncoder(f)
-	enc.SetIndent("", "  ")
-	if err := enc.Encode(cfg); err != nil {
-		_ = f.Close()
-		_ = os.Remove(f.Name())
-		return err
-	}
-
-	if err := f.Close(); err != nil {
-		_ = os.Remove(f.Name())
-		return err
-	}
-
-	if err := os.Chmod(f.Name(), 0o644); err != nil {
-		_ = os.Remove(f.Name())
-		return err
-	}
-
-	return os.Rename(f.Name(), s.path)
-}
-
-func (s *aliasStore) ResolveName(name model.Name) (model.Name, bool, error) {
-	// If a local model exists, do not allow alias shadowing (highest priority).
-	exists, err := localModelExists(name)
-	if err != nil {
-		return name, false, err
-	}
-	if exists {
-		return name, false, nil
-	}
-
-	key := normalizeAliasKey(name)
-
-	s.mu.RLock()
-	entry, exactMatch := s.entries[key]
-	var prefixMatch *aliasEntry
-	if !exactMatch {
-		// Try prefix matching - prefixEntries is sorted longest-first
-		nameStr := strings.ToLower(displayAliasName(name))
-		for i := range s.prefixEntries {
-			prefix := strings.ToLower(s.prefixEntries[i].Alias)
-			if strings.HasPrefix(nameStr, prefix) {
-				prefixMatch = &s.prefixEntries[i]
-				break // First match is longest due to sorting
-			}
-		}
-	}
-	s.mu.RUnlock()
-
-	if !exactMatch && prefixMatch == nil {
-		return name, false, nil
-	}
-
-	var current string
-	var visited map[string]struct{}
-
-	if exactMatch {
-		visited = map[string]struct{}{key: {}}
-		current = entry.Target
-	} else {
-		// For prefix match, use the target as-is
-		visited = map[string]struct{}{}
-		current = prefixMatch.Target
-	}
-
-	targetKey := normalizeAliasKeyString(current)
-
-	for {
-		targetName := model.ParseName(current)
-		if !targetName.IsValid() {
-			return name, false, fmt.Errorf("alias target %q is invalid", current)
-		}
-
-		if _, seen := visited[targetKey]; seen {
-			return name, false, errAliasCycle
-		}
-		visited[targetKey] = struct{}{}
-
-		s.mu.RLock()
-		next, ok := s.entries[targetKey]
-		s.mu.RUnlock()
-		if !ok {
-			return targetName, true, nil
-		}
-
-		current = next.Target
-		targetKey = normalizeAliasKeyString(current)
-	}
-}
-
-func (s *aliasStore) Set(alias, target model.Name, prefixMatching bool) error {
-	targetKey := normalizeAliasKey(target)
-
-	s.mu.Lock()
-	defer s.mu.Unlock()
-
-	if prefixMatching {
-		// For prefix aliases, we skip cycle detection since prefix matching
-		// works differently and the target is a specific model
-		aliasStr := displayAliasName(alias)
-
-		// Remove any existing prefix entry with the same alias
-		for i, e := range s.prefixEntries {
-			if strings.EqualFold(e.Alias, aliasStr) {
-				s.prefixEntries = append(s.prefixEntries[:i], s.prefixEntries[i+1:]...)
-				break
-			}
-		}
-
-		s.prefixEntries = append(s.prefixEntries, aliasEntry{
-			Alias:          aliasStr,
-			Target:         displayAliasName(target),
-			PrefixMatching: true,
-		})
-		s.sortPrefixEntriesLocked()
-		return s.saveLocked()
-	}
-
-	aliasKey := normalizeAliasKey(alias)
-
-	if aliasKey == targetKey {
-		return fmt.Errorf("alias cannot point to itself")
-	}
-
-	visited := map[string]struct{}{aliasKey: {}}
-	currentKey := targetKey
-	for {
-		if _, seen := visited[currentKey]; seen {
-			return errAliasCycle
-		}
-		visited[currentKey] = struct{}{}
-
-		next, ok := s.entries[currentKey]
-		if !ok {
-			break
-		}
-		currentKey = normalizeAliasKeyString(next.Target)
-	}
-
-	s.entries[aliasKey] = aliasEntry{
-		Alias:  displayAliasName(alias),
-		Target: displayAliasName(target),
-	}
-
-	return s.saveLocked()
-}
-
-func (s *aliasStore) Delete(alias model.Name) (bool, error) {
-	aliasKey := normalizeAliasKey(alias)
-
-	s.mu.Lock()
-	defer s.mu.Unlock()
-
-	// Try exact match first
-	if _, ok := s.entries[aliasKey]; ok {
-		delete(s.entries, aliasKey)
-		return true, s.saveLocked()
-	}
-
-	// Try prefix entries
-	aliasStr := displayAliasName(alias)
-	for i, e := range s.prefixEntries {
-		if strings.EqualFold(e.Alias, aliasStr) {
-			s.prefixEntries = append(s.prefixEntries[:i], s.prefixEntries[i+1:]...)
-			return true, s.saveLocked()
-		}
-	}
-
-	return false, nil
-}
-
-// DeleteByString deletes an alias by its raw string value, useful for prefix
-// aliases that may not be valid model names.
-func (s *aliasStore) DeleteByString(alias string) (bool, error) {
-	alias = strings.TrimSpace(alias)
-	aliasLower := strings.ToLower(alias)
-
-	s.mu.Lock()
-	defer s.mu.Unlock()
-
-	// Try prefix entries first (since this is mainly for prefix aliases)
-	for i, e := range s.prefixEntries {
-		if strings.EqualFold(e.Alias, alias) {
-			s.prefixEntries = append(s.prefixEntries[:i], s.prefixEntries[i+1:]...)
-			return true, s.saveLocked()
-		}
-	}
-
-	// Also check exact entries by normalized key
-	if _, ok := s.entries[aliasLower]; ok {
-		delete(s.entries, aliasLower)
-		return true, s.saveLocked()
-	}
-
-	return false, nil
-}
-
-func (s *aliasStore) List() []aliasEntry {
-	s.mu.RLock()
-	defer s.mu.RUnlock()
-
-	entries := make([]aliasEntry, 0, len(s.entries)+len(s.prefixEntries))
-	for _, entry := range s.entries {
-		entries = append(entries, entry)
-	}
-	entries = append(entries, s.prefixEntries...)
-
-	sort.Slice(entries, func(i, j int) bool {
-		return strings.Compare(entries[i].Alias, entries[j].Alias) < 0
-	})
-	return entries
-}
-
-func normalizeAliasKey(name model.Name) string {
-	return strings.ToLower(displayAliasName(name))
-}
-
-func (s *aliasStore) sortPrefixEntriesLocked() {
-	sort.Slice(s.prefixEntries, func(i, j int) bool {
-		// Sort by length descending (longest prefix first)
-		return len(s.prefixEntries[i].Alias) > len(s.prefixEntries[j].Alias)
-	})
-}
-
-func normalizeAliasKeyString(value string) string {
-	n := model.ParseName(value)
-	if !n.IsValid() {
-		return strings.ToLower(strings.TrimSpace(value))
-	}
-	return normalizeAliasKey(n)
-}
-
-func displayAliasName(n model.Name) string {
-	display := n.DisplayShortest()
-	if strings.EqualFold(n.Tag, "latest") {
-		if idx := strings.LastIndex(display, ":"); idx != -1 {
-			return display[:idx]
-		}
-	}
-	return display
-}
-
-func localModelExists(name model.Name) (bool, error) {
-	manifests, err := manifest.Manifests(true)
-	if err != nil {
-		return false, err
-	}
-	needle := name.String()
-	for existing := range manifests {
-		if strings.EqualFold(existing.String(), needle) {
-			return true, nil
-		}
-	}
-	return false, nil
-}
-
-func routerConfigPath() string {
-	home, err := os.UserHomeDir()
-	if err != nil {
-		return filepath.Join(".ollama", routerConfigFilename)
-	}
-	return filepath.Join(home, ".ollama", routerConfigFilename)
-}
-
-func (s *Server) aliasStore() (*aliasStore, error) {
-	s.aliasesOnce.Do(func() {
-		s.aliases, s.aliasesErr = newAliasStore(routerConfigPath())
-	})
-
-	return s.aliases, s.aliasesErr
-}
-
-func (s *Server) resolveModelAliasName(name model.Name) (model.Name, bool, error) {
-	store, err := s.aliasStore()
-	if err != nil {
-		return name, false, err
-	}
-
-	if store == nil {
-		return name, false, nil
-	}
-
-	return store.ResolveName(name)
-}

server/inference_request_log.go

@@ -0,0 +1,144 @@
+package server
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync/atomic"
+	"time"
+
+	"github.com/gin-gonic/gin"
+
+	"github.com/ollama/ollama/envconfig"
+)
+
+type inferenceRequestLogger struct {
+	dir     string
+	counter uint64
+}
+
+func newInferenceRequestLogger() (*inferenceRequestLogger, error) {
+	dir, err := os.MkdirTemp("", "ollama-request-logs-*")
+	if err != nil {
+		return nil, err
+	}
+
+	return &inferenceRequestLogger{dir: dir}, nil
+}
+
+func (s *Server) initRequestLogging() error {
+	if !envconfig.DebugLogRequests() {
+		return nil
+	}
+
+	requestLogger, err := newInferenceRequestLogger()
+	if err != nil {
+		return fmt.Errorf("enable OLLAMA_DEBUG_LOG_REQUESTS: %w", err)
+	}
+
+	s.requestLogger = requestLogger
+	slog.Info(fmt.Sprintf("request debug logging enabled; inference request logs will be stored in %s and include request bodies and replay curl commands", requestLogger.dir))
+
+	return nil
+}
+
+func (s *Server) withInferenceRequestLogging(route string, handlers ...gin.HandlerFunc) []gin.HandlerFunc {
+	if s.requestLogger == nil {
+		return handlers
+	}
+
+	return append([]gin.HandlerFunc{s.requestLogger.middleware(route)}, handlers...)
+}
+
+func (l *inferenceRequestLogger) middleware(route string) gin.HandlerFunc {
+	return func(c *gin.Context) {
+		if c.Request == nil {
+			c.Next()
+			return
+		}
+
+		method := c.Request.Method
+		host := c.Request.Host
+		scheme := "http"
+		if c.Request.TLS != nil {
+			scheme = "https"
+		}
+		contentType := c.GetHeader("Content-Type")
+
+		var body []byte
+		if c.Request.Body != nil {
+			var err error
+			body, err = io.ReadAll(c.Request.Body)
+			c.Request.Body = io.NopCloser(bytes.NewReader(body))
+			if err != nil {
+				slog.Warn("failed to read request body for debug logging", "route", route, "error", err)
+			}
+		}
+
+		c.Next()
+		l.log(route, method, scheme, host, contentType, body)
+	}
+}
+
+func (l *inferenceRequestLogger) log(route, method, scheme, host, contentType string, body []byte) {
+	if l == nil || l.dir == "" {
+		return
+	}
+
+	if contentType == "" {
+		contentType = "application/json"
+	}
+	if host == "" || scheme == "" {
+		base := envconfig.Host()
+		if host == "" {
+			host = base.Host
+		}
+		if scheme == "" {
+			scheme = base.Scheme
+		}
+	}
+
+	routeForFilename := sanitizeRouteForFilename(route)
+	timestamp := fmt.Sprintf("%s-%06d", time.Now().UTC().Format("20060102T150405.000000000Z"), atomic.AddUint64(&l.counter, 1))
+	bodyFilename := fmt.Sprintf("%s_%s_body.json", timestamp, routeForFilename)
+	curlFilename := fmt.Sprintf("%s_%s_request.sh", timestamp, routeForFilename)
+	bodyPath := filepath.Join(l.dir, bodyFilename)
+	curlPath := filepath.Join(l.dir, curlFilename)
+
+	if err := os.WriteFile(bodyPath, body, 0o600); err != nil {
+		slog.Warn("failed to write debug request body", "route", route, "error", err)
+		return
+	}
+
+	url := fmt.Sprintf("%s://%s%s", scheme, host, route)
+	curl := fmt.Sprintf("#!/bin/sh\nSCRIPT_DIR=\"$(CDPATH= cd -- \"$(dirname -- \"$0\")\" && pwd)\"\ncurl --request %s --url %q --header %q --data-binary @\"${SCRIPT_DIR}/%s\"\n", method, url, "Content-Type: "+contentType, bodyFilename)
+	if err := os.WriteFile(curlPath, []byte(curl), 0o600); err != nil {
+		slog.Warn("failed to write debug request replay command", "route", route, "error", err)
+		return
+	}
+
+	slog.Info(fmt.Sprintf("logged to %s, replay using curl with `sh %s`", bodyPath, curlPath))
+}
+
+func sanitizeRouteForFilename(route string) string {
+	route = strings.TrimPrefix(route, "/")
+	if route == "" {
+		return "root"
+	}
+
+	var b strings.Builder
+	b.Grow(len(route))
+	for _, r := range route {
+		if ('a' <= r && r <= 'z') || ('A' <= r && r <= 'Z') || ('0' <= r && r <= '9') {
+			b.WriteRune(r)
+		} else {
+			b.WriteByte('_')
+		}
+	}
+
+	return b.String()
+}

server/prompt.go

@@ -27,14 +27,12 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 	// Clip images are represented as 768 tokens, each an embedding
 	imageNumTokens := 768
 
-	n := len(msgs) - 1
-	// in reverse, find all messages that fit into context window
-	for i := n; i >= 0; i-- {
-		// always include the last message
-		if i == n {
-			continue
-		}
+	lastMsgIdx := len(msgs) - 1
+	currMsgIdx := 0
 
+	// Start with all messages and remove from the front until it fits in context
+	for i := 0; i <= lastMsgIdx; i++ {
+		// Collect system messages from the portion we're about to skip
 		system = make([]api.Message, 0)
 		for j := range i {
 			if msgs[j].Role == "system" {
@@ -54,20 +52,26 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 
 		ctxLen := len(s)
 		if m.ProjectorPaths != nil {
-			for _, m := range msgs[i:] {
-				ctxLen += imageNumTokens * len(m.Images)
+			for _, msg := range msgs[i:] {
+				ctxLen += imageNumTokens * len(msg.Images)
 			}
 		}
 
-		if truncate && ctxLen > opts.NumCtx {
-			slog.Debug("truncating input messages which exceed context length", "truncated", len(msgs[i:]))
+		if !truncate || ctxLen <= opts.NumCtx {
+			currMsgIdx = i
+			break
+		}
+
+		// Must always include at least the last message
+		if i == lastMsgIdx {
+			currMsgIdx = lastMsgIdx
 			break
-		} else {
-			n = i
 		}
 	}
 
-	currMsgIdx := n
+	if currMsgIdx > 0 {
+		slog.Debug("truncating input messages which exceed context length", "truncated", len(msgs[currMsgIdx:]))
+	}
 
 	for cnt, msg := range msgs[currMsgIdx:] {
 		if slices.Contains(m.Config.ModelFamilies, "mllama") && len(msg.Images) > 1 {

server/prompt_test.go

@@ -2,6 +2,7 @@ package server
 
 import (
 	"bytes"
+	"context"
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
@@ -264,3 +265,68 @@ func TestChatPrompt(t *testing.T) {
 		})
 	}
 }
+
+func TestChatPromptTokenizeCalls(t *testing.T) {
+	tmpl, err := template.Parse(`
+{{- if .System }}{{ .System }} {{ end }}
+{{- if .Prompt }}{{ .Prompt }} {{ end }}
+{{- if .Response }}{{ .Response }} {{ end }}`)
+	if err != nil {
+		t.Fatal(err)
+	}
+	model := Model{Template: tmpl}
+
+	cases := []struct {
+		name         string
+		limit        int
+		msgs         []api.Message
+		maxTokenizes int
+	}{
+		{
+			name:  "all messages fit",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "user", Content: "message 1"},
+				{Role: "assistant", Content: "response 1"},
+				{Role: "user", Content: "message 2"},
+				{Role: "assistant", Content: "response 2"},
+				{Role: "user", Content: "message 3"},
+			},
+			maxTokenizes: 1,
+		},
+		{
+			name:  "truncate to last message",
+			limit: 5,
+			msgs: []api.Message{
+				{Role: "user", Content: "message 1"},
+				{Role: "assistant", Content: "response 1"},
+				{Role: "user", Content: "message 2"},
+				{Role: "assistant", Content: "response 2"},
+				{Role: "user", Content: "message 3"},
+			},
+			maxTokenizes: 5,
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			tokenizeCount := 0
+			countingTokenize := func(ctx context.Context, s string) ([]int, error) {
+				tokenizeCount++
+				tokens, err := mockRunner{}.Tokenize(ctx, s)
+				return tokens, err
+			}
+
+			opts := api.Options{Runner: api.Runner{NumCtx: tt.limit}}
+			think := false
+			_, _, err := chatPrompt(t.Context(), &model, countingTokenize, &opts, tt.msgs, nil, &api.ThinkValue{Value: think}, true)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			if tokenizeCount > tt.maxTokenizes {
+				t.Errorf("tokenize called %d times, expected at most %d", tokenizeCount, tt.maxTokenizes)
+			}
+		})
+	}
+}

server/routes.go

@@ -22,7 +22,6 @@ import (
 	"os/signal"
 	"slices"
 	"strings"
-	"sync"
 	"sync/atomic"
 	"syscall"
 	"time"
@@ -82,9 +81,7 @@ type Server struct {
 	addr          net.Addr
 	sched         *Scheduler
 	defaultNumCtx int
-	aliasesOnce   sync.Once
-	aliases       *aliasStore
-	aliasesErr    error
+	requestLogger *inferenceRequestLogger
 }
 
 func init() {
@@ -195,16 +192,9 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	resolvedName, _, err := s.resolveModelAliasName(name)
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-	name = resolvedName
-
 	// We cannot currently consolidate this into GetModel because all we'll
 	// induce infinite recursion given the current code structure.
-	name, err = getExistingName(name)
+	name, err := getExistingName(name)
 	if err != nil {
 		c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Model)})
 		return
@@ -1591,30 +1581,27 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.POST("/api/blobs/:digest", s.CreateBlobHandler)
 	r.HEAD("/api/blobs/:digest", s.HeadBlobHandler)
 	r.POST("/api/copy", s.CopyHandler)
-	r.GET("/api/experimental/aliases", s.ListAliasesHandler)
-	r.POST("/api/experimental/aliases", s.CreateAliasHandler)
-	r.DELETE("/api/experimental/aliases", s.DeleteAliasHandler)
 
 	// Inference
 	r.GET("/api/ps", s.PsHandler)
-	r.POST("/api/generate", s.GenerateHandler)
-	r.POST("/api/chat", s.ChatHandler)
+	r.POST("/api/generate", s.withInferenceRequestLogging("/api/generate", s.GenerateHandler)...)
+	r.POST("/api/chat", s.withInferenceRequestLogging("/api/chat", s.ChatHandler)...)
 	r.POST("/api/embed", s.EmbedHandler)
 	r.POST("/api/embeddings", s.EmbeddingsHandler)
 
 	// Inference (OpenAI compatibility)
-	r.POST("/v1/chat/completions", middleware.ChatMiddleware(), s.ChatHandler)
-	r.POST("/v1/completions", middleware.CompletionsMiddleware(), s.GenerateHandler)
+	r.POST("/v1/chat/completions", s.withInferenceRequestLogging("/v1/chat/completions", middleware.ChatMiddleware(), s.ChatHandler)...)
+	r.POST("/v1/completions", s.withInferenceRequestLogging("/v1/completions", middleware.CompletionsMiddleware(), s.GenerateHandler)...)
 	r.POST("/v1/embeddings", middleware.EmbeddingsMiddleware(), s.EmbedHandler)
 	r.GET("/v1/models", middleware.ListMiddleware(), s.ListHandler)
 	r.GET("/v1/models/:model", middleware.RetrieveMiddleware(), s.ShowHandler)
-	r.POST("/v1/responses", middleware.ResponsesMiddleware(), s.ChatHandler)
+	r.POST("/v1/responses", s.withInferenceRequestLogging("/v1/responses", middleware.ResponsesMiddleware(), s.ChatHandler)...)
 	// OpenAI-compatible image generation endpoints
 	r.POST("/v1/images/generations", middleware.ImageGenerationsMiddleware(), s.GenerateHandler)
 	r.POST("/v1/images/edits", middleware.ImageEditsMiddleware(), s.GenerateHandler)
 
 	// Inference (Anthropic compatibility)
-	r.POST("/v1/messages", middleware.AnthropicMessagesMiddleware(), s.ChatHandler)
+	r.POST("/v1/messages", s.withInferenceRequestLogging("/v1/messages", middleware.AnthropicMessagesMiddleware(), s.ChatHandler)...)
 
 	if rc != nil {
 		// wrap old with new
@@ -1664,6 +1651,9 @@ func Serve(ln net.Listener) error {
 	}
 
 	s := &Server{addr: ln.Addr()}
+	if err := s.initRequestLogging(); err != nil {
+		return err
+	}
 
 	var rc *ollama.Registry
 	if useClient2 {
@@ -1964,20 +1954,13 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
-	resolvedName, _, err := s.resolveModelAliasName(name)
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-	name = resolvedName
-
-	name, err = getExistingName(name)
+	name, err := getExistingName(name)
 	if err != nil {
 		c.JSON(http.StatusBadRequest, gin.H{"error": "model is required"})
 		return
 	}
 
-	m, err := GetModel(name.String())
+	m, err := GetModel(req.Model)
 	if err != nil {
 		switch {
 		case os.IsNotExist(err):

server/routes_aliases.go

@@ -1,159 +0,0 @@
-package server
-
-import (
-	"errors"
-	"fmt"
-	"io"
-	"net/http"
-	"strings"
-
-	"github.com/gin-gonic/gin"
-
-	"github.com/ollama/ollama/types/model"
-)
-
-type aliasListResponse struct {
-	Aliases []aliasEntry `json:"aliases"`
-}
-
-type aliasDeleteRequest struct {
-	Alias string `json:"alias"`
-}
-
-func (s *Server) ListAliasesHandler(c *gin.Context) {
-	store, err := s.aliasStore()
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-
-	var aliases []aliasEntry
-	if store != nil {
-		aliases = store.List()
-	}
-
-	c.JSON(http.StatusOK, aliasListResponse{Aliases: aliases})
-}
-
-func (s *Server) CreateAliasHandler(c *gin.Context) {
-	var req aliasEntry
-	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
-		return
-	} else if err != nil {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-		return
-	}
-
-	req.Alias = strings.TrimSpace(req.Alias)
-	req.Target = strings.TrimSpace(req.Target)
-	if req.Alias == "" || req.Target == "" {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "alias and target are required"})
-		return
-	}
-
-	// Target must always be a valid model name
-	targetName := model.ParseName(req.Target)
-	if !targetName.IsValid() {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("target %q is invalid", req.Target)})
-		return
-	}
-
-	var aliasName model.Name
-	if req.PrefixMatching {
-		// For prefix aliases, we still parse the alias to normalize it,
-		// but we allow any non-empty string since prefix patterns may not be valid model names
-		aliasName = model.ParseName(req.Alias)
-		// Even if not valid as a model name, we accept it for prefix matching
-	} else {
-		aliasName = model.ParseName(req.Alias)
-		if !aliasName.IsValid() {
-			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("alias %q is invalid", req.Alias)})
-			return
-		}
-
-		if normalizeAliasKey(aliasName) == normalizeAliasKey(targetName) {
-			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "alias cannot point to itself"})
-			return
-		}
-
-		exists, err := localModelExists(aliasName)
-		if err != nil {
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-			return
-		}
-		if exists {
-			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("alias %q conflicts with existing model", req.Alias)})
-			return
-		}
-	}
-
-	store, err := s.aliasStore()
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-
-	if err := store.Set(aliasName, targetName, req.PrefixMatching); err != nil {
-		status := http.StatusInternalServerError
-		if errors.Is(err, errAliasCycle) {
-			status = http.StatusBadRequest
-		}
-		c.AbortWithStatusJSON(status, gin.H{"error": err.Error()})
-		return
-	}
-
-	resp := aliasEntry{
-		Alias:          displayAliasName(aliasName),
-		Target:         displayAliasName(targetName),
-		PrefixMatching: req.PrefixMatching,
-	}
-	if req.PrefixMatching && !aliasName.IsValid() {
-		// For prefix aliases that aren't valid model names, use the raw alias
-		resp.Alias = req.Alias
-	}
-	c.JSON(http.StatusOK, resp)
-}
-
-func (s *Server) DeleteAliasHandler(c *gin.Context) {
-	var req aliasDeleteRequest
-	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
-		return
-	} else if err != nil {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-		return
-	}
-
-	req.Alias = strings.TrimSpace(req.Alias)
-	if req.Alias == "" {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "alias is required"})
-		return
-	}
-
-	store, err := s.aliasStore()
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-
-	aliasName := model.ParseName(req.Alias)
-	var deleted bool
-	if aliasName.IsValid() {
-		deleted, err = store.Delete(aliasName)
-	} else {
-		// For invalid model names (like prefix aliases), try deleting by raw string
-		deleted, err = store.DeleteByString(req.Alias)
-	}
-
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-	if !deleted {
-		c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("alias %q not found", req.Alias)})
-		return
-	}
-
-	c.JSON(http.StatusOK, gin.H{"deleted": true})
-}

server/routes_aliases_test.go

@@ -1,426 +0,0 @@
-package server
-
-import (
-	"encoding/json"
-	"net/http"
-	"net/http/httptest"
-	"net/url"
-	"path/filepath"
-	"testing"
-
-	"github.com/gin-gonic/gin"
-
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/types/model"
-)
-
-func TestAliasShadowingRejected(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-	t.Setenv("HOME", t.TempDir())
-
-	s := Server{}
-	w := createRequest(t, s.CreateHandler, api.CreateRequest{
-		Model:      "shadowed-model",
-		RemoteHost: "example.com",
-		From:       "test",
-		Info: map[string]any{
-			"capabilities": []string{"completion"},
-		},
-		Stream: &stream,
-	})
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d", w.Code)
-	}
-
-	w = createRequest(t, s.CreateAliasHandler, aliasEntry{Alias: "shadowed-model", Target: "other-model"})
-	if w.Code != http.StatusBadRequest {
-		t.Fatalf("expected status 400, got %d", w.Code)
-	}
-}
-
-func TestAliasResolvesForChatRemote(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-	t.Setenv("HOME", t.TempDir())
-
-	var remoteModel string
-	rs := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		var req api.ChatRequest
-		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
-			t.Fatal(err)
-		}
-		remoteModel = req.Model
-
-		w.Header().Set("Content-Type", "application/json")
-		resp := api.ChatResponse{
-			Model:      req.Model,
-			Done:       true,
-			DoneReason: "load",
-		}
-		if err := json.NewEncoder(w).Encode(&resp); err != nil {
-			t.Fatal(err)
-		}
-	}))
-	defer rs.Close()
-
-	p, err := url.Parse(rs.URL)
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	t.Setenv("OLLAMA_REMOTES", p.Hostname())
-
-	s := Server{}
-	w := createRequest(t, s.CreateHandler, api.CreateRequest{
-		Model:      "target-model",
-		RemoteHost: rs.URL,
-		From:       "test",
-		Info: map[string]any{
-			"capabilities": []string{"completion"},
-		},
-		Stream: &stream,
-	})
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d", w.Code)
-	}
-
-	w = createRequest(t, s.CreateAliasHandler, aliasEntry{Alias: "alias-model", Target: "target-model"})
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d", w.Code)
-	}
-
-	w = createRequest(t, s.ChatHandler, api.ChatRequest{
-		Model:    "alias-model",
-		Messages: []api.Message{{Role: "user", Content: "hi"}},
-		Stream:   &stream,
-	})
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d", w.Code)
-	}
-
-	var resp api.ChatResponse
-	if err := json.NewDecoder(w.Body).Decode(&resp); err != nil {
-		t.Fatal(err)
-	}
-
-	if resp.Model != "alias-model" {
-		t.Fatalf("expected response model to be alias-model, got %q", resp.Model)
-	}
-
-	if remoteModel != "test" {
-		t.Fatalf("expected remote model to be 'test', got %q", remoteModel)
-	}
-}
-
-func TestPrefixAliasBasicMatching(t *testing.T) {
-	tmpDir := t.TempDir()
-	store, err := newAliasStore(filepath.Join(tmpDir, "server.json"))
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	// Create a prefix alias: "myprefix-" -> "targetmodel"
-	targetName := model.ParseName("targetmodel")
-
-	// Set a prefix alias (using "myprefix-" as the pattern)
-	store.mu.Lock()
-	store.prefixEntries = append(store.prefixEntries, aliasEntry{
-		Alias:          "myprefix-",
-		Target:         "targetmodel",
-		PrefixMatching: true,
-	})
-	store.mu.Unlock()
-
-	// Test that "myprefix-foo" resolves to "targetmodel"
-	testName := model.ParseName("myprefix-foo")
-	resolved, wasResolved, err := store.ResolveName(testName)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !wasResolved {
-		t.Fatal("expected name to be resolved")
-	}
-	if resolved.DisplayShortest() != targetName.DisplayShortest() {
-		t.Fatalf("expected resolved name to be %q, got %q", targetName.DisplayShortest(), resolved.DisplayShortest())
-	}
-
-	// Test that "otherprefix-foo" does not resolve
-	otherName := model.ParseName("otherprefix-foo")
-	_, wasResolved, err = store.ResolveName(otherName)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if wasResolved {
-		t.Fatal("expected name not to be resolved")
-	}
-
-	// Test that exact alias takes precedence
-	exactAlias := model.ParseName("myprefix-exact")
-	exactTarget := model.ParseName("exacttarget")
-	if err := store.Set(exactAlias, exactTarget, false); err != nil {
-		t.Fatal(err)
-	}
-
-	resolved, wasResolved, err = store.ResolveName(exactAlias)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !wasResolved {
-		t.Fatal("expected name to be resolved")
-	}
-	if resolved.DisplayShortest() != exactTarget.DisplayShortest() {
-		t.Fatalf("expected resolved name to be %q (exact match), got %q", exactTarget.DisplayShortest(), resolved.DisplayShortest())
-	}
-}
-
-func TestPrefixAliasLongestMatchWins(t *testing.T) {
-	tmpDir := t.TempDir()
-	store, err := newAliasStore(filepath.Join(tmpDir, "server.json"))
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	// Add two prefix aliases with overlapping patterns
-	store.mu.Lock()
-	store.prefixEntries = []aliasEntry{
-		{Alias: "abc-", Target: "short-target", PrefixMatching: true},
-		{Alias: "abc-def-", Target: "long-target", PrefixMatching: true},
-	}
-	store.sortPrefixEntriesLocked()
-	store.mu.Unlock()
-
-	// "abc-def-ghi" should match the longer prefix "abc-def-"
-	testName := model.ParseName("abc-def-ghi")
-	resolved, wasResolved, err := store.ResolveName(testName)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !wasResolved {
-		t.Fatal("expected name to be resolved")
-	}
-	expectedLongTarget := model.ParseName("long-target")
-	if resolved.DisplayShortest() != expectedLongTarget.DisplayShortest() {
-		t.Fatalf("expected resolved name to be %q (longest prefix match), got %q", expectedLongTarget.DisplayShortest(), resolved.DisplayShortest())
-	}
-
-	// "abc-xyz" should match the shorter prefix "abc-"
-	testName2 := model.ParseName("abc-xyz")
-	resolved, wasResolved, err = store.ResolveName(testName2)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !wasResolved {
-		t.Fatal("expected name to be resolved")
-	}
-	expectedShortTarget := model.ParseName("short-target")
-	if resolved.DisplayShortest() != expectedShortTarget.DisplayShortest() {
-		t.Fatalf("expected resolved name to be %q, got %q", expectedShortTarget.DisplayShortest(), resolved.DisplayShortest())
-	}
-}
-
-func TestPrefixAliasChain(t *testing.T) {
-	tmpDir := t.TempDir()
-	store, err := newAliasStore(filepath.Join(tmpDir, "server.json"))
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	// Create a chain: prefix "test-" -> "intermediate" -> "final"
-	intermediate := model.ParseName("intermediate")
-	final := model.ParseName("final")
-
-	// Add prefix alias
-	store.mu.Lock()
-	store.prefixEntries = []aliasEntry{
-		{Alias: "test-", Target: "intermediate", PrefixMatching: true},
-	}
-	store.mu.Unlock()
-
-	// Add exact alias for the intermediate step
-	if err := store.Set(intermediate, final, false); err != nil {
-		t.Fatal(err)
-	}
-
-	// "test-foo" should resolve through the chain to "final"
-	testName := model.ParseName("test-foo")
-	resolved, wasResolved, err := store.ResolveName(testName)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !wasResolved {
-		t.Fatal("expected name to be resolved")
-	}
-	if resolved.DisplayShortest() != final.DisplayShortest() {
-		t.Fatalf("expected resolved name to be %q, got %q", final.DisplayShortest(), resolved.DisplayShortest())
-	}
-}
-
-func TestPrefixAliasCRUD(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-	t.Setenv("HOME", t.TempDir())
-
-	s := Server{}
-
-	// Create a prefix alias via API
-	w := createRequest(t, s.CreateAliasHandler, aliasEntry{
-		Alias:          "myprefix-",
-		Target:         "llama2",
-		PrefixMatching: true,
-	})
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
-	}
-
-	var createResp aliasEntry
-	if err := json.NewDecoder(w.Body).Decode(&createResp); err != nil {
-		t.Fatal(err)
-	}
-	if !createResp.PrefixMatching {
-		t.Fatal("expected prefix_matching to be true in response")
-	}
-
-	// List aliases and verify the prefix alias is included
-	w = createRequest(t, s.ListAliasesHandler, nil)
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d", w.Code)
-	}
-
-	var listResp aliasListResponse
-	if err := json.NewDecoder(w.Body).Decode(&listResp); err != nil {
-		t.Fatal(err)
-	}
-
-	found := false
-	for _, a := range listResp.Aliases {
-		if a.PrefixMatching && a.Target == "llama2" {
-			found = true
-			break
-		}
-	}
-	if !found {
-		t.Fatal("expected to find prefix alias in list")
-	}
-
-	// Delete the prefix alias
-	w = createRequest(t, s.DeleteAliasHandler, aliasDeleteRequest{Alias: "myprefix-"})
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
-	}
-
-	// Verify it's deleted
-	w = createRequest(t, s.ListAliasesHandler, nil)
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d", w.Code)
-	}
-
-	if err := json.NewDecoder(w.Body).Decode(&listResp); err != nil {
-		t.Fatal(err)
-	}
-
-	for _, a := range listResp.Aliases {
-		if a.PrefixMatching {
-			t.Fatal("expected prefix alias to be deleted")
-		}
-	}
-}
-
-func TestPrefixAliasCaseInsensitive(t *testing.T) {
-	tmpDir := t.TempDir()
-	store, err := newAliasStore(filepath.Join(tmpDir, "server.json"))
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	// Add a prefix alias with mixed case
-	store.mu.Lock()
-	store.prefixEntries = []aliasEntry{
-		{Alias: "MyPrefix-", Target: "targetmodel", PrefixMatching: true},
-	}
-	store.mu.Unlock()
-
-	// Test that matching is case-insensitive
-	testName := model.ParseName("myprefix-foo")
-	resolved, wasResolved, err := store.ResolveName(testName)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !wasResolved {
-		t.Fatal("expected name to be resolved (case-insensitive)")
-	}
-	expectedTarget := model.ParseName("targetmodel")
-	if resolved.DisplayShortest() != expectedTarget.DisplayShortest() {
-		t.Fatalf("expected resolved name to be %q, got %q", expectedTarget.DisplayShortest(), resolved.DisplayShortest())
-	}
-
-	// Test uppercase request
-	testName2 := model.ParseName("MYPREFIX-BAR")
-	_, wasResolved, err = store.ResolveName(testName2)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !wasResolved {
-		t.Fatal("expected name to be resolved (uppercase)")
-	}
-}
-
-func TestPrefixAliasLocalModelPrecedence(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-	t.Setenv("HOME", t.TempDir())
-
-	s := Server{}
-
-	// Create a local model that would match a prefix alias
-	w := createRequest(t, s.CreateHandler, api.CreateRequest{
-		Model:      "myprefix-localmodel",
-		RemoteHost: "example.com",
-		From:       "test",
-		Info: map[string]any{
-			"capabilities": []string{"completion"},
-		},
-		Stream: &stream,
-	})
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
-	}
-
-	// Create a prefix alias that would match the local model name
-	w = createRequest(t, s.CreateAliasHandler, aliasEntry{
-		Alias:          "myprefix-",
-		Target:         "someothermodel",
-		PrefixMatching: true,
-	})
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
-	}
-
-	// Verify that resolving "myprefix-localmodel" returns the local model, not the alias target
-	store, err := s.aliasStore()
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	localModelName := model.ParseName("myprefix-localmodel")
-	resolved, wasResolved, err := store.ResolveName(localModelName)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if wasResolved {
-		t.Fatalf("expected local model to take precedence (wasResolved should be false), but got resolved to %q", resolved.DisplayShortest())
-	}
-	if resolved.DisplayShortest() != localModelName.DisplayShortest() {
-		t.Fatalf("expected resolved name to be local model %q, got %q", localModelName.DisplayShortest(), resolved.DisplayShortest())
-	}
-
-	// Also verify that a non-local model matching the prefix DOES resolve to the alias target
-	nonLocalName := model.ParseName("myprefix-nonexistent")
-	resolved, wasResolved, err = store.ResolveName(nonLocalName)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !wasResolved {
-		t.Fatal("expected non-local model to resolve via prefix alias")
-	}
-	expectedTarget := model.ParseName("someothermodel")
-	if resolved.DisplayShortest() != expectedTarget.DisplayShortest() {
-		t.Fatalf("expected resolved name to be %q, got %q", expectedTarget.DisplayShortest(), resolved.DisplayShortest())
-	}
-}

server/routes_request_log_test.go

@@ -0,0 +1,128 @@
+package server
+
+import (
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+)
+
+func TestInferenceRequestLoggerMiddlewareWritesReplayArtifacts(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	logDir := t.TempDir()
+	requestLogger := &inferenceRequestLogger{dir: logDir}
+
+	const route = "/v1/chat/completions"
+	const requestBody = `{"model":"test-model","messages":[{"role":"user","content":"hello"}]}`
+
+	var bodySeenByHandler string
+
+	r := gin.New()
+	r.POST(route, requestLogger.middleware(route), func(c *gin.Context) {
+		body, err := io.ReadAll(c.Request.Body)
+		if err != nil {
+			t.Fatalf("failed to read body in handler: %v", err)
+		}
+
+		bodySeenByHandler = string(body)
+		c.Status(http.StatusOK)
+	})
+
+	req := httptest.NewRequest(http.MethodPost, route, strings.NewReader(requestBody))
+	req.Host = "127.0.0.1:11434"
+	req.Header.Set("Content-Type", "application/json")
+	w := httptest.NewRecorder()
+	r.ServeHTTP(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	if bodySeenByHandler != requestBody {
+		t.Fatalf("handler body mismatch:\nexpected: %s\ngot: %s", requestBody, bodySeenByHandler)
+	}
+
+	bodyFiles, err := filepath.Glob(filepath.Join(logDir, "*_v1_chat_completions_body.json"))
+	if err != nil {
+		t.Fatalf("failed to glob body logs: %v", err)
+	}
+	if len(bodyFiles) != 1 {
+		t.Fatalf("expected 1 body log, got %d (%v)", len(bodyFiles), bodyFiles)
+	}
+
+	curlFiles, err := filepath.Glob(filepath.Join(logDir, "*_v1_chat_completions_request.sh"))
+	if err != nil {
+		t.Fatalf("failed to glob curl logs: %v", err)
+	}
+	if len(curlFiles) != 1 {
+		t.Fatalf("expected 1 curl log, got %d (%v)", len(curlFiles), curlFiles)
+	}
+
+	bodyData, err := os.ReadFile(bodyFiles[0])
+	if err != nil {
+		t.Fatalf("failed to read body log: %v", err)
+	}
+	if string(bodyData) != requestBody {
+		t.Fatalf("body log mismatch:\nexpected: %s\ngot: %s", requestBody, string(bodyData))
+	}
+
+	curlData, err := os.ReadFile(curlFiles[0])
+	if err != nil {
+		t.Fatalf("failed to read curl log: %v", err)
+	}
+
+	curlString := string(curlData)
+	if !strings.Contains(curlString, "http://127.0.0.1:11434"+route) {
+		t.Fatalf("curl log does not contain expected route URL: %s", curlString)
+	}
+
+	bodyFileName := filepath.Base(bodyFiles[0])
+	if !strings.Contains(curlString, "@\"${SCRIPT_DIR}/"+bodyFileName+"\"") {
+		t.Fatalf("curl log does not reference sibling body file: %s", curlString)
+	}
+}
+
+func TestNewInferenceRequestLoggerCreatesDirectory(t *testing.T) {
+	requestLogger, err := newInferenceRequestLogger()
+	if err != nil {
+		t.Fatalf("expected no error creating request logger: %v", err)
+	}
+	t.Cleanup(func() {
+		_ = os.RemoveAll(requestLogger.dir)
+	})
+
+	if requestLogger == nil || requestLogger.dir == "" {
+		t.Fatalf("expected request logger directory to be set")
+	}
+
+	info, err := os.Stat(requestLogger.dir)
+	if err != nil {
+		t.Fatalf("expected directory to exist: %v", err)
+	}
+	if !info.IsDir() {
+		t.Fatalf("expected %q to be a directory", requestLogger.dir)
+	}
+}
+
+func TestSanitizeRouteForFilename(t *testing.T) {
+	tests := []struct {
+		route string
+		want  string
+	}{
+		{route: "/api/generate", want: "api_generate"},
+		{route: "/v1/chat/completions", want: "v1_chat_completions"},
+		{route: "/v1/messages", want: "v1_messages"},
+	}
+
+	for _, tt := range tests {
+		if got := sanitizeRouteForFilename(tt.route); got != tt.want {
+			t.Fatalf("sanitizeRouteForFilename(%q) = %q, want %q", tt.route, got, tt.want)
+		}
+	}
+}

server/sched.go

@@ -417,9 +417,9 @@ func (s *Scheduler) load(req *LlmRequest, f *ggml.GGML, systemInfo ml.SystemInfo
 		numParallel = 1
 	}
 
-	// `mllama`, `qwen3vl`, and `qwen3vlmoe` are snowflakes and uses an encoder cache which cannot be used with num_parallel > 1
+	// Some architectures are not safe with num_parallel > 1.
 	// ref: https://github.com/ollama/ollama/issues/4165
-	if slices.Contains([]string{"mllama", "qwen3vl", "qwen3vlmoe"}, req.model.Config.ModelFamily) && numParallel != 1 {
+	if slices.Contains([]string{"mllama", "qwen3vl", "qwen3vlmoe", "qwen3next", "lfm2", "lfm2moe"}, req.model.Config.ModelFamily) && numParallel != 1 {
 		numParallel = 1
 		slog.Warn("model architecture does not currently support parallel requests", "architecture", req.model.Config.ModelFamily)
 	}