middleware: use HTTP status code for Anthropic error mapping

Use w.ResponseWriter.Status() instead of parsing StatusCode from JSON
payload. routes.go typically sends errors as gin.H{"error": "..."}
without a StatusCode field, causing all errors to be mapped to
"api_error" instead of the appropriate type (not_found_error,
invalid_request_error, etc.).

Added tests to verify error handling for common routes.go patterns.

anthropic/anthropic.go

@@ -0,0 +1,769 @@
+package anthropic
+
+import (
+	"crypto/rand"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// Error types matching Anthropic API
+type Error struct {
+	Type    string `json:"type"`
+	Message string `json:"message"`
+}
+
+type ErrorResponse struct {
+	Type      string `json:"type"` // always "error"
+	Error     Error  `json:"error"`
+	RequestID string `json:"request_id,omitempty"`
+}
+
+// NewError creates a new ErrorResponse with the appropriate error type based on HTTP status code
+func NewError(code int, message string) ErrorResponse {
+	var etype string
+	switch code {
+	case http.StatusBadRequest:
+		etype = "invalid_request_error"
+	case http.StatusUnauthorized:
+		etype = "authentication_error"
+	case http.StatusForbidden:
+		etype = "permission_error"
+	case http.StatusNotFound:
+		etype = "not_found_error"
+	case http.StatusTooManyRequests:
+		etype = "rate_limit_error"
+	case http.StatusServiceUnavailable, 529:
+		etype = "overloaded_error"
+	default:
+		etype = "api_error"
+	}
+
+	return ErrorResponse{
+		Type:      "error",
+		Error:     Error{Type: etype, Message: message},
+		RequestID: generateID("req"),
+	}
+}
+
+// Request types
+
+// MessagesRequest represents an Anthropic Messages API request
+type MessagesRequest struct {
+	Model         string          `json:"model"`
+	MaxTokens     int             `json:"max_tokens"`
+	Messages      []MessageParam  `json:"messages"`
+	System        any             `json:"system,omitempty"` // string or []ContentBlock
+	Stream        bool            `json:"stream,omitempty"`
+	Temperature   *float64        `json:"temperature,omitempty"`
+	TopP          *float64        `json:"top_p,omitempty"`
+	TopK          *int            `json:"top_k,omitempty"`
+	StopSequences []string        `json:"stop_sequences,omitempty"`
+	Tools         []Tool          `json:"tools,omitempty"`
+	ToolChoice    *ToolChoice     `json:"tool_choice,omitempty"`
+	Thinking      *ThinkingConfig `json:"thinking,omitempty"`
+	Metadata      *Metadata       `json:"metadata,omitempty"`
+}
+
+// MessageParam represents a message in the request
+type MessageParam struct {
+	Role    string `json:"role"`    // "user" or "assistant"
+	Content any    `json:"content"` // string or []ContentBlock
+}
+
+// ContentBlock represents a content block in a message.
+// Text and Thinking use pointers so they serialize as the field being present (even if empty)
+// only when set, which is required for SDK streaming accumulation.
+type ContentBlock struct {
+	Type string `json:"type"` // text, image, tool_use, tool_result, thinking
+
+	// For text blocks - pointer so field only appears when set (SDK requires it for accumulation)
+	Text *string `json:"text,omitempty"`
+
+	// For image blocks
+	Source *ImageSource `json:"source,omitempty"`
+
+	// For tool_use blocks
+	ID    string `json:"id,omitempty"`
+	Name  string `json:"name,omitempty"`
+	Input any    `json:"input,omitempty"`
+
+	// For tool_result blocks
+	ToolUseID string `json:"tool_use_id,omitempty"`
+	Content   any    `json:"content,omitempty"` // string or []ContentBlock
+	IsError   bool   `json:"is_error,omitempty"`
+
+	// For thinking blocks - pointer so field only appears when set (SDK requires it for accumulation)
+	Thinking  *string `json:"thinking,omitempty"`
+	Signature string  `json:"signature,omitempty"`
+}
+
+// ImageSource represents the source of an image
+type ImageSource struct {
+	Type      string `json:"type"` // "base64" or "url"
+	MediaType string `json:"media_type,omitempty"`
+	Data      string `json:"data,omitempty"`
+	URL       string `json:"url,omitempty"`
+}
+
+// Tool represents a tool definition
+type Tool struct {
+	Type        string          `json:"type,omitempty"` // "custom" for user-defined tools
+	Name        string          `json:"name"`
+	Description string          `json:"description,omitempty"`
+	InputSchema json.RawMessage `json:"input_schema,omitempty"`
+}
+
+// ToolChoice controls how the model uses tools
+type ToolChoice struct {
+	Type                   string `json:"type"` // "auto", "any", "tool", "none"
+	Name                   string `json:"name,omitempty"`
+	DisableParallelToolUse bool   `json:"disable_parallel_tool_use,omitempty"`
+}
+
+// ThinkingConfig controls extended thinking
+type ThinkingConfig struct {
+	Type         string `json:"type"` // "enabled" or "disabled"
+	BudgetTokens int    `json:"budget_tokens,omitempty"`
+}
+
+// Metadata for the request
+type Metadata struct {
+	UserID string `json:"user_id,omitempty"`
+}
+
+// Response types
+
+// MessagesResponse represents an Anthropic Messages API response
+type MessagesResponse struct {
+	ID           string         `json:"id"`
+	Type         string         `json:"type"` // "message"
+	Role         string         `json:"role"` // "assistant"
+	Model        string         `json:"model"`
+	Content      []ContentBlock `json:"content"`
+	StopReason   string         `json:"stop_reason,omitempty"`
+	StopSequence string         `json:"stop_sequence,omitempty"`
+	Usage        Usage          `json:"usage"`
+}
+
+// Usage contains token usage information
+type Usage struct {
+	InputTokens  int `json:"input_tokens"`
+	OutputTokens int `json:"output_tokens"`
+}
+
+// Streaming event types
+
+// MessageStartEvent is sent at the start of streaming
+type MessageStartEvent struct {
+	Type    string           `json:"type"` // "message_start"
+	Message MessagesResponse `json:"message"`
+}
+
+// ContentBlockStartEvent signals the start of a content block
+type ContentBlockStartEvent struct {
+	Type         string       `json:"type"` // "content_block_start"
+	Index        int          `json:"index"`
+	ContentBlock ContentBlock `json:"content_block"`
+}
+
+// ContentBlockDeltaEvent contains incremental content updates
+type ContentBlockDeltaEvent struct {
+	Type  string `json:"type"` // "content_block_delta"
+	Index int    `json:"index"`
+	Delta Delta  `json:"delta"`
+}
+
+// Delta represents an incremental update
+type Delta struct {
+	Type        string `json:"type"` // "text_delta", "input_json_delta", "thinking_delta", "signature_delta"
+	Text        string `json:"text,omitempty"`
+	PartialJSON string `json:"partial_json,omitempty"`
+	Thinking    string `json:"thinking,omitempty"`
+	Signature   string `json:"signature,omitempty"`
+}
+
+// ContentBlockStopEvent signals the end of a content block
+type ContentBlockStopEvent struct {
+	Type  string `json:"type"` // "content_block_stop"
+	Index int    `json:"index"`
+}
+
+// MessageDeltaEvent contains updates to the message
+type MessageDeltaEvent struct {
+	Type  string       `json:"type"` // "message_delta"
+	Delta MessageDelta `json:"delta"`
+	Usage DeltaUsage   `json:"usage"`
+}
+
+// MessageDelta contains stop information
+type MessageDelta struct {
+	StopReason   string `json:"stop_reason,omitempty"`
+	StopSequence string `json:"stop_sequence,omitempty"`
+}
+
+// DeltaUsage contains cumulative token usage
+type DeltaUsage struct {
+	OutputTokens int `json:"output_tokens"`
+}
+
+// MessageStopEvent signals the end of the message
+type MessageStopEvent struct {
+	Type string `json:"type"` // "message_stop"
+}
+
+// PingEvent is a keepalive event
+type PingEvent struct {
+	Type string `json:"type"` // "ping"
+}
+
+// StreamErrorEvent is an error during streaming
+type StreamErrorEvent struct {
+	Type  string `json:"type"` // "error"
+	Error Error  `json:"error"`
+}
+
+// FromMessagesRequest converts an Anthropic MessagesRequest to an Ollama api.ChatRequest
+func FromMessagesRequest(r MessagesRequest) (*api.ChatRequest, error) {
+	var messages []api.Message
+
+	if r.System != nil {
+		switch sys := r.System.(type) {
+		case string:
+			if sys != "" {
+				messages = append(messages, api.Message{Role: "system", Content: sys})
+			}
+		case []any:
+			// System can be an array of content blocks
+			var content strings.Builder
+			for _, block := range sys {
+				if blockMap, ok := block.(map[string]any); ok {
+					if blockMap["type"] == "text" {
+						if text, ok := blockMap["text"].(string); ok {
+							content.WriteString(text)
+						}
+					}
+				}
+			}
+			if content.Len() > 0 {
+				messages = append(messages, api.Message{Role: "system", Content: content.String()})
+			}
+		}
+	}
+
+	for _, msg := range r.Messages {
+		converted, err := convertMessage(msg)
+		if err != nil {
+			return nil, err
+		}
+		messages = append(messages, converted...)
+	}
+
+	options := make(map[string]any)
+
+	options["num_predict"] = r.MaxTokens
+
+	if r.Temperature != nil {
+		options["temperature"] = *r.Temperature
+	}
+
+	if r.TopP != nil {
+		options["top_p"] = *r.TopP
+	}
+
+	if r.TopK != nil {
+		options["top_k"] = *r.TopK
+	}
+
+	if len(r.StopSequences) > 0 {
+		options["stop"] = r.StopSequences
+	}
+
+	var tools api.Tools
+	for _, t := range r.Tools {
+		tool, err := convertTool(t)
+		if err != nil {
+			return nil, err
+		}
+		tools = append(tools, tool)
+	}
+
+	var think *api.ThinkValue
+	if r.Thinking != nil && r.Thinking.Type == "enabled" {
+		think = &api.ThinkValue{Value: true}
+	}
+
+	stream := r.Stream
+
+	return &api.ChatRequest{
+		Model:    r.Model,
+		Messages: messages,
+		Options:  options,
+		Stream:   &stream,
+		Tools:    tools,
+		Think:    think,
+	}, nil
+}
+
+// convertMessage converts an Anthropic MessageParam to Ollama api.Message(s)
+func convertMessage(msg MessageParam) ([]api.Message, error) {
+	var messages []api.Message
+	role := strings.ToLower(msg.Role)
+
+	switch content := msg.Content.(type) {
+	case string:
+		messages = append(messages, api.Message{Role: role, Content: content})
+
+	case []any:
+		var textContent strings.Builder
+		var images []api.ImageData
+		var toolCalls []api.ToolCall
+		var thinking string
+		var toolResults []api.Message
+
+		for _, block := range content {
+			blockMap, ok := block.(map[string]any)
+			if !ok {
+				return nil, errors.New("invalid content block format")
+			}
+
+			blockType, _ := blockMap["type"].(string)
+
+			switch blockType {
+			case "text":
+				if text, ok := blockMap["text"].(string); ok {
+					textContent.WriteString(text)
+				}
+
+			case "image":
+				source, ok := blockMap["source"].(map[string]any)
+				if !ok {
+					return nil, errors.New("invalid image source")
+				}
+
+				sourceType, _ := source["type"].(string)
+				if sourceType == "base64" {
+					data, _ := source["data"].(string)
+					decoded, err := base64.StdEncoding.DecodeString(data)
+					if err != nil {
+						return nil, fmt.Errorf("invalid base64 image data: %w", err)
+					}
+					images = append(images, decoded)
+				} else {
+					return nil, fmt.Errorf("invalid image source type: %s. Only base64 images are supported.", sourceType)
+				}
+				// URL images would need to be fetched - skip for now
+
+			case "tool_use":
+				id, ok := blockMap["id"].(string)
+				if !ok {
+					return nil, errors.New("tool_use block missing required 'id' field")
+				}
+				name, ok := blockMap["name"].(string)
+				if !ok {
+					return nil, errors.New("tool_use block missing required 'name' field")
+				}
+				tc := api.ToolCall{
+					ID: id,
+					Function: api.ToolCallFunction{
+						Name: name,
+					},
+				}
+				if input, ok := blockMap["input"].(map[string]any); ok {
+					tc.Function.Arguments = api.ToolCallFunctionArguments(input)
+				}
+				toolCalls = append(toolCalls, tc)
+
+			case "tool_result":
+				toolUseID, _ := blockMap["tool_use_id"].(string)
+				var resultContent string
+
+				switch c := blockMap["content"].(type) {
+				case string:
+					resultContent = c
+				case []any:
+					for _, cb := range c {
+						if cbMap, ok := cb.(map[string]any); ok {
+							if cbMap["type"] == "text" {
+								if text, ok := cbMap["text"].(string); ok {
+									resultContent += text
+								}
+							}
+						}
+					}
+				}
+
+				toolResults = append(toolResults, api.Message{
+					Role:       "tool",
+					Content:    resultContent,
+					ToolCallID: toolUseID,
+				})
+
+			case "thinking":
+				if t, ok := blockMap["thinking"].(string); ok {
+					thinking = t
+				}
+			}
+		}
+
+		if textContent.Len() > 0 || len(images) > 0 || len(toolCalls) > 0 || thinking != "" {
+			m := api.Message{
+				Role:      role,
+				Content:   textContent.String(),
+				Images:    images,
+				ToolCalls: toolCalls,
+				Thinking:  thinking,
+			}
+			messages = append(messages, m)
+		}
+
+		// Add tool results as separate messages
+		messages = append(messages, toolResults...)
+
+	default:
+		return nil, fmt.Errorf("invalid message content type: %T", content)
+	}
+
+	return messages, nil
+}
+
+// convertTool converts an Anthropic Tool to an Ollama api.Tool
+func convertTool(t Tool) (api.Tool, error) {
+	var params api.ToolFunctionParameters
+	if len(t.InputSchema) > 0 {
+		if err := json.Unmarshal(t.InputSchema, &params); err != nil {
+			return api.Tool{}, fmt.Errorf("invalid input_schema for tool %q: %w", t.Name, err)
+		}
+	}
+
+	return api.Tool{
+		Type: "function",
+		Function: api.ToolFunction{
+			Name:        t.Name,
+			Description: t.Description,
+			Parameters:  params,
+		},
+	}, nil
+}
+
+// ToMessagesResponse converts an Ollama api.ChatResponse to an Anthropic MessagesResponse
+func ToMessagesResponse(id string, r api.ChatResponse) MessagesResponse {
+	var content []ContentBlock
+
+	if r.Message.Thinking != "" {
+		content = append(content, ContentBlock{
+			Type:     "thinking",
+			Thinking: ptr(r.Message.Thinking),
+		})
+	}
+
+	if r.Message.Content != "" {
+		content = append(content, ContentBlock{
+			Type: "text",
+			Text: ptr(r.Message.Content),
+		})
+	}
+
+	for _, tc := range r.Message.ToolCalls {
+		content = append(content, ContentBlock{
+			Type:  "tool_use",
+			ID:    tc.ID,
+			Name:  tc.Function.Name,
+			Input: tc.Function.Arguments,
+		})
+	}
+
+	stopReason := mapStopReason(r.DoneReason, len(r.Message.ToolCalls) > 0)
+
+	return MessagesResponse{
+		ID:         id,
+		Type:       "message",
+		Role:       "assistant",
+		Model:      r.Model,
+		Content:    content,
+		StopReason: stopReason,
+		Usage: Usage{
+			InputTokens:  r.Metrics.PromptEvalCount,
+			OutputTokens: r.Metrics.EvalCount,
+		},
+	}
+}
+
+// mapStopReason converts Ollama done_reason to Anthropic stop_reason
+func mapStopReason(reason string, hasToolCalls bool) string {
+	if hasToolCalls {
+		return "tool_use"
+	}
+
+	switch reason {
+	case "stop":
+		return "end_turn"
+	case "length":
+		return "max_tokens"
+	default:
+		if reason != "" {
+			return "stop_sequence"
+		}
+		return ""
+	}
+}
+
+// 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
+	thinkingStarted bool
+	thinkingDone    bool
+	textStarted     bool
+	toolCallsSent   map[string]bool
+}
+
+func NewStreamConverter(id, model string) *StreamConverter {
+	return &StreamConverter{
+		ID:            id,
+		Model:         model,
+		firstWrite:    true,
+		toolCallsSent: make(map[string]bool),
+	}
+}
+
+// StreamEvent represents a streaming event to be sent to the client
+type StreamEvent struct {
+	Event string
+	Data  any
+}
+
+// Process converts an Ollama ChatResponse to Anthropic streaming events
+func (c *StreamConverter) Process(r api.ChatResponse) []StreamEvent {
+	var events []StreamEvent
+
+	if c.firstWrite {
+		c.firstWrite = false
+		c.inputTokens = r.Metrics.PromptEvalCount
+
+		events = append(events, StreamEvent{
+			Event: "message_start",
+			Data: MessageStartEvent{
+				Type: "message_start",
+				Message: MessagesResponse{
+					ID:      c.ID,
+					Type:    "message",
+					Role:    "assistant",
+					Model:   c.Model,
+					Content: []ContentBlock{},
+					Usage: Usage{
+						InputTokens:  c.inputTokens,
+						OutputTokens: 0,
+					},
+				},
+			},
+		})
+	}
+
+	if r.Message.Thinking != "" && !c.thinkingDone {
+		if !c.thinkingStarted {
+			c.thinkingStarted = true
+			events = append(events, StreamEvent{
+				Event: "content_block_start",
+				Data: ContentBlockStartEvent{
+					Type:  "content_block_start",
+					Index: c.contentIndex,
+					ContentBlock: ContentBlock{
+						Type:     "thinking",
+						Thinking: ptr(""),
+					},
+				},
+			})
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type:     "thinking_delta",
+					Thinking: r.Message.Thinking,
+				},
+			},
+		})
+	}
+
+	if r.Message.Content != "" {
+		if c.thinkingStarted && !c.thinkingDone {
+			c.thinkingDone = true
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+			c.contentIndex++
+		}
+
+		if !c.textStarted {
+			c.textStarted = true
+			events = append(events, StreamEvent{
+				Event: "content_block_start",
+				Data: ContentBlockStartEvent{
+					Type:  "content_block_start",
+					Index: c.contentIndex,
+					ContentBlock: ContentBlock{
+						Type: "text",
+						Text: ptr(""),
+					},
+				},
+			})
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type: "text_delta",
+					Text: r.Message.Content,
+				},
+			},
+		})
+	}
+
+	for _, tc := range r.Message.ToolCalls {
+		if c.toolCallsSent[tc.ID] {
+			continue
+		}
+
+		if c.textStarted {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+			c.contentIndex++
+			c.textStarted = false
+		}
+
+		argsJSON, err := json.Marshal(tc.Function.Arguments)
+		if err != nil {
+			slog.Error("failed to marshal tool arguments", "error", err, "tool_id", tc.ID)
+			continue
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_start",
+			Data: ContentBlockStartEvent{
+				Type:  "content_block_start",
+				Index: c.contentIndex,
+				ContentBlock: ContentBlock{
+					Type:  "tool_use",
+					ID:    tc.ID,
+					Name:  tc.Function.Name,
+					Input: map[string]any{},
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type:        "input_json_delta",
+					PartialJSON: string(argsJSON),
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "content_block_stop",
+			Data: ContentBlockStopEvent{
+				Type:  "content_block_stop",
+				Index: c.contentIndex,
+			},
+		})
+
+		c.toolCallsSent[tc.ID] = true
+		c.contentIndex++
+	}
+
+	if r.Done {
+		if c.textStarted {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+		} else if c.thinkingStarted && !c.thinkingDone {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+		}
+
+		c.outputTokens = r.Metrics.EvalCount
+		stopReason := mapStopReason(r.DoneReason, len(c.toolCallsSent) > 0)
+
+		events = append(events, StreamEvent{
+			Event: "message_delta",
+			Data: MessageDeltaEvent{
+				Type: "message_delta",
+				Delta: MessageDelta{
+					StopReason: stopReason,
+				},
+				Usage: DeltaUsage{
+					OutputTokens: c.outputTokens,
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "message_stop",
+			Data: MessageStopEvent{
+				Type: "message_stop",
+			},
+		})
+	}
+
+	return events
+}
+
+// generateID generates a unique ID with the given prefix using crypto/rand
+func generateID(prefix string) string {
+	b := make([]byte, 12)
+	if _, err := rand.Read(b); err != nil {
+		// Fallback to time-based ID if crypto/rand fails
+		return fmt.Sprintf("%s_%d", prefix, time.Now().UnixNano())
+	}
+	return fmt.Sprintf("%s_%x", prefix, b)
+}
+
+// GenerateMessageID generates a unique message ID
+func GenerateMessageID() string {
+	return generateID("msg")
+}
+
+// ptr returns a pointer to the given string value
+func ptr(s string) *string {
+	return &s
+}

anthropic/anthropic_test.go

@@ -0,0 +1,940 @@
+package anthropic
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+const (
+	testImage = `iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=`
+)
+
+func TestFromMessagesRequest_Basic(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Model != "test-model" {
+		t.Errorf("expected model 'test-model', got %q", result.Model)
+	}
+
+	if len(result.Messages) != 1 {
+		t.Fatalf("expected 1 message, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Role != "user" || result.Messages[0].Content != "Hello" {
+		t.Errorf("unexpected message: %+v", result.Messages[0])
+	}
+
+	if numPredict, ok := result.Options["num_predict"].(int); !ok || numPredict != 1024 {
+		t.Errorf("expected num_predict 1024, got %v", result.Options["num_predict"])
+	}
+}
+
+func TestFromMessagesRequest_WithSystemPrompt(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		System:    "You are a helpful assistant.",
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Role != "system" || result.Messages[0].Content != "You are a helpful assistant." {
+		t.Errorf("unexpected system message: %+v", result.Messages[0])
+	}
+}
+
+func TestFromMessagesRequest_WithSystemPromptArray(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		System: []any{
+			map[string]any{"type": "text", "text": "You are helpful."},
+			map[string]any{"type": "text", "text": " Be concise."},
+		},
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Content != "You are helpful. Be concise." {
+		t.Errorf("unexpected system message content: %q", result.Messages[0].Content)
+	}
+}
+
+func TestFromMessagesRequest_WithOptions(t *testing.T) {
+	temp := 0.7
+	topP := 0.9
+	topK := 40
+	req := MessagesRequest{
+		Model:         "test-model",
+		MaxTokens:     2048,
+		Messages:      []MessageParam{{Role: "user", Content: "Hello"}},
+		Temperature:   &temp,
+		TopP:          &topP,
+		TopK:          &topK,
+		StopSequences: []string{"\n", "END"},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Options["temperature"] != 0.7 {
+		t.Errorf("expected temperature 0.7, got %v", result.Options["temperature"])
+	}
+	if result.Options["top_p"] != 0.9 {
+		t.Errorf("expected top_p 0.9, got %v", result.Options["top_p"])
+	}
+	if result.Options["top_k"] != 40 {
+		t.Errorf("expected top_k 40, got %v", result.Options["top_k"])
+	}
+	if diff := cmp.Diff([]string{"\n", "END"}, result.Options["stop"]); diff != "" {
+		t.Errorf("stop sequences mismatch: %s", diff)
+	}
+}
+
+func TestFromMessagesRequest_WithImage(t *testing.T) {
+	imgData, _ := base64.StdEncoding.DecodeString(testImage)
+
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "user",
+				Content: []any{
+					map[string]any{"type": "text", "text": "What's in this image?"},
+					map[string]any{
+						"type": "image",
+						"source": map[string]any{
+							"type":       "base64",
+							"media_type": "image/png",
+							"data":       testImage,
+						},
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 1 {
+		t.Fatalf("expected 1 message, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Content != "What's in this image?" {
+		t.Errorf("expected content 'What's in this image?', got %q", result.Messages[0].Content)
+	}
+
+	if len(result.Messages[0].Images) != 1 {
+		t.Fatalf("expected 1 image, got %d", len(result.Messages[0].Images))
+	}
+
+	if string(result.Messages[0].Images[0]) != string(imgData) {
+		t.Error("image data mismatch")
+	}
+}
+
+func TestFromMessagesRequest_WithToolUse(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "What's the weather in Paris?"},
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type":  "tool_use",
+						"id":    "call_123",
+						"name":  "get_weather",
+						"input": map[string]any{"location": "Paris"},
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if len(result.Messages[1].ToolCalls) != 1 {
+		t.Fatalf("expected 1 tool call, got %d", len(result.Messages[1].ToolCalls))
+	}
+
+	tc := result.Messages[1].ToolCalls[0]
+	if tc.ID != "call_123" {
+		t.Errorf("expected tool call ID 'call_123', got %q", tc.ID)
+	}
+	if tc.Function.Name != "get_weather" {
+		t.Errorf("expected tool name 'get_weather', got %q", tc.Function.Name)
+	}
+}
+
+func TestFromMessagesRequest_WithToolResult(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "user",
+				Content: []any{
+					map[string]any{
+						"type":        "tool_result",
+						"tool_use_id": "call_123",
+						"content":     "The weather in Paris is sunny, 22°C",
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 1 {
+		t.Fatalf("expected 1 message, got %d", len(result.Messages))
+	}
+
+	msg := result.Messages[0]
+	if msg.Role != "tool" {
+		t.Errorf("expected role 'tool', got %q", msg.Role)
+	}
+	if msg.ToolCallID != "call_123" {
+		t.Errorf("expected tool_call_id 'call_123', got %q", msg.ToolCallID)
+	}
+	if msg.Content != "The weather in Paris is sunny, 22°C" {
+		t.Errorf("unexpected content: %q", msg.Content)
+	}
+}
+
+func TestFromMessagesRequest_WithTools(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Tools: []Tool{
+			{
+				Name:        "get_weather",
+				Description: "Get current weather",
+				InputSchema: json.RawMessage(`{"type":"object","properties":{"location":{"type":"string"}},"required":["location"]}`),
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Tools) != 1 {
+		t.Fatalf("expected 1 tool, got %d", len(result.Tools))
+	}
+
+	tool := result.Tools[0]
+	if tool.Type != "function" {
+		t.Errorf("expected type 'function', got %q", tool.Type)
+	}
+	if tool.Function.Name != "get_weather" {
+		t.Errorf("expected name 'get_weather', got %q", tool.Function.Name)
+	}
+	if tool.Function.Description != "Get current weather" {
+		t.Errorf("expected description 'Get current weather', got %q", tool.Function.Description)
+	}
+}
+
+func TestFromMessagesRequest_WithThinking(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Thinking:  &ThinkingConfig{Type: "enabled", BudgetTokens: 1000},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Think == nil {
+		t.Fatal("expected Think to be set")
+	}
+	if v, ok := result.Think.Value.(bool); !ok || !v {
+		t.Errorf("expected Think.Value to be true, got %v", result.Think.Value)
+	}
+}
+
+// TestFromMessagesRequest_ThinkingOnlyBlock verifies that messages containing only
+// a thinking block (no text, images, or tool calls) are preserved and not dropped.
+func TestFromMessagesRequest_ThinkingOnlyBlock(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type":     "thinking",
+						"thinking": "Let me think about this...",
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	assistantMsg := result.Messages[1]
+	if assistantMsg.Thinking != "Let me think about this..." {
+		t.Errorf("expected thinking content, got %q", assistantMsg.Thinking)
+	}
+}
+
+func TestFromMessagesRequest_ToolUseMissingID(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type": "tool_use",
+						"name": "get_weather",
+					},
+				},
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for missing tool_use id")
+	}
+	if err.Error() != "tool_use block missing required 'id' field" {
+		t.Errorf("unexpected error message: %v", err)
+	}
+}
+
+func TestFromMessagesRequest_ToolUseMissingName(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type": "tool_use",
+						"id":   "call_123",
+					},
+				},
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for missing tool_use name")
+	}
+	if err.Error() != "tool_use block missing required 'name' field" {
+		t.Errorf("unexpected error message: %v", err)
+	}
+}
+
+func TestFromMessagesRequest_InvalidToolSchema(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Tools: []Tool{
+			{
+				Name:        "bad_tool",
+				InputSchema: json.RawMessage(`{invalid json`),
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for invalid tool schema")
+	}
+}
+
+func TestToMessagesResponse_Basic(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: "Hello there!",
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics: api.Metrics{
+			PromptEvalCount: 10,
+			EvalCount:       5,
+		},
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if result.ID != "msg_123" {
+		t.Errorf("expected ID 'msg_123', got %q", result.ID)
+	}
+	if result.Type != "message" {
+		t.Errorf("expected type 'message', got %q", result.Type)
+	}
+	if result.Role != "assistant" {
+		t.Errorf("expected role 'assistant', got %q", result.Role)
+	}
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "text" || result.Content[0].Text == nil || *result.Content[0].Text != "Hello there!" {
+		t.Errorf("unexpected content: %+v", result.Content[0])
+	}
+	if result.StopReason != "end_turn" {
+		t.Errorf("expected stop_reason 'end_turn', got %q", result.StopReason)
+	}
+	if result.Usage.InputTokens != 10 || result.Usage.OutputTokens != 5 {
+		t.Errorf("unexpected usage: %+v", result.Usage)
+	}
+}
+
+func TestToMessagesResponse_WithToolCalls(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_123",
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"location": "Paris"},
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "tool_use" {
+		t.Errorf("expected type 'tool_use', got %q", result.Content[0].Type)
+	}
+	if result.Content[0].ID != "call_123" {
+		t.Errorf("expected ID 'call_123', got %q", result.Content[0].ID)
+	}
+	if result.Content[0].Name != "get_weather" {
+		t.Errorf("expected name 'get_weather', got %q", result.Content[0].Name)
+	}
+	if result.StopReason != "tool_use" {
+		t.Errorf("expected stop_reason 'tool_use', got %q", result.StopReason)
+	}
+}
+
+func TestToMessagesResponse_WithThinking(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:     "assistant",
+			Content:  "The answer is 42.",
+			Thinking: "Let me think about this...",
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if len(result.Content) != 2 {
+		t.Fatalf("expected 2 content blocks, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "thinking" {
+		t.Errorf("expected first block type 'thinking', got %q", result.Content[0].Type)
+	}
+	if result.Content[0].Thinking == nil || *result.Content[0].Thinking != "Let me think about this..." {
+		t.Errorf("unexpected thinking content: %v", result.Content[0].Thinking)
+	}
+	if result.Content[1].Type != "text" {
+		t.Errorf("expected second block type 'text', got %q", result.Content[1].Type)
+	}
+}
+
+func TestMapStopReason(t *testing.T) {
+	tests := []struct {
+		reason       string
+		hasToolCalls bool
+		want         string
+	}{
+		{"stop", false, "end_turn"},
+		{"length", false, "max_tokens"},
+		{"stop", true, "tool_use"},
+		{"other", false, "stop_sequence"},
+		{"", false, ""},
+	}
+
+	for _, tt := range tests {
+		got := mapStopReason(tt.reason, tt.hasToolCalls)
+		if got != tt.want {
+			t.Errorf("mapStopReason(%q, %v) = %q, want %q", tt.reason, tt.hasToolCalls, got, tt.want)
+		}
+	}
+}
+
+func TestNewError(t *testing.T) {
+	tests := []struct {
+		code int
+		want string
+	}{
+		{400, "invalid_request_error"},
+		{401, "authentication_error"},
+		{403, "permission_error"},
+		{404, "not_found_error"},
+		{429, "rate_limit_error"},
+		{500, "api_error"},
+		{503, "overloaded_error"},
+		{529, "overloaded_error"},
+	}
+
+	for _, tt := range tests {
+		result := NewError(tt.code, "test message")
+		if result.Type != "error" {
+			t.Errorf("NewError(%d) type = %q, want 'error'", tt.code, result.Type)
+		}
+		if result.Error.Type != tt.want {
+			t.Errorf("NewError(%d) error.type = %q, want %q", tt.code, result.Error.Type, tt.want)
+		}
+		if result.Error.Message != "test message" {
+			t.Errorf("NewError(%d) message = %q, want 'test message'", tt.code, result.Error.Message)
+		}
+		if result.RequestID == "" {
+			t.Errorf("NewError(%d) request_id should not be empty", tt.code)
+		}
+	}
+}
+
+func TestGenerateMessageID(t *testing.T) {
+	id1 := GenerateMessageID()
+	id2 := GenerateMessageID()
+
+	if id1 == "" {
+		t.Error("GenerateMessageID returned empty string")
+	}
+	if id1 == id2 {
+		t.Error("GenerateMessageID returned duplicate IDs")
+	}
+	if len(id1) < 10 {
+		t.Errorf("GenerateMessageID returned short ID: %q", id1)
+	}
+	if id1[:4] != "msg_" {
+		t.Errorf("GenerateMessageID should start with 'msg_', got %q", id1[:4])
+	}
+}
+
+func TestStreamConverter_Basic(t *testing.T) {
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	// First chunk
+	resp1 := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: "Hello",
+		},
+		Metrics: api.Metrics{PromptEvalCount: 10},
+	}
+
+	events1 := conv.Process(resp1)
+	if len(events1) < 3 {
+		t.Fatalf("expected at least 3 events for first chunk, got %d", len(events1))
+	}
+
+	// Should have message_start, content_block_start, content_block_delta
+	if events1[0].Event != "message_start" {
+		t.Errorf("expected first event 'message_start', got %q", events1[0].Event)
+	}
+	if events1[1].Event != "content_block_start" {
+		t.Errorf("expected second event 'content_block_start', got %q", events1[1].Event)
+	}
+	if events1[2].Event != "content_block_delta" {
+		t.Errorf("expected third event 'content_block_delta', got %q", events1[2].Event)
+	}
+
+	// Final chunk
+	resp2 := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: " world!",
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics:    api.Metrics{EvalCount: 5},
+	}
+
+	events2 := conv.Process(resp2)
+
+	// Should have content_block_delta, content_block_stop, message_delta, message_stop
+	hasStop := false
+	for _, e := range events2 {
+		if e.Event == "message_stop" {
+			hasStop = true
+		}
+	}
+	if !hasStop {
+		t.Error("expected message_stop event in final chunk")
+	}
+}
+
+func TestStreamConverter_WithToolCalls(t *testing.T) {
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_123",
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"location": "Paris"},
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics:    api.Metrics{PromptEvalCount: 10, EvalCount: 5},
+	}
+
+	events := conv.Process(resp)
+
+	hasToolStart := false
+	hasToolDelta := false
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					hasToolStart = true
+				}
+			}
+		}
+		if e.Event == "content_block_delta" {
+			if delta, ok := e.Data.(ContentBlockDeltaEvent); ok {
+				if delta.Delta.Type == "input_json_delta" {
+					hasToolDelta = true
+				}
+			}
+		}
+	}
+
+	if !hasToolStart {
+		t.Error("expected tool_use content_block_start event")
+	}
+	if !hasToolDelta {
+		t.Error("expected input_json_delta event")
+	}
+}
+
+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")
+
+	// Create a channel which cannot be JSON marshaled
+	unmarshalable := make(chan int)
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_bad",
+					Function: api.ToolCallFunction{
+						Name:      "bad_function",
+						Arguments: map[string]any{"channel": unmarshalable},
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	// Should not panic and should skip the unmarshalable tool call
+	events := conv.Process(resp)
+
+	// Verify no tool_use block was started (since marshal failed before block start)
+	hasToolStart := false
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					hasToolStart = true
+				}
+			}
+		}
+	}
+
+	if hasToolStart {
+		t.Error("expected no tool_use block when arguments cannot be marshaled")
+	}
+}
+
+func TestStreamConverter_MultipleToolCallsWithMixedValidity(t *testing.T) {
+	// Test that valid tool calls still work when mixed with invalid ones
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	unmarshalable := make(chan int)
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_good",
+					Function: api.ToolCallFunction{
+						Name:      "good_function",
+						Arguments: map[string]any{"location": "Paris"},
+					},
+				},
+				{
+					ID: "call_bad",
+					Function: api.ToolCallFunction{
+						Name:      "bad_function",
+						Arguments: map[string]any{"channel": unmarshalable},
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	events := conv.Process(resp)
+
+	// Count tool_use blocks - should only have 1 (the valid one)
+	toolStartCount := 0
+	toolDeltaCount := 0
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					toolStartCount++
+					if start.ContentBlock.Name != "good_function" {
+						t.Errorf("expected tool name 'good_function', got %q", start.ContentBlock.Name)
+					}
+				}
+			}
+		}
+		if e.Event == "content_block_delta" {
+			if delta, ok := e.Data.(ContentBlockDeltaEvent); ok {
+				if delta.Delta.Type == "input_json_delta" {
+					toolDeltaCount++
+				}
+			}
+		}
+	}
+
+	if toolStartCount != 1 {
+		t.Errorf("expected 1 tool_use block, got %d", toolStartCount)
+	}
+	if toolDeltaCount != 1 {
+		t.Errorf("expected 1 input_json_delta, got %d", toolDeltaCount)
+	}
+}
+
+// TestContentBlockJSON_EmptyFieldsPresent verifies that empty text and thinking fields
+// are serialized in JSON output. The Anthropic SDK requires these fields to be present
+// (even when empty) in content_block_start events to properly accumulate streaming deltas.
+// Without these fields, the SDK throws: "TypeError: unsupported operand type(s) for +=: 'NoneType' and 'str'"
+func TestContentBlockJSON_EmptyFieldsPresent(t *testing.T) {
+	tests := []struct {
+		name     string
+		block    ContentBlock
+		wantKeys []string
+	}{
+		{
+			name: "text block includes empty text field",
+			block: ContentBlock{
+				Type: "text",
+				Text: ptr(""),
+			},
+			wantKeys: []string{"type", "text"},
+		},
+		{
+			name: "thinking block includes empty thinking field",
+			block: ContentBlock{
+				Type:     "thinking",
+				Thinking: ptr(""),
+			},
+			wantKeys: []string{"type", "thinking"},
+		},
+		{
+			name: "text block with content",
+			block: ContentBlock{
+				Type: "text",
+				Text: ptr("hello"),
+			},
+			wantKeys: []string{"type", "text"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			data, err := json.Marshal(tt.block)
+			if err != nil {
+				t.Fatalf("failed to marshal: %v", err)
+			}
+
+			var result map[string]any
+			if err := json.Unmarshal(data, &result); err != nil {
+				t.Fatalf("failed to unmarshal: %v", err)
+			}
+
+			for _, key := range tt.wantKeys {
+				if _, ok := result[key]; !ok {
+					t.Errorf("expected key %q to be present in JSON output, got: %s", key, string(data))
+				}
+			}
+		})
+	}
+}
+
+// TestStreamConverter_ContentBlockStartIncludesEmptyFields verifies that content_block_start
+// 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")
+
+		resp := api.ChatResponse{
+			Model:   "test-model",
+			Message: api.Message{Role: "assistant", Content: "hello"},
+		}
+
+		events := conv.Process(resp)
+
+		var foundTextStart bool
+		for _, e := range events {
+			if e.Event == "content_block_start" {
+				if start, ok := e.Data.(ContentBlockStartEvent); ok {
+					if start.ContentBlock.Type == "text" {
+						foundTextStart = true
+						// Marshal and verify the text field is present
+						data, _ := json.Marshal(start)
+						var result map[string]any
+						json.Unmarshal(data, &result)
+						cb := result["content_block"].(map[string]any)
+						if _, ok := cb["text"]; !ok {
+							t.Error("content_block_start for text should include 'text' field")
+						}
+					}
+				}
+			}
+		}
+
+		if !foundTextStart {
+			t.Error("expected text content_block_start event")
+		}
+	})
+
+	t.Run("thinking block start includes empty thinking", func(t *testing.T) {
+		conv := NewStreamConverter("msg_123", "test-model")
+
+		resp := api.ChatResponse{
+			Model:   "test-model",
+			Message: api.Message{Role: "assistant", Thinking: "let me think..."},
+		}
+
+		events := conv.Process(resp)
+
+		var foundThinkingStart bool
+		for _, e := range events {
+			if e.Event == "content_block_start" {
+				if start, ok := e.Data.(ContentBlockStartEvent); ok {
+					if start.ContentBlock.Type == "thinking" {
+						foundThinkingStart = true
+						data, _ := json.Marshal(start)
+						var result map[string]any
+						json.Unmarshal(data, &result)
+						cb := result["content_block"].(map[string]any)
+						if _, ok := cb["thinking"]; !ok {
+							t.Error("content_block_start for thinking should include 'thinking' field")
+						}
+					}
+				}
+			}
+		}
+
+		if !foundThinkingStart {
+			t.Error("expected thinking content_block_start event")
+		}
+	})
+}

api/types.go

@@ -554,6 +554,9 @@ type CreateRequest struct {
 	Renderer string `json:"renderer,omitempty"`
 	Parser   string `json:"parser,omitempty"`
 
+	// Requires is the minimum version of Ollama required by the model.
+	Requires string `json:"requires,omitempty"`
+
 	// Info is a map of additional information for the model
 	Info map[string]any `json:"info,omitempty"`
 
@@ -604,6 +607,7 @@ type ShowResponse struct {
 	Tensors       []Tensor           `json:"tensors,omitempty"`
 	Capabilities  []model.Capability `json:"capabilities,omitempty"`
 	ModifiedAt    time.Time          `json:"modified_at,omitempty"`
+	Requires      string             `json:"requires,omitempty"`
 }
 
 // CopyRequest is the request passed to [Client.Copy].

cmd/cmd.go

@@ -943,6 +943,9 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 			rows = append(rows, []string{"", "parameters", resp.Details.ParameterSize})
 		}
 		rows = append(rows, []string{"", "quantization", resp.Details.QuantizationLevel})
+		if resp.Requires != "" {
+			rows = append(rows, []string{"", "requires", resp.Requires})
+		}
 		return
 	})
 

cmd/cmd_test.go

@@ -291,6 +291,31 @@ Weigh anchor!
 			t.Errorf("unexpected output (-want +got):\n%s", diff)
 		}
 	})
+
+	t.Run("min version", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			Requires: "0.14.0",
+		}, false, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test      
+    parameters      7B        
+    quantization    FP16      
+    requires        0.14.0    
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
 }
 
 func TestDeleteHandler(t *testing.T) {

docs/README.md

@@ -14,6 +14,7 @@
 * [API Reference](https://docs.ollama.com/api)
 * [Modelfile Reference](https://docs.ollama.com/modelfile)
 * [OpenAI Compatibility](https://docs.ollama.com/api/openai-compatibility)
+* [Anthropic Compatibility](./api/anthropic-compatibility.mdx)
 
 ### Resources
 

docs/api/anthropic-compatibility.mdx

@@ -0,0 +1,406 @@
+---
+title: Anthropic compatibility
+---
+
+Ollama provides compatibility with the [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) to help connect existing applications to Ollama, including tools like Claude Code.
+
+## Recommended models
+
+For coding use cases, models like `glm-4.7:cloud`, `minimax-m2.1:cloud`, and `qwen3-coder` are recommended.
+
+Pull a model before use:
+```shell
+ollama pull qwen3-coder
+ollama pull glm-4.7:cloud
+```
+
+## Usage
+
+### Environment variables
+
+To use Ollama with tools that expect the Anthropic API (like Claude Code), set these environment variables:
+
+```shell
+export ANTHROPIC_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama  # required but ignored
+```
+
+### Simple `/v1/messages` example
+
+<CodeGroup dropdown>
+
+```python basic.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',  # required but ignored
+)
+
+message = client.messages.create(
+    model='qwen3-coder',
+    max_tokens=1024,
+    messages=[
+        {'role': 'user', 'content': 'Hello, how are you?'}
+    ]
+)
+print(message.content[0].text)
+```
+
+```javascript basic.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama", // required but ignored
+});
+
+const message = await anthropic.messages.create({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  messages: [{ role: "user", content: "Hello, how are you?" }],
+});
+
+console.log(message.content[0].text);
+```
+
+```shell basic.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-H "x-api-key: ollama" \
+-H "anthropic-version: 2023-06-01" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "messages": [{ "role": "user", "content": "Hello, how are you?" }]
+}'
+```
+
+</CodeGroup>
+
+### Streaming example
+
+<CodeGroup dropdown>
+
+```python streaming.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',
+)
+
+with client.messages.stream(
+    model='qwen3-coder',
+    max_tokens=1024,
+    messages=[{'role': 'user', 'content': 'Count from 1 to 10'}]
+) as stream:
+    for text in stream.text_stream:
+        print(text, end='', flush=True)
+```
+
+```javascript streaming.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama",
+});
+
+const stream = await anthropic.messages.stream({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  messages: [{ role: "user", content: "Count from 1 to 10" }],
+});
+
+for await (const event of stream) {
+  if (
+    event.type === "content_block_delta" &&
+    event.delta.type === "text_delta"
+  ) {
+    process.stdout.write(event.delta.text);
+  }
+}
+```
+
+```shell streaming.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "stream": true,
+  "messages": [{ "role": "user", "content": "Count from 1 to 10" }]
+}'
+```
+
+</CodeGroup>
+
+### Tool calling example
+
+<CodeGroup dropdown>
+
+```python tools.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',
+)
+
+message = client.messages.create(
+    model='qwen3-coder',
+    max_tokens=1024,
+    tools=[
+        {
+            'name': 'get_weather',
+            'description': 'Get the current weather in a location',
+            'input_schema': {
+                'type': 'object',
+                'properties': {
+                    'location': {
+                        'type': 'string',
+                        'description': 'The city and state, e.g. San Francisco, CA'
+                    }
+                },
+                'required': ['location']
+            }
+        }
+    ],
+    messages=[{'role': 'user', 'content': "What's the weather in San Francisco?"}]
+)
+
+for block in message.content:
+    if block.type == 'tool_use':
+        print(f'Tool: {block.name}')
+        print(f'Input: {block.input}')
+```
+
+```javascript tools.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama",
+});
+
+const message = await anthropic.messages.create({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  tools: [
+    {
+      name: "get_weather",
+      description: "Get the current weather in a location",
+      input_schema: {
+        type: "object",
+        properties: {
+          location: {
+            type: "string",
+            description: "The city and state, e.g. San Francisco, CA",
+          },
+        },
+        required: ["location"],
+      },
+    },
+  ],
+  messages: [{ role: "user", content: "What's the weather in San Francisco?" }],
+});
+
+for (const block of message.content) {
+  if (block.type === "tool_use") {
+    console.log("Tool:", block.name);
+    console.log("Input:", block.input);
+  }
+}
+```
+
+```shell tools.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "tools": [
+    {
+      "name": "get_weather",
+      "description": "Get the current weather in a location",
+      "input_schema": {
+        "type": "object",
+        "properties": {
+          "location": {
+            "type": "string",
+            "description": "The city and state"
+          }
+        },
+        "required": ["location"]
+      }
+    }
+  ],
+  "messages": [{ "role": "user", "content": "What is the weather in San Francisco?" }]
+}'
+```
+
+</CodeGroup>
+
+## Using with Claude Code
+
+[Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend:
+
+```shell
+ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+```
+
+Or set the environment variables in your shell profile:
+
+```shell
+export ANTHROPIC_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama
+```
+
+Then run Claude Code with any Ollama model:
+
+```shell
+# Local models
+claude --model qwen3-coder
+claude --model gpt-oss:20b
+
+# Cloud models
+claude --model glm-4.7:cloud
+claude --model minimax-m2.1:cloud
+```
+
+## Endpoints
+
+### `/v1/messages`
+
+#### Supported features
+
+- [x] Messages
+- [x] Streaming
+- [x] System prompts
+- [x] Multi-turn conversations
+- [x] Vision (images)
+- [x] Tools (function calling)
+- [x] Tool results
+- [x] Thinking/extended thinking
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `max_tokens`
+- [x] `messages`
+  - [x] Text `content`
+  - [x] Image `content` (base64)
+  - [x] Array of content blocks
+  - [x] `tool_use` blocks
+  - [x] `tool_result` blocks
+  - [x] `thinking` blocks
+- [x] `system` (string or array)
+- [x] `stream`
+- [x] `temperature`
+- [x] `top_p`
+- [x] `top_k`
+- [x] `stop_sequences`
+- [x] `tools`
+- [x] `thinking`
+- [ ] `tool_choice`
+- [ ] `metadata`
+
+#### Supported response fields
+
+- [x] `id`
+- [x] `type`
+- [x] `role`
+- [x] `model`
+- [x] `content` (text, tool_use, thinking blocks)
+- [x] `stop_reason` (end_turn, max_tokens, tool_use)
+- [x] `usage` (input_tokens, output_tokens)
+
+#### Streaming events
+
+- [x] `message_start`
+- [x] `content_block_start`
+- [x] `content_block_delta` (text_delta, input_json_delta, thinking_delta)
+- [x] `content_block_stop`
+- [x] `message_delta`
+- [x] `message_stop`
+- [x] `ping`
+- [x] `error`
+
+## Models
+
+Ollama supports both local and cloud models.
+
+### Local models
+
+Pull a local model before use:
+
+```shell
+ollama pull qwen3-coder
+```
+
+Recommended local models:
+- `qwen3-coder` - Excellent for coding tasks
+- `gpt-oss:20b` - Strong general-purpose model
+
+### Cloud models
+
+Cloud models are available immediately without pulling:
+
+- `glm-4.7:cloud` - High-performance cloud model
+- `minimax-m2.1:cloud` - Fast cloud model
+
+### Default model names
+
+For tooling that relies on default Anthropic model names such as `claude-3-5-sonnet`, use `ollama cp` to copy an existing model name:
+
+```shell
+ollama cp qwen3-coder claude-3-5-sonnet
+```
+
+Afterwards, this new model name can be specified in the `model` field:
+
+```shell
+curl http://localhost:11434/v1/messages \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "claude-3-5-sonnet",
+        "max_tokens": 1024,
+        "messages": [
+            {
+                "role": "user",
+                "content": "Hello!"
+            }
+        ]
+    }'
+```
+
+## Differences from the Anthropic API
+
+### Behavior differences
+
+- API key is accepted but not validated
+- `anthropic-version` header is accepted but not used
+- Token counts are approximations based on the underlying model's tokenizer
+
+### Not supported
+
+The following Anthropic API features are not currently supported:
+
+| Feature | Description |
+|---------|-------------|
+| `/v1/messages/count_tokens` | Token counting endpoint |
+| `tool_choice` | Forcing specific tool use or disabling tools |
+| `metadata` | Request metadata (user_id) |
+| Prompt caching | `cache_control` blocks for caching prefixes |
+| Batches API | `/v1/messages/batches` for async batch processing |
+| Citations | `citations` content blocks |
+| PDF support | `document` content blocks with PDF files |
+| Server-sent errors | `error` events during streaming (errors return HTTP status) |
+
+### Partial support
+
+| Feature | Status |
+|---------|--------|
+| Image content | Base64 images supported; URL images not supported |
+| Extended thinking | Basic support; `budget_tokens` accepted but not enforced |

docs/docs.json

@@ -32,7 +32,9 @@
     "codeblocks": "system"
   },
   "contextual": {
-    "options": ["copy"]
+    "options": [
+      "copy"
+    ]
   },
   "navbar": {
     "links": [
@@ -52,7 +54,9 @@
       "display": "simple"
     },
     "examples": {
-      "languages": ["curl"]
+      "languages": [
+        "curl"
+      ]
     }
   },
   "redirects": [
@@ -97,6 +101,7 @@
           {
             "group": "Integrations",
             "pages": [
+              "/integrations/claude-code",
               "/integrations/vscode",
               "/integrations/jetbrains",
               "/integrations/codex",
@@ -139,7 +144,8 @@
               "/api/streaming",
               "/api/usage",
               "/api/errors",
-              "/api/openai-compatibility"
+              "/api/openai-compatibility",
+              "/api/anthropic-compatibility"
             ]
           },
           {

docs/faq.mdx

@@ -14,11 +14,11 @@ curl -fsSL https://ollama.com/install.sh | sh
 
 ## How can I view the logs?
 
-Review the [Troubleshooting](./troubleshooting.md) docs for more about using logs.
+Review the [Troubleshooting](./troubleshooting) docs for more about using logs.
 
 ## Is my GPU compatible with Ollama?
 
-Please refer to the [GPU docs](./gpu.md).
+Please refer to the [GPU docs](./gpu).
 
 ## How can I specify the context window size?
 

docs/gpu.mdx

@@ -33,7 +33,7 @@ Check your compute compatibility to see if your card is supported:
 | 5.0                | GeForce GTX         | `GTX 750 Ti` `GTX 750` `NVS 810`                                                                                               |
 |                    | Quadro              | `K2200` `K1200` `K620` `M1200` `M520` `M5000M` `M4000M` `M3000M` `M2000M` `M1000M` `K620M` `M600M` `M500M`                     |
 
-For building locally to support older GPUs, see [developer.md](./development.md#linux-cuda-nvidia)
+For building locally to support older GPUs, see [developer](./development#linux-cuda-nvidia)
 
 ### GPU Selection
 
@@ -54,7 +54,7 @@ sudo modprobe nvidia_uvm`
 
 Ollama supports the following AMD GPUs via the ROCm library:
 
-> [!NOTE]
+> **NOTE:**
 > Additional AMD GPU support is provided by the Vulkan Library - see below.
 
 
@@ -132,9 +132,9 @@ Ollama supports GPU acceleration on Apple devices via the Metal API.
 
 ## Vulkan GPU Support
 
-> [!NOTE]
+> **NOTE:**
 > Vulkan is currently an Experimental feature.  To enable, you must set OLLAMA_VULKAN=1 for the Ollama server as
-described in the [FAQ](faq.md#how-do-i-configure-ollama-server)
+described in the [FAQ](faq#how-do-i-configure-ollama-server)
 
 Additional GPU support on Windows and Linux is provided via
 [Vulkan](https://www.vulkan.org/). On Windows most GPU vendors drivers come
@@ -161,6 +161,6 @@ sudo setcap cap_perfmon+ep /usr/local/bin/ollama
 
 To select specific Vulkan GPU(s), you can set the environment variable
 `GGML_VK_VISIBLE_DEVICES` to one or more numeric IDs on the Ollama server as
-described in the [FAQ](faq.md#how-do-i-configure-ollama-server). If you
+described in the [FAQ](faq#how-do-i-configure-ollama-server). If you
 encounter any problems with Vulkan based GPUs, you can disable all Vulkan GPUs
 by setting `GGML_VK_VISIBLE_DEVICES=-1` 

docs/integrations/claude-code.mdx

@@ -0,0 +1,69 @@
+---
+title: Claude Code
+---
+
+## Install
+
+Install [Claude Code](https://code.claude.com/docs/en/overview):
+
+<CodeGroup>
+
+```shell macOS / Linux
+curl -fsSL https://claude.ai/install.sh | bash
+```
+
+```powershell Windows
+irm https://claude.ai/install.ps1 | iex
+```
+
+</CodeGroup>
+
+## Usage with Ollama
+
+Claude Code connects to Ollama using the Anthropic-compatible API.
+
+1. Set the environment variables:
+
+```shell
+export ANTHROPIC_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama
+```
+
+2. Run Claude Code with an Ollama model:
+
+```shell
+claude --model qwen3-coder
+```
+
+Or run with environment variables inline:
+
+```shell
+ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+```
+
+## Connecting to ollama.com
+
+1. Create an [API key](https://ollama.com/settings/keys) on ollama.com
+2. Set the environment variables:
+
+```shell
+export ANTHROPIC_BASE_URL=https://ollama.com
+export ANTHROPIC_API_KEY=<your-api-key>
+```
+
+3. Run Claude Code with a cloud model:
+
+```shell
+claude --model glm-4.7:cloud
+```
+
+## Recommended Models
+
+### Cloud models
+- `glm-4.7:cloud` - High-performance cloud model
+- `minimax-m2.1:cloud` - Fast cloud model
+- `qwen3-coder:480b` - Large coding model
+
+### Local models
+- `qwen3-coder` - Excellent for coding tasks
+- `gpt-oss:20b` - Strong general-purpose model

docs/modelfile.mdx

@@ -41,6 +41,7 @@ INSTRUCTION arguments
 | [`ADAPTER`](#adapter)               | Defines the (Q)LoRA adapters to apply to the model.            |
 | [`LICENSE`](#license)               | Specifies the legal license.                                   |
 | [`MESSAGE`](#message)               | Specify message history.                                       |
+| [`REQUIRES`](#requires)             | Specify the minimum version of Ollama required by the model.   |
 
 ## Examples
 
@@ -248,6 +249,16 @@ MESSAGE user Is Ontario in Canada?
 MESSAGE assistant yes
 ```
 
+### REQUIRES
+
+The `REQUIRES` instruction allows you to specify the minimum version of Ollama required by the model.
+
+```
+REQUIRES <version>
+```
+
+The version should be a valid Ollama version (e.g. 0.14.0).
+
 ## Notes
 
 - the **`Modelfile` is not case sensitive**. In the examples, uppercase instructions are used to make it easier to distinguish it from arguments.

docs/troubleshooting.mdx

@@ -87,7 +87,7 @@ When Ollama starts up, it takes inventory of the GPUs present in the system to d
 
 ### Linux NVIDIA Troubleshooting
 
-If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker.md](./docker.md)
+If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker](./docker)
 
 Sometimes the Ollama can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
 

go.mod

@@ -15,8 +15,8 @@ require (
 	github.com/spf13/cobra v1.7.0
 	github.com/stretchr/testify v1.9.0
 	github.com/x448/float16 v0.8.4
-	golang.org/x/sync v0.12.0
-	golang.org/x/sys v0.36.0
+	golang.org/x/sync v0.17.0
+	golang.org/x/sys v0.37.0
 )
 
 require (
@@ -29,7 +29,8 @@ require (
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
 	github.com/tkrajina/typescriptify-golang-structs v0.2.0
 	golang.org/x/image v0.22.0
-	golang.org/x/tools v0.30.0
+	golang.org/x/mod v0.30.0
+	golang.org/x/tools v0.38.0
 	gonum.org/v1/gonum v0.15.0
 )
 
@@ -76,11 +77,11 @@ require (
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.12 // indirect
 	golang.org/x/arch v0.8.0 // indirect
-	golang.org/x/crypto v0.36.0
+	golang.org/x/crypto v0.43.0
 	golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa // indirect
-	golang.org/x/net v0.38.0 // indirect
-	golang.org/x/term v0.30.0
-	golang.org/x/text v0.23.0
+	golang.org/x/net v0.46.0 // indirect
+	golang.org/x/term v0.36.0
+	golang.org/x/text v0.30.0
 	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -224,8 +224,8 @@ golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACk
 golang.org/x/crypto v0.0.0-20190510104115-cbcb75029529/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
-golang.org/x/crypto v0.36.0 h1:AnAEvhDddvBdpY+uR+MyHmuZzzNqXSe/GvuDeob5L34=
-golang.org/x/crypto v0.36.0/go.mod h1:Y4J0ReaxCR1IMaabaSMugxJES1EpwhBHhv2bDHklZvc=
+golang.org/x/crypto v0.43.0 h1:dduJYIi3A3KOfdGOHX8AVZ/jGiyPa3IbBozJ5kNuE04=
+golang.org/x/crypto v0.43.0/go.mod h1:BFbav4mRNlXJL4wNeejLpWxB7wMbc79PdRGhWKncxR0=
 golang.org/x/exp v0.0.0-20180321215751-8460e604b9de/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20180807140117-3d87b88a115f/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190121172915-509febef88a4/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
@@ -255,6 +255,8 @@ golang.org/x/mod v0.1.1-0.20191105210325-c90efee705ee/go.mod h1:QqPTAvyqsEbceGzB
 golang.org/x/mod v0.2.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.3.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.4.2/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
+golang.org/x/mod v0.30.0 h1:fDEXFVZ/fmCKProc/yAXXUijritrDzahmwwefnjoPFk=
+golang.org/x/mod v0.30.0/go.mod h1:lAsf5O2EvJeSFMiBxXDki7sCgAxEUcZHXoXMKT4GJKc=
 golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20180826012351-8a410e7b638d/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20190108225652-1e06a53dbb7e/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
@@ -267,8 +269,8 @@ golang.org/x/net v0.0.0-20200822124328-c89045814202/go.mod h1:/O7V0waA8r7cgGh81R
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210405180319-a5a99cb37ef4/go.mod h1:p54w0d4576C0XHj96bSt6lcn1PtDYWL6XObtHCRCNQM=
 golang.org/x/net v0.0.0-20210614182718-04defd469f4e/go.mod h1:9nx3DQGgdP8bBQD5qxJ1jj9UTztislL4KSBs9R2vV5Y=
-golang.org/x/net v0.38.0 h1:vRMAPTMaeGqVhG5QyLJHqNDwecKTomGeqbnfZyKlBI8=
-golang.org/x/net v0.38.0/go.mod h1:ivrbrMbzFq5J41QOQh0siUuly180yBYtLp+CKbEaFx8=
+golang.org/x/net v0.46.0 h1:giFlY12I07fugqwPuWJi68oOnpfqFnJIJzaIIm2JVV4=
+golang.org/x/net v0.46.0/go.mod h1:Q9BGdFy1y4nkUwiLvT5qtyhAnEHgnQ/zd8PfU6nc210=
 golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U=
 golang.org/x/oauth2 v0.0.0-20200107190931-bf48bf16ab8d/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20180314180146-1d60e4601c6f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -278,8 +280,8 @@ golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20210220032951-036812b2e83c/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sync v0.12.0 h1:MHc5BpPuC30uJk597Ri8TV3CNZcTLu6B6z4lJy+g6Jw=
-golang.org/x/sync v0.12.0/go.mod h1:1dzgHSNfp02xaA81J2MS99Qcpr2w7fw1gpm99rleRqA=
+golang.org/x/sync v0.17.0 h1:l60nONMj9l5drqw6jlhIELNv9I0A4OFgRsG9k2oT9Ug=
+golang.org/x/sync v0.17.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
 golang.org/x/sys v0.0.0-20180830151530-49385e6e1522/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190312061237-fead79001313/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
@@ -295,17 +297,17 @@ golang.org/x/sys v0.0.0-20210510120138-977fb7262007/go.mod h1:oPkhp1MJrh7nUepCBc
 golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.36.0 h1:KVRy2GtZBrk1cBYA7MKu5bEZFxQk4NIDV6RLVcC8o0k=
-golang.org/x/sys v0.36.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
+golang.org/x/sys v0.37.0 h1:fdNQudmxPjkdUTPnLn5mdQv7Zwvbvpaxqs831goi9kQ=
+golang.org/x/sys v0.37.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.30.0 h1:PQ39fJZ+mfadBm0y5WlL4vlM7Sx1Hgf13sMIY2+QS9Y=
-golang.org/x/term v0.30.0/go.mod h1:NYYFdzHoI5wRh/h5tDMdMqCqPJZEuNqVR5xJLd/n67g=
+golang.org/x/term v0.36.0 h1:zMPR+aF8gfksFprF/Nc/rd1wRS1EI6nDBGyWAvDzx2Q=
+golang.org/x/term v0.36.0/go.mod h1:Qu394IJq6V6dCBRgwqshf3mPF85AqzYEzofzRdZkWss=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.5/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
-golang.org/x/text v0.23.0 h1:D71I7dUrlY+VX0gQShAThNGHFxZ13dGLBHQLVl1mJlY=
-golang.org/x/text v0.23.0/go.mod h1:/BLNzu4aZCJ1+kcD0DNRotWKage4q2rGVAg4o22unh4=
+golang.org/x/text v0.30.0 h1:yznKA/E9zq54KzlzBEAWn1NXSQ8DIp/NYMy88xJjl4k=
+golang.org/x/text v0.30.0/go.mod h1:yDdHFIX9t+tORqspjENWgzaCVXgk0yYnYuSZ8UzzBVM=
 golang.org/x/tools v0.0.0-20180525024113-a5b4c53f6e8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
@@ -319,8 +321,8 @@ golang.org/x/tools v0.0.0-20200130002326-2f3ba24bd6e7/go.mod h1:TB2adYChydJhpapK
 golang.org/x/tools v0.0.0-20200619180055-7c47624df98f/go.mod h1:EkVYQZoAsY45+roYkvgYkIh4xh/qjgUK9TdY2XT94GE=
 golang.org/x/tools v0.0.0-20210106214847-113979e3529a/go.mod h1:emZCQorbCU4vsT4fOWvOPXz4eW1wZW4PmDk9uLelYpA=
 golang.org/x/tools v0.1.4/go.mod h1:o0xws9oXOQQZyjljx8fwUC0k7L1pTE6eaCbjGeHmOkk=
-golang.org/x/tools v0.30.0 h1:BgcpHewrV5AUp2G9MebG4XPFI1E2W41zU1SaqVA9vJY=
-golang.org/x/tools v0.30.0/go.mod h1:c347cR/OJfw5TI+GfX7RUPNMdDRRbjvYTS0jPyvsVtY=
+golang.org/x/tools v0.38.0 h1:Hx2Xv8hISq8Lm16jvBZ2VQf+RLmbd7wVUsALibYI/IQ=
+golang.org/x/tools v0.38.0/go.mod h1:yEsQ/d/YK8cjh0L6rZlY8tgtlKiBNTL14pGDJPJpYQs=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=

llama/patches/0024-GPU-discovery-enhancements.patch

@@ -20,10 +20,10 @@ fix vulkan PCI ID and ID handling
  ggml/src/ggml-cuda/vendors/hip.h     |   3 +
  ggml/src/ggml-impl.h                 |   8 +
  ggml/src/ggml-metal/ggml-metal.cpp   |   2 +
- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 169 ++++++++-
- ggml/src/mem_hip.cpp                 | 529 +++++++++++++++++++++++++++
- ggml/src/mem_nvml.cpp                | 209 +++++++++++
- 9 files changed, 976 insertions(+), 17 deletions(-)
+ ggml/src/ggml-vulkan/ggml-vulkan.cpp | 169 +++++++-
+ ggml/src/mem_hip.cpp                 | 558 +++++++++++++++++++++++++++
+ ggml/src/mem_nvml.cpp                | 209 ++++++++++
+ 9 files changed, 1005 insertions(+), 17 deletions(-)
  create mode 100644 ggml/src/mem_hip.cpp
  create mode 100644 ggml/src/mem_nvml.cpp
 
@@ -58,7 +58,7 @@ index d55aed348..99ae293cc 100644
  
  set_target_properties(ggml-base PROPERTIES
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 6852d2e20..48cdb1dcf 100644
+index 6852d2e20..334a30135 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -267,6 +267,16 @@ static ggml_cuda_device_info ggml_cuda_init() {
@@ -109,7 +109,7 @@ index 6852d2e20..48cdb1dcf 100644
 +
 +#if defined(GGML_USE_HIP)
 +    if (ggml_hip_mgmt_init() == 0) {
-+        int status = ggml_hip_get_device_memory(ctx->pci_bus_id.c_str(), free, total);
++        int status = ggml_hip_get_device_memory(ctx->pci_bus_id.c_str(), free, total, ctx->integrated != 0);
 +        if (status == 0) {
 +            GGML_LOG_DEBUG("%s device %s utilizing AMD specific memory reporting free: %zu total: %zu\n", __func__, ctx->pci_bus_id.c_str(), *free, *total);
 +            ggml_hip_mgmt_release();
@@ -204,7 +204,7 @@ index 4e162258d..d89e35a8e 100644
  #define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
  #define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
 diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
-index fe57d4c58..1c07e767a 100644
+index fe57d4c58..dba8f4695 100644
 --- a/ggml/src/ggml-impl.h
 +++ b/ggml/src/ggml-impl.h
 @@ -677,6 +677,14 @@ static inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph * cgraph,
@@ -216,7 +216,7 @@ index fe57d4c58..1c07e767a 100644
 +GGML_API int ggml_nvml_get_device_memory(const char *uuid, size_t *free, size_t *total);
 +GGML_API void ggml_nvml_release();
 +GGML_API int ggml_hip_mgmt_init();
-+GGML_API int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total);
++GGML_API int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total, bool is_integrated_gpu);
 +GGML_API void ggml_hip_mgmt_release();
 +
  #ifdef __cplusplus
@@ -243,7 +243,7 @@ index ba95b4acc..f6f8f7a10 100644
          /* .async                 = */ true,
          /* .host_buffer           = */ false,
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index 5349bce24..d43d46d1d 100644
+index 5349bce24..0103fd03a 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 @@ -236,6 +236,7 @@ class vk_memory_logger;
@@ -334,7 +334,7 @@ index 5349bce24..d43d46d1d 100644
 +        switch (props2.properties.vendorID) {
 +        case VK_VENDOR_ID_AMD:
 +            if (ggml_hip_mgmt_init() == 0) {
-+                int status = ggml_hip_get_device_memory(ctx->pci_id != "" ? ctx->pci_id.c_str() : ctx->uuid.c_str(), free, total);
++                int status = ggml_hip_get_device_memory(ctx->pci_id != "" ? ctx->pci_id.c_str() : ctx->uuid.c_str(), free, total, ctx->is_integrated_gpu);
 +                if (status == 0) {
 +                    GGML_LOG_DEBUG("%s device %s utilizing AMD specific memory reporting free: %zu total: %zu\n", __func__, ctx->pci_id != "" ? ctx->pci_id.c_str() : ctx->uuid.c_str(), *free, *total);
 +                    ggml_hip_mgmt_release();
@@ -505,10 +505,10 @@ index 5349bce24..d43d46d1d 100644
          }
 diff --git a/ggml/src/mem_hip.cpp b/ggml/src/mem_hip.cpp
 new file mode 100644
-index 000000000..c1949b899
+index 000000000..23c765806
 --- /dev/null
 +++ b/ggml/src/mem_hip.cpp
-@@ -0,0 +1,529 @@
+@@ -0,0 +1,558 @@
 +#include "ggml.h"
 +#include "ggml-impl.h"
 +
@@ -842,7 +842,7 @@ index 000000000..c1949b899
 +    if (gpus != NULL) gpus->pVtbl->Release(gpus); \
 +    if (gpu != NULL) gpu->pVtbl->Release(gpu)
 +
-+int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total) {
++int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total, bool is_integrated_gpu) {
 +    std::lock_guard<std::mutex> lock(ggml_adlx_lock);
 +    if (adlx.handle == NULL) {
 +        GGML_LOG_INFO("%s ADLX was not initialized\n", __func__);
@@ -966,13 +966,16 @@ index 000000000..c1949b899
 +    return 0;
 +}
 +void ggml_hip_mgmt_release() {}
-+int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total) {
++int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total, bool is_integrated_gpu) {
 +    GGML_LOG_INFO("%s searching for device %s\n", __func__, id);
 +    const std::string drmDeviceGlob = "/sys/class/drm/card*/device/uevent";
 +    const std::string drmTotalMemoryFile = "mem_info_vram_total";
 +    const std::string drmUsedMemoryFile = "mem_info_vram_used";
++    const std::string drmGTTTotalMemoryFile = "mem_info_gtt_total";
++    const std::string drmGTTUsedMemoryFile = "mem_info_gtt_used";
 +    const std::string drmUeventPCISlotLabel = "PCI_SLOT_NAME=";
 +
++
 +    glob_t glob_result;
 +    glob(drmDeviceGlob.c_str(), GLOB_NOSORT, NULL, &glob_result);
 +
@@ -1006,7 +1009,6 @@ index 000000000..c1949b899
 +
 +                    uint64_t memory;
 +                    totalFileStream >> memory;
-+                    *total = memory;
 +
 +                    std::string usedFile = dir + "/" + drmUsedMemoryFile;
 +                    std::ifstream usedFileStream(usedFile.c_str());
@@ -1019,6 +1021,33 @@ index 000000000..c1949b899
 +
 +                    uint64_t memoryUsed;
 +                    usedFileStream >> memoryUsed;
++
++                    if (is_integrated_gpu) {
++                        std::string totalFile = dir + "/" + drmGTTTotalMemoryFile;
++                        std::ifstream totalFileStream(totalFile.c_str());
++                        if (!totalFileStream.is_open()) {
++                            GGML_LOG_DEBUG("%s Failed to read sysfs node %s\n", __func__, totalFile.c_str());
++                            file.close();
++                            globfree(&glob_result);
++                            return 1;
++                        }
++                        uint64_t gtt;
++                        totalFileStream >> gtt;
++                        std::string usedFile = dir + "/" + drmGTTUsedMemoryFile;
++                        std::ifstream usedFileStream(usedFile.c_str());
++                        if (!usedFileStream.is_open()) {
++                            GGML_LOG_DEBUG("%s Failed to read sysfs node %s\n", __func__, usedFile.c_str());
++                            file.close();
++                            globfree(&glob_result);
++                            return 1;
++                        }
++                        uint64_t gttUsed;
++                        usedFileStream >> gttUsed;
++                        memory += gtt;
++                        memoryUsed += gttUsed;
++                    }
++
++                    *total = memory;
 +                    *free = memory - memoryUsed;
 +
 +                    file.close();

llama/patches/0028-Add-memory-detection-using-DXGI-PDH.patch

@@ -24,12 +24,12 @@ index 99ae293cc..9a134b7af 100644
  
  set_target_properties(ggml-base PROPERTIES
 diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
-index 1c07e767a..0da3e065b 100644
+index dba8f4695..7e17032c7 100644
 --- a/ggml/src/ggml-impl.h
 +++ b/ggml/src/ggml-impl.h
 @@ -684,6 +684,9 @@ GGML_API void ggml_nvml_release();
  GGML_API int ggml_hip_mgmt_init();
- GGML_API int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total);
+ GGML_API int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total, bool is_integrated_gpu);
  GGML_API void ggml_hip_mgmt_release();
 +GGML_API int ggml_dxgi_pdh_init();
 +GGML_API int ggml_dxgi_pdh_get_device_memory(const char* luid, size_t *free, size_t *total, bool is_integrated_gpu);
@@ -38,7 +38,7 @@ index 1c07e767a..0da3e065b 100644
  #ifdef __cplusplus
  }
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index d43d46d1d..df79f9f79 100644
+index 0103fd03a..9cc4ebdef 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 @@ -74,6 +74,7 @@ DispatchLoaderDynamic & ggml_vk_default_dispatcher();

llama/patches/0029-ggml-cuda-skip-large-batches.patch

@@ -10,7 +10,7 @@ fallback to cpu
  1 file changed, 3 insertions(+)
 
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 48cdb1dcf..3102d7ea7 100644
+index 334a30135..5c9dfd032 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -4633,6 +4633,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g

llm/server.go

@@ -524,8 +524,13 @@ func (s *llamaServer) Load(ctx context.Context, systemInfo ml.SystemInfo, system
 	// Use the size of one layer as a buffer
 	layers := s.ggml.Tensors().GroupLayers()
 	if blk0, ok := layers["blk.0"]; ok {
+		buffer := blk0.Size() + kv[0]
 		for i := range gpus {
-			gpus[i].FreeMemory -= blk0.Size() + kv[0]
+			if gpus[i].FreeMemory > buffer {
+				gpus[i].FreeMemory -= buffer
+			} else {
+				gpus[i].FreeMemory = 0
+			}
 		}
 	} else {
 		slog.Warn("model missing blk.0 layer size")
@@ -575,7 +580,11 @@ func (s *llamaServer) Load(ctx context.Context, systemInfo ml.SystemInfo, system
 			projectorGPU = firstIntegrated
 		}
 
-		gpus[projectorGPU].FreeMemory -= projectorWeights
+		if gpus[projectorGPU].FreeMemory > projectorWeights {
+			gpus[projectorGPU].FreeMemory -= projectorWeights
+		} else {
+			gpus[projectorGPU].FreeMemory = 0
+		}
 	}
 
 	var kvTotal uint64

middleware/anthropic.go

@@ -0,0 +1,149 @@
+package middleware
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+
+	"github.com/gin-gonic/gin"
+
+	"github.com/ollama/ollama/anthropic"
+	"github.com/ollama/ollama/api"
+)
+
+// AnthropicWriter wraps the response writer to transform Ollama responses to Anthropic format
+type AnthropicWriter struct {
+	BaseWriter
+	stream    bool
+	id        string
+	model     string
+	converter *anthropic.StreamConverter
+}
+
+func (w *AnthropicWriter) writeError(data []byte) (int, error) {
+	var errData struct {
+		Error string `json:"error"`
+	}
+	if err := json.Unmarshal(data, &errData); err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err := json.NewEncoder(w.ResponseWriter).Encode(anthropic.NewError(w.ResponseWriter.Status(), errData.Error))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *AnthropicWriter) writeEvent(eventType string, data any) error {
+	d, err := json.Marshal(data)
+	if err != nil {
+		return err
+	}
+	_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("event: %s\ndata: %s\n\n", eventType, d)))
+	if err != nil {
+		return err
+	}
+	if f, ok := w.ResponseWriter.(http.Flusher); ok {
+		f.Flush()
+	}
+	return nil
+}
+
+func (w *AnthropicWriter) writeResponse(data []byte) (int, error) {
+	var chatResponse api.ChatResponse
+	err := json.Unmarshal(data, &chatResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	if w.stream {
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+
+		events := w.converter.Process(chatResponse)
+		for _, event := range events {
+			if err := w.writeEvent(event.Event, event.Data); err != nil {
+				return 0, err
+			}
+		}
+		return len(data), nil
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	response := anthropic.ToMessagesResponse(w.id, chatResponse)
+	return len(data), json.NewEncoder(w.ResponseWriter).Encode(response)
+}
+
+func (w *AnthropicWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+
+	return w.writeResponse(data)
+}
+
+// AnthropicMessagesMiddleware handles Anthropic Messages API requests
+func AnthropicMessagesMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req anthropic.MessagesRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Model == "" {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "model is required"))
+			return
+		}
+
+		if req.MaxTokens <= 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "max_tokens is required and must be positive"))
+			return
+		}
+
+		if len(req.Messages) == 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, "messages is required"))
+			return
+		}
+
+		chatReq, err := anthropic.FromMessagesRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, anthropic.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, anthropic.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		messageID := anthropic.GenerateMessageID()
+
+		w := &AnthropicWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			stream:     req.Stream,
+			id:         messageID,
+			model:      req.Model,
+			converter:  anthropic.NewStreamConverter(messageID, req.Model),
+		}
+
+		if req.Stream {
+			c.Writer.Header().Set("Content-Type", "text/event-stream")
+			c.Writer.Header().Set("Cache-Control", "no-cache")
+			c.Writer.Header().Set("Connection", "keep-alive")
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}

middleware/anthropic_test.go

@@ -0,0 +1,573 @@
+package middleware
+
+import (
+	"bytes"
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/anthropic"
+	"github.com/ollama/ollama/api"
+)
+
+func captureAnthropicRequest(capturedRequest any) gin.HandlerFunc {
+	return func(c *gin.Context) {
+		bodyBytes, _ := io.ReadAll(c.Request.Body)
+		c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+		_ = json.Unmarshal(bodyBytes, capturedRequest)
+		c.Next()
+	}
+}
+
+func TestAnthropicMessagesMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.ChatRequest
+		err  anthropic.ErrorResponse
+	}
+
+	var capturedRequest *api.ChatRequest
+	stream := true
+
+	testCases := []testCase{
+		{
+			name: "basic message",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with system prompt",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"system": "You are helpful.",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "system", Content: "You are helpful."},
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with options",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 2048,
+				"temperature": 0.7,
+				"top_p": 0.9,
+				"top_k": 40,
+				"stop_sequences": ["\n", "END"],
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{
+					"num_predict": 2048,
+					"temperature": 0.7,
+					"top_p":       0.9,
+					"top_k":       40,
+					"stop":        []string{"\n", "END"},
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "streaming",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"stream": true,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &stream,
+			},
+		},
+		{
+			name: "with tools",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "What's the weather?"}
+				],
+				"tools": [{
+					"name": "get_weather",
+					"description": "Get current weather",
+					"input_schema": {
+						"type": "object",
+						"properties": {
+							"location": {"type": "string"}
+						},
+						"required": ["location"]
+					}
+				}]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "What's the weather?"},
+				},
+				Tools: []api.Tool{
+					{
+						Type: "function",
+						Function: api.ToolFunction{
+							Name:        "get_weather",
+							Description: "Get current weather",
+							Parameters: api.ToolFunctionParameters{
+								Type:     "object",
+								Required: []string{"location"},
+								Properties: map[string]api.ToolProperty{
+									"location": {Type: api.PropertyType{"string"}},
+								},
+							},
+						},
+					},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with tool result",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "What's the weather?"},
+					{"role": "assistant", "content": [
+						{"type": "tool_use", "id": "call_123", "name": "get_weather", "input": {"location": "Paris"}}
+					]},
+					{"role": "user", "content": [
+						{"type": "tool_result", "tool_use_id": "call_123", "content": "Sunny, 22°C"}
+					]}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "What's the weather?"},
+					{
+						Role: "assistant",
+						ToolCalls: []api.ToolCall{
+							{
+								ID: "call_123",
+								Function: api.ToolCallFunction{
+									Name:      "get_weather",
+									Arguments: api.ToolCallFunctionArguments{"location": "Paris"},
+								},
+							},
+						},
+					},
+					{Role: "tool", Content: "Sunny, 22°C", ToolCallID: "call_123"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+			},
+		},
+		{
+			name: "with thinking enabled",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"thinking": {"type": "enabled", "budget_tokens": 1000},
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello"},
+				},
+				Options: map[string]any{"num_predict": 1024},
+				Stream:  &False,
+				Think:   &api.ThinkValue{Value: true},
+			},
+		},
+		{
+			name: "missing model error",
+			body: `{
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "model is required",
+				},
+			},
+		},
+		{
+			name: "missing max_tokens error",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "max_tokens is required and must be positive",
+				},
+			},
+		},
+		{
+			name: "missing messages error",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "messages is required",
+				},
+			},
+		},
+		{
+			name: "tool_use missing id error",
+			body: `{
+				"model": "test-model",
+				"max_tokens": 1024,
+				"messages": [
+					{"role": "assistant", "content": [
+						{"type": "tool_use", "name": "test"}
+					]}
+				]
+			}`,
+			err: anthropic.ErrorResponse{
+				Type: "error",
+				Error: anthropic.Error{
+					Type:    "invalid_request_error",
+					Message: "tool_use block missing required 'id' field",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware(), captureAnthropicRequest(&capturedRequest))
+	router.Handle(http.MethodPost, "/v1/messages", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			defer func() { capturedRequest = nil }()
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if tc.err.Type != "" {
+				// Expect error
+				if resp.Code == http.StatusOK {
+					t.Fatalf("expected error response, got 200 OK")
+				}
+				var errResp anthropic.ErrorResponse
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatalf("failed to unmarshal error: %v", err)
+				}
+				if errResp.Type != tc.err.Type {
+					t.Errorf("expected error type %q, got %q", tc.err.Type, errResp.Type)
+				}
+				if errResp.Error.Type != tc.err.Error.Type {
+					t.Errorf("expected error.type %q, got %q", tc.err.Error.Type, errResp.Error.Type)
+				}
+				if errResp.Error.Message != tc.err.Error.Message {
+					t.Errorf("expected error.message %q, got %q", tc.err.Error.Message, errResp.Error.Message)
+				}
+				return
+			}
+
+			if resp.Code != http.StatusOK {
+				t.Fatalf("unexpected status code: %d, body: %s", resp.Code, resp.Body.String())
+			}
+
+			if capturedRequest == nil {
+				t.Fatal("request was not captured")
+			}
+
+			// Compare relevant fields
+			if capturedRequest.Model != tc.req.Model {
+				t.Errorf("model mismatch: got %q, want %q", capturedRequest.Model, tc.req.Model)
+			}
+
+			if diff := cmp.Diff(tc.req.Messages, capturedRequest.Messages); diff != "" {
+				t.Errorf("messages mismatch (-want +got):\n%s", diff)
+			}
+
+			if tc.req.Stream != nil && capturedRequest.Stream != nil {
+				if *tc.req.Stream != *capturedRequest.Stream {
+					t.Errorf("stream mismatch: got %v, want %v", *capturedRequest.Stream, *tc.req.Stream)
+				}
+			}
+
+			if tc.req.Think != nil {
+				if capturedRequest.Think == nil {
+					t.Error("expected Think to be set")
+				} else if capturedRequest.Think.Value != tc.req.Think.Value {
+					t.Errorf("Think mismatch: got %v, want %v", capturedRequest.Think.Value, tc.req.Think.Value)
+				}
+			}
+		})
+	}
+}
+
+func TestAnthropicMessagesMiddleware_Headers(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	t.Run("streaming sets correct headers", func(t *testing.T) {
+		router := gin.New()
+		router.Use(AnthropicMessagesMiddleware())
+		router.POST("/v1/messages", func(c *gin.Context) {
+			// Check headers were set
+			if c.Writer.Header().Get("Content-Type") != "text/event-stream" {
+				t.Errorf("expected Content-Type text/event-stream, got %q", c.Writer.Header().Get("Content-Type"))
+			}
+			if c.Writer.Header().Get("Cache-Control") != "no-cache" {
+				t.Errorf("expected Cache-Control no-cache, got %q", c.Writer.Header().Get("Cache-Control"))
+			}
+			c.Status(http.StatusOK)
+		})
+
+		body := `{"model": "test", "max_tokens": 100, "stream": true, "messages": [{"role": "user", "content": "Hi"}]}`
+		req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+		req.Header.Set("Content-Type", "application/json")
+
+		resp := httptest.NewRecorder()
+		router.ServeHTTP(resp, req)
+	})
+}
+
+func TestAnthropicMessagesMiddleware_InvalidJSON(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware())
+	router.POST("/v1/messages", func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	})
+
+	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(`{invalid json`))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusBadRequest {
+		t.Errorf("expected status 400, got %d", resp.Code)
+	}
+
+	var errResp anthropic.ErrorResponse
+	if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+		t.Fatalf("failed to unmarshal error: %v", err)
+	}
+
+	if errResp.Type != "error" {
+		t.Errorf("expected type 'error', got %q", errResp.Type)
+	}
+	if errResp.Error.Type != "invalid_request_error" {
+		t.Errorf("expected error type 'invalid_request_error', got %q", errResp.Error.Type)
+	}
+}
+
+func TestAnthropicWriter_NonStreaming(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	router := gin.New()
+	router.Use(AnthropicMessagesMiddleware())
+	router.POST("/v1/messages", func(c *gin.Context) {
+		// Simulate Ollama response
+		resp := api.ChatResponse{
+			Model: "test-model",
+			Message: api.Message{
+				Role:    "assistant",
+				Content: "Hello there!",
+			},
+			Done:       true,
+			DoneReason: "stop",
+			Metrics: api.Metrics{
+				PromptEvalCount: 10,
+				EvalCount:       5,
+			},
+		}
+		data, _ := json.Marshal(resp)
+		c.Writer.WriteHeader(http.StatusOK)
+		_, _ = c.Writer.Write(data)
+	})
+
+	body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
+	req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+
+	resp := httptest.NewRecorder()
+	router.ServeHTTP(resp, req)
+
+	if resp.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", resp.Code)
+	}
+
+	var result anthropic.MessagesResponse
+	if err := json.Unmarshal(resp.Body.Bytes(), &result); err != nil {
+		t.Fatalf("failed to unmarshal response: %v", err)
+	}
+
+	if result.Type != "message" {
+		t.Errorf("expected type 'message', got %q", result.Type)
+	}
+	if result.Role != "assistant" {
+		t.Errorf("expected role 'assistant', got %q", result.Role)
+	}
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Text == nil || *result.Content[0].Text != "Hello there!" {
+		t.Errorf("expected text 'Hello there!', got %v", result.Content[0].Text)
+	}
+	if result.StopReason != "end_turn" {
+		t.Errorf("expected stop_reason 'end_turn', got %q", result.StopReason)
+	}
+	if result.Usage.InputTokens != 10 {
+		t.Errorf("expected input_tokens 10, got %d", result.Usage.InputTokens)
+	}
+	if result.Usage.OutputTokens != 5 {
+		t.Errorf("expected output_tokens 5, got %d", result.Usage.OutputTokens)
+	}
+}
+
+// TestAnthropicWriter_ErrorFromRoutes tests error handling when routes.go sends
+// gin.H{"error": "message"} without a StatusCode field (which is the common case)
+func TestAnthropicWriter_ErrorFromRoutes(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	tests := []struct {
+		name           string
+		statusCode     int
+		errorPayload   any
+		wantErrorType  string
+		wantMessage    string
+	}{
+		// routes.go sends errors without StatusCode in JSON, so we must use HTTP status
+		{
+			name:          "404 with gin.H error (model not found)",
+			statusCode:    http.StatusNotFound,
+			errorPayload:  gin.H{"error": "model 'nonexistent' not found"},
+			wantErrorType: "not_found_error",
+			wantMessage:   "model 'nonexistent' not found",
+		},
+		{
+			name:          "400 with gin.H error (bad request)",
+			statusCode:    http.StatusBadRequest,
+			errorPayload:  gin.H{"error": "model is required"},
+			wantErrorType: "invalid_request_error",
+			wantMessage:   "model is required",
+		},
+		{
+			name:          "500 with gin.H error (internal error)",
+			statusCode:    http.StatusInternalServerError,
+			errorPayload:  gin.H{"error": "something went wrong"},
+			wantErrorType: "api_error",
+			wantMessage:   "something went wrong",
+		},
+		{
+			name:       "404 with api.StatusError",
+			statusCode: http.StatusNotFound,
+			errorPayload: api.StatusError{
+				StatusCode:   http.StatusNotFound,
+				ErrorMessage: "model not found via StatusError",
+			},
+			wantErrorType: "not_found_error",
+			wantMessage:   "model not found via StatusError",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			router := gin.New()
+			router.Use(AnthropicMessagesMiddleware())
+			router.POST("/v1/messages", func(c *gin.Context) {
+				// Simulate what routes.go does - set status and write error JSON
+				data, _ := json.Marshal(tt.errorPayload)
+				c.Writer.WriteHeader(tt.statusCode)
+				_, _ = c.Writer.Write(data)
+			})
+
+			body := `{"model": "test-model", "max_tokens": 100, "messages": [{"role": "user", "content": "Hi"}]}`
+			req, _ := http.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(body))
+			req.Header.Set("Content-Type", "application/json")
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			if resp.Code != tt.statusCode {
+				t.Errorf("expected status %d, got %d", tt.statusCode, resp.Code)
+			}
+
+			var errResp anthropic.ErrorResponse
+			if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+				t.Fatalf("failed to unmarshal error response: %v\nbody: %s", err, resp.Body.String())
+			}
+
+			if errResp.Type != "error" {
+				t.Errorf("expected type 'error', got %q", errResp.Type)
+			}
+			if errResp.Error.Type != tt.wantErrorType {
+				t.Errorf("expected error type %q, got %q", tt.wantErrorType, errResp.Error.Type)
+			}
+			if errResp.Error.Message != tt.wantMessage {
+				t.Errorf("expected message %q, got %q", tt.wantMessage, errResp.Error.Message)
+			}
+		})
+	}
+}

ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu

@@ -4436,7 +4436,7 @@ static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t *
 
 #if defined(GGML_USE_HIP)
     if (ggml_hip_mgmt_init() == 0) {
-        int status = ggml_hip_get_device_memory(ctx->pci_bus_id.c_str(), free, total);
+        int status = ggml_hip_get_device_memory(ctx->pci_bus_id.c_str(), free, total, ctx->integrated != 0);
         if (status == 0) {
             GGML_LOG_DEBUG("%s device %s utilizing AMD specific memory reporting free: %zu total: %zu\n", __func__, ctx->pci_bus_id.c_str(), *free, *total);
             ggml_hip_mgmt_release();

ml/backend/ggml/ggml/src/ggml-impl.h

@@ -682,7 +682,7 @@ GGML_API int ggml_nvml_init();
 GGML_API int ggml_nvml_get_device_memory(const char *uuid, size_t *free, size_t *total);
 GGML_API void ggml_nvml_release();
 GGML_API int ggml_hip_mgmt_init();
-GGML_API int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total);
+GGML_API int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total, bool is_integrated_gpu);
 GGML_API void ggml_hip_mgmt_release();
 GGML_API int ggml_dxgi_pdh_init();
 GGML_API int ggml_dxgi_pdh_get_device_memory(const char* luid, size_t *free, size_t *total, bool is_integrated_gpu);

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

@@ -13710,7 +13710,7 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
         switch (props2.properties.vendorID) {
         case VK_VENDOR_ID_AMD:
             if (ggml_hip_mgmt_init() == 0) {
-                int status = ggml_hip_get_device_memory(ctx->pci_id != "" ? ctx->pci_id.c_str() : ctx->uuid.c_str(), free, total);
+                int status = ggml_hip_get_device_memory(ctx->pci_id != "" ? ctx->pci_id.c_str() : ctx->uuid.c_str(), free, total, ctx->is_integrated_gpu);
                 if (status == 0) {
                     GGML_LOG_DEBUG("%s device %s utilizing AMD specific memory reporting free: %zu total: %zu\n", __func__, ctx->pci_id != "" ? ctx->pci_id.c_str() : ctx->uuid.c_str(), *free, *total);
                     ggml_hip_mgmt_release();

ml/backend/ggml/ggml/src/mem_hip.cpp

@@ -331,7 +331,7 @@ void ggml_hip_mgmt_release() {
     if (gpus != NULL) gpus->pVtbl->Release(gpus); \
     if (gpu != NULL) gpu->pVtbl->Release(gpu)
 
-int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total) {
+int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total, bool is_integrated_gpu) {
     std::lock_guard<std::mutex> lock(ggml_adlx_lock);
     if (adlx.handle == NULL) {
         GGML_LOG_INFO("%s ADLX was not initialized\n", __func__);
@@ -455,13 +455,16 @@ int ggml_hip_mgmt_init() {
     return 0;
 }
 void ggml_hip_mgmt_release() {}
-int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total) {
+int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total, bool is_integrated_gpu) {
     GGML_LOG_INFO("%s searching for device %s\n", __func__, id);
     const std::string drmDeviceGlob = "/sys/class/drm/card*/device/uevent";
     const std::string drmTotalMemoryFile = "mem_info_vram_total";
     const std::string drmUsedMemoryFile = "mem_info_vram_used";
+    const std::string drmGTTTotalMemoryFile = "mem_info_gtt_total";
+    const std::string drmGTTUsedMemoryFile = "mem_info_gtt_used";
     const std::string drmUeventPCISlotLabel = "PCI_SLOT_NAME=";
 
+
     glob_t glob_result;
     glob(drmDeviceGlob.c_str(), GLOB_NOSORT, NULL, &glob_result);
 
@@ -495,7 +498,6 @@ int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total) {
 
                     uint64_t memory;
                     totalFileStream >> memory;
-                    *total = memory;
 
                     std::string usedFile = dir + "/" + drmUsedMemoryFile;
                     std::ifstream usedFileStream(usedFile.c_str());
@@ -508,6 +510,33 @@ int ggml_hip_get_device_memory(const char *id, size_t *free, size_t *total) {
 
                     uint64_t memoryUsed;
                     usedFileStream >> memoryUsed;
+
+                    if (is_integrated_gpu) {
+                        std::string totalFile = dir + "/" + drmGTTTotalMemoryFile;
+                        std::ifstream totalFileStream(totalFile.c_str());
+                        if (!totalFileStream.is_open()) {
+                            GGML_LOG_DEBUG("%s Failed to read sysfs node %s\n", __func__, totalFile.c_str());
+                            file.close();
+                            globfree(&glob_result);
+                            return 1;
+                        }
+                        uint64_t gtt;
+                        totalFileStream >> gtt;
+                        std::string usedFile = dir + "/" + drmGTTUsedMemoryFile;
+                        std::ifstream usedFileStream(usedFile.c_str());
+                        if (!usedFileStream.is_open()) {
+                            GGML_LOG_DEBUG("%s Failed to read sysfs node %s\n", __func__, usedFile.c_str());
+                            file.close();
+                            globfree(&glob_result);
+                            return 1;
+                        }
+                        uint64_t gttUsed;
+                        usedFileStream >> gttUsed;
+                        memory += gtt;
+                        memoryUsed += gttUsed;
+                    }
+
+                    *total = memory;
                     *free = memory - memoryUsed;
 
                     file.close();

parser/parser.go

@@ -17,6 +17,7 @@ import (
 	"strings"
 	"sync"
 
+	"golang.org/x/mod/semver"
 	"golang.org/x/sync/errgroup"
 	"golang.org/x/text/encoding/unicode"
 	"golang.org/x/text/transform"
@@ -104,6 +105,16 @@ func (f Modelfile) CreateRequest(relativeDir string) (*api.CreateRequest, error)
 			req.Renderer = c.Args
 		case "parser":
 			req.Parser = c.Args
+		case "requires":
+			// golang.org/x/mod/semver requires "v" prefix
+			requires := c.Args
+			if !strings.HasPrefix(requires, "v") {
+				requires = "v" + requires
+			}
+			if !semver.IsValid(requires) {
+				return nil, fmt.Errorf("requires must be a valid semver (e.g. 0.14.0)")
+			}
+			req.Requires = strings.TrimPrefix(requires, "v")
 		case "message":
 			role, msg, _ := strings.Cut(c.Args, ": ")
 			messages = append(messages, api.Message{Role: role, Content: msg})
@@ -322,7 +333,7 @@ func (c Command) String() string {
 	switch c.Name {
 	case "model":
 		fmt.Fprintf(&sb, "FROM %s", c.Args)
-	case "license", "template", "system", "adapter", "renderer", "parser":
+	case "license", "template", "system", "adapter", "renderer", "parser", "requires":
 		fmt.Fprintf(&sb, "%s %s", strings.ToUpper(c.Name), quote(c.Args))
 	case "message":
 		role, message, _ := strings.Cut(c.Args, ": ")
@@ -348,7 +359,7 @@ const (
 var (
 	errMissingFrom        = errors.New("no FROM line")
 	errInvalidMessageRole = errors.New("message role must be one of \"system\", \"user\", or \"assistant\"")
-	errInvalidCommand     = errors.New("command must be one of \"from\", \"license\", \"template\", \"system\", \"adapter\", \"renderer\", \"parser\", \"parameter\", or \"message\"")
+	errInvalidCommand     = errors.New("command must be one of \"from\", \"license\", \"template\", \"system\", \"adapter\", \"renderer\", \"parser\", \"parameter\", \"message\", or \"requires\"")
 )
 
 type ParserError struct {
@@ -608,7 +619,7 @@ func isValidMessageRole(role string) bool {
 
 func isValidCommand(cmd string) bool {
 	switch strings.ToLower(cmd) {
-	case "from", "license", "template", "system", "adapter", "renderer", "parser", "parameter", "message":
+	case "from", "license", "template", "system", "adapter", "renderer", "parser", "parameter", "message", "requires":
 		return true
 	default:
 		return false

server/create.go

@@ -61,6 +61,7 @@ func (s *Server) CreateHandler(c *gin.Context) {
 
 	config.Renderer = r.Renderer
 	config.Parser = r.Parser
+	config.Requires = r.Requires
 
 	for v := range r.Files {
 		if !fs.ValidPath(v) {
@@ -120,7 +121,7 @@ func (s *Server) CreateHandler(c *gin.Context) {
 					ch <- gin.H{"error": err.Error()}
 				}
 
-				if err == nil && !remote && (config.Renderer == "" || config.Parser == "") {
+				if err == nil && !remote && (config.Renderer == "" || config.Parser == "" || config.Requires == "") {
 					manifest, mErr := ParseNamedManifest(fromName)
 					if mErr == nil && manifest.Config.Digest != "" {
 						configPath, pErr := GetBlobsPath(manifest.Config.Digest)
@@ -134,6 +135,9 @@ func (s *Server) CreateHandler(c *gin.Context) {
 									if config.Parser == "" {
 										config.Parser = baseConfig.Parser
 									}
+									if config.Requires == "" {
+										config.Requires = baseConfig.Requires
+									}
 								}
 								cfgFile.Close()
 							}

server/routes.go

@@ -1106,6 +1106,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		Messages:     msgs,
 		Capabilities: m.Capabilities(),
 		ModifiedAt:   manifest.fi.ModTime(),
+		Requires:     m.Config.Requires,
 	}
 
 	if m.Config.RemoteHost != "" {
@@ -1534,6 +1535,9 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.GET("/v1/models/:model", middleware.RetrieveMiddleware(), s.ShowHandler)
 	r.POST("/v1/responses", middleware.ResponsesMiddleware(), s.ChatHandler)
 
+	// Inference (Anthropic compatibility)
+	r.POST("/v1/messages", middleware.AnthropicMessagesMiddleware(), s.ChatHandler)
+
 	if rc != nil {
 		// wrap old with new
 		rs := &registry.Local{

types/model/config.go

@@ -9,6 +9,7 @@ type ConfigV2 struct {
 	FileType      string   `json:"file_type"`  // shown as Quantization Level
 	Renderer      string   `json:"renderer,omitempty"`
 	Parser        string   `json:"parser,omitempty"`
+	Requires      string   `json:"requires,omitempty"`
 
 	RemoteHost  string `json:"remote_host,omitempty"`
 	RemoteModel string `json:"remote_model,omitempty"`