model: fix issues with spm tokenizer for Gemma 3 (#10081)

The sliding window cache trims entries that are outside the window for
the latest token. This works when we are extending the cache, such as
when the conversation continues. However, if we have a partial overlap
in conversation (including the BOS tokens), then we resume from a past
point in the conversation and the needed tokens are no longer stored
in memory. This verifies that the new window overlaps with the old one
before reusing the cache.

Co-authored-by: Jesse Gross <jesse@ollama.com>

README.md

@@ -325,6 +325,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [RWKV-Runner](https://github.com/josStorer/RWKV-Runner) (RWKV offline LLM deployment tool, also usable as a client for ChatGPT and Ollama)
 - [Ollama Grid Search](https://github.com/dezoito/ollama-grid-search) (app to evaluate and compare models)
 - [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
+- [Casibase](https://casibase.org) (An open source AI knowledge base and dialogue system combining the latest RAG, SSO, ollama support and multiple large language models.)
 - [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
 - [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama)
 - [Shinkai Desktop](https://github.com/dcSpark/shinkai-apps) (Two click install Local AI using Ollama + Files + RAG)

api/types.go

@@ -12,6 +12,7 @@ import (
 	"time"
 
 	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/types/model"
 )
 
 // StatusError is an error with an HTTP status code and message.
@@ -81,7 +82,7 @@ type GenerateRequest struct {
 
 	// Options lists model-specific options. For example, temperature can be
 	// set through this field, if the model supports it.
-	Options map[string]interface{} `json:"options"`
+	Options map[string]any `json:"options"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].
@@ -106,7 +107,7 @@ type ChatRequest struct {
 	Tools `json:"tools,omitempty"`
 
 	// Options lists model-specific options.
-	Options map[string]interface{} `json:"options"`
+	Options map[string]any `json:"options"`
 }
 
 type Tools []Tool
@@ -260,7 +261,7 @@ type EmbedRequest struct {
 	Truncate *bool `json:"truncate,omitempty"`
 
 	// Options lists model-specific options.
-	Options map[string]interface{} `json:"options"`
+	Options map[string]any `json:"options"`
 }
 
 // EmbedResponse is the response from [Client.Embed].
@@ -286,7 +287,7 @@ type EmbeddingRequest struct {
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
 	// Options lists model-specific options.
-	Options map[string]interface{} `json:"options"`
+	Options map[string]any `json:"options"`
 }
 
 // EmbeddingResponse is the response from [Client.Embeddings].
@@ -332,7 +333,7 @@ type ShowRequest struct {
 	Template string `json:"template"`
 	Verbose  bool   `json:"verbose"`
 
-	Options map[string]interface{} `json:"options"`
+	Options map[string]any `json:"options"`
 
 	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
@@ -340,17 +341,18 @@ type ShowRequest struct {
 
 // ShowResponse is the response returned from [Client.Show].
 type ShowResponse struct {
-	License       string         `json:"license,omitempty"`
-	Modelfile     string         `json:"modelfile,omitempty"`
-	Parameters    string         `json:"parameters,omitempty"`
-	Template      string         `json:"template,omitempty"`
-	System        string         `json:"system,omitempty"`
-	Details       ModelDetails   `json:"details,omitempty"`
-	Messages      []Message      `json:"messages,omitempty"`
-	ModelInfo     map[string]any `json:"model_info,omitempty"`
-	ProjectorInfo map[string]any `json:"projector_info,omitempty"`
-	Tensors       []Tensor       `json:"tensors,omitempty"`
-	ModifiedAt    time.Time      `json:"modified_at,omitempty"`
+	License       string             `json:"license,omitempty"`
+	Modelfile     string             `json:"modelfile,omitempty"`
+	Parameters    string             `json:"parameters,omitempty"`
+	Template      string             `json:"template,omitempty"`
+	System        string             `json:"system,omitempty"`
+	Details       ModelDetails       `json:"details,omitempty"`
+	Messages      []Message          `json:"messages,omitempty"`
+	ModelInfo     map[string]any     `json:"model_info,omitempty"`
+	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
+	Tensors       []Tensor           `json:"tensors,omitempty"`
+	Capabilities  []model.Capability `json:"capabilities,omitempty"`
+	ModifiedAt    time.Time          `json:"modified_at,omitempty"`
 }
 
 // CopyRequest is the request passed to [Client.Copy].
@@ -503,7 +505,7 @@ func (m *Metrics) Summary() {
 	}
 }
 
-func (opts *Options) FromMap(m map[string]interface{}) error {
+func (opts *Options) FromMap(m map[string]any) error {
 	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
 	typeOpts := reflect.TypeOf(opts).Elem()   // types of the fields in the options struct
 
@@ -560,12 +562,12 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 				}
 				field.SetString(val)
 			case reflect.Slice:
-				// JSON unmarshals to []interface{}, not []string
-				val, ok := val.([]interface{})
+				// JSON unmarshals to []any, not []string
+				val, ok := val.([]any)
 				if !ok {
 					return fmt.Errorf("option %q must be of type array", key)
 				}
-				// convert []interface{} to []string
+				// convert []any to []string
 				slice := make([]string, len(val))
 				for i, item := range val {
 					str, ok := item.(string)
@@ -672,7 +674,7 @@ func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 }
 
 // FormatParams converts specified parameter options to their correct types
-func FormatParams(params map[string][]string) (map[string]interface{}, error) {
+func FormatParams(params map[string][]string) (map[string]any, error) {
 	opts := Options{}
 	valueOpts := reflect.ValueOf(&opts).Elem() // names of the fields in the options struct
 	typeOpts := reflect.TypeOf(opts)           // types of the fields in the options struct
@@ -686,7 +688,7 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
 		}
 	}
 
-	out := make(map[string]interface{})
+	out := make(map[string]any)
 	// iterate params and set values based on json struct tags
 	for key, vals := range params {
 		if opt, ok := jsonOpts[key]; !ok {

api/types_test.go

@@ -134,7 +134,7 @@ func TestUseMmapParsingFromJSON(t *testing.T) {
 
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {
-			var oMap map[string]interface{}
+			var oMap map[string]any
 			err := json.Unmarshal([]byte(test.req), &oMap)
 			require.NoError(t, err)
 			opts := DefaultOptions()

benchmark/server_benchmark_test.go

@@ -92,7 +92,7 @@ func BenchmarkColdStart(b *testing.B) {
 				req := &api.GenerateRequest{
 					Model:   m,
 					Prompt:  tt.prompt,
-					Options: map[string]interface{}{"num_predict": tt.maxTokens, "temperature": 0.1},
+					Options: map[string]any{"num_predict": tt.maxTokens, "temperature": 0.1},
 				}
 
 				runGenerateBenchmark(b, ctx, client, req)
@@ -155,7 +155,7 @@ func warmup(client *api.Client, model string, prompt string, b *testing.B) {
 			&api.GenerateRequest{
 				Model:   model,
 				Prompt:  prompt,
-				Options: map[string]interface{}{"num_predict": 50, "temperature": 0.1},
+				Options: map[string]any{"num_predict": 50, "temperature": 0.1},
 			},
 			func(api.GenerateResponse) error { return nil },
 		)

cmd/cmd.go

@@ -18,6 +18,7 @@ import (
 	"os/signal"
 	"path/filepath"
 	"runtime"
+	"slices"
 	"sort"
 	"strconv"
 	"strings"
@@ -267,7 +268,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	opts := runOptions{
 		Model:    args[0],
 		WordWrap: os.Getenv("TERM") == "xterm-256color",
-		Options:  map[string]interface{}{},
+		Options:  map[string]any{},
 	}
 
 	format, err := cmd.Flags().GetString("format")
@@ -339,6 +340,11 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
+	opts.MultiModal = slices.Contains(info.Capabilities, model.CapabilityVision)
+
+	// TODO: remove the projector info and vision info checks below,
+	// these are left in for backwards compatibility with older servers
+	// that don't have the capabilities field in the model info
 	if len(info.ProjectorInfo) != 0 {
 		opts.MultiModal = true
 	}
@@ -669,6 +675,15 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 		return
 	})
 
+	if len(resp.Capabilities) > 0 {
+		tableRender("Capabilities", func() (rows [][]string) {
+			for _, capability := range resp.Capabilities {
+				rows = append(rows, []string{"", capability.String()})
+			}
+			return
+		})
+	}
+
 	if resp.ProjectorInfo != nil {
 		tableRender("Projector", func() (rows [][]string) {
 			arch := resp.ProjectorInfo["general.architecture"].(string)
@@ -837,7 +852,7 @@ type runOptions struct {
 	Format      string
 	System      string
 	Images      []api.ImageData
-	Options     map[string]interface{}
+	Options     map[string]any
 	MultiModal  bool
 	KeepAlive   *api.Duration
 }

cmd/cmd_test.go

@@ -16,6 +16,7 @@ import (
 	"github.com/spf13/cobra"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/types/model"
 )
 
 func TestShowInfo(t *testing.T) {
@@ -260,6 +261,34 @@ Weigh anchor!
 			t.Errorf("unexpected output (-want +got):\n%s", diff)
 		}
 	})
+
+	t.Run("capabilities", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			Capabilities: []model.Capability{model.CapabilityVision, model.CapabilityTools},
+		}, false, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := "  Model\n" +
+			"    architecture    test    \n" +
+			"    parameters      7B      \n" +
+			"    quantization    FP16    \n" +
+			"\n" +
+			"  Capabilities\n" +
+			"    vision    \n" +
+			"    tools     \n" +
+			"\n"
+
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
 }
 
 func TestDeleteHandler(t *testing.T) {

convert/sentencepiece/sentencepiece_model.pb.go

@@ -1360,7 +1360,7 @@ func file_sentencepiece_model_proto_rawDescGZIP() []byte {
 
 var file_sentencepiece_model_proto_enumTypes = make([]protoimpl.EnumInfo, 2)
 var file_sentencepiece_model_proto_msgTypes = make([]protoimpl.MessageInfo, 6)
-var file_sentencepiece_model_proto_goTypes = []interface{}{
+var file_sentencepiece_model_proto_goTypes = []any{
 	(TrainerSpec_ModelType)(0),         // 0: sentencepiece.TrainerSpec.ModelType
 	(ModelProto_SentencePiece_Type)(0), // 1: sentencepiece.ModelProto.SentencePiece.Type
 	(*TrainerSpec)(nil),                // 2: sentencepiece.TrainerSpec
@@ -1392,7 +1392,7 @@ func file_sentencepiece_model_proto_init() {
 		return
 	}
 	if !protoimpl.UnsafeEnabled {
-		file_sentencepiece_model_proto_msgTypes[0].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[0].Exporter = func(v any, i int) any {
 			switch v := v.(*TrainerSpec); i {
 			case 0:
 				return &v.state
@@ -1406,7 +1406,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[1].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[1].Exporter = func(v any, i int) any {
 			switch v := v.(*NormalizerSpec); i {
 			case 0:
 				return &v.state
@@ -1420,7 +1420,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[2].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[2].Exporter = func(v any, i int) any {
 			switch v := v.(*SelfTestData); i {
 			case 0:
 				return &v.state
@@ -1434,7 +1434,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[3].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[3].Exporter = func(v any, i int) any {
 			switch v := v.(*ModelProto); i {
 			case 0:
 				return &v.state
@@ -1448,7 +1448,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[4].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[4].Exporter = func(v any, i int) any {
 			switch v := v.(*SelfTestData_Sample); i {
 			case 0:
 				return &v.state
@@ -1460,7 +1460,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[5].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[5].Exporter = func(v any, i int) any {
 			switch v := v.(*ModelProto_SentencePiece); i {
 			case 0:
 				return &v.state

discover/cpu_common.go

@@ -12,7 +12,7 @@ func IsNUMA() bool {
 		// numa support in llama.cpp is linux only
 		return false
 	}
-	ids := map[string]interface{}{}
+	ids := map[string]any{}
 	packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")
 	for _, packageId := range packageIds {
 		id, err := os.ReadFile(packageId)

docs/api.md

@@ -1217,7 +1217,7 @@ Show information about a model including details, modelfile, template, parameter
 
 ```shell
 curl http://localhost:11434/api/show -d '{
-  "model": "llama3.2"
+  "model": "llava"
 }'
 ```
 
@@ -1260,7 +1260,11 @@ curl http://localhost:11434/api/show -d '{
     "tokenizer.ggml.pre": "llama-bpe",
     "tokenizer.ggml.token_type": [],        // populates if `verbose=true`
     "tokenizer.ggml.tokens": []             // populates if `verbose=true`
-  }
+  },
+  "capabilities": [
+    "completion",
+    "vision"
+  ],
 }
 ```
 

format/time_test.go

@@ -5,7 +5,7 @@ import (
 	"time"
 )
 
-func assertEqual(t *testing.T, a interface{}, b interface{}) {
+func assertEqual(t *testing.T, a any, b any) {
 	if a != b {
 		t.Errorf("Assert failed, expected %v, got %v", b, a)
 	}

integration/basic_test.go

@@ -22,7 +22,7 @@ func TestOrcaMiniBlueSky(t *testing.T) {
 		Model:  "orca-mini",
 		Prompt: "why is the sky blue?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},
@@ -39,7 +39,7 @@ func TestUnicode(t *testing.T) {
 		Model:  "deepseek-coder-v2:16b-lite-instruct-q2_K",
 		Prompt: "天空为什么是蓝色的?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 			// Workaround deepseek context shifting bug
@@ -61,7 +61,7 @@ func TestExtendedUnicodeOutput(t *testing.T) {
 		Model:  "gemma2:2b",
 		Prompt: "Output some smily face emoji",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},
@@ -96,7 +96,7 @@ func TestUnicodeModelDir(t *testing.T) {
 		Model:  "orca-mini",
 		Prompt: "why is the sky blue?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},

integration/concurrency_test.go

@@ -25,7 +25,7 @@ func TestMultiModelConcurrency(t *testing.T) {
 				Prompt:    "why is the ocean blue?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
@@ -34,7 +34,7 @@ func TestMultiModelConcurrency(t *testing.T) {
 				Prompt:    "what is the origin of the us thanksgiving holiday?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},

integration/context_test.go

@@ -23,7 +23,7 @@ func TestLongInputContext(t *testing.T) {
 		Model:  "llama2",
 		Prompt: "Oh, don’t speak to me of Austria. Perhaps I don’t understand things, but Austria never has wished, and does not wish, for war. She is betraying us! Russia alone must save Europe. Our gracious sovereign recognizes his high vocation and will be true to it. That is the one thing I have faith in! Our good and wonderful sovereign has to perform the noblest role on earth, and he is so virtuous and noble that God will not forsake him. He will fulfill his vocation and crush the hydra of revolution, which has become more terrible than ever in the person of this murderer and villain! We alone must avenge the blood of the just one.... Whom, I ask you, can we rely on?... England with her commercial spirit will not and cannot understand the Emperor Alexander’s loftiness of soul. She has refused to evacuate Malta. She wanted to find, and still seeks, some secret motive in our actions. What answer did Novosíltsev get? None. The English have not understood and cannot understand the self-abnegation of our Emperor who wants nothing for himself, but only desires the good of mankind. And what have they promised? Nothing! And what little they have promised they will not perform! Prussia has always declared that Buonaparte is invincible, and that all Europe is powerless before him.... And I don’t believe a word that Hardenburg says, or Haugwitz either. This famous Prussian neutrality is just a trap. I have faith only in God and the lofty destiny of our adored monarch. He will save Europe! What country is this referring to?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 			"num_ctx":     128,
@@ -50,7 +50,7 @@ func TestContextExhaustion(t *testing.T) {
 		Model:  "llama2",
 		Prompt: "Write me a story with a ton of emojis?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 			"num_ctx":     128,

integration/llm_image_test.go

@@ -19,7 +19,7 @@ func TestIntegrationLlava(t *testing.T) {
 		Model:  "llava:7b",
 		Prompt: "what does the text in this image say?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"seed":        42,
 			"temperature": 0.0,
 		},
@@ -47,7 +47,7 @@ func TestIntegrationMllama(t *testing.T) {
 		Model:  "x/llama3.2-vision",
 		Prompt: "what does the text in this image say?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"seed":        42,
 			"temperature": 0.0,
 		},
@@ -75,7 +75,7 @@ func TestIntegrationSplitBatch(t *testing.T) {
 		System: "Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed aliquet, justo in malesuada lobortis, odio ligula volutpat quam, quis faucibus ipsum magna quis sapien. Aliquam in venenatis diam, eu viverra magna. Phasellus imperdiet hendrerit volutpat. Vivamus sem ex, facilisis placerat felis non, dictum elementum est. Phasellus aliquam imperdiet lacus, eget placerat ligula sodales vel. Pellentesque nec auctor mi. Curabitur arcu nisi, faucibus eget nunc id, viverra interdum mi. Curabitur ornare ipsum ex, ac euismod ex aliquam in. Vestibulum id magna at purus accumsan fermentum. Proin scelerisque posuere nunc quis interdum. Maecenas sed mollis nisl. Etiam vitae ipsum interdum, placerat est quis, tincidunt velit. Nullam tempor nibh non lorem volutpat efficitur. Cras laoreet diam imperdiet ipsum auctor bibendum. Suspendisse ultrices urna sed metus sagittis suscipit. Quisque ullamcorper aliquam nibh ut mollis. Aenean dapibus mauris pharetra, venenatis elit ac, hendrerit odio. Cras vestibulum erat tempor, lobortis justo eu, lobortis ipsum. Nam laoreet dapibus sem. Proin vel diam ultrices, elementum ante et, ornare lectus. Proin eu accumsan nisl. Praesent ac ex vitae ipsum vulputate tristique facilisis sit amet lacus. Nullam faucibus magna a pellentesque pretium. Nunc lacinia ullamcorper sollicitudin. Donec vitae accumsan turpis, sed porttitor est. Donec porttitor mi vitae augue faucibus, vel mollis diam tincidunt.",
 		Prompt: "what does the text in this image say?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"seed":        42,
 			"temperature": 0.0,
 		},

integration/llm_test.go

@@ -20,7 +20,7 @@ var (
 			Model:  "orca-mini",
 			Prompt: "why is the ocean blue?",
 			Stream: &stream,
-			Options: map[string]interface{}{
+			Options: map[string]any{
 				"seed":        42,
 				"temperature": 0.0,
 			},
@@ -28,7 +28,7 @@ var (
 			Model:  "orca-mini",
 			Prompt: "what is the origin of the us thanksgiving holiday?",
 			Stream: &stream,
-			Options: map[string]interface{}{
+			Options: map[string]any{
 				"seed":        42,
 				"temperature": 0.0,
 			},

integration/max_queue_test.go

@@ -32,7 +32,7 @@ func TestMaxQueue(t *testing.T) {
 	req := api.GenerateRequest{
 		Model:  "orca-mini",
 		Prompt: "write a long historical fiction story about christopher columbus.  use at least 10 facts from his actual journey",
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"seed":        42,
 			"temperature": 0.0,
 		},

integration/utils_test.go

@@ -291,7 +291,7 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 				Prompt:    "why is the ocean blue?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
@@ -300,7 +300,7 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 				Prompt:    "why is the color of dirt brown?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
@@ -309,7 +309,7 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 				Prompt:    "what is the origin of the us thanksgiving holiday?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
@@ -318,7 +318,7 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 				Prompt:    "what is the origin of independence day?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
@@ -327,7 +327,7 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 				Prompt:    "what is the composition of air?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},

kvcache/cache.go

@@ -62,6 +62,11 @@ type Cache interface {
 	// CopyPrefix copies tokens in the range [0, len) from srcSeq to dstSeq
 	CopyPrefix(srcSeq, dstSeq int, len int32)
 
+	// CanResume returns true if the cache can continue with the next token at
+	// the given position and sequence. Assumes that the caller has already
+	// verified the contents of the cache.
+	CanResume(seq int, pos int32) bool
+
 	// Remove deletes tokens in the range [beginIndex, endIndex) from seq. Set
 	// endIndex to math.MaxInt32 to remove everything starting at beginIndex.
 	//

kvcache/causal.go

@@ -581,6 +581,35 @@ func (c *Causal) CopyPrefix(srcSeq, dstSeq int, len int32) {
 	c.cellRanges[dstSeq] = seqRange
 }
 
+func (c *Causal) CanResume(seq int, pos int32) bool {
+	if c.windowSize == math.MaxInt32 {
+		return true
+	}
+
+	seqRange, ok := c.cellRanges[seq]
+	if !ok {
+		return false
+	}
+
+	// for sliding window, check that the window of the new sequence is contained in
+	// the window of what we are storing
+	var last int32 = -1
+	for i := seqRange.min; i <= seqRange.max; i++ {
+		if slices.Contains(c.cells[i].sequences, seq) {
+			last = max(last, c.cells[i].pos)
+		}
+	}
+
+	if last == -1 {
+		return false
+	}
+
+	lastWindowStart := max(0, last-c.windowSize)
+	posWindowStart := max(0, pos-c.windowSize)
+
+	return posWindowStart >= lastWindowStart
+}
+
 func (c *Causal) shift(seq int, beginIndex, offset int32) error {
 	if c.shiftFn == nil {
 		return ErrNotSupported
@@ -635,6 +664,12 @@ func (c *Causal) shift(seq int, beginIndex, offset int32) error {
 }
 
 func (c *Causal) Remove(seq int, beginIndex, endIndex int32) error {
+	// TODO(jessegross): We should check to see if removing the middle of the sequence will
+	// cause the sliding window to encompass tokens that we no longer have. If so, then we
+	// should return an error, which will trigger the runner to evaluate the full history and
+	// rebuild the window. However, if we have multimodal inputs in our history, this reuse
+	// results in use after free, so we don't do it for now.
+
 	var offset int32
 	if endIndex != math.MaxInt32 {
 		offset = beginIndex - endIndex
@@ -649,8 +684,7 @@ func (c *Causal) Remove(seq int, beginIndex, endIndex int32) error {
 			} else {
 				if c.cells[i].pos >= endIndex {
 					if slices.ContainsFunc(c.cells[i].sequences, func(s int) bool { return s != seq }) {
-						// TODO(jessegross): Need to be careful about data shared between sequences
-						return errors.New("shifting on cells shared by multiple sequences not yet implemented")
+						return errors.New("shifting cells shared by multiple sequences not supported")
 					}
 
 					c.cells[i].pos += offset

kvcache/causal_test.go

@@ -300,6 +300,77 @@ func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase)
 	}
 }
 
+func TestCanResume(t *testing.T) {
+	backend := &testBackend{}
+	windowSize := int32(4)
+	cache := NewSWACache(windowSize, nil)
+	defer cache.Close()
+
+	cache.Init(backend, ml.DTypeF16, 1, 16, 16)
+
+	context := backend.NewContext()
+	defer context.Close()
+
+	err := cache.StartForward(context, input.Batch{
+		Positions: []int32{0, 1, 2, 3},
+		Sequences: []int{0, 0, 0, 0},
+	})
+	if err != nil {
+		t.Fatalf("StartForward failed: %v", err)
+	}
+
+	cache.SetLayer(0)
+	tensor, _ := context.FromFloatSlice([]float32{1, 2, 3, 4}, 1, 1, 4)
+	cache.Put(context, tensor, tensor)
+
+	// with window size 4, nothing has slid out of the window yet
+	if !cache.CanResume(0, 0) {
+		t.Errorf("CanResume(0, 0) = false, want true (within window)")
+	}
+	if !cache.CanResume(0, 1) {
+		t.Errorf("CanResume(0, 1) = false, want true (within window)")
+	}
+	if !cache.CanResume(0, 2) {
+		t.Errorf("CanResume(0, 2) = false, want true (within window)")
+	}
+	if !cache.CanResume(0, 3) {
+		t.Errorf("CanResume(0, 3) = false, want true (latest position)")
+	}
+
+	// shift window by adding position 4
+	err = cache.StartForward(context, input.Batch{
+		Positions: []int32{4, 5},
+		Sequences: []int{0, 0},
+	})
+	if err != nil {
+		t.Fatalf("StartForward failed: %v", err)
+	}
+
+	cache.SetLayer(0)
+	tensor, _ = context.FromFloatSlice([]float32{5, 6}, 1, 1, 2)
+	cache.Put(context, tensor, tensor)
+
+	// only the latest position has overlapping windows
+	if cache.CanResume(0, 0) {
+		t.Errorf("after shift: CanResume(0, 0) = true, want false (outside window)")
+	}
+	if cache.CanResume(0, 1) {
+		t.Errorf("after shift: CanResume(0, 1) = true, want false (outside window)")
+	}
+	if cache.CanResume(0, 2) {
+		t.Errorf("after shift: CanResume(0, 2) = true, want false (outside window)")
+	}
+	if cache.CanResume(0, 3) {
+		t.Errorf("after shift: CanResume(0, 3) = true, want false (outside window)")
+	}
+	if cache.CanResume(0, 4) {
+		t.Errorf("after shift: CanResume(0, 4) = true, want false (outside window)")
+	}
+	if !cache.CanResume(0, 5) {
+		t.Errorf("after shift: CanResume(0, 5) = false, want true (latest position)")
+	}
+}
+
 type testBackend struct{}
 
 func (b *testBackend) Config() ml.Config {
@@ -462,7 +533,7 @@ func (t *testTensor) Conv2D(ctx ml.Context, weight ml.Tensor, s0, s1, p0, p1, d0
 	panic("not implemented")
 }
 
-func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, config ml.RoPEConfig) ml.Tensor {
+func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, dim, ropeType uint32, base, scale float32) ml.Tensor {
 	panic("not implemented")
 }
 

kvcache/encoder.go

@@ -134,6 +134,10 @@ func (c *EncoderCache) CopyPrefix(srcSeq, dstSeq int, len int32) {
 	panic("encoder cache does not support multiple sequences")
 }
 
+func (c *EncoderCache) CanResume(seq int, pos int32) bool {
+	return true
+}
+
 func (c *EncoderCache) Remove(seq int, beginIndex, endIndex int32) error {
 	if c.encoderPos >= beginIndex && c.encoderPos < endIndex {
 		c.encoderCached = false

kvcache/wrapper.go

@@ -87,6 +87,16 @@ func (c *WrapperCache) CopyPrefix(srcSeq, dstSeq int, len int32) {
 	}
 }
 
+func (c *WrapperCache) CanResume(seq int, pos int32) bool {
+	for _, cache := range c.caches {
+		if !cache.CanResume(seq, pos) {
+			return false
+		}
+	}
+
+	return true
+}
+
 func (c *WrapperCache) Remove(seq int, beginIndex, endIndex int32) error {
 	// If the one of these fails, the caller is supposed to retry with endIndex set to math.MaxInt32, which should not fail
 	for _, cache := range c.caches {

ml/backend.go

@@ -118,53 +118,6 @@ type Context interface {
 	Layer(int) Context
 }
 
-// RopeType represents different RoPE (Rotary Position Embedding) implementation types
-type RopeType int
-
-// Available RoPE implementation types
-const (
-	RopeTypeNormal RopeType = iota // Standard RoPE implementation
-	RopeTypeNeox                   // NeoX-style RoPE implementation
-	RopeTypeMRoPE                  // Multimodal RoPE implementation
-	RopeTypeVision                 // Vision-specific RoPE implementation
-)
-
-type YarnConfig struct {
-	YarnCtxTrain   int     // Context size used during training (for YaRN scaling)
-	YarnExtFactor  float32 // Extension factor for YaRN
-	YarnAttnFactor float32 // Attention scaling factor for YaRN
-	YarnBetaFast   float32 // Fast decay parameter for YaRN
-	YarnBetaSlow   float32 // Slow decay parameter for YaRN
-}
-
-// DefaultYarnConfig returns a default configuration for YaRN (Yet Another Rope Extension)
-func DefaultYarnConfig(nCtx int32) *YarnConfig {
-	return &YarnConfig{
-		YarnCtxTrain:   int(nCtx),
-		YarnExtFactor:  0.0,
-		YarnAttnFactor: 1.0,
-		YarnBetaFast:   32.0,
-		YarnBetaSlow:   1.0,
-	}
-}
-
-// RoPEConfig holds configuration for Rotary Position Embedding
-type RoPEConfig struct {
-	// Dim is the dimensionality for applying rotary embeddings
-	Dim uint32
-
-	// Type specifies the RoPE implementation variant
-	Type RopeType
-
-	// Base controls frequency decay for the embeddings
-	Base float32
-
-	// Scale allows scaling the effective context length
-	Scale float32
-
-	*YarnConfig
-}
-
 type Tensor interface {
 	Dim(n int) int
 	Stride(n int) int
@@ -188,7 +141,7 @@ type Tensor interface {
 	AvgPool2D(ctx Context, k, s int, p float32) Tensor
 	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 
-	RoPE(ctx Context, positionIDs, ropeFactors Tensor, config RoPEConfig) Tensor
+	RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim, ropeType uint32, base, scale float32) Tensor
 
 	Tanh(ctx Context) Tensor
 	GELU(ctx Context) Tensor

ml/backend/ggml/ggml.go

@@ -907,8 +907,6 @@ func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	}
 }
 
-// GGML RoPE types
-// These are the types used in the C implementation of RoPE
 const (
 	ropeTypeNorm   C.int = 0
 	ropeTypeNeox   C.int = 2
@@ -916,8 +914,7 @@ const (
 	ropeTypeVision C.int = 24
 )
 
-// RoPE applies Rotary Position Embeddings to the tensor
-func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, config ml.RoPEConfig) ml.Tensor {
+func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDim, ropeType uint32, ropeBase, ropeScale float32) ml.Tensor {
 	if ropeFactors == nil {
 		ropeFactors = &Tensor{b: t.b}
 	}
@@ -927,41 +924,19 @@ func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, config
 		dequant = C.ggml_cast(ctx.(*Context).ctx, t.t, C.GGML_TYPE_F32)
 	}
 
-	if config.YarnConfig == nil {
-		config.YarnConfig = ml.DefaultYarnConfig(131072) // 131072 is the default for LLaMA, so it is common at the time of writing
-	}
-
-	// Map Go RopeType to C implementation constants
-	var ropeTypeC C.int
-	switch config.Type {
-	case ml.RopeTypeNormal:
-		ropeTypeC = ropeTypeNorm
-	case ml.RopeTypeNeox:
-		ropeTypeC = ropeTypeNeox
-	case ml.RopeTypeMRoPE:
-		ropeTypeC = ropeTypeMrope
-	case ml.RopeTypeVision:
-		ropeTypeC = ropeTypeVision
-	default:
-		ropeTypeC = ropeTypeNorm
-	}
-
 	return &Tensor{
 		b: t.b,
 		t: C.ggml_rope_ext(
-			ctx.(*Context).ctx,
-			dequant,
-			positionIDs.(*Tensor).t,
-			ropeFactors.(*Tensor).t,
-			C.int(config.Dim),
-			ropeTypeC,
-			C.int(config.YarnCtxTrain),
-			C.float(config.Base),
-			C.float(config.Scale),
-			C.float(config.YarnExtFactor),
-			C.float(config.YarnAttnFactor),
-			C.float(config.YarnBetaFast),
-			C.float(config.YarnBetaSlow),
+			ctx.(*Context).ctx, dequant, positionIDs.(*Tensor).t, ropeFactors.(*Tensor).t,
+			C.int(ropeDim),
+			C.int(ropeType),
+			131072, // YaRN n_ctx_train
+			C.float(ropeBase),
+			C.float(ropeScale),
+			0.,  // YaRN ext_factor
+			1.,  // YaRN attn_factor
+			32., // YaRN beta_fast
+			1.,  // YaRN beta_slow
 		),
 	}
 }

model/models/gemma2/model.go

@@ -13,11 +13,10 @@ import (
 type Options struct {
 	hiddenSize, numHeads, numKVHeads int
 	attnKeyLen, attnValLen           int
-	eps                              float32
+	eps, ropeBase, ropeScale         float32
 	attnLogitSoftcap                 float32
 	finalLogitSoftcap                float32
 	largeModelScaling                bool
-	ropeConfig                       ml.RoPEConfig
 }
 
 type Model struct {
@@ -39,7 +38,6 @@ const (
 func New(c ml.Config) (model.Model, error) {
 	m := Model{
 		SentencePieceModel: model.NewSentencePieceModel(
-			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
 			&model.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),
@@ -56,15 +54,10 @@ func New(c ml.Config) (model.Model, error) {
 			attnKeyLen:        int(c.Uint("attention.key_length")),
 			attnValLen:        int(c.Uint("attention.value_length")),
 			eps:               c.Float("attention.layer_norm_rms_epsilon"),
+			ropeBase:          c.Float("rope.freq_base", 10000.0),
+			ropeScale:         c.Float("rope.freq_scale", 1.0),
 			attnLogitSoftcap:  c.Float("attn_logit_softcapping"),
 			finalLogitSoftcap: c.Float("final_logit_softcapping"),
-			ropeConfig: ml.RoPEConfig{
-				Base:       c.Float("rope.freq_base", 10000.0),
-				Scale:      c.Float("rope.freq_scale", 1.0),
-				Dim:        c.Uint("attention.key_length"),
-				Type:       ml.RopeTypeNormal,
-				YarnConfig: ml.DefaultYarnConfig(int32(c.Uint("context_length", 131072))),
-			},
 		},
 	}
 
@@ -84,10 +77,11 @@ type SelfAttention struct {
 
 func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
+	ropeType := uint32(2)
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
-	q = q.RoPE(ctx, positionIDs, nil, opts.ropeConfig)
+	q = q.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, opts.ropeBase, opts.ropeScale)
 
 	if opts.largeModelScaling {
 		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
@@ -97,7 +91,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, opts.attnKeyLen, opts.numKVHeads, batchSize)
-	k = k.RoPE(ctx, positionIDs, nil, opts.ropeConfig)
+	k = k.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, opts.ropeBase, opts.ropeScale)
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)
@@ -127,7 +121,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 }
 
 func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return key.RoPE(ctx, shift, nil, m.ropeConfig), nil
+	return key.RoPE(ctx, shift, nil, uint32(m.Options.attnKeyLen), uint32(2), m.Options.ropeBase, m.Options.ropeScale), nil
 }
 
 type MLP struct {

model/models/gemma3/model.go

@@ -55,7 +55,6 @@ func (p *MultiModalProjector) Forward(ctx ml.Context, visionOutputs ml.Tensor, i
 func New(c ml.Config) (model.Model, error) {
 	m := Model{
 		SentencePieceModel: model.NewSentencePieceModel(
-			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
 			&model.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),

model/models/gemma3/model_text.go

@@ -13,11 +13,9 @@ import (
 type TextOptions struct {
 	hiddenSize, numHeads, numKVHeads int
 	attnKeyLen, attnValLen           int
-	eps                              float32
+	eps, ropeScale                   float32
+	ropeLocalBase, ropeGlobalBase    float32
 	largeModelScaling                bool
-
-	ropeLocalConfig  ml.RoPEConfig
-	ropeGlobalConfig ml.RoPEConfig
 }
 
 type TextModel struct {
@@ -47,7 +45,6 @@ func newTextModel(c ml.Config) *TextModel {
 
 	m := TextModel{
 		SentencePieceModel: model.NewSentencePieceModel(
-			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
 			&model.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),
@@ -58,27 +55,15 @@ func newTextModel(c ml.Config) *TextModel {
 		),
 		Layers: make([]TextLayer, numBlocks),
 		TextOptions: &TextOptions{
-			hiddenSize: int(c.Uint("embedding_length")),
-			numHeads:   int(c.Uint("attention.head_count")),
-			numKVHeads: int(c.Uint("attention.head_count_kv")),
-			attnKeyLen: int(c.Uint("attention.key_length", 256)),
-			attnValLen: int(c.Uint("attention.value_length", 256)),
-			eps:        c.Float("attention.layer_norm_rms_epsilon", 1e-06),
-
-			ropeLocalConfig: ml.RoPEConfig{
-				Base:       c.Float("rope.local.freq_base", 10000.0),
-				Scale:      c.Float("rope.freq_scale", 1.0),
-				Dim:        c.Uint("attention.key_length", 256),
-				Type:       ml.RopeTypeNeox,
-				YarnConfig: ml.DefaultYarnConfig(int32(c.Uint("context_length", 131072))),
-			},
-			ropeGlobalConfig: ml.RoPEConfig{
-				Base:       c.Float("rope.global.freq_base", 1000000.0),
-				Scale:      c.Float("rope.freq_scale", 1.0),
-				Dim:        c.Uint("attention.key_length", 256),
-				Type:       ml.RopeTypeNeox,
-				YarnConfig: ml.DefaultYarnConfig(int32(c.Uint("context_length", 131072))),
-			},
+			hiddenSize:     int(c.Uint("embedding_length")),
+			numHeads:       int(c.Uint("attention.head_count")),
+			numKVHeads:     int(c.Uint("attention.head_count_kv")),
+			attnKeyLen:     int(c.Uint("attention.key_length", 256)),
+			attnValLen:     int(c.Uint("attention.value_length", 256)),
+			eps:            c.Float("attention.layer_norm_rms_epsilon", 1e-06),
+			ropeLocalBase:  c.Float("rope.local.freq_base", 10000.0),
+			ropeGlobalBase: c.Float("rope.global.freq_base", 1000000.0),
+			ropeScale:      c.Float("rope.freq_scale", 1.0),
 		},
 	}
 
@@ -100,16 +85,17 @@ type TextSelfAttention struct {
 
 func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *TextOptions) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
+	ropeType := uint32(2)
 
-	ropeConfig := opts.ropeLocalConfig
+	ropeBase := opts.ropeLocalBase
 	if (layer+1)%gemmaGlobalCacheCount == 0 {
-		ropeConfig = opts.ropeGlobalConfig
+		ropeBase = opts.ropeGlobalBase
 	}
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
 	q = sa.QueryNorm.Forward(ctx, q, opts.eps)
-	q = q.RoPE(ctx, positionIDs, nil, ropeConfig)
+	q = q.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, ropeBase, opts.ropeScale)
 
 	if opts.largeModelScaling {
 		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
@@ -120,7 +106,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, pos
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, opts.attnKeyLen, opts.numKVHeads, batchSize)
 	k = sa.KeyNorm.Forward(ctx, k, opts.eps)
-	k = k.RoPE(ctx, positionIDs, nil, ropeConfig)
+	k = k.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, ropeBase, opts.ropeScale)
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)
@@ -133,12 +119,12 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, pos
 }
 
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	ropeConfig := m.ropeLocalConfig
+	ropeBase := m.TextOptions.ropeLocalBase
 	if (layer+1)%gemmaGlobalCacheCount == 0 {
-		ropeConfig = m.ropeGlobalConfig
+		ropeBase = m.TextOptions.ropeGlobalBase
 	}
 
-	return key.RoPE(ctx, shift, nil, ropeConfig), nil
+	return key.RoPE(ctx, shift, nil, uint32(m.TextOptions.attnKeyLen), uint32(2), ropeBase, m.TextOptions.ropeScale), nil
 }
 
 type TextMLP struct {

model/models/llama/model.go

@@ -14,8 +14,8 @@ import (
 
 type Options struct {
 	hiddenSize, numHeads, numKVHeads int
-	eps                              float32
-	ropeConfig                       ml.RoPEConfig
+	eps, ropeBase, ropeScale         float32
+	ropeDim                          uint32
 }
 
 type Model struct {
@@ -54,13 +54,9 @@ func New(c ml.Config) (model.Model, error) {
 			numHeads:   int(c.Uint("attention.head_count")),
 			numKVHeads: int(c.Uint("attention.head_count_kv")),
 			eps:        c.Float("attention.layer_norm_rms_epsilon"),
-			ropeConfig: ml.RoPEConfig{
-				Base:       c.Float("rope.freq_base"),
-				Scale:      c.Float("rope.freq_scale", 1),
-				Dim:        c.Uint("rope.dimension_count"),
-				Type:       ml.RopeTypeNormal,
-				YarnConfig: ml.DefaultYarnConfig(int32(c.Uint("context_length", 131072))),
-			},
+			ropeBase:   c.Float("rope.freq_base"),
+			ropeScale:  c.Float("rope.freq_scale", 1),
+			ropeDim:    c.Uint("rope.dimension_count"),
 		},
 	}
 
@@ -80,14 +76,15 @@ type SelfAttention struct {
 func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
 	headDim := opts.hiddenSize / opts.numHeads
+	ropeType := uint32(0)
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	q = q.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeConfig)
+	q = q.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
 
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	k = k.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeConfig)
+	k = k.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -100,7 +97,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 }
 
 func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return key.RoPE(ctx, shift, m.Layers[layer].SelfAttention.RopeFactors, m.ropeConfig), nil
+	return key.RoPE(ctx, shift, m.Layers[layer].SelfAttention.RopeFactors, uint32(0), m.ropeDim, m.ropeBase, m.ropeScale), nil
 }
 
 type MLP struct {

model/models/mllama/model_text.go

@@ -20,14 +20,15 @@ type TextSelfAttention struct {
 func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions, _ ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
 	headDim := opts.hiddenSize / opts.numHeads
+	ropeType := uint32(0)
 
 	query := sa.Query.Forward(ctx, hiddenState)
 	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	query = query.RoPE(ctx, positions, sa.RopeFactors, opts.ropeConfig)
+	query = query.RoPE(ctx, positions, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
 
 	key := sa.Key.Forward(ctx, hiddenState)
 	key = key.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	key = key.RoPE(ctx, positions, sa.RopeFactors, opts.ropeConfig)
+	key = key.RoPE(ctx, positions, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
 
 	value := sa.Value.Forward(ctx, hiddenState)
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -42,7 +43,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions, _ m
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
 	// This will only get called for layers in the cache, which are just the self attention layers
 	if sa, ok := m.Transformer.Layers[layer].(*TextSelfAttentionDecoderLayer); ok {
-		return key.RoPE(ctx, shift, sa.SelfAttention.RopeFactors, m.ropeConfig), nil
+		return key.RoPE(ctx, shift, sa.SelfAttention.RopeFactors, m.ropeDim, uint32(0), m.ropeBase, m.ropeScale), nil
 	}
 
 	return key, nil
@@ -197,8 +198,8 @@ func (d *TextDecoder) Forward(ctx ml.Context, hiddenState, positionIDs, outputs,
 
 type TextModelOptions struct {
 	hiddenSize, numHeads, numKVHeads int
-	eps                              float32
-	ropeConfig                       ml.RoPEConfig
+	eps, ropeBase, ropeScale         float32
+	ropeDim                          uint32
 
 	crossAttentionLayers []uint32
 }
@@ -239,14 +240,10 @@ func newTextModel(c ml.Config) *TextModel {
 			numHeads:             int(c.Uint("attention.head_count")),
 			numKVHeads:           int(c.Uint("attention.head_count_kv")),
 			eps:                  c.Float("attention.layer_norm_rms_epsilon"),
+			ropeBase:             c.Float("rope.freq_base"),
+			ropeScale:            c.Float("rope.freq_scale", 1),
+			ropeDim:              c.Uint("rope.dimension_count"),
 			crossAttentionLayers: c.Uints("attention.cross_attention_layers"),
-			ropeConfig: ml.RoPEConfig{
-				Base:       c.Float("rope.freq_base"),
-				Scale:      c.Float("rope.freq_scale", 1),
-				Dim:        c.Uint("rope.dimension_count"),
-				Type:       ml.RopeTypeNormal,
-				YarnConfig: ml.DefaultYarnConfig(int32(c.Uint("context_length", 131072))),
-			},
 		},
 	}
 }

model/process_text_spm.go

@@ -1,29 +1,23 @@
 package model
 
 import (
-	"iter"
+	"container/heap"
+	"fmt"
 	"log/slog"
+	"strconv"
 	"strings"
-
-	"github.com/dlclark/regexp2"
-	queue "github.com/emirpasic/gods/v2/queues/priorityqueue"
 )
 
 const spmWhitespaceSep = "▁"
 
-func replaceWhitespaceBySeperator(s string) string {
-	return strings.ReplaceAll(s, " ", spmWhitespaceSep)
-}
-
 type SentencePieceModel struct {
 	maxTokenLen int
-	pre         *regexp2.Regexp
 	vocab       *Vocabulary
 }
 
 var _ TextProcessor = (*SentencePieceModel)(nil)
 
-func NewSentencePieceModel(pre string, vocab *Vocabulary) SentencePieceModel {
+func NewSentencePieceModel(vocab *Vocabulary) SentencePieceModel {
 	slog.Debug("Tokens", "num tokens", len(vocab.Values), "vals", vocab.Values[:5], "scores", vocab.Scores[:5], "types", vocab.Types[:5])
 
 	counter := map[int]int{}
@@ -44,7 +38,6 @@ func NewSentencePieceModel(pre string, vocab *Vocabulary) SentencePieceModel {
 
 	return SentencePieceModel{
 		maxTokenLen: maxTokenLen,
-		pre:         regexp2.MustCompile(pre, regexp2.Unicode|regexp2.RE2),
 		vocab:       vocab,
 	}
 }
@@ -53,20 +46,9 @@ func (spm SentencePieceModel) Is(id int32, special Special) bool {
 	return spm.vocab.Is(id, special)
 }
 
-func (spm *SentencePieceModel) split(s string) iter.Seq[string] {
-	return func(yield func(string) bool) {
-		for m, _ := spm.pre.FindStringMatch(s); m != nil; m, _ = spm.pre.FindNextMatch(m) {
-			if !yield(m.String()) {
-				break
-			}
-		}
-	}
-}
-
 func (spm SentencePieceModel) Encode(s string, addSpecial bool) ([]int32, error) {
 	fragments := []fragment{{value: s}}
 	for _, special := range spm.vocab.SpecialVocabulary() {
-		// TODO: process special tokens concurrently
 		id := spm.vocab.Encode(special)
 		for i := 0; i < len(fragments); i++ {
 			frag := fragments[i]
@@ -91,7 +73,6 @@ func (spm SentencePieceModel) Encode(s string, addSpecial bool) ([]int32, error)
 			fragments = append(fragments[:i], append(middle, fragments[i+1:]...)...)
 		}
 	}
-	slog.Debug("fragments", "frags", fragments)
 
 	var ids []int32
 	for _, frag := range fragments {
@@ -100,105 +81,96 @@ func (spm SentencePieceModel) Encode(s string, addSpecial bool) ([]int32, error)
 			continue
 		}
 
-		for split := range spm.split(frag.value) {
-			split = replaceWhitespaceBySeperator(split)
+		text := strings.ReplaceAll(frag.value, " ", spmWhitespaceSep)
 
-			var sb strings.Builder
-			sb.Write([]byte(split))
-			if id := spm.vocab.Encode(sb.String()); id >= 0 {
-				ids = append(ids, id)
-				continue
+		if id := spm.vocab.Encode(text); id >= 0 {
+			ids = append(ids, id)
+			continue
+		}
+
+		q := &queue{}
+		heap.Init(q)
+
+		runes := []rune(text)
+		merges := make([]merge, len(runes))
+		for r := range runes {
+			merges[r] = merge{
+				p:     r - 1,
+				n:     r + 1,
+				runes: []rune{runes[r]},
 			}
+		}
 
-			runes := []rune(sb.String())
-			pq := queue.NewWith(func(a, b any) int {
-				priA := a.(*candidate)
-				priB := b.(*candidate)
-				if priA.score > priB.score || (priA.score == priB.score && priA.a < priB.a) {
-					return -1
-				}
-				return 1
-			})
-
-			merges := make([]merge, len(runes))
-			for r := range runes {
-				merges[r] = merge{
-					p:     r - 1,
-					n:     r + 1,
-					runes: []rune{runes[r]},
-				}
-			}
-
-			slog.Debug("tokenizer", "merges", merges)
-
-			pairwise := func(a, b int) *candidate {
-				if a < 0 || b >= len(runes) {
-					return nil
-				}
-
-				left, right := string(merges[a].runes), string(merges[b].runes)
-				if id := spm.vocab.Encode(left + right); id >= 0 {
-					return &candidate{
-						a:     a,
-						b:     b,
-						score: spm.vocab.Scores[id],
-					}
-				}
+		pairwise := func(a, b int) *candidate {
+			if a < 0 || b >= len(runes) {
 				return nil
 			}
 
-			for i := range len(runes) - 1 {
-				if pair := pairwise(i, i+1); pair != nil {
-					pq.Enqueue(pair)
+			left, right := string(merges[a].runes), string(merges[b].runes)
+			if id := spm.vocab.Encode(left + right); id >= 0 {
+				return &candidate{
+					a:     a,
+					b:     b,
+					score: spm.vocab.Scores[id],
+					size:  len(left) + len(right),
 				}
 			}
 
-			pqv := pq.Values()
-			for _, v := range pqv {
-				e := v.(*candidate)
-				slog.Debug("candidate", "candidate", e)
+			return nil
+		}
+
+		for i := range len(runes) - 1 {
+			if pair := pairwise(i, i+1); pair != nil {
+				heap.Push(q, pair)
+			}
+		}
+
+		for q.Len() > 0 {
+			pair := heap.Pop(q).(*candidate)
+			left, right := merges[pair.a], merges[pair.b]
+
+			if string(left.runes) == "" || string(right.runes) == "" || len(string(left.runes))+len(string(right.runes)) != pair.size {
+				continue
 			}
 
-			for !pq.Empty() {
-				v, _ := pq.Dequeue()
-				pair := v.(*candidate)
-				left, right := merges[pair.a], merges[pair.b]
+			merges[pair.a].runes = append(left.runes, right.runes...)
+			merges[pair.b].runes = nil
+			merges[pair.a].n = right.n
+			if right.n < len(merges) {
+				merges[right.n].p = pair.a
+			}
 
-				slog.Debug("pair", "left", left, "right", right)
-				if len(left.runes) == 0 || len(right.runes) == 0 {
+			if pair := pairwise(merges[pair.a].p, pair.a); pair != nil {
+				heap.Push(q, pair)
+			}
+
+			if pair := pairwise(pair.a, merges[pair.a].n); pair != nil {
+				heap.Push(q, pair)
+			}
+		}
+
+		for _, merge := range merges {
+			if token := string(merge.runes); token != "" {
+				id := spm.vocab.Encode(token)
+
+				if id >= 0 {
+					ids = append(ids, id)
 					continue
 				}
 
-				if id := spm.vocab.Encode(string(left.runes) + string(right.runes)); id < 0 {
-					continue
-				}
-
-				merges[pair.a].runes = append(left.runes, right.runes...)
-				merges[pair.b].runes = nil
-				merges[pair.a].n = right.n
-				if right.n < len(merges) {
-					merges[right.n].p = pair.a
-				}
-
-				if pair := pairwise(merges[pair.a].p, pair.a); pair != nil {
-					pq.Enqueue(pair)
-				}
-
-				if pair := pairwise(pair.a, merges[pair.a].n); pair != nil {
-					pq.Enqueue(pair)
-				}
-			}
-
-			slog.Debug("merges", "merges", merges)
-
-			for _, merge := range merges {
-				if len(merge.runes) > 0 {
-					if id := spm.vocab.Encode(string(merge.runes)); id >= 0 {
-						ids = append(ids, id)
+				// Fallback to byte tokenization
+				var result []int32
+				for _, b := range []byte(token) {
+					byteToken := fmt.Sprintf("<0x%02X>", b)
+					unknownID := spm.vocab.Encode(byteToken)
+					if unknownID >= 0 {
+						result = append(result, unknownID)
 					} else {
-						slog.Debug("missing token", "token", string(merge.runes))
+						slog.Debug("unknown byte token", "byte", b, "token", byteToken)
 					}
 				}
+
+				ids = append(ids, result...)
 			}
 		}
 	}
@@ -229,6 +201,30 @@ func (spm SentencePieceModel) Encode(s string, addSpecial bool) ([]int32, error)
 type candidate struct {
 	a, b  int
 	score float32
+	size  int
+}
+
+type queue []*candidate
+
+func (q queue) Len() int { return len(q) }
+
+func (q queue) Less(i, j int) bool {
+	return (q[i].score > q[j].score) || (q[i].score == q[j].score && q[i].a < q[j].a)
+}
+
+func (q queue) Swap(i, j int) { q[i], q[j] = q[j], q[i] }
+
+func (q *queue) Push(x interface{}) {
+	item := x.(*candidate)
+	*q = append(*q, item)
+}
+
+func (q *queue) Pop() interface{} {
+	old := *q
+	n := len(old)
+	item := old[n-1]
+	*q = old[0 : n-1]
+	return item
 }
 
 func (spm SentencePieceModel) Decode(ids []int32) (string, error) {
@@ -236,11 +232,26 @@ func (spm SentencePieceModel) Decode(ids []int32) (string, error) {
 	for _, id := range ids {
 		data := spm.vocab.Decode(id)
 		data = strings.ReplaceAll(data, spmWhitespaceSep, " ")
-		if _, err := sb.WriteString(data); err != nil {
-			return "", err
+
+		// For tokenizers that use byte tokens like "<0xEA>"
+		// convert them to the partial unicode character
+		// so they are buffered correctly by the runner instead
+		// of being sent back to the api as "<0xEA>"
+		if len(data) == 6 && strings.HasPrefix(data, "<0x") && strings.HasSuffix(data, ">") {
+			byteVal, err := strconv.ParseUint(data[1:5], 0, 8)
+			if err != nil {
+				return "", fmt.Errorf("failed to parse hex byte: %v", err)
+			}
+
+			if err := sb.WriteByte(byte(byteVal)); err != nil {
+				return "", err
+			}
+		} else {
+			if _, err := sb.WriteString(data); err != nil {
+				return "", err
+			}
 		}
 	}
 
-	slog.Debug("decoded", "ids", ids, "text", sb.String())
 	return sb.String(), nil
 }

model/process_text_spm_test.go

@@ -25,8 +25,6 @@ func loadSentencePieceVocab(t *testing.T) SentencePieceModel {
 		t.Fatal(err)
 	}
 
-	preTokenizer := `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`
-
 	var v Vocabulary
 
 	for _, piece := range spm.GetPieces() {
@@ -47,7 +45,7 @@ func loadSentencePieceVocab(t *testing.T) SentencePieceModel {
 		}
 	}
 
-	return NewSentencePieceModel(preTokenizer, &v)
+	return NewSentencePieceModel(&v)
 }
 
 func TestSentencePieceEncode(t *testing.T) {
@@ -116,3 +114,59 @@ func TestSentencePieceEncode(t *testing.T) {
 		}
 	})
 }
+
+func TestSentencePieceModelDecodeByteTokens(t *testing.T) {
+	vocab := &Vocabulary{
+		Values: []string{
+			"normal",
+			"<0xEA>",
+			"<0x41>",
+			"<0xC3>",
+			"<0xA3>",
+		},
+		Types: []uint32{
+			TOKEN_TYPE_NORMAL,
+			TOKEN_TYPE_BYTE,
+			TOKEN_TYPE_BYTE,
+			TOKEN_TYPE_BYTE,
+			TOKEN_TYPE_BYTE,
+		},
+		Scores: []float32{0, 0, 0, 0, 0},
+	}
+
+	spm := NewSentencePieceModel(vocab)
+
+	tests := []struct {
+		name     string
+		ids      []int32
+		expected string
+	}{
+		{
+			name:     "single byte token",
+			ids:      []int32{1},
+			expected: "\xea",
+		},
+		{
+			name:     "ASCII byte token",
+			ids:      []int32{2},
+			expected: "A",
+		},
+		{
+			name:     "multiple byte tokens forming UTF-8 character",
+			ids:      []int32{3, 4},
+			expected: "ã",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result, err := spm.Decode(tt.ids)
+			if err != nil {
+				t.Errorf("failed to decode token IDs %v: %v", tt.ids, err)
+			}
+			if result != tt.expected {
+				t.Errorf("got %q, want %q", result, tt.expected)
+			}
+		})
+	}
+}

openai/openai.go

@@ -23,10 +23,10 @@ import (
 var finishReasonToolCalls = "tool_calls"
 
 type Error struct {
-	Message string      `json:"message"`
-	Type    string      `json:"type"`
-	Param   interface{} `json:"param"`
-	Code    *string     `json:"code"`
+	Message string  `json:"message"`
+	Type    string  `json:"type"`
+	Param   any     `json:"param"`
+	Code    *string `json:"code"`
 }
 
 type ErrorResponse struct {
@@ -465,7 +465,7 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 		}
 	}
 
-	options := make(map[string]interface{})
+	options := make(map[string]any)
 
 	switch stop := r.Stop.(type) {
 	case string:

openai/openai_test.go

@@ -219,7 +219,7 @@ func TestChatMiddleware(t *testing.T) {
 							{
 								Function: api.ToolCallFunction{
 									Name: "get_current_weather",
-									Arguments: map[string]interface{}{
+									Arguments: map[string]any{
 										"location": "Paris, France",
 										"format":   "celsius",
 									},

runner/ollamarunner/cache.go

@@ -118,6 +118,10 @@ func (c *InputCache) LoadCacheSlot(prompt []input.Input) (*InputCacheSlot, []inp
 	}
 
 	if c.cache != nil {
+		if numPast > 0 && !c.cache.CanResume(slot.Id, numPast) {
+			numPast = 0
+		}
+
 		err = c.cache.Remove(slot.Id, numPast, math.MaxInt32)
 		if err != nil {
 			// Some models don't support partial erasure
@@ -225,6 +229,8 @@ func countCommonPrefix(a []input.Input, b []input.Input) int32 {
 	return count
 }
 
+// TODO(jessegross): If we need to reprocess the inputs we should ensure that
+// we don't split up a SameBatch
 func (c *InputCache) ShiftDiscard(inputLen int32, numKeep int32) int32 {
 	targetFree := (c.numCtx - numKeep) / 2
 	targetFree = max(targetFree, 1)

runner/ollamarunner/cache_test.go

@@ -451,6 +451,7 @@ func (m *mockCache) Close()
 func (m *mockCache) StartForward(ctx ml.Context, batch input.Batch) error                          { return nil }
 func (m *mockCache) CopyPrefix(srcSeq, dstSeq int, len int32)                                      {}
 func (m *mockCache) SetConfig(ml.CacheConfig)                                                      {}
+func (m *mockCache) CanResume(seq int, pos int32) bool                                             { return true }
 
 func TestShiftCacheSlot(t *testing.T) {
 	tests := []struct {

runner/ollamarunner/runner.go

@@ -115,16 +115,41 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 		params.numKeep = int32(len(inputs))
 	}
 
-	// TODO(jessegross): We should ensure that we always leave minBatch of context space to shift,
-	// otherwise we might truncate or split the batch against the model's wishes
-
 	// Ensure that at least 1 input can be discarded during shift
 	params.numKeep = min(params.numKeep, s.cache.numCtx-1)
 
 	if int32(len(inputs)) > s.cache.numCtx {
 		discard := int32(len(inputs)) - s.cache.numCtx
+		promptStart := params.numKeep + discard
+
+		// If we need to truncate in the middle of a unbreakable batch, remove the entire batch
+		sameBatch := 0
+		for i, inp := range inputs {
+			if sameBatch > 0 {
+				sameBatch--
+
+				if promptStart == int32(i) {
+					promptStart++
+				}
+			} else if promptStart == int32(i) {
+				break
+			}
+
+			if inp.SameBatch != 0 {
+				if int32(i) < params.numKeep {
+					return nil, fmt.Errorf("SameBatch may not be specified within numKeep (index: %v numKeep: %v SameBatch: %v)", i, params.numKeep, inp.SameBatch)
+				}
+
+				sameBatch = inp.SameBatch
+			}
+		}
+
+		if promptStart >= int32(len(inputs)) {
+			return nil, errors.New("entire prompt removed by truncation")
+		}
+
 		newInputs := inputs[:params.numKeep]
-		newInputs = append(newInputs, inputs[params.numKeep+discard:]...)
+		newInputs = append(newInputs, inputs[promptStart:]...)
 
 		slog.Warn("truncating input prompt", "limit", s.cache.numCtx, "prompt", len(inputs), "keep", params.numKeep, "new", len(newInputs))
 		inputs = newInputs

server/images.go

@@ -35,17 +35,11 @@ var (
 	errCapabilityCompletion = errors.New("completion")
 	errCapabilityTools      = errors.New("tools")
 	errCapabilityInsert     = errors.New("insert")
+	errCapabilityVision     = errors.New("vision")
+	errCapabilityEmbedding  = errors.New("embedding")
 	errInsecureProtocol     = errors.New("insecure protocol http")
 )
 
-type Capability string
-
-const (
-	CapabilityCompletion = Capability("completion")
-	CapabilityTools      = Capability("tools")
-	CapabilityInsert     = Capability("insert")
-)
-
 type registryOptions struct {
 	Insecure bool
 	Username string
@@ -66,52 +60,83 @@ type Model struct {
 	System         string
 	License        []string
 	Digest         string
-	Options        map[string]interface{}
+	Options        map[string]any
 	Messages       []api.Message
 
 	Template *template.Template
 }
 
+// Capabilities returns the capabilities that the model supports
+func (m *Model) Capabilities() []model.Capability {
+	capabilities := []model.Capability{}
+
+	// Check for completion capability
+	r, err := os.Open(m.ModelPath)
+	if err == nil {
+		defer r.Close()
+
+		f, _, err := ggml.Decode(r, 0)
+		if err == nil {
+			if _, ok := f.KV()[fmt.Sprintf("%s.pooling_type", f.KV().Architecture())]; ok {
+				capabilities = append(capabilities, model.CapabilityEmbedding)
+			} else {
+				capabilities = append(capabilities, model.CapabilityCompletion)
+			}
+			if _, ok := f.KV()[fmt.Sprintf("%s.vision.block_count", f.KV().Architecture())]; ok {
+				capabilities = append(capabilities, model.CapabilityVision)
+			}
+		} else {
+			slog.Error("couldn't decode ggml", "error", err)
+		}
+	} else {
+		slog.Error("couldn't open model file", "error", err)
+	}
+
+	if m.Template == nil {
+		return capabilities
+	}
+
+	// Check for tools capability
+	if slices.Contains(m.Template.Vars(), "tools") {
+		capabilities = append(capabilities, model.CapabilityTools)
+	}
+
+	// Check for insert capability
+	if slices.Contains(m.Template.Vars(), "suffix") {
+		capabilities = append(capabilities, model.CapabilityInsert)
+	}
+
+	return capabilities
+}
+
 // CheckCapabilities checks if the model has the specified capabilities returning an error describing
 // any missing or unknown capabilities
-func (m *Model) CheckCapabilities(caps ...Capability) error {
+func (m *Model) CheckCapabilities(want ...model.Capability) error {
+	available := m.Capabilities()
 	var errs []error
-	for _, cap := range caps {
-		switch cap {
-		case CapabilityCompletion:
-			r, err := os.Open(m.ModelPath)
-			if err != nil {
-				slog.Error("couldn't open model file", "error", err)
-				continue
-			}
-			defer r.Close()
 
-			// TODO(mxyng): decode the GGML into model to avoid doing this multiple times
-			f, _, err := ggml.Decode(r, 0)
-			if err != nil {
-				slog.Error("couldn't decode ggml", "error", err)
-				continue
-			}
+	// Map capabilities to their corresponding error
+	capToErr := map[model.Capability]error{
+		model.CapabilityCompletion: errCapabilityCompletion,
+		model.CapabilityTools:      errCapabilityTools,
+		model.CapabilityInsert:     errCapabilityInsert,
+		model.CapabilityVision:     errCapabilityVision,
+		model.CapabilityEmbedding:  errCapabilityEmbedding,
+	}
 
-			if _, ok := f.KV()[fmt.Sprintf("%s.pooling_type", f.KV().Architecture())]; ok {
-				errs = append(errs, errCapabilityCompletion)
-			}
-		case CapabilityTools:
-			if !slices.Contains(m.Template.Vars(), "tools") {
-				errs = append(errs, errCapabilityTools)
-			}
-		case CapabilityInsert:
-			vars := m.Template.Vars()
-			if !slices.Contains(vars, "suffix") {
-				errs = append(errs, errCapabilityInsert)
-			}
-		default:
+	for _, cap := range want {
+		err, ok := capToErr[cap]
+		if !ok {
 			slog.Error("unknown capability", "capability", cap)
 			return fmt.Errorf("unknown capability: %s", cap)
 		}
+
+		if !slices.Contains(available, cap) {
+			errs = append(errs, err)
+		}
 	}
 
-	if err := errors.Join(errs...); err != nil {
+	if len(errs) > 0 {
 		return fmt.Errorf("%w %w", errCapabilities, errors.Join(errs...))
 	}
 

server/images_test.go

@@ -0,0 +1,360 @@
+package server
+
+import (
+	"bytes"
+	"encoding/binary"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/ollama/ollama/template"
+	"github.com/ollama/ollama/types/model"
+)
+
+// Constants for GGUF magic bytes and version
+var (
+	ggufMagic = []byte{0x47, 0x47, 0x55, 0x46} // "GGUF"
+	ggufVer   = uint32(3)                      // Version 3
+)
+
+// Helper function to create mock GGUF data
+func createMockGGUFData(architecture string, vision bool) []byte {
+	var buf bytes.Buffer
+
+	// Write GGUF header
+	buf.Write(ggufMagic)
+	binary.Write(&buf, binary.LittleEndian, ggufVer)
+
+	// Write tensor count (0 for our test)
+	var numTensors uint64 = 0
+	binary.Write(&buf, binary.LittleEndian, numTensors)
+
+	// Calculate number of metadata entries
+	numMetaEntries := uint64(1) // architecture entry
+	if vision {
+		numMetaEntries++
+	}
+	// Add embedding entry if architecture is "bert"
+	if architecture == "bert" {
+		numMetaEntries++
+	}
+	binary.Write(&buf, binary.LittleEndian, numMetaEntries)
+
+	// Write architecture metadata
+	archKey := "general.architecture"
+	keyLen := uint64(len(archKey))
+	binary.Write(&buf, binary.LittleEndian, keyLen)
+	buf.WriteString(archKey)
+
+	// String type (8)
+	var strType uint32 = 8
+	binary.Write(&buf, binary.LittleEndian, strType)
+
+	// String length
+	strLen := uint64(len(architecture))
+	binary.Write(&buf, binary.LittleEndian, strLen)
+	buf.WriteString(architecture)
+
+	if vision {
+		visionKey := architecture + ".vision.block_count"
+		keyLen = uint64(len(visionKey))
+		binary.Write(&buf, binary.LittleEndian, keyLen)
+		buf.WriteString(visionKey)
+
+		// uint32 type (4)
+		var uint32Type uint32 = 4
+		binary.Write(&buf, binary.LittleEndian, uint32Type)
+
+		// uint32 value (1)
+		var countVal uint32 = 1
+		binary.Write(&buf, binary.LittleEndian, countVal)
+	}
+	// Write embedding metadata if architecture is "bert"
+	if architecture == "bert" {
+		poolKey := architecture + ".pooling_type"
+		keyLen = uint64(len(poolKey))
+		binary.Write(&buf, binary.LittleEndian, keyLen)
+		buf.WriteString(poolKey)
+
+		// uint32 type (4)
+		var uint32Type uint32 = 4
+		binary.Write(&buf, binary.LittleEndian, uint32Type)
+
+		// uint32 value (1)
+		var poolingVal uint32 = 1
+		binary.Write(&buf, binary.LittleEndian, poolingVal)
+	}
+
+	return buf.Bytes()
+}
+
+func TestModelCapabilities(t *testing.T) {
+	// Create a temporary directory for test files
+	tempDir, err := os.MkdirTemp("", "model_capabilities_test")
+	if err != nil {
+		t.Fatalf("Failed to create temp directory: %v", err)
+	}
+	defer os.RemoveAll(tempDir)
+
+	// Create different types of mock model files
+	completionModelPath := filepath.Join(tempDir, "model.bin")
+	visionModelPath := filepath.Join(tempDir, "vision_model.bin")
+	embeddingModelPath := filepath.Join(tempDir, "embedding_model.bin")
+	// Create a simple model file for tests that don't depend on GGUF content
+	simpleModelPath := filepath.Join(tempDir, "simple_model.bin")
+
+	err = os.WriteFile(completionModelPath, createMockGGUFData("llama", false), 0o644)
+	if err != nil {
+		t.Fatalf("Failed to create completion model file: %v", err)
+	}
+	err = os.WriteFile(visionModelPath, createMockGGUFData("llama", true), 0o644)
+	if err != nil {
+		t.Fatalf("Failed to create completion model file: %v", err)
+	}
+	err = os.WriteFile(embeddingModelPath, createMockGGUFData("bert", false), 0o644)
+	if err != nil {
+		t.Fatalf("Failed to create embedding model file: %v", err)
+	}
+	err = os.WriteFile(simpleModelPath, []byte("dummy model data"), 0o644)
+	if err != nil {
+		t.Fatalf("Failed to create simple model file: %v", err)
+	}
+
+	toolsInsertTemplate, err := template.Parse("{{ .prompt }}{{ if .tools }}{{ .tools }}{{ end }}{{ if .suffix }}{{ .suffix }}{{ end }}")
+	if err != nil {
+		t.Fatalf("Failed to parse template: %v", err)
+	}
+	chatTemplate, err := template.Parse("{{ .prompt }}")
+	if err != nil {
+		t.Fatalf("Failed to parse template: %v", err)
+	}
+	toolsTemplate, err := template.Parse("{{ .prompt }}{{ if .tools }}{{ .tools }}{{ end }}")
+	if err != nil {
+		t.Fatalf("Failed to parse template: %v", err)
+	}
+
+	testModels := []struct {
+		name         string
+		model        Model
+		expectedCaps []model.Capability
+	}{
+		{
+			name: "model with completion capability",
+			model: Model{
+				ModelPath: completionModelPath,
+				Template:  chatTemplate,
+			},
+			expectedCaps: []model.Capability{model.CapabilityCompletion},
+		},
+
+		{
+			name: "model with completion, tools, and insert capability",
+			model: Model{
+				ModelPath: completionModelPath,
+				Template:  toolsInsertTemplate,
+			},
+			expectedCaps: []model.Capability{model.CapabilityCompletion, model.CapabilityTools, model.CapabilityInsert},
+		},
+		{
+			name: "model with tools and insert capability",
+			model: Model{
+				ModelPath: simpleModelPath,
+				Template:  toolsInsertTemplate,
+			},
+			expectedCaps: []model.Capability{model.CapabilityTools, model.CapabilityInsert},
+		},
+		{
+			name: "model with tools capability",
+			model: Model{
+				ModelPath: simpleModelPath,
+				Template:  toolsTemplate,
+			},
+			expectedCaps: []model.Capability{model.CapabilityTools},
+		},
+		{
+			name: "model with vision capability",
+			model: Model{
+				ModelPath: visionModelPath,
+				Template:  chatTemplate,
+			},
+			expectedCaps: []model.Capability{model.CapabilityCompletion, model.CapabilityVision},
+		},
+		{
+			name: "model with vision, tools, and insert capability",
+			model: Model{
+				ModelPath: visionModelPath,
+				Template:  toolsInsertTemplate,
+			},
+			expectedCaps: []model.Capability{model.CapabilityCompletion, model.CapabilityVision, model.CapabilityTools, model.CapabilityInsert},
+		},
+		{
+			name: "model with embedding capability",
+			model: Model{
+				ModelPath: embeddingModelPath,
+				Template:  chatTemplate,
+			},
+			expectedCaps: []model.Capability{model.CapabilityEmbedding},
+		},
+	}
+
+	// compare two slices of model.Capability regardless of order
+	compareCapabilities := func(a, b []model.Capability) bool {
+		if len(a) != len(b) {
+			return false
+		}
+
+		aCount := make(map[model.Capability]int)
+		for _, cap := range a {
+			aCount[cap]++
+		}
+
+		bCount := make(map[model.Capability]int)
+		for _, cap := range b {
+			bCount[cap]++
+		}
+
+		for cap, count := range aCount {
+			if bCount[cap] != count {
+				return false
+			}
+		}
+
+		return true
+	}
+
+	for _, tt := range testModels {
+		t.Run(tt.name, func(t *testing.T) {
+			// Test Capabilities method
+			caps := tt.model.Capabilities()
+			if !compareCapabilities(caps, tt.expectedCaps) {
+				t.Errorf("Expected capabilities %v, got %v", tt.expectedCaps, caps)
+			}
+		})
+	}
+}
+
+func TestModelCheckCapabilities(t *testing.T) {
+	// Create a temporary directory for test files
+	tempDir, err := os.MkdirTemp("", "model_check_capabilities_test")
+	if err != nil {
+		t.Fatalf("Failed to create temp directory: %v", err)
+	}
+	defer os.RemoveAll(tempDir)
+
+	visionModelPath := filepath.Join(tempDir, "vision_model.bin")
+	simpleModelPath := filepath.Join(tempDir, "model.bin")
+	embeddingModelPath := filepath.Join(tempDir, "embedding_model.bin")
+
+	err = os.WriteFile(simpleModelPath, []byte("dummy model data"), 0o644)
+	if err != nil {
+		t.Fatalf("Failed to create simple model file: %v", err)
+	}
+	err = os.WriteFile(visionModelPath, createMockGGUFData("llama", true), 0o644)
+	if err != nil {
+		t.Fatalf("Failed to create vision model file: %v", err)
+	}
+	err = os.WriteFile(embeddingModelPath, createMockGGUFData("bert", false), 0o644)
+	if err != nil {
+		t.Fatalf("Failed to create embedding model file: %v", err)
+	}
+
+	toolsInsertTemplate, err := template.Parse("{{ .prompt }}{{ if .tools }}{{ .tools }}{{ end }}{{ if .suffix }}{{ .suffix }}{{ end }}")
+	if err != nil {
+		t.Fatalf("Failed to parse template: %v", err)
+	}
+	chatTemplate, err := template.Parse("{{ .prompt }}")
+	if err != nil {
+		t.Fatalf("Failed to parse template: %v", err)
+	}
+	toolsTemplate, err := template.Parse("{{ .prompt }}{{ if .tools }}{{ .tools }}{{ end }}")
+	if err != nil {
+		t.Fatalf("Failed to parse template: %v", err)
+	}
+
+	tests := []struct {
+		name           string
+		model          Model
+		checkCaps      []model.Capability
+		expectedErrMsg string
+	}{
+		{
+			name: "completion model without tools capability",
+			model: Model{
+				ModelPath: simpleModelPath,
+				Template:  chatTemplate,
+			},
+			checkCaps:      []model.Capability{model.CapabilityTools},
+			expectedErrMsg: "does not support tools",
+		},
+		{
+			name: "model with all needed capabilities",
+			model: Model{
+				ModelPath: simpleModelPath,
+				Template:  toolsInsertTemplate,
+			},
+			checkCaps: []model.Capability{model.CapabilityTools, model.CapabilityInsert},
+		},
+		{
+			name: "model missing insert capability",
+			model: Model{
+				ModelPath: simpleModelPath,
+				Template:  toolsTemplate,
+			},
+			checkCaps:      []model.Capability{model.CapabilityInsert},
+			expectedErrMsg: "does not support insert",
+		},
+		{
+			name: "model missing vision capability",
+			model: Model{
+				ModelPath: simpleModelPath,
+				Template:  toolsTemplate,
+			},
+			checkCaps:      []model.Capability{model.CapabilityVision},
+			expectedErrMsg: "does not support vision",
+		},
+		{
+			name: "model with vision capability",
+			model: Model{
+				ModelPath: visionModelPath,
+				Template:  chatTemplate,
+			},
+			checkCaps: []model.Capability{model.CapabilityVision},
+		},
+		{
+			name: "model with embedding capability",
+			model: Model{
+				ModelPath: embeddingModelPath,
+				Template:  chatTemplate,
+			},
+			checkCaps: []model.Capability{model.CapabilityEmbedding},
+		},
+		{
+			name: "unknown capability",
+			model: Model{
+				ModelPath: simpleModelPath,
+				Template:  chatTemplate,
+			},
+			checkCaps:      []model.Capability{"unknown"},
+			expectedErrMsg: "unknown capability",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			// Test CheckCapabilities method
+			err := tt.model.CheckCapabilities(tt.checkCaps...)
+			if tt.expectedErrMsg == "" {
+				if err != nil {
+					t.Errorf("Expected no error, got: %v", err)
+				}
+			} else {
+				if err == nil {
+					t.Errorf("Expected error containing %q, got nil", tt.expectedErrMsg)
+				} else if !strings.Contains(err.Error(), tt.expectedErrMsg) {
+					t.Errorf("Expected error containing %q, got: %v", tt.expectedErrMsg, err)
+				}
+			}
+		})
+	}
+}

server/routes.go

@@ -72,7 +72,7 @@ var (
 	errBadTemplate = errors.New("template error")
 )
 
-func modelOptions(model *Model, requestOpts map[string]interface{}) (api.Options, error) {
+func modelOptions(model *Model, requestOpts map[string]any) (api.Options, error) {
 	opts := api.DefaultOptions()
 	if err := opts.FromMap(model.Options); err != nil {
 		return api.Options{}, err
@@ -87,7 +87,7 @@ func modelOptions(model *Model, requestOpts map[string]interface{}) (api.Options
 
 // scheduleRunner schedules a runner after validating inputs such as capabilities and model options.
 // It returns the allocated runner, model instance, and consolidated options if successful and error otherwise.
-func (s *Server) scheduleRunner(ctx context.Context, name string, caps []Capability, requestOpts map[string]any, keepAlive *api.Duration) (llm.LlamaServer, *Model, *api.Options, error) {
+func (s *Server) scheduleRunner(ctx context.Context, name string, caps []model.Capability, requestOpts map[string]any, keepAlive *api.Duration) (llm.LlamaServer, *Model, *api.Options, error) {
 	if name == "" {
 		return nil, nil, nil, fmt.Errorf("model %w", errRequired)
 	}
@@ -144,7 +144,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	model, err := GetModel(name.String())
+	m, err := GetModel(name.String())
 	if err != nil {
 		switch {
 		case errors.Is(err, fs.ErrNotExist):
@@ -159,7 +159,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 
 	// expire the runner
 	if req.Prompt == "" && req.KeepAlive != nil && int(req.KeepAlive.Seconds()) == 0 {
-		s.sched.expireRunner(model)
+		s.sched.expireRunner(m)
 
 		c.JSON(http.StatusOK, api.GenerateResponse{
 			Model:      req.Model,
@@ -176,9 +176,9 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	caps := []Capability{CapabilityCompletion}
+	caps := []model.Capability{model.CapabilityCompletion}
 	if req.Suffix != "" {
-		caps = append(caps, CapabilityInsert)
+		caps = append(caps, model.CapabilityInsert)
 	}
 
 	r, m, opts, err := s.scheduleRunner(c.Request.Context(), name.String(), caps, req.Options, req.KeepAlive)
@@ -203,7 +203,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	isMllama := checkMllamaModelFamily(model)
+	isMllama := checkMllamaModelFamily(m)
 	if isMllama && len(req.Images) > 1 {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "this model only supports one image: more than one image sent"})
 		return
@@ -211,7 +211,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 
 	images := make([]llm.ImageData, len(req.Images))
 	for i := range req.Images {
-		if isMllama && len(model.ProjectorPaths) > 0 {
+		if isMllama && len(m.ProjectorPaths) > 0 {
 			data, opts, err := mllama.Preprocess(bytes.NewReader(req.Images[i]))
 			if err != nil {
 				c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": "error processing image"})
@@ -422,7 +422,7 @@ func (s *Server) EmbedHandler(c *gin.Context) {
 		return
 	}
 
-	r, m, opts, err := s.scheduleRunner(c.Request.Context(), name.String(), []Capability{}, req.Options, req.KeepAlive)
+	r, m, opts, err := s.scheduleRunner(c.Request.Context(), name.String(), []model.Capability{}, req.Options, req.KeepAlive)
 	if err != nil {
 		handleScheduleError(c, req.Model, err)
 		return
@@ -530,7 +530,7 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 		return
 	}
 
-	r, _, _, err := s.scheduleRunner(c.Request.Context(), name.String(), []Capability{}, req.Options, req.KeepAlive)
+	r, _, _, err := s.scheduleRunner(c.Request.Context(), name.String(), []model.Capability{}, req.Options, req.KeepAlive)
 	if err != nil {
 		handleScheduleError(c, req.Model, err)
 		return
@@ -813,19 +813,20 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	}
 
 	resp := &api.ShowResponse{
-		License:    strings.Join(m.License, "\n"),
-		System:     m.System,
-		Template:   m.Template.String(),
-		Details:    modelDetails,
-		Messages:   msgs,
-		ModifiedAt: manifest.fi.ModTime(),
+		License:      strings.Join(m.License, "\n"),
+		System:       m.System,
+		Template:     m.Template.String(),
+		Details:      modelDetails,
+		Messages:     msgs,
+		Capabilities: m.Capabilities(),
+		ModifiedAt:   manifest.fi.ModTime(),
 	}
 
 	var params []string
 	cs := 30
 	for k, v := range m.Options {
 		switch val := v.(type) {
-		case []interface{}:
+		case []any:
 			for _, nv := range val {
 				params = append(params, fmt.Sprintf("%-*s %#v", cs, k, nv))
 			}
@@ -1335,7 +1336,7 @@ func Serve(ln net.Listener) error {
 	return nil
 }
 
-func waitForStream(c *gin.Context, ch chan interface{}) {
+func waitForStream(c *gin.Context, ch chan any) {
 	c.Header("Content-Type", "application/json")
 	for resp := range ch {
 		switch r := resp.(type) {
@@ -1468,9 +1469,9 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
-	caps := []Capability{CapabilityCompletion}
+	caps := []model.Capability{model.CapabilityCompletion}
 	if len(req.Tools) > 0 {
-		caps = append(caps, CapabilityTools)
+		caps = append(caps, model.CapabilityTools)
 	}
 
 	name := model.ParseName(req.Model)

server/sched.go

@@ -20,6 +20,7 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/types/model"
 )
 
 type LlmRequest struct {
@@ -37,7 +38,7 @@ type Scheduler struct {
 	pendingReqCh  chan *LlmRequest
 	finishedReqCh chan *LlmRequest
 	expiredCh     chan *runnerRef
-	unloadedCh    chan interface{}
+	unloadedCh    chan any
 
 	loaded   map[string]*runnerRef
 	loadedMu sync.Mutex
@@ -67,7 +68,7 @@ func InitScheduler(ctx context.Context) *Scheduler {
 		pendingReqCh:  make(chan *LlmRequest, maxQueue),
 		finishedReqCh: make(chan *LlmRequest, maxQueue),
 		expiredCh:     make(chan *runnerRef, maxQueue),
-		unloadedCh:    make(chan interface{}, maxQueue),
+		unloadedCh:    make(chan any, maxQueue),
 		loaded:        make(map[string]*runnerRef),
 		newServerFn:   llm.NewLlamaServer,
 		getGpuFn:      discover.GetGPUInfo,
@@ -195,7 +196,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 					}
 
 					// Embedding models should always be loaded with parallel=1
-					if pending.model.CheckCapabilities(CapabilityCompletion) != nil {
+					if pending.model.CheckCapabilities(model.CapabilityCompletion) != nil {
 						numParallel = 1
 					}
 
@@ -617,8 +618,8 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 // a before and after GPU memory allocation.  The returned channel
 // will be notified when we're done waiting, or have timed out and should
 // proceed anyway
-func (runner *runnerRef) waitForVRAMRecovery() chan interface{} {
-	finished := make(chan interface{}, 1)
+func (runner *runnerRef) waitForVRAMRecovery() chan any {
+	finished := make(chan any, 1)
 
 	// CPU or Metal don't need checking, so no waiting required
 	// windows can page VRAM, only cuda currently can report accurate used vram usage

types/model/capability.go

@@ -0,0 +1,15 @@
+package model
+
+type Capability string
+
+const (
+	CapabilityCompletion = Capability("completion")
+	CapabilityTools      = Capability("tools")
+	CapabilityInsert     = Capability("insert")
+	CapabilityVision     = Capability("vision")
+	CapabilityEmbedding  = Capability("embedding")
+)
+
+func (c Capability) String() string {
+	return string(c)
+}