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>

CMakeLists.txt

@@ -86,9 +86,9 @@ if(CMAKE_CUDA_COMPILER)
     )
 endif()
 
-set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a):xnack[+-]$"
+set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a|1200|1201):xnack[+-]$"
     CACHE STRING
-    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a):xnack[+-]$\"."
+    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a|1200|1201):xnack[+-]$\"."
 )
 
 check_language(HIP)
@@ -97,7 +97,7 @@ if(CMAKE_HIP_COMPILER)
 
     find_package(hip REQUIRED)
     if(NOT AMDGPU_TARGETS)
-        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012])$")
+        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012]|120[01])$")
     elseif(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
         list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
     endif()

CMakePresets.json

@@ -56,7 +56,7 @@
       "name": "ROCm 6",
       "inherits": [ "ROCm" ],
       "cacheVariables": {
-        "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
+        "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
       }
     }
   ],

README.md

@@ -285,6 +285,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
 - [HTML UI](https://github.com/rtcfirefly/ollama-ui)
 - [Saddle](https://github.com/jikkuatwork/saddle)
+- [TagSpaces](https://www.tagspaces.org) (A platform for file based apps, [utilizing Ollama](https://docs.tagspaces.org/ai/) for the generation of tags and descriptions)
 - [Chatbot UI](https://github.com/ivanfioravanti/chatbot-ollama)
 - [Chatbot UI v2](https://github.com/mckaywrigley/chatbot-ui)
 - [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
@@ -324,6 +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)
@@ -394,6 +396,8 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Reins](https://github.com/ibrahimcetin/reins) (Easily tweak parameters, customize system prompts per chat, and enhance your AI experiments with reasoning model support.)
 - [Ellama](https://github.com/zeozeozeo/ellama) (Friendly native app to chat with an Ollama instance)
 - [screenpipe](https://github.com/mediar-ai/screenpipe) Build agents powered by your screen history
+- [Ollamb](https://github.com/hengkysteen/ollamb) (Simple yet rich in features, cross-platform built with Flutter and designed for Ollama. Try the [web demo](https://hengkysteen.github.io/demo/ollamb/).)
+- [Writeopia](https://github.com/Writeopia/Writeopia) (Text editor with integration with Ollama)
 
 ### Cloud
 
@@ -433,7 +437,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [SwollamaCLI](https://github.com/marcusziade/Swollama) bundled with the Swollama Swift package. [Demo](https://github.com/marcusziade/Swollama?tab=readme-ov-file#cli-usage)
 - [aichat](https://github.com/sigoden/aichat) All-in-one LLM CLI tool featuring Shell Assistant, Chat-REPL, RAG, AI tools & agents, with access to OpenAI, Claude, Gemini, Ollama, Groq, and more.
 - [PowershAI](https://github.com/rrg92/powershai) PowerShell module that brings AI to terminal on Windows, including support for Ollama
+- [DeepShell](https://github.com/Abyss-c0re/deepshell) Your self-hosted AI assistant. Interactive Shell, Files and Folders analysis.
 - [orbiton](https://github.com/xyproto/orbiton) Configuration-free text editor and IDE with support for tab completion with Ollama.
+- [orca-cli](https://github.com/molbal/orca-cli) Ollama Registry CLI Application - Browse, pull and download models from Ollama Registry in your terminal.
 
 ### Apple Vision Pro
 

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)

discover/gpu_linux.go

@@ -111,6 +111,7 @@ func GetCPUDetails() ([]CPU, error) {
 	if err != nil {
 		return nil, err
 	}
+	defer file.Close()
 	return linuxCPUDetails(file)
 }
 
@@ -168,13 +169,11 @@ func linuxCPUDetails(file io.Reader) ([]CPU, error) {
 	for id, s := range socketByID {
 		s.CoreCount = len(coreBySocket[id])
 		s.ThreadCount = 0
-		for _, tc := range threadsByCoreBySocket[id] {
-			s.ThreadCount += tc
-		}
 
 		// This only works if HT is enabled, consider a more reliable model, maybe cache size comparisons?
 		efficiencyCoreCount := 0
 		for _, threads := range threadsByCoreBySocket[id] {
+			s.ThreadCount += threads
 			if threads == 1 {
 				efficiencyCoreCount++
 			}

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"
+  ],
 }
 ```
 

docs/faq.md

@@ -20,7 +20,13 @@ Please refer to the [GPU docs](./gpu.md).
 
 ## How can I specify the context window size?
 
-By default, Ollama uses a context window size of 2048 tokens. This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context length to 8K, use: `OLLAMA_CONTEXT_LENGTH=8192 ollama serve`.
+By default, Ollama uses a context window size of 2048 tokens. 
+
+This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use: 
+
+```shell
+OLLAMA_CONTEXT_LENGTH=8192 ollama serve
+```
 
 To change this when using `ollama run`, use `/set parameter`:
 

docs/troubleshooting.md

@@ -9,7 +9,7 @@ cat ~/.ollama/logs/server.log
 On **Linux** systems with systemd, the logs can be found with this command:
 
 ```shell
-journalctl -u ollama --no-pager
+journalctl -u ollama --no-pager --follow --pager-end 
 ```
 
 When you run Ollama in a **container**, the logs go to stdout/stderr in the container:

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)
 	}

fs/ggml/ggml.go

@@ -413,7 +413,7 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	}, offset, nil
 }
 
-func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialOffload, fullOffload uint64) {
+func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
 	embedding := f.KV().EmbeddingLength()
 	heads := f.KV().HeadCount()
 	headsKV := f.KV().HeadCountKV()
@@ -426,7 +426,10 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	layers := f.Tensors().GroupLayers()
 
 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
-	kv = uint64(float64(context*f.KV().BlockCount()*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	kv = make([]uint64, f.KV().BlockCount())
+	for i := range kv {
+		kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	}
 
 	switch f.KV().Architecture() {
 	case "llama":
@@ -460,16 +463,14 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	case "mllama":
 		var visionTokens, tiles uint64 = 1601, 4
 
-		if crossAttentionLayers, ok := f.KV()["mllama.attention.cross_attention_layers"].(*array); ok {
-			kv = headsKV *
-				(embeddingHeadsK + embeddingHeadsV) * // one for K, one for V
-				(2* // sizeof(float16)
-					(f.KV().BlockCount()-uint64(crossAttentionLayers.size))* // num non-cross attention layers
-					context +
-					4* // sizeof(float32)
-						uint64(crossAttentionLayers.size)* // num cross attention layers
-						visionTokens*
-						tiles)
+		crossAttentionLayers := f.KV().Uints("attention.cross_attention_layers")
+		for i := range kv {
+			if slices.Contains(crossAttentionLayers, uint32(i)) {
+				kv[i] = headsKV * (embeddingHeadsK + embeddingHeadsV) *
+					4 * // sizeof(float32)
+					visionTokens *
+					tiles
+			}
 		}
 
 		fullOffload = max(
@@ -505,6 +506,20 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 				4*embeddingHeadsK*context*8+
 				embedding*embeddingHeadsK*heads*9/16,
 		)
+
+		// Gemma2 also has sliding window attention but we only have an optimized implementation in the Ollama
+		// engine. Gemma3 always uses the Ollama engine.
+		if f.KV().Architecture() == "gemma3" {
+			const gemma3GlobalCacheCount = 6
+			slidingWindow := (uint64(numParallel) * uint64(f.KV().Uint("attention.sliding_window"))) + batch
+			for i := range kv {
+				// Every 6th layer is a global layer, which is the full context size that has already been set. The other
+				// layers are the smaller local (sliding) layers.
+				if (i+1)%gemma3GlobalCacheCount != 0 {
+					kv[i] = uint64(float64(slidingWindow*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+				}
+			}
+		}
 	case "command-r":
 		fullOffload = max(
 			4*batch*(embedding+vocab),

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

@@ -119,10 +119,10 @@ func (c *Causal) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity
 	}
 
 	var cacheSize int
-	if c.windowSize == math.MaxInt32 || capacity < int(c.windowSize)+maxBatch {
+	if c.windowSize == math.MaxInt32 || capacity < int(c.windowSize) {
 		cacheSize = maxSequences * capacity
 	} else {
-		cacheSize = maxSequences * (int(c.windowSize) + maxBatch)
+		cacheSize = (maxSequences * int(c.windowSize)) + maxBatch
 	}
 	cacheSize = roundUp(cacheSize, c.config.CachePadding)
 	c.cells = make([]cacheCell, cacheSize)
@@ -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 {
@@ -362,7 +433,6 @@ func (c *testContext) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
 }
 
 func (c *testContext) Input() ml.Context    { return c }
-func (c *testContext) Output() ml.Context   { return c }
 func (c *testContext) Layer(int) ml.Context { return c }
 
 func (c *testContext) Forward(...ml.Tensor) ml.Context { return c }

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 {

llama/llama.go

@@ -166,6 +166,10 @@ func (c *Context) KvCacheDefrag() {
 	C.llama_kv_cache_defrag(c.c)
 }
 
+func (c *Context) KvCacheCanShift() bool {
+	return bool(C.llama_kv_cache_can_shift(c.c))
+}
+
 // Get the embeddings for a sequence id
 func (c *Context) GetEmbeddingsSeq(seqId int) []float32 {
 	e := unsafe.Pointer(C.llama_get_embeddings_seq(c.c, C.int(seqId)))

llama/patches/0022-add-rdna4-support.patch

@@ -0,0 +1,103 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Saman <saman.khatir@amd.com>
+Date: Wed, 19 Mar 2025 14:02:26 -0700
+Subject: [PATCH] add rdna4 support
+
+---
+ ggml/src/ggml-cuda/common.cuh    | 6 ++++--
+ ggml/src/ggml-cuda/mmq.cu        | 2 +-
+ ggml/src/ggml-cuda/mmq.cuh       | 4 ++--
+ ggml/src/ggml-cuda/mmvq.cu       | 4 ++--
+ ggml/src/ggml-cuda/vendors/hip.h | 4 ++++
+ 5 files changed, 13 insertions(+), 7 deletions(-)
+
+diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
+index adf0d3ec..b24593fc 100644
+--- a/ggml/src/ggml-cuda/common.cuh
++++ b/ggml/src/ggml-cuda/common.cuh
+@@ -61,11 +61,13 @@
+ #define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
+ #define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
+ #define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
++#define GGML_CUDA_CC_RDNA4      (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000
+ 
+ #define GGML_CUDA_CC_IS_RDNA(cc)  (cc >= GGML_CUDA_CC_RDNA1)
+ #define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
+ #define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
+-#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3)
++#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4)
++#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
+ #define GGML_CUDA_CC_IS_GCN(cc)   (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA)
+ #define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
+ 
+@@ -386,7 +388,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
+ #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+ #if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
+     c = __builtin_amdgcn_sdot4(a, b, c, false);
+-#elif defined(RDNA3)
++#elif defined(RDNA3) || defined(RDNA4)
+     c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
+ #elif defined(__gfx1010__) || defined(__gfx900__)
+     int tmp1;
+diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
+index 10f2ebb1..933d945c 100644
+--- a/ggml/src/ggml-cuda/mmq.cu
++++ b/ggml/src/ggml-cuda/mmq.cu
+@@ -149,5 +149,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
+         return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+     }
+ 
+-    return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
++    return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+ }
+diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
+index 0451c65f..66ce2bc9 100644
+--- a/ggml/src/ggml-cuda/mmq.cuh
++++ b/ggml/src/ggml-cuda/mmq.cuh
+@@ -2577,9 +2577,9 @@ static __device__ void mul_mat_q_process_tile(
+ 
+ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
+ #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+-#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
++#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+     __launch_bounds__(WARP_SIZE*nwarps, 2)
+-#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
++#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+ #else
+ #if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+     __launch_bounds__(WARP_SIZE*nwarps, 1)
+diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
+index 4fb466ca..23ae7abc 100644
+--- a/ggml/src/ggml-cuda/mmvq.cu
++++ b/ggml/src/ggml-cuda/mmvq.cu
+@@ -62,13 +62,13 @@ static __global__ void mul_mat_vec_q(
+ 
+     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
+ 
+-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
++#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4))
+     constexpr int nwarps              = 1;
+     constexpr int rows_per_cuda_block = 1;
+ #else
+     constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
+     constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
+-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
++#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3) && !defined(RDNA4)
+ 
+     const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+     const     int row0 = rows_per_cuda_block*blockIdx.x;
+diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
+index 81964611..a62544b5 100644
+--- a/ggml/src/ggml-cuda/vendors/hip.h
++++ b/ggml/src/ggml-cuda/vendors/hip.h
+@@ -150,6 +150,10 @@
+ #define CDNA
+ #endif
+ 
++#if defined(__gfx1200__) || defined(__gfx1201__)
++#define RDNA4
++#endif
++
+ #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
+     defined(__gfx1150__) || defined(__gfx1151__)
+ #define RDNA3

llm/memory.go

@@ -15,12 +15,12 @@ import (
 )
 
 // This algorithm looks for a complete fit to determine if we need to unload other models
-func PredictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, projectors []string, opts api.Options) (bool, uint64) {
+func PredictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (bool, uint64) {
 	// Split up the GPUs by type and try them
 	var estimatedVRAM uint64
 	for _, gpus := range allGpus.ByLibrary() {
 		var layerCount int
-		estimate := EstimateGPULayers(gpus, f, projectors, opts)
+		estimate := EstimateGPULayers(gpus, f, projectors, opts, numParallel)
 		layerCount, estimatedVRAM = estimate.Layers, estimate.VRAMSize
 		if opts.NumGPU < 0 {
 			if layerCount > 0 && layerCount >= int(f.KV().BlockCount()+1) {
@@ -71,7 +71,7 @@ type MemoryEstimate struct {
 
 // Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
 // The GPUs provided must all be the same Library
-func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []string, opts api.Options) MemoryEstimate {
+func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []string, opts api.Options, numParallel int) MemoryEstimate {
 	// Graph size for a partial offload, applies to all GPUs
 	var graphPartialOffload uint64
 
@@ -137,13 +137,19 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		}
 	}
 
-	kv, graphPartialOffload, graphFullOffload := f.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)), kvct)
+	kv, graphPartialOffload, graphFullOffload := f.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)), numParallel, kvct)
 
-	// KV is proportional to the number of layers
-	layerSize += kv / f.KV().BlockCount()
+	if len(kv) > 0 {
+		layerSize += kv[0]
+	}
+
+	var kvTotal uint64
+	for _, kvLayer := range kv {
+		kvTotal += kvLayer
+	}
 
 	if graphPartialOffload == 0 {
-		graphPartialOffload = f.KV().GQA() * kv / 6
+		graphPartialOffload = f.KV().GQA() * kvTotal / 6
 	}
 	if graphFullOffload == 0 {
 		graphFullOffload = graphPartialOffload
@@ -217,7 +223,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		// Some models have inconsistent layer sizes
 		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
 			layerSize = blk.Size()
-			layerSize += kv / f.KV().BlockCount()
+			layerSize += kv[i]
 			memoryWeights += blk.Size()
 		}
 
@@ -315,7 +321,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		layersRequested:     opts.NumGPU,
 		layersModel:         int(f.KV().BlockCount()) + 1,
 		availableList:       availableList,
-		kv:                  kv,
+		kv:                  kvTotal,
 		allocationsList:     allocationsList,
 		memoryWeights:       memoryWeights,
 		memoryLayerOutput:   memoryLayerOutput,
@@ -374,7 +380,7 @@ func (m MemoryEstimate) LogValue() slog.Value {
 			slog.Group(
 				"weights",
 				// memory of the weights
-				"total", format.HumanBytes2(m.memoryWeights),
+				"total", format.HumanBytes2(m.memoryWeights+m.memoryLayerOutput),
 				// memory of repeating layers
 				"repeating", format.HumanBytes2(m.memoryWeights),
 				// memory of non-repeating layers

llm/memory_test.go

@@ -61,7 +61,7 @@ func TestEstimateGPULayers(t *testing.T) {
 	projectors := []string{}
 	opts := api.DefaultOptions()
 	t.Run("cpu", func(t *testing.T) {
-		estimate := EstimateGPULayers(gpus, ggml, projectors, opts)
+		estimate := EstimateGPULayers(gpus, ggml, projectors, opts, 1)
 		assert.Equal(t, 0, estimate.Layers)
 		assert.Equal(t, uint64(0), estimate.Graph)
 	})
@@ -112,7 +112,7 @@ func TestEstimateGPULayers(t *testing.T) {
 			gpus[1].FreeMemory += gpuMinimumMemory + layerSize + s.layer1*layerSize + 1
 			gpus[0].FreeMemory += max(graphFullOffload, graphPartialOffload)
 			gpus[1].FreeMemory += max(graphFullOffload, graphPartialOffload)
-			estimate := EstimateGPULayers(gpus, ggml, projectors, opts)
+			estimate := EstimateGPULayers(gpus, ggml, projectors, opts, 1)
 			assert.Equal(t, int(s.expect0+s.expect1), estimate.Layers, "scenario %d: %v", i, s)
 			assert.Equal(t, fmt.Sprintf("%d,%d", s.expect0, s.expect1), estimate.TensorSplit, "scenario %d: %v", i, s)
 			var layerSums uint64

llm/server.go

@@ -109,7 +109,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		gpus = discover.GetCPUInfo()
 	}
 
-	estimate := EstimateGPULayers(gpus, f, projectors, opts)
+	estimate := EstimateGPULayers(gpus, f, projectors, opts, numParallel)
 	if len(gpus) > 1 || gpus[0].Library != "cpu" {
 		switch {
 		case gpus[0].Library == "metal" && estimate.VRAMSize > systemTotalMemory:

ml/backend.go

@@ -110,12 +110,10 @@ type Context interface {
 	MaxGraphNodes() int
 	Close()
 
-	// Input returns a context appropriate for creating input tensors
+	// Input returns a context appropriate for creating tensors that are
+	// inputs to the model (which includes things like output locations)
 	Input() Context
 
-	// Output returns a context appropriate for creating output tensors
-	Output() Context
-
 	// Layer returns a context appropriate for creating intermediate tensors
 	Layer(int) Context
 }

ml/backend/ggml/ggml.go

@@ -48,9 +48,6 @@ type Backend struct {
 	// input is the backend used for inputs
 	input *C.struct_ggml_backend_buffer_type
 
-	// output is the backend used for outputs
-	output *C.struct_ggml_backend_buffer_type
-
 	// layers is the backend used for repeating layers
 	layers map[int]*C.struct_ggml_backend_buffer_type
 
@@ -400,8 +397,7 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 			C.size_t(maxGraphNodes),
 			C._Bool(len(gpus) > 1 && slices.Contains(gpus, output.d)),
 		),
-		input:  deviceBufferTypes[input.d],
-		output: deviceBufferTypes[output.d],
+		input: deviceBufferTypes[input.d],
 		layers: func() map[int]*C.struct_ggml_backend_buffer_type {
 			m := make(map[int]*C.struct_ggml_backend_buffer_type)
 			for i, layer := range layers {
@@ -482,19 +478,6 @@ func (c Context) Input() ml.Context {
 	return &c
 }
 
-func (c Context) Output() ml.Context {
-	if c.b.output != nil {
-		return &Context{
-			b:             c.b,
-			ctx:           c.ctx,
-			buft:          c.b.output,
-			maxGraphNodes: c.maxGraphNodes,
-		}
-	}
-
-	return &c
-}
-
 func (c Context) Layer(i int) ml.Context {
 	if buft, ok := c.b.layers[i]; ok {
 		return &Context{

ml/backend/ggml/ggml/src/ggml-cuda/common.cuh

@@ -61,11 +61,13 @@
 #define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
 #define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
 #define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
+#define GGML_CUDA_CC_RDNA4      (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000
 
 #define GGML_CUDA_CC_IS_RDNA(cc)  (cc >= GGML_CUDA_CC_RDNA1)
 #define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
 #define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
-#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3)
+#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4)
+#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_GCN(cc)   (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA)
 #define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
 
@@ -386,7 +388,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
-#elif defined(RDNA3)
+#elif defined(RDNA3) || defined(RDNA4)
     c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
 #elif defined(__gfx1010__) || defined(__gfx900__)
     int tmp1;

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

@@ -149,5 +149,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
     }
 
-    return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }

ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh

@@ -2577,9 +2577,9 @@ static __device__ void mul_mat_q_process_tile(
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
     __launch_bounds__(WARP_SIZE*nwarps, 2)
-#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
 #else
 #if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
     __launch_bounds__(WARP_SIZE*nwarps, 1)

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

@@ -62,13 +62,13 @@ static __global__ void mul_mat_vec_q(
 
     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4))
     constexpr int nwarps              = 1;
     constexpr int rows_per_cuda_block = 1;
 #else
     constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
     constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3) && !defined(RDNA4)
 
     const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
     const     int row0 = rows_per_cuda_block*blockIdx.x;

ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h

@@ -150,6 +150,10 @@
 #define CDNA
 #endif
 
+#if defined(__gfx1200__) || defined(__gfx1201__)
+#define RDNA4
+#endif
+
 #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
     defined(__gfx1150__) || defined(__gfx1151__)
 #define RDNA3

model/models/gemma2/model.go

@@ -38,7 +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"),

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

@@ -45,7 +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"),

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/llamarunner/cache.go

@@ -213,8 +213,16 @@ func (c *InputCache) ShiftDiscard(inputLen int, numKeep int) int {
 	return discard
 }
 
-// Frees up space in the KV cache by deleting the oldest half of history and shifting
-// the newest half into that space (saving numKeep inputs at the beginning).
+type ErrReprocessInputs struct {
+	Inputs []input
+}
+
+func (e *ErrReprocessInputs) Error() string {
+	return fmt.Sprintf("kv cache shift not supported, inputs need reprocessing (input count: %v)", len(e.Inputs))
+}
+
+// ShiftCacheSlot frees up space in the KV cache by deleting the oldest half of history
+// and shifting the newest half into that space (saving numKeep inputs at the beginning).
 //
 // Assumes that at least 1 entry can be freed up by shifting (i.e. numKeep < numCtx)
 func (c *InputCache) ShiftCacheSlot(slot *InputCacheSlot, numKeep int) error {
@@ -222,7 +230,8 @@ func (c *InputCache) ShiftCacheSlot(slot *InputCacheSlot, numKeep int) error {
 		return fmt.Errorf("unable to shift context - keep exceeds context (keep: %v context: %v)", numKeep, c.numCtx)
 	}
 
-	discard := c.ShiftDiscard(len(slot.Inputs), numKeep)
+	inputLen := len(slot.Inputs)
+	discard := c.ShiftDiscard(inputLen, numKeep)
 
 	if discard <= 0 {
 		return nil
@@ -231,16 +240,42 @@ func (c *InputCache) ShiftCacheSlot(slot *InputCacheSlot, numKeep int) error {
 	slog.Debug("context limit hit - shifting", "id", slot.Id, "limit", c.numCtx, "input", len(slot.Inputs),
 		"keep", numKeep, "discard", discard)
 
-	// TODO (jessegross): KV cache removal can fail for certain types of models
-	if !c.lc.KvCacheSeqRm(slot.Id, numKeep, numKeep+discard) {
-		return fmt.Errorf("unable to remove old kv cache entries (id: %v, keep: %v discard: %v)", slot.Id, numKeep, discard)
-	}
-	c.lc.KvCacheSeqAdd(slot.Id, numKeep+discard, len(slot.Inputs), -discard)
+	var shiftFailed bool
 
-	for i := numKeep + discard; i < len(slot.Inputs); i++ {
+	if c.lc.KvCacheCanShift() {
+		// For models that support shifting, attempt to shift the KV cache
+		if !c.lc.KvCacheSeqRm(slot.Id, numKeep, numKeep+discard) {
+			shiftFailed = true
+			slog.Debug("kv cache removal not supported, clearing cache and returning inputs for reprocessing", "id", slot.Id)
+		} else {
+			c.lc.KvCacheSeqAdd(slot.Id, numKeep+discard, inputLen, -discard)
+		}
+	} else {
+		// For models that don't support shifting
+		shiftFailed = true
+		slog.Debug("kv cache cannot shift, clearing cache and returning inputs for reprocessing", "id", slot.Id)
+	}
+
+	if shiftFailed {
+		// Create new input slice with preserved tokens (numKeep + remaining tokens after discard)
+		newInputs := make([]input, numKeep+inputLen-(numKeep+discard))
+		copy(newInputs[:numKeep], slot.Inputs[:numKeep])
+		copy(newInputs[numKeep:], slot.Inputs[numKeep+discard:])
+
+		// Clear the entire KV cache
+		_ = c.lc.KvCacheSeqRm(slot.Id, 0, -1)
+		// Reset the slot inputs since we've cleared the cache
+		slot.Inputs = []input{}
+
+		// Return error with inputs that need to be reprocessed
+		return &ErrReprocessInputs{Inputs: newInputs}
+	}
+
+	// Standard shift succeeded - update input array
+	for i := numKeep + discard; i < inputLen; i++ {
 		slot.Inputs[i-discard] = slot.Inputs[i]
 	}
-	slot.Inputs = slot.Inputs[:len(slot.Inputs)-discard]
+	slot.Inputs = slot.Inputs[:inputLen-discard]
 
 	return nil
 }

runner/llamarunner/runner.go

@@ -389,7 +389,15 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 				if len(seq.pendingInputs) == 0 {
 					err := s.cache.ShiftCacheSlot(seq.cache, seq.numKeep)
 					if err != nil {
-						return err
+						var reprocess *ErrReprocessInputs
+						if errors.As(err, &reprocess) {
+							// Prepend these inputs to the sequence's inputs queue for reprocessing
+							seq.inputs = append(reprocess.Inputs, seq.inputs...)
+							// Continue processing as normal
+							continue
+						} else {
+							return err
+						}
 					}
 				} else {
 					break
@@ -599,7 +607,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 		if errors.Is(err, context.Canceled) {
 			slog.Info("aborting completion request due to client closing the connection")
 		} else {
-			slog.Error("Failed to acquire semaphore", "error", err)
+			http.Error(w, fmt.Sprintf("Failed to acquire semaphore: %v", err), http.StatusInternalServerError)
 		}
 		return
 	}
@@ -611,6 +619,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 			seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs, true)
 			if err != nil {
 				s.mu.Unlock()
+				s.seqsSem.Release(1)
 				http.Error(w, fmt.Sprintf("Failed to load cache: %v", err), http.StatusInternalServerError)
 				return
 			}
@@ -626,6 +635,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 	s.mu.Unlock()
 
 	if !found {
+		s.seqsSem.Release(1)
 		http.Error(w, "could not find an available sequence", http.StatusInternalServerError)
 		return
 	}
@@ -691,7 +701,7 @@ func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
 		if errors.Is(err, context.Canceled) {
 			slog.Info("aborting embeddings request due to client closing the connection")
 		} else {
-			slog.Error("Failed to acquire semaphore", "error", err)
+			http.Error(w, fmt.Sprintf("Failed to acquire semaphore: %v", err), http.StatusInternalServerError)
 		}
 		return
 	}
@@ -703,6 +713,7 @@ func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
 			seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs, false)
 			if err != nil {
 				s.mu.Unlock()
+				s.seqsSem.Release(1)
 				http.Error(w, fmt.Sprintf("Failed to load cache: %v", err), http.StatusInternalServerError)
 				return
 			}
@@ -715,6 +726,7 @@ func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
 	s.mu.Unlock()
 
 	if !found {
+		s.seqsSem.Release(1)
 		http.Error(w, "could not find an available sequence", http.StatusInternalServerError)
 		return
 	}

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)
@@ -239,6 +245,14 @@ func (c *InputCache) ShiftDiscard(inputLen int32, numKeep int32) int32 {
 	return discard
 }
 
+type ErrReprocessInputs struct {
+	Inputs []input.Input
+}
+
+func (e *ErrReprocessInputs) Error() string {
+	return fmt.Sprintf("kv cache shift not supported, inputs need reprocessing (input count: %v)", len(e.Inputs))
+}
+
 // Frees up space in the KV cache by deleting the oldest half of history and shifting
 // the newest half into that space (saving numKeep inputs at the beginning).
 //
@@ -258,11 +272,23 @@ func (c *InputCache) ShiftCacheSlot(slot *InputCacheSlot, numKeep int32) error {
 	slog.Debug("context limit hit - shifting", "id", slot.Id, "limit", c.numCtx, "input", len(slot.Inputs),
 		"keep", numKeep, "discard", discard)
 
-	// TODO (jessegross): KV cache removal can fail for certain types of models
 	if c.cache != nil {
 		err := c.cache.Remove(slot.Id, numKeep, numKeep+discard)
 		if err != nil {
-			return fmt.Errorf("unable to remove old kv cache entries (id: %v, keep: %v discard: %v): %w", slot.Id, numKeep, discard, err)
+			slog.Debug("kv cache removal unsupported, clearing cache and returning inputs for reprocessing",
+				"id", slot.Id, "error", err)
+
+			// Create new input slice with preserved tokens (numKeep + remaining tokens after discard)
+			newInputs := make([]input.Input, numKeep+inputLen-(numKeep+discard))
+			copy(newInputs[:numKeep], slot.Inputs[:numKeep])
+			copy(newInputs[numKeep:], slot.Inputs[numKeep+discard:])
+
+			// Reset the cache
+			_ = c.cache.Remove(slot.Id, 0, -1)
+			slot.Inputs = []input.Input{}
+
+			// Return error with inputs that need to be reprocessed
+			return &ErrReprocessInputs{Inputs: newInputs}
 		}
 	}
 

runner/ollamarunner/cache_test.go

@@ -1,10 +1,13 @@
 package ollamarunner
 
 import (
+	"errors"
+	"fmt"
 	"image"
 	"testing"
 	"time"
 
+	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model/input"
 )
 
@@ -425,3 +428,92 @@ func TestLoadCacheSlot(t *testing.T) {
 		})
 	}
 }
+
+// Mock implementation of the Cache interface
+type mockCache struct {
+	shouldFail bool
+}
+
+// Implement only the methods needed for the test
+func (m *mockCache) Remove(seq int, beginIndex, endIndex int32) error {
+	if m.shouldFail {
+		return fmt.Errorf("mock cache removal error")
+	}
+	return nil
+}
+
+// Stub implementations for other interface methods
+func (m *mockCache) SetLayer(layer int)                                                            {}
+func (m *mockCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor)                          { return nil, nil, nil }
+func (m *mockCache) Put(ctx ml.Context, key, value ml.Tensor)                                      {}
+func (m *mockCache) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity, maxBatch int) {}
+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 {
+		name          string
+		numCtx        int32
+		inputs        []input.Input
+		numKeep       int32
+		cacheErr      bool
+		wantErr       any
+		wantInputsLen int
+	}{
+		{
+			name:          "Normal shift",
+			numCtx:        10,
+			inputs:        []input.Input{{Token: 1}, {Token: 2}, {Token: 3}, {Token: 4}, {Token: 5}, {Token: 6}, {Token: 7}, {Token: 8}, {Token: 9}, {Token: 10}},
+			numKeep:       2,
+			cacheErr:      false, // No error
+			wantErr:       nil,
+			wantInputsLen: 6, // After discarding 4 tokens
+		},
+		{
+			name:          "Cache removal fails",
+			numCtx:        10,
+			inputs:        []input.Input{{Token: 1}, {Token: 2}, {Token: 3}, {Token: 4}, {Token: 5}, {Token: 6}, {Token: 7}, {Token: 8}, {Token: 9}, {Token: 10}},
+			numKeep:       2,
+			cacheErr:      true,
+			wantErr:       &ErrReprocessInputs{},
+			wantInputsLen: 0, // Original inputs should be cleared
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			mock := &mockCache{shouldFail: tt.cacheErr}
+			c := InputCache{
+				numCtx: tt.numCtx,
+				cache:  mock,
+			}
+			slot := &InputCacheSlot{
+				Id:     123,
+				Inputs: make([]input.Input, len(tt.inputs)),
+			}
+			copy(slot.Inputs, tt.inputs)
+
+			err := c.ShiftCacheSlot(slot, tt.numKeep)
+
+			if tt.wantErr != nil {
+				if err == nil {
+					t.Errorf("Expected error but got nil")
+					return
+				}
+
+				if !errors.As(err, &tt.wantErr) {
+					t.Errorf("Expected error of type %T but got %T: %v", tt.wantErr, err, err)
+				}
+			} else if err != nil {
+				t.Errorf("Unexpected error: %v", err)
+			}
+
+			if len(slot.Inputs) != tt.wantInputsLen {
+				t.Errorf("Slot inputs length after operation: got %v, want %v", len(slot.Inputs), tt.wantInputsLen)
+			}
+		})
+	}
+}

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
@@ -267,6 +292,9 @@ type Server struct {
 	// KV cache
 	cache *InputCache
 
+	// next sequence for prompt processing to avoid starvation
+	nextSeq int
+
 	// multimodalHash generates hashes for comparing equality
 	// of non-text data
 	multimodalHash maphash.Hash
@@ -351,14 +379,19 @@ func (s *Server) processBatch() error {
 	var batchInputs []int32
 	var batch input.Batch
 
-	for i, seq := range s.seqs {
+	resumeSeq := -1
+	seqIdx := s.nextSeq - 1
+	for range s.seqs {
+		seqIdx = (seqIdx + 1) % len(s.seqs)
+		seq := s.seqs[seqIdx]
+
 		if seq == nil {
 			continue
 		}
 
 		// if past the num predict limit
 		if seq.numPredict > 0 && seq.numPredicted >= seq.numPredict {
-			s.removeSequence(i, "limit")
+			s.removeSequence(seqIdx, "limit")
 			continue
 		}
 
@@ -369,16 +402,23 @@ func (s *Server) processBatch() error {
 
 		batchSize := s.batchSize
 
-		for j, inp := range seq.inputs {
+		for i, inp := range seq.inputs {
 			// If we are required to put following inputs into a single batch then extend the
 			// batch size. Since we are only extending the size the minimum amount possible, this
-			// will cause a break if we have pending inputs.
+			// will cause a break if we have existing inputs.
 			minBatch := 1 + inp.SameBatch
 			if minBatch > batchSize {
 				batchSize = minBatch
 			}
 
-			if len(seq.pendingInputs)+minBatch > batchSize {
+			// Stop if the required batch would put us over the total batch size (including tokens
+			// added by other sequences). If we haven't been able to add anything yet then pick up
+			// here again for the next batch to avoid starvation, though we can opportunistically
+			// check if other sequences can still squeeze something in.
+			if len(batchInputs)+minBatch > batchSize {
+				if len(seq.pendingInputs) == 0 && resumeSeq == -1 {
+					resumeSeq = seqIdx
+				}
 				break
 			}
 
@@ -392,7 +432,15 @@ func (s *Server) processBatch() error {
 
 				err := s.cache.ShiftCacheSlot(seq.cache, seq.numKeep)
 				if err != nil {
-					return err
+					var reprocess *ErrReprocessInputs
+					if errors.As(err, &reprocess) {
+						// Prepend these inputs to the sequence's inputs queue for reprocessing
+						seq.inputs = append(reprocess.Inputs, seq.inputs...)
+						// Skip this sequence but continue processing the rest
+						continue
+					} else {
+						return err
+					}
 				}
 			}
 
@@ -405,7 +453,7 @@ func (s *Server) processBatch() error {
 			batch.Sequences = append(batch.Sequences, seq.cache.Id)
 
 			seq.iBatch = len(batch.Outputs)
-			if j+1 == len(seq.inputs) {
+			if i+1 == len(seq.inputs) {
 				batch.Outputs = append(batch.Outputs, int32(len(batchInputs)-1))
 			}
 			seq.pendingInputs = append(seq.pendingInputs, inp)
@@ -414,6 +462,12 @@ func (s *Server) processBatch() error {
 		seq.inputs = seq.inputs[len(seq.pendingInputs):]
 	}
 
+	if resumeSeq != -1 {
+		s.nextSeq = resumeSeq
+	} else {
+		s.nextSeq = seqIdx + 1
+	}
+
 	if len(batchInputs) == 0 {
 		return nil
 	}
@@ -588,7 +642,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 		if errors.Is(err, context.Canceled) {
 			slog.Info("aborting completion request due to client closing the connection")
 		} else {
-			slog.Error("Failed to acquire semaphore", "error", err)
+			http.Error(w, fmt.Sprintf("Failed to acquire semaphore: %v", err), http.StatusInternalServerError)
 		}
 		return
 	}
@@ -600,6 +654,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 			seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs)
 			if err != nil {
 				s.mu.Unlock()
+				s.seqsSem.Release(1)
 				http.Error(w, fmt.Sprintf("Failed to load cache: %v", err), http.StatusInternalServerError)
 				return
 			}
@@ -613,6 +668,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 	s.mu.Unlock()
 
 	if !found {
+		s.seqsSem.Release(1)
 		http.Error(w, "could not find an available sequence", http.StatusInternalServerError)
 		return
 	}

server/download.go

@@ -29,8 +29,9 @@ import (
 const maxRetries = 6
 
 var (
-	errMaxRetriesExceeded = errors.New("max retries exceeded")
-	errPartStalled        = errors.New("part stalled")
+	errMaxRetriesExceeded   = errors.New("max retries exceeded")
+	errPartStalled          = errors.New("part stalled")
+	errMaxRedirectsExceeded = errors.New("maximum redirects exceeded (10) for directURL")
 )
 
 var blobDownloadManager sync.Map
@@ -236,7 +237,7 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 
 			newOpts.CheckRedirect = func(req *http.Request, via []*http.Request) error {
 				if len(via) > 10 {
-					return errors.New("maximum redirects exceeded (10) for directURL")
+					return errMaxRedirectsExceeded
 				}
 
 				// if the hostname is the same, allow the redirect

server/images.go

@@ -35,14 +35,9 @@ var (
 	errCapabilityCompletion = errors.New("completion")
 	errCapabilityTools      = errors.New("tools")
 	errCapabilityInsert     = errors.New("insert")
-)
-
-type Capability string
-
-const (
-	CapabilityCompletion = Capability("completion")
-	CapabilityTools      = Capability("tools")
-	CapabilityInsert     = Capability("insert")
+	errCapabilityVision     = errors.New("vision")
+	errCapabilityEmbedding  = errors.New("embedding")
+	errInsecureProtocol     = errors.New("insecure protocol http")
 )
 
 type registryOptions struct {
@@ -65,52 +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...))
 	}
 
@@ -479,7 +505,7 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	fn(api.ProgressResponse{Status: "retrieving manifest"})
 
 	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
-		return errors.New("insecure protocol http")
+		return errInsecureProtocol
 	}
 
 	manifest, _, err := GetManifest(mp)
@@ -543,7 +569,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	}
 
 	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
-		return errors.New("insecure protocol http")
+		return errInsecureProtocol
 	}
 
 	fn(api.ProgressResponse{Status: "pulling manifest"})

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/internal/client/ollama/registry.go

@@ -421,14 +421,6 @@ func (r *Registry) Push(ctx context.Context, name string, p *PushParams) error {
 	return err
 }
 
-func canRetry(err error) bool {
-	var re *Error
-	if !errors.As(err, &re) {
-		return false
-	}
-	return re.Status >= 500
-}
-
 // trackingReader is an io.Reader that tracks the number of bytes read and
 // calls the update function with the layer, the number of bytes read.
 //
@@ -514,13 +506,40 @@ func (r *Registry) Pull(ctx context.Context, name string) error {
 				break
 			}
 
+			cacheKey := fmt.Sprintf(
+				"v1 pull chunksum %s %s %d-%d",
+				l.Digest,
+				cs.Digest,
+				cs.Chunk.Start,
+				cs.Chunk.End,
+			)
+			cacheKeyDigest := blob.DigestFromBytes(cacheKey)
+			_, err := c.Get(cacheKeyDigest)
+			if err == nil {
+				received.Add(cs.Chunk.Size())
+				t.update(l, cs.Chunk.Size(), ErrCached)
+				continue
+			}
+
 			wg.Add(1)
 			g.Go(func() (err error) {
 				defer func() {
 					if err == nil {
+						// Ignore cache key write errors for now. We've already
+						// reported to trace that the chunk is complete.
+						//
+						// Ideally, we should only report completion to trace
+						// after successful cache commit. This current approach
+						// works but could trigger unnecessary redownloads if
+						// the checkpoint key is missing on next pull.
+						//
+						// Not incorrect, just suboptimal - fix this in a
+						// future update.
+						_ = blob.PutBytes(c, cacheKeyDigest, cacheKey)
+
 						received.Add(cs.Chunk.Size())
 					} else {
-						err = fmt.Errorf("error downloading %s: %w", cs.Digest.Short(), err)
+						t.update(l, 0, err)
 					}
 					wg.Done()
 				}()
@@ -563,7 +582,7 @@ func (r *Registry) Pull(ctx context.Context, name string) error {
 		return err
 	}
 	if received.Load() != expected {
-		return fmt.Errorf("%w: received %d/%d", ErrIncomplete, received.Load(), expected)
+		return fmt.Errorf("%w: received %d/%d bytes", ErrIncomplete, received.Load(), expected)
 	}
 
 	md := blob.DigestFromBytes(m.Data)
@@ -608,6 +627,30 @@ func (m *Manifest) Layer(d blob.Digest) *Layer {
 	return nil
 }
 
+func (m *Manifest) All() iter.Seq[*Layer] {
+	return func(yield func(*Layer) bool) {
+		if !yield(m.Config) {
+			return
+		}
+		for _, l := range m.Layers {
+			if !yield(l) {
+				return
+			}
+		}
+	}
+}
+
+func (m *Manifest) Size() int64 {
+	var size int64
+	if m.Config != nil {
+		size += m.Config.Size
+	}
+	for _, l := range m.Layers {
+		size += l.Size
+	}
+	return size
+}
+
 // MarshalJSON implements json.Marshaler.
 //
 // NOTE: It adds an empty config object to the manifest, which is required by
@@ -750,20 +793,32 @@ func (r *Registry) chunksums(ctx context.Context, name string, l *Layer) iter.Se
 			return
 		}
 
-		// A chunksums response is a sequence of chunksums in a
-		// simple, easy to parse line-oriented format.
+		// The response is a sequence of chunksums.
 		//
-		// Example:
+		// Chunksums are chunks of a larger blob that can be
+		// downloaded and verified independently.
 		//
-		//     >> GET /v2/<namespace>/<model>/chunksums/<digest>
+		// The chunksums endpoint is a GET request that returns a
+		// sequence of chunksums in the following format:
 		//
-		//     << HTTP/1.1 200 OK
-		//     << Content-Location: <blobURL>
-		//     <<
-		//     << <digest> <start>-<end>
-		//     << ...
+		//     > GET /v2/<namespace>/<model>/chunksums/<digest>
 		//
-		// The blobURL is the URL to download the chunks from.
+		//     < HTTP/1.1 200 OK
+		//     < Content-Location: <blobURL>
+		//     <
+		//     < <digest> <start>-<end>
+		//     < ...
+		//
+		// The <blobURL> is the URL to download the chunks from and
+		// each <digest> is the digest of the chunk, and <start>-<end>
+		// is the range the chunk in the blob.
+		//
+		// Ranges may be used directly in Range headers like
+		// "bytes=<start>-<end>".
+		//
+		// The chunksums returned are guaranteed to be contiguous and
+		// include all bytes of the layer. If the stream is cut short,
+		// clients should retry.
 
 		chunksumsURL := fmt.Sprintf("%s://%s/v2/%s/%s/chunksums/%s",
 			scheme,

server/internal/client/ollama/registry_test.go

@@ -9,17 +9,14 @@ import (
 	"fmt"
 	"io"
 	"io/fs"
-	"math/rand/v2"
+	"net"
 	"net/http"
 	"net/http/httptest"
 	"os"
-	"path"
 	"reflect"
-	"slices"
 	"strings"
-	"sync"
+	"sync/atomic"
 	"testing"
-	"time"
 
 	"github.com/ollama/ollama/server/internal/cache/blob"
 	"github.com/ollama/ollama/server/internal/testutil"
@@ -338,15 +335,8 @@ func TestPushCommitRoundtripError(t *testing.T) {
 	}
 }
 
-func checkNotExist(t *testing.T, err error) {
-	t.Helper()
-	if !errors.Is(err, fs.ErrNotExist) {
-		t.Fatalf("err = %v; want fs.ErrNotExist", err)
-	}
-}
-
 func TestRegistryPullInvalidName(t *testing.T) {
-	rc, _ := newClient(t, nil)
+	rc, _ := newRegistryClient(t, nil)
 	err := rc.Pull(t.Context(), "://")
 	if !errors.Is(err, ErrNameInvalid) {
 		t.Errorf("err = %v; want %v", err, ErrNameInvalid)
@@ -362,197 +352,16 @@ func TestRegistryPullInvalidManifest(t *testing.T) {
 	}
 
 	for _, resp := range cases {
-		rc, _ := newClient(t, func(w http.ResponseWriter, r *http.Request) {
+		rc, _ := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
 			io.WriteString(w, resp)
 		})
-		err := rc.Pull(t.Context(), "x")
+		err := rc.Pull(t.Context(), "http://example.com/a/b")
 		if !errors.Is(err, ErrManifestInvalid) {
 			t.Errorf("err = %v; want invalid manifest", err)
 		}
 	}
 }
 
-func TestRegistryPullNotCached(t *testing.T) {
-	check := testutil.Checker(t)
-
-	var c *blob.DiskCache
-	var rc *Registry
-
-	d := blob.DigestFromBytes("some data")
-	rc, c = newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		if strings.Contains(r.URL.Path, "/blobs/") {
-			io.WriteString(w, "some data")
-			return
-		}
-		fmt.Fprintf(w, `{"layers":[{"digest":%q,"size":9}]}`, d)
-	})
-
-	// Confirm that the layer does not exist locally
-	_, err := rc.ResolveLocal("model")
-	checkNotExist(t, err)
-
-	_, err = c.Get(d)
-	checkNotExist(t, err)
-
-	err = rc.Pull(t.Context(), "model")
-	check(err)
-
-	mw, err := rc.Resolve(t.Context(), "model")
-	check(err)
-	mg, err := rc.ResolveLocal("model")
-	check(err)
-	if !reflect.DeepEqual(mw, mg) {
-		t.Errorf("mw = %v; mg = %v", mw, mg)
-	}
-
-	// Confirm successful download
-	info, err := c.Get(d)
-	check(err)
-	if info.Digest != d {
-		t.Errorf("info.Digest = %v; want %v", info.Digest, d)
-	}
-	if info.Size != 9 {
-		t.Errorf("info.Size = %v; want %v", info.Size, 9)
-	}
-
-	data, err := os.ReadFile(c.GetFile(d))
-	check(err)
-	if string(data) != "some data" {
-		t.Errorf("data = %q; want %q", data, "exists")
-	}
-}
-
-func TestRegistryPullCached(t *testing.T) {
-	cached := blob.DigestFromBytes("exists")
-	rc, _ := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		if strings.Contains(r.URL.Path, "/blobs/") {
-			w.WriteHeader(499) // should not be called
-			return
-		}
-		if strings.Contains(r.URL.Path, "/manifests/") {
-			fmt.Fprintf(w, `{"layers":[{"digest":%q,"size":6}]}`, cached)
-		}
-	})
-
-	var errs []error
-	var reads []int64
-	ctx := WithTrace(t.Context(), &Trace{
-		Update: func(d *Layer, n int64, err error) {
-			t.Logf("update %v %d %v", d, n, err)
-			reads = append(reads, n)
-			errs = append(errs, err)
-		},
-	})
-
-	ctx, cancel := context.WithTimeout(ctx, 3*time.Second)
-	defer cancel()
-
-	err := rc.Pull(ctx, "single")
-	testutil.Check(t, err)
-
-	want := []int64{0, 6}
-	if !errors.Is(errors.Join(errs...), ErrCached) {
-		t.Errorf("errs = %v; want %v", errs, ErrCached)
-	}
-	if !slices.Equal(reads, want) {
-		t.Errorf("pairs = %v; want %v", reads, want)
-	}
-}
-
-func TestRegistryPullManifestNotFound(t *testing.T) {
-	rc, _ := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		w.WriteHeader(http.StatusNotFound)
-	})
-	err := rc.Pull(t.Context(), "notfound")
-	checkErrCode(t, err, 404, "")
-}
-
-func TestRegistryPullResolveRemoteError(t *testing.T) {
-	rc, _ := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		w.WriteHeader(http.StatusInternalServerError)
-		io.WriteString(w, `{"errors":[{"code":"an_error"}]}`)
-	})
-	err := rc.Pull(t.Context(), "single")
-	checkErrCode(t, err, 500, "an_error")
-}
-
-func TestRegistryPullResolveRoundtripError(t *testing.T) {
-	rc, _ := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		if strings.Contains(r.URL.Path, "/manifests/") {
-			w.WriteHeader(499) // force RoundTrip error
-			return
-		}
-	})
-	err := rc.Pull(t.Context(), "single")
-	if !errors.Is(err, errRoundTrip) {
-		t.Errorf("err = %v; want %v", err, errRoundTrip)
-	}
-}
-
-// TestRegistryPullMixedCachedNotCached tests that cached layers do not
-// interfere with pulling layers that are not cached
-func TestRegistryPullMixedCachedNotCached(t *testing.T) {
-	x := blob.DigestFromBytes("xxxxxx")
-	e := blob.DigestFromBytes("exists")
-	y := blob.DigestFromBytes("yyyyyy")
-
-	for i := range 10 {
-		t.Logf("iteration %d", i)
-
-		digests := []blob.Digest{x, e, y}
-
-		rand.Shuffle(len(digests), func(i, j int) {
-			digests[i], digests[j] = digests[j], digests[i]
-		})
-
-		manifest := fmt.Sprintf(`{
-			"layers": [
-				{"digest":"%s","size":6},
-				{"digest":"%s","size":6},
-				{"digest":"%s","size":6}
-			]
-		}`, digests[0], digests[1], digests[2])
-
-		rc, c := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-			switch path.Base(r.URL.Path) {
-			case "latest":
-				io.WriteString(w, manifest)
-			case x.String():
-				io.WriteString(w, "xxxxxx")
-			case e.String():
-				io.WriteString(w, "exists")
-			case y.String():
-				io.WriteString(w, "yyyyyy")
-			default:
-				panic(fmt.Sprintf("unexpected request: %v", r))
-			}
-		})
-
-		ctx := WithTrace(t.Context(), &Trace{
-			Update: func(l *Layer, n int64, err error) {
-				t.Logf("update %v %d %v", l, n, err)
-			},
-		})
-
-		// Check that we pull all layers that we can.
-
-		err := rc.Pull(ctx, "mixed")
-		if err != nil {
-			t.Fatal(err)
-		}
-
-		for _, d := range digests {
-			info, err := c.Get(d)
-			if err != nil {
-				t.Fatalf("Get(%v): %v", d, err)
-			}
-			if info.Size != 6 {
-				t.Errorf("info.Size = %v; want %v", info.Size, 6)
-			}
-		}
-	}
-}
-
 func TestRegistryResolveByDigest(t *testing.T) {
 	check := testutil.Checker(t)
 
@@ -590,26 +399,6 @@ func TestInsecureSkipVerify(t *testing.T) {
 	testutil.Check(t, err)
 }
 
-func TestCanRetry(t *testing.T) {
-	cases := []struct {
-		err  error
-		want bool
-	}{
-		{nil, false},
-		{errors.New("x"), false},
-		{ErrCached, false},
-		{ErrManifestInvalid, false},
-		{ErrNameInvalid, false},
-		{&Error{Status: 100}, false},
-		{&Error{Status: 500}, true},
-	}
-	for _, tt := range cases {
-		if got := canRetry(tt.err); got != tt.want {
-			t.Errorf("CanRetry(%v) = %v; want %v", tt.err, got, tt.want)
-		}
-	}
-}
-
 func TestErrorUnmarshal(t *testing.T) {
 	cases := []struct {
 		name    string
@@ -761,17 +550,23 @@ func TestParseNameExtended(t *testing.T) {
 
 func TestUnlink(t *testing.T) {
 	t.Run("found by name", func(t *testing.T) {
-		rc, _ := newClient(t, nil)
+		check := testutil.Checker(t)
+
+		rc, _ := newRegistryClient(t, nil)
+		// make a blob and link it
+		d := blob.DigestFromBytes("{}")
+		err := blob.PutBytes(rc.Cache, d, "{}")
+		check(err)
+		err = rc.Cache.Link("registry.ollama.ai/library/single:latest", d)
+		check(err)
 
 		// confirm linked
-		_, err := rc.ResolveLocal("single")
-		if err != nil {
-			t.Errorf("unexpected error: %v", err)
-		}
+		_, err = rc.ResolveLocal("single")
+		check(err)
 
 		// unlink
 		_, err = rc.Unlink("single")
-		testutil.Check(t, err)
+		check(err)
 
 		// confirm unlinked
 		_, err = rc.ResolveLocal("single")
@@ -780,7 +575,7 @@ func TestUnlink(t *testing.T) {
 		}
 	})
 	t.Run("not found by name", func(t *testing.T) {
-		rc, _ := newClient(t, nil)
+		rc, _ := newRegistryClient(t, nil)
 		ok, err := rc.Unlink("manifestNotFound")
 		if err != nil {
 			t.Fatal(err)
@@ -791,78 +586,368 @@ func TestUnlink(t *testing.T) {
 	})
 }
 
-func TestPullChunksums(t *testing.T) {
-	check := testutil.Checker(t)
+// Many tests from here out, in this file are based on a single blob, "abc",
+// with the checksum of its sha256 hash. The checksum is:
+//
+//	"abc" -> sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad
+//
+// Using the literal value instead of a constant with fmt.Xprintf calls proved
+// to be the most readable and maintainable approach. The sum is consistently
+// used in the tests and unique so searches do not yield false positives.
 
-	content := "hello"
-	var chunksums string
-	contentDigest := func() blob.Digest {
-		return blob.DigestFromBytes(content)
+func checkRequest(t *testing.T, req *http.Request, method, path string) {
+	t.Helper()
+	if got := req.URL.Path; got != path {
+		t.Errorf("URL = %q, want %q", got, path)
 	}
-	rc, c := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		switch {
-		case strings.Contains(r.URL.Path, "/manifests/latest"):
-			fmt.Fprintf(w, `{"layers":[{"digest":%q,"size":%d}]}`, contentDigest(), len(content))
-		case strings.HasSuffix(r.URL.Path, "/chunksums/"+contentDigest().String()):
-			loc := fmt.Sprintf("http://blob.store/v2/library/test/blobs/%s", contentDigest())
-			w.Header().Set("Content-Location", loc)
-			io.WriteString(w, chunksums)
-		case strings.Contains(r.URL.Path, "/blobs/"+contentDigest().String()):
-			http.ServeContent(w, r, contentDigest().String(), time.Time{}, strings.NewReader(content))
-		default:
-			t.Errorf("unexpected request: %v", r)
-			http.NotFound(w, r)
-		}
-	})
+	if req.Method != method {
+		t.Errorf("Method = %q, want %q", req.Method, method)
+	}
+}
 
-	rc.MaxStreams = 1        // prevent concurrent chunk downloads
-	rc.ChunkingThreshold = 1 // for all blobs to be chunked
+func newRegistryClient(t *testing.T, h http.HandlerFunc) (*Registry, context.Context) {
+	s := httptest.NewServer(h)
+	t.Cleanup(s.Close)
+	cache, err := blob.Open(t.TempDir())
+	if err != nil {
+		t.Fatal(err)
+	}
 
-	var mu sync.Mutex
-	var reads []int64
 	ctx := WithTrace(t.Context(), &Trace{
 		Update: func(l *Layer, n int64, err error) {
-			t.Logf("Update: %v %d %v", l, n, err)
-			mu.Lock()
-			reads = append(reads, n)
-			mu.Unlock()
+			t.Log("trace:", l.Digest.Short(), n, err)
 		},
 	})
 
-	chunksums = fmt.Sprintf("%s 0-2\n%s 3-4\n",
-		blob.DigestFromBytes("hel"),
-		blob.DigestFromBytes("lo"),
-	)
-	err := rc.Pull(ctx, "test")
-	check(err)
-	wantReads := []int64{
-		0, // initial signaling of layer pull starting
-		3, // first chunk read
-		2, // second chunk read
-	}
-	if !slices.Equal(reads, wantReads) {
-		t.Errorf("reads = %v; want %v", reads, wantReads)
+	rc := &Registry{
+		Cache: cache,
+		HTTPClient: &http.Client{Transport: &http.Transport{
+			Dial: func(network, addr string) (net.Conn, error) {
+				return net.Dial(network, s.Listener.Addr().String())
+			},
+		}},
 	}
+	return rc, ctx
+}
 
-	mw, err := rc.Resolve(t.Context(), "test")
-	check(err)
-	mg, err := rc.ResolveLocal("test")
-	check(err)
-	if !reflect.DeepEqual(mw, mg) {
-		t.Errorf("mw = %v; mg = %v", mw, mg)
-	}
-	for i := range mg.Layers {
-		_, err = c.Get(mg.Layers[i].Digest)
-		if err != nil {
-			t.Errorf("Get(%v): %v", mg.Layers[i].Digest, err)
+func TestPullChunked(t *testing.T) {
+	var steps atomic.Int64
+	c, ctx := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		switch steps.Add(1) {
+		case 1:
+			checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+			io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+		case 2:
+			checkRequest(t, r, "GET", "/v2/library/abc/chunksums/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			w.Header().Set("Content-Location", "http://blob.store/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			fmt.Fprintf(w, "%s 0-1\n", blob.DigestFromBytes("ab"))
+			fmt.Fprintf(w, "%s 2-2\n", blob.DigestFromBytes("c"))
+		case 3, 4:
+			checkRequest(t, r, "GET", "/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			switch rng := r.Header.Get("Range"); rng {
+			case "bytes=0-1":
+				io.WriteString(w, "ab")
+			case "bytes=2-2":
+				t.Logf("writing c")
+				io.WriteString(w, "c")
+			default:
+				t.Errorf("unexpected range %q", rng)
+			}
+		default:
+			t.Errorf("unexpected steps %d: %v", steps.Load(), r)
+			http.Error(w, "unexpected steps", http.StatusInternalServerError)
 		}
-	}
+	})
 
-	// missing chunks
-	content = "llama"
-	chunksums = fmt.Sprintf("%s 0-1\n", blob.DigestFromBytes("ll"))
-	err = rc.Pull(ctx, "missingchunks")
-	if err == nil {
-		t.Error("expected error because of missing chunks")
+	c.ChunkingThreshold = 1 // force chunking
+
+	err := c.Pull(ctx, "http://o.com/library/abc")
+	testutil.Check(t, err)
+
+	_, err = c.Cache.Resolve("o.com/library/abc:latest")
+	testutil.Check(t, err)
+
+	if g := steps.Load(); g != 4 {
+		t.Fatalf("got %d steps, want 4", g)
+	}
+}
+
+func TestPullCached(t *testing.T) {
+	c, ctx := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+		io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+	})
+
+	check := testutil.Checker(t)
+
+	// Premeptively cache the blob
+	d, err := blob.ParseDigest("sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+	check(err)
+	err = blob.PutBytes(c.Cache, d, []byte("abc"))
+	check(err)
+
+	// Pull only the manifest, which should be enough to resolve the cached blob
+	err = c.Pull(ctx, "http://o.com/library/abc")
+	check(err)
+}
+
+func TestPullManifestError(t *testing.T) {
+	c, ctx := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+		w.WriteHeader(http.StatusNotFound)
+		io.WriteString(w, `{"errors":[{"code":"MANIFEST_UNKNOWN"}]}`)
+	})
+
+	err := c.Pull(ctx, "http://o.com/library/abc")
+	if err == nil {
+		t.Fatalf("expected error")
+	}
+	var got *Error
+	if !errors.Is(err, ErrModelNotFound) {
+		t.Fatalf("err = %v, want %v", got, ErrModelNotFound)
+	}
+}
+
+func TestPullLayerError(t *testing.T) {
+	c, ctx := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+		io.WriteString(w, `!`)
+	})
+
+	err := c.Pull(ctx, "http://o.com/library/abc")
+	if err == nil {
+		t.Fatalf("expected error")
+	}
+	var want *json.SyntaxError
+	if !errors.As(err, &want) {
+		t.Fatalf("err = %T, want %T", err, want)
+	}
+}
+
+func TestPullLayerChecksumError(t *testing.T) {
+	var step atomic.Int64
+	c, _ := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		switch step.Add(1) {
+		case 1:
+			checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+			io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+		case 2:
+			checkRequest(t, r, "GET", "/v2/library/abc/chunksums/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			w.Header().Set("Content-Location", "http://blob.store/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			fmt.Fprintf(w, "%s 0-1\n", blob.DigestFromBytes("ab"))
+			fmt.Fprintf(w, "%s 2-2\n", blob.DigestFromBytes("c"))
+		case 3:
+			w.WriteHeader(http.StatusNotFound)
+			io.WriteString(w, `{"errors":[{"code":"BLOB_UNKNOWN"}]}`)
+		case 4:
+			io.WriteString(w, "c")
+		default:
+			t.Errorf("unexpected steps %d: %v", step.Load(), r)
+			http.Error(w, "unexpected steps", http.StatusInternalServerError)
+		}
+	})
+
+	c.MaxStreams = 1
+	c.ChunkingThreshold = 1 // force chunking
+
+	var written atomic.Int64
+	ctx := WithTrace(t.Context(), &Trace{
+		Update: func(l *Layer, n int64, err error) {
+			t.Log("trace:", l.Digest.Short(), n, err)
+			written.Add(n)
+		},
+	})
+
+	err := c.Pull(ctx, "http://o.com/library/abc")
+	var got *Error
+	if !errors.As(err, &got) || got.Code != "BLOB_UNKNOWN" {
+		t.Fatalf("err = %v, want %v", err, got)
+	}
+
+	if g := written.Load(); g != 1 {
+		t.Fatalf("wrote %d bytes, want 1", g)
+	}
+}
+
+func TestPullChunksumStreamError(t *testing.T) {
+	var step atomic.Int64
+	c, ctx := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		switch step.Add(1) {
+		case 1:
+			checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+			io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+		case 2:
+			w.Header().Set("Content-Location", "http://blob.store/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+
+			// Write one valid chunksum and one invalid chunksum
+			fmt.Fprintf(w, "%s 0-1\n", blob.DigestFromBytes("ab")) // valid
+			fmt.Fprint(w, "sha256:!")                              // invalid
+		case 3:
+			io.WriteString(w, "ab")
+		default:
+			t.Errorf("unexpected steps %d: %v", step.Load(), r)
+			http.Error(w, "unexpected steps", http.StatusInternalServerError)
+		}
+	})
+
+	c.ChunkingThreshold = 1 // force chunking
+
+	got := c.Pull(ctx, "http://o.com/library/abc")
+	if !errors.Is(got, ErrIncomplete) {
+		t.Fatalf("err = %v, want %v", got, ErrIncomplete)
+	}
+}
+
+type flushAfterWriter struct {
+	w io.Writer
+}
+
+func (f *flushAfterWriter) Write(p []byte) (n int, err error) {
+	n, err = f.w.Write(p)
+	f.w.(http.Flusher).Flush() // panic if not a flusher
+	return
+}
+
+func TestPullChunksumStreaming(t *testing.T) {
+	csr, csw := io.Pipe()
+	defer csw.Close()
+
+	var step atomic.Int64
+	c, _ := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		switch step.Add(1) {
+		case 1:
+			checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+			io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+		case 2:
+			w.Header().Set("Content-Location", "http://blob.store/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			fw := &flushAfterWriter{w} // ensure client gets data as it arrives by aggressively flushing
+			_, err := io.Copy(fw, csr)
+			if err != nil {
+				t.Errorf("copy: %v", err)
+			}
+		case 3:
+			io.WriteString(w, "ab")
+		case 4:
+			io.WriteString(w, "c")
+		default:
+			t.Errorf("unexpected steps %d: %v", step.Load(), r)
+			http.Error(w, "unexpected steps", http.StatusInternalServerError)
+		}
+	})
+
+	c.ChunkingThreshold = 1 // force chunking
+
+	update := make(chan int64, 1)
+	ctx := WithTrace(t.Context(), &Trace{
+		Update: func(l *Layer, n int64, err error) {
+			t.Log("trace:", l.Digest.Short(), n, err)
+			if n > 0 {
+				update <- n
+			}
+		},
+	})
+
+	errc := make(chan error, 1)
+	go func() {
+		errc <- c.Pull(ctx, "http://o.com/library/abc")
+	}()
+
+	// Send first chunksum and ensure it kicks off work immediately
+	fmt.Fprintf(csw, "%s 0-1\n", blob.DigestFromBytes("ab"))
+	if g := <-update; g != 2 {
+		t.Fatalf("got %d, want 2", g)
+	}
+
+	// now send the second chunksum and ensure it kicks off work immediately
+	fmt.Fprintf(csw, "%s 2-2\n", blob.DigestFromBytes("c"))
+	if g := <-update; g != 1 {
+		t.Fatalf("got %d, want 1", g)
+	}
+	csw.Close()
+	testutil.Check(t, <-errc)
+}
+
+func TestPullChunksumsCached(t *testing.T) {
+	var step atomic.Int64
+	c, _ := newRegistryClient(t, func(w http.ResponseWriter, r *http.Request) {
+		switch step.Add(1) {
+		case 1:
+			checkRequest(t, r, "GET", "/v2/library/abc/manifests/latest")
+			io.WriteString(w, `{"layers":[{"size":3,"digest":"sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad"}]}`)
+		case 2:
+			w.Header().Set("Content-Location", "http://blob.store/v2/library/abc/blobs/sha256:ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad")
+			fmt.Fprintf(w, "%s 0-1\n", blob.DigestFromBytes("ab"))
+			fmt.Fprintf(w, "%s 2-2\n", blob.DigestFromBytes("c"))
+		case 3, 4:
+			switch rng := r.Header.Get("Range"); rng {
+			case "bytes=0-1":
+				io.WriteString(w, "ab")
+			case "bytes=2-2":
+				io.WriteString(w, "c")
+			default:
+				t.Errorf("unexpected range %q", rng)
+			}
+		default:
+			t.Errorf("unexpected steps %d: %v", step.Load(), r)
+			http.Error(w, "unexpected steps", http.StatusInternalServerError)
+		}
+	})
+
+	c.MaxStreams = 1        // force serial processing of chunksums
+	c.ChunkingThreshold = 1 // force chunking
+
+	ctx, cancel := context.WithCancel(t.Context())
+	defer cancel()
+
+	// Cancel the pull after the first chunksum is processed, but before
+	// the second chunksum is processed (which is waiting because
+	// MaxStreams=1). This should cause the second chunksum to error out
+	// leaving the blob incomplete.
+	ctx = WithTrace(ctx, &Trace{
+		Update: func(l *Layer, n int64, err error) {
+			if n > 0 {
+				cancel()
+			}
+		},
+	})
+	err := c.Pull(ctx, "http://o.com/library/abc")
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("err = %v, want %v", err, context.Canceled)
+	}
+
+	_, err = c.Cache.Resolve("o.com/library/abc:latest")
+	if !errors.Is(err, fs.ErrNotExist) {
+		t.Fatalf("err = %v, want nil", err)
+	}
+
+	// Reset state and pull again to ensure the blob chunks that should
+	// have been cached are, and the remaining chunk was downloaded, making
+	// the blob complete.
+	step.Store(0)
+	var written atomic.Int64
+	var cached atomic.Int64
+	ctx = WithTrace(t.Context(), &Trace{
+		Update: func(l *Layer, n int64, err error) {
+			t.Log("trace:", l.Digest.Short(), n, err)
+			if errors.Is(err, ErrCached) {
+				cached.Add(n)
+			}
+			written.Add(n)
+		},
+	})
+
+	check := testutil.Checker(t)
+
+	err = c.Pull(ctx, "http://o.com/library/abc")
+	check(err)
+
+	_, err = c.Cache.Resolve("o.com/library/abc:latest")
+	check(err)
+
+	if g := written.Load(); g != 3 {
+		t.Fatalf("wrote %d bytes, want 3", g)
+	}
+	if g := cached.Load(); g != 2 { // "ab" should have been cached
+		t.Fatalf("cached %d bytes, want 3", g)
 	}
 }

server/modelpath.go

@@ -31,9 +31,10 @@ const (
 
 var (
 	ErrInvalidImageFormat  = errors.New("invalid image format")
+	ErrInvalidDigestFormat = errors.New("invalid digest format")
 	ErrInvalidProtocol     = errors.New("invalid protocol scheme")
 	ErrInsecureProtocol    = errors.New("insecure protocol http")
-	ErrInvalidDigestFormat = errors.New("invalid digest format")
+	ErrModelPathInvalid    = errors.New("invalid model path")
 )
 
 func ParseModelPath(name string) ModelPath {
@@ -73,8 +74,6 @@ func ParseModelPath(name string) ModelPath {
 	return mp
 }
 
-var errModelPathInvalid = errors.New("invalid model path")
-
 func (mp ModelPath) GetNamespaceRepository() string {
 	return fmt.Sprintf("%s/%s", mp.Namespace, mp.Repository)
 }

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
@@ -777,7 +777,7 @@ func (s *Server) ShowHandler(c *gin.Context) {
 func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	name := model.ParseName(req.Model)
 	if !name.IsValid() {
-		return nil, errModelPathInvalid
+		return nil, ErrModelPathInvalid
 	}
 	name, err := getExistingName(name)
 	if err != nil {
@@ -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
@@ -711,7 +712,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuIn
 			req.opts.NumCtx = req.origNumCtx * p
 			if !envconfig.SchedSpread() {
 				for _, g := range sgl {
-					if ok, estimatedVRAM = llm.PredictServerFit([]discover.GpuInfo{g}, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+					if ok, estimatedVRAM = llm.PredictServerFit([]discover.GpuInfo{g}, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p); ok {
 						slog.Info("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "parallel", p, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
 						*numParallel = p
 						return []discover.GpuInfo{g}
@@ -727,7 +728,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuIn
 		// Now try all the GPUs
 		for _, p := range numParallelToTry {
 			req.opts.NumCtx = req.origNumCtx * p
-			if ok, estimatedVRAM = llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+			if ok, estimatedVRAM = llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p); ok {
 				slog.Info("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "parallel", p, "required", format.HumanBytes2(estimatedVRAM))
 				*numParallel = p
 				return sgl
@@ -750,7 +751,7 @@ func pickBestPartialFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.Gp
 	var bestEstimate uint64
 	var bestFit int
 	for i, gl := range byLibrary {
-		_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
+		_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, *numParallel)
 		if estimatedVRAM > bestEstimate {
 			bestEstimate = estimatedVRAM
 			bestFit = i
@@ -825,7 +826,7 @@ func (s *Scheduler) expireRunner(model *Model) {
 // If not, pick a runner to unload, else return nil and the request can be loaded
 func (s *Scheduler) maybeFindCPURunnerToUnload(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList) *runnerRef {
 	slog.Debug("evaluating if CPU model load will fit in available system memory")
-	estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts)
+	estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts, req.opts.NumCtx/req.origNumCtx)
 	if estimate.TotalSize <= gpus[0].FreeMemory {
 		slog.Debug("cpu inference mode, model fits in available system memory", "model", format.HumanBytes2(estimate.TotalSize), "available", format.HumanBytes2(gpus[0].FreeMemory))
 		return nil

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)
+}