wip?

Co-authored-by: Patrick Devine <patrick@infrahq.com>

.github/workflows/test.yaml

@@ -237,5 +237,5 @@ jobs:
       - uses: actions/checkout@v4
       - name: Verify patches apply cleanly and do not change files
         run: |
-          make -f Makefile.sync clean sync
-          git diff --compact-summary --exit-code
+          make -f Makefile.sync clean checkout apply-patches sync
+          git diff --compact-summary --exit-code

CMakeLists.txt

@@ -24,6 +24,7 @@ set(GGML_LLAMAFILE ON)
 set(GGML_CUDA_PEER_MAX_BATCH_SIZE 128)
 set(GGML_CUDA_GRAPHS ON)
 set(GGML_CUDA_FA ON)
+set(GGML_CUDA_COMPRESSION_MODE default)
 
 if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
     OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+"))
@@ -86,9 +87,9 @@ 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 +98,7 @@ 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

@@ -21,14 +21,16 @@
       "name": "CUDA 11",
       "inherits": [ "CUDA" ],
       "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50;52;53;60;61;70;75;80;86"
+        "CMAKE_CUDA_ARCHITECTURES": "50;52;53;60;61;70;75;80;86",
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets"
       }
     },
     {
       "name": "CUDA 12",
       "inherits": [ "CUDA" ],
       "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50;60;61;70;75;80;86;87;89;90;90a;120"
+        "CMAKE_CUDA_ARCHITECTURES": "50;60;61;70;75;80;86;87;89;90;90a;120",
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets"
       }
     },
     {
@@ -56,7 +58,7 @@
       "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-"
       }
     }
   ],

CONTRIBUTING.md

@@ -51,7 +51,7 @@ see if the change were accepted.
 
 The title should look like:
 
-   <package>: <short description>
+    <package>: <short description>
 
 The package is the most affected Go package. If the change does not affect Go
 code, then use the directory name instead. Changes to a single well-known

Dockerfile

@@ -104,8 +104,8 @@ COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
 FROM --platform=linux/arm64 scratch AS arm64
 COPY --from=cuda-11 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_v11
 COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
-COPY --from=jetpack-5 dist/lib/ollama/cuda_v11 lib/ollama/cuda_jetpack5
-COPY --from=jetpack-6 dist/lib/ollama/cuda_v12 lib/ollama/cuda_jetpack6
+COPY --from=jetpack-5 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_jetpack5
+COPY --from=jetpack-6 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_jetpack6
 
 FROM scratch AS rocm
 COPY --from=rocm-6 dist/lib/ollama/rocm /lib/ollama/rocm

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggerganov/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=d7cfe1ffe0f435d0048a6058d529daf76e072d9c
+FETCH_HEAD=2016f07bd106c73699ecbaace80f55db5ed95dac
 
 .PHONY: help
 help:
@@ -15,18 +15,18 @@ help:
 	@echo "    make -f $(lastword $(MAKEFILE_LIST)) clean sync"
 
 .PHONY: sync
-sync: llama/build-info.cpp llama/llama.cpp ml/backend/ggml/ggml apply-patches
+sync: llama/build-info.cpp llama/llama.cpp ml/backend/ggml/ggml
 
 .PHONY: llama/build-info.cpp
 llama/build-info.cpp: llama/build-info.cpp.in
 	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' $< > $@
 
 .PHONY: llama/llama.cpp
-llama/llama.cpp: llama/vendor/ apply-patches
+llama/llama.cpp: llama/vendor/
 	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
 
-.PHONY: ml/backend/ggml/ggml apply-patches
-ml/backend/ggml/ggml: llama/vendor/ggml/ apply-patches
+.PHONY: ml/backend/ggml/ggml
+ml/backend/ggml/ggml: llama/vendor/ggml/
 	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
 
 PATCHES=$(wildcard llama/patches/*.patch)

README.md

@@ -285,12 +285,13 @@ 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)
 - [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)
 - [Ollamac](https://github.com/kevinhermawan/Ollamac)
-- [big-AGI](https://github.com/enricoros/big-AGI/blob/main/docs/config-local-ollama.md)
+- [big-AGI](https://github.com/enricoros/big-AGI)
 - [Cheshire Cat assistant framework](https://github.com/cheshire-cat-ai/core)
 - [Amica](https://github.com/semperai/amica)
 - [chatd](https://github.com/BruceMacD/chatd)
@@ -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)
@@ -346,7 +348,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [PartCAD](https://github.com/openvmp/partcad/) (CAD model generation with OpenSCAD and CadQuery)
 - [Ollama4j Web UI](https://github.com/ollama4j/ollama4j-web-ui) - Java-based Web UI for Ollama built with Vaadin, Spring Boot and Ollama4j
 - [PyOllaMx](https://github.com/kspviswa/pyOllaMx) - macOS application capable of chatting with both Ollama and Apple MLX models.
-- [Claude Dev](https://github.com/saoudrizwan/claude-dev) - VSCode extension for multi-file/whole-repo coding
+- [Cline](https://github.com/cline/cline) - Formerly known as Claude Dev is a VSCode extension for multi-file/whole-repo coding
 - [Cherry Studio](https://github.com/kangfenmao/cherry-studio) (Desktop client with Ollama support)
 - [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
 - [Archyve](https://github.com/nickthecook/archyve) (RAG-enabling document library)
@@ -394,6 +396,9 @@ 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)
+- [AppFlowy](https://github.com/AppFlowy-IO/AppFlowy) (AI collaborative workspace with Ollama, cross-platform and self-hostable)
 
 ### Cloud
 
@@ -433,7 +438,10 @@ 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.
+- [GGUF-to-Ollama](https://github.com/jonathanhecl/gguf-to-ollama) - Importing GGUF to Ollama made easy (multiplatform)
 
 ### 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.
@@ -75,13 +76,13 @@ type GenerateRequest struct {
 	// this request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
-	// Images is an optional list of base64-encoded images accompanying this
+	// Images is an optional list of raw image bytes accompanying this
 	// request, for multimodal models.
 	Images []ImageData `json:"images,omitempty"`
 
 	// 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
@@ -162,19 +163,65 @@ func (t *ToolCallFunctionArguments) String() string {
 
 type Tool struct {
 	Type     string       `json:"type"`
+	Items    any          `json:"items,omitempty"`
 	Function ToolFunction `json:"function"`
 }
 
+// PropertyType can be either a string or an array of strings
+type PropertyType []string
+
+// UnmarshalJSON implements the json.Unmarshaler interface
+func (pt *PropertyType) UnmarshalJSON(data []byte) error {
+	// Try to unmarshal as a string first
+	var s string
+	if err := json.Unmarshal(data, &s); err == nil {
+		*pt = []string{s}
+		return nil
+	}
+
+	// If that fails, try to unmarshal as an array of strings
+	var a []string
+	if err := json.Unmarshal(data, &a); err != nil {
+		return err
+	}
+	*pt = a
+	return nil
+}
+
+// MarshalJSON implements the json.Marshaler interface
+func (pt PropertyType) MarshalJSON() ([]byte, error) {
+	if len(pt) == 1 {
+		// If there's only one type, marshal as a string
+		return json.Marshal(pt[0])
+	}
+	// Otherwise marshal as an array
+	return json.Marshal([]string(pt))
+}
+
+// String returns a string representation of the PropertyType
+func (pt PropertyType) String() string {
+	if len(pt) == 0 {
+		return ""
+	}
+	if len(pt) == 1 {
+		return pt[0]
+	}
+	return fmt.Sprintf("%v", []string(pt))
+}
+
 type ToolFunction struct {
 	Name        string `json:"name"`
 	Description string `json:"description"`
 	Parameters  struct {
 		Type       string   `json:"type"`
+		Defs       any      `json:"$defs,omitempty"`
+		Items      any      `json:"items,omitempty"`
 		Required   []string `json:"required"`
 		Properties map[string]struct {
-			Type        string   `json:"type"`
-			Description string   `json:"description"`
-			Enum        []string `json:"enum,omitempty"`
+			Type        PropertyType `json:"type"`
+			Items       any          `json:"items,omitempty"`
+			Description string       `json:"description"`
+			Enum        []any        `json:"enum,omitempty"`
 		} `json:"properties"`
 	} `json:"parameters"`
 }
@@ -260,7 +307,7 @@ 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 +333,7 @@ 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 +379,7 @@ 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 +387,18 @@ 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 +551,7 @@ 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 +608,12 @@ 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 +720,7 @@ 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 +734,7 @@ 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()
@@ -231,3 +231,144 @@ func TestMessage_UnmarshalJSON(t *testing.T) {
 		}
 	}
 }
+
+func TestToolFunction_UnmarshalJSON(t *testing.T) {
+	tests := []struct {
+		name    string
+		input   string
+		wantErr string
+	}{
+		{
+			name: "valid enum with same types",
+			input: `{
+				"name": "test",
+				"description": "test function",
+				"parameters": {
+					"type": "object",
+					"required": ["test"],
+					"properties": {
+						"test": {
+							"type": "string",
+							"description": "test prop",
+							"enum": ["a", "b", "c"]
+						}
+					}
+				}
+			}`,
+			wantErr: "",
+		},
+		{
+			name: "empty enum array",
+			input: `{
+				"name": "test",
+				"description": "test function",
+				"parameters": {
+					"type": "object",
+					"required": ["test"],
+					"properties": {
+						"test": {
+							"type": "string",
+							"description": "test prop",
+							"enum": []
+						}
+					}
+				}
+			}`,
+			wantErr: "",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var tf ToolFunction
+			err := json.Unmarshal([]byte(tt.input), &tf)
+
+			if tt.wantErr != "" {
+				require.Error(t, err)
+				assert.Contains(t, err.Error(), tt.wantErr)
+			} else {
+				require.NoError(t, err)
+			}
+		})
+	}
+}
+
+func TestPropertyType_UnmarshalJSON(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    string
+		expected PropertyType
+	}{
+		{
+			name:     "string type",
+			input:    `"string"`,
+			expected: PropertyType{"string"},
+		},
+		{
+			name:     "array of types",
+			input:    `["string", "number"]`,
+			expected: PropertyType{"string", "number"},
+		},
+		{
+			name:     "array with single type",
+			input:    `["string"]`,
+			expected: PropertyType{"string"},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var pt PropertyType
+			if err := json.Unmarshal([]byte(test.input), &pt); err != nil {
+				t.Errorf("Unexpected error: %v", err)
+			}
+
+			if len(pt) != len(test.expected) {
+				t.Errorf("Length mismatch: got %v, expected %v", len(pt), len(test.expected))
+			}
+
+			for i, v := range pt {
+				if v != test.expected[i] {
+					t.Errorf("Value mismatch at index %d: got %v, expected %v", i, v, test.expected[i])
+				}
+			}
+		})
+	}
+}
+
+func TestPropertyType_MarshalJSON(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    PropertyType
+		expected string
+	}{
+		{
+			name:     "single type",
+			input:    PropertyType{"string"},
+			expected: `"string"`,
+		},
+		{
+			name:     "multiple types",
+			input:    PropertyType{"string", "number"},
+			expected: `["string","number"]`,
+		},
+		{
+			name:     "empty type",
+			input:    PropertyType{},
+			expected: `[]`,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			data, err := json.Marshal(test.input)
+			if err != nil {
+				t.Errorf("Unexpected error: %v", err)
+			}
+
+			if string(data) != test.expected {
+				t.Errorf("Marshaled data mismatch: got %v, expected %v", string(data), test.expected)
+			}
+		})
+	}
+}

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)
@@ -793,13 +808,38 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 
 	fn := func(resp api.ProgressResponse) error {
 		if resp.Digest != "" {
+			if resp.Completed == 0 {
+				// This is the initial status update for the
+				// layer, which the server sends before
+				// beginning the download, for clients to
+				// compute total size and prepare for
+				// downloads, if needed.
+				//
+				// Skipping this here to avoid showing a 0%
+				// progress bar, which *should* clue the user
+				// into the fact that many things are being
+				// downloaded and that the current active
+				// download is not that last. However, in rare
+				// cases it seems to be triggering to some, and
+				// it isn't worth explaining, so just ignore
+				// and regress to the old UI that keeps giving
+				// you the "But wait, there is more!" after
+				// each "100% done" bar, which is "better."
+				return nil
+			}
+
 			if spinner != nil {
 				spinner.Stop()
 			}
 
 			bar, ok := bars[resp.Digest]
 			if !ok {
-				bar = progress.NewBar(fmt.Sprintf("pulling %s...", resp.Digest[7:19]), resp.Total, resp.Completed)
+				name, isDigest := strings.CutPrefix(resp.Digest, "sha256:")
+				name = strings.TrimSpace(name)
+				if isDigest {
+					name = name[:min(12, len(name))]
+				}
+				bar = progress.NewBar(fmt.Sprintf("pulling %s:", name), resp.Total, resp.Completed)
 				bars[resp.Digest] = bar
 				p.Add(resp.Digest, bar)
 			}
@@ -819,11 +859,7 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 	}
 
 	request := api.PullRequest{Name: args[0], Insecure: insecure}
-	if err := client.Pull(cmd.Context(), &request, fn); err != nil {
-		return err
-	}
-
-	return nil
+	return client.Pull(cmd.Context(), &request, fn)
 }
 
 type generateContextKey string
@@ -837,7 +873,7 @@ type runOptions struct {
 	Format      string
 	System      string
 	Images      []api.ImageData
-	Options     map[string]interface{}
+	Options     map[string]any
 	MultiModal  bool
 	KeepAlive   *api.Duration
 }
@@ -1366,12 +1402,12 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_NOPRUNE"],
 				envVars["OLLAMA_ORIGINS"],
 				envVars["OLLAMA_SCHED_SPREAD"],
-				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
 				envVars["OLLAMA_KV_CACHE_TYPE"],
 				envVars["OLLAMA_LLM_LIBRARY"],
 				envVars["OLLAMA_GPU_OVERHEAD"],
 				envVars["OLLAMA_LOAD_TIMEOUT"],
+				envVars["OLLAMA_CONTEXT_LENGTH"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)

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) {
@@ -386,6 +415,7 @@ func TestGetModelfileName(t *testing.T) {
 				if err != nil {
 					t.Fatalf("temp modelfile creation failed: %v", err)
 				}
+				defer tempFile.Close()
 
 				expectedFilename = tempFile.Name()
 				err = cmd.Flags().Set("file", expectedFilename)

cmd/interactive.go

@@ -503,6 +503,7 @@ func normalizeFilePath(fp string) string {
 		"\\\\", "\\", // Escaped backslash
 		"\\*", "*", // Escaped asterisk
 		"\\?", "?", // Escaped question mark
+		"\\~", "~", // Escaped tilde
 	).Replace(fp)
 }
 

convert/convert.go

@@ -7,6 +7,7 @@ import (
 	"io"
 	"io/fs"
 	"log/slog"
+	"slices"
 	"strings"
 
 	"github.com/ollama/ollama/fs/ggml"
@@ -84,14 +85,6 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-func (ModelParameters) writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
-	return ggml.WriteGGUF(ws, kv, ts)
-}
-
-func (AdapterParameters) writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
-	return ggml.WriteGGUF(ws, kv, ts)
-}
-
 type ModelConverter interface {
 	// KV maps parameters to LLM key-values
 	KV(*Tokenizer) ggml.KV
@@ -103,8 +96,6 @@ type ModelConverter interface {
 
 	// specialTokenTypes returns any special token types the model uses
 	specialTokenTypes() []string
-	// writeFile writes the model to the provided io.WriteSeeker
-	writeFile(io.WriteSeeker, ggml.KV, []ggml.Tensor) error
 }
 
 type moreParser interface {
@@ -119,8 +110,6 @@ type AdapterConverter interface {
 	// Replacements returns a list of string pairs to replace in tensor names.
 	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
 	Replacements() []string
-
-	writeFile(io.WriteSeeker, ggml.KV, []ggml.Tensor) error
 }
 
 func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV ggml.KV) error {
@@ -158,7 +147,7 @@ func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV ggml.KV) error {
 		return err
 	}
 
-	return conv.writeFile(ws, conv.KV(baseKV), conv.Tensors(ts))
+	return writeFile(ws, conv.KV(baseKV), conv.Tensors(ts))
 }
 
 // Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
@@ -182,8 +171,12 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 
 	var conv ModelConverter
 	switch p.Architectures[0] {
-	case "LlamaForCausalLM", "MistralForCausalLM":
+	case "LlamaForCausalLM":
 		conv = &llamaModel{}
+	case "Llama4ForConditionalGeneration":
+		conv = &llama4Model{}
+	case "Mistral3ForConditionalGeneration":
+		conv = &mistral3Model{}
 	case "MixtralForCausalLM":
 		conv = &mixtralModel{}
 	case "GemmaForCausalLM":
@@ -246,5 +239,13 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		return err
 	}
 
-	return conv.writeFile(ws, conv.KV(t), conv.Tensors(ts))
+	return writeFile(ws, conv.KV(t), conv.Tensors(ts))
+}
+
+func writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
+	for i := range ts {
+		ts[i].Shape = slices.Clone(ts[i].Shape)
+		slices.Reverse(ts[i].Shape)
+	}
+	return ggml.WriteGGUF(ws, kv, ts)
 }

convert/convert_llama.go

@@ -28,12 +28,12 @@ type llamaModel struct {
 	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
 	RopeTheta             float32 `json:"rope_theta"`
 	RopeScaling           struct {
-		Type                            string  `json:"type"`
-		RopeType                        string  `json:"rope_type"`
-		Factor                          float32 `json:"factor"`
-		LowFrequencyFactor              float32 `json:"low_freq_factor"`
-		HighFrequencyFactor             float32 `json:"high_freq_factor"`
-		OriginalMaxPositionalEmbeddings uint32  `json:"original_max_positional_embeddings"`
+		Type                          string  `json:"type"`
+		RopeType                      string  `json:"rope_type"`
+		Factor                        float32 `json:"factor"`
+		LowFrequencyFactor            float32 `json:"low_freq_factor"`
+		HighFrequencyFactor           float32 `json:"high_freq_factor"`
+		OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
 
 		factors ropeFactor
 	} `json:"rope_scaling"`
@@ -42,6 +42,8 @@ type llamaModel struct {
 	LayerNormEpsilon float32 `json:"layer_norm_epsilon"`
 	NormEpsilon      float32 `json:"norm_epsilon"`
 	HeadDim          uint32  `json:"head_dim"`
+
+	skipRepack bool
 }
 
 var _ ModelConverter = (*llamaModel)(nil)
@@ -70,6 +72,10 @@ func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
 		kv["llama.rope.dimension_count"] = p.HiddenSize / headCount
 	}
 
+	if p.HeadDim > 0 {
+		kv["llama.attention.head_dim"] = p.HeadDim
+	}
+
 	if p.RopeTheta > 0 {
 		kv["llama.rope.freq_base"] = p.RopeTheta
 	}
@@ -84,7 +90,7 @@ func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
 			factorLow := cmp.Or(p.RopeScaling.LowFrequencyFactor, 1.0)
 			factorHigh := cmp.Or(p.RopeScaling.HighFrequencyFactor, 4.0)
 
-			original := cmp.Or(p.RopeScaling.OriginalMaxPositionalEmbeddings, 8192)
+			original := cmp.Or(p.RopeScaling.OriginalMaxPositionEmbeddings, 8192)
 			lambdaLow := float32(original) / factorLow
 			lambdaHigh := float32(original) / factorHigh
 
@@ -133,9 +139,10 @@ func (p *llamaModel) Tensors(ts []Tensor) []ggml.Tensor {
 	}
 
 	for _, t := range ts {
-		if strings.HasSuffix(t.Name(), "attn_q.weight") ||
-			strings.HasSuffix(t.Name(), "attn_k.weight") {
-			t.SetRepacker(p.repack)
+		if strings.HasSuffix(t.Name(), "attn_q.weight") || strings.HasSuffix(t.Name(), "attn_k.weight") {
+			if !p.skipRepack {
+				t.SetRepacker(p.repack)
+			}
 		}
 
 		out = append(out, ggml.Tensor{

convert/convert_llama4.go

@@ -0,0 +1,169 @@
+package convert
+
+import (
+	"slices"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type llama4Model struct {
+	ModelParameters
+	TextModel struct {
+		llamaModel
+		NumExpertsPerToken     uint32 `json:"num_experts_per_tok"`
+		NumLocalExperts        uint32 `json:"num_local_experts"`
+		InterleaveMOELayerStep uint32 `json:"interleave_moe_layer_step"`
+		UseQKNorm              bool   `json:"use_qk_norm"`
+		IntermediateSizeMLP    uint32 `json:"intermediate_size_mlp"`
+		AttentionChunkSize     uint32 `json:"attention_chunk_size"`
+	} `json:"text_config"`
+	VisionModel struct {
+		NumHiddenLayers   uint32  `json:"num_hidden_layers"`
+		HiddenSize        uint32  `json:"hidden_size"`
+		IntermediateSize  uint32  `json:"intermediate_size"`
+		NumAttentionHeads uint32  `json:"num_attention_heads"`
+		ImageSize         uint32  `json:"image_size"`
+		PatchSize         uint32  `json:"patch_size"`
+		RopeTheta         float32 `json:"rope_theta"`
+		NormEpsilon       float32 `json:"norm_eps"`
+		PixelShuffleRatio float32 `json:"pixel_shuffle_ratio"`
+	} `json:"vision_config"`
+}
+
+// KV implements ModelConverter.
+func (p *llama4Model) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "llama4"
+
+	for k, v := range p.TextModel.KV(t) {
+		if strings.HasPrefix(k, "llama.") {
+			kv[strings.ReplaceAll(k, "llama.", "llama4.")] = v
+		}
+	}
+
+	kv["llama4.feed_forward_length"] = p.TextModel.IntermediateSizeMLP
+	kv["llama4.expert_feed_forward_length"] = p.TextModel.IntermediateSize
+
+	kv["llama4.expert_count"] = p.TextModel.NumLocalExperts
+	kv["llama4.expert_used_count"] = p.TextModel.NumExpertsPerToken
+	kv["llama4.interleave_moe_layer_step"] = p.TextModel.InterleaveMOELayerStep
+	kv["llama4.use_qk_norm"] = p.TextModel.UseQKNorm
+	kv["llama4.attention.chunk_size"] = p.TextModel.AttentionChunkSize
+
+	kv["llama4.vision.block_count"] = p.VisionModel.NumHiddenLayers
+	kv["llama4.vision.embedding_length"] = p.VisionModel.HiddenSize
+	kv["llama4.vision.feed_forward_length"] = p.VisionModel.IntermediateSize
+	kv["llama4.vision.attention.head_count"] = p.VisionModel.NumAttentionHeads
+	kv["llama4.vision.image_size"] = p.VisionModel.ImageSize
+	kv["llama4.vision.patch_size"] = p.VisionModel.PatchSize
+	kv["llama4.vision.rope.freq_base"] = p.VisionModel.RopeTheta
+	kv["llama4.vision.layer_norm_epsilon"] = p.VisionModel.NormEpsilon
+	kv["llama4.vision.pixel_shuffle_ratio"] = p.VisionModel.PixelShuffleRatio
+	return kv
+}
+
+// Replacements implements ModelConverter.
+func (p *llama4Model) Replacements() []string {
+	return append(
+		p.TextModel.Replacements(),
+		"language_model.", "",
+		"vision_model", "v",
+		"multi_modal_projector", "mm",
+		"feed_forward.down_proj", "ffn_down",
+		"feed_forward.up_proj", "ffn_up",
+		"feed_forward.gate_proj", "ffn_gate",
+		"feed_forward.", "ffn_",
+		"shared_expert.down_proj", "down_shexp",
+		"shared_expert.gate_proj", "gate_shexp",
+		"shared_expert.up_proj", "up_shexp",
+		"experts.down_proj", "down_exps.weight",
+		"experts.gate_up_proj", "gate_up_exps.weight",
+		"router", "gate_inp",
+		"patch_embedding.linear", "patch_embedding",
+	)
+}
+
+// Tensors implements ModelConverter.
+func (p *llama4Model) Tensors(ts []Tensor) []ggml.Tensor {
+	var out []ggml.Tensor
+
+	var textTensors []Tensor
+	for _, t := range ts {
+		if strings.HasPrefix(t.Name(), "v.") || strings.HasPrefix(t.Name(), "mm.") {
+			out = append(out, ggml.Tensor{
+				Name:     t.Name(),
+				Kind:     t.Kind(),
+				Shape:    t.Shape(),
+				WriterTo: t,
+			})
+		} else if strings.Contains(t.Name(), "ffn_gate_up_exps") {
+			// gate and up projectors are fused
+			// dims[1], dims[2] must be swapped
+			// [experts, hidden_size, intermediate_size * 2] --> [experts, intermediate_size, hidden_size]
+			halfDim := int(t.Shape()[2]) / 2
+
+			newShape := slices.Clone(t.Shape())
+			newShape[1], newShape[2] = newShape[2]/2, newShape[1]
+			for i, name := range []string{"ffn_gate_exps", "ffn_up_exps"} {
+				// clone tensor since we need separate repackers
+				tt := t.Clone()
+				tt.SetRepacker(p.repack(nil, nil, tensor.S(i*halfDim, (i+1)*halfDim)))
+				out = append(out, ggml.Tensor{
+					Name:     strings.ReplaceAll(tt.Name(), "ffn_gate_up_exps", name),
+					Kind:     tt.Kind(),
+					Shape:    newShape,
+					WriterTo: tt,
+				})
+			}
+		} else if strings.Contains(t.Name(), "ffn_down_exps") {
+			// dims[1], dims[2] must be swapped
+			// [experts, intermediate_size, hidden_size] --> [experts, hidden_size, intermediate_size]
+			t.SetRepacker(p.repack())
+			newShape := slices.Clone(t.Shape())
+			newShape[1], newShape[2] = newShape[2], newShape[1]
+			out = append(out, ggml.Tensor{
+				Name:     t.Name(),
+				Kind:     t.Kind(),
+				Shape:    newShape,
+				WriterTo: t,
+			})
+		} else {
+			textTensors = append(textTensors, t)
+		}
+	}
+
+	p.TextModel.skipRepack = true
+	out = append(out, p.TextModel.Tensors(textTensors)...)
+	return out
+}
+
+func (p *llama4Model) repack(slice ...tensor.Slice) Repacker {
+	return func(name string, data []float32, shape []uint64) ([]float32, error) {
+		dims := make([]int, len(shape))
+		for i, dim := range shape {
+			dims[i] = int(dim)
+		}
+
+		var t tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+		t, err := t.Slice(slice...)
+		if err != nil {
+			return nil, err
+		}
+
+		if err := t.T(0, 2, 1); err != nil {
+			return nil, err
+		}
+
+		t = tensor.Materialize(t)
+		// flatten tensor so it can be return as a vector
+		if err := t.Reshape(t.Shape().TotalSize()); err != nil {
+			return nil, err
+		}
+
+		return native.VectorF32(t.(*tensor.Dense))
+	}
+}

convert/convert_mistral.go

@@ -0,0 +1,190 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type mistral3Model struct {
+	ModelParameters
+	ImageTokenIndex    uint32 `json:"image_token_index"`
+	SpatialMergeSize   uint32 `json:"spatial_merge_size"`
+	VisionFeatureLayer int32  `json:"vision_feature_layer"`
+	TextModel          struct {
+		NumHiddenLayers       uint32  `json:"num_hidden_layers"`
+		MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+		HiddenSize            uint32  `json:"hidden_size"`
+		IntermediateSize      uint32  `json:"intermediate_size"`
+		NumAttentionHeads     uint32  `json:"num_attention_heads"`
+		NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+		RopeTheta             float32 `json:"rope_theta"`
+		RMSNormEPS            float32 `json:"rms_norm_eps"`
+		HeadDim               uint32  `json:"head_dim"`
+		SlidingWindow         *uint32 `json:"sliding_window"`
+		HiddenAct             string  `json:"hidden_act"`
+		VocabSize             uint32  `json:"vocab_size"`
+	} `json:"text_config"`
+	VisionModel struct {
+		NumAttentionHeads uint32  `json:"num_attention_heads"`
+		NumHiddenLayers   uint32  `json:"num_hidden_layers"`
+		HiddenSize        uint32  `json:"hidden_size"`
+		IntermediateSize  uint32  `json:"intermediate_size"`
+		ImageSize         uint32  `json:"image_size"`
+		NumChannels       uint32  `json:"num_channels"`
+		PatchSize         uint32  `json:"patch_size"`
+		HeadDim           uint32  `json:"head_dim"`
+		HiddenAct         string  `json:"hidden_act"`
+		RopeTheta         float32 `json:"rope_theta"`
+	} `json:"vision_config"`
+	MultiModalProjectorBias bool   `json:"multimodal_projector_bias"`
+	ProjectorHiddenAct      string `json:"projector_hidden_act"`
+}
+
+func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "mistral3"
+	kv["mistral3.vocab_size"] = p.TextModel.VocabSize
+
+	// Text configuration
+	kv["mistral3.block_count"] = p.TextModel.NumHiddenLayers
+	kv["mistral3.context_length"] = p.TextModel.MaxPositionEmbeddings
+	kv["mistral3.embedding_length"] = p.TextModel.HiddenSize
+	kv["mistral3.feed_forward_length"] = p.TextModel.IntermediateSize
+	kv["mistral3.attention.head_count"] = p.TextModel.NumAttentionHeads
+	kv["mistral3.attention.head_count_kv"] = p.TextModel.NumKeyValueHeads
+	kv["mistral3.attention.layer_norm_rms_epsilon"] = p.TextModel.RMSNormEPS
+	kv["mistral3.attention.key_length"] = p.TextModel.HeadDim
+	kv["mistral3.attention.value_length"] = p.TextModel.HeadDim
+	kv["mistral3.rope.dimension_count"] = p.TextModel.HiddenSize / p.TextModel.NumHiddenLayers
+	kv["mistral3.rope.freq_base"] = p.TextModel.RopeTheta
+
+	// Vision configuration
+	kv["mistral3.vision.block_count"] = p.VisionModel.NumHiddenLayers
+	kv["mistral3.vision.embedding_length"] = p.VisionModel.HiddenSize
+	kv["mistral3.vision.feed_forward_length"] = p.VisionModel.IntermediateSize
+	kv["mistral3.vision.attention.head_count"] = p.VisionModel.NumAttentionHeads
+	kv["mistral3.vision.attention.key_length"] = p.VisionModel.HeadDim
+	kv["mistral3.vision.image_size"] = p.VisionModel.ImageSize
+	kv["mistral3.vision.patch_size"] = p.VisionModel.PatchSize
+	kv["mistral3.vision.num_channels"] = p.VisionModel.NumChannels
+	// kv["mistral3.vision.attention.layer_norm_epsilon"] = 1e-05 // Default value
+	kv["mistral3.vision.rope.freq_base"] = p.VisionModel.RopeTheta
+
+	// Multimodal configuration
+	kv["mistral3.image_token_index"] = p.ImageTokenIndex
+	kv["mistral3.spatial_merge_size"] = p.SpatialMergeSize
+
+	kv["mistral3.mm.projector_bias"] = p.MultiModalProjectorBias
+
+	if p.ProjectorHiddenAct != "" {
+		kv["mistral3.mm.projector_hidden_act"] = p.ProjectorHiddenAct
+	}
+
+	return kv
+}
+
+func (p *mistral3Model) Tensors(ts []Tensor) []ggml.Tensor {
+	var out []ggml.Tensor
+
+	for _, t := range ts {
+		if !strings.HasPrefix(t.Name(), "v.") {
+			if strings.HasSuffix(t.Name(), ".attn_q.weight") ||
+				strings.HasSuffix(t.Name(), ".attn_k.weight") {
+				t.SetRepacker(p.repack)
+			}
+		}
+
+		out = append(out, ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *mistral3Model) Replacements() []string {
+	return []string{
+		"language_model.model.norm", "output_norm",
+		"language_model.model.", "",
+		"language_model.", "",
+		"layers", "blk",
+		"transformer.layers", "blk",
+		"vision_tower", "v",
+		"ln_pre", "encoder_norm",
+		"input_layernorm", "attn_norm",
+		"post_attention_layernorm", "ffn_norm",
+		"embed_tokens", "token_embd",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.down_proj", "ffn_down",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.up_proj", "ffn_up",
+		"attention.q_proj", "attn_q",
+		"attention.k_proj", "attn_k",
+		"attention.v_proj", "attn_v",
+		"attention.o_proj", "attn_output",
+		"attention_norm", "attn_norm",
+		"feed_forward.gate_proj", "ffn_gate",
+		"feed_forward.down_proj", "ffn_down",
+		"feed_forward.up_proj", "ffn_up",
+		"multi_modal_projector", "mm",
+		"ffn_norm", "ffn_norm",
+		"lm_head", "output",
+	}
+}
+
+func (p *mistral3Model) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	var dims []int
+	for _, dim := range shape {
+		dims = append(dims, int(dim))
+	}
+
+	var heads uint32
+	if strings.HasSuffix(name, ".attn_q.weight") {
+		heads = p.TextModel.NumAttentionHeads
+	} else if strings.HasSuffix(name, ".attn_k.weight") {
+		heads = cmp.Or(p.TextModel.NumKeyValueHeads, p.TextModel.NumAttentionHeads)
+	} else {
+		return nil, fmt.Errorf("unknown tensor for repack: %s", name)
+	}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+	if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}

convert/convert_phi3.go

@@ -118,6 +118,5 @@ func (p *phi3Model) Replacements() []string {
 type ropeFactor []float32
 
 func (r ropeFactor) WriteTo(w io.Writer) (int64, error) {
-	err := binary.Write(w, binary.LittleEndian, r)
-	return 0, err
+	return 0, binary.Write(w, binary.LittleEndian, r)
 }

convert/convert_test.go

@@ -11,7 +11,6 @@ import (
 	"io"
 	"io/fs"
 	"log/slog"
-	"math"
 	"os"
 	"path/filepath"
 	"slices"
@@ -48,7 +47,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
 	}
 	t.Cleanup(func() { r.Close() })
 
-	m, _, err := ggml.Decode(r, math.MaxInt)
+	m, _, err := ggml.Decode(r, -1)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -332,7 +331,7 @@ func TestConvertAdapter(t *testing.T) {
 			}
 			defer r.Close()
 
-			m, _, err := ggml.Decode(r, math.MaxInt)
+			m, _, err := ggml.Decode(r, -1)
 			if err != nil {
 				t.Fatal(err)
 			}

convert/reader.go

@@ -11,14 +11,15 @@ type Tensor interface {
 	Name() string
 	Shape() []uint64
 	Kind() uint32
-	SetRepacker(repacker)
+	SetRepacker(Repacker)
 	WriteTo(io.Writer) (int64, error)
+	Clone() Tensor
 }
 
 type tensorBase struct {
-	name  string
-	shape []uint64
-	repacker
+	name     string
+	shape    []uint64
+	repacker Repacker
 }
 
 func (t tensorBase) Name() string {
@@ -36,7 +37,8 @@ const (
 
 func (t tensorBase) Kind() uint32 {
 	if strings.HasSuffix(t.name, ".ffn_gate_inp.weight") ||
-		t.name == "token_types.weight" {
+		t.name == "token_types.weight" ||
+		t.name == "v.positional_embedding_vlm" {
 		// these tensors are always F32
 		return 0
 	}
@@ -51,21 +53,18 @@ func (t tensorBase) Kind() uint32 {
 	}
 }
 
-func (t *tensorBase) SetRepacker(fn repacker) {
+func (t *tensorBase) SetRepacker(fn Repacker) {
 	t.repacker = fn
 }
 
-type repacker func(string, []float32, []uint64) ([]float32, error)
+type Repacker func(string, []float32, []uint64) ([]float32, error)
 
 func parseTensors(fsys fs.FS, replacer *strings.Replacer) ([]Tensor, error) {
 	patterns := []struct {
 		Pattern string
 		Func    func(fs.FS, *strings.Replacer, ...string) ([]Tensor, error)
 	}{
-		{"model-*-of-*.safetensors", parseSafetensors},
-		{"model.safetensors", parseSafetensors},
-		{"adapters.safetensors", parseSafetensors},
-		{"adapter_model.safetensors", parseSafetensors},
+		{"*.safetensors", parseSafetensors},
 		{"pytorch_model-*-of-*.bin", parseTorch},
 		{"pytorch_model.bin", parseTorch},
 		{"consolidated.*.pth", parseTorch},

convert/reader_safetensors.go

@@ -94,6 +94,21 @@ type safetensor struct {
 	*tensorBase
 }
 
+func (st safetensor) Clone() Tensor {
+	return &safetensor{
+		fs:     st.fs,
+		path:   st.path,
+		dtype:  st.dtype,
+		offset: st.offset,
+		size:   st.size,
+		tensorBase: &tensorBase{
+			name:     st.name,
+			repacker: st.repacker,
+			shape:    slices.Clone(st.shape),
+		},
+	}
+}
+
 func (st safetensor) WriteTo(w io.Writer) (int64, error) {
 	f, err := st.fs.Open(st.path)
 	if err != nil {

convert/reader_torch.go

@@ -43,6 +43,17 @@ type torch struct {
 	*tensorBase
 }
 
+func (t torch) Clone() Tensor {
+	return torch{
+		storage: t.storage,
+		tensorBase: &tensorBase{
+			name:     t.name,
+			shape:    t.shape,
+			repacker: t.repacker,
+		},
+	}
+}
+
 func (pt torch) WriteTo(w io.Writer) (int64, error) {
 	return 0, nil
 }

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

@@ -173,7 +173,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ##### Response
 
-```json
+```json5
 {
   "model": "codellama:code",
   "created_at": "2024-07-22T20:47:51.147561Z",
@@ -1217,13 +1217,13 @@ Show information about a model including details, modelfile, template, parameter
 
 ```shell
 curl http://localhost:11434/api/show -d '{
-  "model": "llama3.2"
+  "model": "llava"
 }'
 ```
 
 #### Response
 
-```json
+```json5
 {
   "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llava:latest\n\nFROM /Users/matt/.ollama/models/blobs/sha256:200765e1283640ffbd013184bf496e261032fa75b99498a9613be4e94d63ad52\nTEMPLATE \"\"\"{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: \"\"\"\nPARAMETER num_ctx 4096\nPARAMETER stop \"\u003c/s\u003e\"\nPARAMETER stop \"USER:\"\nPARAMETER stop \"ASSISTANT:\"",
   "parameters": "num_keep                       24\nstop                           \"<|start_header_id|>\"\nstop                           \"<|end_header_id|>\"\nstop                           \"<|eot_id|>\"",
@@ -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

@@ -20,12 +20,18 @@ 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 4096 tokens, unless you have a single GPU with <= 4 GB of VRAM, in which case it will default to 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`:
 
 ```shell
-/set parameter num_ctx 4096
+/set parameter num_ctx 8192
 ```
 
 When using the API, specify the `num_ctx` parameter:
@@ -35,7 +41,7 @@ curl http://localhost:11434/api/generate -d '{
   "model": "llama3.2",
   "prompt": "Why is the sky blue?",
   "options": {
-    "num_ctx": 4096
+    "num_ctx": 8192
   }
 }'
 ```

docs/template.md

@@ -12,7 +12,7 @@ A basic Go template consists of three main parts:
 
 Here's an example of a simple chat template:
 
-```gotmpl
+```go
 {{- range .Messages }}
 {{ .Role }}: {{ .Content }}
 {{- end }}
@@ -162,6 +162,6 @@ CodeLlama [7B](https://ollama.com/library/codellama:7b-code) and [13B](https://o
 
 Codestral [22B](https://ollama.com/library/codestral:22b) supports fill-in-middle.
 
-```gotmpl
+```go
 [SUFFIX]{{ .Suffix }}[PREFIX] {{ .Prompt }}
 ```

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:
@@ -26,7 +26,6 @@ When you run Ollama on **Windows**, there are a few different locations. You can
 - `explorer %LOCALAPPDATA%\Ollama` to view logs.  The most recent server logs will be in `server.log` and older logs will be in `server-#.log` 
 - `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
-- `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
 
 To enable additional debug logging to help troubleshoot problems, first **Quit the running app from the tray menu** then in a powershell terminal
 
@@ -69,10 +68,6 @@ If you run into problems on Linux and want to install an older version, or you'd
 curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.5.7 sh
 ```
 
-## Linux tmp noexec 
-
-If your system is configured with the "noexec" flag where Ollama stores its temporary executable files, you can specify an alternate location by setting OLLAMA_TMPDIR to a location writable by the user ollama runs as. For example OLLAMA_TMPDIR=/usr/share/ollama/
-
 ## Linux docker
 
 If Ollama initially works on the GPU in a docker container, but then switches to running on CPU after some period of time with errors in the server log reporting GPU discovery failures, this can be resolved by disabling systemd cgroup management in Docker.  Edit `/etc/docker/daemon.json` on the host and add `"exec-opts": ["native.cgroupdriver=cgroupfs"]` to the docker configuration.

docs/windows.md

@@ -62,7 +62,6 @@ the explorer window by hitting `<Ctrl>+R` and type in:
     - *upgrade.log* contains log output for upgrades
 - `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` contains models and configuration
-- `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories
 
 ## Uninstall
 

envconfig/config.go

@@ -169,7 +169,7 @@ var (
 	// Enable the new Ollama engine
 	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 	// ContextLength sets the default context length
-	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 2048)
+	ContextLength = Int64("OLLAMA_CONTEXT_LENGTH", -1)
 )
 
 func String(s string) func() string {
@@ -227,6 +227,20 @@ func Uint64(key string, defaultValue uint64) func() uint64 {
 	}
 }
 
+func Int64(key string, defaultValue int64) func() int64 {
+	return func() int64 {
+		if s := Var(key); s != "" {
+			if n, err := strconv.ParseInt(s, 10, 64); err != nil {
+				slog.Warn("invalid environment variable, using default", "key", key, "value", s, "default", defaultValue)
+			} else {
+				return n
+			}
+		}
+
+		return defaultValue
+	}
+}
+
 // Set aside VRAM per GPU
 var GpuOverhead = Uint64("OLLAMA_GPU_OVERHEAD", 0)
 
@@ -255,7 +269,7 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowedOrigins(), "A comma separated list of allowed origins"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
-		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 2048)"},
+		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default 4096 or 2048 with low VRAM)"},
 		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
 
 		// Informational

envconfig/config_test.go

@@ -278,8 +278,8 @@ func TestVar(t *testing.T) {
 }
 
 func TestContextLength(t *testing.T) {
-	cases := map[string]uint{
-		"":     2048,
+	cases := map[string]int64{
+		"":     -1,
 		"4096": 4096,
 	}
 

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/config.go

@@ -0,0 +1,13 @@
+package fs
+
+type Config interface {
+	Architecture() string
+	String(string, ...string) string
+	Uint(string, ...uint32) uint32
+	Float(string, ...float32) float32
+	Bool(string, ...bool) bool
+
+	Strings(string, ...[]string) []string
+	Ints(string, ...[]int32) []int32
+	Floats(string, ...[]float32) []float32
+}

fs/ggml/ggml.go

@@ -33,7 +33,7 @@ func (kv KV) Kind() string {
 }
 
 func (kv KV) ParameterCount() uint64 {
-	return keyValue[uint64](kv, "general.parameter_count")
+	return keyValue(kv, "general.parameter_count", uint64(0))
 }
 
 func (kv KV) FileType() fileType {
@@ -105,39 +105,42 @@ func (kv KV) Bool(key string, defaultValue ...bool) bool {
 }
 
 func (kv KV) Strings(key string, defaultValue ...[]string) []string {
-	r := keyValue(kv, key, &array{})
-	s := make([]string, r.size)
-	for i := range r.size {
-		s[i] = r.values[i].(string)
-	}
+	return keyValue(kv, key, &array[string]{values: append(defaultValue, []string(nil))[0]}).values
+}
 
-	return s
+func (kv KV) Ints(key string, defaultValue ...[]int32) []int32 {
+	return keyValue(kv, key, &array[int32]{values: append(defaultValue, []int32(nil))[0]}).values
 }
 
 func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
-	r := keyValue(kv, key, &array{})
-	s := make([]uint32, r.size)
-	for i := range r.size {
-		s[i] = uint32(r.values[i].(int32))
-	}
-
-	return s
+	return keyValue(kv, key, &array[uint32]{values: append(defaultValue, []uint32(nil))[0]}).values
 }
 
 func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
-	r := keyValue(kv, key, &array{})
-	s := make([]float32, r.size)
-	for i := range r.size {
-		s[i] = float32(r.values[i].(float32))
-	}
-	return s
+	return keyValue(kv, key, &array[float32]{values: append(defaultValue, []float32(nil))[0]}).values
 }
 
 func (kv KV) OllamaEngineRequired() bool {
-	return kv.Architecture() == "gemma3"
+	return slices.Contains([]string{
+		"gemma3",
+		"mistral3",
+		"llama4",
+	}, kv.Architecture())
 }
 
-func keyValue[T string | uint32 | uint64 | float32 | *array | bool](kv KV, key string, defaultValue ...T) T {
+type valueTypes interface {
+	uint8 | int8 | uint16 | int16 |
+		uint32 | int32 | uint64 | int64 |
+		string | float32 | float64 | bool
+}
+
+type arrayValueTypes interface {
+	*array[uint8] | *array[int8] | *array[uint16] | *array[int16] |
+		*array[uint32] | *array[int32] | *array[uint64] | *array[int64] |
+		*array[string] | *array[float32] | *array[float64] | *array[bool]
+}
+
+func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ...T) T {
 	if !strings.HasPrefix(key, "tokenizer.") && !strings.HasPrefix(key, "general.") {
 		key = kv.Architecture() + "." + key
 	}
@@ -372,13 +375,8 @@ func DetectContentType(b []byte) string {
 // Decode decodes a GGML model from the given reader.
 //
 // It collects array values for arrays with a size less than or equal to
-// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
-// the maxArraySize is negative, all arrays are collected.
+// maxArraySize. If the maxArraySize is negative, all arrays are collected.
 func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
-	if maxArraySize == 0 {
-		maxArraySize = 1024
-	}
-
 	rs = bufioutil.NewBufferedSeeker(rs, 32<<10)
 
 	var magic uint32
@@ -413,11 +411,11 @@ 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()
-	vocab := uint64(f.KV()["tokenizer.ggml.tokens"].(*array).size)
+	vocab := uint64(f.KV()["tokenizer.ggml.tokens"].(*array[string]).size)
 
 	embeddingHeads := f.KV().EmbeddingHeadCount()
 	embeddingHeadsK := f.KV().EmbeddingHeadCountK()
@@ -426,10 +424,13 @@ 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":
+	case "llama", "llama4":
 		fullOffload = max(
 			4*batch*(1+4*embedding+context*(1+heads)),
 			4*batch*(embedding+vocab),
@@ -443,7 +444,7 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 
 		if ffnGateExpsWeight, ok := layers["blk.0"]["ffn_gate_exps.weight"]; ok {
 			// mixtral 8x22b
-			ff := uint64(f.KV()["llama.feed_forward_length"].(uint32))
+			ff := uint64(f.KV().Uint("feed_forward_length"))
 			partialOffload = max(
 				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embeddingHeads*headsKV),
 				4*(context*batch*heads+context*embeddingHeads*headsKV+batch*1024+embeddingHeads*headsKV*batch),
@@ -460,16 +461,14 @@ 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().Ints("attention.cross_attention_layers")
+		for i := range kv {
+			if slices.Contains(crossAttentionLayers, int32(i)) {
+				kv[i] = headsKV * (embeddingHeadsK + embeddingHeadsV) *
+					4 * // sizeof(float32)
+					visionTokens *
+					tiles
+			}
 		}
 
 		fullOffload = max(
@@ -505,6 +504,20 @@ 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),
@@ -623,10 +636,13 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
 			embeddingLength*numPatches*maxNumTiles +
 			9*embeddingLength*numPaddedPatches*maxNumTiles +
 			numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount)
-	case "gemma3":
+	case "gemma3", "mistral3":
 		graphSize = 4 * (imageSize*imageSize*numChannels +
 			embeddingLength*patchSize +
 			numPatches*numPatches*headCount)
+	case "llama4":
+		// vision graph is computed independently in the same schedule
+		// and is negligible compared to the worst case text graph
 	}
 
 	return weights, graphSize

fs/ggml/ggml_test.go

@@ -2,6 +2,7 @@ package ggml
 
 import (
 	"maps"
+	"math"
 	"slices"
 	"strconv"
 	"strings"
@@ -210,3 +211,61 @@ func TestTensorTypes(t *testing.T) {
 		})
 	}
 }
+
+func TestKeyValue(t *testing.T) {
+	kv := KV{
+		"general.architecture": "test",
+		"test.strings":         &array[string]{size: 3, values: []string{"a", "b", "c"}},
+		"test.float32s":        &array[float32]{size: 3, values: []float32{1.0, 2.0, 3.0}},
+		"test.int32s":          &array[int32]{size: 3, values: []int32{1, 2, 3}},
+		"test.uint32s":         &array[uint32]{size: 3, values: []uint32{1, 2, 3}},
+	}
+
+	if diff := cmp.Diff(kv.Strings("strings"), []string{"a", "b", "c"}); diff != "" {
+		t.Errorf("unexpected strings (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Strings("nonexistent.strings"), []string(nil)); diff != "" {
+		t.Errorf("unexpected strings (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Strings("default.strings", []string{"ollama"}), []string{"ollama"}); diff != "" {
+		t.Errorf("unexpected strings (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Floats("float32s"), []float32{1.0, 2.0, 3.0}); diff != "" {
+		t.Errorf("unexpected float32s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Floats("nonexistent.float32s"), []float32(nil)); diff != "" {
+		t.Errorf("unexpected float32s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Floats("default.float32s", []float32{math.MaxFloat32}), []float32{math.MaxFloat32}); diff != "" {
+		t.Errorf("unexpected float32s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Ints("int32s"), []int32{1, 2, 3}); diff != "" {
+		t.Errorf("unexpected int8s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Ints("nonexistent.int32s"), []int32(nil)); diff != "" {
+		t.Errorf("unexpected int8s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Ints("default.int32s", []int32{math.MaxInt32}), []int32{math.MaxInt32}); diff != "" {
+		t.Errorf("unexpected int8s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Uints("uint32s"), []uint32{1, 2, 3}); diff != "" {
+		t.Errorf("unexpected uint8s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Uints("nonexistent.uint32s"), []uint32(nil)); diff != "" {
+		t.Errorf("unexpected uint8s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Uints("default.uint32s", []uint32{math.MaxUint32}), []uint32{math.MaxUint32}); diff != "" {
+		t.Errorf("unexpected uint8s (-got +want):\n%s", diff)
+	}
+}

fs/ggml/gguf.go

@@ -36,10 +36,6 @@ type containerGGUF struct {
 	maxArraySize int
 }
 
-func (c *containerGGUF) canCollectArray(size int) bool {
-	return c.maxArraySize < 0 || size <= c.maxArraySize
-}
-
 func (c *containerGGUF) Name() string {
 	return "gguf"
 }
@@ -235,10 +231,7 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 	// patch KV with parameter count
 	llm.kv["general.parameter_count"] = llm.parameters
 
-	alignment, ok := llm.kv["general.alignment"].(uint32)
-	if !ok {
-		alignment = 32
-	}
+	alignment := llm.kv.Uint("general.alignment", 32)
 
 	offset, err := rs.Seek(0, io.SeekCurrent)
 	if err != nil {
@@ -298,6 +291,23 @@ func readGGUFV1String(llm *gguf, r io.Reader) (string, error) {
 	return b.String(), nil
 }
 
+func readGGUFV1StringsData(llm *gguf, r io.Reader, a *array[string]) (any, error) {
+	for i := range a.size {
+		if a.values != nil {
+			e, err := readGGUFV1String(llm, r)
+			if err != nil {
+				return nil, err
+			}
+
+			a.values[i] = e
+		} else {
+			discardGGUFString(llm, r)
+		}
+	}
+
+	return a, nil
+}
+
 func discardGGUFString(llm *gguf, r io.Reader) error {
 	buf := llm.scratch[:8]
 	_, err := io.ReadFull(r, buf)
@@ -355,78 +365,44 @@ func writeGGUFString(w io.Writer, s string) error {
 	return err
 }
 
-type array struct {
-	size   int
-	values []any
-}
-
-func (a *array) MarshalJSON() ([]byte, error) {
-	return json.Marshal(a.values)
-}
-
-func readGGUFV1Array(llm *gguf, r io.Reader) (*array, error) {
-	t, err := readGGUF[uint32](llm, r)
-	if err != nil {
-		return nil, err
-	}
-
-	n, err := readGGUF[uint32](llm, r)
-	if err != nil {
-		return nil, err
-	}
-
-	a := &array{size: int(n)}
-	if llm.canCollectArray(int(n)) {
-		a.values = make([]any, 0, int(n))
-	}
-
-	for i := range n {
-		var e any
-		switch t {
-		case ggufTypeUint8:
-			e, err = readGGUF[uint8](llm, r)
-		case ggufTypeInt8:
-			e, err = readGGUF[int8](llm, r)
-		case ggufTypeUint16:
-			e, err = readGGUF[uint16](llm, r)
-		case ggufTypeInt16:
-			e, err = readGGUF[int16](llm, r)
-		case ggufTypeUint32:
-			e, err = readGGUF[uint32](llm, r)
-		case ggufTypeInt32:
-			e, err = readGGUF[int32](llm, r)
-		case ggufTypeUint64:
-			e, err = readGGUF[uint64](llm, r)
-		case ggufTypeInt64:
-			e, err = readGGUF[int64](llm, r)
-		case ggufTypeFloat32:
-			e, err = readGGUF[float32](llm, r)
-		case ggufTypeFloat64:
-			e, err = readGGUF[float64](llm, r)
-		case ggufTypeBool:
-			e, err = readGGUF[bool](llm, r)
-		case ggufTypeString:
-			e, err = readGGUFV1String(llm, r)
-		default:
-			return nil, fmt.Errorf("invalid array type: %d", t)
-		}
-		if err != nil {
-			return nil, err
-		}
-
+func readGGUFStringsData(llm *gguf, r io.Reader, a *array[string]) (any, error) {
+	for i := range a.size {
 		if a.values != nil {
+			e, err := readGGUFString(llm, r)
+			if err != nil {
+				return nil, err
+			}
+
 			a.values[i] = e
+		} else {
+			discardGGUFString(llm, r)
 		}
 	}
 
 	return a, nil
 }
 
-func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
-	if llm.Version == 1 {
-		return readGGUFV1Array(llm, r)
-	}
+type array[T any] struct {
+	// size is the actual size of the array
+	size int
 
+	// values is the array of values. this is nil if the array is larger than configured maxSize
+	values []T
+}
+
+func (a *array[T]) MarshalJSON() ([]byte, error) {
+	return json.Marshal(a.values)
+}
+
+func newArray[T any](size, maxSize int) *array[T] {
+	a := array[T]{size: size}
+	if maxSize < 0 || size <= maxSize {
+		a.values = make([]T, size)
+	}
+	return &a
+}
+
+func readGGUFArray(llm *gguf, r io.Reader) (any, error) {
 	t, err := readGGUF[uint32](llm, r)
 	if err != nil {
 		return nil, err
@@ -437,45 +413,55 @@ func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
 		return nil, err
 	}
 
-	a := &array{size: int(n)}
-	if llm.canCollectArray(int(n)) {
-		a.values = make([]any, int(n))
-	}
-
-	for i := range n {
-		var e any
-		switch t {
-		case ggufTypeUint8:
-			e, err = readGGUF[uint8](llm, r)
-		case ggufTypeInt8:
-			e, err = readGGUF[int8](llm, r)
-		case ggufTypeUint16:
-			e, err = readGGUF[uint16](llm, r)
-		case ggufTypeInt16:
-			e, err = readGGUF[int16](llm, r)
-		case ggufTypeUint32:
-			e, err = readGGUF[uint32](llm, r)
-		case ggufTypeInt32:
-			e, err = readGGUF[int32](llm, r)
-		case ggufTypeUint64:
-			e, err = readGGUF[uint64](llm, r)
-		case ggufTypeInt64:
-			e, err = readGGUF[int64](llm, r)
-		case ggufTypeFloat32:
-			e, err = readGGUF[float32](llm, r)
-		case ggufTypeFloat64:
-			e, err = readGGUF[float64](llm, r)
-		case ggufTypeBool:
-			e, err = readGGUF[bool](llm, r)
-		case ggufTypeString:
-			if a.values != nil {
-				e, err = readGGUFString(llm, r)
-			} else {
-				err = discardGGUFString(llm, r)
-			}
-		default:
-			return nil, fmt.Errorf("invalid array type: %d", t)
+	switch t {
+	case ggufTypeUint8:
+		a := newArray[uint8](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeInt8:
+		a := newArray[int8](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeUint16:
+		a := newArray[uint16](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeInt16:
+		a := newArray[int16](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeUint32:
+		a := newArray[uint32](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeInt32:
+		a := newArray[int32](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeUint64:
+		a := newArray[uint64](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeInt64:
+		a := newArray[int64](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeFloat32:
+		a := newArray[float32](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeFloat64:
+		a := newArray[float64](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeBool:
+		a := newArray[bool](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeString:
+		a := newArray[string](int(n), llm.maxArraySize)
+		if llm.Version == 1 {
+			return readGGUFV1StringsData(llm, r, a)
 		}
+
+		return readGGUFStringsData(llm, r, a)
+	default:
+		return nil, fmt.Errorf("invalid array type: %d", t)
+	}
+}
+
+func readGGUFArrayData[T any](llm *gguf, r io.Reader, a *array[T]) (any, error) {
+	for i := range a.size {
+		e, err := readGGUF[T](llm, r)
 		if err != nil {
 			return nil, err
 		}
@@ -506,6 +492,8 @@ func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
 }
 
 func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
+	alignment := kv.Uint("general.alignment", 32)
+
 	if err := binary.Write(ws, binary.LittleEndian, []byte("GGUF")); err != nil {
 		return err
 	}
@@ -543,16 +531,15 @@ func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
 
 	var s uint64
 	for _, t := range ts {
-		t.Offset = s
+		t.Offset = s + uint64(ggufPadding(int64(s), int64(alignment)))
 		if err := ggufWriteTensorInfo(ws, t); err != nil {
 			return err
 		}
 		s += t.Size()
 	}
 
-	var alignment int64 = 32
 	for _, t := range ts {
-		if err := ggufWriteTensor(ws, t, alignment); err != nil {
+		if err := ggufWriteTensor(ws, t, int64(alignment)); err != nil {
 			return err
 		}
 	}
@@ -629,8 +616,8 @@ func ggufWriteTensorInfo(ws io.WriteSeeker, t Tensor) error {
 		return err
 	}
 
-	for i := range len(t.Shape) {
-		if err := binary.Write(ws, binary.LittleEndian, t.Shape[len(t.Shape)-i-1]); err != nil {
+	for _, n := range t.Shape {
+		if err := binary.Write(ws, binary.LittleEndian, n); err != nil {
 			return err
 		}
 	}

grammar/bench_test.go

@@ -1,22 +0,0 @@
-//go:build go1.24
-
-package grammar
-
-import "testing"
-
-func BenchmarkFromSchema(b *testing.B) {
-	for tt := range testCases(b) {
-		b.Run("", func(b *testing.B) {
-			s := []byte(tt.schema)
-
-			b.ReportAllocs()
-			for b.Loop() {
-				_, err := FromSchema(nil, s)
-				if err != nil {
-					b.Fatalf("GrammarFromSchema: %v", err)
-				}
-			}
-		})
-		return
-	}
-}

grammar/grammar.go

@@ -1,227 +0,0 @@
-package grammar
-
-import (
-	"bytes"
-	"encoding/json"
-	"fmt"
-	"iter"
-	"strconv"
-
-	"github.com/ollama/ollama/grammar/jsonschema"
-)
-
-const jsonTerms = `
-# Unicode
-#
-# Unicode characters can be specified directly in the grammar, for example
-# hiragana ::= [ぁ-ゟ], or with escapes: 8-bit (\xXX), 16-bit (\uXXXX) or 32-bit
-# (\UXXXXXXXX).
-unicode ::= \x{hex}{2} | \u{hex}{4} | \U{hex}{8}
-
-# JSON grammar from RFC 7159
-null    ::= "null"
-object  ::= "{" (kv ("," kv)*)? "}"
-array   ::= "[" (value ("," value)*)? "]"
-kv      ::= string ":" value
-integer ::= "0" | [1-9] [0-9]*
-number  ::= "-"? integer frac? exp?
-frac    ::= "." [0-9]+
-exp     ::= ("e" | "E") ("+" | "-") [0-9]+
-string  ::= "\"" char* "\""
-escape  ::= ["/" | "b" | "f" | "n" | "r" | "t" | unicode]
-char    ::= [^"\\] | escape
-space   ::= (" " | "\t" | "\n" | "\r")*
-hex     ::= [0-9] | [a-f] | [A-F]
-boolean ::= "true" | "false"
-value   ::= object | array | string | number | boolean | "null"
-
-# User-defined
-`
-
-// FromSchema generates a grammar from a JSON schema.
-func FromSchema(buf []byte, jsonSchema []byte) ([]byte, error) {
-	var s *jsonschema.Schema
-	if err := json.Unmarshal(jsonSchema, &s); err != nil {
-		return nil, err
-	}
-
-	var g builder
-
-	// "root" is the only rule that is guaranteed to exist, so we start
-	// with its length for padding, and then adjust it as we go.
-	g.pad = len("root")
-	for id := range dependencies("root", s) {
-		g.pad = max(g.pad, len(id))
-	}
-
-	g.b.WriteString(jsonTerms)
-
-	ids := make(map[*jsonschema.Schema]string)
-	for id, s := range dependencies("root", s) {
-		ids[s] = id
-		g.define(id)
-		if err := fromSchema(&g, ids, s); err != nil {
-			return nil, err
-		}
-	}
-	g.define("root")
-	if err := fromSchema(&g, ids, s); err != nil {
-		return nil, err
-	}
-	g.define("") // finalize the last rule
-	return g.b.Bytes(), nil
-}
-
-func fromSchema(g *builder, ids map[*jsonschema.Schema]string, s *jsonschema.Schema) error {
-	switch typ := s.EffectiveType(); typ {
-	case "array":
-		if len(s.PrefixItems) == 0 && s.Items == nil {
-			g.u("array")
-		} else {
-			g.q("[")
-			for i, s := range s.PrefixItems {
-				if i > 0 {
-					g.q(",")
-				}
-				g.u(ids[s])
-			}
-			if s.Items != nil {
-				g.u("(")
-				if len(s.PrefixItems) > 0 {
-					g.q(",")
-				}
-				g.u(ids[s.Items])
-				g.u(")*")
-			}
-			g.q("]")
-		}
-	case "object":
-		if len(s.Properties) == 0 {
-			g.u("object")
-		} else {
-			g.q("{")
-			for i, p := range s.Properties {
-				name := ids[p]
-				if i > 0 {
-					g.q(",")
-				}
-				g.q(p.Name)
-				g.q(":")
-				g.u(name)
-			}
-			g.q("}")
-		}
-	case "number":
-		buildConstrainedNumber(g, s)
-	case "string":
-		if len(s.Enum) == 0 {
-			g.u("string")
-		} else {
-			g.u("(")
-			for i, e := range s.Enum {
-				if i > 0 {
-					g.q("|")
-				}
-				g.q(string(e))
-			}
-			g.u(")")
-		}
-	case "boolean", "value", "null", "integer":
-		g.u(typ)
-	default:
-		return fmt.Errorf("%s: unsupported type %q", s.Name, typ)
-	}
-	return nil
-}
-
-// dependencies returns a sequence of all child dependencies of the schema in
-// post-order.
-//
-// The first value is the id/pointer to the dependency, and the second value
-// is the schema.
-func dependencies(id string, s *jsonschema.Schema) iter.Seq2[string, *jsonschema.Schema] {
-	return func(yield func(string, *jsonschema.Schema) bool) {
-		for i, p := range s.Properties {
-			id := fmt.Sprintf("%s_%d", id, i)
-			for did, d := range dependencies(id, p) {
-				if !yield(did, d) {
-					return
-				}
-			}
-			if !yield(id, p) {
-				return
-			}
-		}
-		for i, p := range s.PrefixItems {
-			id := fmt.Sprintf("tuple_%d", i)
-			for did, d := range dependencies(id, p) {
-				id := fmt.Sprintf("%s_%s", id, did)
-				if !yield(id, d) {
-					return
-				}
-			}
-			if !yield(id, p) {
-				return
-			}
-		}
-		if s.Items != nil {
-			id := fmt.Sprintf("%s_tuple_%d", id, len(s.PrefixItems))
-			for did, d := range dependencies(id, s.Items) {
-				if !yield(did, d) {
-					return
-				}
-			}
-			if !yield(id, s.Items) {
-				return
-			}
-		}
-	}
-}
-
-type builder struct {
-	b     bytes.Buffer
-	pad   int
-	rules int
-	items int
-}
-
-// define terminates the current rule, if any, and then either starts a new
-// rule or does nothing else if the name is empty.
-func (b *builder) define(name string) {
-	if b.rules > 0 {
-		b.b.WriteString(";\n")
-	}
-	if name == "" {
-		return
-	}
-	fmt.Fprintf(&b.b, "% -*s", b.pad, name)
-	b.b.WriteString(" ::=")
-	b.rules++
-	b.items = 0
-}
-
-// quote appends a terminal to the current rule.
-func (b *builder) q(s string) {
-	if b.items > 0 {
-		b.b.WriteString(" ")
-	}
-	b.b.WriteString(" ")
-	b.b.WriteString(strconv.Quote(s))
-}
-
-// u appends a non-terminal to the current rule.
-func (b *builder) u(s string) {
-	if b.items > 0 {
-		b.b.WriteString(" ")
-	}
-	b.b.WriteString(" ")
-	b.b.WriteString(s)
-}
-
-func buildConstrainedNumber(b *builder, s *jsonschema.Schema) {
-	if s.Minimum == 0 && s.Maximum == 0 {
-		b.u("TODO")
-	} else {
-		b.u("number")
-	}
-}

grammar/grammar_test.go

@@ -1,75 +0,0 @@
-package grammar
-
-import (
-	"bufio"
-	"cmp"
-	"iter"
-	"strings"
-	"testing"
-
-	_ "embed"
-
-	"github.com/ollama/ollama/grammar/internal/diff"
-)
-
-func TestFromSchema(t *testing.T) {
-	for tt := range testCases(t) {
-		t.Run(tt.name, func(t *testing.T) {
-			g, err := FromSchema(nil, []byte(tt.schema))
-			if err != nil {
-				t.Fatalf("FromSchema: %v", err)
-			}
-			got := string(g)
-			got = strings.TrimPrefix(got, jsonTerms)
-			if got != tt.want {
-				t.Logf("schema:\n%s", tt.schema)
-				t.Fatal(string(diff.Diff("got", []byte(got), "want", []byte(tt.want))))
-			}
-		})
-	}
-}
-
-type testCase struct {
-	name   string
-	schema string
-	want   string
-}
-
-//go:embed testdata/schemas.txt
-var tests string
-
-func testCases(t testing.TB) iter.Seq[testCase] {
-	t.Helper()
-	return func(yield func(testCase) bool) {
-		t.Helper()
-		sc := bufio.NewScanner(strings.NewReader(tests))
-		name := ""
-		for sc.Scan() {
-			line := strings.TrimSpace(sc.Text())
-			if line == "" {
-				name = ""
-				continue
-			}
-			if line[0] == '#' {
-				name = cmp.Or(name, strings.TrimSpace(line[1:]))
-				continue
-			}
-			s := sc.Text()
-			g := ""
-			for sc.Scan() {
-				line = strings.TrimSpace(sc.Text())
-				if line == "" || line[0] == '#' {
-					break
-				}
-				g += sc.Text() + "\n"
-			}
-			if !yield(testCase{name, s, g}) {
-				return
-			}
-			name = strings.TrimSpace(strings.TrimPrefix(line, "#"))
-		}
-		if err := sc.Err(); err != nil {
-			t.Fatalf("error reading tests: %v", err)
-		}
-	}
-}

grammar/internal/diff/diff.go

@@ -1,261 +0,0 @@
-// Copyright 2022 The Go Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-package diff
-
-import (
-	"bytes"
-	"fmt"
-	"sort"
-	"strings"
-)
-
-// A pair is a pair of values tracked for both the x and y side of a diff.
-// It is typically a pair of line indexes.
-type pair struct{ x, y int }
-
-// Diff returns an anchored diff of the two texts old and new
-// in the “unified diff” format. If old and new are identical,
-// Diff returns a nil slice (no output).
-//
-// Unix diff implementations typically look for a diff with
-// the smallest number of lines inserted and removed,
-// which can in the worst case take time quadratic in the
-// number of lines in the texts. As a result, many implementations
-// either can be made to run for a long time or cut off the search
-// after a predetermined amount of work.
-//
-// In contrast, this implementation looks for a diff with the
-// smallest number of “unique” lines inserted and removed,
-// where unique means a line that appears just once in both old and new.
-// We call this an “anchored diff” because the unique lines anchor
-// the chosen matching regions. An anchored diff is usually clearer
-// than a standard diff, because the algorithm does not try to
-// reuse unrelated blank lines or closing braces.
-// The algorithm also guarantees to run in O(n log n) time
-// instead of the standard O(n²) time.
-//
-// Some systems call this approach a “patience diff,” named for
-// the “patience sorting” algorithm, itself named for a solitaire card game.
-// We avoid that name for two reasons. First, the name has been used
-// for a few different variants of the algorithm, so it is imprecise.
-// Second, the name is frequently interpreted as meaning that you have
-// to wait longer (to be patient) for the diff, meaning that it is a slower algorithm,
-// when in fact the algorithm is faster than the standard one.
-func Diff(oldName string, old []byte, newName string, new []byte) []byte {
-	if bytes.Equal(old, new) {
-		return nil
-	}
-	x := lines(old)
-	y := lines(new)
-
-	// Print diff header.
-	var out bytes.Buffer
-	fmt.Fprintf(&out, "diff %s %s\n", oldName, newName)
-	fmt.Fprintf(&out, "--- %s\n", oldName)
-	fmt.Fprintf(&out, "+++ %s\n", newName)
-
-	// Loop over matches to consider,
-	// expanding each match to include surrounding lines,
-	// and then printing diff chunks.
-	// To avoid setup/teardown cases outside the loop,
-	// tgs returns a leading {0,0} and trailing {len(x), len(y)} pair
-	// in the sequence of matches.
-	var (
-		done  pair     // printed up to x[:done.x] and y[:done.y]
-		chunk pair     // start lines of current chunk
-		count pair     // number of lines from each side in current chunk
-		ctext []string // lines for current chunk
-	)
-	for _, m := range tgs(x, y) {
-		if m.x < done.x {
-			// Already handled scanning forward from earlier match.
-			continue
-		}
-
-		// Expand matching lines as far as possible,
-		// establishing that x[start.x:end.x] == y[start.y:end.y].
-		// Note that on the first (or last) iteration we may (or definitely do)
-		// have an empty match: start.x==end.x and start.y==end.y.
-		start := m
-		for start.x > done.x && start.y > done.y && x[start.x-1] == y[start.y-1] {
-			start.x--
-			start.y--
-		}
-		end := m
-		for end.x < len(x) && end.y < len(y) && x[end.x] == y[end.y] {
-			end.x++
-			end.y++
-		}
-
-		// Emit the mismatched lines before start into this chunk.
-		// (No effect on first sentinel iteration, when start = {0,0}.)
-		for _, s := range x[done.x:start.x] {
-			ctext = append(ctext, "-"+s)
-			count.x++
-		}
-		for _, s := range y[done.y:start.y] {
-			ctext = append(ctext, "+"+s)
-			count.y++
-		}
-
-		// If we're not at EOF and have too few common lines,
-		// the chunk includes all the common lines and continues.
-		const C = 3 // number of context lines
-		if (end.x < len(x) || end.y < len(y)) &&
-			(end.x-start.x < C || (len(ctext) > 0 && end.x-start.x < 2*C)) {
-			for _, s := range x[start.x:end.x] {
-				ctext = append(ctext, " "+s)
-				count.x++
-				count.y++
-			}
-			done = end
-			continue
-		}
-
-		// End chunk with common lines for context.
-		if len(ctext) > 0 {
-			n := end.x - start.x
-			if n > C {
-				n = C
-			}
-			for _, s := range x[start.x : start.x+n] {
-				ctext = append(ctext, " "+s)
-				count.x++
-				count.y++
-			}
-			done = pair{start.x + n, start.y + n}
-
-			// Format and emit chunk.
-			// Convert line numbers to 1-indexed.
-			// Special case: empty file shows up as 0,0 not 1,0.
-			if count.x > 0 {
-				chunk.x++
-			}
-			if count.y > 0 {
-				chunk.y++
-			}
-			fmt.Fprintf(&out, "@@ -%d,%d +%d,%d @@\n", chunk.x, count.x, chunk.y, count.y)
-			for _, s := range ctext {
-				out.WriteString(s)
-			}
-			count.x = 0
-			count.y = 0
-			ctext = ctext[:0]
-		}
-
-		// If we reached EOF, we're done.
-		if end.x >= len(x) && end.y >= len(y) {
-			break
-		}
-
-		// Otherwise start a new chunk.
-		chunk = pair{end.x - C, end.y - C}
-		for _, s := range x[chunk.x:end.x] {
-			ctext = append(ctext, " "+s)
-			count.x++
-			count.y++
-		}
-		done = end
-	}
-
-	return out.Bytes()
-}
-
-// lines returns the lines in the file x, including newlines.
-// If the file does not end in a newline, one is supplied
-// along with a warning about the missing newline.
-func lines(x []byte) []string {
-	l := strings.SplitAfter(string(x), "\n")
-	if l[len(l)-1] == "" {
-		l = l[:len(l)-1]
-	} else {
-		// Treat last line as having a message about the missing newline attached,
-		// using the same text as BSD/GNU diff (including the leading backslash).
-		l[len(l)-1] += "\n\\ No newline at end of file\n"
-	}
-	return l
-}
-
-// tgs returns the pairs of indexes of the longest common subsequence
-// of unique lines in x and y, where a unique line is one that appears
-// once in x and once in y.
-//
-// The longest common subsequence algorithm is as described in
-// Thomas G. Szymanski, “A Special Case of the Maximal Common
-// Subsequence Problem,” Princeton TR #170 (January 1975),
-// available at https://research.swtch.com/tgs170.pdf.
-func tgs(x, y []string) []pair {
-	// Count the number of times each string appears in a and b.
-	// We only care about 0, 1, many, counted as 0, -1, -2
-	// for the x side and 0, -4, -8 for the y side.
-	// Using negative numbers now lets us distinguish positive line numbers later.
-	m := make(map[string]int)
-	for _, s := range x {
-		if c := m[s]; c > -2 {
-			m[s] = c - 1
-		}
-	}
-	for _, s := range y {
-		if c := m[s]; c > -8 {
-			m[s] = c - 4
-		}
-	}
-
-	// Now unique strings can be identified by m[s] = -1+-4.
-	//
-	// Gather the indexes of those strings in x and y, building:
-	//	xi[i] = increasing indexes of unique strings in x.
-	//	yi[i] = increasing indexes of unique strings in y.
-	//	inv[i] = index j such that x[xi[i]] = y[yi[j]].
-	var xi, yi, inv []int
-	for i, s := range y {
-		if m[s] == -1+-4 {
-			m[s] = len(yi)
-			yi = append(yi, i)
-		}
-	}
-	for i, s := range x {
-		if j, ok := m[s]; ok && j >= 0 {
-			xi = append(xi, i)
-			inv = append(inv, j)
-		}
-	}
-
-	// Apply Algorithm A from Szymanski's paper.
-	// In those terms, A = J = inv and B = [0, n).
-	// We add sentinel pairs {0,0}, and {len(x),len(y)}
-	// to the returned sequence, to help the processing loop.
-	J := inv
-	n := len(xi)
-	T := make([]int, n)
-	L := make([]int, n)
-	for i := range T {
-		T[i] = n + 1
-	}
-	for i := range n {
-		k := sort.Search(n, func(k int) bool {
-			return T[k] >= J[i]
-		})
-		T[k] = J[i]
-		L[i] = k + 1
-	}
-	k := 0
-	for _, v := range L {
-		if k < v {
-			k = v
-		}
-	}
-	seq := make([]pair, 2+k)
-	seq[1+k] = pair{len(x), len(y)} // sentinel at end
-	lastj := n
-	for i := n - 1; i >= 0; i-- {
-		if L[i] == k && J[i] < lastj {
-			seq[k] = pair{xi[i], yi[J[i]]}
-			k--
-		}
-	}
-	seq[0] = pair{0, 0} // sentinel at start
-	return seq
-}

grammar/internal/diff/diff_test.go

@@ -1,44 +0,0 @@
-// Copyright 2022 The Go Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-package diff
-
-import (
-	"bytes"
-	"path/filepath"
-	"testing"
-
-	"golang.org/x/tools/txtar"
-)
-
-func clean(text []byte) []byte {
-	text = bytes.ReplaceAll(text, []byte("$\n"), []byte("\n"))
-	text = bytes.TrimSuffix(text, []byte("^D\n"))
-	return text
-}
-
-func Test(t *testing.T) {
-	files, _ := filepath.Glob("testdata/*.txt")
-	if len(files) == 0 {
-		t.Fatalf("no testdata")
-	}
-
-	for _, file := range files {
-		t.Run(filepath.Base(file), func(t *testing.T) {
-			a, err := txtar.ParseFile(file)
-			if err != nil {
-				t.Fatal(err)
-			}
-			if len(a.Files) != 3 || a.Files[2].Name != "diff" {
-				t.Fatalf("%s: want three files, third named \"diff\"", file)
-			}
-			diffs := Diff(a.Files[0].Name, clean(a.Files[0].Data), a.Files[1].Name, clean(a.Files[1].Data))
-			want := clean(a.Files[2].Data)
-			if !bytes.Equal(diffs, want) {
-				t.Fatalf("%s: have:\n%s\nwant:\n%s\n%s", file,
-					diffs, want, Diff("have", diffs, "want", want))
-			}
-		})
-	}
-}

grammar/internal/diff/testdata/allnew.txt

@@ -1,13 +0,0 @@
--- old --
--- new --
-a
-b
-c
--- diff --
-diff old new
---- old
-+++ new
-@@ -0,0 +1,3 @@
-+a
-+b
-+c

grammar/internal/diff/testdata/allold.txt

@@ -1,13 +0,0 @@
--- old --
-a
-b
-c
--- new --
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,3 +0,0 @@
--a
--b
--c

grammar/internal/diff/testdata/basic.txt

@@ -1,35 +0,0 @@
-Example from Hunt and McIlroy, “An Algorithm for Differential File Comparison.”
-https://www.cs.dartmouth.edu/~doug/diff.pdf
-
--- old --
-a
-b
-c
-d
-e
-f
-g
--- new --
-w
-a
-b
-x
-y
-z
-e
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,7 +1,7 @@
-+w
- a
- b
--c
--d
-+x
-+y
-+z
- e
--f
--g

grammar/internal/diff/testdata/dups.txt

@@ -1,40 +0,0 @@
--- old --
-a
-
-b
-
-c
-
-d
-
-e
-
-f
--- new --
-a
-
-B
-
-C
-
-d
-
-e
-
-f
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,8 +1,8 @@
- a
- $
--b
--
--c
-+B
-+
-+C
- $
- d
- $

grammar/internal/diff/testdata/end.txt

@@ -1,38 +0,0 @@
--- old --
-1
-2
-3
-4
-5
-6
-7
-eight
-nine
-ten
-eleven
--- new --
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
--- diff --
-diff old new
---- old
-+++ new
-@@ -5,7 +5,6 @@
- 5
- 6
- 7
--eight
--nine
--ten
--eleven
-+8
-+9
-+10

grammar/internal/diff/testdata/eof.txt

@@ -1,9 +0,0 @@
--- old --
-a
-b
-c^D
--- new --
-a
-b
-c^D
--- diff --

grammar/internal/diff/testdata/eof1.txt

@@ -1,18 +0,0 @@
--- old --
-a
-b
-c
--- new --
-a
-b
-c^D
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,3 +1,3 @@
- a
- b
--c
-+c
-\ No newline at end of file

grammar/internal/diff/testdata/eof2.txt

@@ -1,18 +0,0 @@
--- old --
-a
-b
-c^D
--- new --
-a
-b
-c
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,3 +1,3 @@
- a
- b
--c
-\ No newline at end of file
-+c

grammar/internal/diff/testdata/long.txt

@@ -1,62 +0,0 @@
--- old --
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
-12
-13
-14
-14½
-15
-16
-17
-18
-19
-20
--- new --
-1
-2
-3
-4
-5
-6
-8
-9
-10
-11
-12
-13
-14
-17
-18
-19
-20
--- diff --
-diff old new
---- old
-+++ new
-@@ -4,7 +4,6 @@
- 4
- 5
- 6
--7
- 8
- 9
- 10
-@@ -12,9 +11,6 @@
- 12
- 13
- 14
--14½
--15
--16
- 17
- 18
- 19

grammar/internal/diff/testdata/same.txt

@@ -1,5 +0,0 @@
--- old --
-hello world
--- new --
-hello world
--- diff --

grammar/internal/diff/testdata/start.txt

@@ -1,34 +0,0 @@
--- old --
-e
-pi
-4
-5
-6
-7
-8
-9
-10
--- new --
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,5 +1,6 @@
--e
--pi
-+1
-+2
-+3
- 4
- 5
- 6

grammar/internal/diff/testdata/triv.txt

@@ -1,40 +0,0 @@
-Another example from Hunt and McIlroy,
-“An Algorithm for Differential File Comparison.”
-https://www.cs.dartmouth.edu/~doug/diff.pdf
-
-Anchored diff gives up on finding anything,
-since there are no unique lines.
-
--- old --
-a
-b
-c
-a
-b
-b
-a
--- new --
-c
-a
-b
-a
-b
-c
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,7 +1,6 @@
--a
--b
--c
--a
--b
--b
--a
-+c
-+a
-+b
-+a
-+b
-+c

grammar/jsonschema/decode.go

@@ -1,171 +0,0 @@
-package jsonschema
-
-import (
-	"bytes"
-	"encoding/json"
-	"errors"
-)
-
-// Schema holds a JSON schema.
-type Schema struct {
-	// Name is the name of the property. For the parent/root property, this
-	// is "root". For child properties, this is the name of the property.
-	Name string `json:"-"`
-
-	// Type is the type of the property.
-	//
-	// TODO: Union types (e.g. make this a []string).
-	Type string
-
-	// PrefixItems is a list of schemas for each item in a tuple. By
-	// default, the tuple is "closed." unless Items is set to true or a
-	// valid Schema.
-	PrefixItems []*Schema
-
-	// Items is the schema for each item in a list.
-	//
-	// If it is missing, or its JSON value is "null" or "false", it is nil.
-	// If the JSON value is "true", it is set to the empty Schema. If the
-	// JSON value is an object, it will be decoded as a Schema.
-	Items *Schema
-
-	// MinItems specifies the minimum number of items allowed in a list.
-	MinItems int
-
-	// MaxItems specifies the maximum number of items allowed in a list.
-	MaxItems int
-
-	// Properties is the schema for each property of an object.
-	Properties []*Schema
-
-	// Format is the format of the property. This is used to validate the
-	// property against a specific format.
-	//
-	// It is the callers responsibility to validate the property against
-	// the format.
-	Format string
-
-	// Minimum specifies the minimum value for numeric properties.
-	Minimum float64
-
-	// Maximum specifies the maximum value for numeric properties.
-	Maximum float64
-
-	// Enum is a list of valid values for the property.
-	Enum []json.RawMessage
-}
-
-func (s *Schema) UnmarshalJSON(data []byte) error {
-	type S Schema
-	w := struct {
-		Properties props
-		Items      items
-		*S
-	}{
-		S: (*S)(s),
-	}
-	if err := json.Unmarshal(data, &w); err != nil {
-		return err
-	}
-	if w.Items.set {
-		s.Items = &w.Items.Schema
-	}
-	s.Properties = w.Properties
-	return nil
-}
-
-type items struct {
-	Schema
-	set bool
-}
-
-func (s *items) UnmarshalJSON(data []byte) error {
-	switch b := data[0]; b {
-	case 't':
-		*s = items{set: true}
-	case '{':
-		type I items
-		if err := json.Unmarshal(data, (*I)(s)); err != nil {
-			return err
-		}
-		s.set = true
-	case 'n', 'f':
-	default:
-		return errors.New("invalid Items")
-	}
-	return nil
-}
-
-// EffectiveType returns the effective type of the schema. If the Type field is
-// not empty, it is returned; otherwise:
-//
-//   - If the schema has both Properties and Items, it returns an empty string.
-//   - If the schema has Properties, it returns "object".
-//   - If the schema has Items, it returns "array".
-//   - If the schema has neither Properties nor Items, it returns "value".
-//
-// The returned string is never empty.
-func (d *Schema) EffectiveType() string {
-	if d.Type == "" {
-		if len(d.Properties) > 0 {
-			return "object"
-		}
-		if len(d.PrefixItems) > 0 || d.Items != nil {
-			return "array"
-		}
-		return "value"
-	}
-	return d.Type
-}
-
-// props is an ordered list of properties. The order of the properties
-// is the order in which they were defined in the schema.
-type props []*Schema
-
-var _ json.Unmarshaler = (*props)(nil)
-
-func (v *props) UnmarshalJSON(data []byte) error {
-	if len(data) == 0 {
-		return nil
-	}
-	if data[0] != '{' {
-		return errors.New("expected object")
-	}
-
-	d := json.NewDecoder(bytes.NewReader(data))
-
-	// TODO(bmizerany): Consider DisallowUnknownFields. Currently, we, like
-	// llama.cpp, ignore unknown fields, which could be lead to unexpected
-	// behavior for clients of this package, since they may not be aware
-	// that "additionalFields", "itemsPrefix", etc, are being ignored.
-	//
-	// For now, just do what llama.cpp does.
-
-	t, err := d.Token()
-	if err != nil {
-		return err
-	}
-	if t != json.Delim('{') {
-		return errors.New("expected object")
-	}
-	for d.More() {
-		// Use the first token (map key) as the property name, then
-		// decode the rest of the object fields into a Schema and
-		// append.
-		t, err := d.Token()
-		if err != nil {
-			return err
-		}
-		if t == json.Delim('}') {
-			return nil
-		}
-		s := &Schema{
-			Name: t.(string),
-		}
-		if err := d.Decode(s); err != nil {
-			return err
-		}
-		*v = append(*v, s)
-	}
-	return nil
-}

grammar/jsonschema/decode_test.go

@@ -1,104 +0,0 @@
-package jsonschema
-
-import (
-	"encoding/json"
-	"reflect"
-	"strings"
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-)
-
-const testSchemaBasic = `
-{
-  "properties": {
-    "tupleClosedEmpty":   { "prefixItems": [] },
-    "tupleClosedMissing": { "prefixItems": [{}] },
-    "tupleClosedNull":    { "prefixItems": [{}], "items": null },
-    "tupleClosedFalse":   { "prefixItems": [{}], "items": false },
-    "tupleOpenTrue":      { "prefixItems": [{}], "items": true },
-    "tupleOpenEmpty":     { "prefixItems": [{}], "items": {} },
-    "tupleOpenTyped":     { "prefixItems": [{}], "items": {"type": "boolean"} },
-    "tupleOpenMax":       { "prefixItems": [{}], "items": true, "maxItems": 3},
-
-    "array": { "items": {"type": "number"} },
-
-    "null": { "type": "null" },
-    "string": { "type": "string" },
-    "boolean": { "type": "boolean" }
-  }
-}
-`
-
-func TestSchemaUnmarshal(t *testing.T) {
-	var got *Schema
-	if err := json.Unmarshal([]byte(testSchemaBasic), &got); err != nil {
-		t.Fatalf("Unmarshal: %v", err)
-	}
-	want := &Schema{
-		Properties: []*Schema{
-			{Name: "tupleClosedEmpty", PrefixItems: []*Schema{}, Items: nil},
-			{Name: "tupleClosedMissing", PrefixItems: []*Schema{{}}, Items: nil},
-			{Name: "tupleClosedNull", PrefixItems: []*Schema{{}}, Items: nil},
-			{Name: "tupleClosedFalse", PrefixItems: []*Schema{{}}, Items: nil},
-
-			{Name: "tupleOpenTrue", PrefixItems: []*Schema{{}}, Items: &Schema{}},
-			{Name: "tupleOpenEmpty", PrefixItems: []*Schema{{}}, Items: &Schema{}},
-			{Name: "tupleOpenTyped", PrefixItems: []*Schema{{}}, Items: &Schema{Type: "boolean"}},
-			{Name: "tupleOpenMax", PrefixItems: []*Schema{{}}, Items: &Schema{}, MaxItems: 3},
-
-			{Name: "array", Items: &Schema{Type: "number"}},
-
-			{Name: "null", Type: "null"},
-			{Name: "string", Type: "string"},
-			{Name: "boolean", Type: "boolean"},
-		},
-	}
-
-	if diff := cmp.Diff(want, got); diff != "" {
-		t.Errorf("(-want, +got)\n%s", diff)
-	}
-}
-
-func TestEffectiveType(t *testing.T) {
-	const schema = `
-		{"properties": {
-			"o": {"type": "object"},
-			"a": {"type": "array"},
-			"n": {"type": "number"},
-			"s": {"type": "string"},
-			"z": {"type": "null"},
-			"b": {"type": "boolean"},
-
-			"t0": {"prefixItems": [{}], "items": {"type": "number"}},
-			"t1": {"items": {"type": "number"}, "maxItems": 3},
-
-			"v": {"maxItems": 3}
-		}}
-	`
-
-	var s *Schema
-	if err := json.Unmarshal([]byte(schema), &s); err != nil {
-		t.Fatalf("json.Unmarshal: %v", err)
-	}
-
-	var got []string
-	for _, p := range s.Properties {
-		got = append(got, p.EffectiveType())
-	}
-
-	want := strings.Fields(`
-		object
-		array
-		number
-		string
-		null
-		boolean
-		array
-		array
-		value
-	`)
-	if !reflect.DeepEqual(want, got) {
-		t.Errorf("\ngot:\n\t%v\nwant:\n\t%v", got, want)
-	}
-}

grammar/testdata/schemas.txt

@@ -1,76 +0,0 @@
-# This file holds tests for JSON schema to EBNF grammar conversions.
-#
-# The format is a JSON schema, followed by the expected EBNF grammar. Each test
-# MAY be preceded by a comment that describes the test (e.g. the test name), followed by
-# the JSON schema and the expected EBNF grammar. If no comment is present, the test
-# name the tests number in the file (e.g. "#0", "#1", etc.)
-#
-# Blank lines signify the end or start of a new test. Comments can be added
-# anywhere in the file, but they must be preceded by a '#' character and start at
-# the beginning of the line.
-
-# default
-{}
-root ::= value;
-
-{"properties": {}}
-root ::= value;
-
-# array
-{"properties": {"a": {"type": "array", "items": {"type": "string"}}}}
-root_0_tuple_0 ::= string;
-root_0         ::= "[" ( root_0_tuple_0 )* "]";
-root           ::= "{" "a" ":" root_0 "}";
-
-# array with nested array
-{"type": "array", "items": {"type": "array", "items": {"type": "string"}}}
-root_tuple_0_tuple_0 ::= string;
-root_tuple_0         ::= "[" ( root_tuple_0_tuple_0 )* "]";
-root                 ::= "[" ( root_tuple_0 )* "]";
-
-# object
-{"properties": {"e": {}}}
-root_0 ::= value;
-root   ::= "{" "e" ":" root_0 "}";
-
-# object with nested object
-{"properties": {"o": {"type": "object", "properties": {"e": {}}}}}
-root_0_0 ::= value;
-root_0   ::= "{" "e" ":" root_0_0 "}";
-root     ::= "{" "o" ":" root_0 "}";
-
-# boolean
-{"type": "boolean"}
-root ::= boolean;
-
-# number
-{"properties": {"n": {"type": "number", "minimum": 123, "maximum": 4567}}}
-root_0 ::= number;
-root   ::= "{" "n" ":" root_0 "}";
-
-# string
-{"type": "string"}
-root ::= string;
-
-# string with enum
-{"type": "string", "enum": ["a", "b", "c"]}
-root ::= ( "\"a\"" "|" "\"b\"" "|" "\"c\"" );
-
-# spaces in key
-{"properties": {"a b": {}}}
-root_0 ::= value;
-root   ::= "{" "a b" ":" root_0 "}";
-
-# issue7978
-{ "type": "object", "properties": { "steps": { "type": "array", "items": { "type": "object", "properties": { "explanation": { "type": "string" }, "output": { "type": "string" } }, "required": [ "explanation", "output" ], "additionalProperties": false } }, "final_answer": { "type": "string" } }, "required": [ "steps", "final_answer" ], "additionalProperties": false }
-root_0_tuple_0_0 ::= string;
-root_0_tuple_0_1 ::= string;
-root_0_tuple_0   ::= "{" "explanation" ":" root_0_tuple_0_0 "," "output" ":" root_0_tuple_0_1 "}";
-root_0           ::= "[" ( root_0_tuple_0 )* "]";
-root_1           ::= string;
-root             ::= "{" "steps" ":" root_0 "," "final_answer" ":" root_1 "}";
-
-# !! # special characters in key
-# !! {"properties": {"a!b": {}}}
-# !! !invalid character '!' in key
-# !! 

integration/api_test.go

@@ -0,0 +1,412 @@
+//go:build integration
+
+package integration
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"math/rand"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestAPIGenerate(t *testing.T) {
+	initialTimeout := 60 * time.Second
+	streamTimeout := 30 * time.Second
+	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
+	defer cancel()
+	// Set up the test data
+	req := api.GenerateRequest{
+		Model:  smol,
+		Prompt: "why is the sky blue? be brief",
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+	anyResp := []string{"rayleigh", "scattering"}
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+
+	tests := []struct {
+		name   string
+		stream bool
+	}{
+		{
+			name:   "stream",
+			stream: true,
+		},
+		{
+			name:   "no_stream",
+			stream: false,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			stallTimer := time.NewTimer(initialTimeout)
+			var buf bytes.Buffer
+			fn := func(response api.GenerateResponse) error {
+				// Fields that must always be present
+				if response.Model == "" {
+					t.Errorf("response missing model: %#v", response)
+				}
+				if response.Done {
+					// Required fields for final updates:
+					if response.DoneReason == "" && *req.Stream {
+						// TODO - is the lack of done reason on non-stream a bug?
+						t.Errorf("final response missing done_reason: %#v", response)
+					}
+					if response.Metrics.TotalDuration == 0 {
+						t.Errorf("final response missing total_duration: %#v", response)
+					}
+					if response.Metrics.LoadDuration == 0 {
+						t.Errorf("final response missing load_duration: %#v", response)
+					}
+					if response.Metrics.PromptEvalDuration == 0 {
+						t.Errorf("final response missing prompt_eval_duration: %#v", response)
+					}
+					if response.Metrics.EvalCount == 0 {
+						t.Errorf("final response missing eval_count: %#v", response)
+					}
+					if response.Metrics.EvalDuration == 0 {
+						t.Errorf("final response missing eval_duration: %#v", response)
+					}
+					if len(response.Context) == 0 {
+						t.Errorf("final response missing context: %#v", response)
+					}
+
+					// Note: caching can result in no prompt eval count, so this can't be verified reliably
+					// if response.Metrics.PromptEvalCount == 0 {
+					// 	t.Errorf("final response missing prompt_eval_count: %#v", response)
+					// }
+
+				} // else incremental response, nothing to check right now...
+				buf.Write([]byte(response.Response))
+				if !stallTimer.Reset(streamTimeout) {
+					return fmt.Errorf("stall was detected while streaming response, aborting")
+				}
+				return nil
+			}
+
+			done := make(chan int)
+			var genErr error
+			go func() {
+				req.Stream = &test.stream
+				req.Options["seed"] = rand.Int() // bust cache for prompt eval results
+				genErr = client.Generate(ctx, &req, fn)
+				done <- 0
+			}()
+
+			select {
+			case <-stallTimer.C:
+				if buf.Len() == 0 {
+					t.Errorf("generate never started.  Timed out after :%s", initialTimeout.String())
+				} else {
+					t.Errorf("generate stalled.  Response so far:%s", buf.String())
+				}
+			case <-done:
+				if genErr != nil {
+					t.Fatalf("failed with %s request prompt %s ", req.Model, req.Prompt)
+				}
+				// Verify the response contains the expected data
+				response := buf.String()
+				atLeastOne := false
+				for _, resp := range anyResp {
+					if strings.Contains(strings.ToLower(response), resp) {
+						atLeastOne = true
+						break
+					}
+				}
+				if !atLeastOne {
+					t.Errorf("none of %v found in %s", anyResp, response)
+				}
+			case <-ctx.Done():
+				t.Error("outer test context done while waiting for generate")
+			}
+		})
+	}
+
+	// Validate PS while we're at it...
+	resp, err := client.ListRunning(ctx)
+	if err != nil {
+		t.Fatalf("list models API error: %s", err)
+	}
+	if resp == nil || len(resp.Models) == 0 {
+		t.Fatalf("list models API returned empty list while model should still be loaded")
+	}
+	// Find the model we just loaded and verify some attributes
+	found := false
+	for _, model := range resp.Models {
+		if strings.Contains(model.Name, req.Model) {
+			found = true
+			if model.Model == "" {
+				t.Errorf("model field omitted: %#v", model)
+			}
+			if model.Size == 0 {
+				t.Errorf("size omitted: %#v", model)
+			}
+			if model.Digest == "" {
+				t.Errorf("digest omitted: %#v", model)
+			}
+			verifyModelDetails(t, model.Details)
+			var nilTime time.Time
+			if model.ExpiresAt == nilTime {
+				t.Errorf("expires_at omitted: %#v", model)
+			}
+			// SizeVRAM could be zero.
+		}
+	}
+	if !found {
+		t.Errorf("unable to locate running model: %#v", resp)
+	}
+}
+
+func TestAPIChat(t *testing.T) {
+	initialTimeout := 60 * time.Second
+	streamTimeout := 30 * time.Second
+	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
+	defer cancel()
+	// Set up the test data
+	req := api.ChatRequest{
+		Model: smol,
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: "why is the sky blue?  be brief",
+			},
+		},
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+	anyResp := []string{"rayleigh", "scattering"}
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+
+	tests := []struct {
+		name   string
+		stream bool
+	}{
+		{
+			name:   "stream",
+			stream: true,
+		},
+		{
+			name:   "no_stream",
+			stream: false,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			stallTimer := time.NewTimer(initialTimeout)
+			var buf bytes.Buffer
+			fn := func(response api.ChatResponse) error {
+				// Fields that must always be present
+				if response.Model == "" {
+					t.Errorf("response missing model: %#v", response)
+				}
+				if response.Done {
+					// Required fields for final updates:
+					var nilTime time.Time
+					if response.CreatedAt == nilTime {
+						t.Errorf("final response missing total_duration: %#v", response)
+					}
+					if response.DoneReason == "" {
+						t.Errorf("final response missing done_reason: %#v", response)
+					}
+					if response.Metrics.TotalDuration == 0 {
+						t.Errorf("final response missing total_duration: %#v", response)
+					}
+					if response.Metrics.LoadDuration == 0 {
+						t.Errorf("final response missing load_duration: %#v", response)
+					}
+					if response.Metrics.PromptEvalDuration == 0 {
+						t.Errorf("final response missing prompt_eval_duration: %#v", response)
+					}
+					if response.Metrics.EvalCount == 0 {
+						t.Errorf("final response missing eval_count: %#v", response)
+					}
+					if response.Metrics.EvalDuration == 0 {
+						t.Errorf("final response missing eval_duration: %#v", response)
+					}
+
+					if response.Metrics.PromptEvalCount == 0 {
+						t.Errorf("final response missing prompt_eval_count: %#v", response)
+					}
+				} // else incremental response, nothing to check right now...
+				buf.Write([]byte(response.Message.Content))
+				if !stallTimer.Reset(streamTimeout) {
+					return fmt.Errorf("stall was detected while streaming response, aborting")
+				}
+				return nil
+			}
+
+			done := make(chan int)
+			var genErr error
+			go func() {
+				req.Stream = &test.stream
+				req.Options["seed"] = rand.Int() // bust cache for prompt eval results
+				genErr = client.Chat(ctx, &req, fn)
+				done <- 0
+			}()
+
+			select {
+			case <-stallTimer.C:
+				if buf.Len() == 0 {
+					t.Errorf("chat never started.  Timed out after :%s", initialTimeout.String())
+				} else {
+					t.Errorf("chat stalled.  Response so far:%s", buf.String())
+				}
+			case <-done:
+				if genErr != nil {
+					t.Fatalf("failed with %s request prompt %v", req.Model, req.Messages)
+				}
+				// Verify the response contains the expected data
+				response := buf.String()
+				atLeastOne := false
+				for _, resp := range anyResp {
+					if strings.Contains(strings.ToLower(response), resp) {
+						atLeastOne = true
+						break
+					}
+				}
+				if !atLeastOne {
+					t.Errorf("none of %v found in %s", anyResp, response)
+				}
+			case <-ctx.Done():
+				t.Error("outer test context done while waiting for chat")
+			}
+		})
+	}
+}
+
+func TestAPIListModels(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	// Make sure we have at least one model so an empty list can be considered a failure
+	if err := PullIfMissing(ctx, client, smol); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+
+	resp, err := client.List(ctx)
+	if err != nil {
+		t.Fatalf("unable to list models: %s", err)
+	}
+	if len(resp.Models) == 0 {
+		t.Fatalf("list should not be empty")
+	}
+	model := resp.Models[0]
+	if model.Name == "" {
+		t.Errorf("first model name empty: %#v", model)
+	}
+	var nilTime time.Time
+	if model.ModifiedAt == nilTime {
+		t.Errorf("first model modified_at empty: %#v", model)
+	}
+	if model.Size == 0 {
+		t.Errorf("first model size empty: %#v", model)
+	}
+	if model.Digest == "" {
+		t.Errorf("first model digest empty: %#v", model)
+	}
+	verifyModelDetails(t, model.Details)
+}
+
+func verifyModelDetails(t *testing.T, details api.ModelDetails) {
+	if details.Format == "" {
+		t.Errorf("first model details.format empty: %#v", details)
+	}
+	if details.Family == "" {
+		t.Errorf("first model details.family empty: %#v", details)
+	}
+	if details.ParameterSize == "" {
+		t.Errorf("first model details.parameter_size empty: %#v", details)
+	}
+	if details.QuantizationLevel == "" {
+		t.Errorf("first model details.quantization_level empty: %#v", details)
+	}
+}
+
+func TestAPIShowModel(t *testing.T) {
+	modelName := "llama3.2"
+	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	if err := PullIfMissing(ctx, client, modelName); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+	resp, err := client.Show(ctx, &api.ShowRequest{Name: modelName})
+	if err != nil {
+		t.Fatalf("unable to show model: %s", err)
+	}
+	if resp.License == "" {
+		t.Errorf("%s missing license: %#v", modelName, resp)
+	}
+	if resp.Modelfile == "" {
+		t.Errorf("%s missing modelfile: %#v", modelName, resp)
+	}
+	if resp.Parameters == "" {
+		t.Errorf("%s missing parameters: %#v", modelName, resp)
+	}
+	if resp.Template == "" {
+		t.Errorf("%s missing template: %#v", modelName, resp)
+	}
+	// llama3 omits system
+	verifyModelDetails(t, resp.Details)
+	// llama3 ommits messages
+	if len(resp.ModelInfo) == 0 {
+		t.Errorf("%s missing model_info: %#v", modelName, resp)
+	}
+	// llama3 omits projectors
+	var nilTime time.Time
+	if resp.ModifiedAt == nilTime {
+		t.Errorf("%s missing modified_at: %#v", modelName, resp)
+	}
+}
+
+func TestAPIEmbeddings(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	req := api.EmbeddingRequest{
+		Model:  "orca-mini",
+		Prompt: "why is the sky blue?",
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+
+	resp, err := client.Embeddings(ctx, &req)
+	if err != nil {
+		t.Fatalf("embeddings call failed %s", err)
+	}
+	if len(resp.Embedding) == 0 {
+		t.Errorf("zero length embedding response")
+	}
+}

integration/basic_test.go

@@ -14,15 +14,15 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-func TestOrcaMiniBlueSky(t *testing.T) {
+func TestBlueSky(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{
-		Model:  "orca-mini",
+		Model:  smol,
 		Prompt: "why is the sky blue?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},
@@ -31,6 +31,7 @@ func TestOrcaMiniBlueSky(t *testing.T) {
 }
 
 func TestUnicode(t *testing.T) {
+	skipUnderMinVRAM(t, 6)
 	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
 	defer cancel()
 	// Set up the test data
@@ -39,7 +40,7 @@ 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 +62,7 @@ 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,
 		},
@@ -93,10 +94,10 @@ func TestUnicodeModelDir(t *testing.T) {
 	defer cancel()
 
 	req := api.GenerateRequest{
-		Model:  "orca-mini",
+		Model:  smol,
 		Prompt: "why is the sky blue?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},

integration/concurrency_test.go

@@ -21,11 +21,11 @@ func TestMultiModelConcurrency(t *testing.T) {
 	var (
 		req = [2]api.GenerateRequest{
 			{
-				Model:     "orca-mini",
+				Model:     "llama3.2:1b",
 				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,
 				},
@@ -67,7 +67,7 @@ func TestMultiModelConcurrency(t *testing.T) {
 	wg.Wait()
 }
 
-func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
+func TestIntegrationConcurrentPredict(t *testing.T) {
 	req, resp := GenerateRequests()
 	reqLimit := len(req)
 	iterLimit := 5
@@ -117,6 +117,9 @@ func TestMultiModelStress(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
+	if maxVram < 2*format.GibiByte {
+		t.Skip("VRAM less than 2G, skipping model stress tests")
+	}
 
 	type model struct {
 		name string
@@ -125,8 +128,8 @@ func TestMultiModelStress(t *testing.T) {
 
 	smallModels := []model{
 		{
-			name: "orca-mini",
-			size: 2992 * format.MebiByte,
+			name: "llama3.2:1b",
+			size: 2876 * format.MebiByte,
 		},
 		{
 			name: "phi",

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

@@ -12,58 +12,51 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-func TestIntegrationLlava(t *testing.T) {
-	image, err := base64.StdEncoding.DecodeString(imageEncoding)
-	require.NoError(t, err)
-	req := api.GenerateRequest{
-		Model:  "llava:7b",
-		Prompt: "what does the text in this image say?",
-		Stream: &stream,
-		Options: map[string]interface{}{
-			"seed":        42,
-			"temperature": 0.0,
+func TestVisionModels(t *testing.T) {
+	skipUnderMinVRAM(t, 6)
+	type testCase struct {
+		model string
+	}
+	testCases := []testCase{
+		{
+			model: "llava:7b",
 		},
-		Images: []api.ImageData{
-			image,
+		{
+			model: "llama3.2-vision",
+		},
+		{
+			model: "gemma3",
 		},
 	}
 
-	// Note: sometimes it returns "the ollamas" sometimes "the ollams"
-	resp := "the ollam"
-	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-	require.NoError(t, PullIfMissing(ctx, client, req.Model))
-	// llava models on CPU can be quite slow to start,
-	DoGenerate(ctx, t, client, req, []string{resp}, 120*time.Second, 30*time.Second)
-}
+	for _, v := range testCases {
+		t.Run(v.model, func(t *testing.T) {
+			image, err := base64.StdEncoding.DecodeString(imageEncoding)
+			require.NoError(t, err)
+			req := api.GenerateRequest{
+				Model:  v.model,
+				Prompt: "what does the text in this image say?",
+				Stream: &stream,
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+				Images: []api.ImageData{
+					image,
+				},
+			}
+			ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
+			defer cancel()
+			client, _, cleanup := InitServerConnection(ctx, t)
 
-func TestIntegrationMllama(t *testing.T) {
-	image, err := base64.StdEncoding.DecodeString(imageEncoding)
-	require.NoError(t, err)
-	req := api.GenerateRequest{
-		// TODO fix up once we publish the final image
-		Model:  "x/llama3.2-vision",
-		Prompt: "what does the text in this image say?",
-		Stream: &stream,
-		Options: map[string]interface{}{
-			"seed":        42,
-			"temperature": 0.0,
-		},
-		Images: []api.ImageData{
-			image,
-		},
+			// Note: sometimes it returns "the ollamas" sometimes "the ollams"
+			resp := "the ollam"
+			defer cleanup()
+			require.NoError(t, PullIfMissing(ctx, client, req.Model))
+			// llava models on CPU can be quite slow to start
+			DoGenerate(ctx, t, client, req, []string{resp}, 240*time.Second, 30*time.Second)
+		})
 	}
-
-	resp := "the ollamas"
-	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-	require.NoError(t, PullIfMissing(ctx, client, req.Model))
-	// mllama models on CPU can be quite slow to start,
-	DoGenerate(ctx, t, client, req, []string{resp}, 240*time.Second, 30*time.Second)
 }
 
 func TestIntegrationSplitBatch(t *testing.T) {
@@ -75,7 +68,7 @@ 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

@@ -17,30 +17,30 @@ var (
 	stream = false
 	req    = [2]api.GenerateRequest{
 		{
-			Model:  "orca-mini",
+			Model:  smol,
 			Prompt: "why is the ocean blue?",
 			Stream: &stream,
-			Options: map[string]interface{}{
+			Options: map[string]any{
 				"seed":        42,
 				"temperature": 0.0,
 			},
 		}, {
-			Model:  "orca-mini",
+			Model:  smol,
 			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,
 			},
 		},
 	}
 	resp = [2][]string{
-		{"sunlight"},
+		{"sunlight", "scattering", "interact"},
 		{"england", "english", "massachusetts", "pilgrims"},
 	}
 )
 
-func TestIntegrationSimpleOrcaMini(t *testing.T) {
+func TestIntegrationSimple(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
 	defer cancel()
 	GenerateTestHelper(ctx, t, req[0], resp[0])

integration/max_queue_test.go

@@ -30,9 +30,9 @@ func TestMaxQueue(t *testing.T) {
 	t.Setenv("OLLAMA_MAX_QUEUE", strconv.Itoa(threadCount))
 
 	req := api.GenerateRequest{
-		Model:  "orca-mini",
+		Model:  smol,
 		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,
 		},
@@ -52,8 +52,8 @@ func TestMaxQueue(t *testing.T) {
 	embedCtx := ctx
 
 	var genwg sync.WaitGroup
+	genwg.Add(1)
 	go func() {
-		genwg.Add(1)
 		defer genwg.Done()
 		slog.Info("Starting generate request")
 		DoGenerate(ctx, t, client, req, resp, 45*time.Second, 5*time.Second)
@@ -61,7 +61,7 @@ func TestMaxQueue(t *testing.T) {
 	}()
 
 	// Give the generate a chance to get started before we start hammering on embed requests
-	time.Sleep(5 * time.Millisecond)
+	time.Sleep(10 * time.Millisecond)
 
 	threadCount += 10 // Add a few extra to ensure we push the queue past its limit
 	busyCount := 0
@@ -71,8 +71,8 @@ func TestMaxQueue(t *testing.T) {
 	counterMu := sync.Mutex{}
 	var embedwg sync.WaitGroup
 	for i := 0; i < threadCount; i++ {
+		embedwg.Add(1)
 		go func(i int) {
-			embedwg.Add(1)
 			defer embedwg.Done()
 			slog.Info("embed started", "id", i)
 			embedReq := api.EmbeddingRequest{

integration/model_arch_test.go

@@ -0,0 +1,195 @@
+//go:build integration && models
+
+package integration
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io/ioutil"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/format"
+)
+
+var (
+	started    = time.Now()
+	chatModels = []string{
+		"granite3-moe:latest",
+		"granite-code:latest",
+		"nemotron-mini:latest",
+		"command-r:latest",
+		"gemma2:latest",
+		"gemma:latest",
+		"internlm2:latest",
+		"phi3.5:latest",
+		"phi3:latest",
+		// "phi:latest", // flaky, sometimes generates no response on first query
+		"stablelm2:latest", // Predictions are off, crashes on small VRAM GPUs
+		"falcon:latest",
+		"falcon2:latest",
+		"minicpm-v:latest",
+		"mistral:latest",
+		"orca-mini:latest",
+		"llama2:latest",
+		"llama3.1:latest",
+		"llama3.2:latest",
+		"llama3.2-vision:latest",
+		"qwen2.5-coder:latest",
+		"qwen:latest",
+		"solar-pro:latest",
+	}
+)
+
+func getTimeouts(t *testing.T) (soft time.Duration, hard time.Duration) {
+	deadline, hasDeadline := t.Deadline()
+	if !hasDeadline {
+		return 8 * time.Minute, 10 * time.Minute
+	} else if deadline.Compare(time.Now().Add(2*time.Minute)) <= 0 {
+		t.Skip("too little time")
+		return time.Duration(0), time.Duration(0)
+	}
+	return -time.Since(deadline.Add(-2 * time.Minute)), -time.Since(deadline.Add(-20 * time.Second))
+}
+
+func TestModelsGenerate(t *testing.T) {
+	softTimeout, hardTimeout := getTimeouts(t)
+	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
+	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	// TODO use info API eventually
+	var maxVram uint64
+	var err error
+	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
+		maxVram, err = strconv.ParseUint(s, 10, 64)
+		if err != nil {
+			t.Fatalf("invalid  OLLAMA_MAX_VRAM %v", err)
+		}
+	} else {
+		slog.Warn("No VRAM info available, testing all models, so larger ones might timeout...")
+	}
+
+	for _, model := range chatModels {
+		t.Run(model, func(t *testing.T) {
+			if time.Now().Sub(started) > softTimeout {
+				t.Skip("skipping remaining tests to avoid excessive runtime")
+			}
+			if err := PullIfMissing(ctx, client, model); err != nil {
+				t.Fatalf("pull failed %s", err)
+			}
+			if maxVram > 0 {
+				resp, err := client.List(ctx)
+				if err != nil {
+					t.Fatalf("list models failed %v", err)
+				}
+				for _, m := range resp.Models {
+					if m.Name == model && float32(m.Size)*1.2 > float32(maxVram) {
+						t.Skipf("model %s is too large for available VRAM: %s > %s", model, format.HumanBytes(m.Size), format.HumanBytes(int64(maxVram)))
+					}
+				}
+			}
+			// TODO - fiddle with context size
+			req := api.GenerateRequest{
+				Model:  model,
+				Prompt: "why is the sky blue?",
+				Options: map[string]interface{}{
+					"temperature": 0,
+					"seed":        123,
+				},
+			}
+			anyResp := []string{"rayleigh", "scattering", "atmosphere", "nitrogen", "oxygen"}
+			DoGenerate(ctx, t, client, req, anyResp, 120*time.Second, 30*time.Second)
+		})
+	}
+}
+
+func TestModelsEmbed(t *testing.T) {
+	softTimeout, hardTimeout := getTimeouts(t)
+	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	// TODO use info API eventually
+	var maxVram uint64
+	var err error
+	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
+		maxVram, err = strconv.ParseUint(s, 10, 64)
+		if err != nil {
+			t.Fatalf("invalid  OLLAMA_MAX_VRAM %v", err)
+		}
+	} else {
+		slog.Warn("No VRAM info available, testing all models, so larger ones might timeout...")
+	}
+
+	data, err := ioutil.ReadFile(filepath.Join("testdata", "embed.json"))
+	if err != nil {
+		t.Fatalf("failed to open test data file: %s", err)
+	}
+	testCase := map[string][]float64{}
+	err = json.Unmarshal(data, &testCase)
+	if err != nil {
+		t.Fatalf("failed to load test data: %s", err)
+	}
+	for model, expected := range testCase {
+
+		t.Run(model, func(t *testing.T) {
+			if time.Now().Sub(started) > softTimeout {
+				t.Skip("skipping remaining tests to avoid excessive runtime")
+			}
+			if err := PullIfMissing(ctx, client, model); err != nil {
+				t.Fatalf("pull failed %s", err)
+			}
+			if maxVram > 0 {
+				resp, err := client.List(ctx)
+				if err != nil {
+					t.Fatalf("list models failed %v", err)
+				}
+				for _, m := range resp.Models {
+					if m.Name == model && float32(m.Size)*1.2 > float32(maxVram) {
+						t.Skipf("model %s is too large for available VRAM: %s > %s", model, format.HumanBytes(m.Size), format.HumanBytes(int64(maxVram)))
+					}
+				}
+			}
+			req := api.EmbeddingRequest{
+				Model:  model,
+				Prompt: "why is the sky blue?",
+				Options: map[string]interface{}{
+					"temperature": 0,
+					"seed":        123,
+				},
+			}
+			resp, err := client.Embeddings(ctx, &req)
+			if err != nil {
+				t.Fatalf("embeddings call failed %s", err)
+			}
+			if len(resp.Embedding) == 0 {
+				t.Errorf("zero length embedding response")
+			}
+			if len(expected) != len(resp.Embedding) {
+				expStr := make([]string, len(resp.Embedding))
+				for i, v := range resp.Embedding {
+					expStr[i] = fmt.Sprintf("%0.6f", v)
+				}
+				// When adding new models, use this output to populate the testdata/embed.json
+				fmt.Printf("expected\n%s\n", strings.Join(expStr, ", "))
+				t.Fatalf("expected %d, got %d", len(expected), len(resp.Embedding))
+			}
+			sim := cosineSimilarity(resp.Embedding, expected)
+			if sim < 0.99 {
+				t.Fatalf("expected %v, got %v (similarity: %f)", expected[0:5], resp.Embedding[0:5], sim)
+			}
+		})
+	}
+
+}

integration/utils_test.go

@@ -24,9 +24,14 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/app/lifecycle"
+	"github.com/ollama/ollama/format"
 	"github.com/stretchr/testify/require"
 )
 
+const (
+	smol = "llama3.2:1b"
+)
+
 func Init() {
 	lifecycle.InitLogging()
 }
@@ -140,7 +145,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 
 	showCtx, cancel := context.WithDeadlineCause(
 		ctx,
-		time.Now().Add(10*time.Second),
+		time.Now().Add(20*time.Second),
 		fmt.Errorf("show for existing model %s took too long", modelName),
 	)
 	defer cancel()
@@ -157,7 +162,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 	}
 	slog.Info("model missing", "model", modelName)
 
-	stallDuration := 30 * time.Second // This includes checksum verification, which can take a while on larger models
+	stallDuration := 60 * time.Second // This includes checksum verification, which can take a while on larger models, and slower systems
 	stallTimer := time.NewTimer(stallDuration)
 	fn := func(resp api.ProgressResponse) error {
 		// fmt.Print(".")
@@ -283,51 +288,51 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 }
 
 // Generate a set of requests
-// By default each request uses orca-mini as the model
+// By default each request uses llama3.2 as the model
 func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 	return []api.GenerateRequest{
 			{
-				Model:     "orca-mini",
+				Model:     smol,
 				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,
 				},
 			}, {
-				Model:     "orca-mini",
+				Model:     smol,
 				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,
 				},
 			}, {
-				Model:     "orca-mini",
+				Model:     smol,
 				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,
 				},
 			}, {
-				Model:     "orca-mini",
+				Model:     smol,
 				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,
 				},
 			}, {
-				Model:     "orca-mini",
+				Model:     smol,
 				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,
 				},
@@ -341,3 +346,15 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 			{"nitrogen", "oxygen", "carbon", "dioxide"},
 		}
 }
+
+func skipUnderMinVRAM(t *testing.T, gb uint64) {
+	// TODO use info API in the future
+	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
+		maxVram, err := strconv.ParseUint(s, 10, 64)
+		require.NoError(t, err)
+		// Don't hammer on small VRAM cards...
+		if maxVram < gb*format.GibiByte {
+			t.Skip("skipping with small VRAM to avoid timeouts")
+		}
+	}
+}

kvcache/cache.go

@@ -56,12 +56,18 @@ type Cache interface {
 
 	// StartForward is called before the start of the model's forward pass.
 	// For each token in the coming batch, there must be a corresponding
-	// entry in positions and seqs.
-	StartForward(ctx ml.Context, batch input.Batch) error
+	// entry in positions and seqs. reserve is to preallocate memory
+	// without actually storing data in the cache.
+	StartForward(ctx ml.Context, batch input.Batch, reserve bool) error
 
 	// 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

@@ -21,6 +21,7 @@ type shiftFn func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, e
 type Causal struct {
 	DType      ml.DType
 	windowSize int32
+	chunkSize  int32
 
 	opts CausalOptions
 
@@ -97,6 +98,17 @@ func NewSWACache(windowSize int32, shift shiftFn) *Causal {
 	}
 }
 
+func NewChunkedAttentionCache(chunkSize int32, shift shiftFn) *Causal {
+	return &Causal{
+		windowSize: math.MaxInt32,
+		chunkSize:  chunkSize,
+		shiftFn:    shift,
+		ctxs:       make(map[int]ml.Context),
+		keys:       make(map[int]ml.Tensor),
+		values:     make(map[int]ml.Tensor),
+	}
+}
+
 func (c *Causal) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity, maxBatch int) {
 	if c.config == nil {
 		var config ml.CacheConfig
@@ -119,10 +131,10 @@ 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)
@@ -146,51 +158,60 @@ func (c *Causal) Close() {
 	}
 }
 
-func (c *Causal) StartForward(ctx ml.Context, batch input.Batch) error {
+func (c *Causal) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
 	c.curBatchSize = len(batch.Positions)
 	c.curSequences = batch.Sequences
 	c.curPositions = batch.Positions
 	c.opts.Except = nil
 
-	c.updateSlidingWindow()
+	if !reserve {
+		c.updateSlidingWindow()
+
+		var err error
+		c.curLoc, err = c.findStartLoc()
+		if errors.Is(err, ErrKvCacheFull) {
+			c.defrag()
+			c.curLoc, err = c.findStartLoc()
+		}
+		if err != nil {
+			return err
+		}
+
+		c.curCellRange = newRange()
+		for i, pos := range batch.Positions {
+			seq := batch.Sequences[i]
+
+			c.cells[c.curLoc+i] = cacheCell{pos: pos, sequences: []int{seq}}
+
+			seqRange, ok := c.cellRanges[seq]
+			if !ok {
+				seqRange = newRange()
+			}
+
+			if c.curLoc+i > seqRange.max {
+				seqRange.max = c.curLoc + i
+			}
+			if seqRange.max > c.curCellRange.max {
+				c.curCellRange.max = seqRange.max
+			}
+
+			if c.curLoc+i < seqRange.min {
+				seqRange.min = c.curLoc + i
+			}
+			if seqRange.min < c.curCellRange.min {
+				c.curCellRange.min = seqRange.min
+			}
+			c.cellRanges[seq] = seqRange
+		}
+	} else {
+		// If we are reserving memory, don't update any of the cache metadata but set the size
+		// to the worst case.
+		c.curLoc = 0
+		c.curCellRange.min = 0
+		c.curCellRange.max = len(c.cells) - 1
+	}
 
 	var err error
-	c.curLoc, err = c.findStartLoc()
-	if errors.Is(err, ErrKvCacheFull) {
-		c.defrag()
-		c.curLoc, err = c.findStartLoc()
-	}
-	if err != nil {
-		return err
-	}
-
-	c.curCellRange = newRange()
-	for i, pos := range batch.Positions {
-		seq := batch.Sequences[i]
-
-		c.cells[c.curLoc+i] = cacheCell{pos: pos, sequences: []int{seq}}
-
-		seqRange, ok := c.cellRanges[seq]
-		if !ok {
-			seqRange = newRange()
-		}
-
-		if c.curLoc+i > seqRange.max {
-			seqRange.max = c.curLoc + i
-		}
-		if seqRange.max > c.curCellRange.max {
-			c.curCellRange.max = seqRange.max
-		}
-
-		if c.curLoc+i < seqRange.min {
-			seqRange.min = c.curLoc + i
-		}
-		if seqRange.min < c.curCellRange.min {
-			c.curCellRange.min = seqRange.min
-		}
-		c.cellRanges[seq] = seqRange
-	}
-
 	c.curMask, err = c.buildMask(ctx)
 
 	return err
@@ -291,6 +312,7 @@ func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
 		for j := c.curCellRange.min; j <= c.curCellRange.max; j++ {
 			if !slices.Contains(c.cells[j].sequences, c.curSequences[i]) ||
 				(enabled && c.cells[j].pos > c.curPositions[i]) ||
+				c.chunkSize > 0 && c.cells[j].pos < c.curPositions[i]-c.curPositions[i]%c.chunkSize ||
 				c.cells[j].pos < c.curPositions[i]-c.windowSize {
 				mask[i*length+(j-c.curCellRange.min)] = float32(math.Inf(-1))
 			}
@@ -581,6 +603,35 @@ 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 +686,12 @@ 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 +706,7 @@ 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

@@ -86,6 +86,64 @@ func TestSWA(t *testing.T) {
 	testCache(t, backend, cache, tests)
 }
 
+func TestChunkedAttention(t *testing.T) {
+	cache := NewChunkedAttentionCache(2, nil)
+	defer cache.Close()
+
+	var b testBackend
+	cache.Init(&b, ml.DTypeF16, 1, 16, 16)
+
+	x := float32(math.Inf(-1))
+
+	testCache(
+		t, &b, cache,
+		[]testCase{
+			{
+				name:          "FirstBatch",
+				in:            []float32{1, 2, 3, 4},
+				inShape:       []int{1, 1, 4},
+				seqs:          []int{0, 0, 0, 0},
+				pos:           []int32{0, 1, 2, 3},
+				expected:      []float32{1, 2, 3, 4},
+				expectedShape: []int{1, 1, 4},
+				expectedMask: []float32{
+					0, x, x, x,
+					0, 0, x, x,
+					x, x, 0, x,
+					x, x, 0, 0,
+				},
+			},
+			{
+				name:          "SecondBatch",
+				in:            []float32{5, 6, 7},
+				inShape:       []int{1, 1, 3},
+				seqs:          []int{0, 0, 0},
+				pos:           []int32{4, 5, 6},
+				expected:      []float32{1, 2, 3, 4, 5, 6, 7},
+				expectedShape: []int{1, 1, 7},
+				expectedMask: []float32{
+					x, x, x, x, 0, x, x,
+					x, x, x, x, 0, 0, x,
+					x, x, x, x, x, x, 0,
+				},
+			},
+			{
+				name:          "ThirdBatch",
+				in:            []float32{8, 9},
+				inShape:       []int{1, 1, 2},
+				seqs:          []int{0, 0},
+				pos:           []int32{7, 8},
+				expected:      []float32{1, 2, 3, 4, 5, 6, 7, 8, 9},
+				expectedShape: []int{1, 1, 9},
+				expectedMask: []float32{
+					x, x, x, x, x, x, 0, 0, x,
+					x, x, x, x, x, x, x, x, 0,
+				},
+			},
+		},
+	)
+}
+
 func TestSequences(t *testing.T) {
 	backend := &testBackend{}
 	cache := NewCausalCache(nil)
@@ -280,7 +338,7 @@ func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase)
 			context := backend.NewContext()
 			defer context.Close()
 
-			err := cache.StartForward(context, input.Batch{Positions: test.pos, Sequences: test.seqs})
+			err := cache.StartForward(context, input.Batch{Positions: test.pos, Sequences: test.seqs}, false)
 			if err != nil {
 				panic(err)
 			}
@@ -293,21 +351,94 @@ func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase)
 
 			context.Forward(out, mask).Compute(out, mask)
 
-			if !slices.Equal(out.Floats(), test.expected) || !slices.Equal(out.Shape(), test.expectedShape) || !slices.Equal(mask.Floats(), test.expectedMask) {
-				t.Errorf("TestCache: have %v (shape %v); want %v (shape %v); mask: have %v (shape %v) want %v", out.Floats(), out.Shape(), test.expected, test.expectedShape, mask.Floats(), mask.Shape(), test.expectedMask)
+			if !slices.Equal(out.Floats(), test.expected) {
+				t.Errorf("TestCache: have %v; want %v", out.Floats(), test.expected)
+			}
+
+			if !slices.Equal(out.Shape(), test.expectedShape) {
+				t.Errorf("TestCache: has shape %v; want %v", out.Shape(), test.expectedShape)
+			}
+
+			if !slices.Equal(mask.Floats(), test.expectedMask) {
+				t.Errorf("TestCache: have mask: have %v want %v", mask.Floats(), test.expectedMask)
 			}
 		})
 	}
 }
 
-type testBackend struct{}
+func TestCanResume(t *testing.T) {
+	backend := &testBackend{}
+	windowSize := int32(4)
+	cache := NewSWACache(windowSize, nil)
+	defer cache.Close()
 
-func (b *testBackend) Config() ml.Config {
-	panic("not implemented")
+	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},
+	}, false)
+	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},
+	}, false)
+	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)")
+	}
 }
 
-func (b *testBackend) Get(name string) ml.Tensor {
-	panic("not implemented")
+type testBackend struct {
+	ml.Backend
 }
 
 func (b *testBackend) NewContext() ml.Context {
@@ -318,12 +449,10 @@ func (b *testBackend) NewContextSize(int) ml.Context {
 	return &testContext{}
 }
 
-func (b *testBackend) SystemInfo() string {
-	return "not implemented"
+type testContext struct {
+	ml.Context
 }
 
-type testContext struct{}
-
 func (c *testContext) Empty(dtype ml.DType, shape ...int) ml.Tensor {
 	total := 0
 
@@ -361,14 +490,26 @@ func (c *testContext) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
 	return out, nil
 }
 
+func (c *testContext) Arange(start, stop, step float32, dtype ml.DType) ml.Tensor {
+	s := make([]float32, 0, int((stop-start)/step))
+	for i := start; i < stop; i += step {
+		s = append(s, i)
+	}
+
+	out, _ := c.FromFloatSlice(s, len(s))
+	out.(*testTensor).dtype = dtype
+	return out
+}
+
 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 }
 
 func (c *testContext) Compute(...ml.Tensor) {}
 
+func (c *testContext) Reserve() error { return nil }
+
 func (c *testContext) MaxGraphNodes() int {
 	return 10
 }
@@ -376,6 +517,8 @@ func (c *testContext) MaxGraphNodes() int {
 func (c *testContext) Close() {}
 
 type testTensor struct {
+	ml.Tensor
+
 	dtype       ml.DType
 	elementSize int
 	data        []float32
@@ -403,16 +546,20 @@ func (t *testTensor) DType() ml.DType {
 	return t.dtype
 }
 
-func (t *testTensor) Bytes() []byte {
-	panic("not implemented")
-}
-
 func (t *testTensor) Floats() []float32 {
 	out := make([]float32, len(t.data))
 	copy(out, t.data)
 	return out
 }
 
+func (t *testTensor) Neg(ctx ml.Context) ml.Tensor {
+	out := ctx.Empty(t.DType(), t.Shape()...).(*testTensor)
+	for i := range out.data {
+		out.data[i] = -t.data[i]
+	}
+	return out
+}
+
 func (t *testTensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	out := ctx.Empty(t.DType(), t.Shape()...).(*testTensor)
 
@@ -423,66 +570,6 @@ func (t *testTensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	return out
 }
 
-func (t *testTensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Mulmat(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) MulmatFullPrec(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Softmax(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) LayerNorm(ctx ml.Context, weight, bias ml.Tensor, eps float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) RMSNorm(ctx ml.Context, weight ml.Tensor, eps float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Scale(ctx ml.Context, s float64) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) AvgPool1D(ctx ml.Context, k, s, p int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Conv2D(ctx ml.Context, weight ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, dim, ropeType uint32, base, scale float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Tanh(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) GELU(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) SILU(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Reshape(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
 func (t *testTensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	offset /= t.elementSize
 
@@ -505,38 +592,6 @@ func (t *testTensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	return view
 }
 
-func (t *testTensor) Permute(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Contiguous(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Set(ctx ml.Context, t2 ml.Tensor, offset int, strides ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Unpad(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Stack(ctx ml.Context, dim int, s ...ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Concat(ctx ml.Context, t2 ml.Tensor, dim int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Rows(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
 func (t *testTensor) Copy(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	copy(t2.(*testTensor).data, t.data)
 	return nil

kvcache/encoder.go

@@ -27,6 +27,11 @@ type EncoderCache struct {
 	// anything will be stored)
 	curPos int32
 
+	// curReserve indicates that this forward pass is only for
+	// memory reservation and we should not update our metadata
+	// based on it.
+	curReserve bool
+
 	// ** cache metadata **
 
 	// was something stored in the cache?
@@ -83,12 +88,14 @@ func (c *EncoderCache) Close() {
 	}
 }
 
-func (c *EncoderCache) StartForward(ctx ml.Context, batch input.Batch) error {
+func (c *EncoderCache) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
 	// We work with the most recent image
 	if len(batch.Multimodal) > 0 {
 		c.curPos = batch.Positions[batch.Multimodal[len(batch.Multimodal)-1].Index]
 	}
 
+	c.curReserve = reserve
+
 	return nil
 }
 
@@ -105,8 +112,10 @@ func (c *EncoderCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
 }
 
 func (c *EncoderCache) Put(ctx ml.Context, key, value ml.Tensor) {
-	c.encoderPos = c.curPos
-	c.encoderCached = true
+	if !c.curReserve {
+		c.encoderPos = c.curPos
+		c.encoderCached = true
+	}
 
 	if c.config.PermutedV {
 		value = value.Permute(ctx, 1, 2, 0, 3)
@@ -134,6 +143,10 @@ 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

@@ -41,9 +41,9 @@ func (c *WrapperCache) Close() {
 	}
 }
 
-func (c *WrapperCache) StartForward(ctx ml.Context, batch input.Batch) error {
+func (c *WrapperCache) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
 	for i, cache := range c.caches {
-		err := cache.StartForward(ctx, batch)
+		err := cache.StartForward(ctx, batch, reserve)
 		if err != nil {
 			// unwind on error - Remove with endIndex set to math.MaxInt32 does not fail
 			for j := i - 1; j >= 0; j-- {
@@ -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/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "d7cfe1ffe0f435d0048a6058d529daf76e072d9c";
+char const *LLAMA_COMMIT = "2016f07bd106c73699ecbaace80f55db5ed95dac";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/.rsync-filter

@@ -13,6 +13,7 @@ include include/llama-*.*
 include examples/
 include examples/llava/
 include examples/llava/clip.*
+include examples/llava/clip-impl.*
 include examples/llava/llava.*
 include src/
 include src/llama.*

llama/llama.cpp/common/common.cpp

@@ -7,10 +7,6 @@
 
 #include "common.h"
 #include "log.h"
-// Change JSON_ASSERT from assert() to GGML_ASSERT:
-#define JSON_ASSERT GGML_ASSERT
-#include "json.hpp"
-#include "json-schema-to-grammar.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -52,47 +48,11 @@
 #include <sys/stat.h>
 #include <unistd.h>
 #endif
-#if defined(LLAMA_USE_CURL)
-#include <curl/curl.h>
-#include <curl/easy.h>
-#include <future>
-#endif
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-#if defined(LLAMA_USE_CURL)
-#ifdef __linux__
-#include <linux/limits.h>
-#elif defined(_WIN32)
-#   if !defined(PATH_MAX)
-#   define PATH_MAX MAX_PATH
-#   endif
-#else
-#include <sys/syslimits.h>
-#endif
-#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
-
-//
-// CURL utils
-//
-
-using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
-
-// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
-struct curl_slist_ptr {
-    struct curl_slist * ptr = nullptr;
-    ~curl_slist_ptr() {
-        if (ptr) {
-            curl_slist_free_all(ptr);
-        }
-    }
-};
-#endif // LLAMA_USE_CURL
-
-using json = nlohmann::ordered_json;
-
 //
 // CPU utils
 //
@@ -483,6 +443,11 @@ void string_replace_all(std::string & s, const std::string & search, const std::
     s = std::move(builder);
 }
 
+std::string regex_escape(const std::string & s) {
+    static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
+    return std::regex_replace(s, special_chars, "\\$0");
+}
+
 std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
     std::ostringstream result;
     for (size_t i = 0; i < values.size(); ++i) {
@@ -865,7 +830,7 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#ifdef __linux__
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else {
@@ -875,7 +840,9 @@ std::string fs_get_cache_directory() {
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
         cache_directory = std::getenv("LOCALAPPDATA");
-#endif // __linux__
+#else
+#  error Unknown architecture
+#endif
         cache_directory = ensure_trailing_slash(cache_directory);
         cache_directory += "llama.cpp";
     }
@@ -896,22 +863,14 @@ std::string fs_get_cache_file(const std::string & filename) {
 //
 // Model utils
 //
+
 struct common_init_result common_init_from_params(common_params & params) {
     common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
 
-    llama_model * model = nullptr;
-
-    if (!params.hf_repo.empty() && !params.hf_file.empty()) {
-        model = common_load_model_from_hf(params.hf_repo, params.hf_file, params.model, params.hf_token, mparams);
-    } else if (!params.model_url.empty()) {
-        model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams);
-    } else {
-        model = llama_model_load_from_file(params.model.c_str(), mparams);
-    }
-
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
     if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.c_str());
+        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
         return iparams;
     }
 
@@ -946,13 +905,13 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str());
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
         llama_model_free(model);
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
-        LOG_WRN("%s: KV cache shifting is not supported for this model, disabling KV cache shifting\n", __func__);
+    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
+        LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
 
@@ -1029,6 +988,8 @@ struct common_init_result common_init_from_params(common_params & params) {
     if (params.warmup) {
         LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
+        llama_set_warmup(lctx, true);
+
         std::vector<llama_token> tmp;
         llama_token bos = llama_vocab_bos(vocab);
         llama_token eos = llama_vocab_eos(vocab);
@@ -1056,9 +1017,10 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_decoder(model)) {
             llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
         }
-        llama_kv_cache_clear(lctx);
+        llama_kv_self_clear(lctx);
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
+        llama_set_warmup(lctx, false);
     }
 
     iparams.model.reset(model);
@@ -1067,6 +1029,19 @@ struct common_init_result common_init_from_params(common_params & params) {
     return iparams;
 }
 
+std::string get_model_endpoint() {
+    const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
+    // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
+    const char * hf_endpoint_env = getenv("HF_ENDPOINT");
+    const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env;
+    std::string model_endpoint = "https://huggingface.co/";
+    if (endpoint_env) {
+        model_endpoint = endpoint_env;
+        if (model_endpoint.back() != '/') model_endpoint += '/';
+    }
+    return model_endpoint;
+}
+
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
     llama_clear_adapter_lora(ctx);
     for (auto & la : lora) {
@@ -1082,15 +1057,18 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     if (!params.devices.empty()) {
         mparams.devices = params.devices.data();
     }
+
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
+
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
+
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;
     } else {
@@ -1098,6 +1076,13 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.kv_overrides = params.kv_overrides.data();
     }
 
+    if (params.tensor_buft_overrides.empty()) {
+        mparams.tensor_buft_overrides = NULL;
+    } else {
+        GGML_ASSERT(params.tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern");
+        mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
+    }
+
     return mparams;
 }
 
@@ -1157,451 +1142,6 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
     return tpp;
 }
 
-#ifdef LLAMA_USE_CURL
-
-#define CURL_MAX_RETRY 3
-#define CURL_RETRY_DELAY_SECONDS 2
-
-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
-    int remaining_attempts = max_attempts;
-
-    while (remaining_attempts > 0) {
-        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
-
-        CURLcode res = curl_easy_perform(curl);
-        if (res == CURLE_OK) {
-            return true;
-        }
-
-        int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000;
-        LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
-
-        remaining_attempts--;
-        std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
-    }
-
-    LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
-
-    return false;
-}
-
-static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
-    // Initialize libcurl
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    if (!curl) {
-        LOG_ERR("%s: error initializing libcurl\n", __func__);
-        return false;
-    }
-
-    bool force_download = false;
-
-    // Set the URL, allow to follow http redirection
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
-
-    // Check if hf-token or bearer-token was specified
-    if (!hf_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + hf_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-        curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
-    }
-
-#if defined(_WIN32)
-    // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
-    //   operating system. Currently implemented under MS-Windows.
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
-
-    // Check if the file already exists locally
-    auto file_exists = std::filesystem::exists(path);
-
-    // If the file exists, check its JSON metadata companion file.
-    std::string metadata_path = path + ".json";
-    nlohmann::json metadata;
-    std::string etag;
-    std::string last_modified;
-
-    if (file_exists) {
-        // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
-        std::ifstream metadata_in(metadata_path);
-        if (metadata_in.good()) {
-            try {
-                metadata_in >> metadata;
-                LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("url") && metadata.at("url").is_string()) {
-                    auto previous_url = metadata.at("url").get<std::string>();
-                    if (previous_url != url) {
-                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
-                        return false;
-                    }
-                }
-                if (metadata.contains("etag") && metadata.at("etag").is_string()) {
-                    etag = metadata.at("etag");
-                }
-                if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
-                    last_modified = metadata.at("lastModified");
-                }
-            } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-                return false;
-            }
-        }
-    } else {
-        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
-    }
-
-    // Send a HEAD request to retrieve the etag and last-modified headers
-    struct common_load_model_from_url_headers {
-        std::string etag;
-        std::string last_modified;
-    };
-
-    common_load_model_from_url_headers headers;
-
-    {
-        typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
-        auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
-            common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
-
-            static std::regex header_regex("([^:]+): (.*)\r\n");
-            static std::regex etag_regex("ETag", std::regex_constants::icase);
-            static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
-
-            std::string header(buffer, n_items);
-            std::smatch match;
-            if (std::regex_match(header, match, header_regex)) {
-                const std::string & key = match[1];
-                const std::string & value = match[2];
-                if (std::regex_match(key, match, etag_regex)) {
-                    headers->etag = value;
-                } else if (std::regex_match(key, match, last_modified_regex)) {
-                    headers->last_modified = value;
-                }
-            }
-            return n_items;
-        };
-
-        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
-        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
-        curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
-        curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
-
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
-        if (!was_perform_successful) {
-            return false;
-        }
-
-        long http_code = 0;
-        curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code != 200) {
-            // HEAD not supported, we don't know if the file has changed
-            // force trigger downloading
-            force_download = true;
-            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
-        }
-    }
-
-    bool should_download = !file_exists || force_download;
-    if (!should_download) {
-        if (!etag.empty() && etag != headers.etag) {
-            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
-            should_download = true;
-        } else if (!last_modified.empty() && last_modified != headers.last_modified) {
-            LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str());
-            should_download = true;
-        }
-    }
-    if (should_download) {
-        std::string path_temporary = path + ".downloadInProgress";
-        if (file_exists) {
-            LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
-            if (remove(path.c_str()) != 0) {
-                LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
-                return false;
-            }
-        }
-
-        // Set the output file
-
-        struct FILE_deleter {
-            void operator()(FILE * f) const {
-                fclose(f);
-            }
-        };
-
-        std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "wb"));
-        if (!outfile) {
-            LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str());
-            return false;
-        }
-
-        typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
-        auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t {
-            return fwrite(data, size, nmemb, (FILE *)fd);
-        };
-        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L);
-        curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-        curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get());
-
-        //  display download progress
-        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L);
-
-        // helper function to hide password in URL
-        auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
-            std::size_t protocol_pos = url.find("://");
-            if (protocol_pos == std::string::npos) {
-                return url;  // Malformed URL
-            }
-
-            std::size_t at_pos = url.find('@', protocol_pos + 3);
-            if (at_pos == std::string::npos) {
-                return url;  // No password in URL
-            }
-
-            return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
-        };
-
-        // start the download
-        LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
-            llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
-        if (!was_perform_successful) {
-            return false;
-        }
-
-        long http_code = 0;
-        curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code < 200 || http_code >= 400) {
-            LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
-            return false;
-        }
-
-        // Causes file to be closed explicitly here before we rename it.
-        outfile.reset();
-
-        // Write the updated JSON metadata file.
-        metadata.update({
-            {"url", url},
-            {"etag", headers.etag},
-            {"lastModified", headers.last_modified}
-        });
-        std::ofstream(metadata_path) << metadata.dump(4);
-        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
-
-        if (rename(path_temporary.c_str(), path.c_str()) != 0) {
-            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
-            return false;
-        }
-    }
-
-    return true;
-}
-
-struct llama_model * common_load_model_from_url(
-        const std::string & model_url,
-        const std::string & local_path,
-        const std::string & hf_token,
-        const struct llama_model_params & params) {
-    // Basic validation of the model_url
-    if (model_url.empty()) {
-        LOG_ERR("%s: invalid model_url\n", __func__);
-        return NULL;
-    }
-
-    if (!common_download_file(model_url, local_path, hf_token)) {
-        return NULL;
-    }
-
-    // check for additional GGUFs split to download
-    int n_split = 0;
-    {
-        struct gguf_init_params gguf_params = {
-            /*.no_alloc = */ true,
-            /*.ctx      = */ NULL,
-        };
-        auto * ctx_gguf = gguf_init_from_file(local_path.c_str(), gguf_params);
-        if (!ctx_gguf) {
-            LOG_ERR("\n%s:  failed to load input GGUF from %s\n", __func__, local_path.c_str());
-            return NULL;
-        }
-
-        auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
-        if (key_n_split >= 0) {
-            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
-        }
-
-        gguf_free(ctx_gguf);
-    }
-
-    if (n_split > 1) {
-        char split_prefix[PATH_MAX] = {0};
-        char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
-
-        // Verify the first split file format
-        // and extract split URL and PATH prefixes
-        {
-            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), local_path.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, local_path.c_str(), n_split);
-                return NULL;
-            }
-
-            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model_url.c_str(), n_split);
-                return NULL;
-            }
-        }
-
-        // Prepare download in parallel
-        std::vector<std::future<bool>> futures_download;
-        for (int idx = 1; idx < n_split; idx++) {
-            futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split, hf_token](int download_idx) -> bool {
-                char split_path[PATH_MAX] = {0};
-                llama_split_path(split_path, sizeof(split_path), split_prefix, download_idx, n_split);
-
-                char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
-                llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split);
-
-                return common_download_file(split_url, split_path, hf_token);
-            }, idx));
-        }
-
-        // Wait for all downloads to complete
-        for (auto & f : futures_download) {
-            if (!f.get()) {
-                return NULL;
-            }
-        }
-    }
-
-    return llama_model_load_from_file(local_path.c_str(), params);
-}
-
-struct llama_model * common_load_model_from_hf(
-        const std::string & repo,
-        const std::string & remote_path,
-        const std::string & local_path,
-        const std::string & hf_token,
-        const struct llama_model_params & params) {
-    // construct hugging face model url:
-    //
-    //  --repo ggml-org/models --file tinyllama-1.1b/ggml-model-f16.gguf
-    //    https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf
-    //
-    //  --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf
-    //    https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf
-    //
-
-    std::string model_url = "https://huggingface.co/";
-    model_url += repo;
-    model_url += "/resolve/main/";
-    model_url += remote_path;
-
-    return common_load_model_from_url(model_url, local_path, hf_token, params);
-}
-
-/**
- * Allow getting the HF file from the HF repo with tag (like ollama), for example:
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
- * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
- *
- * Return pair of <repo, file> (with "repo" already having tag removed)
- *
- * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
- */
-std::pair<std::string, std::string> common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & hf_token) {
-    auto parts = string_split<std::string>(hf_repo_with_tag, ':');
-    std::string tag = parts.size() > 1 ? parts.back() : "latest";
-    std::string hf_repo = parts[0];
-    if (string_split<std::string>(hf_repo, '/').size() != 2) {
-        throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
-    }
-
-    // fetch model info from Hugging Face Hub API
-    json model_info;
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    std::string res_str;
-    std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
-    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
-    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
-        static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
-        return size * nmemb;
-    };
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
-#if defined(_WIN32)
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
-    if (!hf_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + hf_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-    }
-    // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
-    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
-
-    CURLcode res = curl_easy_perform(curl.get());
-
-    if (res != CURLE_OK) {
-        throw std::runtime_error("error: cannot make GET request to HF API");
-    }
-
-    long res_code;
-    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
-    if (res_code == 200) {
-        model_info = json::parse(res_str);
-    } else if (res_code == 401) {
-        throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
-    } else {
-        throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
-    }
-
-    // check response
-    if (!model_info.contains("ggufFile")) {
-        throw std::runtime_error("error: model does not have ggufFile");
-    }
-    json & gguf_file = model_info.at("ggufFile");
-    if (!gguf_file.contains("rfilename")) {
-        throw std::runtime_error("error: ggufFile does not have rfilename");
-    }
-
-    return std::make_pair(hf_repo, gguf_file.at("rfilename"));
-}
-
-#else
-
-struct llama_model * common_load_model_from_url(
-        const std::string & /*model_url*/,
-        const std::string & /*local_path*/,
-        const std::string & /*hf_token*/,
-        const struct llama_model_params & /*params*/) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
-    return nullptr;
-}
-
-struct llama_model * common_load_model_from_hf(
-        const std::string & /*repo*/,
-        const std::string & /*remote_path*/,
-        const std::string & /*local_path*/,
-        const std::string & /*hf_token*/,
-        const struct llama_model_params & /*params*/) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
-    return nullptr;
-}
-
-std::pair<std::string, std::string> common_get_hf_file(const std::string &, const std::string &) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
-    return std::make_pair("", "");
-}
-
-#endif // LLAMA_USE_CURL
-
 //
 // Batch utils
 //
@@ -2025,4 +1565,3 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 
     return result;
 }
-

llama/llama.cpp/common/common.h

@@ -110,9 +110,17 @@ enum common_conversation_mode {
     COMMON_CONVERSATION_MODE_AUTO     = 2,
 };
 
+enum common_grammar_trigger_type {
+    COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN,
+    COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
+    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
+    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+};
+
 struct common_grammar_trigger {
-    std::string word;
-    bool at_start;
+    common_grammar_trigger_type type;
+    std::string value;
+    llama_token token = LLAMA_TOKEN_NULL;
 };
 
 // sampling parameters
@@ -163,8 +171,7 @@ struct common_params_sampling {
 
     std::string                         grammar; // optional BNF-like grammar to constrain sampling
     bool                                grammar_lazy = false;
-    std::vector<common_grammar_trigger> grammar_trigger_words;  // optional trigger words to trigger lazy grammar
-    std::vector<llama_token>            grammar_trigger_tokens; // optional trigger tokens to trigger lazy grammar and print trigger special tokens.
+    std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
     std::set<llama_token>               preserved_tokens;
 
     std::vector<llama_logit_bias> logit_bias; // logit biases to apply
@@ -173,6 +180,13 @@ struct common_params_sampling {
     std::string print() const;
 };
 
+struct common_params_model {
+    std::string path    = ""; // model local path                                           // NOLINT
+    std::string url     = ""; // model url to download                                      // NOLINT
+    std::string hf_repo = ""; // HF repo                                                    // NOLINT
+    std::string hf_file = ""; // HF file                                                    // NOLINT
+};
+
 struct common_params_speculative {
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
@@ -186,19 +200,13 @@ struct common_params_speculative {
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
-    std::string hf_repo = ""; // HF repo                                                     // NOLINT
-    std::string hf_file = ""; // HF file                                                     // NOLINT
-
-    std::string model = "";     // draft model for speculative decoding                      // NOLINT
-    std::string model_url = ""; // model url to download                                     // NOLINT
+    struct common_params_model model;
 };
 
 struct common_params_vocoder {
-    std::string hf_repo = ""; // HF repo                                                     // NOLINT
-    std::string hf_file = ""; // HF file                                                     // NOLINT
+    struct common_params_model model;
 
-    std::string model     = ""; // model path                                                // NOLINT
-    std::string model_url = ""; // model url to download                                     // NOLINT
+    std::string speaker_file = ""; // speaker file path                                      // NOLINT
 
     bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
 };
@@ -254,13 +262,12 @@ struct common_params {
     struct common_params_speculative speculative;
     struct common_params_vocoder     vocoder;
 
-    std::string model                = ""; // model path                                                    // NOLINT
+    struct common_params_model model;
+
     std::string model_alias          = ""; // model alias                                                   // NOLINT
-    std::string model_url            = ""; // model url to download                                         // NOLINT
     std::string hf_token             = ""; // HF token                                                      // NOLINT
-    std::string hf_repo              = ""; // HF repo                                                       // NOLINT
-    std::string hf_file              = ""; // HF file                                                       // NOLINT
     std::string prompt               = "";                                                                  // NOLINT
+    std::string system_prompt        = "";                                                                  // NOLINT
     std::string prompt_file          = ""; // store the external prompt file name                           // NOLINT
     std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state             // NOLINT
     std::string input_prefix         = ""; // string to prefix user inputs with                             // NOLINT
@@ -272,6 +279,7 @@ struct common_params {
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
 
     bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply)
     std::vector<common_adapter_lora_info> lora_adapters; // lora adapter path with user defined scale
@@ -325,13 +333,15 @@ struct common_params {
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
 
+    bool single_turn       = false; // single turn chat conversation
+
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
 
     common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
 
     // multimodal models (see examples/llava)
-    std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
+    struct common_params_model mmproj;
     std::vector<std::string> image; // path to image file(s)
 
     // embedding
@@ -391,8 +401,6 @@ struct common_params {
     int32_t i_pos  = -1;  // position of the passkey in the junk text
 
     // imatrix params
-    std::string out_file = "imatrix.dat"; // save the resulting imatrix to this file
-
     int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
     int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
     int32_t i_chunk     =  0; // start processing from this chunk
@@ -404,16 +412,16 @@ struct common_params {
     int n_pca_batch = 100;
     int n_pca_iterations = 1000;
     dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
-    std::string cvector_outfile       = "control_vector.gguf";
     std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
     std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
 
     bool spm_infill = false; // suffix/prefix/middle pattern for infill
 
-    std::string lora_outfile = "ggml-lora-merged-f16.gguf";
-
     // batched-bench params
     bool batched_bench_output_jsonl = false;
+
+    // common params
+    std::string out_file; // output filename for all example programs
 };
 
 // call once at the start of a program if it uses libcommon
@@ -453,6 +461,8 @@ std::string string_repeat(const std::string & str, size_t n);
 
 void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
 
+std::string regex_escape(const std::string & s);
+
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
     static_assert(!std::is_same<T, std::string>::value, "Please use the specialized version for std::string");
@@ -530,26 +540,11 @@ struct llama_model_params     common_model_params_to_llama  (      common_params
 struct llama_context_params   common_context_params_to_llama(const common_params & params);
 struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 
-struct llama_model * common_load_model_from_url(
-    const std::string & model_url,
-    const std::string & local_path,
-    const std::string & hf_token,
-    const struct llama_model_params & params);
-
-struct llama_model * common_load_model_from_hf(
-    const std::string & repo,
-    const std::string & remote_path,
-    const std::string & local_path,
-    const std::string & hf_token,
-    const struct llama_model_params & params);
-
-std::pair<std::string, std::string> common_get_hf_file(
-    const std::string & hf_repo_with_tag,
-    const std::string & hf_token);
-
 // clear LoRA adapters from context, then apply new list of adapters
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
 
+std::string                   get_model_endpoint();
+
 //
 // Batch utils
 //

llama/llama.cpp/common/json-schema-to-grammar.cpp

@@ -264,7 +264,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
     throw std::runtime_error("At least one of min_value or max_value must be set");
 }
 
-const std::string SPACE_RULE = "| \" \" | \"\\n\" [ \\t]{0,20}";
+const std::string SPACE_RULE = "| \" \" | \"\\n\"{1,2} [ \\t]{0,20}";
 
 struct BuiltinRule {
     std::string content;
@@ -764,11 +764,10 @@ private:
 public:
     SchemaConverter(
         const std::function<json(const std::string &)> & fetch_json,
-        bool dotall,
-        bool compact_spaces)
+        bool dotall)
           : _fetch_json(fetch_json), _dotall(dotall)
     {
-        _rules["space"] = compact_spaces ? "\" \"?" : SPACE_RULE;
+        _rules["space"] = SPACE_RULE;
     }
 
     void resolve_refs(json & schema, const std::string & url) {
@@ -1007,7 +1006,7 @@ std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
 }
 
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options) {
-    SchemaConverter converter([&](const std::string &) { return json(); }, options.dotall, options.compact_spaces);
+    SchemaConverter converter([&](const std::string &) { return json(); }, options.dotall);
     common_grammar_builder builder {
         /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
             return converter._add_rule(name, rule);

llama/llama.cpp/common/json-schema-to-grammar.h

@@ -16,7 +16,6 @@ struct common_grammar_builder {
 
 struct common_grammar_options {
     bool dotall = false;
-    bool compact_spaces = false;
 };
 
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options = {});

llama/llama.cpp/common/sampling.cpp

@@ -4,6 +4,7 @@
 
 #include <cmath>
 #include <unordered_map>
+#include <algorithm>
 
 // the ring buffer works similarly to std::deque, but with a fixed capacity
 // TODO: deduplicate with llama-impl.h
@@ -159,17 +160,57 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
     } else {
-        std::vector<const char *> trigger_words;
-        trigger_words.reserve(params.grammar_trigger_words.size());
-        for (const auto & str : params.grammar_trigger_words) {
-            trigger_words.push_back(str.word.c_str());
+        std::vector<std::string> patterns_at_start;
+        std::vector<std::string> patterns_anywhere;
+        std::vector<llama_token> trigger_tokens;
+        for (const auto & trigger : params.grammar_triggers) {
+            switch (trigger.type) {
+                case COMMON_GRAMMAR_TRIGGER_TYPE_WORD:
+                {
+                    const auto & word = trigger.value;
+                    patterns_anywhere.push_back(regex_escape(word));
+                    break;
+                }
+                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN:
+                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START:
+                {
+                    const auto & pattern = trigger.value;
+                    (trigger.type == COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START ? patterns_at_start : patterns_anywhere).push_back(pattern);
+                    break;
+                }
+                case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN:
+                {
+                    const auto token = trigger.token;
+                    trigger_tokens.push_back(token);
+                    break;
+                }
+                default:
+                    GGML_ASSERT(false && "unknown trigger type");
+            }
+        }
+
+        std::vector<std::string> trigger_patterns;
+        if (!patterns_at_start.empty()) {
+            trigger_patterns.push_back("^(" + string_join(patterns_at_start, "|") + ")[\\s\\S]*");
+        }
+        if (!patterns_anywhere.empty()) {
+            trigger_patterns.push_back("^[\\s\\S]*?(" + string_join(patterns_anywhere, "|") + ")[\\s\\S]*");
+        }
+
+        std::vector<const char *> trigger_patterns_c;
+        trigger_patterns_c.reserve(trigger_patterns.size());
+        for (const auto & regex : trigger_patterns) {
+            trigger_patterns_c.push_back(regex.c_str());
         }
 
         grmr = params.grammar_lazy
-             ? llama_sampler_init_grammar_lazy(vocab, params.grammar.c_str(), "root",
-                                               trigger_words.data(), trigger_words.size(),
-                                               params.grammar_trigger_tokens.data(), params.grammar_trigger_tokens.size())
+             ? llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
+                                                        trigger_patterns_c.data(), trigger_patterns_c.size(),
+                                                        trigger_tokens.data(), trigger_tokens.size())
              :      llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
+        if (!grmr) {
+            return nullptr;
+        }
     }
 
     auto * result = new common_sampler {

llama/llama.cpp/examples/llava/clip-impl.h

@@ -0,0 +1,341 @@
+#include "ggml.h"
+#include "gguf.h"
+#include "clip.h"
+
+#include "clip.h"
+
+#include <climits>
+#include <cstdarg>
+#include <string>
+#include <map>
+#include <sstream>
+#include <vector>
+#include <memory>
+
+// Internal header for clip.cpp
+
+#define KEY_FTYPE               "general.file_type"
+#define KEY_NAME                "general.name"
+#define KEY_DESCRIPTION         "general.description"
+#define KEY_HAS_TEXT_ENC        "clip.has_text_encoder"
+#define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
+#define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
+#define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
+#define KEY_HAS_GLM_PROJ        "clip.has_glm_projector"
+#define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
+#define KEY_HAS_QWEN2VL_MERGER  "clip.has_qwen2vl_merger"
+#define KEY_USE_GELU            "clip.use_gelu"
+#define KEY_USE_SILU            "clip.use_silu"
+#define KEY_N_EMBD              "clip.%s.embedding_length"
+#define KEY_N_FF                "clip.%s.feed_forward_length"
+#define KEY_N_BLOCK             "clip.%s.block_count"
+#define KEY_N_HEAD              "clip.%s.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
+#define KEY_PROJ_DIM            "clip.%s.projection_dim"
+#define KEY_TOKENS              "tokenizer.ggml.tokens"
+#define KEY_N_POSITIONS         "clip.text.context_length"
+#define KEY_IMAGE_SIZE          "clip.vision.image_size"
+#define KEY_PATCH_SIZE          "clip.vision.patch_size"
+#define KEY_IMAGE_MEAN          "clip.vision.image_mean"
+#define KEY_IMAGE_STD           "clip.vision.image_std"
+#define KEY_PROJ_TYPE           "clip.projector_type"
+#define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
+
+#define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
+#define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
+#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
+
+
+//
+// tensor name constants
+//
+
+#define TN_POS_EMBD        "%s.position_embd.weight"
+#define TN_CLASS_EMBD      "v.class_embd"
+#define TN_PATCH_EMBD      "v.patch_embd.weight"  // not rename tensor with ".0" postfix for backwrad compat
+#define TN_PATCH_EMBD_1    "v.patch_embd.weight.1"
+#define TN_PATCH_BIAS      "v.patch_embd.bias"
+#define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
+#define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
+#define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
+#define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
+#define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
+#define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
+#define TN_LN_1            "%s.blk.%d.ln1.%s"
+#define TN_LN_2            "%s.blk.%d.ln2.%s"
+#define TN_LN_PRE          "%s.pre_ln.%s"
+#define TN_LN_POST         "%s.post_ln.%s"
+#define TN_LLAVA_PROJ      "mm.%d.%s"
+#define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
+#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
+#define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
+#define TN_IMAGE_NEWLINE   "model.image_newline"
+#define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
+#define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
+
+// mimicpmv
+#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
+#define TN_MINICPMV_QUERY      "resampler.query"
+#define TN_MINICPMV_PROJ       "resampler.proj.weight"
+#define TN_MINICPMV_KV_PROJ    "resampler.kv.weight"
+#define TN_MINICPMV_ATTN       "resampler.attn.%s.%s"
+#define TN_MINICPMV_LN         "resampler.ln_%s.%s"
+
+#define TN_GLM_ADAPER_CONV      "adapter.conv.%s"
+#define TN_GLM_ADAPTER_LINEAR   "adapter.linear.linear.%s"
+#define TN_GLM_ADAPTER_NORM_1   "adapter.linear.norm1.%s"
+#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
+#define TN_GLM_ADAPTER_GATE     "adapter.linear.gate.%s"
+#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
+#define TN_GLM_BOI_W            "adapter.boi"
+#define TN_GLM_EOI_W            "adapter.eoi"
+
+enum projector_type {
+    PROJECTOR_TYPE_MLP,
+    PROJECTOR_TYPE_MLP_NORM,
+    PROJECTOR_TYPE_LDP,
+    PROJECTOR_TYPE_LDPV2,
+    PROJECTOR_TYPE_RESAMPLER,
+    PROJECTOR_TYPE_GLM_EDGE,
+    PROJECTOR_TYPE_MERGER,
+    PROJECTOR_TYPE_GEMMA3,
+    PROJECTOR_TYPE_UNKNOWN,
+};
+
+static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
+    { PROJECTOR_TYPE_MLP,       "mlp" },
+    { PROJECTOR_TYPE_LDP,       "ldp" },
+    { PROJECTOR_TYPE_LDPV2,     "ldpv2"},
+    { PROJECTOR_TYPE_RESAMPLER, "resampler"},
+    { PROJECTOR_TYPE_GLM_EDGE,  "adapter"},
+    { PROJECTOR_TYPE_MERGER,    "qwen2vl_merger"},
+    { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
+};
+
+static projector_type clip_projector_type_from_string(const std::string & str) {
+    for (const auto & pair : PROJECTOR_TYPE_NAMES) {
+        if (pair.second == str) {
+            return pair.first;
+        }
+    }
+    return PROJECTOR_TYPE_UNKNOWN;
+}
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
+//
+// logging
+//
+
+static void clip_log_callback_default(enum ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    fputs(text, stderr);
+    fflush(stderr);
+}
+
+struct clip_logger_state {
+    ggml_log_level verbosity_thold;
+    ggml_log_callback log_callback;
+    void * log_callback_user_data;
+};
+
+extern struct clip_logger_state g_logger_state;
+
+static void clip_log_internal_v(enum ggml_log_level level, const char * format, va_list args) {
+    if (format == NULL) {
+        return;
+    }
+    va_list args_copy;
+    va_copy(args_copy, args);
+    char buffer[128];
+    int len = vsnprintf(buffer, 128, format, args);
+    if (len < 128) {
+        g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
+    } else {
+        char * buffer2 = (char *) calloc(len + 1, sizeof(char));
+        vsnprintf(buffer2, len + 1, format, args_copy);
+        buffer2[len] = 0;
+        g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
+        free(buffer2);
+    }
+    va_end(args_copy);
+}
+
+static void clip_log_internal(enum ggml_log_level level, const char * format, ...) {
+    va_list args;
+    va_start(args, format);
+    clip_log_internal_v(level, format, args);
+    va_end(args);
+}
+
+#define LOG_TMPL(level, ...) \
+    do { \
+        if ((level) >= g_logger_state.verbosity_thold) { \
+            clip_log_internal((level), __VA_ARGS__); \
+        } \
+    } while (0)
+#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
+#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)
+#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
+#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  __VA_ARGS__)
+
+//
+// cpp wrappers
+//
+
+// wrapper for clip_image_size
+struct clip_image_size_deleter {
+    void operator()(clip_image_size * val) { clip_image_size_free(val); }
+};
+typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
+
+// wrapper for clip_image_u8
+struct clip_image_u8_deleter {
+    void operator()(clip_image_u8 * val) { clip_image_u8_free(val); }
+};
+typedef std::unique_ptr<clip_image_u8, clip_image_u8_deleter> clip_image_u8_ptr;
+
+// wrapper for clip_image_f32
+struct clip_image_f32_deleter {
+    void operator()(clip_image_f32 * val) { clip_image_f32_free(val); }
+};
+typedef std::unique_ptr<clip_image_f32, clip_image_f32_deleter> clip_image_f32_ptr;
+
+struct clip_image_u8_batch {
+    std::vector<clip_image_u8_ptr> entries;
+};
+
+struct clip_image_f32_batch {
+    std::vector<clip_image_f32_ptr> entries;
+};
+
+//
+// common utils
+//
+
+static std::string string_format(const char * fmt, ...) {
+    va_list ap;
+    va_list ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    GGML_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), buf.size());
+}
+
+static void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return;
+    }
+    std::string builder;
+    builder.reserve(s.length());
+    size_t pos = 0;
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
+    }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
+}
+
+// split string by a `std::string delim` instead of `char delim`
+static std::vector<std::string> string_split_str(std::string s, const std::string & delimiter) {
+    std::vector<std::string> tokens;
+    size_t pos = 0;
+    std::string token;
+    while ((pos = s.find(delimiter)) != std::string::npos) {
+        token = s.substr(0, pos);
+        tokens.push_back(token);
+        s.erase(0, pos + delimiter.length());
+    }
+    tokens.push_back(s);
+    return tokens;
+}
+
+//
+// gguf utils
+//
+
+static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
+    switch (type) {
+        case GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
+        case GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
+        case GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
+        case GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
+        case GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
+        case GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
+        case GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
+        case GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
+        case GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
+        case GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
+        case GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
+        default:                return string_format("unknown type %d", type);
+    }
+}
+
+static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
+    const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
+
+    switch (type) {
+        case GGUF_TYPE_STRING:
+            return gguf_get_val_str(ctx_gguf, i);
+        case GGUF_TYPE_ARRAY:
+            {
+                const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
+                int arr_n = gguf_get_arr_n(ctx_gguf, i);
+                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
+                std::stringstream ss;
+                ss << "[";
+                for (int j = 0; j < arr_n; j++) {
+                    if (arr_type == GGUF_TYPE_STRING) {
+                        std::string val = gguf_get_arr_str(ctx_gguf, i, j);
+                        // escape quotes
+                        string_replace_all(val, "\\", "\\\\");
+                        string_replace_all(val, "\"", "\\\"");
+                        ss << '"' << val << '"';
+                    } else if (arr_type == GGUF_TYPE_ARRAY) {
+                        ss << "???";
+                    } else {
+                        ss << gguf_data_to_str(arr_type, data, j);
+                    }
+                    if (j < arr_n - 1) {
+                        ss << ", ";
+                    }
+                }
+                ss << "]";
+                return ss.str();
+            }
+        default:
+            return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
+    }
+}
+
+//
+// API used internally with mtmd
+//
+
+projector_type clip_get_projector_type(const struct clip_ctx * ctx);

llama/llama.cpp/examples/llava/clip.h

@@ -1,6 +1,7 @@
 #ifndef CLIP_H
 #define CLIP_H
 
+#include "ggml.h"
 #include <stddef.h>
 #include <stdint.h>
 
@@ -29,27 +30,28 @@ struct clip_image_size {
     int height;
 };
 
-struct clip_image_u8_batch {
-    struct clip_image_u8 * data;
-    size_t size;
+struct clip_image_f32;
+struct clip_image_u8_batch;
+struct clip_image_f32_batch;
+
+struct clip_context_params {
+    bool use_gpu;
+    enum ggml_log_level verbosity;
 };
 
-struct clip_image_f32_batch {
-    struct clip_image_f32 * data;
-    size_t size;
-};
+// deprecated, use clip_init
+CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
 
-CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
-CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
+CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params);
 
 CLIP_API void clip_free(struct clip_ctx * ctx);
 
 CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
 CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);
 
-CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
+CLIP_API int32_t clip_get_image_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_get_patch_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_get_hidden_size(const struct clip_ctx * ctx);
 
 // TODO: should be enum, not string
 CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
@@ -65,16 +67,30 @@ CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
 CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
 CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip);
 
-CLIP_API struct clip_image_size * clip_image_size_init();
-CLIP_API struct clip_image_u8  * clip_image_u8_init ();
-CLIP_API struct clip_image_f32 * clip_image_f32_init();
+CLIP_API struct clip_image_size      * clip_image_size_init();
+CLIP_API struct clip_image_u8        * clip_image_u8_init ();
+CLIP_API struct clip_image_f32       * clip_image_f32_init();
+CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(); // only used by libllava
 
+// nx, ny are the output image dimensions
+CLIP_API unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny);
+
+CLIP_API void clip_image_size_free (struct clip_image_size * img_size);
 CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
 CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
 CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
 CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
 
-/** build image from pixels decoded by other libraries instead of stb_image.h for better performance. The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes */
+// use for accessing underlay data of clip_image_f32_batch
+CLIP_API size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size()
+CLIP_API size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx
+CLIP_API size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
+CLIP_API struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
+
+/**
+ * Build image from pixels decoded by other libraries instead of stb_image.h for better performance.
+ * The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes
+ */
 CLIP_API void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img);
 
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
@@ -95,6 +111,8 @@ CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out
 CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_llava(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_gemma3(const struct clip_ctx * ctx);
 
 CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);
 

llama/llama.cpp/examples/llava/llava.cpp

@@ -10,6 +10,7 @@
 #include <cstring>
 #include <limits>
 #include <vector>
+#include <memory>
 
 #if defined(LLAVA_LOG_OFF)
 #   define LOG_INF(...)
@@ -45,6 +46,17 @@ struct clip_image_grid_shape {
     int second;
 };
 
+// convenience cpp wrapper
+struct clip_image_f32_batch_deleter {
+    void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); }
+};
+typedef std::unique_ptr<clip_image_f32_batch, clip_image_f32_batch_deleter> clip_image_f32_batch_ptr;
+
+struct clip_image_size_deleter {
+    void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); }
+};
+typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
+
 /**
  * Selects the best resolution from a list of possible resolutions based on the original size.
  *
@@ -105,8 +117,8 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
         struct ggml_context * ctx;
     } model;
 
-    const int32_t image_size = clip_image_size(ctx_clip);
-    const int32_t patch_size = clip_patch_size(ctx_clip);
+    const int32_t image_size = clip_get_image_size(ctx_clip);
+    const int32_t patch_size = clip_get_patch_size(ctx_clip);
 
     int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
 
@@ -246,12 +258,9 @@ static clip_image_f32 * reshape_by_patch(clip_image_f32 * image, int patch_size)
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
     // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
-    clip_image_f32_batch img_res_v;
-    img_res_v.size = 0;
-    img_res_v.data = nullptr;
-    if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
+    clip_image_f32_batch_ptr img_res_v(clip_image_f32_batch_init());
+    if (!clip_image_preprocess(ctx_clip, img, img_res_v.get())) {
         LOG_ERR("%s: unable to preprocess image\n", __func__);
-        delete[] img_res_v.data;
         return false;
     }
 
@@ -259,66 +268,72 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
     const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
 
+    const size_t n_imgs = clip_image_f32_batch_n_images(img_res_v.get());
+
     if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) {
         std::vector<float *> image_embd_v;
-        image_embd_v.resize(img_res_v.size);
-        struct clip_image_size * load_image_size = clip_image_size_init();
+        image_embd_v.resize(n_imgs);
+        clip_image_size load_image_size;
 
-        for (size_t i = 0; i < img_res_v.size; i++) {
+        for (size_t i = 0; i < n_imgs; i++) {
             const int64_t t_img_enc_step_start_us = ggml_time_us();
-            image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
-            int patch_size=14;
-            load_image_size->width = img_res_v.data[i].nx;
-            load_image_size->height = img_res_v.data[i].ny;
-            clip_add_load_image_size(ctx_clip, load_image_size);
+            int nx = clip_image_f32_batch_nx(img_res_v.get(), i);
+            int ny = clip_image_f32_batch_ny(img_res_v.get(), i);
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, nx, ny));
+            int patch_size = 14;
+            load_image_size.width = nx;
+            load_image_size.height = ny;
+            clip_add_load_image_size(ctx_clip, &load_image_size);
 
             bool encoded = false;
+            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
             if (clip_is_qwen2vl(ctx_clip)) {
-                encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
+                encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]);
             }
             else {
-                encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+                encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(img_res, patch_size), image_embd_v[i]);
             }
 
             if (!encoded) {
-                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
                 return false;
             }
             const int64_t t_img_enc_steop_batch_us = ggml_time_us();
-            LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
+            LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)n_imgs, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
         }
         const int64_t t_img_enc_batch_us = ggml_time_us();
-        LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
 
         int n_img_pos_out = 0;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
+            int nx = clip_image_f32_batch_nx(img_res_v.get(), i);
+            int ny = clip_image_f32_batch_ny(img_res_v.get(), i);
+            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
             std::memcpy(
                 image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
                 image_embd_v[i],
-                clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
-            n_img_pos_out += clip_n_patches_by_img(ctx_clip, &img_res_v.data[i]);
+                clip_embd_nbytes_by_img(ctx_clip, nx, ny));
+            n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res);
         }
         *n_img_pos = n_img_pos_out;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
             free(image_embd_v[i]);
         }
         image_embd_v.clear();
-        load_image_size->width = img->nx;
-        load_image_size->height = img->ny;
-        clip_add_load_image_size(ctx_clip, load_image_size);
-        LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
-        delete[] img_res_v.data;
-        img_res_v.size = 0;
-        img_res_v.data = nullptr;
+        load_image_size.width = img->nx;
+        load_image_size.height = img->ny;
+        clip_add_load_image_size(ctx_clip, &load_image_size);
+        LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size.width, load_image_size.height);
     }
     else if (clip_is_glm(ctx_clip)){
         struct clip_image_size * load_image_size = clip_image_size_init();
-        load_image_size->width = img_res_v.data[0].nx;
-        load_image_size->height = img_res_v.data[0].ny;
+        load_image_size->width  = clip_image_f32_batch_nx(img_res_v.get(), 0);
+        load_image_size->height = clip_image_f32_batch_ny(img_res_v.get(), 0);
         clip_add_load_image_size(ctx_clip, load_image_size);
 
-        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd);
-        int pos = int(load_image_size->width/clip_patch_size(ctx_clip)/2);
+        clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
+        int pos = int(load_image_size->width/clip_get_patch_size(ctx_clip)/2);
         *n_img_pos = (pos * pos + 2);
         if (!encoded){
             LOG_ERR("Unable to encode image \n");
@@ -328,8 +343,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
     else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
         // flat / default llava-1.5 type embedding
         *n_img_pos = clip_n_patches(ctx_clip);
-        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
-        delete[] img_res_v.data;
+        clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); // image_embd shape is 576 x 4096
         if (!encoded) {
             LOG_ERR("Unable to encode image\n");
 
@@ -340,17 +355,18 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         // spatial_unpad llava-1.6 type embedding
         // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
         std::vector<float *> image_embd_v;
-        image_embd_v.resize(img_res_v.size);
-        for (size_t i = 0; i < img_res_v.size; i++) {
+        image_embd_v.resize(n_imgs);
+        for (size_t i = 0; i < n_imgs; i++) {
+            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
             image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
-            const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
             if (!encoded) {
-                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
                 return false;
             }
         }
         const int64_t t_img_enc_batch_us = ggml_time_us();
-        LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
 
         const int32_t * image_grid = clip_image_grid(ctx_clip);
         const size_t num_gridpoints = get_clip_image_grid_size(ctx_clip);
@@ -360,12 +376,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
             grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
         }
 
-        // free all img_res_v - not needed anymore
-        delete[] img_res_v.data;
-        img_res_v.size = 0;
-        img_res_v.data = nullptr;
-
-        const int32_t image_size = clip_image_size(ctx_clip);
+        const int32_t image_size = clip_get_image_size(ctx_clip);
 
         struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
 

llama/llama.cpp/include/llama.h

@@ -60,6 +60,7 @@ extern "C" {
     struct llama_model;
     struct llama_context;
     struct llama_sampler;
+    struct llama_kv_cache;
 
     typedef int32_t llama_pos;
     typedef int32_t llama_token;
@@ -106,6 +107,10 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
         LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM  = 28,
         LLAMA_VOCAB_PRE_TYPE_GPT4O          = 29,
+        LLAMA_VOCAB_PRE_TYPE_SUPERBPE       = 30,
+        LLAMA_VOCAB_PRE_TYPE_TRILLION       = 31,
+        LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
+        LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
     };
 
     enum llama_rope_type {
@@ -277,10 +282,18 @@ extern "C" {
         };
     };
 
+    struct llama_model_tensor_buft_override {
+        const char * pattern;
+        ggml_backend_buffer_type_t buft;
+    };
+
     struct llama_model_params {
         // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used)
         ggml_backend_dev_t * devices;
 
+        // NULL-terminated list of buffer types to use for tensors that match a pattern
+        const struct llama_model_tensor_buft_override * tensor_buft_overrides;
+
         int32_t n_gpu_layers; // number of layers to store in VRAM
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
@@ -356,17 +369,18 @@ extern "C" {
 
     // model quantization parameters
     typedef struct llama_model_quantize_params {
-        int32_t nthread;                     // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
-        enum llama_ftype ftype;              // quantize to this llama_ftype
-        enum ggml_type output_tensor_type;   // output tensor type
-        enum ggml_type token_embedding_type; // token embeddings tensor type
-        bool allow_requantize;               // allow quantizing non-f32/f16 tensors
-        bool quantize_output_tensor;         // quantize output.weight
-        bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
-        bool pure;                           // quantize all tensors to the default type
-        bool keep_split;                     // quantize to the same number of shards
-        void * imatrix;                      // pointer to importance matrix data
-        void * kv_overrides;                 // pointer to vector containing overrides
+        int32_t nthread;                      // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype ftype;               // quantize to this llama_ftype
+        enum ggml_type output_tensor_type;    // output tensor type
+        enum ggml_type token_embedding_type;  // token embeddings tensor type
+        bool allow_requantize;                // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor;          // quantize output.weight
+        bool only_copy;                       // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                            // quantize all tensors to the default type
+        bool keep_split;                      // quantize to the same number of shards
+        void * imatrix;                       // pointer to importance matrix data
+        void * kv_overrides;                  // pointer to vector containing overrides
+        void * tensor_types;                  // pointer to vector containing tensor types
     } llama_model_quantize_params;
 
     typedef struct llama_logit_bias {
@@ -475,7 +489,8 @@ extern "C" {
     DEPRECATED(LLAMA_API int32_t llama_n_vocab    (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead");
 
     LLAMA_API const struct llama_model * llama_get_model   (const struct llama_context * ctx);
-    LLAMA_API enum llama_pooling_type    llama_pooling_type(const struct llama_context * ctx);
+    LLAMA_API    struct llama_kv_cache * llama_get_kv_self (      struct llama_context * ctx);
+    LLAMA_API  enum llama_pooling_type   llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type
 
     LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
     LLAMA_API enum llama_rope_type       llama_model_rope_type(const struct llama_model * model);
@@ -592,7 +607,7 @@ extern "C" {
     // KV cache
     //
 
-    // TODO: remove llama_kv_cache_view_* API
+    // TODO: start using struct llama_kv_cache
 
     // Information associated with an individual cell in the KV cache view.
     struct llama_kv_cache_view_cell {
@@ -647,13 +662,19 @@ extern "C" {
 
     // Returns the number of tokens in the KV cache (slow, use only for debug)
     // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
-    LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx);
+    LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx);
+
+    DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx),
+            "use llama_kv_self_n_tokens instead");
 
     // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
-    LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx);
+    LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx);
+
+    DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx),
+            "use llama_kv_self_used_cells instead");
 
     // Clear the KV cache - both cell info is erased and KV data is zeroed
-    LLAMA_API void llama_kv_cache_clear(
+    LLAMA_API void llama_kv_self_clear(
             struct llama_context * ctx);
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
@@ -661,7 +682,7 @@ extern "C" {
     // seq_id < 0 : match any sequence
     // p0 < 0     : [0,  p1]
     // p1 < 0     : [p0, inf)
-    LLAMA_API bool llama_kv_cache_seq_rm(
+    LLAMA_API bool llama_kv_self_seq_rm(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -671,7 +692,7 @@ extern "C" {
     // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_cp(
+    LLAMA_API void llama_kv_self_seq_cp(
             struct llama_context * ctx,
                     llama_seq_id   seq_id_src,
                     llama_seq_id   seq_id_dst,
@@ -679,17 +700,17 @@ extern "C" {
                        llama_pos   p1);
 
     // Removes all tokens that do not belong to the specified sequence
-    LLAMA_API void llama_kv_cache_seq_keep(
+    LLAMA_API void llama_kv_self_seq_keep(
             struct llama_context * ctx,
                     llama_seq_id   seq_id);
 
     // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
     // If the KV cache is RoPEd, the KV data is updated accordingly:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
+    //   - explicitly with llama_kv_self_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_add(
+    LLAMA_API void llama_kv_self_seq_add(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -699,10 +720,10 @@ extern "C" {
     // Integer division of the positions by factor of `d > 1`
     // If the KV cache is RoPEd, the KV data is updated accordingly:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
+    //   - explicitly with llama_kv_self_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_div(
+    LLAMA_API void llama_kv_self_seq_div(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -710,24 +731,76 @@ extern "C" {
                              int   d);
 
     // Returns the largest position present in the KV cache for the specified sequence
-    LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
+    LLAMA_API llama_pos llama_kv_self_seq_pos_max(
             struct llama_context * ctx,
-                    llama_seq_id   seq_id);
-
-    // TODO: the llama_kv_cache_defrag and llama_kv_cache_update API tightly couples llama_context with llama_kv_cache
-    //       how to avoid this?
+                     llama_seq_id   seq_id);
 
     // Defragment the KV cache
     // This will be applied:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
-    LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx);
-
-    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
+    //   - explicitly with llama_kv_self_update()
+    LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx);
 
     // Check if the context supports KV cache shifting
-    LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx);
+    LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx);
+
+    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
+    LLAMA_API void llama_kv_self_update(struct llama_context * ctx);
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_clear(
+            struct llama_context * ctx),
+            "use llama_kv_self_clear instead");
+
+    DEPRECATED(LLAMA_API bool llama_kv_cache_seq_rm(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1),
+            "use llama_kv_self_seq_rm instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_cp(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id_src,
+                    llama_seq_id   seq_id_dst,
+                       llama_pos   p0,
+                       llama_pos   p1),
+            "use llama_kv_self_seq_cp instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_keep(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id),
+            "use llama_kv_self_seq_keep instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_add(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                       llama_pos   delta),
+            "use llama_kv_self_seq_add instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_div(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                             int   d),
+            "use llama_kv_self_seq_div instead");
+
+    DEPRECATED(LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id),
+            "use llama_kv_self_seq_pos_max instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx),
+            "use llama_kv_self_defrag instead");
+
+    DEPRECATED(LLAMA_API bool llama_kv_cache_can_shift(const struct llama_context * ctx),
+            "use llama_kv_self_can_shift instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_update(struct llama_context * ctx),
+            "use llama_kv_self_update instead");
+
 
     //
     // State / sessions
@@ -891,6 +964,10 @@ extern "C" {
     // If set to true, the model will only attend to the past tokens
     LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn);
 
+    // Set whether the model is in warmup mode or not
+    // If true, all model tensors are activated during llama_decode() to load and cache their weights.
+    LLAMA_API void llama_set_warmup(struct llama_context * ctx, bool warmup);
+
     // Set abort callback
     LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data);
 
@@ -1206,22 +1283,38 @@ extern "C" {
                                float   tau,
                                float   eta);
 
+    /// @details Intializes a GBNF grammar, see grammars/README.md for details.
+    /// @param vocab The vocabulary that this grammar will be used with.
+    /// @param grammar_str The production rules for the grammar, encoded as a string. Returns an empty grammar if empty. Returns NULL if parsing of grammar_str fails.
+    /// @param grammar_root The name of the start symbol for the grammar.
     LLAMA_API struct llama_sampler * llama_sampler_init_grammar(
             const struct llama_vocab * vocab,
                           const char * grammar_str,
                           const char * grammar_root);
 
-    /// @details Lazy grammar sampler, introduced in https://github.com/ggml-org/llama.cpp/pull/9639
-    /// @param trigger_words A list of words that will trigger the grammar sampler. This may be updated to a loose regex syntax (w/ ^) in a near future.
-    /// @param trigger_tokens A list of tokens that will trigger the grammar sampler.
-    LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy(
+    DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy(
             const struct llama_vocab * vocab,
                           const char * grammar_str,
                           const char * grammar_root,
                          const char ** trigger_words,
                                 size_t num_trigger_words,
                    const llama_token * trigger_tokens,
-                                size_t num_trigger_tokens);
+                                size_t num_trigger_tokens),
+        "use llama_sampler_init_grammar_lazy_patterns instead");
+
+
+    /// @details Lazy grammar sampler, introduced in https://github.com/ggml-org/llama.cpp/pull/9639
+    /// @param trigger_patterns A list of patterns that will trigger the grammar sampler. Pattern will be matched from the start of the generation output, and grammar sampler will be fed content starting from its first match group.
+    /// @param trigger_tokens A list of tokens that will trigger the grammar sampler. Grammar sampler will be fed content starting from the trigger token included.
+    LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy_patterns(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens);
+
 
     /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first.
     LLAMA_API struct llama_sampler * llama_sampler_init_penalties(

llama/llama.cpp/src/llama-adapter.cpp

@@ -4,14 +4,13 @@
 #include "llama-mmap.h"
 #include "llama-model.h"
 
-#include <algorithm>
 #include <map>
 #include <cassert>
 #include <stdexcept>
 
 // vec
 
-struct ggml_tensor * llama_adapter_cvec::tensor_for(int il) const {
+ggml_tensor * llama_adapter_cvec::tensor_for(int il) const {
     if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) {
         return nullptr;
     }
@@ -19,7 +18,7 @@ struct ggml_tensor * llama_adapter_cvec::tensor_for(int il) const {
     return tensors[il];
 }
 
-struct ggml_tensor * llama_adapter_cvec::apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const {
+ggml_tensor * llama_adapter_cvec::apply_to(ggml_context * ctx, ggml_tensor * cur, int  il) const {
     ggml_tensor * layer_dir = tensor_for(il);
     if (layer_dir != nullptr) {
         cur = ggml_add(ctx, cur, layer_dir);
@@ -40,7 +39,7 @@ bool llama_adapter_cvec::init(const llama_model & model) {
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
-            struct ggml_init_params params = {
+            ggml_init_params params = {
                 /*.mem_size   =*/ hparams.n_layer*ggml_tensor_overhead(),
                 /*.mem_buffer =*/ NULL,
                 /*.no_alloc   =*/ true,
@@ -91,7 +90,7 @@ bool llama_adapter_cvec::init(const llama_model & model) {
     return true;
 }
 
-int32_t llama_adapter_cvec::apply(
+bool llama_adapter_cvec::apply(
         const llama_model & model,
         const float * data,
         size_t len,
@@ -104,17 +103,17 @@ int32_t llama_adapter_cvec::apply(
         // disable the current control vector (but leave allocated for later)
         layer_start = -1;
         layer_end   = -1;
-        return 0;
+        return true;
     }
 
     if (n_embd != (int) hparams.n_embd) {
         LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__);
-        return 1;
+        return false;
     }
 
     if (tensors.empty()) {
         if (!init(model)) {
-            return 1;
+            return false;
         }
     }
 
@@ -130,12 +129,12 @@ int32_t llama_adapter_cvec::apply(
         }
     }
 
-    return 0;
+    return true;
 }
 
 // lora
 
-llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor * w) {
+llama_adapter_lora_weight * llama_adapter_lora::get_weight(ggml_tensor * w) {
     const std::string name(w->name);
 
     const auto pos = ab_map.find(name);
@@ -146,11 +145,11 @@ llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor *
     return nullptr;
 }
 
-static void llama_adapter_lora_init_impl(struct llama_model & model, const char * path_lora, struct llama_adapter_lora & adapter) {
+static void llama_adapter_lora_init_impl(llama_model & model, const char * path_lora, llama_adapter_lora & adapter) {
     LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
 
     ggml_context * ctx_init;
-    struct gguf_init_params meta_gguf_params = {
+    gguf_init_params meta_gguf_params = {
         /* .no_alloc = */ true,
         /* .ctx      = */ &ctx_init,
     };
@@ -201,7 +200,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
             // add a new context
-            struct ggml_init_params params = {
+            ggml_init_params params = {
                 /*.mem_size   =*/ n_tensors*ggml_tensor_overhead(),
                 /*.mem_buffer =*/ NULL,
                 /*.no_alloc   =*/ true,
@@ -248,6 +247,26 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
         }
     }
 
+    // get extra buffer types of the CPU
+    // TODO: a more general solution for non-CPU extra buft should be imlpemented in the future
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/12593#pullrequestreview-2718659948
+    std::vector<ggml_backend_buffer_type_t> buft_extra;
+    {
+        auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+
+        auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+            ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
+
+        if (ggml_backend_dev_get_extra_bufts_fn) {
+            ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
+            while (extra_bufts && *extra_bufts) {
+                buft_extra.emplace_back(*extra_bufts);
+                ++extra_bufts;
+            }
+        }
+    }
+
     // add tensors
     for (auto & it : ab_map) {
         const std::string & name = it.first;
@@ -264,7 +283,23 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
             throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong base model?)");
         }
 
-        struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
+        auto * buft = ggml_backend_buffer_get_type(model_tensor->buffer);
+
+        // do not load loras to extra buffer types (i.e. bufts for repacking) -> use the CPU in that case
+        for (auto & ex : buft_extra) {
+            if (ex == buft) {
+                LLAMA_LOG_WARN("%s: lora for '%s' cannot use buft '%s', fallback to CPU\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
+
+                auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+                buft = ggml_backend_dev_buffer_type(cpu_dev);
+
+                break;
+            }
+        }
+
+        LLAMA_LOG_DEBUG("%s: lora for '%s' -> '%s'\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
+
+        ggml_context * dev_ctx = ctx_for_buft(buft);
         // validate tensor shape
         if (is_token_embd) {
             // expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd()
@@ -281,8 +316,8 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
         }
 
         // save tensor to adapter
-        struct ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
-        struct ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
+        ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
+        ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
         ggml_set_name(tensor_a, w.a->name);
         ggml_set_name(tensor_b, w.b->name);
         adapter.ab_map[name] = llama_adapter_lora_weight(tensor_a, tensor_b);
@@ -308,7 +343,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
     {
         llama_file gguf_file(path_lora, "rb");
         std::vector<uint8_t> read_buf;
-        auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) {
+        auto set_tensor = [&](ggml_tensor * orig, ggml_tensor * dev) {
             size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name));
             size_t size = ggml_nbytes(orig);
             read_buf.resize(size);
@@ -327,8 +362,8 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
     LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
 }
 
-struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model, const char * path_lora) {
-    struct llama_adapter_lora * adapter = new llama_adapter_lora();
+llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * path_lora) {
+    llama_adapter_lora * adapter = new llama_adapter_lora();
 
     try {
         llama_adapter_lora_init_impl(*model, path_lora, *adapter);
@@ -342,6 +377,6 @@ struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model,
     return nullptr;
 }
 
-void llama_adapter_lora_free(struct llama_adapter_lora * adapter) {
+void llama_adapter_lora_free(llama_adapter_lora * adapter) {
     delete adapter;
 }

llama/llama.cpp/src/llama-adapter.h

@@ -15,11 +15,11 @@
 //
 
 struct llama_adapter_cvec {
-    struct ggml_tensor * tensor_for(int il) const;
+    ggml_tensor * tensor_for(int il) const;
 
-    struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const;
+    ggml_tensor * apply_to(ggml_context * ctx, ggml_tensor * cur, int  il) const;
 
-    int32_t apply(
+    bool apply(
             const llama_model & model,
             const float * data,
             size_t len,
@@ -36,7 +36,7 @@ private:
     std::vector<ggml_context_ptr> ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
 
-    std::vector<struct ggml_tensor *> tensors; // per layer
+    std::vector<ggml_tensor *> tensors; // per layer
 };
 
 //
@@ -44,8 +44,8 @@ private:
 //
 
 struct llama_adapter_lora_weight {
-    struct ggml_tensor * a = nullptr;
-    struct ggml_tensor * b = nullptr;
+    ggml_tensor * a = nullptr;
+    ggml_tensor * b = nullptr;
 
     // get actual scale based on rank and alpha
     float get_scale(float alpha, float adapter_scale) const {
@@ -55,12 +55,12 @@ struct llama_adapter_lora_weight {
     }
 
     llama_adapter_lora_weight() = default;
-    llama_adapter_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b) : a(a), b(b) {}
+    llama_adapter_lora_weight(ggml_tensor * a, ggml_tensor * b) : a(a), b(b) {}
 };
 
 struct llama_adapter_lora {
     // map tensor name to lora_a_b
-    std::unordered_map<std::string, struct llama_adapter_lora_weight> ab_map;
+    std::unordered_map<std::string, llama_adapter_lora_weight> ab_map;
 
     std::vector<ggml_context_ptr> ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
@@ -70,5 +70,7 @@ struct llama_adapter_lora {
     llama_adapter_lora() = default;
     ~llama_adapter_lora() = default;
 
-    llama_adapter_lora_weight * get_weight(struct ggml_tensor * w);
+    llama_adapter_lora_weight * get_weight(ggml_tensor * w);
 };
+
+using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;

llama/llama.cpp/src/llama-arch.cpp

@@ -7,6 +7,7 @@
 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_LLAMA,            "llama"            },
     { LLM_ARCH_MLLAMA,           "mllama"           },
+    { LLM_ARCH_LLAMA4,           "llama4"           },
     { LLM_ARCH_DECI,             "deci"             },
     { LLM_ARCH_FALCON,           "falcon"           },
     { LLM_ARCH_GROK,             "grok"             },
@@ -26,6 +27,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_QWEN2,            "qwen2"            },
     { LLM_ARCH_QWEN2MOE,         "qwen2moe"         },
     { LLM_ARCH_QWEN2VL,          "qwen2vl"          },
+    { LLM_ARCH_QWEN3,            "qwen3"            },
+    { LLM_ARCH_QWEN3MOE,         "qwen3moe"         },
     { LLM_ARCH_PHI2,             "phi2"             },
     { LLM_ARCH_PHI3,             "phi3"             },
     { LLM_ARCH_PHIMOE,           "phimoe"           },
@@ -52,6 +55,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DEEPSEEK,         "deepseek"         },
     { LLM_ARCH_DEEPSEEK2,        "deepseek2"        },
     { LLM_ARCH_CHATGLM,          "chatglm"          },
+    { LLM_ARCH_GLM4,             "glm4"             },
     { LLM_ARCH_BITNET,           "bitnet"           },
     { LLM_ARCH_T5,               "t5"               },
     { LLM_ARCH_T5ENCODER,        "t5encoder"        },
@@ -60,11 +64,16 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_EXAONE,           "exaone"           },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
     { LLM_ARCH_RWKV6QWEN2,       "rwkv6qwen2"       },
+    { LLM_ARCH_RWKV7,            "rwkv7"            },
+    { LLM_ARCH_ARWKV7,           "arwkv7"           },
     { LLM_ARCH_GRANITE,          "granite"          },
     { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
     { LLM_ARCH_CHAMELEON,        "chameleon"        },
     { LLM_ARCH_SOLAR,            "solar"            },
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
+    { LLM_ARCH_PLM,              "plm"              },
+    { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
+    { LLM_ARCH_MISTRAL3,         "mistral3"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -73,6 +82,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_GENERAL_ARCHITECTURE,         "general.architecture"                  },
     { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version"          },
     { LLM_KV_GENERAL_ALIGNMENT,            "general.alignment"                     },
+    { LLM_KV_GENERAL_FILE_TYPE,            "general.file_type"                     },
     { LLM_KV_GENERAL_NAME,                 "general.name"                          },
     { LLM_KV_GENERAL_AUTHOR,               "general.author"                        },
     { LLM_KV_GENERAL_VERSION,              "general.version"                       },
@@ -111,25 +121,32 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_RESIDUAL_SCALE,                    "%s.residual_scale"                    },
     { LLM_KV_EMBEDDING_SCALE,                   "%s.embedding_scale"                   },
     { LLM_KV_TOKEN_SHIFT_COUNT,                 "%s.token_shift_count"                 },
+    { LLM_KV_INTERLEAVE_MOE_LAYER_STEP,         "%s.interleave_moe_layer_step"         },
 
-    { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"             },
-    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"          },
-    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"         },
-    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"              },
-    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"             },
-    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"           },
-    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"     },
-    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon" },
-    { LLM_KV_ATTENTION_GROUPNORM_EPS,          "%s.attention.group_norm_epsilon"     },
-    { LLM_KV_ATTENTION_GROUPNORM_GROUPS,       "%s.attention.group_norm_groups"      },
-    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                 },
-    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"            },
-    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"           },
-    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
-    { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"         },
-    { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                  },
-    { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,  "%s.attention.block_skip_connection"  },
-    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, "%s.attention.cross_attention_layers" },
+    { LLM_KV_ATTENTION_HEAD_COUNT,                   "%s.attention.head_count"                   },
+    { LLM_KV_ATTENTION_HEAD_COUNT_KV,                "%s.attention.head_count_kv"                },
+    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,               "%s.attention.max_alibi_bias"               },
+    { LLM_KV_ATTENTION_CLAMP_KQV,                    "%s.attention.clamp_kqv"                    },
+    { LLM_KV_ATTENTION_KEY_LENGTH,                   "%s.attention.key_length"                   },
+    { LLM_KV_ATTENTION_VALUE_LENGTH,                 "%s.attention.value_length"                 },
+    { LLM_KV_ATTENTION_LAYERNORM_EPS,                "%s.attention.layer_norm_epsilon"           },
+    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,            "%s.attention.layer_norm_rms_epsilon"       },
+    { LLM_KV_ATTENTION_GROUPNORM_EPS,                "%s.attention.group_norm_epsilon"           },
+    { LLM_KV_ATTENTION_GROUPNORM_GROUPS,             "%s.attention.group_norm_groups"            },
+    { LLM_KV_ATTENTION_CAUSAL,                       "%s.attention.causal"                       },
+    { LLM_KV_ATTENTION_Q_LORA_RANK,                  "%s.attention.q_lora_rank"                  },
+    { LLM_KV_ATTENTION_KV_LORA_RANK,                 "%s.attention.kv_lora_rank"                 },
+    { LLM_KV_ATTENTION_DECAY_LORA_RANK,              "%s.attention.decay_lora_rank"              },
+    { LLM_KV_ATTENTION_ICLR_LORA_RANK,               "%s.attention.iclr_lora_rank"               },
+    { LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, "%s.attention.value_residual_mix_lora_rank" },
+    { LLM_KV_ATTENTION_GATE_LORA_RANK,               "%s.attention.gate_lora_rank"               },
+    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
+    { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
+    { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
+    { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
+    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,       "%s.attention.cross_attention_layers"       },
+    { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
+    { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -228,6 +245,35 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_LLAMA4,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
+            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
+            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,  "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
+        },
+    },
     {
         LLM_ARCH_MLLAMA,
         {
@@ -593,6 +639,45 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
         },
     },
+    {
+        LLM_ARCH_QWEN3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_QWEN3MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_PHI2,
         {
@@ -810,9 +895,12 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
         {
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
             { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
             { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
             { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
             { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
             { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
             { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
             { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
@@ -1056,6 +1144,8 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_ATTN_Q_B,           "blk.%d.attn_q_b" },
             { LLM_TENSOR_ATTN_KV_A_MQA,      "blk.%d.attn_kv_a_mqa" },
             { LLM_TENSOR_ATTN_KV_B,          "blk.%d.attn_kv_b" },
+            { LLM_TENSOR_ATTN_K_B,           "blk.%d.attn_k_b" },
+            { LLM_TENSOR_ATTN_V_B,           "blk.%d.attn_v_b" },
             { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
             { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
             { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
@@ -1072,6 +1162,22 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
         },
     },
+    {
+        LLM_ARCH_PLM,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_KV_A_MQA,      "blk.%d.attn_kv_a_mqa" },
+            { LLM_TENSOR_ATTN_KV_A_NORM,     "blk.%d.attn_kv_a_norm" },
+            { LLM_TENSOR_ATTN_KV_B,          "blk.%d.attn_kv_b" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_CHATGLM,
         {
@@ -1090,6 +1196,25 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
         },
     },
+    {
+        LLM_ARCH_GLM4,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
     {
         LLM_ARCH_BITNET,
         {
@@ -1274,6 +1399,74 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,                    "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_RWKV7,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,                "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM,           "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,               "output_norm" },
+            { LLM_TENSOR_OUTPUT,                    "output" },
+            { LLM_TENSOR_ATTN_NORM,                 "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_NORM_2,               "blk.%d.attn_norm_2" },
+            { LLM_TENSOR_TIME_MIX_W0,               "blk.%d.time_mix_w0" },
+            { LLM_TENSOR_TIME_MIX_W1,               "blk.%d.time_mix_w1" },
+            { LLM_TENSOR_TIME_MIX_W2,               "blk.%d.time_mix_w2" },
+            { LLM_TENSOR_TIME_MIX_A0,               "blk.%d.time_mix_a0" },
+            { LLM_TENSOR_TIME_MIX_A1,               "blk.%d.time_mix_a1" },
+            { LLM_TENSOR_TIME_MIX_A2,               "blk.%d.time_mix_a2" },
+            { LLM_TENSOR_TIME_MIX_V0,               "blk.%d.time_mix_v0" },
+            { LLM_TENSOR_TIME_MIX_V1,               "blk.%d.time_mix_v1" },
+            { LLM_TENSOR_TIME_MIX_V2,               "blk.%d.time_mix_v2" },
+            { LLM_TENSOR_TIME_MIX_G1,               "blk.%d.time_mix_g1" },
+            { LLM_TENSOR_TIME_MIX_G2,               "blk.%d.time_mix_g2" },
+            { LLM_TENSOR_TIME_MIX_K_K,              "blk.%d.time_mix_k_k" },
+            { LLM_TENSOR_TIME_MIX_K_A,              "blk.%d.time_mix_k_a" },
+            { LLM_TENSOR_TIME_MIX_R_K,              "blk.%d.time_mix_r_k" },
+            { LLM_TENSOR_TIME_MIX_LERP_FUSED,       "blk.%d.time_mix_lerp_fused" },
+            { LLM_TENSOR_TIME_MIX_KEY,              "blk.%d.time_mix_key" },
+            { LLM_TENSOR_TIME_MIX_VALUE,            "blk.%d.time_mix_value" },
+            { LLM_TENSOR_TIME_MIX_RECEPTANCE,       "blk.%d.time_mix_receptance" },
+            { LLM_TENSOR_TIME_MIX_LN,               "blk.%d.time_mix_ln" },
+            { LLM_TENSOR_TIME_MIX_OUTPUT,           "blk.%d.time_mix_output" },
+            { LLM_TENSOR_CHANNEL_MIX_LERP_K,        "blk.%d.channel_mix_lerp_k" },
+            { LLM_TENSOR_CHANNEL_MIX_KEY,           "blk.%d.channel_mix_key" },
+            { LLM_TENSOR_CHANNEL_MIX_VALUE,         "blk.%d.channel_mix_value" },
+        },
+    },
+    {
+        LLM_ARCH_ARWKV7,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,                "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM,           "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,               "output_norm" },
+            { LLM_TENSOR_OUTPUT,                    "output" },
+            { LLM_TENSOR_ATTN_NORM,                 "blk.%d.attn_norm" },
+            { LLM_TENSOR_TIME_MIX_W0,               "blk.%d.time_mix_w0" },
+            { LLM_TENSOR_TIME_MIX_W1,               "blk.%d.time_mix_w1" },
+            { LLM_TENSOR_TIME_MIX_W2,               "blk.%d.time_mix_w2" },
+            { LLM_TENSOR_TIME_MIX_A0,               "blk.%d.time_mix_a0" },
+            { LLM_TENSOR_TIME_MIX_A1,               "blk.%d.time_mix_a1" },
+            { LLM_TENSOR_TIME_MIX_A2,               "blk.%d.time_mix_a2" },
+            { LLM_TENSOR_TIME_MIX_V0,               "blk.%d.time_mix_v0" },
+            { LLM_TENSOR_TIME_MIX_V1,               "blk.%d.time_mix_v1" },
+            { LLM_TENSOR_TIME_MIX_V2,               "blk.%d.time_mix_v2" },
+            { LLM_TENSOR_TIME_MIX_G1,               "blk.%d.time_mix_g1" },
+            { LLM_TENSOR_TIME_MIX_G2,               "blk.%d.time_mix_g2" },
+            { LLM_TENSOR_TIME_MIX_K_K,              "blk.%d.time_mix_k_k" },
+            { LLM_TENSOR_TIME_MIX_K_A,              "blk.%d.time_mix_k_a" },
+            { LLM_TENSOR_TIME_MIX_R_K,              "blk.%d.time_mix_r_k" },
+            { LLM_TENSOR_TIME_MIX_LERP_FUSED,       "blk.%d.time_mix_lerp_fused" },
+            { LLM_TENSOR_TIME_MIX_KEY,              "blk.%d.time_mix_key" },
+            { LLM_TENSOR_TIME_MIX_VALUE,            "blk.%d.time_mix_value" },
+            { LLM_TENSOR_TIME_MIX_RECEPTANCE,       "blk.%d.time_mix_receptance" },
+            { LLM_TENSOR_TIME_MIX_LN,               "blk.%d.time_mix_ln" },
+            { LLM_TENSOR_TIME_MIX_OUTPUT,           "blk.%d.time_mix_output" },
+            { LLM_TENSOR_FFN_NORM,                  "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,                  "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,                  "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,                    "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_GRANITE,
         {
@@ -1371,6 +1564,45 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_POS_NET_ATTN_OUT,  "posnet.%d.attn_output" },
         },
     },
+    {
+        LLM_ARCH_BAILINGMOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ROPE_FREQS,         "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+        },
+    },
+    {
+        LLM_ARCH_MISTRAL3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,  "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
+            { LLM_TENSOR_ATTN_NORM,   "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,      "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,      "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,      "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,    "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,    "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,    "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_UP,      "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,    "blk.%d.ffn_down" },
+        }
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -1408,23 +1640,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_K,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_V,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_QKV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_OUT,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP_SHEXP,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_A,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_K_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_V_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_V,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -1451,6 +1668,12 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SSM_OUT,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_A1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_A2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_V1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_V2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_G1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_G2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_DECAY_W1,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_DECAY_W2,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_KEY,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -1469,6 +1692,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_TIME_MIX_LN,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_CHANNEL_MIX_LERP_K,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_CHANNEL_MIX_LERP_R,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_K_K,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_K_A,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_R_K,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_TIME_MIX_LERP_W,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_K,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_V,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
@@ -1476,6 +1702,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_TIME_MIX_LERP_G,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_FUSED,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_DECAY,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_W0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_A0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_V0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_FIRST,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_RWKV_WKV6}},
     {LLM_TENSOR_ATTN_NORM,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_ATTN_NORM_2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},

llama/llama.cpp/src/llama-arch.h

@@ -10,6 +10,7 @@
 
 enum llm_arch {
     LLM_ARCH_LLAMA,
+    LLM_ARCH_LLAMA4,
     LLM_ARCH_MLLAMA,
     LLM_ARCH_DECI,
     LLM_ARCH_FALCON,
@@ -30,6 +31,8 @@ enum llm_arch {
     LLM_ARCH_QWEN2,
     LLM_ARCH_QWEN2MOE,
     LLM_ARCH_QWEN2VL,
+    LLM_ARCH_QWEN3,
+    LLM_ARCH_QWEN3MOE,
     LLM_ARCH_PHI2,
     LLM_ARCH_PHI3,
     LLM_ARCH_PHIMOE,
@@ -56,6 +59,7 @@ enum llm_arch {
     LLM_ARCH_DEEPSEEK,
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
+    LLM_ARCH_GLM4,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
@@ -64,11 +68,16 @@ enum llm_arch {
     LLM_ARCH_EXAONE,
     LLM_ARCH_RWKV6,
     LLM_ARCH_RWKV6QWEN2,
+    LLM_ARCH_RWKV7,
+    LLM_ARCH_ARWKV7,
     LLM_ARCH_GRANITE,
     LLM_ARCH_GRANITE_MOE,
     LLM_ARCH_CHAMELEON,
     LLM_ARCH_SOLAR,
     LLM_ARCH_WAVTOKENIZER_DEC,
+    LLM_ARCH_MISTRAL3,
+    LLM_ARCH_PLM,
+    LLM_ARCH_BAILINGMOE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -77,6 +86,7 @@ enum llm_kv {
     LLM_KV_GENERAL_ARCHITECTURE,
     LLM_KV_GENERAL_QUANTIZATION_VERSION,
     LLM_KV_GENERAL_ALIGNMENT,
+    LLM_KV_GENERAL_FILE_TYPE,
     LLM_KV_GENERAL_NAME,
     LLM_KV_GENERAL_AUTHOR,
     LLM_KV_GENERAL_VERSION,
@@ -115,6 +125,7 @@ enum llm_kv {
     LLM_KV_RESIDUAL_SCALE,
     LLM_KV_EMBEDDING_SCALE,
     LLM_KV_TOKEN_SHIFT_COUNT,
+    LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
 
     LLM_KV_ATTENTION_HEAD_COUNT,
     LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -129,11 +140,17 @@ enum llm_kv {
     LLM_KV_ATTENTION_CAUSAL,
     LLM_KV_ATTENTION_Q_LORA_RANK,
     LLM_KV_ATTENTION_KV_LORA_RANK,
+    LLM_KV_ATTENTION_DECAY_LORA_RANK,
+    LLM_KV_ATTENTION_ICLR_LORA_RANK,
+    LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK,
+    LLM_KV_ATTENTION_GATE_LORA_RANK,
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
     LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,
+    LLM_KV_ATTENTION_KEY_LENGTH_MLA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -247,6 +264,8 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_NORM,
     LLM_TENSOR_ATTN_K_NORM,
     LLM_TENSOR_LAYER_OUT_NORM,
+    LLM_TENSOR_POST_ATTN_NORM,
+    LLM_TENSOR_POST_MLP_NORM,
     LLM_TENSOR_SSM_IN,
     LLM_TENSOR_SSM_CONV1D,
     LLM_TENSOR_SSM_X,
@@ -254,8 +273,20 @@ enum llm_tensor {
     LLM_TENSOR_SSM_A,
     LLM_TENSOR_SSM_D,
     LLM_TENSOR_SSM_OUT,
+    LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
     LLM_TENSOR_TIME_MIX_W2,
+    LLM_TENSOR_TIME_MIX_A0,
+    LLM_TENSOR_TIME_MIX_A1,
+    LLM_TENSOR_TIME_MIX_A2,
+    LLM_TENSOR_TIME_MIX_V0,
+    LLM_TENSOR_TIME_MIX_V1,
+    LLM_TENSOR_TIME_MIX_V2,
+    LLM_TENSOR_TIME_MIX_G1,
+    LLM_TENSOR_TIME_MIX_G2,
+    LLM_TENSOR_TIME_MIX_K_K,
+    LLM_TENSOR_TIME_MIX_K_A,
+    LLM_TENSOR_TIME_MIX_R_K,
     LLM_TENSOR_TIME_MIX_LERP_X,
     LLM_TENSOR_TIME_MIX_LERP_W,
     LLM_TENSOR_TIME_MIX_LERP_K,
@@ -282,6 +313,8 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_B,
     LLM_TENSOR_ATTN_KV_A_MQA,
     LLM_TENSOR_ATTN_KV_B,
+    LLM_TENSOR_ATTN_K_B,
+    LLM_TENSOR_ATTN_V_B,
     LLM_TENSOR_ATTN_Q_A_NORM,
     LLM_TENSOR_ATTN_KV_A_NORM,
     LLM_TENSOR_ATTN_SUB_NORM,

llama/llama.cpp/src/llama-batch.h

@@ -42,9 +42,9 @@ struct llama_sbatch {
     bool logits_all; // TODO: remove once lctx.logits_all is removed too
 
     // sorted indices into the batch
-    std::vector<size_t> ids;
+    std::vector<int64_t> ids;
     // batch indices of the output
-    std::vector<size_t> out_ids;
+    std::vector<int64_t> out_ids;
     std::vector<llama_sbatch_seq> seq;
 
     const llama_batch * batch = nullptr;

llama/llama.cpp/src/llama-chat.cpp

@@ -4,6 +4,7 @@
 
 #include <map>
 #include <sstream>
+#include <algorithm>
 
 #if __cplusplus >= 202000L
     #define LU8(x) (const char*)(u8##x)
@@ -58,6 +59,9 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
     { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
     { "megrez",            LLM_CHAT_TEMPLATE_MEGREZ            },
+    { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
+    { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
+    { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -167,6 +171,12 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_GIGACHAT;
     } else if (tmpl_contains("<|role_start|>")) {
         return LLM_CHAT_TEMPLATE_MEGREZ;
+    } else if (tmpl_contains(" Ассистент:")) {
+        return LLM_CHAT_TEMPLATE_YANDEX;
+    } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("'HUMAN'")) {
+        return LLM_CHAT_TEMPLATE_BAILING;
+    } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
+        return LLM_CHAT_TEMPLATE_LLAMA4;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -566,7 +576,51 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|role_start|>assistant<|role_end|>";
         }
-    } else {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_YANDEX) {
+        // Yandex template ("\n\n" is defined as EOT token)
+
+        ss << "<s>";
+
+        for (size_t i = 0; i < chat.size(); i++) {
+            std::string role(chat[i]->role);
+            if (role == "user") {
+                ss << " Пользователь: " << chat[i]->content << "\n\n";
+            } else if (role == "assistant") {
+                ss << " Ассистент: " << chat[i]->content << "\n\n";
+            }
+        }
+
+        // Add generation prompt if needed
+        if (add_ass) {
+            ss << " Ассистент:[SEP]";
+        }
+    }  else if (tmpl == LLM_CHAT_TEMPLATE_BAILING) {
+        // Bailing (Ling) template
+        for (auto message : chat) {
+            std::string role(message->role);
+
+            if (role == "user") {
+                role = "HUMAN";
+            } else {
+                std::transform(role.begin(), role.end(), role.begin(), ::toupper);
+            }
+
+            ss << "<role>" << role << "</role>" << message->content;
+        }
+
+        if (add_ass) {
+            ss << "<role>ASSISTANT</role>";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_LLAMA4) {
+        // Llama 4
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|header_start|>" << role << "<|header_end|>\n\n" << trim(message->content) << "<|eot|>";
+        }
+        if (add_ass) {
+            ss << "<|header_start|>assistant<|header_end|>\n\n";
+        }
+    }  else {
         // template not supported
         return -1;
     }
@@ -584,4 +638,3 @@ int32_t llama_chat_builtin_templates(const char ** output, size_t len) {
     }
     return (int32_t) LLM_CHAT_TEMPLATES.size();
 }
-

llama/llama.cpp/src/llama-chat.h

@@ -38,6 +38,9 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_GRANITE,
     LLM_CHAT_TEMPLATE_GIGACHAT,
     LLM_CHAT_TEMPLATE_MEGREZ,
+    LLM_CHAT_TEMPLATE_YANDEX,
+    LLM_CHAT_TEMPLATE_BAILING,
+    LLM_CHAT_TEMPLATE_LLAMA4,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

llama/llama.cpp/src/llama-context.h

@@ -3,66 +3,216 @@
 #include "llama.h"
 #include "llama-batch.h"
 #include "llama-cparams.h"
-#include "llama-model.h"
-#include "llama-kv-cache.h"
+#include "llama-graph.h"
 #include "llama-adapter.h"
+#include "llama-kv-cache.h"
 
 #include "ggml-cpp.h"
 
 #include <map>
-#include <unordered_map>
 #include <vector>
-#include <set>
+
+struct llama_model;
+struct llama_kv_cache;
+
+class llama_io_read_i;
+class llama_io_write_i;
 
 struct llama_context {
-    llama_context(const llama_model & model)
-        : model(model)
-        , t_start_us(model.t_start_us)
-        , t_load_us(model.t_load_us) {}
+    // init scheduler and compute buffers, reserve worst-case graphs
+    llama_context(
+            const llama_model & model,
+                  llama_context_params params);
 
-    const struct llama_model & model;
+    ~llama_context();
 
-    struct llama_cparams      cparams;
-    struct llama_sbatch       sbatch;  // TODO: revisit if needed
-    struct llama_kv_cache     kv_self;
-    struct llama_adapter_cvec cvec;
+    void synchronize();
 
-    std::unordered_map<struct llama_adapter_lora *, float> lora;
+    const llama_model & get_model() const;
 
-    std::vector<ggml_backend_ptr> backends;
-    std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
+    uint32_t n_ctx()         const;
+    uint32_t n_ctx_per_seq() const;
+    uint32_t n_batch()       const;
+    uint32_t n_ubatch()      const;
+    uint32_t n_seq_max()     const;
 
-    ggml_backend_t backend_cpu = nullptr;
+    uint32_t n_threads()       const;
+    uint32_t n_threads_batch() const;
 
-    ggml_threadpool_t threadpool       = nullptr;
-    ggml_threadpool_t threadpool_batch = nullptr;
+          llama_kv_cache * get_kv_self();
+    const llama_kv_cache * get_kv_self() const;
 
-    bool has_evaluated_once = false;
+    void kv_self_update();
 
-    mutable int64_t t_start_us;
-    mutable int64_t t_load_us;
-    mutable int64_t t_p_eval_us = 0;
-    mutable int64_t t_eval_us   = 0;
+    enum llama_pooling_type pooling_type() const;
 
-    mutable int64_t t_compute_start_us = 0;
-    mutable int64_t n_queued_tokens = 0;
+    float * get_logits();
+    float * get_logits_ith(int32_t i);
 
-    mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
-    mutable int32_t n_eval   = 0; // number of eval calls
+    float * get_embeddings();
+    float * get_embeddings_ith(int32_t i);
+    float * get_embeddings_seq(llama_seq_id seq_id);
 
-    // host buffer for the model output (logits and embeddings)
-    ggml_backend_buffer_ptr buf_output;
+    void attach_threadpool(
+            ggml_threadpool_t threadpool,
+            ggml_threadpool_t threadpool_batch);
+
+    void detach_threadpool();
+
+    void set_n_threads(int32_t n_threads, int32_t n_threads_batch);
+
+    void set_abort_callback(bool (*abort_callback)(void * data), void * abort_callback_data);
+
+    void set_embeddings (bool value);
+    void set_causal_attn(bool value);
+    void set_warmup(bool value);
+    void set_cross_attn(bool value);
+
+    void set_adapter_lora(
+            llama_adapter_lora * adapter,
+            float scale);
+
+    bool rm_adapter_lora(
+            llama_adapter_lora * adapter);
+
+    void clear_adapter_lora();
+
+    bool apply_adapter_cvec(
+            const float * data,
+                 size_t   len,
+                int32_t   n_embd,
+                int32_t   il_start,
+                int32_t   il_end);
+
+    int encode(llama_batch & inp_batch);
+    int decode(llama_batch & inp_batch);
+
+    //
+    // state save/load
+    //
+
+    size_t state_get_size();
+    size_t state_get_data(      uint8_t * dst, size_t size);
+    size_t state_set_data(const uint8_t * src, size_t size);
+
+    size_t state_seq_get_size(llama_seq_id seq_id);
+    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size);
+    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size);
+
+    bool state_load_file(
+            const char * filepath,
+           llama_token * tokens_out,
+                size_t   n_token_capacity,
+                size_t * n_token_count_out);
+
+    bool state_save_file(
+            const char * filepath,
+     const llama_token * tokens,
+                size_t   n_token_count);
+
+    size_t state_seq_load_file(
+          llama_seq_id   seq_id,
+            const char * filepath,
+           llama_token * tokens_out,
+                size_t   n_token_capacity,
+                size_t * n_token_count_out);
+
+    size_t state_seq_save_file(
+          llama_seq_id   seq_id,
+            const char * filepath,
+     const llama_token * tokens,
+                size_t   n_token_count);
+
+    //
+    // perf
+    //
+
+    llama_perf_context_data perf_get_data() const;
+    void perf_reset();
+
+private:
+    //
+    // output
+    //
+
+    // Make sure enough space is available for outputs.
+    // Returns max number of outputs for which space was reserved.
+    int32_t output_reserve(int32_t n_outputs);
+
+    // make the outputs have the same order they had in the user-provided batch
+    // TODO: maybe remove this
+    void output_reorder();
+
+    //
+    // graph
+    //
+
+    int32_t graph_max_nodes() const;
+
+    // zero-out inputs and create the ctx_compute for the compute graph
+    ggml_cgraph * graph_init();
+
+    llm_graph_result_ptr graph_build(
+            ggml_context * ctx,
+             ggml_cgraph * gf,
+      const llama_ubatch & ubatch,
+          llm_graph_type   gtype);
+
+    // returns the result of ggml_backend_sched_graph_compute_async execution
+    ggml_status graph_compute(
+            ggml_cgraph * gf,
+                   bool   batched);
+
+    llm_graph_cb graph_get_cb() const;
+
+    // used by kv_self_update()
+    ggml_tensor * build_rope_shift(
+        ggml_context * ctx0,
+        ggml_tensor * cur,
+        ggml_tensor * shift,
+        ggml_tensor * factors,
+              float   freq_base,
+              float   freq_scale,
+        ggml_backend_buffer * bbuf) const;
+
+    llm_graph_result_ptr build_kv_self_shift(
+            ggml_context * ctx0,
+            ggml_cgraph * gf) const;
+
+    llm_graph_result_ptr build_kv_self_defrag(
+            ggml_context * ctx0,
+            ggml_cgraph * gf,
+            const std::vector<struct llama_kv_defrag_move> & moves) const;
+
+    // TODO: read/write lora adapters and cvec
+    size_t state_write_data(llama_io_write_i & io);
+    size_t state_read_data (llama_io_read_i  & io);
+
+    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id);
+    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id);
+
+    //
+    // members
+    //
+
+    const llama_model & model;
+
+    llama_cparams       cparams;
+    llama_adapter_cvec  cvec;
+    llama_adapter_loras loras;
+    llama_sbatch        sbatch;
+
+    llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably
+
+    std::unique_ptr<llama_kv_cache_unified> kv_self;
+
+    // TODO: remove
+    bool logits_all = false;
 
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     size_t  logits_size = 0; // capacity (of floats) for logits
     float * logits      = nullptr;
 
-    std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
-    size_t  output_size = 0; // capacity (of tokens positions) for the output buffers
-    int32_t n_outputs   = 0; // number of actually-used outputs in the current ubatch or last logical batch
-
-    bool logits_all = false;
-
     // embeddings output (2-dimensional array: [n_outputs][n_embd])
     // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE
     size_t  embd_size = 0; // capacity (of floats) for embeddings
@@ -72,59 +222,47 @@ struct llama_context {
     // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
     std::map<llama_seq_id, std::vector<float>> embd_seq;
 
-    // whether we are computing encoder output or decoder output
-    bool is_encoding = false;
+    int32_t n_outputs     = 0; // number of actually-used outputs in the current ubatch or last logical batch
+    int32_t n_outputs_max = 0; // capacity (of tokens positions) for the output buffers
 
-    // TODO: find a better way to accommodate mutli-dimension position encoding methods
-    // number of position id each token get, 1 for each token in most cases.
-    // when using m-rope, it will be 3 position ids per token to representing 3 dimension coordinate.
-    int n_pos_per_token = 1;
+    std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
 
-    // output of the encoder part of the encoder-decoder models
-    std::vector<float> embd_enc;
-    std::vector<std::set<llama_seq_id>> seq_ids_enc;
-
-    // memory buffers used to evaluate the model
-    std::vector<uint8_t> buf_compute_meta;
     ggml_backend_sched_ptr sched;
 
+    ggml_backend_t backend_cpu = nullptr;
+    std::vector<ggml_backend_ptr> backends;
+
+    ggml_context_ptr ctx_compute;
+
+    ggml_threadpool_t threadpool       = nullptr;
+    ggml_threadpool_t threadpool_batch = nullptr;
+
     ggml_abort_callback abort_callback      = nullptr;
     void *              abort_callback_data = nullptr;
 
-    // input tensors
-    struct ggml_tensor * inp_tokens;        // I32 [n_batch]
-    struct ggml_tensor * inp_embd;          // F32 [n_embd, n_batch]
-    struct ggml_tensor * inp_pos;           // I32 [n_batch]
-    struct ggml_tensor * inp_out_ids;       // I32 [n_outputs]
-    struct ggml_tensor * inp_KQ_mask;       // F32 [kv_size, n_batch]
-    struct ggml_tensor * inp_KQ_mask_swa;   // F32 [kv_size, n_batch]
-    struct ggml_tensor * inp_K_shift;       // I32 [kv_size]
-    struct ggml_tensor * inp_mean;          // F32 [n_batch, n_batch]
-    struct ggml_tensor * inp_cls;           // I32 [n_batch]
-    struct ggml_tensor * inp_s_copy;        // I32 [kv_size]
-    struct ggml_tensor * inp_s_mask;        // F32 [1, n_kv]
-    struct ggml_tensor * inp_s_seq;         // I32 [n_kv, n_batch]
-    struct ggml_tensor * inp_pos_bucket;    // I32 [n_batch|n_kv, n_batch]
-    struct ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
-    struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
+    std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
 
-    struct ggml_tensor * inp_cross_attn_state; // F32 [4, n_embd, 1061]
+    // buffer types used for the compute buffer of each backend
+    std::vector<ggml_backend_t>             backend_ptrs;
+    std::vector<ggml_backend_buffer_type_t> backend_buft;
+
+    // memory buffers used to evaluate the model
+    std::vector<uint8_t> buf_compute_meta;
+
+    // host buffer for the model output (logits and embeddings)
+    ggml_backend_buffer_ptr buf_output;
+
+    bool has_evaluated_once = false;
+
+    // perf
+    mutable int64_t t_start_us  = 0;
+    mutable int64_t t_load_us   = 0;
+    mutable int64_t t_p_eval_us = 0;
+    mutable int64_t t_eval_us   = 0;
+
+    mutable int64_t t_compute_start_us = 0;
+    mutable int64_t n_queued_tokens    = 0;
+
+    mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
+    mutable int32_t n_eval   = 0; // number of eval calls
 };
-
-// TODO: make these methods of llama_context
-void llama_set_k_shift(struct llama_context & lctx);
-
-void llama_set_s_copy(struct llama_context & lctx);
-
-void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch);
-
-// Make sure enough space is available for outputs.
-// Returns max number of outputs for which space was reserved.
-size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs);
-
-// make the outputs have the same order they had in the user-provided batch
-void llama_output_reorder(struct llama_context & ctx);
-
-// For internal test use
-// TODO: remove
-const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(struct llama_context * ctx);

llama/llama.cpp/src/llama-cparams.h

@@ -30,6 +30,7 @@ struct llama_cparams {
     bool flash_attn;
     bool no_perf;
     bool cross_attn;
+    bool warmup;
 
     enum llama_pooling_type pooling_type;
 

llama/llama.cpp/src/llama-grammar.cpp

@@ -907,6 +907,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
 
 struct llama_grammar * llama_grammar_init_impl(
         const struct llama_vocab * vocab,
+        const struct ollama_vocab * ollama_vocab,
         const llama_grammar_element ** rules,
         size_t n_rules,
         size_t start_rule_index) {
@@ -962,6 +963,7 @@ struct llama_grammar * llama_grammar_init_impl(
     // then the pointers would be invalidated when the local vec_rules goes out of scope.
     return new llama_grammar {
         vocab,
+        ollama_vocab,
         std::move(vec_rules),
         std::move(stacks),
         /* .partial_utf8 = */     {},
@@ -969,28 +971,25 @@ struct llama_grammar * llama_grammar_init_impl(
         /* .awaiting_trigger = */ false,
         /* .trigger_buffer = */   "",
         /* .trigger_tokens   = */ {},
-        /* .trigger_words    = */ {},
+        /* .trigger_patterns    = */ {},
     };
 }
 
 struct llama_grammar * llama_grammar_init_impl(
         const struct llama_vocab * vocab,
+        const struct ollama_vocab * ollama_vocab,
                       const char * grammar_str,
                       const char * grammar_root,
                               bool lazy,
-                     const char ** trigger_words,
-                            size_t num_trigger_words,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
                const llama_token * trigger_tokens,
                             size_t num_trigger_tokens) {
     llama_grammar_parser parser;
 
     // if there is a grammar, parse it
-    if (!parser.parse(grammar_str)) {
-        return nullptr;
-    }
-
-    // will be empty (default) if there are parse errors
-    if (parser.rules.empty()) {
+    // rules will be empty (default) if there are parse errors
+    if (!parser.parse(grammar_str) || parser.rules.empty()) {
         fprintf(stderr, "%s: failed to parse grammar\n", __func__);
         return nullptr;
     }
@@ -1054,14 +1053,16 @@ struct llama_grammar * llama_grammar_init_impl(
     } while (true);
 
     std::vector<llama_token>    vec_trigger_tokens;
-    std::vector<std::string> vec_trigger_words;
+    std::vector<llama_grammar_trigger_pattern> vec_trigger_patterns;
     for (size_t i = 0; i < num_trigger_tokens; i++) {
         GGML_ASSERT(trigger_tokens != nullptr);
         vec_trigger_tokens.push_back(trigger_tokens[i]);
     }
-    for (size_t i = 0; i < num_trigger_words; i++) {
-        GGML_ASSERT(trigger_words != nullptr);
-        vec_trigger_words.push_back(trigger_words[i]);
+    for (size_t i = 0; i < num_trigger_patterns; i++) {
+        GGML_ASSERT(trigger_patterns != nullptr);
+        auto & trigger = vec_trigger_patterns.emplace_back();
+        trigger.pattern = trigger_patterns[i];
+        trigger.regex = std::regex(trigger.pattern);
     }
 
     // Important: vec_rules has to be moved here, not copied, because stacks contains
@@ -1069,6 +1070,7 @@ struct llama_grammar * llama_grammar_init_impl(
     // then the pointers would be invalidated when the local vec_rules goes out of scope.
     return new llama_grammar {
         vocab,
+        ollama_vocab,
         std::move(vec_rules),
         std::move(stacks),
         /* .partial_utf8 = */     {},
@@ -1076,7 +1078,7 @@ struct llama_grammar * llama_grammar_init_impl(
         /* .awaiting_trigger = */ lazy,
         /* .trigger_buffer = */   "",
         std::move(vec_trigger_tokens),
-        std::move(vec_trigger_words),
+        std::move(vec_trigger_patterns),
     };
 }
 
@@ -1089,8 +1091,9 @@ void llama_grammar_free_impl(struct llama_grammar * grammar) {
 }
 
 struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & grammar) {
-    llama_grammar * result = new llama_grammar {
+    auto * result = new llama_grammar {
         grammar.vocab,
+        grammar.o_vocab,
         grammar.rules,
         grammar.stacks,
         grammar.partial_utf8,
@@ -1098,7 +1101,7 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
         grammar.awaiting_trigger,
         grammar.trigger_buffer,
         grammar.trigger_tokens,
-        grammar.trigger_words,
+        grammar.trigger_patterns,
     };
 
     // redirect elements in stacks to point to new rules
@@ -1118,7 +1121,6 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
 }
 
 void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_data_array * cur_p) {
-    GGML_ASSERT(grammar.vocab != nullptr);
 
     if (grammar.awaiting_trigger) {
         return;
@@ -1140,9 +1142,13 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
 
     for (size_t i = 0; i < cur_p->size; ++i) {
         const llama_token id      = cur_p->data[i].id;
-        const std::string & piece = grammar.vocab->token_to_piece(id);
+        const std::string piece = grammar.o_vocab ?
+            grammar.o_vocab->token_to_piece(id) :
+            grammar.vocab->token_to_piece(id);
 
-        if (grammar.vocab->is_eog(id)) {
+        const bool is_eog = grammar.o_vocab ? grammar.o_vocab->is_eog(id) : grammar.vocab->is_eog(id);
+
+        if (is_eog) {
             if (!allow_eog) {
                 cur_p->data[i].logit = -INFINITY;
             }
@@ -1161,9 +1167,10 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
 }
 
 void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token) {
-    GGML_ASSERT(grammar.vocab != nullptr);
 
-    const auto & piece = grammar.vocab->token_to_piece(token);
+    const std::string piece = grammar.o_vocab ?
+        grammar.o_vocab->token_to_piece(token) :
+        grammar.vocab->token_to_piece(token);
 
     if (grammar.awaiting_trigger) {
         if (std::find(grammar.trigger_tokens.begin(), grammar.trigger_tokens.end(), token) != grammar.trigger_tokens.end()) {
@@ -1173,16 +1180,18 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
             LLAMA_LOG_DEBUG("Grammar triggered on token %u (`%s`)", token, piece.c_str());
             return;
         } else {
-            // TODO: consider a smarter incremental substring search algorithm (store last position to search from).
             grammar.trigger_buffer += piece;
-            for (const auto & word : grammar.trigger_words) {
-                auto pos = grammar.trigger_buffer.find(word);
-                if (pos != std::string::npos) {
+
+            std::smatch match;
+            for (const auto & trigger_pattern : grammar.trigger_patterns) {
+                if (std::regex_match(grammar.trigger_buffer, match, trigger_pattern.regex)) {
                     grammar.awaiting_trigger = false;
-                    auto constrained_str = grammar.trigger_buffer.substr(pos);
+                    // get from the first match to the end of the string
+                    auto constrained_str = grammar.trigger_buffer.substr(match.position(1));
+                    // std::string constrained_str(match[1].first, grammar.trigger_buffer.end());
                     grammar.trigger_buffer.clear();
                     llama_grammar_accept_str(grammar, constrained_str);
-                    LLAMA_LOG_DEBUG("Grammar triggered on word `%s`", word.c_str());
+                    LLAMA_LOG_DEBUG("Grammar triggered on regex: '%s'\n", constrained_str.c_str());
                     return;
                 }
             }
@@ -1191,13 +1200,14 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
         }
     }
 
-    if (grammar.vocab->is_eog(token)) {
+    const bool is_eog = grammar.o_vocab ? grammar.o_vocab->is_eog(token) : grammar.vocab->is_eog(token);
+    if (is_eog) {
         for (const auto & stack : grammar.stacks) {
             if (stack.empty()) {
                 return;
             }
         }
-        GGML_ABORT("fatal error");
+        GGML_ABORT("grammar error: end of grammar token received but grammar stack is not empty");
     }
 
     llama_grammar_accept_str(grammar, piece);
@@ -1217,3 +1227,28 @@ void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string
         throw std::runtime_error("Unexpected empty grammar stack after accepting piece: " + piece);
     }
 }
+
+
+const std::string & ollama_vocab::token_to_piece(const uint32_t token) const {
+    try {
+        return token_to_piece_map.at(token);
+    } catch (const std::out_of_range&) {
+        throw std::runtime_error("Token not found in vocabulary: " + std::to_string(token));
+    }
+}
+
+void ollama_vocab::add_token_pieces(const uint32_t* tokens, size_t n_tokens, const char** pieces) {
+    for (size_t i = 0; i < n_tokens; i++) {
+        token_to_piece_map[tokens[i]] = pieces[i];
+    }
+}
+
+bool ollama_vocab::is_eog(const uint32_t token) const {
+    return special_eog_ids.count(token) > 0;
+}
+
+void ollama_vocab::set_eog_tokens(const uint32_t* tokens, size_t n_tokens) {
+    for (size_t i = 0; i < n_tokens; i++) {
+        special_eog_ids.insert(tokens[i]);
+    }
+}

llama/llama.cpp/src/llama-grammar.h

@@ -3,10 +3,22 @@
 #include "llama.h"
 
 #include <map>
+#include <regex>
 #include <string>
 #include <vector>
+#include <set>
 
 struct llama_vocab;
+struct ollama_vocab {
+    std::map<uint32_t, std::string> token_to_piece_map;
+    std::set<uint32_t> special_eog_ids;
+
+    const std::string & token_to_piece(const uint32_t token) const;
+    void add_token_pieces(const uint32_t* tokens, size_t n_tokens, const char** pieces);
+    void set_eog_tokens(const uint32_t* tokens, size_t n_tokens);
+    bool is_eog(const uint32_t token) const;
+
+};
 
 // grammar element type
 enum llama_gretype {
@@ -105,9 +117,15 @@ struct llama_grammar_parser {
     void print(FILE * file);
 };
 
+struct llama_grammar_trigger_pattern {
+    std::string pattern;
+    std::regex  regex;
+};
+
 struct llama_grammar {
     // note: allow null vocab for testing (not great)
     const llama_vocab * vocab;
+    const ollama_vocab * o_vocab;
 
     const llama_grammar_rules  rules;  // TODO: shared ptr
           llama_grammar_stacks stacks;
@@ -122,7 +140,10 @@ struct llama_grammar {
     bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only
     std::string              trigger_buffer;           // Output buffered by lazy grammar. Will be cleared once trigger is found.
     std::vector<llama_token> trigger_tokens;           // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
-    std::vector<std::string> trigger_words;
+    std::vector<llama_grammar_trigger_pattern>
+                             trigger_patterns;         // Regular expressions that trigger a lazy grammar. Must be a full match of the entire generated
+                                                       // string, and the grammar will be given the string from the first match group onwards.
+
 };
 
 //
@@ -132,17 +153,19 @@ struct llama_grammar {
 // note: needed for tests (not great)
 struct llama_grammar * llama_grammar_init_impl(
         const struct llama_vocab * vocab,
+        const struct ollama_vocab * ollama_vocab,
         const llama_grammar_element ** rules,
         size_t n_rules,
         size_t start_rule_index);
 
 struct llama_grammar * llama_grammar_init_impl(
         const struct llama_vocab * vocab,
+        const struct ollama_vocab * ollama_vocab,
                       const char * grammar_str,
                       const char * grammar_root,
                               bool lazy,
-                     const char ** trigger_words,
-                            size_t num_trigger_words,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
                const llama_token * trigger_tokens,
                             size_t num_trigger_tokens);
 

llama/llama.cpp/src/llama-graph.h

@@ -0,0 +1,608 @@
+#pragma once
+
+#include "llama-arch.h"
+#include "llama-hparams.h"
+#include "llama-adapter.h"
+
+#include <cstdint>
+#include <vector>
+#include <memory>
+#include <set>
+#include <functional>
+
+struct ggml_cgraph;
+struct ggml_context;
+struct ggml_tensor;
+
+struct llama_ubatch;
+struct llama_cparams;
+
+class llama_memory_i;
+class llama_kv_cache_unified;
+
+// certain models (typically multi-modal) can produce different types of graphs
+enum llm_graph_type {
+    LLM_GRAPH_TYPE_DEFAULT,
+    LLM_GRAPH_TYPE_ENCODER,
+    LLM_GRAPH_TYPE_DECODER,
+};
+
+enum llm_ffn_op_type {
+    LLM_FFN_SILU,
+    LLM_FFN_GELU,
+    LLM_FFN_RELU,
+    LLM_FFN_RELU_SQR,
+    LLM_FFN_SWIGLU,
+};
+
+enum llm_ffn_gate_type {
+    LLM_FFN_SEQ,
+    LLM_FFN_PAR, // ffn_gate is parallel to ffn_up
+};
+
+enum llm_norm_type {
+    LLM_NORM,
+    LLM_NORM_RMS,
+    LLM_NORM_GROUP,
+};
+
+// TODO: tmp - need something better to pass the data from the encoder to the decoder
+struct llama_cross {
+    // the output embeddings from the encoder as a ggml tensor
+    // TODO: this needs more work to be correct, for now copy the embeddings data to host memory
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/11213#discussion_r1969892524
+    //ggml_tensor * t_embd = nullptr;
+
+    int64_t n_embd = 0;
+    int64_t n_enc  = 0;
+
+    // embeddings data copied to host memory (tmp)
+    std::vector<float> v_embd;
+
+    // needed to construct the cross-attention mask in the decoder
+    std::vector<std::set<llama_seq_id>> seq_ids_enc;
+};
+
+//
+// llm_graph_input
+//
+
+class llm_graph_input_i {
+public:
+    virtual ~llm_graph_input_i() = default;
+
+    virtual void set_input(const llama_ubatch * ubatch) = 0;
+};
+
+using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
+
+
+class llm_graph_input_embd : public llm_graph_input_i {
+public:
+    llm_graph_input_embd()          = default;
+    virtual ~llm_graph_input_embd() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * tokens = nullptr; // I32 [n_batch]
+    ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]
+    ggml_tensor * cross_attn_state; // F32 [4, n_embd, 1061]
+};
+
+class llm_graph_input_pos : public llm_graph_input_i {
+public:
+    llm_graph_input_pos(int64_t n_pos_per_token) : n_pos_per_token(n_pos_per_token) {}
+    virtual ~llm_graph_input_pos() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * pos = nullptr; // I32 [n_batch]
+
+    const int64_t n_pos_per_token = 1;
+};
+
+// temperature tuning, used by llama4
+class llm_graph_input_attn_temp : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_temp(int64_t n_pos_per_token, uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
+        : n_pos_per_token(n_pos_per_token), n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
+    virtual ~llm_graph_input_attn_temp() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * attn_scale = nullptr; // F32 [n_batch]
+
+    const int64_t n_pos_per_token = 1;
+
+    const uint32_t n_attn_temp_floor_scale;
+    const float    f_attn_temp_scale;
+};
+
+class llm_graph_input_pos_bucket : public llm_graph_input_i {
+public:
+    llm_graph_input_pos_bucket(const llama_hparams & hparams) : hparams(hparams) {}
+    virtual ~llm_graph_input_pos_bucket() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * pos_bucket = nullptr; // I32 [n_batch, n_batch]
+
+    const llama_hparams & hparams;
+};
+
+class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
+public:
+    llm_graph_input_pos_bucket_kv(
+            const llama_hparams & hparams,
+            const llama_kv_cache_unified * kv_self) : hparams(hparams), kv_self(kv_self) {}
+    virtual ~llm_graph_input_pos_bucket_kv() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch]
+
+    const llama_hparams & hparams;
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_out_ids : public llm_graph_input_i {
+public:
+    llm_graph_input_out_ids(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            int32_t n_outputs) : hparams(hparams), cparams(cparams), n_outputs(n_outputs) {}
+    virtual ~llm_graph_input_out_ids() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * out_ids; // I32 [n_outputs]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+
+    const int32_t n_outputs;
+};
+
+class llm_graph_input_mean : public llm_graph_input_i {
+public:
+    llm_graph_input_mean(const llama_cparams & cparams) : cparams(cparams) {}
+    virtual ~llm_graph_input_mean() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * mean; // F32 [n_batch, n_batch]
+
+    const llama_cparams & cparams;
+};
+
+class llm_graph_input_cls : public llm_graph_input_i {
+public:
+    llm_graph_input_cls(const llama_cparams & cparams) : cparams(cparams) {}
+    virtual ~llm_graph_input_cls() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * cls; // I32 [n_batch]
+
+    const llama_cparams & cparams;
+};
+
+class llm_graph_input_s_copy : public llm_graph_input_i {
+public:
+    llm_graph_input_s_copy(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    virtual ~llm_graph_input_s_copy() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * s_copy; // I32 [kv_size]
+
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_s_mask : public llm_graph_input_i {
+public:
+    llm_graph_input_s_mask(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    virtual ~llm_graph_input_s_mask() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * s_mask; // F32 [1, n_kv]
+
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_cross_embd : public llm_graph_input_i {
+public:
+    llm_graph_input_cross_embd(
+            const llama_cross * cross) : cross(cross) {}
+    virtual ~llm_graph_input_cross_embd() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * cross_embd; // F32 [n_embd, n_outputs_enc]
+
+    const llama_cross * cross;
+};
+
+class llm_graph_input_attn_no_cache : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_no_cache(const llama_hparams & hparams, const llama_cparams & cparams) :
+        hparams(hparams),
+        cparams(cparams) {
+    }
+    ~llm_graph_input_attn_no_cache() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * get_kq_mask() const { return kq_mask_cnv; }
+
+    ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch]
+    ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+};
+
+class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_kv_unified(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_kv_cache_unified * kv_self) :
+        hparams(hparams),
+        cparams(cparams),
+        kv_self(kv_self) {
+    }
+    ~llm_graph_input_attn_kv_unified() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }
+    ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
+
+    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch]
+    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_attn_cross : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_cross(const llama_cross * cross) : cross(cross) {}
+    ~llm_graph_input_attn_cross() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * get_kq_mask_cross() const { return cross_kq_mask_cnv; }
+
+    ggml_tensor * cross_kq_mask     = nullptr; // F32 [n_outputs_enc, n_batch]
+    ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch]
+
+    const llama_cross * cross = nullptr;
+};
+
+class llm_graph_input_cross_attn_state : public llm_graph_input_i {
+public:
+    llm_graph_input_cross_attn_state()          = default;
+    virtual ~llm_graph_input_cross_attn_state() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * cross_attn_state; // F32 [4, n_embd, 1061]
+};
+
+//
+// llm_graph_result
+//
+
+// these objects deliver the result from the graph build process back to the llama_context
+// note that the input tensors created for the graph are referenced here - the goal is to be able to populate their
+//   specific data, by calling the set_inputs() method
+// along with the input tensors, the object also provides commonly used outputs tensors, such as logits, embeddings, etc.
+//   these are used by the llama_context to extact the relevant data, based on the compute parameters
+
+class llm_graph_result_i {
+public:
+    virtual ~llm_graph_result_i() = default;
+
+    virtual ggml_tensor * get_logits()      = 0;
+    virtual ggml_tensor * get_embd()        = 0;
+    virtual ggml_tensor * get_embd_pooled() = 0;
+
+    virtual void set_inputs(const llama_ubatch * ubatch) = 0;
+};
+
+using llm_graph_result_ptr = std::unique_ptr<llm_graph_result_i>;
+
+
+class llm_graph_result : public llm_graph_result_i {
+public:
+    virtual ~llm_graph_result() = default;
+
+    ggml_tensor * get_logits()      override { return t_logits; }
+    ggml_tensor * get_embd()        override { return t_embd; }
+    ggml_tensor * get_embd_pooled() override { return t_embd_pooled; }
+
+    void set_inputs(const llama_ubatch * ubatch) override {
+        for (auto & input : inputs) {
+            input->set_input(ubatch);
+        }
+    }
+
+    llm_graph_input_i * add_input(llm_graph_input_ptr input) {
+        inputs.emplace_back(std::move(input));
+        return inputs.back().get();
+    }
+
+    // important graph nodes
+    ggml_tensor * t_logits      = nullptr;
+    ggml_tensor * t_embd        = nullptr;
+    ggml_tensor * t_embd_pooled = nullptr;
+
+    std::vector<llm_graph_input_ptr> inputs;
+};
+
+//
+// llm_graph_context
+//
+
+// callback that allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
+using llm_graph_cb = std::function<void(const llama_ubatch & ubatch, ggml_tensor * cur, const char * name, int il)>;
+
+struct llm_graph_params {
+    ggml_context * ctx;
+
+    const llm_arch arch;
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+    const llama_ubatch  & ubatch;
+
+    ggml_backend_sched * sched;
+    ggml_backend * backend_cpu;
+
+    const llama_adapter_cvec  * cvec;
+    const llama_adapter_loras * loras;
+    const llama_memory_i      * memory;
+    const llama_cross         * cross;
+
+    int32_t n_outputs;
+
+    const llm_graph_cb & cb;
+};
+
+struct llm_graph_context {
+    const llm_arch arch;
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+    const llama_ubatch  & ubatch;
+
+    const int64_t n_embd;
+    const int64_t n_layer;
+    const int64_t n_rot;
+    const int64_t n_ctx;       // user-specified context size (can be different from n_ctx_train)
+    const int64_t n_ctx_per_seq;
+    const int64_t n_head;
+    const int64_t n_head_kv;
+    const int64_t n_embd_head_k;
+    const int64_t n_embd_k_gqa;
+    const int64_t n_embd_head_v;
+    const int64_t n_embd_v_gqa;
+    const int64_t n_expert;
+    const int64_t n_expert_used;
+
+    const float freq_base;
+    const float freq_scale;
+    const float ext_factor;
+    const float attn_factor;
+    const float beta_fast;
+    const float beta_slow;
+    const float norm_eps;
+    const float norm_rms_eps;
+
+    const int32_t n_tokens;
+    const int32_t n_outputs;
+    const int32_t n_ctx_orig; // yarn
+
+    const enum llama_pooling_type pooling_type;
+    const enum llama_rope_type    rope_type;
+
+    ggml_context * ctx0 = nullptr;
+
+    ggml_backend_sched * sched;
+
+    ggml_backend * backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
+
+    const llama_adapter_cvec  * cvec;
+    const llama_adapter_loras * loras;
+    const llama_memory_i      * memory;
+    const llama_cross         * cross;
+
+    const llm_graph_cb & cb_func;
+
+    std::unique_ptr<llm_graph_result> res;
+
+    llm_graph_context(const llm_graph_params & params);
+
+    int64_t n_pos_per_token() const;
+
+    void cb(ggml_tensor * cur, const char * name, int il) const;
+
+    //
+    // common
+    //
+
+    ggml_tensor * build_cvec(
+             ggml_tensor * cur,
+                     int   il) const;
+
+    // do mat_mul, while optionally apply lora
+    ggml_tensor * build_lora_mm(
+              ggml_tensor * w,
+              ggml_tensor * cur) const;
+
+    // do mat_mul_id, while optionally apply lora
+    ggml_tensor * build_lora_mm_id(
+              ggml_tensor * w,   // ggml_tensor * as
+              ggml_tensor * cur, // ggml_tensor * b
+              ggml_tensor * ids) const;
+
+    ggml_tensor * build_norm(
+             ggml_tensor * cur,
+             ggml_tensor * mw,
+             ggml_tensor * mb,
+           llm_norm_type   type,
+                     int   il) const;
+
+    ggml_tensor * build_ffn(
+             ggml_tensor * cur,
+             ggml_tensor * up,
+             ggml_tensor * up_b,
+             ggml_tensor * up_s,
+             ggml_tensor * gate,
+             ggml_tensor * gate_b,
+             ggml_tensor * gate_s,
+             ggml_tensor * down,
+             ggml_tensor * down_b,
+             ggml_tensor * down_s,
+             ggml_tensor * act_scales,
+         llm_ffn_op_type   type_op,
+       llm_ffn_gate_type   type_gate,
+                     int   il) const;
+
+    ggml_tensor * build_moe_ffn(
+             ggml_tensor * cur,
+             ggml_tensor * gate_inp,
+             ggml_tensor * up_exps,
+             ggml_tensor * gate_exps,
+             ggml_tensor * down_exps,
+             ggml_tensor * exp_probs_b,
+                 int64_t   n_expert,
+                 int64_t   n_expert_used,
+         llm_ffn_op_type   type_op,
+                    bool   norm_w,
+                    bool   scale_w,
+                   float   w_scale,
+            llama_expert_gating_func_type gating_op,
+                     int   il) const;
+
+    //
+    // inputs
+    //
+
+    ggml_tensor * build_inp_embd(ggml_tensor * tok_embd) const;
+    ggml_tensor * build_inp_pos() const;
+    ggml_tensor * build_inp_attn_scale() const;
+    ggml_tensor * build_inp_out_ids() const;
+    ggml_tensor * build_inp_mean() const;
+    ggml_tensor * build_inp_cls() const;
+    ggml_tensor * build_inp_s_copy() const;
+    ggml_tensor * build_inp_s_mask() const;
+    ggml_tensor * build_inp_cross_attn_state() const;
+
+    ggml_tensor * build_inp_cross_embd() const;
+    ggml_tensor * build_inp_pos_bucket_enc() const;
+    ggml_tensor * build_inp_pos_bucket_dec() const;
+    ggml_tensor * build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const;
+
+    //
+    // attention
+    //
+
+    ggml_tensor * build_attn_mha(
+             ggml_cgraph * gf,
+             ggml_tensor * q,     // [n_embd_head_q, n_tokens, n_head_q]
+             ggml_tensor * k,     // [n_embd_head_k, n_tokens, n_head_k]
+             ggml_tensor * v,     // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false)
+             ggml_tensor * kq_b,
+             ggml_tensor * kq_mask,
+             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                    bool   v_trans,
+                   float   kq_scale) const;
+
+    llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_no_cache * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  float   kq_scale,
+                    int   il) const;
+
+    llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_kv_unified * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  float   kq_scale,
+                    int   il) const;
+
+    llm_graph_input_attn_cross * build_attn_inp_cross() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_cross * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  float   kq_scale,
+                    int   il) const;
+
+    //
+    // recurrent
+    //
+
+    ggml_tensor * build_copy_mask_state(
+             ggml_cgraph * gf,
+             ggml_tensor * s,
+             ggml_tensor * state_copy,
+             ggml_tensor * state_mask,
+                 int32_t   n_state,
+                 int32_t   n_seqs) const;
+
+    ggml_tensor * build_rwkv_token_shift_load(
+             ggml_cgraph * gf,
+             ggml_tensor * state_copy,
+             ggml_tensor * state_mask,
+      const llama_ubatch & ubatch,
+                     int   il) const;
+
+    ggml_tensor * build_rwkv_token_shift_store(
+             ggml_tensor * token_shift,
+      const llama_ubatch & ubatch,
+                     int   il) const;
+
+    //
+    // pooling
+    //
+
+    void build_pooling(
+            ggml_cgraph * gf,
+            ggml_tensor * cls,
+            ggml_tensor * cls_b,
+            ggml_tensor * cls_out,
+            ggml_tensor * cls_out_b) const;
+};

llama/llama.cpp/src/llama-hparams.cpp

@@ -2,8 +2,6 @@
 
 #include "ggml.h"
 
-#include <algorithm>
-
 uint32_t llama_hparams::n_head(uint32_t il) const {
     if (il < n_layer) {
         return n_head_arr[il];
@@ -80,6 +78,14 @@ bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {
     GGML_ABORT("fatal error");
 }
 
+bool llama_hparams::is_swa(uint32_t il) const {
+    if (il < n_layer) {
+        return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1);
+    }
+
+    GGML_ABORT("fatal error");
+}
+
 bool llama_hparams::cross_attention_layers(uint32_t il) const {
     return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
-}
+}

llama/llama.cpp/src/llama-hparams.h

@@ -2,6 +2,8 @@
 
 #include "llama.h"
 
+#include <algorithm>
+
 #include <array>
 
 // bump if necessary
@@ -36,6 +38,7 @@ struct llama_hparams {
     uint32_t n_layer;
     uint32_t n_rot;
     uint32_t n_swa = 0; // sliding window attention (SWA)
+    uint32_t n_swa_pattern = 1; // by default, all layers use non-sliding-window attention
     uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
     uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
     uint32_t n_expert = 0;
@@ -43,6 +46,10 @@ struct llama_hparams {
     uint32_t n_rel_attn_bkts = 0;
     uint32_t n_vocab = 0;
 
+    // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
+    uint32_t n_embd_head_k_mla = 0;
+    uint32_t n_embd_head_v_mla = 0;
+
     // for WavTokenizer
     struct llama_hparams_posnet   posnet;
     struct llama_hparams_convnext convnext;
@@ -79,10 +86,16 @@ struct llama_hparams {
     uint32_t time_decay_extra_dim   = 0;
     uint32_t wkv_head_size          = 0;
     uint32_t token_shift_count      = 2;
+    uint32_t n_lora_decay           = 0;
+    uint32_t n_lora_iclr            = 0;
+    uint32_t n_lora_value_res_mix   = 0;
+    uint32_t n_lora_gate            = 0;
 
     float    rope_attn_factor = 1.0f;
     float    rope_freq_base_train;
+    float    rope_freq_base_train_swa;
     float    rope_freq_scale_train;
+    float    rope_freq_scale_train_swa;
     uint32_t n_ctx_orig_yarn;
     float    rope_yarn_log_mul;
 
@@ -109,6 +122,14 @@ struct llama_hparams {
     bool use_alibi     = false;
     bool attn_soft_cap = false;
 
+    uint32_t n_moe_layer_step        = 0;
+    bool     use_kq_norm             = true;
+    uint32_t n_attn_chunk            = 0;
+    // values below seems to be fixed on llama4
+    uint32_t n_no_rope_layer_step    = 4;
+    uint32_t n_attn_temp_floor_scale = 8192;
+    float    f_attn_temp_scale       = 0.1;
+
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
@@ -143,6 +164,8 @@ struct llama_hparams {
 
     // cross attention layers
     bool cross_attention_layers(uint32_t il) const;
+
+    bool is_swa(uint32_t il) const;
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");