update patches

export ThinkingParser

.github/workflows/release.yaml

@@ -432,6 +432,22 @@ jobs:
           docker buildx imagetools inspect ollama/ollama:${{ steps.metadata.outputs.version }}
         working-directory: ${{ runner.temp }}
 
+  # Trigger downstream release process
+  trigger:
+    runs-on: ubuntu-latest
+    environment: release
+    needs: [darwin-build, windows-build, windows-depends]
+    steps:
+      - name: Trigger downstream release process
+        run: |
+          curl -L \
+            -X POST \
+            -H "Accept: application/vnd.github+json" \
+            -H "Authorization: Bearer ${{ secrets.RELEASE_TOKEN }}" \
+            -H "X-GitHub-Api-Version: 2022-11-28" \
+            https://api.github.com/repos/ollama/${{ vars.RELEASE_REPO }}/dispatches \
+            -d "{\"event_type\": \"trigger-workflow\", \"client_payload\": {\"run_id\": \"${GITHUB_RUN_ID}\", \"version\": \"${GITHUB_REF_NAME#v}\"}}"
+
   # Aggregate all the assets and ship a release
   release:
     needs: [darwin-sign, windows-sign, linux-build]

.golangci.yaml

@@ -19,8 +19,8 @@ linters:
     - nolintlint
     - nosprintfhostport
     - staticcheck
-    - tenv
     - unconvert
+    - usetesting
     - wastedassign
     - whitespace
   disable:

CMakeLists.txt

@@ -51,6 +51,8 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/include
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu/amx)
 
+add_compile_definitions(NDEBUG)
+
 set(GGML_CPU ON)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
 set_property(TARGET ggml PROPERTY EXCLUDE_FROM_ALL TRUE)

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggerganov/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=2016f07bd106c73699ecbaace80f55db5ed95dac
+FETCH_HEAD=1caae7fc6c77551cb1066515e0f414713eebb367
 
 .PHONY: help
 help:
@@ -15,11 +15,13 @@ help:
 	@echo "    make -f $(lastword $(MAKEFILE_LIST)) clean sync"
 
 .PHONY: sync
-sync: llama/build-info.cpp llama/llama.cpp ml/backend/ggml/ggml
+sync: llama/build-info.cpp ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal
 
-.PHONY: llama/build-info.cpp
-llama/build-info.cpp: llama/build-info.cpp.in
-	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' $< > $@
+llama/build-info.cpp: llama/build-info.cpp.in llama/llama.cpp
+	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' <$< >$@
+
+ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal: ml/backend/ggml/ggml
+	go generate ./$(@D)
 
 .PHONY: llama/llama.cpp
 llama/llama.cpp: llama/vendor/
@@ -30,12 +32,13 @@ ml/backend/ggml/ggml: llama/vendor/ggml/
 	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
 
 PATCHES=$(wildcard llama/patches/*.patch)
+PATCHED=$(join $(dir $(PATCHES)), $(addsuffix ed, $(addprefix ., $(notdir $(PATCHES)))))
 
 .PHONY: apply-patches
 .NOTPARALLEL:
-apply-patches: $(addsuffix ed, $(PATCHES))
+apply-patches: $(PATCHED)
 
-%.patched: %.patch
+llama/patches/.%.patched: llama/patches/%.patch
 	@if git -c user.name=nobody -c 'user.email=<>' -C $(WORKDIR) am -3 $(realpath $<); then touch $@; else git -C $(WORKDIR) am --abort; exit 1; fi
 
 .PHONY: checkout
@@ -57,4 +60,4 @@ format-patches: llama/patches
 
 .PHONE: clean
 clean: checkout
-	$(RM) $(addsuffix ed, $(PATCHES))
+	$(RM) llama/patches/.*.patched

README.md

@@ -61,6 +61,8 @@ Here are some example models that can be downloaded:
 | QwQ                | 32B        | 20GB  | `ollama run qwq`                 |
 | DeepSeek-R1        | 7B         | 4.7GB | `ollama run deepseek-r1`         |
 | DeepSeek-R1        | 671B       | 404GB | `ollama run deepseek-r1:671b`    |
+| Llama 4            | 109B       | 67GB  | `ollama run llama4:scout`        |
+| Llama 4            | 400B       | 245GB | `ollama run llama4:maverick`     |
 | Llama 3.3          | 70B        | 43GB  | `ollama run llama3.3`            |
 | Llama 3.2          | 3B         | 2.0GB | `ollama run llama3.2`            |
 | Llama 3.2          | 1B         | 1.3GB | `ollama run llama3.2:1b`         |
@@ -77,7 +79,7 @@ Here are some example models that can be downloaded:
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`           |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored`   |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`               |
-| Granite-3.2         | 8B         | 4.9GB | `ollama run granite3.2`          |
+| Granite-3.3         | 8B         | 4.9GB | `ollama run granite3.3`          |
 
 > [!NOTE]
 > You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
@@ -312,6 +314,8 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Ollama Basic Chat: Uses HyperDiv Reactive UI](https://github.com/rapidarchitect/ollama_basic_chat)
 - [Ollama-chats RPG](https://github.com/drazdra/ollama-chats)
 - [IntelliBar](https://intellibar.app/) (AI-powered assistant for macOS)
+- [Jirapt](https://github.com/AliAhmedNada/jirapt) (Jira Integration to generate issues, tasks, epics)
+- [ojira](https://github.com/AliAhmedNada/ojira) (Jira chrome plugin to easily generate descriptions for tasks)
 - [QA-Pilot](https://github.com/reid41/QA-Pilot) (Interactive chat tool that can leverage Ollama models for rapid understanding and navigation of GitHub code repositories)
 - [ChatOllama](https://github.com/sugarforever/chat-ollama) (Open Source Chatbot based on Ollama with Knowledge Bases)
 - [CRAG Ollama Chat](https://github.com/Nagi-ovo/CRAG-Ollama-Chat) (Simple Web Search with Corrective RAG)
@@ -394,12 +398,15 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [1Panel](https://github.com/1Panel-dev/1Panel/) (Web-based Linux Server Management Tool)
 - [AstrBot](https://github.com/Soulter/AstrBot/) (User-friendly LLM-based multi-platform chatbot with a WebUI, supporting RAG, LLM agents, and plugins integration)
 - [Reins](https://github.com/ibrahimcetin/reins) (Easily tweak parameters, customize system prompts per chat, and enhance your AI experiments with reasoning model support.)
+- [Flufy](https://github.com/Aharon-Bensadoun/Flufy) (A beautiful chat interface for interacting with Ollama's API. Built with React, TypeScript, and Material-UI.)
 - [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)
 - [Lumina](https://github.com/cushydigit/lumina.git) (A lightweight, minimal React.js frontend for interacting with Ollama servers)
+- [Tiny Notepad](https://pypi.org/project/tiny-notepad) (A lightweight, notepad-like interface to chat with ollama available on PyPI)
+- [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.) 
 
 ### Cloud
 
@@ -443,6 +450,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
+- [AWS-Strands-With-Ollama](https://github.com/rapidarchitect/ollama_strands) - AWS Strands Agents with Ollama Examples
 
 ### Apple Vision Pro
 
@@ -469,7 +477,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 ### Libraries
 
-- [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/integrations/chat/ollama/) with [example](https://js.langchain.com/docs/tutorials/local_rag/)
+- [LangChain](https://python.langchain.com/docs/integrations/chat/ollama/) and [LangChain.js](https://js.langchain.com/docs/integrations/chat/ollama/) with [example](https://js.langchain.com/docs/tutorials/local_rag/)
 - [Firebase Genkit](https://firebase.google.com/docs/genkit/plugins/ollama)
 - [crewAI](https://github.com/crewAIInc/crewAI)
 - [Yacana](https://remembersoftwares.github.io/yacana/) (User-friendly multi-agent framework for brainstorming and executing predetermined flows with built-in tool integration)
@@ -522,6 +530,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Abso](https://github.com/lunary-ai/abso) (OpenAI-compatible TypeScript SDK for any LLM provider)
 - [Nichey](https://github.com/goodreasonai/nichey) is a Python package for generating custom wikis for your research topic
 - [Ollama for D](https://github.com/kassane/ollama-d)
+- [OllamaPlusPlus](https://github.com/HardCodeDev777/OllamaPlusPlus) (Very simple C++ library for Ollama)
 
 ### Mobile
 
@@ -578,6 +587,8 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Simple-Discord-AI](https://github.com/zyphixor/simple-discord-ai)
 - [LLM Telegram Bot](https://github.com/innightwolfsleep/llm_telegram_bot) (telegram bot, primary for RP. Oobabooga-like buttons, [A1111](https://github.com/AUTOMATIC1111/stable-diffusion-webui) API integration e.t.c)
 - [mcp-llm](https://github.com/sammcj/mcp-llm) (MCP Server to allow LLMs to call other LLMs)
+- [SimpleOllamaUnity](https://github.com/HardCodeDev777/SimpleOllamaUnity) (Unity Engine extension for communicating with Ollama in a few lines of code. Also works at runtime)
+- [UnityCodeLama](https://github.com/HardCodeDev777/UnityCodeLama) (Unity Edtior tool to analyze scripts via Ollama)
 
 ### Supported backends
 

api/client.go

@@ -24,7 +24,10 @@ import (
 	"net/http"
 	"net/url"
 	"runtime"
+	"strconv"
+	"time"
 
+	"github.com/ollama/ollama/auth"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/version"
@@ -76,6 +79,14 @@ func NewClient(base *url.URL, http *http.Client) *Client {
 	}
 }
 
+func getAuthorizationToken(ctx context.Context, challenge string) (string, error) {
+	token, err := auth.Sign(ctx, []byte(challenge))
+	if err != nil {
+		return "", err
+	}
+	return token, nil
+}
+
 func (c *Client) do(ctx context.Context, method, path string, reqData, respData any) error {
 	var reqBody io.Reader
 	var data []byte
@@ -97,6 +108,21 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	}
 
 	requestURL := c.base.JoinPath(path)
+
+	var token string
+	if envconfig.UseAuth() || c.base.Hostname() == "ollama.com" {
+		now := strconv.FormatInt(time.Now().Unix(), 10)
+		chal := fmt.Sprintf("%s,%s?ts=%s", method, path, now)
+		token, err = getAuthorizationToken(ctx, chal)
+		if err != nil {
+			return err
+		}
+
+		q := requestURL.Query()
+		q.Set("ts", now)
+		requestURL.RawQuery = q.Encode()
+	}
+
 	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), reqBody)
 	if err != nil {
 		return err
@@ -106,6 +132,10 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	request.Header.Set("Accept", "application/json")
 	request.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))
 
+	if token != "" {
+		request.Header.Set("Authorization", token)
+	}
+
 	respObj, err := c.http.Do(request)
 	if err != nil {
 		return err
@@ -143,6 +173,22 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 	}
 
 	requestURL := c.base.JoinPath(path)
+
+	var token string
+	if envconfig.UseAuth() || c.base.Hostname() == "ollama.com" {
+		var err error
+		now := strconv.FormatInt(time.Now().Unix(), 10)
+		chal := fmt.Sprintf("%s,%s?ts=%s", method, path, now)
+		token, err = getAuthorizationToken(ctx, chal)
+		if err != nil {
+			return err
+		}
+
+		q := requestURL.Query()
+		q.Set("ts", now)
+		requestURL.RawQuery = q.Encode()
+	}
+
 	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), buf)
 	if err != nil {
 		return err
@@ -152,6 +198,10 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 	request.Header.Set("Accept", "application/x-ndjson")
 	request.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))
 
+	if token != "" {
+		request.Header.Set("Authorization", token)
+	}
+
 	response, err := c.http.Do(request)
 	if err != nil {
 		return err

api/client_test.go

@@ -1,7 +1,6 @@
 package api
 
 import (
-	"context"
 	"encoding/json"
 	"fmt"
 	"net/http"
@@ -137,7 +136,7 @@ func TestClientStream(t *testing.T) {
 			client := NewClient(&url.URL{Scheme: "http", Host: ts.Listener.Addr().String()}, http.DefaultClient)
 
 			var receivedChunks []ChatResponse
-			err := client.stream(context.Background(), http.MethodPost, "/v1/chat", nil, func(chunk []byte) error {
+			err := client.stream(t.Context(), http.MethodPost, "/v1/chat", nil, func(chunk []byte) error {
 				var resp ChatResponse
 				if err := json.Unmarshal(chunk, &resp); err != nil {
 					return fmt.Errorf("failed to unmarshal chunk: %w", err)
@@ -223,7 +222,7 @@ func TestClientDo(t *testing.T) {
 				ID      string `json:"id"`
 				Success bool   `json:"success"`
 			}
-			err := client.do(context.Background(), http.MethodPost, "/v1/messages", nil, &resp)
+			err := client.do(t.Context(), http.MethodPost, "/v1/messages", nil, &resp)
 
 			if tc.wantErr != "" {
 				if err == nil {

api/types.go

@@ -83,6 +83,12 @@ type GenerateRequest struct {
 	// Options lists model-specific options. For example, temperature can be
 	// set through this field, if the model supports it.
 	Options map[string]any `json:"options"`
+
+	// Think controls whether thinking/reasoning models will think before
+	// responding. Needs to be a pointer so we can distinguish between false
+	// (request that thinking _not_ be used) and unset (use the old behavior
+	// before this option was introduced)
+	Think *bool `json:"think,omitempty"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].
@@ -108,6 +114,10 @@ type ChatRequest struct {
 
 	// Options lists model-specific options.
 	Options map[string]any `json:"options"`
+
+	// Think controls whether thinking/reasoning models will think before
+	// responding
+	Think *bool `json:"think,omitempty"`
 }
 
 type Tools []Tool
@@ -126,8 +136,11 @@ func (t Tool) String() string {
 // role ("system", "user", or "assistant"), the content and an optional list
 // of images.
 type Message struct {
-	Role      string      `json:"role"`
-	Content   string      `json:"content"`
+	Role    string `json:"role"`
+	Content string `json:"content"`
+	// Thinking contains the text that was inside thinking tags in the
+	// original model output when ChatRequest.Think is enabled.
+	Thinking  string      `json:"thinking,omitempty"`
 	Images    []ImageData `json:"images,omitempty"`
 	ToolCalls []ToolCall  `json:"tool_calls,omitempty"`
 }
@@ -271,9 +284,6 @@ type Options struct {
 	RepeatPenalty    float32  `json:"repeat_penalty,omitempty"`
 	PresencePenalty  float32  `json:"presence_penalty,omitempty"`
 	FrequencyPenalty float32  `json:"frequency_penalty,omitempty"`
-	Mirostat         int      `json:"mirostat,omitempty"`
-	MirostatTau      float32  `json:"mirostat_tau,omitempty"`
-	MirostatEta      float32  `json:"mirostat_eta,omitempty"`
 	Stop             []string `json:"stop,omitempty"`
 }
 
@@ -283,12 +293,7 @@ type Runner struct {
 	NumBatch  int   `json:"num_batch,omitempty"`
 	NumGPU    int   `json:"num_gpu,omitempty"`
 	MainGPU   int   `json:"main_gpu,omitempty"`
-	LowVRAM   bool  `json:"low_vram,omitempty"`
-	F16KV     bool  `json:"f16_kv,omitempty"` // Deprecated: This option is ignored
-	LogitsAll bool  `json:"logits_all,omitempty"`
-	VocabOnly bool  `json:"vocab_only,omitempty"`
 	UseMMap   *bool `json:"use_mmap,omitempty"`
-	UseMLock  bool  `json:"use_mlock,omitempty"`
 	NumThread int   `json:"num_thread,omitempty"`
 }
 
@@ -452,12 +457,13 @@ type ProcessResponse struct {
 
 // ListModelResponse is a single model description in [ListResponse].
 type ListModelResponse struct {
-	Name       string       `json:"name"`
-	Model      string       `json:"model"`
-	ModifiedAt time.Time    `json:"modified_at"`
-	Size       int64        `json:"size"`
-	Digest     string       `json:"digest"`
-	Details    ModelDetails `json:"details,omitempty"`
+	Name         string             `json:"name"`
+	Model        string             `json:"model"`
+	ModifiedAt   time.Time          `json:"modified_at"`
+	Size         int64              `json:"size"`
+	Digest       string             `json:"digest"`
+	Capabilities []model.Capability `json:"capabilities,omitempty"`
+	Details      ModelDetails       `json:"details,omitempty"`
 }
 
 // ProcessModelResponse is a single model description in [ProcessResponse].
@@ -471,13 +477,6 @@ type ProcessModelResponse struct {
 	SizeVRAM  int64        `json:"size_vram"`
 }
 
-type RetrieveModelResponse struct {
-	Id      string `json:"id"`
-	Object  string `json:"object"`
-	Created int64  `json:"created"`
-	OwnedBy string `json:"owned_by"`
-}
-
 type TokenResponse struct {
 	Token string `json:"token"`
 }
@@ -493,6 +492,10 @@ type GenerateResponse struct {
 	// Response is the textual response itself.
 	Response string `json:"response"`
 
+	// Thinking contains the text that was inside thinking tags in the
+	// original model output when ChatRequest.Think is enabled.
+	Thinking string `json:"thinking,omitempty"`
+
 	// Done specifies if the response is complete.
 	Done bool `json:"done"`
 
@@ -660,9 +663,6 @@ func DefaultOptions() Options {
 		RepeatPenalty:    1.1,
 		PresencePenalty:  0.0,
 		FrequencyPenalty: 0.0,
-		Mirostat:         0,
-		MirostatTau:      5.0,
-		MirostatEta:      0.1,
 		Seed:             -1,
 
 		Runner: Runner{
@@ -671,8 +671,6 @@ func DefaultOptions() Options {
 			NumBatch:  512,
 			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
 			NumThread: 0,  // let the runtime decide
-			LowVRAM:   false,
-			UseMLock:  false,
 			UseMMap:   nil,
 		},
 	}

api/types_test.go

@@ -372,3 +372,50 @@ func TestPropertyType_MarshalJSON(t *testing.T) {
 		})
 	}
 }
+
+func TestThinking_UnmarshalJSON(t *testing.T) {
+	trueVal := true
+	falseVal := false
+
+	tests := []struct {
+		name             string
+		input            string
+		expectedThinking *bool
+		expectedError    bool
+	}{
+		{
+			name:             "true",
+			input:            `{ "think": true }`,
+			expectedThinking: &trueVal,
+		},
+		{
+			name:             "false",
+			input:            `{ "think": false }`,
+			expectedThinking: &falseVal,
+		},
+		{
+			name:             "unset",
+			input:            `{ }`,
+			expectedThinking: nil,
+		},
+		{
+			name:             "invalid",
+			input:            `{ "think": "true" }`,
+			expectedThinking: nil,
+			expectedError:    true,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var req GenerateRequest
+			err := json.Unmarshal([]byte(test.input), &req)
+			if test.expectedError {
+				require.Error(t, err)
+			} else {
+				require.NoError(t, err)
+				assert.Equal(t, test.expectedThinking, req.Think)
+			}
+		})
+	}
+}

app/lifecycle/logging.go

@@ -4,20 +4,14 @@ import (
 	"fmt"
 	"log/slog"
 	"os"
-	"path/filepath"
 	"strconv"
 	"strings"
 
 	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/logutil"
 )
 
 func InitLogging() {
-	level := slog.LevelInfo
-
-	if envconfig.Debug() {
-		level = slog.LevelDebug
-	}
-
 	var logFile *os.File
 	var err error
 	// Detect if we're a GUI app on windows, and if not, send logs to console
@@ -33,20 +27,8 @@ func InitLogging() {
 			return
 		}
 	}
-	handler := slog.NewTextHandler(logFile, &slog.HandlerOptions{
-		Level:     level,
-		AddSource: true,
-		ReplaceAttr: func(_ []string, attr slog.Attr) slog.Attr {
-			if attr.Key == slog.SourceKey {
-				source := attr.Value.Any().(*slog.Source)
-				source.File = filepath.Base(source.File)
-			}
-			return attr
-		},
-	})
-
-	slog.SetDefault(slog.New(handler))
 
+	slog.SetDefault(logutil.NewLogger(logFile, envconfig.LogLevel()))
 	slog.Info("ollama app started")
 }
 

benchmark/server_benchmark_test.go

@@ -78,7 +78,7 @@ func BenchmarkColdStart(b *testing.B) {
 
 	for _, tt := range tests {
 		b.Run(fmt.Sprintf("%s/cold/%s", m, tt.name), func(b *testing.B) {
-			ctx := context.Background()
+			ctx := b.Context()
 
 			// Set number of tokens as our throughput metric
 			b.SetBytes(int64(tt.maxTokens))
@@ -113,7 +113,7 @@ func BenchmarkWarmStart(b *testing.B) {
 
 	for _, tt := range tests {
 		b.Run(fmt.Sprintf("%s/warm/%s", m, tt.name), func(b *testing.B) {
-			ctx := context.Background()
+			ctx := b.Context()
 
 			// Pre-warm the model
 			warmup(client, m, tt.prompt, b)
@@ -140,7 +140,7 @@ func setup(b *testing.B) *api.Client {
 	if err != nil {
 		b.Fatal(err)
 	}
-	if _, err := client.Show(context.Background(), &api.ShowRequest{Model: modelName(b)}); err != nil {
+	if _, err := client.Show(b.Context(), &api.ShowRequest{Model: modelName(b)}); err != nil {
 		b.Fatalf("Model unavailable: %v", err)
 	}
 

cmd/cmd.go

@@ -31,6 +31,7 @@ import (
 	"github.com/olekukonko/tablewriter"
 	"github.com/spf13/cobra"
 	"golang.org/x/crypto/ssh"
+	"golang.org/x/sync/errgroup"
 	"golang.org/x/term"
 
 	"github.com/ollama/ollama/api"
@@ -38,12 +39,31 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
+	"github.com/ollama/ollama/readline"
 	"github.com/ollama/ollama/runner"
 	"github.com/ollama/ollama/server"
 	"github.com/ollama/ollama/types/model"
+	"github.com/ollama/ollama/types/syncmap"
 	"github.com/ollama/ollama/version"
 )
 
+// ensureThinkingSupport emits a warning if the model does not advertise thinking support
+func ensureThinkingSupport(ctx context.Context, client *api.Client, name string) {
+	if name == "" {
+		return
+	}
+	resp, err := client.Show(ctx, &api.ShowRequest{Model: name})
+	if err != nil {
+		return
+	}
+	for _, cap := range resp.Capabilities {
+		if cap == model.CapabilityThinking {
+			return
+		}
+	}
+	fmt.Fprintf(os.Stderr, "warning: model %q does not support thinking output\n", name)
+}
+
 var errModelfileNotFound = errors.New("specified Modelfile wasn't found")
 
 func getModelfileName(cmd *cobra.Command) (string, error) {
@@ -106,7 +126,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	}
 	spinner.Stop()
 
-	req.Name = args[0]
+	req.Model = args[0]
 	quantize, _ := cmd.Flags().GetString("quantize")
 	if quantize != "" {
 		req.Quantize = quantize
@@ -117,34 +137,54 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	if len(req.Files) > 0 {
-		fileMap := map[string]string{}
-		for f, digest := range req.Files {
+	var g errgroup.Group
+	g.SetLimit(max(runtime.GOMAXPROCS(0)-1, 1))
+
+	files := syncmap.NewSyncMap[string, string]()
+	for f, digest := range req.Files {
+		g.Go(func() error {
 			if _, err := createBlob(cmd, client, f, digest, p); err != nil {
 				return err
 			}
-			fileMap[filepath.Base(f)] = digest
-		}
-		req.Files = fileMap
+
+			// TODO: this is incorrect since the file might be in a subdirectory
+			//       instead this should take the path relative to the model directory
+			//       but the current implementation does not allow this
+			files.Store(filepath.Base(f), digest)
+			return nil
+		})
 	}
 
-	if len(req.Adapters) > 0 {
-		fileMap := map[string]string{}
-		for f, digest := range req.Adapters {
+	adapters := syncmap.NewSyncMap[string, string]()
+	for f, digest := range req.Adapters {
+		g.Go(func() error {
 			if _, err := createBlob(cmd, client, f, digest, p); err != nil {
 				return err
 			}
-			fileMap[filepath.Base(f)] = digest
-		}
-		req.Adapters = fileMap
+
+			// TODO: same here
+			adapters.Store(filepath.Base(f), digest)
+			return nil
+		})
 	}
 
+	if err := g.Wait(); err != nil {
+		return err
+	}
+
+	req.Files = files.Items()
+	req.Adapters = adapters.Items()
+
 	bars := make(map[string]*progress.Bar)
 	fn := func(resp api.ProgressResponse) error {
 		if resp.Digest != "" {
 			bar, ok := bars[resp.Digest]
 			if !ok {
-				bar = progress.NewBar(fmt.Sprintf("pulling %s...", resp.Digest[7:19]), resp.Total, resp.Completed)
+				msg := resp.Status
+				if msg == "" {
+					msg = fmt.Sprintf("pulling %s...", resp.Digest[7:19])
+				}
+				bar = progress.NewBar(msg, resp.Total, resp.Completed)
 				bars[resp.Digest] = bar
 				p.Add(resp.Digest, bar)
 			}
@@ -213,7 +253,7 @@ func createBlob(cmd *cobra.Command, client *api.Client, path string, digest stri
 		}
 	}()
 
-	if err = client.CreateBlob(cmd.Context(), digest, io.TeeReader(bin, &pw)); err != nil {
+	if err := client.CreateBlob(cmd.Context(), digest, io.TeeReader(bin, &pw)); err != nil {
 		return "", err
 	}
 	return digest, nil
@@ -243,6 +283,9 @@ func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
 	req := &api.GenerateRequest{
 		Model:     opts.Model,
 		KeepAlive: opts.KeepAlive,
+
+		// pass Think here so we fail before getting to the chat prompt if the model doesn't support it
+		Think: opts.Think,
 	}
 
 	return client.Generate(cmd.Context(), req, func(api.GenerateResponse) error { return nil })
@@ -277,6 +320,22 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 	opts.Format = format
 
+	thinkFlag := cmd.Flags().Lookup("think")
+	if thinkFlag.Changed {
+		think, err := cmd.Flags().GetBool("think")
+		if err != nil {
+			return err
+		}
+		opts.Think = &think
+	} else {
+		opts.Think = nil
+	}
+	hidethinking, err := cmd.Flags().GetBool("hidethinking")
+	if err != nil {
+		return err
+	}
+	opts.HideThinking = hidethinking
+
 	keepAlive, err := cmd.Flags().GetString("keepalive")
 	if err != nil {
 		return err
@@ -340,6 +399,11 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
+	opts.Think, err = inferThinkingOption(&info.Capabilities, &opts, thinkFlag.Changed)
+	if err != nil {
+		return err
+	}
+
 	opts.MultiModal = slices.Contains(info.Capabilities, model.CapabilityVision)
 
 	// TODO: remove the projector info and vision info checks below,
@@ -725,11 +789,38 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 				case float64:
 					v = fmt.Sprintf("%g", vData)
 				case []any:
-					n := 3
-					if len(vData) < n {
-						n = len(vData)
+					targetWidth := 10 // Small width where we are displaying the data in a column
+
+					var itemsToShow int
+					totalWidth := 1 // Start with 1 for opening bracket
+
+					// Find how many we can fit
+					for i := range vData {
+						itemStr := fmt.Sprintf("%v", vData[i])
+						width := runewidth.StringWidth(itemStr)
+
+						// Add separator width (", ") for all items except the first
+						if i > 0 {
+							width += 2
+						}
+
+						// Check if adding this item would exceed our width limit
+						if totalWidth+width > targetWidth && i > 0 {
+							break
+						}
+
+						totalWidth += width
+						itemsToShow++
+					}
+
+					// Format the output
+					if itemsToShow < len(vData) {
+						v = fmt.Sprintf("%v", vData[:itemsToShow])
+						v = strings.TrimSuffix(v, "]")
+						v += fmt.Sprintf(" ...+%d more]", len(vData)-itemsToShow)
+					} else {
+						v = fmt.Sprintf("%v", vData)
 					}
-					v = fmt.Sprintf("%v", vData[:n])
 				default:
 					v = fmt.Sprintf("%T", vData)
 				}
@@ -750,10 +841,19 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 
 	head := func(s string, n int) (rows [][]string) {
 		scanner := bufio.NewScanner(strings.NewReader(s))
-		for scanner.Scan() && (len(rows) < n || n < 0) {
-			if text := scanner.Text(); text != "" {
-				rows = append(rows, []string{"", strings.TrimSpace(text)})
+		count := 0
+		for scanner.Scan() {
+			text := strings.TrimSpace(scanner.Text())
+			if text == "" {
+				continue
 			}
+			count++
+			if n < 0 || count <= n {
+				rows = append(rows, []string{"", text})
+			}
+		}
+		if n >= 0 && count > n {
+			rows = append(rows, []string{"", "..."})
 		}
 		return
 	}
@@ -865,17 +965,19 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 type generateContextKey string
 
 type runOptions struct {
-	Model       string
-	ParentModel string
-	Prompt      string
-	Messages    []api.Message
-	WordWrap    bool
-	Format      string
-	System      string
-	Images      []api.ImageData
-	Options     map[string]any
-	MultiModal  bool
-	KeepAlive   *api.Duration
+	Model        string
+	ParentModel  string
+	Prompt       string
+	Messages     []api.Message
+	WordWrap     bool
+	Format       string
+	System       string
+	Images       []api.ImageData
+	Options      map[string]any
+	MultiModal   bool
+	KeepAlive    *api.Duration
+	Think        *bool
+	HideThinking bool
 }
 
 type displayResponseState struct {
@@ -931,6 +1033,26 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 	}
 }
 
+func thinkingOutputOpeningText(plainText bool) string {
+	text := "Thinking...\n"
+
+	if plainText {
+		return text
+	}
+
+	return readline.ColorGrey + readline.ColorBold + text + readline.ColorDefault + readline.ColorGrey
+}
+
+func thinkingOutputClosingText(plainText bool) string {
+	text := "...done thinking.\n\n"
+
+	if plainText {
+		return text
+	}
+
+	return readline.ColorGrey + readline.ColorBold + text + readline.ColorDefault
+}
+
 func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -958,14 +1080,34 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	var latest api.ChatResponse
 	var fullResponse strings.Builder
 	var role string
+	var thinkTagOpened bool = false
+	var thinkTagClosed bool = false
 
 	fn := func(response api.ChatResponse) error {
-		p.StopAndClear()
+		if response.Message.Content != "" || !opts.HideThinking {
+			p.StopAndClear()
+		}
 
 		latest = response
 
 		role = response.Message.Role
+		if response.Message.Thinking != "" && !opts.HideThinking {
+			if !thinkTagOpened {
+				fmt.Print(thinkingOutputOpeningText(false))
+				thinkTagOpened = true
+			}
+			displayResponse(response.Message.Thinking, opts.WordWrap, state)
+		}
+
 		content := response.Message.Content
+		if thinkTagOpened && !thinkTagClosed && content != "" {
+			fmt.Print(thinkingOutputClosingText(false))
+			thinkTagClosed = true
+		}
+		// purposefully not putting thinking blocks in the response, which would
+		// only be needed if we later added tool calling to the cli (they get
+		// filtered out anyway since current models don't expect them unless you're
+		// about to finish some tool calls)
 		fullResponse.WriteString(content)
 
 		displayResponse(content, opts.WordWrap, state)
@@ -982,6 +1124,7 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		Messages: opts.Messages,
 		Format:   json.RawMessage(opts.Format),
 		Options:  opts.Options,
+		Think:    opts.Think,
 	}
 
 	if opts.KeepAlive != nil {
@@ -1043,13 +1186,32 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 	}()
 
 	var state *displayResponseState = &displayResponseState{}
+	var thinkTagOpened bool = false
+	var thinkTagClosed bool = false
+
+	plainText := !term.IsTerminal(int(os.Stdout.Fd()))
 
 	fn := func(response api.GenerateResponse) error {
-		p.StopAndClear()
-
 		latest = response
 		content := response.Response
 
+		if response.Response != "" || !opts.HideThinking {
+			p.StopAndClear()
+		}
+
+		if response.Thinking != "" && !opts.HideThinking {
+			if !thinkTagOpened {
+				fmt.Print(thinkingOutputOpeningText(plainText))
+				thinkTagOpened = true
+			}
+			displayResponse(response.Thinking, opts.WordWrap, state)
+		}
+
+		if thinkTagOpened && !thinkTagClosed && content != "" {
+			fmt.Print(thinkingOutputClosingText(plainText))
+			thinkTagClosed = true
+		}
+
 		displayResponse(content, opts.WordWrap, state)
 
 		return nil
@@ -1075,6 +1237,7 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		System:    opts.System,
 		Options:   opts.Options,
 		KeepAlive: opts.KeepAlive,
+		Think:     opts.Think,
 	}
 
 	if err := client.Generate(ctx, &request, fn); err != nil {
@@ -1178,11 +1341,11 @@ func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
 		return err
 	}
 	if err := client.Heartbeat(cmd.Context()); err != nil {
-		if !strings.Contains(err.Error(), " refused") {
+		if !(strings.Contains(err.Error(), " refused") || strings.Contains(err.Error(), "could not connect")) {
 			return err
 		}
 		if err := startApp(cmd.Context(), client); err != nil {
-			return errors.New("could not connect to ollama app, is it running?")
+			return fmt.Errorf("ollama server not responding - %w", err)
 		}
 	}
 	return nil
@@ -1260,7 +1423,7 @@ func NewCLI() *cobra.Command {
 	}
 
 	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\"")
-	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_0)")
+	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_K_M)")
 
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
@@ -1290,6 +1453,8 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("insecure", false, "Use an insecure registry")
 	runCmd.Flags().Bool("nowordwrap", false, "Don't wrap words to the next line automatically")
 	runCmd.Flags().String("format", "", "Response format (e.g. json)")
+	runCmd.Flags().Bool("think", false, "Whether to use thinking mode for supported models")
+	runCmd.Flags().Bool("hidethinking", false, "Hide thinking output (if provided)")
 
 	stopCmd := &cobra.Command{
 		Use:     "stop MODEL",
@@ -1341,7 +1506,6 @@ func NewCLI() *cobra.Command {
 		PreRunE: checkServerHeartbeat,
 		RunE:    ListRunningHandler,
 	}
-
 	copyCmd := &cobra.Command{
 		Use:     "cp SOURCE DESTINATION",
 		Short:   "Copy a model",
@@ -1430,3 +1594,45 @@ func NewCLI() *cobra.Command {
 
 	return rootCmd
 }
+
+// If the user has explicitly set thinking options, either through the CLI or
+// through the `/set think` or `set nothink` interactive options, then we
+// respect them. Otherwise, we check model capabilities to see if the model
+// supports thinking. If the model does support thinking, we enable it.
+// Otherwise, we unset the thinking option (which is different than setting it
+// to false).
+//
+// If capabilities are not provided, we fetch them from the server.
+func inferThinkingOption(caps *[]model.Capability, runOpts *runOptions, explicitlySetByUser bool) (*bool, error) {
+	if explicitlySetByUser {
+		return runOpts.Think, nil
+	}
+
+	if caps == nil {
+		client, err := api.ClientFromEnvironment()
+		if err != nil {
+			return nil, err
+		}
+		ret, err := client.Show(context.Background(), &api.ShowRequest{
+			Model: runOpts.Model,
+		})
+		if err != nil {
+			return nil, err
+		}
+		caps = &ret.Capabilities
+	}
+
+	thinkingSupported := false
+	for _, cap := range *caps {
+		if cap == model.CapabilityThinking {
+			thinkingSupported = true
+		}
+	}
+
+	if thinkingSupported {
+		thinking := true
+		return &thinking, nil
+	}
+
+	return nil, nil
+}

cmd/cmd_test.go

@@ -2,7 +2,6 @@ package cmd
 
 import (
 	"bytes"
-	"context"
 	"encoding/json"
 	"io"
 	"net/http"
@@ -226,6 +225,7 @@ Weigh anchor!
   System
     You are a pirate!    
     Ahoy, matey!         
+    ...                  
 
 `
 		if diff := cmp.Diff(expect, b.String()); diff != "" {
@@ -337,7 +337,7 @@ func TestDeleteHandler(t *testing.T) {
 	t.Cleanup(mockServer.Close)
 
 	cmd := &cobra.Command{}
-	cmd.SetContext(context.TODO())
+	cmd.SetContext(t.Context())
 	if err := DeleteHandler(cmd, []string{"test-model"}); err != nil {
 		t.Fatalf("DeleteHandler failed: %v", err)
 	}
@@ -399,11 +399,6 @@ func TestGetModelfileName(t *testing.T) {
 			var expectedFilename string
 
 			if tt.fileExists {
-				tempDir, err := os.MkdirTemp("", "modelfiledir")
-				defer os.RemoveAll(tempDir)
-				if err != nil {
-					t.Fatalf("temp modelfile dir creation failed: %v", err)
-				}
 				var fn string
 				if tt.modelfileName != "" {
 					fn = tt.modelfileName
@@ -411,7 +406,7 @@ func TestGetModelfileName(t *testing.T) {
 					fn = "Modelfile"
 				}
 
-				tempFile, err := os.CreateTemp(tempDir, fn)
+				tempFile, err := os.CreateTemp(t.TempDir(), fn)
 				if err != nil {
 					t.Fatalf("temp modelfile creation failed: %v", err)
 				}
@@ -530,7 +525,7 @@ func TestPushHandler(t *testing.T) {
 
 			cmd := &cobra.Command{}
 			cmd.Flags().Bool("insecure", false, "")
-			cmd.SetContext(context.TODO())
+			cmd.SetContext(t.Context())
 
 			// Redirect stderr to capture progress output
 			oldStderr := os.Stderr
@@ -635,7 +630,7 @@ func TestListHandler(t *testing.T) {
 			t.Setenv("OLLAMA_HOST", mockServer.URL)
 
 			cmd := &cobra.Command{}
-			cmd.SetContext(context.TODO())
+			cmd.SetContext(t.Context())
 
 			// Capture stdout
 			oldStdout := os.Stdout
@@ -690,7 +685,7 @@ func TestCreateHandler(t *testing.T) {
 						return
 					}
 
-					if req.Name != "test-model" {
+					if req.Model != "test-model" {
 						t.Errorf("expected model name 'test-model', got %s", req.Name)
 					}
 
@@ -730,7 +725,7 @@ func TestCreateHandler(t *testing.T) {
 			}))
 			t.Setenv("OLLAMA_HOST", mockServer.URL)
 			t.Cleanup(mockServer.Close)
-			tempFile, err := os.CreateTemp("", "modelfile")
+			tempFile, err := os.CreateTemp(t.TempDir(), "modelfile")
 			if err != nil {
 				t.Fatal(err)
 			}
@@ -750,7 +745,7 @@ func TestCreateHandler(t *testing.T) {
 			}
 
 			cmd.Flags().Bool("insecure", false, "")
-			cmd.SetContext(context.TODO())
+			cmd.SetContext(t.Context())
 
 			// Redirect stderr to capture progress output
 			oldStderr := os.Stderr

cmd/interactive.go

@@ -44,7 +44,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "Use \"\"\" to begin a multi-line message.")
 
 		if opts.MultiModal {
-			fmt.Fprintf(os.Stderr, "Use %s to include .jpg or .png images.\n", filepath.FromSlash("/path/to/file"))
+			fmt.Fprintf(os.Stderr, "Use %s to include .jpg, .png, or .webp images.\n", filepath.FromSlash("/path/to/file"))
 		}
 
 		fmt.Fprintln(os.Stderr, "")
@@ -62,6 +62,8 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "  /set noformat          Disable formatting")
 		fmt.Fprintln(os.Stderr, "  /set verbose           Show LLM stats")
 		fmt.Fprintln(os.Stderr, "  /set quiet             Disable LLM stats")
+		fmt.Fprintln(os.Stderr, "  /set think             Enable thinking")
+		fmt.Fprintln(os.Stderr, "  /set nothink           Disable thinking")
 		fmt.Fprintln(os.Stderr, "")
 	}
 
@@ -128,6 +130,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 	var sb strings.Builder
 	var multiline MultilineState
+	var thinkExplicitlySet bool = opts.Think != nil
 
 	for {
 		line, err := scanner.Readline()
@@ -195,11 +198,19 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			opts.Model = args[1]
 			opts.Messages = []api.Message{}
 			fmt.Printf("Loading model '%s'\n", opts.Model)
+			opts.Think, err = inferThinkingOption(nil, &opts, thinkExplicitlySet)
+			if err != nil {
+				return err
+			}
 			if err := loadOrUnloadModel(cmd, &opts); err != nil {
 				if strings.Contains(err.Error(), "not found") {
 					fmt.Printf("error: %v\n", err)
 					continue
 				}
+				if strings.Contains(err.Error(), "does not support thinking") {
+					fmt.Printf("error: %v\n", err)
+					continue
+				}
 				return err
 			}
 			continue
@@ -260,6 +271,22 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 						return err
 					}
 					fmt.Println("Set 'quiet' mode.")
+				case "think":
+					think := true
+					opts.Think = &think
+					thinkExplicitlySet = true
+					if client, err := api.ClientFromEnvironment(); err == nil {
+						ensureThinkingSupport(cmd.Context(), client, opts.Model)
+					}
+					fmt.Println("Set 'think' mode.")
+				case "nothink":
+					think := false
+					opts.Think = &think
+					thinkExplicitlySet = true
+					if client, err := api.ClientFromEnvironment(); err == nil {
+						ensureThinkingSupport(cmd.Context(), client, opts.Model)
+					}
+					fmt.Println("Set 'nothink' mode.")
 				case "format":
 					if len(args) < 3 || args[2] != "json" {
 						fmt.Println("Invalid or missing format. For 'json' mode use '/set format json'")
@@ -448,6 +475,11 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 			assistant, err := chat(cmd, opts)
 			if err != nil {
+				if strings.Contains(err.Error(), "does not support thinking") {
+					fmt.Printf("error: %v\n", err)
+					sb.Reset()
+					continue
+				}
 				return err
 			}
 			if assistant != nil {
@@ -511,7 +543,7 @@ func extractFileNames(input string) []string {
 	// Regex to match file paths starting with optional drive letter, / ./ \ or .\ and include escaped or unescaped spaces (\ or %20)
 	// and followed by more characters and a file extension
 	// This will capture non filename strings, but we'll check for file existence to remove mismatches
-	regexPattern := `(?:[a-zA-Z]:)?(?:\./|/|\\)[\S\\ ]+?\.(?i:jpg|jpeg|png)\b`
+	regexPattern := `(?:[a-zA-Z]:)?(?:\./|/|\\)[\S\\ ]+?\.(?i:jpg|jpeg|png|webp)\b`
 	re := regexp.MustCompile(regexPattern)
 
 	return re.FindAllString(input, -1)
@@ -531,6 +563,8 @@ func extractFileData(input string) (string, []api.ImageData, error) {
 			return "", imgs, err
 		}
 		fmt.Fprintf(os.Stderr, "Added image '%s'\n", nfp)
+		input = strings.ReplaceAll(input, "'"+nfp+"'", "")
+		input = strings.ReplaceAll(input, "'"+fp+"'", "")
 		input = strings.ReplaceAll(input, fp, "")
 		imgs = append(imgs, data)
 	}
@@ -551,7 +585,7 @@ func getImageData(filePath string) ([]byte, error) {
 	}
 
 	contentType := http.DetectContentType(buf)
-	allowedTypes := []string{"image/jpeg", "image/jpg", "image/png"}
+	allowedTypes := []string{"image/jpeg", "image/jpg", "image/png", "image/webp"}
 	if !slices.Contains(allowedTypes, contentType) {
 		return nil, fmt.Errorf("invalid image type: %s", contentType)
 	}

cmd/interactive_test.go

@@ -1,6 +1,8 @@
 package cmd
 
 import (
+	"os"
+	"path/filepath"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
@@ -10,14 +12,17 @@ func TestExtractFilenames(t *testing.T) {
 	// Unix style paths
 	input := ` some preamble 
  ./relative\ path/one.png inbetween1 ./not a valid two.jpg inbetween2 ./1.svg
-/unescaped space /three.jpeg inbetween3 /valid\ path/dir/four.png "./quoted with spaces/five.JPG`
+/unescaped space /three.jpeg inbetween3 /valid\ path/dir/four.png "./quoted with spaces/five.JPG
+/unescaped space /six.webp inbetween6 /valid\ path/dir/seven.WEBP`
 	res := extractFileNames(input)
-	assert.Len(t, res, 5)
+	assert.Len(t, res, 7)
 	assert.Contains(t, res[0], "one.png")
 	assert.Contains(t, res[1], "two.jpg")
 	assert.Contains(t, res[2], "three.jpeg")
 	assert.Contains(t, res[3], "four.png")
 	assert.Contains(t, res[4], "five.JPG")
+	assert.Contains(t, res[5], "six.webp")
+	assert.Contains(t, res[6], "seven.WEBP")
 	assert.NotContains(t, res[4], '"')
 	assert.NotContains(t, res, "inbetween1")
 	assert.NotContains(t, res, "./1.svg")
@@ -28,10 +33,12 @@ func TestExtractFilenames(t *testing.T) {
  /absolute/nospace/three.jpeg inbetween3 /absolute/with space/four.png inbetween4
 ./relative\ path/five.JPG inbetween5 "./relative with/spaces/six.png inbetween6
 d:\path with\spaces\seven.JPEG inbetween7 c:\users\jdoe\eight.png inbetween8 
- d:\program files\someplace\nine.png inbetween9 "E:\program files\someplace\ten.PNG some ending
+ d:\program files\someplace\nine.png inbetween9 "E:\program files\someplace\ten.PNG
+c:/users/jdoe/eleven.webp inbetween11 c:/program files/someplace/twelve.WebP inbetween12
+d:\path with\spaces\thirteen.WEBP some ending
 `
 	res = extractFileNames(input)
-	assert.Len(t, res, 10)
+	assert.Len(t, res, 13)
 	assert.NotContains(t, res, "inbetween2")
 	assert.Contains(t, res[0], "one.png")
 	assert.Contains(t, res[0], "c:")
@@ -49,4 +56,31 @@ d:\path with\spaces\seven.JPEG inbetween7 c:\users\jdoe\eight.png inbetween8
 	assert.Contains(t, res[8], "d:")
 	assert.Contains(t, res[9], "ten.PNG")
 	assert.Contains(t, res[9], "E:")
+	assert.Contains(t, res[10], "eleven.webp")
+	assert.Contains(t, res[10], "c:")
+	assert.Contains(t, res[11], "twelve.WebP")
+	assert.Contains(t, res[11], "c:")
+	assert.Contains(t, res[12], "thirteen.WEBP")
+	assert.Contains(t, res[12], "d:")
+}
+
+// Ensure that file paths wrapped in single quotes are removed with the quotes.
+func TestExtractFileDataRemovesQuotedFilepath(t *testing.T) {
+	dir := t.TempDir()
+	fp := filepath.Join(dir, "img.jpg")
+	data := make([]byte, 600)
+	copy(data, []byte{
+		0xff, 0xd8, 0xff, 0xe0, 0x00, 0x10, 'J', 'F', 'I', 'F',
+		0x00, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+		0xff, 0xd9,
+	})
+	if err := os.WriteFile(fp, data, 0o600); err != nil {
+		t.Fatalf("failed to write test image: %v", err)
+	}
+
+	input := "before '" + fp + "' after"
+	cleaned, imgs, err := extractFileData(input)
+	assert.NoError(t, err)
+	assert.Len(t, imgs, 1)
+	assert.Equal(t, cleaned, "before  after")
 }

cmd/start_windows.go

@@ -4,17 +4,27 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"log/slog"
 	"os"
 	"os/exec"
+	"path"
 	"path/filepath"
 	"strings"
 	"syscall"
+	"unsafe"
 
 	"github.com/ollama/ollama/api"
+	"golang.org/x/sys/windows"
+)
+
+const (
+	Installer = "OllamaSetup.exe"
 )
 
 func startApp(ctx context.Context, client *api.Client) error {
-	// log.Printf("XXX Attempting to find and start ollama app")
+	if len(isProcRunning(Installer)) > 0 {
+		return fmt.Errorf("upgrade in progress...")
+	}
 	AppName := "ollama app.exe"
 	exe, err := os.Executable()
 	if err != nil {
@@ -56,3 +66,41 @@ func startApp(ctx context.Context, client *api.Client) error {
 	}
 	return waitForServer(ctx, client)
 }
+
+func isProcRunning(procName string) []uint32 {
+	pids := make([]uint32, 2048)
+	var ret uint32
+	if err := windows.EnumProcesses(pids, &ret); err != nil || ret == 0 {
+		slog.Debug("failed to check for running installers", "error", err)
+		return nil
+	}
+	pids = pids[:ret]
+	var matches []uint32
+	for _, pid := range pids {
+		if pid == 0 {
+			continue
+		}
+		hProcess, err := windows.OpenProcess(windows.PROCESS_QUERY_INFORMATION|windows.PROCESS_VM_READ, false, pid)
+		if err != nil {
+			continue
+		}
+		defer windows.CloseHandle(hProcess)
+		var module windows.Handle
+		var cbNeeded uint32
+		cb := (uint32)(unsafe.Sizeof(module))
+		if err := windows.EnumProcessModules(hProcess, &module, cb, &cbNeeded); err != nil {
+			continue
+		}
+		var sz uint32 = 1024 * 8
+		moduleName := make([]uint16, sz)
+		cb = uint32(len(moduleName)) * (uint32)(unsafe.Sizeof(uint16(0)))
+		if err := windows.GetModuleBaseName(hProcess, module, &moduleName[0], cb); err != nil && err != syscall.ERROR_INSUFFICIENT_BUFFER {
+			continue
+		}
+		exeFile := path.Base(strings.ToLower(syscall.UTF16ToString(moduleName)))
+		if strings.EqualFold(exeFile, procName) {
+			matches = append(matches, pid)
+		}
+	}
+	return matches
+}

cmd/warn_thinking_test.go

@@ -0,0 +1,63 @@
+package cmd
+
+import (
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"strings"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/types/model"
+)
+
+// Test that a warning is printed when thinking is requested but not supported.
+func TestWarnMissingThinking(t *testing.T) {
+	cases := []struct {
+		capabilities []model.Capability
+		expectWarn   bool
+	}{
+		{capabilities: []model.Capability{model.CapabilityThinking}, expectWarn: false},
+		{capabilities: []model.Capability{}, expectWarn: true},
+	}
+
+	for _, tc := range cases {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			if r.URL.Path != "/api/show" || r.Method != http.MethodPost {
+				t.Fatalf("unexpected request to %s %s", r.URL.Path, r.Method)
+			}
+			var req api.ShowRequest
+			if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+				t.Fatalf("decode request: %v", err)
+			}
+			resp := api.ShowResponse{Capabilities: tc.capabilities}
+			if err := json.NewEncoder(w).Encode(resp); err != nil {
+				t.Fatalf("encode response: %v", err)
+			}
+		}))
+		defer srv.Close()
+
+		t.Setenv("OLLAMA_HOST", srv.URL)
+		client, err := api.ClientFromEnvironment()
+		if err != nil {
+			t.Fatal(err)
+		}
+		oldStderr := os.Stderr
+		r, w, _ := os.Pipe()
+		os.Stderr = w
+		ensureThinkingSupport(t.Context(), client, "m")
+		w.Close()
+		os.Stderr = oldStderr
+		out, _ := io.ReadAll(r)
+
+		warned := strings.Contains(string(out), "warning:")
+		if tc.expectWarn && !warned {
+			t.Errorf("expected warning, got none")
+		}
+		if !tc.expectWarn && warned {
+			t.Errorf("did not expect warning, got: %s", string(out))
+		}
+	}
+}

convert/convert.go

@@ -1,12 +1,13 @@
 package convert
 
 import (
+	"cmp"
 	"encoding/json"
 	"errors"
 	"fmt"
-	"io"
 	"io/fs"
 	"log/slog"
+	"os"
 	"slices"
 	"strings"
 
@@ -14,13 +15,12 @@ import (
 )
 
 type ModelParameters struct {
-	Architectures []string       `json:"architectures"`
-	VocabSize     uint32         `json:"vocab_size"`
-	TextModel     TextParameters `json:"text_config"`
-}
+	Architectures []string `json:"architectures"`
+	VocabSize     uint32   `json:"vocab_size"`
 
-type TextParameters struct {
-	VocabSize uint32 `json:"vocab_size"`
+	TextModel struct {
+		VocabSize uint32 `json:"vocab_size"`
+	} `json:"text_config"`
 }
 
 type AdapterParameters struct {
@@ -53,8 +53,11 @@ func (ModelParameters) KV(t *Tokenizer) ggml.KV {
 	}
 
 	for _, sv := range t.SpecialVocabulary {
-		kv[fmt.Sprintf("tokenizer.ggml.%s_token_id", sv.Key())] = uint32(sv.ID)
 		kv[fmt.Sprintf("tokenizer.ggml.add_%s_token", sv.Key())] = sv.AddToken
+		kv[fmt.Sprintf("tokenizer.ggml.%s_token_id", sv.Key())] = uint32(sv.ID)
+		if len(sv.IDs) > 0 {
+			kv[fmt.Sprintf("tokenizer.ggml.%s_token_ids", sv.Key())] = sv.IDs
+		}
 	}
 
 	return kv
@@ -89,7 +92,7 @@ type ModelConverter interface {
 	// KV maps parameters to LLM key-values
 	KV(*Tokenizer) ggml.KV
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
-	Tensors([]Tensor) []ggml.Tensor
+	Tensors([]Tensor) []*ggml.Tensor
 	// 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
@@ -106,13 +109,13 @@ type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
 	KV(ggml.KV) ggml.KV
 	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
-	Tensors([]Tensor) []ggml.Tensor
+	Tensors([]Tensor) []*ggml.Tensor
 	// 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
 }
 
-func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV ggml.KV) error {
+func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err
@@ -147,14 +150,14 @@ func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV ggml.KV) error {
 		return err
 	}
 
-	return writeFile(ws, conv.KV(baseKV), conv.Tensors(ts))
+	return writeFile(f, conv.KV(baseKV), conv.Tensors(ts))
 }
 
 // Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
 // and files it finds in the input path.
 // Supported input model formats include safetensors.
 // Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
-func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
+func ConvertModel(fsys fs.FS, f *os.File) error {
 	bts, err := fs.ReadFile(fsys, "config.json")
 	if err != nil {
 		return err
@@ -173,6 +176,8 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 	switch p.Architectures[0] {
 	case "LlamaForCausalLM":
 		conv = &llamaModel{}
+	case "MllamaForConditionalGeneration":
+		conv = &mllamaModel{}
 	case "Llama4ForConditionalGeneration":
 		conv = &llama4Model{}
 	case "Mistral3ForConditionalGeneration":
@@ -189,6 +194,8 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		conv = &phi3Model{}
 	case "Qwen2ForCausalLM":
 		conv = &qwen2Model{}
+	case "Qwen2_5_VLForConditionalGeneration":
+		conv = &qwen25VLModel{}
 	case "BertModel":
 		conv = &bertModel{}
 	case "CohereForCausalLM":
@@ -212,24 +219,22 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		return err
 	}
 
-	vocabSize := int(p.VocabSize)
-	if vocabSize == 0 {
-		tVocabSize := int(p.TextModel.VocabSize)
-		vocabSize = tVocabSize
-	}
+	vocabSize := int(cmp.Or(p.VocabSize, p.TextModel.VocabSize))
 
 	switch {
 	case vocabSize == 0:
-		slog.Warn("vocabulary size was not explicitly set by the model", "default size", len(t.Vocabulary.Tokens))
+		slog.Debug("vocabulary size was not explicitly set by the model", "default size", len(t.Vocabulary.Tokens))
 	case vocabSize > len(t.Vocabulary.Tokens):
-		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
+		slog.Debug("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
 		for i := range vocabSize - len(t.Vocabulary.Tokens) {
 			t.Vocabulary.Tokens = append(t.Vocabulary.Tokens, fmt.Sprintf("[PAD%d]", i))
 			t.Vocabulary.Scores = append(t.Vocabulary.Scores, -1)
 			t.Vocabulary.Types = append(t.Vocabulary.Types, tokenTypeUserDefined)
 		}
 	case vocabSize < len(t.Vocabulary.Tokens):
-		return fmt.Errorf("vocabulary is larger than expected '%d' instead of '%d'", len(t.Vocabulary.Tokens), vocabSize)
+		slog.Debug("vocabulary is larger than expected", "want", vocabSize, "got", len(t.Vocabulary.Tokens))
+		p.VocabSize = uint32(len(t.Vocabulary.Tokens))
+		p.TextModel.VocabSize = uint32(len(t.Vocabulary.Tokens))
 	default:
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}
@@ -239,13 +244,13 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		return err
 	}
 
-	return writeFile(ws, conv.KV(t), conv.Tensors(ts))
+	return writeFile(f, conv.KV(t), conv.Tensors(ts))
 }
 
-func writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
+func writeFile(f *os.File, 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)
+	return ggml.WriteGGUF(f, kv, ts)
 }

convert/convert_bert.go

@@ -132,8 +132,8 @@ func (p *bertModel) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *bertModel) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *bertModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
 	for _, t := range ts {
 		if slices.Contains([]string{
 			"embeddings.position_ids",
@@ -143,7 +143,7 @@ func (p *bertModel) Tensors(ts []Tensor) []ggml.Tensor {
 			continue
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_commandr.go

@@ -43,10 +43,10 @@ func (p *commandrModel) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *commandrModel) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *commandrModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
 	for _, t := range ts {
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_gemma.go

@@ -42,14 +42,14 @@ func (p *gemmaModel) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *gemmaModel) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *gemmaModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
 	for _, t := range ts {
 		if !strings.HasPrefix(t.Name(), "v.") && strings.HasSuffix(t.Name(), "_norm.weight") {
 			t.SetRepacker(p.addOne)
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_gemma2_adapter.go

@@ -21,8 +21,8 @@ func (p *gemma2Adapter) KV(baseKV ggml.KV) ggml.KV {
 	return kv
 }
 
-func (p *gemma2Adapter) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *gemma2Adapter) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -31,7 +31,7 @@ func (p *gemma2Adapter) Tensors(ts []Tensor) []ggml.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_llama.go

@@ -126,11 +126,11 @@ func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *llamaModel) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *llamaModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
 
 	if p.RopeScaling.factors != nil {
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     "rope_freqs.weight",
 			Kind:     0,
 			Shape:    []uint64{uint64(len(p.RopeScaling.factors))},
@@ -139,13 +139,14 @@ 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") {
+		if strings.HasSuffix(t.Name(), "attn_q.weight") || strings.HasSuffix(t.Name(), "attn_k.weight") ||
+			strings.HasSuffix(t.Name(), "attn_q_proj.weight") || strings.HasSuffix(t.Name(), "attn_k_proj.weight") {
 			if !p.skipRepack {
 				t.SetRepacker(p.repack)
 			}
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),
@@ -181,9 +182,9 @@ func (p *llamaModel) repack(name string, data []float32, shape []uint64) ([]floa
 	}
 
 	var heads uint32
-	if strings.HasSuffix(name, "attn_q.weight") {
+	if strings.HasSuffix(name, "attn_q.weight") || strings.HasSuffix(name, "attn_q_proj.weight") {
 		heads = p.NumAttentionHeads
-	} else if strings.HasSuffix(name, "attn_k.weight") {
+	} else if strings.HasSuffix(name, "attn_k.weight") || strings.HasSuffix(name, "attn_k_proj.weight") {
 		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
 	} else {
 		return nil, fmt.Errorf("unknown tensor for repack: %s", name)

convert/convert_llama4.go

@@ -88,13 +88,13 @@ func (p *llama4Model) Replacements() []string {
 }
 
 // Tensors implements ModelConverter.
-func (p *llama4Model) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+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{
+			out = append(out, &ggml.Tensor{
 				Name:     t.Name(),
 				Kind:     t.Kind(),
 				Shape:    t.Shape(),
@@ -112,7 +112,7 @@ func (p *llama4Model) Tensors(ts []Tensor) []ggml.Tensor {
 				// 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{
+				out = append(out, &ggml.Tensor{
 					Name:     strings.ReplaceAll(tt.Name(), "ffn_gate_up_exps", name),
 					Kind:     tt.Kind(),
 					Shape:    newShape,
@@ -125,7 +125,7 @@ func (p *llama4Model) Tensors(ts []Tensor) []ggml.Tensor {
 			t.SetRepacker(p.repack())
 			newShape := slices.Clone(t.Shape())
 			newShape[1], newShape[2] = newShape[2], newShape[1]
-			out = append(out, ggml.Tensor{
+			out = append(out, &ggml.Tensor{
 				Name:     t.Name(),
 				Kind:     t.Kind(),
 				Shape:    newShape,

convert/convert_llama_adapter.go

@@ -29,8 +29,8 @@ func (p *llamaAdapter) KV(baseKV ggml.KV) ggml.KV {
 	return kv
 }
 
-func (p *llamaAdapter) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *llamaAdapter) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -41,7 +41,7 @@ func (p *llamaAdapter) Tensors(ts []Tensor) []ggml.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    shape,

convert/convert_mistral.go

@@ -89,8 +89,8 @@ func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *mistral3Model) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *mistral3Model) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
 
 	for _, t := range ts {
 		if !strings.HasPrefix(t.Name(), "v.") {
@@ -100,7 +100,7 @@ func (p *mistral3Model) Tensors(ts []Tensor) []ggml.Tensor {
 			}
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_mixtral.go

@@ -29,7 +29,7 @@ func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *mixtralModel) Tensors(ts []Tensor) []ggml.Tensor {
+func (p *mixtralModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	oldnew := []string{
 		"model.layers", "blk",
 		"w1", "ffn_gate_exps",
@@ -56,10 +56,10 @@ func (p *mixtralModel) Tensors(ts []Tensor) []ggml.Tensor {
 		return true
 	})
 
-	var out []ggml.Tensor
+	var out []*ggml.Tensor
 	for n, e := range experts {
 		// TODO(mxyng): sanity check experts
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     n,
 			Kind:     e[0].Kind(),
 			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),

convert/convert_mllama.go

@@ -0,0 +1,179 @@
+package convert
+
+import (
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+)
+
+type mllamaModel struct {
+	ModelParameters
+	TextModel struct {
+		llamaModel
+
+		CrossAttentionLayers []int32 `json:"cross_attention_layers"`
+	} `json:"text_config"`
+	VisionModel struct {
+		NumHiddenLayers           uint32  `json:"num_hidden_layers"`
+		NumGlobalLayers           uint32  `json:"num_global_layers"`
+		IntermediateLayersIndices []int32 `json:"intermediate_layers_indices"`
+
+		HiddenSize       uint32 `json:"hidden_size"`
+		IntermediateSize uint32 `json:"intermediate_size"`
+
+		AttentionHeads uint32 `json:"attention_heads"`
+
+		ImageSize   uint32  `json:"image_size"`
+		PatchSize   uint32  `json:"patch_size"`
+		NumChannels uint32  `json:"num_channels"`
+		MaxNumTiles uint32  `json:"max_num_tiles"`
+		NormEpsilon float32 `json:"norm_eps"`
+		RopeTheta   float32 `json:"rope.freq_base"`
+	} `json:"vision_config"`
+}
+
+func (m *mllamaModel) KV(t *Tokenizer) ggml.KV {
+	kv := m.ModelParameters.KV(t)
+	kv["general.architecture"] = "mllama"
+
+	for k, v := range m.TextModel.KV(t) {
+		if strings.HasPrefix(k, "llama.") {
+			kv[strings.ReplaceAll(k, "llama.", "mllama.")] = v
+		}
+	}
+
+	kv["mllama.attention.cross_attention_layers"] = m.TextModel.CrossAttentionLayers
+
+	kv["mllama.vision.block_count"] = m.VisionModel.NumHiddenLayers
+	kv["mllama.vision.global.block_count"] = m.VisionModel.NumGlobalLayers
+	kv["mllama.vision.intermediate_layers_indices"] = m.VisionModel.IntermediateLayersIndices
+
+	kv["mllama.vision.embedding_length"] = m.VisionModel.HiddenSize
+	kv["mllama.vision.feed_forward_length"] = m.VisionModel.IntermediateSize
+
+	kv["mllama.vision.attention.head_count"] = m.VisionModel.AttentionHeads
+	kv["mllama.vision.attention.layer_norm_epsilon"] = m.VisionModel.NormEpsilon
+
+	kv["mllama.vision.image_size"] = m.VisionModel.ImageSize
+	kv["mllama.vision.patch_size"] = m.VisionModel.PatchSize
+	kv["mllama.vision.max_num_tiles"] = m.VisionModel.MaxNumTiles
+	kv["mllama.vision.num_channels"] = m.VisionModel.NumChannels
+
+	return kv
+}
+
+func (m *mllamaModel) Replacements() []string {
+	return append(
+		m.TextModel.Replacements(),
+		"language_model.", "",
+		"gate_attn", "attn_gate",
+		"gate_ffn", "ffn_gate",
+		"cross_attn.", "cross_attn_",
+		"vision_model", "v",
+		"class_embedding", "class_embd",
+		"patch_embedding", "patch_embd",
+		"gated_positional_embedding.tile_embedding", "tile_position_embd",
+		"gated_positional_embedding.embedding", "position_embd.weight",
+		"gated_positional_embedding", "position_embd",
+		"embedding.weight", "weight",
+		"pre_tile_positional_embedding", "pre_tile_position_embd",
+		"post_tile_positional_embedding", "post_tile_position_embd",
+		"layernorm_pre", "pre_ln",
+		"layernorm_post", "post_ln",
+		"global_transformer.layers", "global.blk",
+		"transformer.layers", "blk",
+		"mlp.fc1", "ffn_up",
+		"mlp.fc2", "ffn_down",
+		"multi_modal_projector", "mm.0",
+	)
+}
+
+func (m *mllamaModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+	var text []Tensor
+	for _, t := range ts {
+		if !strings.HasPrefix(t.Name(), "v.") && !strings.HasPrefix(t.Name(), "mm.") {
+			text = append(text, t)
+		} else if t.Name() == "v.position_embd.gate" {
+			for _, name := range []string{"v.position_embd.gate", "v.tile_position_embd.gate"} {
+				tt := t.Clone()
+				tt.SetRepacker(m.repack(name))
+				out = append(out, &ggml.Tensor{
+					Name:     name,
+					Kind:     t.Kind(),
+					Shape:    t.Shape(),
+					WriterTo: tt,
+				})
+			}
+		} else {
+			if t.Name() == "v.pre_tile_position_embd.gate" || t.Name() == "v.post_tile_position_embd.gate" {
+				t.SetRepacker(m.repack(t.Name()))
+			} else if strings.HasSuffix(t.Name(), "attn_q.weight") || strings.HasSuffix(t.Name(), "attn_k.weight") {
+				t.SetRepacker(m.repack(t.Name()))
+			} else if strings.HasSuffix(t.Name(), "attn_gate") || strings.HasSuffix(t.Name(), "ffn_gate") {
+				t.SetRepacker(m.repack(t.Name()))
+			}
+
+			out = append(out, &ggml.Tensor{
+				Name:     t.Name(),
+				Kind:     t.Kind(),
+				Shape:    t.Shape(),
+				WriterTo: t,
+			})
+		}
+	}
+
+	return append(out, m.TextModel.Tensors(text)...)
+}
+
+func (m *mllamaModel) repack(name string) Repacker {
+	return func(_ string, data []float32, shape []uint64) (_ []float32, err 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))
+
+		if strings.HasSuffix(name, "attn_q.weight") || strings.HasSuffix(name, "attn_k.weight") {
+			heads := m.VisionModel.AttentionHeads
+			if err := t.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
+				return nil, err
+			}
+
+			if err := t.T(0, 2, 1, 3); err != nil {
+				return nil, err
+			}
+
+			if err := t.Reshape(dims...); err != nil {
+				return nil, err
+			}
+
+			if err := t.Transpose(); err != nil {
+				return nil, err
+			}
+		} else {
+			t, err = tensor.Tanh(t)
+			if err != nil {
+				return nil, err
+			}
+
+			if name == "v.position_embd.gate" {
+				t, err = tensor.Sub(float32(1), t)
+				if 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_phi3.go

@@ -68,19 +68,19 @@ func (p *phi3Model) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *phi3Model) Tensors(ts []Tensor) []ggml.Tensor {
+func (p *phi3Model) Tensors(ts []Tensor) []*ggml.Tensor {
 	var addRopeFactors sync.Once
 
-	out := make([]ggml.Tensor, 0, len(ts)+2)
+	out := make([]*ggml.Tensor, 0, len(ts)+2)
 	for _, t := range ts {
 		if strings.HasPrefix(t.Name(), "blk.0.") {
 			addRopeFactors.Do(func() {
-				out = append(out, ggml.Tensor{
+				out = append(out, &ggml.Tensor{
 					Name:     "rope_factors_long.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.LongFactor))},
 					WriterTo: p.RopeScaling.LongFactor,
-				}, ggml.Tensor{
+				}, &ggml.Tensor{
 					Name:     "rope_factors_short.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.ShortFactor))},
@@ -89,7 +89,7 @@ func (p *phi3Model) Tensors(ts []Tensor) []ggml.Tensor {
 			})
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_qwen2.go

@@ -15,6 +15,7 @@ type qwen2Model struct {
 		Type                          string     `json:"type"`
 		Factor                        ropeFactor `json:"factor"`
 		OriginalMaxPositionEmbeddings uint32     `json:"original_max_position_embeddings"`
+		MropeSection                  []int32    `json:"mrope_section"`
 	} `json:"rope_scaling"`
 	RMSNormEPS float32 `json:"rms_norm_eps"`
 }
@@ -39,16 +40,18 @@ func (q *qwen2Model) KV(t *Tokenizer) ggml.KV {
 	case "yarn":
 		kv["qwen2.rope.scaling.type"] = q.RopeScaling.Type
 		kv["qwen2.rope.scaling.factor"] = q.RopeScaling.Factor
+	case "mrope", "default":
+		kv["qwen2.rope.mrope_section"] = q.RopeScaling.MropeSection
 	default:
 		panic("unknown rope scaling type")
 	}
 	return kv
 }
 
-func (q *qwen2Model) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (q *qwen2Model) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
 	for _, t := range ts {
-		out = append(out, ggml.Tensor{
+		out = append(out, &ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_qwen25vl.go

@@ -0,0 +1,102 @@
+package convert
+
+import (
+	"cmp"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type qwen25VLModel struct {
+	qwen2Model
+
+	VisionModel struct {
+		Depth               uint32  `json:"depth"`
+		HiddenSize          uint32  `json:"hidden_size"`
+		NumHeads            uint32  `json:"num_heads"`
+		InChannels          uint32  `json:"in_chans"`
+		PatchSize           uint32  `json:"patch_size"`
+		SpatialMergeSize    uint32  `json:"spatial_merge_size"`
+		SpatialPatchSize    uint32  `json:"spatial_patch_size"`
+		WindowSize          uint32  `json:"window_size"`
+		RMSNormEps          float32 `json:"layer_norm_epsilon"`
+		RopeTheta           float32 `json:"rope_theta"`
+		FullAttentionBlocks []int32 `json:"fullatt_block_indexes"`
+		TemporalPatchSize   uint32  `json:"temporal_patch_size"`
+	} `json:"vision_config"`
+}
+
+var _ ModelConverter = (*qwen25VLModel)(nil)
+
+func (q *qwen25VLModel) KV(t *Tokenizer) ggml.KV {
+	kv := q.ModelParameters.KV(t)
+	kv["general.architecture"] = "qwen25vl"
+
+	for k, v := range q.qwen2Model.KV(t) {
+		if strings.HasPrefix(k, "qwen2.") {
+			kv[strings.Replace(k, "qwen2.", "qwen25vl.", 1)] = v
+		}
+	}
+
+	if q.VisionModel.FullAttentionBlocks == nil {
+		kv["qwen25vl.vision.fullatt_block_indexes"] = []int32{7, 15, 23, 31}
+	}
+
+	kv["qwen25vl.vision.block_count"] = cmp.Or(q.VisionModel.Depth, 32)
+	kv["qwen25vl.vision.embedding_length"] = q.VisionModel.HiddenSize
+	kv["qwen25vl.vision.attention.head_count"] = cmp.Or(q.VisionModel.NumHeads, 16)
+	kv["qwen25vl.vision.num_channels"] = q.VisionModel.InChannels
+	kv["qwen25vl.vision.patch_size"] = cmp.Or(q.VisionModel.PatchSize, 14)
+	kv["qwen25vl.vision.spatial_merge_size"] = cmp.Or(q.VisionModel.SpatialMergeSize, 2)
+	kv["qwen25vl.vision.spatial_patch_size"] = q.VisionModel.SpatialPatchSize
+	kv["qwen25vl.vision.window_size"] = cmp.Or(q.VisionModel.WindowSize, 112)
+	kv["qwen25vl.vision.attention.layer_norm_epsilon"] = cmp.Or(q.VisionModel.RMSNormEps, 1e-6)
+	kv["qwen25vl.vision.rope.freq_base"] = cmp.Or(q.VisionModel.RopeTheta, 1e4)
+	kv["qwen25vl.vision.fullatt_block_indexes"] = q.VisionModel.FullAttentionBlocks
+	kv["qwen25vl.vision.temporal_patch_size"] = cmp.Or(q.VisionModel.TemporalPatchSize, 2)
+
+	return kv
+}
+
+func (q *qwen25VLModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+
+	for _, t := range ts {
+		if strings.Contains(t.Name(), "patch_embed.proj") {
+			for t := range splitDim(t, 2,
+				strings.NewReplacer("patch_embed.proj", "patch_embd_0"),
+				strings.NewReplacer("patch_embed.proj", "patch_embd_1"),
+			) {
+				t.Shape = slices.DeleteFunc(t.Shape, func(i uint64) bool { return i == 1 })
+				out = append(out, t)
+			}
+		} else if strings.Contains(t.Name(), "attn.qkv") {
+			out = append(out, slices.Collect(splitDim(t, 0,
+				strings.NewReplacer("attn.qkv", "attn_q"),
+				strings.NewReplacer("attn.qkv", "attn_k"),
+				strings.NewReplacer("attn.qkv", "attn_v"),
+			))...)
+		} else {
+			out = append(out, &ggml.Tensor{
+				Name:     t.Name(),
+				Kind:     t.Kind(),
+				Shape:    t.Shape(),
+				WriterTo: t,
+			})
+		}
+	}
+
+	return out
+}
+
+func (p *qwen25VLModel) Replacements() []string {
+	return append(
+		p.qwen2Model.Replacements(),
+		"visual", "v",
+		"blocks", "blk",
+		"attn.proj", "attn_out",
+		"norm1", "ln1",
+		"norm2", "ln2",
+	)
+}

convert/convert_test.go

@@ -47,7 +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, -1)
+	m, err := ggml.Decode(r, -1)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -130,6 +130,7 @@ func TestConvertModel(t *testing.T) {
 			if err != nil {
 				t.Fatal(err)
 			}
+			defer expectFile.Close()
 
 			var expect map[string]string
 			if err := json.NewDecoder(expectFile).Decode(&expect); err != nil {
@@ -331,7 +332,7 @@ func TestConvertAdapter(t *testing.T) {
 			}
 			defer r.Close()
 
-			m, _, err := ggml.Decode(r, -1)
+			m, err := ggml.Decode(r, -1)
 			if err != nil {
 				t.Fatal(err)
 			}

convert/fs.go

@@ -1,58 +0,0 @@
-package convert
-
-import (
-	"archive/zip"
-	"errors"
-	"io"
-	"io/fs"
-	"os"
-	"path/filepath"
-)
-
-type ZipReader struct {
-	r *zip.Reader
-	p string
-
-	// limit is the maximum size of a file that can be read directly
-	// from the zip archive. Files larger than this size will be extracted
-	limit int64
-}
-
-func NewZipReader(r *zip.Reader, p string, limit int64) fs.FS {
-	return &ZipReader{r, p, limit}
-}
-
-func (z *ZipReader) Open(name string) (fs.File, error) {
-	r, err := z.r.Open(name)
-	if err != nil {
-		return nil, err
-	}
-	defer r.Close()
-
-	if fi, err := r.Stat(); err != nil {
-		return nil, err
-	} else if fi.Size() < z.limit {
-		return r, nil
-	}
-
-	if !filepath.IsLocal(name) {
-		return nil, zip.ErrInsecurePath
-	}
-
-	n := filepath.Join(z.p, name)
-	if _, err := os.Stat(n); errors.Is(err, os.ErrNotExist) {
-		w, err := os.Create(n)
-		if err != nil {
-			return nil, err
-		}
-		defer w.Close()
-
-		if _, err := io.Copy(w, r); err != nil {
-			return nil, err
-		}
-	} else if err != nil {
-		return nil, err
-	}
-
-	return os.Open(n)
-}

convert/reader.go

@@ -38,7 +38,10 @@ const (
 func (t tensorBase) Kind() uint32 {
 	if strings.HasSuffix(t.name, ".ffn_gate_inp.weight") ||
 		t.name == "token_types.weight" ||
-		t.name == "v.positional_embedding_vlm" {
+		t.name == "v.positional_embedding_vlm" ||
+		t.name == "v.tile_position_embd.weight" ||
+		t.name == "v.pre_tile_position_embd.weight" ||
+		t.name == "v.post_tile_position_embd.weight" {
 		// these tensors are always F32
 		return 0
 	}

convert/tensor.go

@@ -0,0 +1,56 @@
+package convert
+
+import (
+	"iter"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+)
+
+// splitDim splits a tensor along a specified dimension into multiple tensors. The dimension
+// is split evenly based on the number of replacers provided.
+func splitDim(t Tensor, dim int, replacers ...*strings.Replacer) iter.Seq[*ggml.Tensor] {
+	return func(yield func(*ggml.Tensor) bool) {
+		for i, replacer := range replacers {
+			shape := slices.Clone(t.Shape())
+			shape[dim] = shape[dim] / uint64(len(replacers))
+
+			slice := slices.Repeat([]tensor.Slice{nil}, len(shape))
+			slice[dim] = tensor.S(i*int(shape[dim]), (i+1)*int(shape[dim]))
+
+			tt := t.Clone()
+			tt.SetRepacker(func(_ string, data []float32, shape []uint64) ([]float32, error) {
+				dims := make([]int, len(shape))
+				for i := range shape {
+					dims[i] = int(shape[i])
+				}
+
+				var t tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+				t, err := t.Slice(slice...)
+				if err != nil {
+					return nil, err
+				}
+
+				t = tensor.Materialize(t)
+				// flatten tensor so it can be written as a vector
+				if err := t.Reshape(t.Shape().TotalSize()); err != nil {
+					return nil, err
+				}
+
+				return native.VectorF32(t.(*tensor.Dense))
+			})
+
+			if !yield(&ggml.Tensor{
+				Name:     replacer.Replace(t.Name()),
+				Kind:     t.Kind(),
+				Shape:    shape,
+				WriterTo: tt,
+			}) {
+				break
+			}
+		}
+	}
+}

convert/tokenizer.go

@@ -110,6 +110,7 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 	}
 
 	if f, err := fsys.Open("tokenizer_config.json"); errors.Is(err, os.ErrNotExist) {
+		// noop
 	} else if err != nil {
 		return nil, err
 	} else {
@@ -171,6 +172,34 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 		}
 	}
 
+	if f, err := fsys.Open("generation_config.json"); errors.Is(err, os.ErrNotExist) {
+	} else if err != nil {
+		return nil, err
+	} else {
+		defer f.Close()
+
+		var p map[string]json.RawMessage
+		if err := json.NewDecoder(f).Decode(&p); err != nil {
+			return nil, err
+		}
+
+		for _, st := range specialTokenTypes {
+			if bts, ok := p[fmt.Sprintf("%s_token_id", st)]; ok {
+				var ids []int32
+				if err := json.Unmarshal(bts, &ids); err != nil {
+					// value is not a list so the existing ID is used
+					continue
+				}
+
+				if i := slices.IndexFunc(t.SpecialVocabulary, func(sv *SpecialVocabulary) bool {
+					return sv.Type == st
+				}); i >= 0 {
+					t.SpecialVocabulary[i].IDs = ids
+				}
+			}
+		}
+	}
+
 	return t, nil
 }
 
@@ -280,6 +309,9 @@ type SpecialVocabulary struct {
 	ID       int
 	Content  string
 	AddToken bool
+
+	// IDs is populated by generation_config.json
+	IDs []int32
 }
 
 func (sv SpecialVocabulary) Key() string {

convert/tokenizer_test.go

@@ -247,6 +247,67 @@ func TestParseTokenizer(t *testing.T) {
 				Pre: "default",
 			},
 		},
+		{
+			name: "generation config eos token ids",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"added_tokens": [
+						{
+							"id": 0,
+							"content": "<bos>",
+							"special": true
+						},
+						{
+							"id": 1,
+							"content": "<eos>",
+							"special": true
+						},
+						{
+							"id": 2,
+							"content": "<eot>",
+							"special": true
+						},
+						{
+							"id": 3,
+							"content": "<eom>",
+							"special": true
+						}
+					],
+					"model": {
+						"vocab": {
+							"<bos>": 0,
+							"<eos>": 1,
+							"<eot>": 2,
+							"<eom>": 3
+						}
+					}
+				}`),
+				"tokenizer_config.json": strings.NewReader(`{
+					"add_bos_token": true,
+					"add_eos_token": false,
+					"bos_token": "<bos>",
+					"eos_token": "<eos>"
+				}`),
+				"generation_config.json": strings.NewReader(`{
+					"bos_token_id": 0,
+					"eos_token_id": [1, 2, 3]
+				}`),
+			}),
+			specialTokenTypes: []string{"pad", "eos", "bos", "unk"},
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model:  "gpt2",
+					Tokens: []string{"<bos>", "<eos>", "<eot>", "<eom>"},
+					Scores: []float32{0, 1, 2, 3},
+					Types:  []int32{3, 3, 3, 3},
+				},
+				SpecialVocabulary: []*SpecialVocabulary{
+					{Type: "eos", Content: "<eos>", ID: 1, IDs: []int32{1, 2, 3}, AddToken: false},
+					{Type: "bos", Content: "<bos>", ID: 0, AddToken: true},
+				},
+				Pre: "default",
+			},
+		},
 	}
 
 	for _, tt := range cases {

discover/gpu.go

@@ -670,7 +670,7 @@ func loadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string, e
 }
 
 func getVerboseState() C.uint16_t {
-	if envconfig.Debug() {
+	if envconfig.LogLevel() < slog.LevelInfo {
 		return C.uint16_t(1)
 	}
 	return C.uint16_t(0)

discover/gpu_info.h

@@ -27,12 +27,14 @@
 
 #endif
 
+#ifndef LOG
 #define LOG(verbose, ...) \
   do { \
     if (verbose) { \
       fprintf(stderr, __VA_ARGS__); \
     } \
   } while (0)
+#endif
 
 #ifdef __cplusplus
 extern "C" {

discover/gpu_info_cudart.c

@@ -1,6 +1,7 @@
 #ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
 
 #include <string.h>
+#include <inttypes.h>
 #include "gpu_info_cudart.h"
 
 void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
@@ -58,7 +59,7 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
     LOG(resp->ch.verbose, "cudaSetDevice err: %d\n", ret);
     UNLOAD_LIBRARY(resp->ch.handle);
     resp->ch.handle = NULL;
-    if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
+    if (ret == CUDART_ERROR_INSUFFICIENT_DRIVER) {
       resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
       return;
     }
@@ -168,9 +169,9 @@ void cudart_bootstrap(cudart_handle_t h, int i, mem_info_t *resp) {
   resp->free = memInfo.free;
   resp->used = memInfo.used;
 
-  LOG(h.verbose, "[%s] CUDA totalMem %lu\n", resp->gpu_id, resp->total);
-  LOG(h.verbose, "[%s] CUDA freeMem %lu\n", resp->gpu_id, resp->free);
-  LOG(h.verbose, "[%s] CUDA usedMem %lu\n", resp->gpu_id, resp->used);
+  LOG(h.verbose, "[%s] CUDA totalMem %" PRId64 "\n", resp->gpu_id, resp->total);
+  LOG(h.verbose, "[%s] CUDA freeMem %" PRId64 "\n", resp->gpu_id, resp->free);
+  LOG(h.verbose, "[%s] CUDA usedMem %" PRId64 "\n", resp->gpu_id, resp->used);
   LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
 }
 
@@ -180,4 +181,4 @@ void cudart_release(cudart_handle_t h) {
   h.handle = NULL;
 }
 
-#endif  // __APPLE__
+#endif  // __APPLE__

discover/gpu_info_nvcuda.c

@@ -1,6 +1,7 @@
 #ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
 
 #include <string.h>
+#include <inttypes.h>
 #include "gpu_info_nvcuda.h"
 
 void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
@@ -193,8 +194,8 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
   resp->total = memInfo.total;
   resp->free = memInfo.free;
 
-  LOG(h.verbose, "[%s] CUDA totalMem %lu mb\n", resp->gpu_id, resp->total / 1024 / 1024);
-  LOG(h.verbose, "[%s] CUDA freeMem %lu mb\n", resp->gpu_id, resp->free / 1024 / 1024);
+  LOG(h.verbose, "[%s] CUDA totalMem %" PRId64 "mb\n", resp->gpu_id, resp->total / 1024 / 1024);
+  LOG(h.verbose, "[%s] CUDA freeMem %" PRId64 "mb\n", resp->gpu_id, resp->free / 1024 / 1024);
   LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
 
   
@@ -247,4 +248,4 @@ void nvcuda_release(nvcuda_handle_t h) {
   h.handle = NULL;
 }
 
-#endif  // __APPLE__
+#endif  // __APPLE__

docs/api.md

@@ -19,7 +19,7 @@
 
 ### Model names
 
-Model names follow a `model:tag` format, where `model` can have an optional namespace such as `example/model`. Some examples are `orca-mini:3b-q4_1` and `llama3:70b`. The tag is optional and, if not provided, will default to `latest`. The tag is used to identify a specific version.
+Model names follow a `model:tag` format, where `model` can have an optional namespace such as `example/model`. Some examples are `orca-mini:3b-q8_0` and `llama3:70b`. The tag is optional and, if not provided, will default to `latest`. The tag is used to identify a specific version.
 
 ### Durations
 
@@ -43,6 +43,7 @@ Generate a response for a given prompt with a provided model. This is a streamin
 - `prompt`: the prompt to generate a response for
 - `suffix`: the text after the model response
 - `images`: (optional) a list of base64-encoded images (for multimodal models such as `llava`)
+- `think`: (for thinking models) should the model think before responding?
 
 Advanced parameters (optional):
 
@@ -394,9 +395,6 @@ curl http://localhost:11434/api/generate -d '{
     "repeat_penalty": 1.2,
     "presence_penalty": 1.5,
     "frequency_penalty": 1.0,
-    "mirostat": 1,
-    "mirostat_tau": 0.8,
-    "mirostat_eta": 0.6,
     "penalize_newline": true,
     "stop": ["\n", "user:"],
     "numa": false,
@@ -404,10 +402,7 @@ curl http://localhost:11434/api/generate -d '{
     "num_batch": 2,
     "num_gpu": 1,
     "main_gpu": 0,
-    "low_vram": false,
-    "vocab_only": false,
     "use_mmap": true,
-    "use_mlock": false,
     "num_thread": 8
   }
 }'
@@ -496,11 +491,13 @@ Generate the next message in a chat with a provided model. This is a streaming e
 - `model`: (required) the [model name](#model-names)
 - `messages`: the messages of the chat, this can be used to keep a chat memory
 - `tools`: list of tools in JSON for the model to use if supported
+- `think`: (for thinking models) should the model think before responding?
 
 The `message` object has the following fields:
 
 - `role`: the role of the message, either `system`, `user`, `assistant`, or `tool`
 - `content`: the content of the message
+- `thinking`: (for thinking models) the model's thinking process
 - `images` (optional): a list of images to include in the message (for multimodal models such as `llava`)
 - `tool_calls` (optional): a list of tools in JSON that the model wants to use
 
@@ -958,19 +955,8 @@ If you are creating a model from a safetensors directory or from a GGUF file, yo
 
 | Type | Recommended |
 | --- | :-: |
-| q2_K | |
-| q3_K_L | |
-| q3_K_M | |
-| q3_K_S | |
-| q4_0 | |
-| q4_1 | |
 | q4_K_M | * |
 | q4_K_S | |
-| q5_0 | |
-| q5_1 | |
-| q5_K_M | |
-| q5_K_S | |
-| q6_K | |
 | q8_0 | * |
 
 ### Examples
@@ -1015,8 +1001,8 @@ Quantize a non-quantized model.
 
 ```shell
 curl http://localhost:11434/api/create -d '{
-  "model": "llama3.1:quantized",
-  "from": "llama3.1:8b-instruct-fp16",
+  "model": "llama3.2:quantized",
+  "from": "llama3.2:3b-instruct-fp16",
   "quantize": "q4_K_M"
 }'
 ```
@@ -1026,12 +1012,14 @@ curl http://localhost:11434/api/create -d '{
 A stream of JSON objects is returned:
 
 ```json
-{"status":"quantizing F16 model to Q4_K_M"}
-{"status":"creating new layer sha256:667b0c1932bc6ffc593ed1d03f895bf2dc8dc6df21db3042284a6f4416b06a29"}
-{"status":"using existing layer sha256:11ce4ee3e170f6adebac9a991c22e22ab3f8530e154ee669954c4bc73061c258"}
-{"status":"using existing layer sha256:0ba8f0e314b4264dfd19df045cde9d4c394a52474bf92ed6a3de22a4ca31a177"}
+{"status":"quantizing F16 model to Q4_K_M","digest":"0","total":6433687776,"completed":12302}
+{"status":"quantizing F16 model to Q4_K_M","digest":"0","total":6433687776,"completed":6433687552}
+{"status":"verifying conversion"}
+{"status":"creating new layer sha256:fb7f4f211b89c6c4928ff4ddb73db9f9c0cfca3e000c3e40d6cf27ddc6ca72eb"}
+{"status":"using existing layer sha256:966de95ca8a62200913e3f8bfbf84c8494536f1b94b49166851e76644e966396"}
+{"status":"using existing layer sha256:fcc5a6bec9daf9b561a68827b67ab6088e1dba9d1fa2a50d7bbcc8384e0a265d"}
+{"status":"using existing layer sha256:a70ff7e570d97baaf4e62ac6e6ad9975e04caa6d900d3742d37698494479e0cd"}
 {"status":"using existing layer sha256:56bb8bd477a519ffa694fc449c2413c6f0e1d3b1c88fa7e3c9d88d3ae49d4dcb"}
-{"status":"creating new layer sha256:455f34728c9b5dd3376378bfb809ee166c145b0b4c1f1a6feca069055066ef9a"}
 {"status":"writing manifest"}
 {"status":"success"}
 ```
@@ -1169,29 +1157,46 @@ A single JSON object will be returned.
 {
   "models": [
     {
-      "name": "codellama:13b",
+
+      "model": "codellama:13b",
       "modified_at": "2023-11-04T14:56:49.277302595-07:00",
       "size": 7365960935,
       "digest": "9f438cb9cd581fc025612d27f7c1a6669ff83a8bb0ed86c94fcf4c5440555697",
+      "capabilities": [
+        "completion"
+      ],
+
       "details": {
+        "parent_model": "",
         "format": "gguf",
-        "family": "llama",
-        "families": null,
-        "parameter_size": "13B",
-        "quantization_level": "Q4_0"
+        "family": "qwen2",
+        "families": [
+          "qwen2"
+        ],
+        "parameter_size": "7.6B",
+        "quantization_level": "Q4_K_M"
       }
     },
     {
-      "name": "llama3:latest",
+
+      "model": "llama4:latest",
       "modified_at": "2023-12-07T09:32:18.757212583-08:00",
       "size": 3825819519,
       "digest": "fe938a131f40e6f6d40083c9f0f430a515233eb2edaa6d72eb85c50d64f2300e",
+      "capabilities": [
+        "completion",
+        "vision"
+      ],
+
       "details": {
+        "parent_model": "",
         "format": "gguf",
         "family": "llama",
-        "families": null,
-        "parameter_size": "7B",
-        "quantization_level": "Q4_0"
+        "families": [
+          "llama"
+        ],
+        "parameter_size": "3.2B",
+        "quantization_level": "Q4_K_M"
       }
     }
   ]

docs/import.md

@@ -132,22 +132,12 @@ success
 
 ### Supported Quantizations
 
-- `q4_0`
-- `q4_1`
-- `q5_0`
-- `q5_1`
 - `q8_0`
 
 #### K-means Quantizations
 
-- `q3_K_S`
-- `q3_K_M`
-- `q3_K_L`
 - `q4_K_S`
 - `q4_K_M`
-- `q5_K_S`
-- `q5_K_M`
-- `q6_K`
 
 
 ## Sharing your model on ollama.com

docs/modelfile.md

@@ -150,9 +150,6 @@ PARAMETER <parameter> <parametervalue>
 
 | Parameter      | Description                                                                                                                                                                                                                                             | Value Type | Example Usage        |
 | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------- |
-| mirostat       | Enable Mirostat sampling for controlling perplexity. (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)                                                                                                                                         | int        | mirostat 0           |
-| mirostat_eta   | Influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. (Default: 0.1)                        | float      | mirostat_eta 0.1     |
-| mirostat_tau   | Controls the balance between coherence and diversity of the output. A lower value will result in more focused and coherent text. (Default: 5.0)                                                                                                         | float      | mirostat_tau 5.0     |
 | num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                    | int        | num_ctx 4096         |
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                           | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                     | float      | repeat_penalty 1.1   |

envconfig/config.go

@@ -149,9 +149,22 @@ func Bool(k string) func() bool {
 	}
 }
 
+// LogLevel returns the log level for the application.
+// Values are 0 or false INFO (Default), 1 or true DEBUG, 2 TRACE
+func LogLevel() slog.Level {
+	level := slog.LevelInfo
+	if s := Var("OLLAMA_DEBUG"); s != "" {
+		if b, _ := strconv.ParseBool(s); b {
+			level = slog.LevelDebug
+		} else if i, _ := strconv.ParseInt(s, 10, 64); i != 0 {
+			level = slog.Level(i * -4)
+		}
+	}
+
+	return level
+}
+
 var (
-	// Debug enabled additional debug information.
-	Debug = Bool("OLLAMA_DEBUG")
 	// FlashAttention enables the experimental flash attention feature.
 	FlashAttention = Bool("OLLAMA_FLASH_ATTENTION")
 	// KvCacheType is the quantization type for the K/V cache.
@@ -170,6 +183,8 @@ var (
 	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 	// ContextLength sets the default context length
 	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
+	// Auth enables authentication between the Ollama client and server
+	UseAuth = Bool("OLLAMA_AUTH")
 )
 
 func String(s string) func() string {
@@ -209,8 +224,6 @@ var (
 	MaxRunners = Uint("OLLAMA_MAX_LOADED_MODELS", 0)
 	// MaxQueue sets the maximum number of queued requests. MaxQueue can be configured via the OLLAMA_MAX_QUEUE environment variable.
 	MaxQueue = Uint("OLLAMA_MAX_QUEUE", 512)
-	// MaxVRAM sets a maximum VRAM override in bytes. MaxVRAM can be configured via the OLLAMA_MAX_VRAM environment variable.
-	MaxVRAM = Uint("OLLAMA_MAX_VRAM", 0)
 )
 
 func Uint64(key string, defaultValue uint64) func() uint64 {
@@ -238,7 +251,7 @@ type EnvVar struct {
 
 func AsMap() map[string]EnvVar {
 	ret := map[string]EnvVar{
-		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
+		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", LogLevel(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
 		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(), "Enabled flash attention"},
 		"OLLAMA_KV_CACHE_TYPE":     {"OLLAMA_KV_CACHE_TYPE", KvCacheType(), "Quantization type for the K/V cache (default: f16)"},
 		"OLLAMA_GPU_OVERHEAD":      {"OLLAMA_GPU_OVERHEAD", GpuOverhead(), "Reserve a portion of VRAM per GPU (bytes)"},

envconfig/config_test.go

@@ -1,11 +1,13 @@
 package envconfig
 
 import (
+	"log/slog"
 	"math"
 	"testing"
 	"time"
 
 	"github.com/google/go-cmp/cmp"
+	"github.com/ollama/ollama/logutil"
 )
 
 func TestHost(t *testing.T) {
@@ -292,3 +294,34 @@ func TestContextLength(t *testing.T) {
 		})
 	}
 }
+
+func TestLogLevel(t *testing.T) {
+	cases := map[string]slog.Level{
+		// Default to INFO
+		"":      slog.LevelInfo,
+		"false": slog.LevelInfo,
+		"f":     slog.LevelInfo,
+		"0":     slog.LevelInfo,
+
+		// True values enable Debug
+		"true": slog.LevelDebug,
+		"t":    slog.LevelDebug,
+
+		// Positive values increase verbosity
+		"1": slog.LevelDebug,
+		"2": logutil.LevelTrace,
+
+		// Negative values decrease verbosity
+		"-1": slog.LevelWarn,
+		"-2": slog.LevelError,
+	}
+
+	for k, v := range cases {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv("OLLAMA_DEBUG", k)
+			if i := LogLevel(); i != v {
+				t.Errorf("%s: expected %d, got %d", k, v, i)
+			}
+		})
+	}
+}

fs/ggml/ggml.go

@@ -15,6 +15,7 @@ import (
 type GGML struct {
 	container
 	model
+	Length int64
 }
 
 type model interface {
@@ -36,12 +37,12 @@ func (kv KV) ParameterCount() uint64 {
 	return keyValue(kv, "general.parameter_count", uint64(0))
 }
 
-func (kv KV) FileType() fileType {
+func (kv KV) FileType() FileType {
 	if t := kv.Uint("general.file_type"); t > 0 {
-		return fileType(t)
+		return FileType(t)
 	}
 
-	return fileTypeUnknown
+	return FileTypeUnknown
 }
 
 func (kv KV) BlockCount() uint64 {
@@ -125,6 +126,8 @@ func (kv KV) OllamaEngineRequired() bool {
 		"gemma3",
 		"mistral3",
 		"llama4",
+		"mllama",
+		"qwen25vl",
 	}, kv.Architecture())
 }
 
@@ -149,7 +152,7 @@ func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ..
 		return val.(T)
 	}
 
-	slog.Warn("key not found", "key", key, "default", defaultValue[0])
+	slog.Debug("key not found", "key", key, "default", defaultValue[0])
 	return defaultValue[0]
 }
 
@@ -226,7 +229,11 @@ func (t Tensor) block() (n int) {
 }
 
 func (t Tensor) blockSize() uint64 {
-	switch t.Kind {
+	return (TensorType)(t.Kind).BlockSize()
+}
+
+func (t TensorType) BlockSize() uint64 {
+	switch t {
 	case
 		0,  // F32
 		1,  // F16
@@ -252,73 +259,77 @@ func (t Tensor) blockSize() uint64 {
 }
 
 func (t Tensor) typeSize() uint64 {
-	blockSize := t.blockSize()
+	return TensorType(t.Kind).TypeSize()
+}
 
-	switch t.Kind {
-	case 0: // FP32
+func (t TensorType) TypeSize() uint64 {
+	blockSize := t.BlockSize()
+
+	switch t {
+	case TensorTypeF32:
 		return 4
-	case 1: // FP16
+	case TensorTypeF16:
 		return 2
-	case 2: // Q4_0
+	case TensorTypeQ4_0:
 		return 2 + blockSize/2
-	case 3: // Q4_1
+	case TensorTypeQ4_1:
 		return 2 + 2 + blockSize/2
-	case 6: // Q5_0
+	case TensorTypeQ5_0:
 		return 2 + 4 + blockSize/2
-	case 7: // Q5_1
+	case TensorTypeQ5_1:
 		return 2 + 2 + 4 + blockSize/2
-	case 8: // Q8_0
+	case TensorTypeQ8_0:
 		return 2 + blockSize
-	case 9: // Q8_1
+	case TensorTypeQ8_1:
 		return 2 + 2 + blockSize
-	case 10: // Q2_K
+	case TensorTypeQ2_K:
 		return blockSize/16 + blockSize/4 + 2 + 2
-	case 11: // Q3_K
+	case TensorTypeQ3_K:
 		return blockSize/8 + blockSize/4 + 12 + 2
-	case 12: // Q4_K
+	case TensorTypeQ4_K:
 		return 2 + 2 + 12 + blockSize/2
-	case 13: // Q5_K
+	case TensorTypeQ5_K:
 		return 2 + 2 + 12 + blockSize/8 + blockSize/2
-	case 14: // Q6_K
+	case TensorTypeQ6_K:
 		return blockSize/2 + blockSize/4 + blockSize/16 + 2
-	case 15: // Q8_K
+	case TensorTypeQ8_K:
 		return 4 + blockSize + 2*blockSize/16
-	case 16: // IQ2_XXS
+	case tensorTypeIQ2_XXS:
 		return 2 + 2*blockSize/8
-	case 17: // IQ2_XS
+	case tensorTypeIQ2_XS:
 		return 2 + 2*blockSize/8 + blockSize/32
-	case 18: // IQ3_XXS
+	case tensorTypeIQ3_XXS:
 		return 2 + blockSize/4 + blockSize/8
-	case 19: // IQ1_S
+	case tensorTypeIQ1_S:
 		return 2 + blockSize/8 + blockSize/16
-	case 20: // IQ4_NL
+	case tensorTypeIQ4_NL:
 		return 2 + blockSize/2
-	case 21: // IQ3_S
+	case tensorTypeIQ3_S:
 		return 2 + blockSize/4 + blockSize/8 + blockSize/32 + 4
-	case 22: // IQ2_S
+	case tensorTypeIQ2_S:
 		return 2 + blockSize/4 + blockSize/16
-	case 23: // IQ4_XS
+	case tensorTypeIQ4_XS:
 		return 2 + 2 + blockSize/2 + blockSize/64
-	case 24: // I8
+	case TensorTypeI8:
 		return 1
-	case 25: // I16
+	case TensorTypeI16:
 		return 2
-	case 26: // I32
+	case TensorTypeI32:
 		return 4
-	case 27: // I64
+	case TensorTypeI64:
 		return 8
-	case 28: // F64
+	case TensorTypeF64:
 		return 8
-	case 29: // IQ1_M
+	case tensorTypeIQ1_M:
 		return blockSize/8 + blockSize/16 + blockSize/32
-	case 30: // BF16
+	case TensorTypeBF16:
 		return 2
 	default:
 		return 0
 	}
 }
 
-func (t Tensor) parameters() uint64 {
+func (t Tensor) Elements() uint64 {
 	var count uint64 = 1
 	for _, n := range t.Shape {
 		count *= n
@@ -327,11 +338,11 @@ func (t Tensor) parameters() uint64 {
 }
 
 func (t Tensor) Size() uint64 {
-	return t.parameters() * t.typeSize() / t.blockSize()
+	return t.Elements() * t.typeSize() / t.blockSize()
 }
 
 func (t Tensor) Type() string {
-	return fileType(t.Kind).String()
+	return TensorType(t.Kind).String()
 }
 
 type container interface {
@@ -376,12 +387,12 @@ func DetectContentType(b []byte) string {
 //
 // It collects array values for arrays with a size less than or equal to
 // maxArraySize. If the maxArraySize is negative, all arrays are collected.
-func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
+func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, error) {
 	rs = bufioutil.NewBufferedSeeker(rs, 32<<10)
 
 	var magic uint32
 	if err := binary.Read(rs, binary.LittleEndian, &magic); err != nil {
-		return nil, 0, err
+		return nil, err
 	}
 
 	var c container
@@ -391,24 +402,25 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	case FILE_MAGIC_GGUF_BE:
 		c = &containerGGUF{ByteOrder: binary.BigEndian, maxArraySize: maxArraySize}
 	default:
-		return nil, 0, errors.New("invalid file magic")
+		return nil, errors.New("invalid file magic")
 	}
 
 	model, err := c.Decode(rs)
 	if err != nil {
-		return nil, 0, err
+		return nil, err
 	}
 
 	offset, err := rs.Seek(0, io.SeekCurrent)
 	if err != nil {
-		return nil, 0, err
+		return nil, err
 	}
 
 	// final model type
 	return &GGML{
 		container: c,
 		model:     model,
-	}, offset, nil
+		Length:    offset,
+	}, nil
 }
 
 func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
@@ -480,7 +492,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 		var ropeFreqsCount uint64
 		if ropeFreqs, ok := f.Tensors().GroupLayers()["rope_freqs"]; ok {
 			if ropeFreqsWeights, ok := ropeFreqs["weights"]; ok {
-				ropeFreqsCount = ropeFreqsWeights.parameters()
+				ropeFreqsCount = ropeFreqsWeights.Elements()
 			}
 		}
 
@@ -640,6 +652,20 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
 		graphSize = 4 * (imageSize*imageSize*numChannels +
 			embeddingLength*patchSize +
 			numPatches*numPatches*headCount)
+	case "qwen25vl":
+		maxPixels := uint64(llm.KV().Uint("vision.max_pixels", 28*28*1280))
+
+		numPatches := maxPixels / (patchSize * patchSize)
+
+		graphSize = 4 * (maxPixels*numChannels + // Original image storage
+			// Normalized pixels
+			maxPixels*numChannels +
+			// Patches storage (numPatches * channels * patchSize^2)
+			numPatches*numChannels*patchSize*patchSize +
+			// Self-attention calculations
+			numPatches*numPatches*headCount +
+			// Additional buffer for processing
+			embeddingLength*numPatches)
 	case "llama4":
 		// vision graph is computed independently in the same schedule
 		// and is negligible compared to the worst case text graph

fs/ggml/gguf.go

@@ -9,8 +9,12 @@ import (
 	"io"
 	"log/slog"
 	"maps"
+	"os"
+	"runtime"
 	"slices"
 	"strings"
+
+	"golang.org/x/sync/errgroup"
 )
 
 type containerGGUF struct {
@@ -225,7 +229,7 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 		}
 
 		llm.tensors = append(llm.tensors, &tensor)
-		llm.parameters += tensor.parameters()
+		llm.parameters += tensor.Elements()
 	}
 
 	// patch KV with parameter count
@@ -488,25 +492,38 @@ func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
 		return err
 	}
 
+	if t == ggufTypeString {
+		for _, e := range any(s).([]string) {
+			if err := binary.Write(w, binary.LittleEndian, uint64(len(e))); err != nil {
+				return err
+			}
+
+			if err := binary.Write(w, binary.LittleEndian, []byte(e)); err != nil {
+				return err
+			}
+		}
+		return nil
+	}
+
 	return binary.Write(w, binary.LittleEndian, s)
 }
 
-func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
+func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 	alignment := kv.Uint("general.alignment", 32)
 
-	if err := binary.Write(ws, binary.LittleEndian, []byte("GGUF")); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, []byte("GGUF")); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, binary.LittleEndian, uint32(3)); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, uint32(3)); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, binary.LittleEndian, uint64(len(ts))); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, uint64(len(ts))); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, binary.LittleEndian, uint64(len(kv))); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, uint64(len(kv))); err != nil {
 		return err
 	}
 
@@ -514,12 +531,12 @@ func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
 	slices.Sort(keys)
 
 	for _, key := range keys {
-		if err := ggufWriteKV(ws, key, kv[key]); err != nil {
+		if err := ggufWriteKV(f, key, kv[key]); err != nil {
 			return err
 		}
 	}
 
-	slices.SortStableFunc(ts, func(a, b Tensor) int {
+	slices.SortStableFunc(ts, func(a, b *Tensor) int {
 		if i, j := a.block(), b.block(); i < 0 && j > 0 {
 			return 1
 		} else if i > 0 && j < 0 {
@@ -530,21 +547,34 @@ func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
 	})
 
 	var s uint64
-	for _, t := range ts {
-		t.Offset = s + uint64(ggufPadding(int64(s), int64(alignment)))
-		if err := ggufWriteTensorInfo(ws, t); err != nil {
+	for i := range ts {
+		ts[i].Offset = s
+		if err := ggufWriteTensorInfo(f, ts[i]); err != nil {
 			return err
 		}
-		s += t.Size()
+		s += ts[i].Size()
+		s += uint64(ggufPadding(int64(s), int64(alignment)))
 	}
 
+	offset, err := f.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
+	}
+	offset += ggufPadding(offset, int64(alignment))
+
+	var g errgroup.Group
+	g.SetLimit(runtime.GOMAXPROCS(0))
+	// TODO consider reducing if tensors size * gomaxprocs is larger than free memory
 	for _, t := range ts {
-		if err := ggufWriteTensor(ws, t, int64(alignment)); err != nil {
+		t := t
+		w := io.NewOffsetWriter(f, offset+int64(t.Offset))
+		g.Go(func() error {
+			_, err := t.WriteTo(w)
 			return err
-		}
+		})
 	}
 
-	return nil
+	return g.Wait()
 }
 
 func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
@@ -559,8 +589,10 @@ func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
 
 	var err error
 	switch v := v.(type) {
-	case uint32:
+	case uint32, FileType:
 		err = writeGGUF(ws, ggufTypeUint32, v)
+	case uint64:
+		err = writeGGUF(ws, ggufTypeUint64, v)
 	case float32:
 		err = writeGGUF(ws, ggufTypeFloat32, v)
 	case bool:
@@ -569,32 +601,20 @@ func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
 		err = writeGGUFString(ws, v)
 	case []int32:
 		err = writeGGUFArray(ws, ggufTypeInt32, v)
+	case *array[int32]:
+		err = writeGGUFArray(ws, ggufTypeInt32, v.values)
 	case []uint32:
 		err = writeGGUFArray(ws, ggufTypeUint32, v)
+	case *array[uint32]:
+		err = writeGGUFArray(ws, ggufTypeUint32, v.values)
 	case []float32:
 		err = writeGGUFArray(ws, ggufTypeFloat32, v)
+	case *array[float32]:
+		err = writeGGUFArray(ws, ggufTypeFloat32, v.values)
 	case []string:
-		if err := binary.Write(ws, binary.LittleEndian, ggufTypeArray); err != nil {
-			return err
-		}
-
-		if err := binary.Write(ws, binary.LittleEndian, ggufTypeString); err != nil {
-			return err
-		}
-
-		if err := binary.Write(ws, binary.LittleEndian, uint64(len(v))); err != nil {
-			return err
-		}
-
-		for _, e := range v {
-			if err := binary.Write(ws, binary.LittleEndian, uint64(len(e))); err != nil {
-				return err
-			}
-
-			if err := binary.Write(ws, binary.LittleEndian, []byte(e)); err != nil {
-				return err
-			}
-		}
+		err = writeGGUFArray(ws, ggufTypeString, v)
+	case *array[string]:
+		err = writeGGUFArray(ws, ggufTypeString, v.values)
 	default:
 		return fmt.Errorf("improper type for '%s'", k)
 	}
@@ -602,7 +622,7 @@ func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
 	return err
 }
 
-func ggufWriteTensorInfo(ws io.WriteSeeker, t Tensor) error {
+func ggufWriteTensorInfo(ws io.WriteSeeker, t *Tensor) error {
 	slog.Debug(t.Name, "kind", t.Kind, "shape", t.Shape, "offset", t.Offset)
 	if err := binary.Write(ws, binary.LittleEndian, uint64(len(t.Name))); err != nil {
 		return err
@@ -629,20 +649,6 @@ func ggufWriteTensorInfo(ws io.WriteSeeker, t Tensor) error {
 	return binary.Write(ws, binary.LittleEndian, t.Offset)
 }
 
-func ggufWriteTensor(ws io.WriteSeeker, t Tensor, alignment int64) error {
-	offset, err := ws.Seek(0, io.SeekCurrent)
-	if err != nil {
-		return err
-	}
-
-	if err := binary.Write(ws, binary.LittleEndian, bytes.Repeat([]byte{0}, int(ggufPadding(offset, alignment)))); err != nil {
-		return err
-	}
-
-	_, err = t.WriteTo(ws)
-	return err
-}
-
 func ggufPadding(offset, align int64) int64 {
 	return (align - offset%align) % align
 }

fs/ggml/gguf_test.go

@@ -0,0 +1,63 @@
+package ggml
+
+import (
+	"bytes"
+	"os"
+	"slices"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+)
+
+func TestWriteGGUF(t *testing.T) {
+	w, err := os.CreateTemp(t.TempDir(), "*.bin")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer w.Close()
+
+	if err := WriteGGUF(w, KV{
+		"general.alignment": uint32(16),
+	}, []*Tensor{
+		{Name: "test.0", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.1", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.2", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.3", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.4", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.5", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	r, err := os.Open(w.Name())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer r.Close()
+
+	ff, err := Decode(r, 0)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if diff := cmp.Diff(ff.KV(), KV{
+		"general.alignment":       uint32(16),
+		"general.parameter_count": uint64(36),
+	}); diff != "" {
+		t.Errorf("Mismatch (-want +got):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(ff.Tensors(), Tensors{
+		Offset: 336,
+		items: []*Tensor{
+			{Name: "test.0", Offset: 0, Shape: []uint64{2, 3}},
+			{Name: "test.1", Offset: 32, Shape: []uint64{2, 3}},
+			{Name: "test.2", Offset: 64, Shape: []uint64{2, 3}},
+			{Name: "test.3", Offset: 96, Shape: []uint64{2, 3}},
+			{Name: "test.4", Offset: 128, Shape: []uint64{2, 3}},
+			{Name: "test.5", Offset: 160, Shape: []uint64{2, 3}},
+		},
+	}, cmp.AllowUnexported(Tensors{})); diff != "" {
+		t.Errorf("Mismatch (-want +got):\n%s", diff)
+	}
+}

fs/ggml/type.go

@@ -1,26 +1,31 @@
 package ggml
 
-import "fmt"
+import (
+	"fmt"
+	"log/slog"
+	"strings"
+)
 
-type fileType uint32
+// FileType is the Go equivalent to llama_ftype used for gguf file typing
+type FileType uint32
 
 const (
-	fileTypeF32 fileType = iota
-	fileTypeF16
+	FileTypeF32 FileType = iota
+	FileTypeF16
 	fileTypeQ4_0
 	fileTypeQ4_1
-	fileTypeQ4_1_F16
-	fileTypeQ4_2 // unused
-	fileTypeQ4_3 // unused
-	fileTypeQ8_0
+	fileTypeQ4_1_F16 // unused by GGML
+	fileTypeQ4_2     // unused by GGML
+	fileTypeQ4_3     // unused by GGML
+	FileTypeQ8_0
 	fileTypeQ5_0
 	fileTypeQ5_1
 	fileTypeQ2_K
 	fileTypeQ3_K_S
 	fileTypeQ3_K_M
 	fileTypeQ3_K_L
-	fileTypeQ4_K_S
-	fileTypeQ4_K_M
+	FileTypeQ4_K_S
+	FileTypeQ4_K_M
 	fileTypeQ5_K_S
 	fileTypeQ5_K_M
 	fileTypeQ6_K
@@ -37,93 +42,62 @@ const (
 	fileTypeIQ2_M
 	fileTypeIQ4_XS
 	fileTypeIQ1_M
-	fileTypeBF16
+	FileTypeBF16
+	fileTypeQ4_0_4_4 // unused by GGML
+	fileTypeQ4_0_4_8 // unused by GGML
+	fileTypeQ4_0_8_8 // unused by GGML
+	fileTypeTQ1_0
+	fileTypeTQ2_0
 
-	fileTypeUnknown
+	FileTypeUnknown = 1024
 )
 
-func ParseFileType(s string) (fileType, error) {
+// ParseFileType parses the provided GGUF file type
+// Only Ollama supported types are considered valid
+func ParseFileType(s string) (FileType, error) {
 	switch s {
 	case "F32":
-		return fileTypeF32, nil
+		return FileTypeF32, nil
 	case "F16":
-		return fileTypeF16, nil
-	case "Q4_0":
-		return fileTypeQ4_0, nil
-	case "Q4_1":
-		return fileTypeQ4_1, nil
-	case "Q4_1_F16":
-		return fileTypeQ4_1_F16, nil
+		return FileTypeF16, nil
 	case "Q8_0":
-		return fileTypeQ8_0, nil
-	case "Q5_0":
-		return fileTypeQ5_0, nil
-	case "Q5_1":
-		return fileTypeQ5_1, nil
-	case "Q2_K":
-		return fileTypeQ2_K, nil
-	case "Q3_K_S":
-		return fileTypeQ3_K_S, nil
-	case "Q3_K_M":
-		return fileTypeQ3_K_M, nil
-	case "Q3_K_L":
-		return fileTypeQ3_K_L, nil
+		return FileTypeQ8_0, nil
 	case "Q4_K_S":
-		return fileTypeQ4_K_S, nil
-	case "Q4_K_M":
-		return fileTypeQ4_K_M, nil
-	case "Q5_K_S":
-		return fileTypeQ5_K_S, nil
-	case "Q5_K_M":
-		return fileTypeQ5_K_M, nil
-	case "Q6_K":
-		return fileTypeQ6_K, nil
-	case "IQ2_XXS":
-		return fileTypeIQ2_XXS, nil
-	case "IQ2_XS":
-		return fileTypeIQ2_XS, nil
-	case "Q2_K_S":
-		return fileTypeQ2_K_S, nil
-	case "IQ3_XS":
-		return fileTypeIQ3_XS, nil
-	case "IQ3_XXS":
-		return fileTypeIQ3_XXS, nil
-	case "IQ1_S":
-		return fileTypeIQ1_S, nil
-	case "IQ4_NL":
-		return fileTypeIQ4_NL, nil
-	case "IQ3_S":
-		return fileTypeIQ3_S, nil
-	case "IQ3_M":
-		return fileTypeIQ3_M, nil
-	case "IQ2_S":
-		return fileTypeIQ2_S, nil
-	case "IQ2_M":
-		return fileTypeIQ2_M, nil
-	case "IQ4_XS":
-		return fileTypeIQ4_XS, nil
-	case "IQ1_M":
-		return fileTypeIQ1_M, nil
+		return FileTypeQ4_K_S, nil
+	case "Q4_K_M", "Q4_K":
+		return FileTypeQ4_K_M, nil
 	case "BF16":
-		return fileTypeBF16, nil
+		return FileTypeBF16, nil
 	default:
-		return fileTypeUnknown, fmt.Errorf("unknown fileType: %s", s)
+		supportedFileTypes := []FileType{
+			FileTypeF32,
+			FileTypeF16,
+			FileTypeQ4_K_S,
+			FileTypeQ4_K_M,
+			FileTypeQ8_0,
+			// fsggml.FileTypeBF16, // TODO
+		}
+		strs := make([]string, len(supportedFileTypes))
+		for i := range supportedFileTypes {
+			strs[i] = supportedFileTypes[i].String()
+		}
+
+		return FileTypeUnknown, fmt.Errorf("unsupported quantization type %s - supported types are %s", s, strings.Join(strs, ", "))
 	}
 }
 
-func (t fileType) String() string {
+func (t FileType) String() string {
+	// Note: this routine will return a broader set of file types for existing models
 	switch t {
-	case fileTypeF32:
+	case FileTypeF32:
 		return "F32"
-	case fileTypeF16:
+	case FileTypeF16:
 		return "F16"
 	case fileTypeQ4_0:
 		return "Q4_0"
 	case fileTypeQ4_1:
 		return "Q4_1"
-	case fileTypeQ4_1_F16:
-		return "Q4_1_F16"
-	case fileTypeQ8_0:
+	case FileTypeQ8_0:
 		return "Q8_0"
 	case fileTypeQ5_0:
 		return "Q5_0"
@@ -137,9 +111,9 @@ func (t fileType) String() string {
 		return "Q3_K_M"
 	case fileTypeQ3_K_L:
 		return "Q3_K_L"
-	case fileTypeQ4_K_S:
+	case FileTypeQ4_K_S:
 		return "Q4_K_S"
-	case fileTypeQ4_K_M:
+	case FileTypeQ4_K_M:
 		return "Q4_K_M"
 	case fileTypeQ5_K_S:
 		return "Q5_K_S"
@@ -147,39 +121,198 @@ func (t fileType) String() string {
 		return "Q5_K_M"
 	case fileTypeQ6_K:
 		return "Q6_K"
-	case fileTypeIQ2_XXS:
-		return "IQ2_XXS"
-	case fileTypeIQ2_XS:
-		return "IQ2_XS"
 	case fileTypeQ2_K_S:
 		return "Q2_K_S"
-	case fileTypeIQ3_XS:
-		return "IQ3_XS"
-	case fileTypeIQ3_XXS:
-		return "IQ3_XXS"
-	case fileTypeIQ1_S:
-		return "IQ1_S"
-	case fileTypeIQ4_NL:
-		return "IQ4_NL"
-	case fileTypeIQ3_S:
-		return "IQ3_S"
-	case fileTypeIQ3_M:
-		return "IQ3_M"
-	case fileTypeIQ2_S:
-		return "IQ2_S"
-	case fileTypeIQ4_XS:
-		return "IQ4_XS"
-	case fileTypeIQ2_M:
-		return "IQ2_M"
-	case fileTypeIQ1_M:
-		return "IQ1_M"
-	case fileTypeBF16:
+	case FileTypeBF16:
 		return "BF16"
 	default:
 		return "unknown"
 	}
 }
 
-func (t fileType) Value() uint32 {
+func (t FileType) Value() uint32 {
 	return uint32(t)
 }
+
+func (ftype FileType) ToTensorType() TensorType {
+	switch ftype {
+	case FileTypeF32:
+		return TensorTypeF32
+	case FileTypeF16:
+		return TensorTypeF16
+	case fileTypeQ4_0:
+		return TensorTypeQ4_0
+	case fileTypeQ4_1:
+		return TensorTypeQ4_1
+	case FileTypeQ8_0:
+		return TensorTypeQ8_0
+	case fileTypeQ5_0:
+		return TensorTypeQ5_0
+	case fileTypeQ5_1:
+		return TensorTypeQ5_1
+	case fileTypeQ2_K:
+		return TensorTypeQ2_K
+	case fileTypeQ3_K_S:
+		return TensorTypeQ3_K
+	case fileTypeQ3_K_M:
+		return TensorTypeQ3_K
+	case fileTypeQ3_K_L:
+		return TensorTypeQ3_K
+	case FileTypeQ4_K_S:
+		return TensorTypeQ4_K
+	case FileTypeQ4_K_M:
+		return TensorTypeQ4_K
+	case fileTypeQ5_K_S:
+		return TensorTypeQ5_K
+	case fileTypeQ5_K_M:
+		return TensorTypeQ5_K
+	case fileTypeQ6_K:
+		return TensorTypeQ6_K
+	case fileTypeQ2_K_S:
+		return TensorTypeQ2_K
+	case FileTypeBF16:
+		return TensorTypeBF16
+	default:
+		slog.Warn("unsupported file type", "type", ftype)
+		return 0 // F32
+	}
+}
+
+// TensorType is equivalent to ggml_type for individual tensor types
+// Note: these are not the same as FileType
+type TensorType uint32
+
+const (
+	TensorTypeF32 TensorType = iota
+	TensorTypeF16
+	TensorTypeQ4_0
+	TensorTypeQ4_1
+	tensorTypeQ4_2 // unused by GGML
+	tensorTypeQ4_3 // unused by GGML
+	TensorTypeQ5_0
+	TensorTypeQ5_1
+	TensorTypeQ8_0
+	TensorTypeQ8_1
+	TensorTypeQ2_K
+	TensorTypeQ3_K
+	TensorTypeQ4_K
+	TensorTypeQ5_K
+	TensorTypeQ6_K
+	TensorTypeQ8_K
+	tensorTypeIQ2_XXS // not supported by ollama
+	tensorTypeIQ2_XS  // not supported by ollama
+	tensorTypeIQ3_XXS // not supported by ollama
+	tensorTypeIQ1_S   // not supported by ollama
+	tensorTypeIQ4_NL  // not supported by ollama
+	tensorTypeIQ3_S   // not supported by ollama
+	tensorTypeIQ2_S   // not supported by ollama
+	tensorTypeIQ4_XS  // not supported by ollama
+	TensorTypeI8
+	TensorTypeI16
+	TensorTypeI32
+	TensorTypeI64
+	TensorTypeF64
+	tensorTypeIQ1_M // not supported by ollama
+	TensorTypeBF16
+	tensorTypeQ4_0_4_4   // unused by GGML
+	tensorTypeQ4_0_4_8   // unused by GGML
+	tensorTypeQ4_0_8_8   // unused by GGML
+	tensorTypeTQ1_0      // not supported by ollama
+	tensorTypeTQ2_0      // not supported by ollama
+	tensorTypeIQ4_NL_4_4 // unused by GGML
+	tensorTypeIQ4_NL_4_8 // unused by GGML
+	tensorTypeIQ4_NL_8_8 // unused by GGML
+)
+
+// ParseFileType parses the provided GGUF file type
+// Only Ollama supported types are considered valid
+func ParseTensorType(s string) (TensorType, error) {
+	switch s {
+	case "F32":
+		return TensorTypeF32, nil
+	case "F16":
+		return TensorTypeF16, nil
+	case "Q4_0":
+		return TensorTypeQ4_0, nil
+	case "Q4_1":
+		return TensorTypeQ4_1, nil
+	case "Q5_0":
+		return TensorTypeQ5_0, nil
+	case "Q5_1":
+		return TensorTypeQ5_1, nil
+	case "Q8_0":
+		return TensorTypeQ8_0, nil
+	case "Q8_1":
+		return TensorTypeQ8_1, nil
+	case "Q2_K":
+		return TensorTypeQ2_K, nil
+	case "Q3_K":
+		return TensorTypeQ3_K, nil
+	case "Q4_K":
+		return TensorTypeQ4_K, nil
+	case "Q5_K":
+		return TensorTypeQ5_K, nil
+	case "Q6_K":
+		return TensorTypeQ6_K, nil
+	case "Q8_K":
+		return TensorTypeQ8_K, nil
+	case "F64":
+		return TensorTypeF64, nil
+	case "BF16":
+		return TensorTypeBF16, nil
+	default:
+		return 0, fmt.Errorf("unsupported quantization type %s", s)
+	}
+}
+
+func (t TensorType) IsQuantized() bool {
+	switch t {
+	case TensorTypeF32, TensorTypeF16, TensorTypeBF16:
+		return false
+	default:
+		return true
+	}
+}
+
+func (t TensorType) RowSize(ne uint64) uint64 {
+	return t.TypeSize() * ne / t.BlockSize()
+}
+
+func (t TensorType) String() string {
+	switch t {
+	case TensorTypeF32:
+		return "F32"
+	case TensorTypeF16:
+		return "F16"
+	case TensorTypeQ4_0:
+		return "Q4_0"
+	case TensorTypeQ4_1:
+		return "Q4_1"
+	case TensorTypeQ5_0:
+		return "Q5_0"
+	case TensorTypeQ5_1:
+		return "Q5_1"
+	case TensorTypeQ8_0:
+		return "Q8_0"
+	case TensorTypeQ8_1:
+		return "Q8_1"
+	case TensorTypeQ2_K:
+		return "Q2_K"
+	case TensorTypeQ3_K:
+		return "Q3_K"
+	case TensorTypeQ4_K:
+		return "Q4_K"
+	case TensorTypeQ5_K:
+		return "Q5_K"
+	case TensorTypeQ6_K:
+		return "Q6_K"
+	case TensorTypeQ8_K:
+		return "Q8_K"
+	case TensorTypeF64:
+		return "F64"
+	case TensorTypeBF16:
+		return "BF16"
+	default:
+		return "unknown"
+	}
+}

integration/llm_image_test.go

@@ -19,7 +19,7 @@ func TestVisionModels(t *testing.T) {
 	}
 	testCases := []testCase{
 		{
-			model: "llava:7b",
+			model: "qwen2.5vl",
 		},
 		{
 			model: "llama3.2-vision",
@@ -60,6 +60,7 @@ func TestVisionModels(t *testing.T) {
 }
 
 func TestIntegrationSplitBatch(t *testing.T) {
+	skipUnderMinVRAM(t, 6)
 	image, err := base64.StdEncoding.DecodeString(imageEncoding)
 	require.NoError(t, err)
 	req := api.GenerateRequest{

integration/model_arch_test.go

@@ -48,17 +48,6 @@ var (
 	}
 )
 
-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)

integration/quantization_test.go

@@ -0,0 +1,130 @@
+//go:build integration && models
+
+package integration
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"log/slog"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestQuantization(t *testing.T) {
+	sourceModels := []string{
+		"qwen2.5:0.5b-instruct-fp16",
+	}
+	quantizations := []string{
+		"Q8_0",
+		"Q4_K_S",
+		"Q4_K_M",
+		"Q4_K",
+	}
+	softTimeout, hardTimeout := getTimeouts(t)
+	started := time.Now()
+	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
+	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	for _, base := range sourceModels {
+		if err := PullIfMissing(ctx, client, base); err != nil {
+			t.Fatalf("pull failed %s", err)
+		}
+		for _, quant := range quantizations {
+			newName := fmt.Sprintf("%s__%s", base, quant)
+			t.Run(newName, func(t *testing.T) {
+				if time.Now().Sub(started) > softTimeout {
+					t.Skip("skipping remaining tests to avoid excessive runtime")
+				}
+				req := &api.CreateRequest{
+					Model:        newName,
+					Quantization: quant,
+					From:         base,
+				}
+				fn := func(resp api.ProgressResponse) error {
+					// fmt.Print(".")
+					return nil
+				}
+				t.Logf("quantizing: %s -> %s", base, quant)
+				if err := client.Create(ctx, req, fn); err != nil {
+					t.Fatalf("create failed %s", err)
+				}
+				defer func() {
+					req := &api.DeleteRequest{
+						Model: newName,
+					}
+					t.Logf("deleting: %s -> %s", base, quant)
+					if err := client.Delete(ctx, req); err != nil {
+						t.Logf("failed to clean up %s: %s", req.Model, err)
+					}
+				}()
+				// Check metadata on the model
+				resp, err := client.Show(ctx, &api.ShowRequest{Name: newName})
+				if err != nil {
+					t.Fatalf("unable to show model: %s", err)
+				}
+				if !strings.Contains(resp.Details.QuantizationLevel, quant) {
+					t.Fatalf("unexpected quantization for %s:\ngot: %s", newName, resp.Details.QuantizationLevel)
+				}
+
+				stream := true
+				genReq := api.GenerateRequest{
+					Model:     newName,
+					Prompt:    "why is the sky blue?",
+					KeepAlive: &api.Duration{Duration: 3 * time.Second},
+					Options: map[string]any{
+						"seed":        42,
+						"temperature": 0.0,
+					},
+					Stream: &stream,
+				}
+				t.Logf("verifying: %s -> %s", base, quant)
+
+				// Some smaller quantizations can cause models to have poor quality
+				// or get stuck in repetition loops, so we stop as soon as we have any matches
+				anyResp := []string{"rayleigh", "scattering", "day", "sun", "moon", "color", "nitrogen", "oxygen"}
+				reqCtx, reqCancel := context.WithCancel(ctx)
+				atLeastOne := false
+				var buf bytes.Buffer
+				genfn := func(response api.GenerateResponse) error {
+					buf.Write([]byte(response.Response))
+					fullResp := strings.ToLower(buf.String())
+					for _, resp := range anyResp {
+						if strings.Contains(fullResp, resp) {
+							atLeastOne = true
+							t.Log(fullResp)
+							reqCancel()
+							break
+						}
+					}
+					return nil
+				}
+
+				done := make(chan int)
+				var genErr error
+				go func() {
+					genErr = client.Generate(reqCtx, &genReq, genfn)
+					done <- 0
+				}()
+
+				select {
+				case <-done:
+					if genErr != nil && !atLeastOne {
+						t.Fatalf("failed with %s request prompt %s ", genReq.Model, genReq.Prompt)
+					}
+				case <-ctx.Done():
+					t.Error("outer test context done while waiting for generate")
+				}
+
+				t.Logf("passed")
+
+			})
+		}
+	}
+}

integration/utils_test.go

@@ -217,6 +217,7 @@ func InitServerConnection(ctx context.Context, t *testing.T) (*api.Client, strin
 					slog.Error("failed to open server log", "logfile", lifecycle.ServerLogFile, "error", err)
 					return
 				}
+				defer fp.Close()
 				data, err := io.ReadAll(fp)
 				if err != nil {
 					slog.Error("failed to read server log", "logfile", lifecycle.ServerLogFile, "error", err)
@@ -358,3 +359,14 @@ func skipUnderMinVRAM(t *testing.T, gb uint64) {
 		}
 	}
 }
+
+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))
+}

kvcache/causal.go

@@ -30,6 +30,11 @@ type Causal struct {
 
 	// ** current forward pass **
 
+	// curReserve indicates that this forward pass is only for
+	// memory reservation and we should not update our metadata
+	// based on it.
+	curReserve bool
+
 	// the active layer for Get and Put
 	curLayer int
 
@@ -159,12 +164,13 @@ func (c *Causal) Close() {
 }
 
 func (c *Causal) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
+	c.curReserve = reserve
 	c.curBatchSize = len(batch.Positions)
 	c.curSequences = batch.Sequences
 	c.curPositions = batch.Positions
 	c.opts.Except = nil
 
-	if !reserve {
+	if !c.curReserve {
 		c.updateSlidingWindow()
 
 		var err error
@@ -211,10 +217,9 @@ func (c *Causal) StartForward(ctx ml.Context, batch input.Batch, reserve bool) e
 		c.curCellRange.max = len(c.cells) - 1
 	}
 
-	var err error
-	c.curMask, err = c.buildMask(ctx)
+	c.curMask = c.buildMask(ctx)
 
-	return err
+	return nil
 }
 
 func newRange() cellRange {
@@ -239,7 +244,7 @@ func (c *Causal) findStartLoc() (int, error) {
 		}
 	}
 
-	return 0, fmt.Errorf("%w (length: %v)", ErrKvCacheFull, len(c.cells))
+	return 0, fmt.Errorf("%w (cache: %v batch: %v)", ErrKvCacheFull, len(c.cells), c.curBatchSize)
 }
 
 func (c *Causal) updateSlidingWindow() {
@@ -297,7 +302,7 @@ func roundUp(length, pad int) int {
 // Builds a mask of history x batch indicating whether for each token in the batch the
 // token in the history should apply. This is based on both the sequence and causality (the
 // position of the history is not ahead of the token in the batch).
-func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
+func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
 	// Align and pad the two dimensions as required by the backend
 	batchSize := roundUp(c.curBatchSize, c.config.MaskBatchPadding)
 
@@ -305,6 +310,11 @@ func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
 	c.curCellRange.max = roundUp(c.curCellRange.max+1, c.config.CachePadding) - 1
 
 	length := c.curCellRange.max - c.curCellRange.min + 1
+
+	if c.curReserve {
+		return ctx.Input().Empty(c.config.MaskDType, length, batchSize)
+	}
+
 	mask := make([]float32, batchSize*length)
 
 	for i := range c.curBatchSize {
@@ -325,10 +335,7 @@ func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
 		mask[i] = float32(math.Inf(-1))
 	}
 
-	maskTensor, err := ctx.Input().FromFloatSlice(mask, length, batchSize)
-	if err != nil {
-		return nil, err
-	}
+	maskTensor := ctx.Input().FromFloatSlice(mask, length, batchSize)
 
 	if c.config.MaskDType != ml.DTypeF32 {
 		out := ctx.Input().Empty(c.config.MaskDType, maskTensor.Shape()...)
@@ -336,7 +343,7 @@ func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
 		maskTensor = out
 	}
 
-	return maskTensor, nil
+	return maskTensor
 }
 
 func (c *Causal) moveCells(ctx ml.Context, src, dst, length int) {
@@ -491,12 +498,7 @@ func (c *Causal) SetCausal(ctx ml.Context, opts CausalOptions) {
 	if !slices.Equal(c.opts.Except, opts.Except) {
 		c.opts = opts
 		if ctx != nil {
-			var err error
-			c.curMask, err = c.buildMask(ctx)
-			if err != nil {
-				// This error should never occur because we have previously built a mask with the same shape
-				panic(fmt.Errorf("SetCausal: %w", err))
-			}
+			c.curMask = c.buildMask(ctx)
 		}
 	}
 }
@@ -652,10 +654,7 @@ func (c *Causal) shift(seq int, beginIndex, offset int32) error {
 		}
 	}
 
-	kShift, err := ctx.Input().FromIntSlice(offsets, len(offsets))
-	if err != nil {
-		return err
-	}
+	kShift := ctx.Input().FromIntSlice(offsets, len(offsets))
 
 	for i, key := range c.keys {
 		if key == nil {

kvcache/causal_test.go

@@ -344,7 +344,7 @@ func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase)
 			}
 
 			cache.SetLayer(0)
-			tensor, _ := context.FromFloatSlice(test.in, test.inShape...)
+			tensor := context.FromFloatSlice(test.in, test.inShape...)
 			cache.Put(context, tensor, tensor)
 
 			out, _, mask := cache.Get(context)
@@ -386,7 +386,7 @@ func TestCanResume(t *testing.T) {
 	}
 
 	cache.SetLayer(0)
-	tensor, _ := context.FromFloatSlice([]float32{1, 2, 3, 4}, 1, 1, 4)
+	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
@@ -413,7 +413,7 @@ func TestCanResume(t *testing.T) {
 	}
 
 	cache.SetLayer(0)
-	tensor, _ = context.FromFloatSlice([]float32{5, 6}, 1, 1, 2)
+	tensor = context.FromFloatSlice([]float32{5, 6}, 1, 1, 2)
 	cache.Put(context, tensor, tensor)
 
 	// only the latest position has overlapping windows
@@ -470,24 +470,24 @@ func (c *testContext) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
 	return c.Empty(dtype, shape...)
 }
 
-func (c *testContext) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
+func (c *testContext) FromFloatSlice(s []float32, shape ...int) ml.Tensor {
 	t := c.Empty(ml.DTypeF32, shape...).(*testTensor)
 
 	copy(t.data, s)
 
-	return t, nil
+	return t
 }
 
-func (c *testContext) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
+func (c *testContext) FromIntSlice(s []int32, shape ...int) ml.Tensor {
 	f := make([]float32, len(s))
 	for i := range f {
 		f[i] = float32(s[i])
 	}
 
-	out, _ := c.FromFloatSlice(f, shape...)
+	out := c.FromFloatSlice(f, shape...)
 	out.(*testTensor).dtype = ml.DTypeI32
 
-	return out, nil
+	return out
 }
 
 func (c *testContext) Arange(start, stop, step float32, dtype ml.DType) ml.Tensor {
@@ -496,7 +496,7 @@ func (c *testContext) Arange(start, stop, step float32, dtype ml.DType) ml.Tenso
 		s = append(s, i)
 	}
 
-	out, _ := c.FromFloatSlice(s, len(s))
+	out := c.FromFloatSlice(s, len(s))
 	out.(*testTensor).dtype = dtype
 	return out
 }
@@ -508,7 +508,7 @@ 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) Reserve() {}
 
 func (c *testContext) MaxGraphNodes() int {
 	return 10

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "2016f07bd106c73699ecbaace80f55db5ed95dac";
+char const *LLAMA_COMMIT = "1caae7fc6c77551cb1066515e0f414713eebb367";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/.rsync-filter

@@ -1,23 +1,39 @@
 protect **/*.go
 include common/
+include common/arg.*
+include common/chat.*
+include common/chat-parser.*
+include common/console.*
 include common/base64.*
 include common/common.*
 include common/json-schema-to-grammar.*
-include common/json.*
+include common/json-partial.*
+include common/regex-partial.*
 include common/log.*
 include common/sampling.*
 include common/stb_image.*
 include include/
 include include/llama.*
 include include/llama-*.*
-include examples/
-include examples/llava/
-include examples/llava/clip.*
-include examples/llava/clip-impl.*
-include examples/llava/llava.*
+include tools/
+include tools/mtmd/
+include tools/mtmd/mtmd.*
+include tools/mtmd/mtmd-helper.*
+include tools/mtmd/mtmd-audio.*
+include tools/mtmd/clip.*
+include tools/mtmd/clip-impl.*
 include src/
 include src/llama.*
 include src/llama-*.*
 include src/unicode-data.*
 include src/unicode.*
+include vendor/
+include vendor/nlohmann
+include vendor/nlohmann/*
+include vendor/miniaudio
+include vendor/miniaudio/*
+include vendor/stb
+include vendor/stb/stb_image.*
+include vendor/minja
+include vendor/minja/*
 exclude *

llama/llama.cpp/common/arg.h

@@ -0,0 +1,89 @@
+#pragma once
+
+#include "common.h"
+
+#include <set>
+#include <string>
+#include <vector>
+
+//
+// CLI argument parsing
+//
+
+struct common_arg {
+    std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
+    std::set<enum llama_example> excludes = {};
+    std::vector<const char *> args;
+    const char * value_hint   = nullptr; // help text or example for arg value
+    const char * value_hint_2 = nullptr; // for second arg value
+    const char * env          = nullptr;
+    std::string help;
+    bool is_sparam = false; // is current arg a sampling param?
+    void (*handler_void)   (common_params & params) = nullptr;
+    void (*handler_string) (common_params & params, const std::string &) = nullptr;
+    void (*handler_str_str)(common_params & params, const std::string &, const std::string &) = nullptr;
+    void (*handler_int)    (common_params & params, int) = nullptr;
+
+    common_arg(
+        const std::initializer_list<const char *> & args,
+        const char * value_hint,
+        const std::string & help,
+        void (*handler)(common_params & params, const std::string &)
+    ) : args(args), value_hint(value_hint), help(help), handler_string(handler) {}
+
+    common_arg(
+        const std::initializer_list<const char *> & args,
+        const char * value_hint,
+        const std::string & help,
+        void (*handler)(common_params & params, int)
+    ) : args(args), value_hint(value_hint), help(help), handler_int(handler) {}
+
+    common_arg(
+        const std::initializer_list<const char *> & args,
+        const std::string & help,
+        void (*handler)(common_params & params)
+    ) : args(args), help(help), handler_void(handler) {}
+
+    // support 2 values for arg
+    common_arg(
+        const std::initializer_list<const char *> & args,
+        const char * value_hint,
+        const char * value_hint_2,
+        const std::string & help,
+        void (*handler)(common_params & params, const std::string &, const std::string &)
+    ) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
+
+    common_arg & set_examples(std::initializer_list<enum llama_example> examples);
+    common_arg & set_excludes(std::initializer_list<enum llama_example> excludes);
+    common_arg & set_env(const char * env);
+    common_arg & set_sparam();
+    bool in_example(enum llama_example ex);
+    bool is_exclude(enum llama_example ex);
+    bool get_value_from_env(std::string & output);
+    bool has_value_from_env();
+    std::string to_string();
+};
+
+struct common_params_context {
+    enum llama_example ex = LLAMA_EXAMPLE_COMMON;
+    common_params & params;
+    std::vector<common_arg> options;
+    void(*print_usage)(int, char **) = nullptr;
+    common_params_context(common_params & params) : params(params) {}
+};
+
+// parse input arguments from CLI
+// if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
+bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
+
+// function to be used by test-arg-parser
+common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
+bool common_has_curl();
+
+struct common_remote_params {
+    std::vector<std::string> headers;
+    long timeout = 0; // CURLOPT_TIMEOUT, in seconds ; 0 means no timeout
+    long max_size = 0; // max size of the response ; unlimited if 0 ; max is 2GB
+};
+// get remote file content, returns <http_code, raw_response_body>
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params);

llama/llama.cpp/common/chat-parser.cpp

@@ -0,0 +1,380 @@
+#include "chat-parser.h"
+#include "common.h"
+#include "log.h"
+#include "regex-partial.h"
+
+#include <optional>
+#include <stdexcept>
+#include <string>
+#include <vector>
+
+using json = nlohmann::ordered_json;
+
+common_chat_msg_parser::common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax)
+    : input_(input), is_partial_(is_partial), syntax_(syntax)
+{
+    result_.role = "assistant";
+
+    while (true) {
+        std::string id = std::to_string(std::rand());
+        if (input.find(id) == std::string::npos) {
+            healing_marker_ = id;
+            break;
+        }
+    }
+}
+
+std::string common_chat_msg_parser::str(const common_string_range & rng) const {
+    GGML_ASSERT(rng.begin <= rng.end);
+    return input_.substr(rng.begin, rng.end - rng.begin);
+}
+
+void common_chat_msg_parser::add_content(const std::string &content) {
+    result_.content += content;
+}
+
+void common_chat_msg_parser::add_reasoning_content(const std::string &reasoning_content) {
+    result_.reasoning_content += reasoning_content;
+}
+
+bool common_chat_msg_parser::add_tool_call(const std::string & name, const std::string & id, const std::string & arguments) {
+    if (name.empty()) {
+        return false;
+    }
+
+    common_chat_tool_call tool_call;
+    tool_call.name = name;
+    tool_call.arguments = arguments;
+    tool_call.id = id;
+
+    // LOG_DBG("Tool call arguments:\n\traw: %s\n\tresult: %s\n", arguments.c_str(), tool_call.arguments.c_str());
+    result_.tool_calls.emplace_back(tool_call);
+    return true;
+}
+bool common_chat_msg_parser::add_tool_call(const json & tool_call) {
+    std::string name = tool_call.contains("name") ? tool_call.at("name") : "";
+    std::string id = tool_call.contains("id") ? tool_call.at("id") : "";
+    std::string arguments = tool_call.contains("arguments") ? tool_call.at("arguments") : "";
+    return add_tool_call(name, id, arguments);
+}
+
+bool common_chat_msg_parser::add_tool_calls(const json & arr) {
+    for (const auto & item : arr) {
+        if (!add_tool_call(item)) {
+            return false;
+        }
+    }
+    return true;
+}
+void common_chat_msg_parser::finish() {
+    if (!is_partial_ && pos_ != input_.size()) {
+        throw std::runtime_error("Unexpected content at end of input");// + input_.substr(pos_));
+    }
+}
+
+bool common_chat_msg_parser::consume_spaces() {
+    const auto length = input_.size();
+    auto consumed = false;
+    while (pos_ < length && std::isspace(input_[pos_])) {
+        ++pos_;
+        consumed = true;
+    }
+    return consumed;
+}
+
+bool common_chat_msg_parser::try_consume_literal(const std::string & literal) {
+    auto pos = pos_;
+    for (auto i = 0u; i < literal.size(); ++i) {
+        if (pos >= input_.size()) {
+            return false;
+        }
+        if (input_[pos] != literal[i]) {
+            return false;
+        }
+        ++pos;
+    }
+    pos_ = pos;
+    return true;
+}
+
+std::optional<common_chat_msg_parser::find_regex_result>  common_chat_msg_parser::try_find_literal(const std::string & literal) {
+    auto idx = input_.find(literal, pos_);
+    if (idx != std::string::npos) {
+        find_regex_result res;
+        res.prelude = input_.substr(pos_, idx - pos_);
+        auto end = idx + literal.size();
+        res.groups.emplace_back(common_string_range{idx, end});
+        move_to(end);
+        return res;
+    }
+    if (is_partial_) {
+        idx = string_find_partial_stop(input_, literal);
+        if (idx != std::string::npos && idx >= pos_) {
+            find_regex_result res;
+            res.prelude = input_.substr(pos_, idx - pos_);
+            auto end = input_.size();
+            res.groups.emplace_back(common_string_range{idx, end});
+            move_to(end);
+            return res;
+        }
+    }
+    return std::nullopt;
+}
+
+void common_chat_msg_parser::consume_literal(const std::string & literal) {
+    if (!try_consume_literal(literal)) {
+        throw common_chat_msg_partial_exception(literal);
+    }
+}
+
+bool common_chat_msg_parser::try_parse_reasoning(const std::string & start_think, const std::string & end_think) {
+    auto handle_reasoning = [&](const std::string & reasoning, bool closed) {
+        auto stripped_reasoning = string_strip(reasoning);
+        if (stripped_reasoning.empty()) {
+            return;
+        }
+        if (syntax_.reasoning_in_content) {
+            add_content(syntax_.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK ? "<think>" : start_think);
+            add_content(stripped_reasoning);
+            if (closed) {
+                add_content(syntax_.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK ? "</think>" : end_think);
+            }
+        } else {
+            add_reasoning_content(stripped_reasoning);
+        }
+    };
+    if (syntax_.reasoning_format != COMMON_REASONING_FORMAT_NONE) {
+        if (syntax_.thinking_forced_open || try_consume_literal(start_think)) {
+            if (auto res = try_find_literal(end_think)) {
+                handle_reasoning(res->prelude, /* closed */ true);
+                consume_spaces();
+                return true;
+            }
+            auto rest = consume_rest();
+            if (!rest.empty()) {
+                handle_reasoning(rest, /* closed */ !is_partial());
+            }
+            // Allow unclosed thinking tags, for now (https://github.com/ggml-org/llama.cpp/issues/13812, https://github.com/ggml-org/llama.cpp/issues/13877)
+            // if (!syntax_.thinking_forced_open) {
+            //     throw common_chat_msg_partial_exception(end_think);
+            // }
+            return true;
+        }
+    }
+    return false;
+}
+
+std::string common_chat_msg_parser::consume_rest() {
+    auto rest = input_.substr(pos_);
+    pos_ = input_.size();
+    return rest;
+}
+
+// Tries to find the regex, consumes it (pos right after it) and gives the prelude (right before it) and the groups to the callback.
+std::optional<common_chat_msg_parser::find_regex_result> common_chat_msg_parser::try_find_regex(const common_regex & regex, size_t from, bool add_prelude_to_content) {
+    auto m = regex.search(input_, from == std::string::npos ? pos_ : from);
+    if (m.type == COMMON_REGEX_MATCH_TYPE_NONE) {
+        return std::nullopt;
+    }
+    auto prelude = input_.substr(pos_, m.groups[0].begin - pos_);
+    pos_ = m.groups[0].end;
+
+    if (add_prelude_to_content) {
+        add_content(prelude);
+    }
+    if (m.type == COMMON_REGEX_MATCH_TYPE_PARTIAL) {
+        if (is_partial()) {
+            throw common_chat_msg_partial_exception(regex.str());
+        }
+        return std::nullopt;
+    }
+    return find_regex_result{prelude, m.groups};
+}
+
+common_chat_msg_parser::find_regex_result common_chat_msg_parser::consume_regex(const common_regex & regex) {
+    if (auto result = try_consume_regex(regex)) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception(regex.str());
+}
+
+std::optional<common_chat_msg_parser::find_regex_result> common_chat_msg_parser::try_consume_regex(const common_regex & regex) {
+    auto m = regex.search(input_, pos_);
+    if (m.type == COMMON_REGEX_MATCH_TYPE_NONE) {
+        return std::nullopt;
+    }
+    if (m.type == COMMON_REGEX_MATCH_TYPE_PARTIAL) {
+        if (is_partial()) {
+            throw common_chat_msg_partial_exception(regex.str());
+        }
+        return std::nullopt;
+    }
+    if (m.groups[0].begin != pos_) {
+        // Didn't match at the current position.
+        return std::nullopt;
+    }
+    pos_ = m.groups[0].end;
+
+    return find_regex_result {
+        /* .prelude = */ "",
+        m.groups,
+    };
+}
+
+std::optional<common_json> common_chat_msg_parser::try_consume_json() {
+    auto it = input_.cbegin() + pos_;
+    const auto end = input_.cend();
+    common_json result;
+    if (!common_json_parse(it, end, healing_marker_, result)) {
+        return std::nullopt;
+    }
+    pos_ = std::distance(input_.cbegin(), it);
+    if (result.healing_marker.marker.empty()) {
+        // No healing marker, just return the parsed json
+        return result;
+    }
+    if (!is_partial()) {
+        throw common_chat_msg_partial_exception("JSON");
+    }
+    return result;
+}
+
+common_json common_chat_msg_parser::consume_json() {
+    if (auto result = try_consume_json()) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception("JSON");
+}
+
+common_chat_msg_parser::consume_json_result common_chat_msg_parser::consume_json_with_dumped_args(
+    const std::vector<std::vector<std::string>> & args_paths,
+    const std::vector<std::vector<std::string>> & content_paths
+) {
+    if (auto result = try_consume_json_with_dumped_args(args_paths, content_paths)) {
+        return *result;
+    }
+    throw common_chat_msg_partial_exception("JSON");
+}
+
+std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parser::try_consume_json_with_dumped_args(
+    const std::vector<std::vector<std::string>> & args_paths,
+    const std::vector<std::vector<std::string>> & content_paths
+) {
+    auto partial = try_consume_json();
+    if (!partial) {
+        return std::nullopt;
+    }
+    auto is_arguments_path = [&](const std::vector<std::string> & path) {
+        return std::find(args_paths.begin(), args_paths.end(), path) != args_paths.end();
+    };
+    auto is_content_path = [&](const std::vector<std::string> & path) {
+        return std::find(content_paths.begin(), content_paths.end(), path) != content_paths.end();
+    };
+
+    if (partial->healing_marker.marker.empty()) {
+        if (args_paths.empty()) {
+            // No arguments to dump, and JSON was parsed fully.
+            return consume_json_result {
+                partial->json,
+                /* .is_partial = */ false,
+            };
+        }
+        if (is_arguments_path({})) {
+            // Entire JSON is the arguments and was parsed fully.
+            return consume_json_result {
+                partial->json.dump(),
+                /* .is_partial = */ false,
+            };
+        }
+    }
+
+    LOG_DBG("Parsed partial JSON: %s (json_healing_marker: %s)\n", partial->json.dump().c_str(), partial->healing_marker.json_dump_marker.c_str());
+
+    auto found_healing_marker = false;
+    std::vector<std::string> path;
+    std::function<json(const json &)> remove_unsupported_healings_and_dump_args = [&](const json & j) -> json {
+        if (is_arguments_path(path)) {
+            auto arguments = j.dump();
+            if (is_partial() && !partial->healing_marker.marker.empty()) {
+                auto idx = arguments.find(partial->healing_marker.json_dump_marker);
+                if (idx != std::string::npos) {
+                    arguments.resize(idx);
+                    found_healing_marker = true;
+                }
+                if (arguments == "\"") {
+                    // This happens because of completing `:"$magic` after `"arguments"`
+                    arguments = "";
+                }
+            }
+            return arguments;
+        }
+        if (is_content_path(path)) {
+            if (!j.is_string()) {
+                throw std::runtime_error("Content path must be a string");
+            }
+            std::string str = j;
+            auto idx = str.find(partial->healing_marker.marker); // not using json_dump_marker as we're inside a string
+            if (idx != std::string::npos) {
+                str.resize(idx);
+                found_healing_marker = true;
+            }
+            return str;
+        }
+        if (j.is_object()) {
+            auto obj = json::object();
+            for (const auto & p : j.items()) {
+                const auto & key = p.key();
+                const auto & value = p.value();
+                const std::string key_str = key; // NOLINT
+                auto idx = key_str.find(healing_marker_);
+                if (idx != std::string::npos) {
+                    found_healing_marker = true;
+                    break;
+                }
+                path.push_back(key_str);
+                if (value.is_string()) {
+                    const std::string value_str = value;
+                    if (value_str.find(healing_marker_) != std::string::npos) {
+                        found_healing_marker = true;
+                        if (is_content_path(path)) {
+                            if (partial->healing_marker.marker == partial->healing_marker.json_dump_marker) {
+                                // The healing occurred inside the string: good. Otherwise we just ditch the entire key/value pair.
+                                obj[key] = remove_unsupported_healings_and_dump_args(value);
+                            }
+                        }
+                        break;
+                    }
+                    obj[key] = value;
+                } else {
+                    obj[key] = remove_unsupported_healings_and_dump_args(value);
+                }
+                path.pop_back();
+            }
+            return obj;
+        }
+        if (j.is_array()) {
+            auto arr = json::array();
+            for (const auto & value : j) {
+                if (value.is_string()) {
+                    std::string str = value;
+                    auto idx = str.find(healing_marker_);
+                    if (idx != std::string::npos) {
+                        // Don't heal array values that aren't in the arguments.
+                        found_healing_marker = true;
+                        break;
+                    }
+                }
+                arr.push_back(remove_unsupported_healings_and_dump_args(value));
+            }
+            return arr;
+        }
+        return j;
+    };
+
+    auto cleaned = remove_unsupported_healings_and_dump_args(partial->json);
+    LOG_DBG("Cleaned up JSON %s to %s (json_healing_marker : '%s')\n", partial->json.dump().c_str(), cleaned.dump().c_str(), partial->healing_marker.json_dump_marker.c_str());
+    return consume_json_result {
+        cleaned,
+        /* .is_partial = */ found_healing_marker,
+    };
+}

llama/llama.cpp/common/chat-parser.h

@@ -0,0 +1,118 @@
+#pragma once
+
+#include "chat.h"
+#include "json-partial.h"
+#include "regex-partial.h"
+
+#include <nlohmann/json.hpp>
+
+#include <optional>
+#include <string>
+#include <vector>
+
+class common_chat_msg_partial_exception : public std::runtime_error {
+  public:
+    common_chat_msg_partial_exception(const std::string & message) : std::runtime_error(message) {}
+};
+
+class common_chat_msg_parser {
+    std::string input_;
+    bool is_partial_;
+    common_chat_syntax syntax_;
+    std::string healing_marker_;
+
+    size_t pos_ = 0;
+    common_chat_msg result_;
+
+  public:
+    common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
+    const std::string & input() const { return input_; }
+    size_t pos() const { return pos_; }
+    const std::string & healing_marker() const { return healing_marker_; }
+    const bool & is_partial() const { return is_partial_; }
+    const common_chat_msg & result() const { return result_; }
+    const common_chat_syntax & syntax() const { return syntax_; }
+
+    void move_to(size_t pos) {
+        if (pos > input_.size()) {
+            throw std::runtime_error("Invalid position!");
+        }
+        pos_ = pos;
+    }
+    void move_back(size_t n) {
+        if (pos_ < n) {
+            throw std::runtime_error("Can't move back that far!");
+        }
+        pos_ -= n;
+    }
+
+    // Get the substring of the input at the given range
+    std::string str(const common_string_range & rng) const;
+
+    // Appends to the result.content field
+    void add_content(const std::string & content);
+
+    // Appends to the result.reasoning_content field
+    void add_reasoning_content(const std::string & reasoning_content);
+
+    // Adds a tool call to the result. If the tool call is too incomplete (e.g. name empty), it won't add anything.
+    bool add_tool_call(const std::string & name, const std::string & id, const std::string & arguments);
+
+    // Adds a tool call using the "name", "id" and "arguments" fields of the json object
+    bool add_tool_call(const nlohmann::ordered_json & tool_call);
+
+    // Adds an array of tool calls using their "name", "id" and "arguments" fields.
+    bool add_tool_calls(const nlohmann::ordered_json & arr);
+
+    void finish();
+
+    bool consume_spaces();
+
+    void consume_literal(const std::string & literal);
+
+    bool try_parse_reasoning(const std::string & start_think, const std::string & end_think);
+
+    std::string consume_rest();
+
+    struct find_regex_result {
+        std::string prelude;
+        std::vector<common_string_range> groups;
+    };
+
+    std::optional<find_regex_result> try_find_regex(const common_regex & regex, size_t from = std::string::npos, bool add_prelude_to_content = true);
+
+    bool try_consume_literal(const std::string & literal);
+
+    std::optional<find_regex_result> try_find_literal(const std::string & literal);
+
+    find_regex_result consume_regex(const common_regex & regex);
+
+    std::optional<find_regex_result> try_consume_regex(const common_regex & regex);
+
+    std::optional<common_json> try_consume_json();
+    common_json consume_json();
+
+    struct consume_json_result {
+        nlohmann::ordered_json value;
+        bool is_partial;
+    };
+
+    /*
+        Consume (possibly partial) json and converts specific subtrees to (possibly truncated) JSON strings.
+
+        By default, object keys can't be truncated, nor can string values (their corresponding key is removed,
+        e.g. `{"foo": "bar", "baz": "b` -> `{"foo": "bar"}`
+
+        But one can allow subpaths to be kept truncated, and possibly json-dumped to truncated json strings
+        - with `content_paths={{"foo"}}` -> `{"foo": "b` -> {"foo": "b"}`
+        - with `args_paths={{"foo"}}` -> `{"foo": {"b` -> `{"foo": "{b"}`
+    */
+    consume_json_result consume_json_with_dumped_args(
+        const std::vector<std::vector<std::string>> & args_paths = {},
+        const std::vector<std::vector<std::string>> & content_paths = {}
+    );
+    std::optional<consume_json_result> try_consume_json_with_dumped_args(
+        const std::vector<std::vector<std::string>> & args_paths = {},
+        const std::vector<std::vector<std::string>> & content_paths = {}
+    );
+};

llama/llama.cpp/common/chat.h

@@ -0,0 +1,202 @@
+// Chat support (incl. tool call grammar constraining & output parsing) w/ generic & custom template handlers.
+
+#pragma once
+
+#include "common.h"
+#include <functional>
+#include <chrono>
+#include <string>
+#include <vector>
+
+struct common_chat_templates;
+
+struct common_chat_tool_call {
+    std::string name;
+    std::string arguments;
+    std::string id;
+
+    bool operator==(const common_chat_tool_call & other) const {
+        return name == other.name && arguments == other.arguments && id == other.id;
+    }
+};
+
+struct common_chat_msg_content_part {
+    std::string type;
+    std::string text;
+
+    bool operator==(const common_chat_msg_content_part & other) const {
+        return type == other.type && text == other.text;
+    }
+};
+
+struct common_chat_msg {
+    std::string role;
+    std::string content;
+    std::vector<common_chat_msg_content_part> content_parts = {};
+    std::vector<common_chat_tool_call> tool_calls = {};
+    std::string reasoning_content;
+    std::string tool_name;
+    std::string tool_call_id;
+
+    template <class T> T to_json_oaicompat() const;
+
+    bool empty() const {
+        return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() && tool_name.empty() && tool_call_id.empty();
+    }
+    void ensure_tool_call_ids_set(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
+        for (auto i = 0u; i < tool_calls.size(); i++) {
+            if (ids_cache.size() <= i) {
+                auto id = tool_calls[i].id;
+                if (id.empty()) {
+                    id = gen_tool_call_id();
+                }
+                ids_cache.push_back(id);
+            }
+            tool_calls[i].id = ids_cache[i];
+        }
+    }
+    bool operator==(const common_chat_msg & other) const {
+        return role == other.role
+            && content == other.content
+            && content_parts == other.content_parts
+            && tool_calls == other.tool_calls
+            && reasoning_content == other.reasoning_content
+            && tool_name == other.tool_name
+            && tool_call_id == other.tool_call_id;
+    }
+    bool operator!=(const common_chat_msg & other) const {
+        return !(*this == other);
+    }
+};
+
+struct common_chat_msg_diff {
+    std::string reasoning_content_delta;
+    std::string content_delta;
+    size_t tool_call_index = std::string::npos;
+    common_chat_tool_call tool_call_delta;
+
+    static std::vector<common_chat_msg_diff> compute_diffs(const common_chat_msg & previous_msg, const common_chat_msg & new_msg);
+
+    bool operator==(const common_chat_msg_diff & other) const {
+        return content_delta == other.content_delta
+        && tool_call_index == other.tool_call_index
+        && tool_call_delta == other.tool_call_delta;
+    }
+};
+
+struct common_chat_tool {
+    std::string name;
+    std::string description;
+    std::string parameters;
+};
+
+enum common_chat_tool_choice {
+    COMMON_CHAT_TOOL_CHOICE_AUTO,
+    COMMON_CHAT_TOOL_CHOICE_REQUIRED,
+    COMMON_CHAT_TOOL_CHOICE_NONE,
+};
+
+enum common_chat_format {
+    COMMON_CHAT_FORMAT_CONTENT_ONLY,
+    COMMON_CHAT_FORMAT_GENERIC,
+    COMMON_CHAT_FORMAT_MISTRAL_NEMO,
+    COMMON_CHAT_FORMAT_LLAMA_3_X,
+    COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
+    COMMON_CHAT_FORMAT_DEEPSEEK_R1,
+    COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
+    COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
+    COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
+    COMMON_CHAT_FORMAT_HERMES_2_PRO,
+    COMMON_CHAT_FORMAT_COMMAND_R7B,
+
+    COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
+};
+
+struct common_chat_templates_inputs {
+    std::vector<common_chat_msg> messages;
+    std::string grammar;
+    std::string json_schema;
+    bool add_generation_prompt = true;
+    bool use_jinja = true;
+    // Parameters below only supported when use_jinja is true
+    std::vector<common_chat_tool> tools;
+    common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
+    bool parallel_tool_calls = false;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    bool enable_thinking = true;
+    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
+};
+
+struct common_chat_params {
+    common_chat_format                  format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    std::string                         prompt;
+    std::string                         grammar;
+    bool                                grammar_lazy = false;
+    bool                                thinking_forced_open = false;
+    std::vector<common_grammar_trigger> grammar_triggers;
+    std::vector<std::string>            preserved_tokens;
+    std::vector<std::string>            additional_stops;
+};
+
+struct common_chat_syntax {
+    common_chat_format       format                = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    common_reasoning_format  reasoning_format      = COMMON_REASONING_FORMAT_NONE;
+    // Whether reasoning_content should be inlined in the content (e.g. for reasoning_format=deepseek in stream mode)
+    bool                     reasoning_in_content  = false;
+    bool                     thinking_forced_open  = false;
+    bool                     parse_tool_calls      = true;
+};
+
+// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
+bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
+
+void common_chat_templates_free(struct common_chat_templates * tmpls);
+
+struct common_chat_templates_deleter { void operator()(common_chat_templates * tmpls) { common_chat_templates_free(tmpls); } };
+
+typedef std::unique_ptr<struct common_chat_templates, common_chat_templates_deleter> common_chat_templates_ptr;
+
+common_chat_templates_ptr common_chat_templates_init(
+                                    const struct llama_model * model,
+                                           const std::string & chat_template_override,
+                                           const std::string & bos_token_override = "",
+                                           const std::string & eos_token_override = "");
+
+bool         common_chat_templates_was_explicit(const struct common_chat_templates * tmpls);
+const char * common_chat_templates_source(const struct common_chat_templates * tmpls, const char * variant = nullptr);
+
+
+struct common_chat_params      common_chat_templates_apply(
+    const struct common_chat_templates * tmpls,
+    const struct common_chat_templates_inputs & inputs);
+
+// Format single message, while taking into account the position of that message in chat history
+std::string common_chat_format_single(
+        const struct common_chat_templates * tmpls,
+        const std::vector<common_chat_msg> & past_msg,
+        const common_chat_msg & new_msg,
+        bool add_ass,
+        bool use_jinja);
+
+// Returns an example of formatted chat
+std::string common_chat_format_example(
+    const struct common_chat_templates * tmpls,
+    bool use_jinja);
+
+const char*               common_chat_format_name(common_chat_format format);
+const char*               common_reasoning_format_name(common_reasoning_format format);
+common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
+
+common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
+
+// Parses a JSON array of messages in OpenAI's chat completion API format.
+// T can be std::string containing JSON or nlohmann::ordered_json
+template <class T> std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const T & messages);
+template <class T> T common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
+
+// Parses a JSON array of tools in OpenAI's chat completion tool call API format.
+// T can be std::string containing JSON or nlohmann::ordered_json
+template <class T> std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const T & tools);
+template <class T> T common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
+
+template <class T> T common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff);

llama/llama.cpp/common/common.cpp

@@ -203,6 +203,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
 
     DWORD p = NORMAL_PRIORITY_CLASS;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      p = BELOW_NORMAL_PRIORITY_CLASS; break;
         case GGML_SCHED_PRIO_NORMAL:   p = NORMAL_PRIORITY_CLASS;       break;
         case GGML_SCHED_PRIO_MEDIUM:   p = ABOVE_NORMAL_PRIORITY_CLASS; break;
         case GGML_SCHED_PRIO_HIGH:     p = HIGH_PRIORITY_CLASS;         break;
@@ -228,6 +229,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
 
     int p = 0;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      p =  5;  break;
         case GGML_SCHED_PRIO_NORMAL:   p =  0;  break;
         case GGML_SCHED_PRIO_MEDIUM:   p = -5;  break;
         case GGML_SCHED_PRIO_HIGH:     p = -10; break;
@@ -443,6 +445,25 @@ void string_replace_all(std::string & s, const std::string & search, const std::
     s = std::move(builder);
 }
 
+bool string_ends_with(const std::string_view & str, const std::string_view & suffix) {
+    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
+}
+size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop) {
+    if (!str.empty() && !stop.empty()) {
+        const char text_last_char = str.back();
+        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
+            if (stop[char_index] == text_last_char) {
+                const auto current_partial = stop.substr(0, char_index + 1);
+                if (string_ends_with(str, current_partial)) {
+                    return str.size() - char_index - 1;
+                }
+            }
+        }
+    }
+
+    return std::string::npos;
+}
+
 std::string regex_escape(const std::string & s) {
     static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
     return std::regex_replace(s, special_chars, "\\$0");
@@ -830,7 +851,7 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else {
@@ -884,13 +905,16 @@ struct common_init_result common_init_from_params(common_params & params) {
             ok = false;
         }
 
-        if (llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, reranking will not work\n", __func__);
-            ok = false;
-        }
+        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
+        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
 
-        if (llama_vocab_sep(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  SEP token, reranking will not work\n", __func__);
+        if (!has_eos && !has_sep) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
+            ok = false;
+        } else if (!has_eos) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
+        } else if (!has_sep) {
+            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
             ok = false;
         }
 
@@ -1083,6 +1107,9 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
     }
 
+    mparams.progress_callback           = params.load_progress_callback;
+    mparams.progress_callback_user_data = params.load_progress_callback_user_data;
+
     return mparams;
 }
 
@@ -1096,7 +1123,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.n_threads         = params.cpuparams.n_threads;
     cparams.n_threads_batch   = params.cpuparams_batch.n_threads == -1 ?
                                 params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
-    cparams.logits_all        = params.logits_all;
     cparams.embeddings        = params.embedding;
     cparams.rope_scaling_type = params.rope_scaling_type;
     cparams.rope_freq_base    = params.rope_freq_base;
@@ -1114,6 +1140,8 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.offload_kqv       = !params.no_kv_offload;
     cparams.flash_attn        = params.flash_attn;
     cparams.no_perf           = params.no_perf;
+    cparams.op_offload        = !params.no_op_offload;
+    cparams.swa_full          = params.swa_full;
 
     if (params.reranking) {
         cparams.embeddings    = true;
@@ -1306,81 +1334,6 @@ std::string common_detokenize(const struct llama_vocab * vocab, const std::vecto
     return text;
 }
 
-//
-// KV cache utils
-//
-
-void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size) {
-    static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
-
-    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
-        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
-
-    llama_kv_cache_view_cell * c_curr = view.cells;
-    llama_seq_id * cs_curr = view.cells_sequences;
-
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        if (i % row_size == 0) {
-            printf("\n%5d: ", i);
-        }
-        int seq_count = 0;
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] >= 0) { seq_count++; }
-        }
-        putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
-    }
-
-    printf("\n=== Done dumping\n");
-}
-
-void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size) {
-    static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
-
-    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
-        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
-
-    std::unordered_map<llama_seq_id, size_t> seqs;
-    llama_kv_cache_view_cell * c_curr = view.cells;
-    llama_seq_id * cs_curr = view.cells_sequences;
-
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] < 0) { continue; }
-            if (seqs.find(cs_curr[j]) == seqs.end()) {
-                if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-                const size_t sz = seqs.size();
-                seqs[cs_curr[j]] = sz;
-            }
-        }
-        if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-    }
-
-    printf("=== Sequence legend: ");
-    for (const auto & it : seqs) {
-        printf("%zu=%d, ", it.second, it.first);
-    }
-    printf("'+'=other sequence ids");
-
-    c_curr = view.cells;
-    cs_curr = view.cells_sequences;
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
-        if (i % row_size == 0) {
-            printf("\n%5d: ", i);
-        }
-        for (int j = 0; j < view.n_seq_max; j++) {
-            if (cs_curr[j] >= 0) {
-                const auto & it = seqs.find(cs_curr[j]);
-                putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
-            } else {
-                putchar('.');
-            }
-        }
-        putchar(' ');
-    }
-
-    printf("\n=== Done dumping\n");
-}
-
 //
 // Embedding utils
 //
@@ -1565,3 +1518,20 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 
     return result;
 }
+
+ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride) {
+    const int64_t ne_datapoint = llama_n_ctx(ctx);
+    const int64_t ndata        = (tokens.size() - ne_datapoint - 1) / stride;
+    ggml_opt_dataset_t result = ggml_opt_dataset_init(
+        GGML_TYPE_I32, GGML_TYPE_I32, ne_datapoint, ne_datapoint, ndata, /*ndata_shard =*/ 1);
+
+    llama_token * data   = (llama_token *) ggml_opt_dataset_data(result)->data;
+    llama_token * labels = (llama_token *) ggml_opt_dataset_labels(result)->data;
+
+    for (int64_t idata = 0; idata < ndata; ++idata) {
+        memcpy(data   + idata*ne_datapoint, tokens.data() + idata*stride + 0, ne_datapoint*sizeof(llama_token));
+        memcpy(labels + idata*ne_datapoint, tokens.data() + idata*stride + 1, ne_datapoint*sizeof(llama_token));
+    }
+
+    return result;
+}

llama/llama.cpp/common/common.go

@@ -1,6 +1,7 @@
 package common
 
-// #cgo CXXFLAGS: -std=c++11
+// #cgo CXXFLAGS: -std=c++17
 // #cgo CPPFLAGS: -I${SRCDIR}/../include
+// #cgo CPPFLAGS: -I${SRCDIR}/../vendor
 // #cgo CPPFLAGS: -I${SRCDIR}/../../../ml/backend/ggml/ggml/include
 import "C"

llama/llama.cpp/common/common.h

@@ -6,6 +6,7 @@
 
 #include <set>
 #include <string>
+#include <string_view>
 #include <vector>
 #include <sstream>
 
@@ -66,7 +67,6 @@ enum llama_example {
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
     LLAMA_EXAMPLE_MAIN,
-    LLAMA_EXAMPLE_INFILL,
     LLAMA_EXAMPLE_EMBEDDING,
     LLAMA_EXAMPLE_PERPLEXITY,
     LLAMA_EXAMPLE_RETRIEVAL,
@@ -76,7 +76,7 @@ enum llama_example {
     LLAMA_EXAMPLE_SERVER,
     LLAMA_EXAMPLE_CVECTOR_GENERATOR,
     LLAMA_EXAMPLE_EXPORT_LORA,
-    LLAMA_EXAMPLE_LLAVA,
+    LLAMA_EXAMPLE_MTMD,
     LLAMA_EXAMPLE_LOOKUP,
     LLAMA_EXAMPLE_PARALLEL,
     LLAMA_EXAMPLE_TTS,
@@ -96,6 +96,7 @@ enum common_sampler_type {
     COMMON_SAMPLER_TYPE_XTC         = 8,
     COMMON_SAMPLER_TYPE_INFILL      = 9,
     COMMON_SAMPLER_TYPE_PENALTIES   = 10,
+    COMMON_SAMPLER_TYPE_TOP_N_SIGMA = 11,
 };
 
 // dimensionality reduction methods, used by cvector-generator
@@ -114,7 +115,7 @@ 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,
+    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
 };
 
 struct common_grammar_trigger {
@@ -161,6 +162,7 @@ struct common_params_sampling {
     std::vector<enum common_sampler_type> samplers = {
         COMMON_SAMPLER_TYPE_PENALTIES,
         COMMON_SAMPLER_TYPE_DRY,
+        COMMON_SAMPLER_TYPE_TOP_N_SIGMA,
         COMMON_SAMPLER_TYPE_TOP_K,
         COMMON_SAMPLER_TYPE_TYPICAL_P,
         COMMON_SAMPLER_TYPE_TOP_P,
@@ -213,7 +215,8 @@ struct common_params_vocoder {
 
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
+    COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
+    COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
 };
 
 struct common_params {
@@ -289,6 +292,7 @@ struct common_params {
     int32_t verbosity                  = 0;
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
+    bool    offline                    = false;
 
     int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
     int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
@@ -321,17 +325,17 @@ struct common_params {
     bool flash_attn        = false; // flash attention
     bool no_perf           = false; // disable performance metrics
     bool ctx_shift         = true;  // context shift on inifinite text generation
+    bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
-    bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
-    bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
+    bool no_op_offload     = false; // globally disable offload host tensor operations to device
 
     bool single_turn       = false; // single turn chat conversation
 
@@ -340,8 +344,10 @@ struct common_params {
 
     common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
 
-    // multimodal models (see examples/llava)
+    // multimodal models (see tools/mtmd)
     struct common_params_model mmproj;
+    bool mmproj_use_gpu = true;     // use GPU for multimodal model
+    bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
 
     // embedding
@@ -364,6 +370,8 @@ struct common_params {
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    int reasoning_budget = -1;
+    bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 
     std::vector<std::string> api_keys;
 
@@ -407,13 +415,14 @@ struct common_params {
 
     bool process_output = false; // collect data for the output tensor
     bool compute_ppl    = true;  // whether to compute perplexity
+    bool parse_special  = false; // whether to parse special tokens during imatrix tokenization
 
     // cvector-generator params
     int n_pca_batch = 100;
     int n_pca_iterations = 1000;
     dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
-    std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
-    std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
+    std::string cvector_positive_file = "tools/cvector-generator/positive.txt";
+    std::string cvector_negative_file = "tools/cvector-generator/negative.txt";
 
     bool spm_infill = false; // suffix/prefix/middle pattern for infill
 
@@ -422,6 +431,11 @@ struct common_params {
 
     // common params
     std::string out_file; // output filename for all example programs
+    // optional callback for model loading progress and cancellation:
+    // called with a progress value between 0.0 and 1.0.
+    // return false from callback to abort model loading or true to continue
+    llama_progress_callback load_progress_callback = NULL;
+    void *                  load_progress_callback_user_data = NULL;
 };
 
 // call once at the start of a program if it uses libcommon
@@ -499,10 +513,9 @@ static bool string_starts_with(const std::string & str,
     return str.rfind(prefix, 0) == 0;
 }
 
-static bool string_ends_with(const std::string & str,
-                               const std::string & suffix) {  // While we wait for C++20's std::string::ends_with...
-    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
-}
+// While we wait for C++20's std::string::ends_with...
+bool string_ends_with(const std::string_view & str, const std::string_view & suffix);
+size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop);
 
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
@@ -611,16 +624,6 @@ std::string common_detokenize(
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
-//
-// KV cache utils
-//
-
-// Dump the KV cache view with the number of sequences per cell.
-void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
-
-// Dump the KV cache view showing individual sequences in each cell (long output).
-void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
-
 //
 // Embedding utils
 //
@@ -662,3 +665,9 @@ const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 
 }
+
+//
+// training utils
+//
+
+ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);

llama/llama.cpp/common/console.cpp

@@ -0,0 +1,504 @@
+#include "console.h"
+#include <vector>
+#include <iostream>
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <fcntl.h>
+#include <io.h>
+#ifndef ENABLE_VIRTUAL_TERMINAL_PROCESSING
+#define ENABLE_VIRTUAL_TERMINAL_PROCESSING 0x0004
+#endif
+#else
+#include <climits>
+#include <sys/ioctl.h>
+#include <unistd.h>
+#include <wchar.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <signal.h>
+#include <termios.h>
+#endif
+
+#define ANSI_COLOR_RED     "\x1b[31m"
+#define ANSI_COLOR_GREEN   "\x1b[32m"
+#define ANSI_COLOR_YELLOW  "\x1b[33m"
+#define ANSI_COLOR_BLUE    "\x1b[34m"
+#define ANSI_COLOR_MAGENTA "\x1b[35m"
+#define ANSI_COLOR_CYAN    "\x1b[36m"
+#define ANSI_COLOR_RESET   "\x1b[0m"
+#define ANSI_BOLD          "\x1b[1m"
+
+namespace console {
+
+    //
+    // Console state
+    //
+
+    static bool      advanced_display = false;
+    static bool      simple_io        = true;
+    static display_t current_display  = reset;
+
+    static FILE*     out              = stdout;
+
+#if defined (_WIN32)
+    static void*     hConsole;
+#else
+    static FILE*     tty              = nullptr;
+    static termios   initial_state;
+#endif
+
+    //
+    // Init and cleanup
+    //
+
+    void init(bool use_simple_io, bool use_advanced_display) {
+        advanced_display = use_advanced_display;
+        simple_io = use_simple_io;
+#if defined(_WIN32)
+        // Windows-specific console initialization
+        DWORD dwMode = 0;
+        hConsole = GetStdHandle(STD_OUTPUT_HANDLE);
+        if (hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(hConsole, &dwMode)) {
+            hConsole = GetStdHandle(STD_ERROR_HANDLE);
+            if (hConsole != INVALID_HANDLE_VALUE && (!GetConsoleMode(hConsole, &dwMode))) {
+                hConsole = nullptr;
+                simple_io = true;
+            }
+        }
+        if (hConsole) {
+            // Check conditions combined to reduce nesting
+            if (advanced_display && !(dwMode & ENABLE_VIRTUAL_TERMINAL_PROCESSING) &&
+                !SetConsoleMode(hConsole, dwMode | ENABLE_VIRTUAL_TERMINAL_PROCESSING)) {
+                advanced_display = false;
+            }
+            // Set console output codepage to UTF8
+            SetConsoleOutputCP(CP_UTF8);
+        }
+        HANDLE hConIn = GetStdHandle(STD_INPUT_HANDLE);
+        if (hConIn != INVALID_HANDLE_VALUE && GetConsoleMode(hConIn, &dwMode)) {
+            // Set console input codepage to UTF16
+            _setmode(_fileno(stdin), _O_WTEXT);
+
+            // Set ICANON (ENABLE_LINE_INPUT) and ECHO (ENABLE_ECHO_INPUT)
+            if (simple_io) {
+                dwMode |= ENABLE_LINE_INPUT | ENABLE_ECHO_INPUT;
+            } else {
+                dwMode &= ~(ENABLE_LINE_INPUT | ENABLE_ECHO_INPUT);
+            }
+            if (!SetConsoleMode(hConIn, dwMode)) {
+                simple_io = true;
+            }
+        }
+        if (simple_io) {
+            _setmode(_fileno(stdin), _O_U8TEXT);
+        }
+#else
+        // POSIX-specific console initialization
+        if (!simple_io) {
+            struct termios new_termios;
+            tcgetattr(STDIN_FILENO, &initial_state);
+            new_termios = initial_state;
+            new_termios.c_lflag &= ~(ICANON | ECHO);
+            new_termios.c_cc[VMIN] = 1;
+            new_termios.c_cc[VTIME] = 0;
+            tcsetattr(STDIN_FILENO, TCSANOW, &new_termios);
+
+            tty = fopen("/dev/tty", "w+");
+            if (tty != nullptr) {
+                out = tty;
+            }
+        }
+
+        setlocale(LC_ALL, "");
+#endif
+    }
+
+    void cleanup() {
+        // Reset console display
+        set_display(reset);
+
+#if !defined(_WIN32)
+        // Restore settings on POSIX systems
+        if (!simple_io) {
+            if (tty != nullptr) {
+                out = stdout;
+                fclose(tty);
+                tty = nullptr;
+            }
+            tcsetattr(STDIN_FILENO, TCSANOW, &initial_state);
+        }
+#endif
+    }
+
+    //
+    // Display and IO
+    //
+
+    // Keep track of current display and only emit ANSI code if it changes
+    void set_display(display_t display) {
+        if (advanced_display && current_display != display) {
+            fflush(stdout);
+            switch(display) {
+                case reset:
+                    fprintf(out, ANSI_COLOR_RESET);
+                    break;
+                case prompt:
+                    fprintf(out, ANSI_COLOR_YELLOW);
+                    break;
+                case user_input:
+                    fprintf(out, ANSI_BOLD ANSI_COLOR_GREEN);
+                    break;
+                case error:
+                    fprintf(out, ANSI_BOLD ANSI_COLOR_RED);
+            }
+            current_display = display;
+            fflush(out);
+        }
+    }
+
+    static char32_t getchar32() {
+#if defined(_WIN32)
+        HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE);
+        wchar_t high_surrogate = 0;
+
+        while (true) {
+            INPUT_RECORD record;
+            DWORD count;
+            if (!ReadConsoleInputW(hConsole, &record, 1, &count) || count == 0) {
+                return WEOF;
+            }
+
+            if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) {
+                wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar;
+                if (wc == 0) {
+                    continue;
+                }
+
+                if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
+                    high_surrogate = wc;
+                    continue;
+                }
+                if ((wc >= 0xDC00) && (wc <= 0xDFFF)) { // Check if wc is a low surrogate
+                    if (high_surrogate != 0) { // Check if we have a high surrogate
+                        return ((high_surrogate - 0xD800) << 10) + (wc - 0xDC00) + 0x10000;
+                    }
+                }
+
+                high_surrogate = 0; // Reset the high surrogate
+                return static_cast<char32_t>(wc);
+            }
+        }
+#else
+        wchar_t wc = getwchar();
+        if (static_cast<wint_t>(wc) == WEOF) {
+            return WEOF;
+        }
+
+#if WCHAR_MAX == 0xFFFF
+        if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
+            wchar_t low_surrogate = getwchar();
+            if ((low_surrogate >= 0xDC00) && (low_surrogate <= 0xDFFF)) { // Check if the next wchar is a low surrogate
+                return (static_cast<char32_t>(wc & 0x03FF) << 10) + (low_surrogate & 0x03FF) + 0x10000;
+            }
+        }
+        if ((wc >= 0xD800) && (wc <= 0xDFFF)) { // Invalid surrogate pair
+            return 0xFFFD; // Return the replacement character U+FFFD
+        }
+#endif
+
+        return static_cast<char32_t>(wc);
+#endif
+    }
+
+    static void pop_cursor() {
+#if defined(_WIN32)
+        if (hConsole != NULL) {
+            CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
+            GetConsoleScreenBufferInfo(hConsole, &bufferInfo);
+
+            COORD newCursorPosition = bufferInfo.dwCursorPosition;
+            if (newCursorPosition.X == 0) {
+                newCursorPosition.X = bufferInfo.dwSize.X - 1;
+                newCursorPosition.Y -= 1;
+            } else {
+                newCursorPosition.X -= 1;
+            }
+
+            SetConsoleCursorPosition(hConsole, newCursorPosition);
+            return;
+        }
+#endif
+        putc('\b', out);
+    }
+
+    static int estimateWidth(char32_t codepoint) {
+#if defined(_WIN32)
+        (void)codepoint;
+        return 1;
+#else
+        return wcwidth(codepoint);
+#endif
+    }
+
+    static int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) {
+#if defined(_WIN32)
+        CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
+        if (!GetConsoleScreenBufferInfo(hConsole, &bufferInfo)) {
+            // go with the default
+            return expectedWidth;
+        }
+        COORD initialPosition = bufferInfo.dwCursorPosition;
+        DWORD nNumberOfChars = length;
+        WriteConsole(hConsole, utf8_codepoint, nNumberOfChars, &nNumberOfChars, NULL);
+
+        CONSOLE_SCREEN_BUFFER_INFO newBufferInfo;
+        GetConsoleScreenBufferInfo(hConsole, &newBufferInfo);
+
+        // Figure out our real position if we're in the last column
+        if (utf8_codepoint[0] != 0x09 && initialPosition.X == newBufferInfo.dwSize.X - 1) {
+            DWORD nNumberOfChars;
+            WriteConsole(hConsole, &" \b", 2, &nNumberOfChars, NULL);
+            GetConsoleScreenBufferInfo(hConsole, &newBufferInfo);
+        }
+
+        int width = newBufferInfo.dwCursorPosition.X - initialPosition.X;
+        if (width < 0) {
+            width += newBufferInfo.dwSize.X;
+        }
+        return width;
+#else
+        // We can trust expectedWidth if we've got one
+        if (expectedWidth >= 0 || tty == nullptr) {
+            fwrite(utf8_codepoint, length, 1, out);
+            return expectedWidth;
+        }
+
+        fputs("\033[6n", tty); // Query cursor position
+        int x1;
+        int y1;
+        int x2;
+        int y2;
+        int results = 0;
+        results = fscanf(tty, "\033[%d;%dR", &y1, &x1);
+
+        fwrite(utf8_codepoint, length, 1, tty);
+
+        fputs("\033[6n", tty); // Query cursor position
+        results += fscanf(tty, "\033[%d;%dR", &y2, &x2);
+
+        if (results != 4) {
+            return expectedWidth;
+        }
+
+        int width = x2 - x1;
+        if (width < 0) {
+            // Calculate the width considering text wrapping
+            struct winsize w;
+            ioctl(STDOUT_FILENO, TIOCGWINSZ, &w);
+            width += w.ws_col;
+        }
+        return width;
+#endif
+    }
+
+    static void replace_last(char ch) {
+#if defined(_WIN32)
+        pop_cursor();
+        put_codepoint(&ch, 1, 1);
+#else
+        fprintf(out, "\b%c", ch);
+#endif
+    }
+
+    static void append_utf8(char32_t ch, std::string & out) {
+        if (ch <= 0x7F) {
+            out.push_back(static_cast<unsigned char>(ch));
+        } else if (ch <= 0x7FF) {
+            out.push_back(static_cast<unsigned char>(0xC0 | ((ch >> 6) & 0x1F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else if (ch <= 0xFFFF) {
+            out.push_back(static_cast<unsigned char>(0xE0 | ((ch >> 12) & 0x0F)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else if (ch <= 0x10FFFF) {
+            out.push_back(static_cast<unsigned char>(0xF0 | ((ch >> 18) & 0x07)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 12) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else {
+            // Invalid Unicode code point
+        }
+    }
+
+    // Helper function to remove the last UTF-8 character from a string
+    static void pop_back_utf8_char(std::string & line) {
+        if (line.empty()) {
+            return;
+        }
+
+        size_t pos = line.length() - 1;
+
+        // Find the start of the last UTF-8 character (checking up to 4 bytes back)
+        for (size_t i = 0; i < 3 && pos > 0; ++i, --pos) {
+            if ((line[pos] & 0xC0) != 0x80) {
+                break; // Found the start of the character
+            }
+        }
+        line.erase(pos);
+    }
+
+    static bool readline_advanced(std::string & line, bool multiline_input) {
+        if (out != stdout) {
+            fflush(stdout);
+        }
+
+        line.clear();
+        std::vector<int> widths;
+        bool is_special_char = false;
+        bool end_of_stream = false;
+
+        char32_t input_char;
+        while (true) {
+            fflush(out); // Ensure all output is displayed before waiting for input
+            input_char = getchar32();
+
+            if (input_char == '\r' || input_char == '\n') {
+                break;
+            }
+
+            if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) {
+                end_of_stream = true;
+                break;
+            }
+
+            if (is_special_char) {
+                set_display(user_input);
+                replace_last(line.back());
+                is_special_char = false;
+            }
+
+            if (input_char == '\033') { // Escape sequence
+                char32_t code = getchar32();
+                if (code == '[' || code == 0x1B) {
+                    // Discard the rest of the escape sequence
+                    while ((code = getchar32()) != (char32_t) WEOF) {
+                        if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') {
+                            break;
+                        }
+                    }
+                }
+            } else if (input_char == 0x08 || input_char == 0x7F) { // Backspace
+                if (!widths.empty()) {
+                    int count;
+                    do {
+                        count = widths.back();
+                        widths.pop_back();
+                        // Move cursor back, print space, and move cursor back again
+                        for (int i = 0; i < count; i++) {
+                            replace_last(' ');
+                            pop_cursor();
+                        }
+                        pop_back_utf8_char(line);
+                    } while (count == 0 && !widths.empty());
+                }
+            } else {
+                int offset = line.length();
+                append_utf8(input_char, line);
+                int width = put_codepoint(line.c_str() + offset, line.length() - offset, estimateWidth(input_char));
+                if (width < 0) {
+                    width = 0;
+                }
+                widths.push_back(width);
+            }
+
+            if (!line.empty() && (line.back() == '\\' || line.back() == '/')) {
+                set_display(prompt);
+                replace_last(line.back());
+                is_special_char = true;
+            }
+        }
+
+        bool has_more = multiline_input;
+        if (is_special_char) {
+            replace_last(' ');
+            pop_cursor();
+
+            char last = line.back();
+            line.pop_back();
+            if (last == '\\') {
+                line += '\n';
+                fputc('\n', out);
+                has_more = !has_more;
+            } else {
+                // llama will just eat the single space, it won't act as a space
+                if (line.length() == 1 && line.back() == ' ') {
+                    line.clear();
+                    pop_cursor();
+                }
+                has_more = false;
+            }
+        } else {
+            if (end_of_stream) {
+                has_more = false;
+            } else {
+                line += '\n';
+                fputc('\n', out);
+            }
+        }
+
+        fflush(out);
+        return has_more;
+    }
+
+    static bool readline_simple(std::string & line, bool multiline_input) {
+#if defined(_WIN32)
+        std::wstring wline;
+        if (!std::getline(std::wcin, wline)) {
+            // Input stream is bad or EOF received
+            line.clear();
+            GenerateConsoleCtrlEvent(CTRL_C_EVENT, 0);
+            return false;
+        }
+
+        int size_needed = WideCharToMultiByte(CP_UTF8, 0, &wline[0], (int)wline.size(), NULL, 0, NULL, NULL);
+        line.resize(size_needed);
+        WideCharToMultiByte(CP_UTF8, 0, &wline[0], (int)wline.size(), &line[0], size_needed, NULL, NULL);
+#else
+        if (!std::getline(std::cin, line)) {
+            // Input stream is bad or EOF received
+            line.clear();
+            return false;
+        }
+#endif
+        if (!line.empty()) {
+            char last = line.back();
+            if (last == '/') { // Always return control on '/' symbol
+                line.pop_back();
+                return false;
+            }
+            if (last == '\\') { // '\\' changes the default action
+                line.pop_back();
+                multiline_input = !multiline_input;
+            }
+        }
+        line += '\n';
+
+        // By default, continue input if multiline_input is set
+        return multiline_input;
+    }
+
+    bool readline(std::string & line, bool multiline_input) {
+        set_display(user_input);
+
+        if (simple_io) {
+            return readline_simple(line, multiline_input);
+        }
+        return readline_advanced(line, multiline_input);
+    }
+
+}

llama/llama.cpp/common/console.h

@@ -0,0 +1,19 @@
+// Console functions
+
+#pragma once
+
+#include <string>
+
+namespace console {
+    enum display_t {
+        reset = 0,
+        prompt,
+        user_input,
+        error
+    };
+
+    void init(bool use_simple_io, bool use_advanced_display);
+    void cleanup();
+    void set_display(display_t display);
+    bool readline(std::string & line, bool multiline_input);
+}

llama/llama.cpp/common/json-partial.cpp

@@ -0,0 +1,256 @@
+#include "json-partial.h"
+
+#include "log.h"
+
+#include <nlohmann/json.hpp>
+
+#include <string>
+
+using json = nlohmann::ordered_json;
+
+enum common_json_stack_element_type {
+    COMMON_JSON_STACK_ELEMENT_OBJECT,
+    COMMON_JSON_STACK_ELEMENT_KEY,
+    COMMON_JSON_STACK_ELEMENT_ARRAY,
+};
+
+struct common_json_stack_element {
+    common_json_stack_element_type type;
+    std::string key;
+};
+
+bool common_json_parse(
+    const std::string & input,
+    const std::string & healing_marker,
+    common_json & out)
+{
+    std::string::const_iterator it = input.begin();
+    const auto end = input.end();
+    return common_json_parse(it, end, healing_marker, out);
+}
+
+bool common_json_parse(
+    std::string::const_iterator & it,
+    const std::string::const_iterator & end,
+    const std::string & healing_marker,
+    common_json & out)
+{
+    // // https://json.nlohmann.me/features/parsing/sax_interface/
+    struct json_error_locator : public nlohmann::json_sax<json> {
+        std::size_t position;
+        bool found_error;
+        std::string last_token;
+        std::string exception_message;
+        std::vector<common_json_stack_element> stack;
+
+        json_error_locator() : position(0), found_error(false) {}
+
+        bool parse_error(std::size_t position, const std::string & last_token, const json::exception & ex) override { // NOLINT
+            this->position = position - 1;
+            this->found_error = true;
+            this->last_token = last_token;
+            this->exception_message = ex.what();
+            return false;
+        }
+        void close_value() {
+            if (!stack.empty() && (stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY)) {
+                stack.pop_back();
+            }
+        }
+        bool null() override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool boolean(bool) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_integer(number_integer_t) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_unsigned(number_unsigned_t) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool number_float(number_float_t, const string_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool string(string_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool binary(binary_t &) override { // NOLINT
+            close_value();
+            return true;
+        }
+        bool start_object(std::size_t) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_OBJECT, ""});
+            return true;
+        }
+        bool end_object() override {
+            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT);
+            stack.pop_back();
+            close_value();
+            return true;
+        }
+        bool key(string_t & key) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_KEY, key});
+            return true;
+        }
+        bool start_array(std::size_t) override { // NOLINT
+            stack.push_back({COMMON_JSON_STACK_ELEMENT_ARRAY, ""});
+            return true;
+        }
+        bool end_array() override {
+            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY);
+            stack.pop_back();
+            close_value();
+            return true;
+        }
+    };
+    json_error_locator err_loc;
+    auto start = it;
+    json::sax_parse(it, end, &err_loc);
+
+    if (err_loc.found_error) {
+        it = start;
+        auto temptative_end = it + err_loc.position;
+        // LOG_DBG("Error at position %zu (is_end = %s): %s\n", err_loc.position, temptative_end == end ? "true" : "false", err_loc.exception_message.c_str());
+
+        auto input = std::string(it, temptative_end);
+        try {
+            out.json = json::parse(input);
+            // out.json = json::parse(it, temptative_end);
+            it = temptative_end;
+            return true;
+        } catch (const std::exception & ex) {
+            // No, needs healing.
+            LOG_DBG("Failed to parse up to error: %s: <<<%s>>>\n", ex.what(), std::string(it, temptative_end).c_str());
+        }
+        auto can_parse = [](const std::string & str) {
+            try {
+                auto _ = json::parse(str); // NOLINT
+                return true;
+            } catch (const std::exception &) {
+                return false;
+            }
+        };
+        if (!healing_marker.empty() && !err_loc.stack.empty()) {
+            std::string str(it, temptative_end);
+            auto last_non_sp_pos = str.find_last_not_of(" \n\r\t");
+            if (last_non_sp_pos == std::string::npos) {
+                throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
+            }
+            auto last_non_sp_char = str[last_non_sp_pos];
+            // Used to detect stops on a number, which may not be complete.
+            auto was_maybe_number = [&]() {
+                if (!str.empty() && std::isspace(str.back())) {
+                    return false;
+                }
+                return std::isdigit(last_non_sp_char) ||
+                    last_non_sp_char == '.' ||
+                    last_non_sp_char == 'e' ||
+                    last_non_sp_char == 'E' ||
+                    last_non_sp_char == '-';
+            };
+
+            std::string closing;
+            for (size_t i = err_loc.stack.size(); i > 0; i--) {
+                auto & el = err_loc.stack[i - 1];
+                if (el.type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
+                    closing += "}";
+                } else if (el.type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
+                    closing += "]";
+                } else if (el.type != COMMON_JSON_STACK_ELEMENT_KEY) {
+                    throw std::runtime_error("Unexpected stack element type");
+                }
+            }
+
+            const auto & magic_seed = out.healing_marker.marker = healing_marker;//"$llama.cpp.json$";
+
+            if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY) {
+                // We're inside an object value
+                if (last_non_sp_char == ':' && can_parse(str + "1" + closing)) {
+                    // Was about to create an object value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + ": 1" + closing)) {
+                    str += (out.healing_marker.json_dump_marker = ":\"" + magic_seed) + "\"" + closing;
+                } else if (last_non_sp_char == '{' && can_parse(str + closing)) {
+                    // Was about to create an object
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
+                } else if (can_parse(str + "\"" + closing)) {
+                    // Was inside an object value string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
+                    // Was inside an object value string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else {
+                    // find last :
+                    auto last_pos = str.find_last_of(':');
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
+                    }
+                    // Cutting back to opening : for object value
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
+                if ((last_non_sp_char == ',' || last_non_sp_char == '[') && can_parse(str + "1" + closing)) {
+                    // Was about to create an array value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + "\"" + closing)) {
+                    // Was inside an array value string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
+                    // Was inside an array value string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else if (!was_maybe_number() && can_parse(str + ", 1" + closing)) {
+                    // Had just finished a value
+                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\"" + closing;
+                } else {
+                    auto last_pos = str.find_last_of("[,");
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON array stopped in an unknown location");
+                    }
+                    // Cutting back to last [ or , for array value
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
+                if ((last_non_sp_char == '{' && can_parse(str + closing)) ||
+                        (last_non_sp_char == ',' && can_parse(str + "\"\": 1" + closing))) {
+                    // Was about to create an object key+value
+                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
+                } else if (!was_maybe_number() && can_parse(str + ",\"\": 1" + closing)) {
+                    // Was about to create an object key+value
+                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\": 1" + closing;
+                } else if (can_parse(str + "\": 1" + closing)) {
+                    // Was inside an object key string
+                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\": 1" + closing;
+                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\": 1" + closing)) {
+                    // Was inside an object key string after an escape
+                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\": 1" + closing;
+                } else {
+                    auto last_pos = str.find_last_of(':');
+                    if (last_pos == std::string::npos) {
+                        throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
+                    }
+                    // fprintf(stderr, "Cutting back to last : for object key+value\n");
+                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
+                }
+            } else {
+                throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
+            }
+            // fprintf(stderr, "HEALED:\nSTRING <<<\n%s\n>>>\n\nmagic_cut: <<<\n%s\n>>>\n\n", str.c_str(), out.healing_marker.json_dump_marker.c_str());
+            out.json = json::parse(str);
+            it = temptative_end;
+            return true;
+        }
+        // TODO: handle unclosed top-level primitive if the stack was empty but we got an error (e.g. "tru", "\"", etc...)
+        // fprintf(stderr, "Closing: TODO\n");
+        return false;
+    }
+    out.json = json::parse(it, end);
+    it = end;
+    return true;
+}

llama/llama.cpp/common/json-partial.h

@@ -0,0 +1,38 @@
+#pragma once
+
+#include <nlohmann/json.hpp>
+
+// Healing marker (empty if the JSON was fully parsed / wasn't healed).
+struct common_healing_marker {
+    // Raw marker.
+    std::string marker;
+
+    // Cutting the `common_json.json.dump()` string at the (only) occurrence of this marker should yield the original partial JSON string (modulo spaces / if it had the same dump format).
+    std::string json_dump_marker;
+};
+
+// Represents a parsed JSON object, with its optional healing marker (a JSON dump fragment that can be used to find the position of healing in the JSON dump string)
+struct common_json {
+    nlohmann::ordered_json json;
+
+    common_healing_marker healing_marker;
+};
+
+// Parse the JSON string, healing (closing) any partial JSON if `healing_marker` is not empty.
+//
+// Healing completes partial JSON strings by adding a (possibly modified) healing marker, then whatever is needed to close the JSON.
+// This allows to parse the resulting healed JSON string, yet be able to cut it again if needed at the healing marker.
+// (this is used when parsing JSON outputs from the models, then crafting partial JSONs for the partial tool calls in OAI format).
+//
+// For instance, parsing `{` with a healing marker `foo` will produce a healed JSON `{"foo":1}`, w/ json_dump_marker = `"foo"` (which can be used to break the JSON again).
+bool common_json_parse(
+    const std::string & input,
+    const std::string & healing_marker,
+    common_json & out);
+
+// Parse the JSON string (see overload above), but advancing an iterator to the end of the input when the (potentially partial) parsing succeeds.
+bool common_json_parse(
+    std::string::const_iterator & it,
+    const std::string::const_iterator & end,
+    const std::string & healing_marker,
+    common_json & out);

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

@@ -1,8 +1,9 @@
 #include "json-schema-to-grammar.h"
 #include "common.h"
 
+#include <nlohmann/json.hpp>
+
 #include <algorithm>
-#include <fstream>
 #include <map>
 #include <regex>
 #include <sstream>
@@ -16,6 +17,9 @@ using json = nlohmann::ordered_json;
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
 
+    if (max_items == 0) {
+        return "";
+    }
     if (min_items == 0 && max_items == 1) {
         return item_rule + "?";
     }

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

@@ -1,9 +1,9 @@
 #pragma once
 
-#include "ggml.h"
-// Change JSON_ASSERT from assert() to GGML_ASSERT:
-#define JSON_ASSERT GGML_ASSERT
-#include "json.hpp"
+#include <nlohmann/json_fwd.hpp>
+
+#include <functional>
+#include <string>
 
 std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
                                    bool force_gbnf = false);

llama/llama.cpp/common/regex-partial.cpp

@@ -0,0 +1,204 @@
+#include "regex-partial.h"
+#include "common.h"
+#include <functional>
+#include <optional>
+
+common_regex::common_regex(const std::string & pattern) :
+    pattern(pattern),
+    rx(pattern),
+    rx_reversed_partial(regex_to_reversed_partial_regex(pattern)) {}
+
+common_regex_match common_regex::search(const std::string & input, size_t pos, bool as_match) const {
+    std::smatch match;
+    if (pos > input.size()) {
+        throw std::runtime_error("Position out of bounds");
+    }
+    auto start = input.begin() + pos;
+    auto found = as_match
+        ? std::regex_match(start, input.end(), match, rx)
+        : std::regex_search(start, input.end(), match, rx);
+    if (found) {
+        common_regex_match res;
+        res.type = COMMON_REGEX_MATCH_TYPE_FULL;
+        for (size_t i = 0; i < match.size(); ++i) {
+            auto begin = pos + match.position(i);
+            res.groups.emplace_back(begin, begin + match.length(i));
+        }
+        return res;
+    }
+    std::match_results<std::string::const_reverse_iterator> srmatch;
+    if (std::regex_match(input.rbegin(), input.rend() - pos, srmatch, rx_reversed_partial)) {
+        auto group = srmatch[1].str();
+        if (group.length() != 0) {
+            auto it = srmatch[1].second.base();
+            // auto position = static_cast<size_t>(std::distance(input.begin(), it));
+            if ((!as_match) || it == input.begin()) {
+                common_regex_match res;
+                res.type = COMMON_REGEX_MATCH_TYPE_PARTIAL;
+                const size_t begin = std::distance(input.begin(), it);
+                const size_t end = input.size();
+                if (begin == std::string::npos || end == std::string::npos || begin > end) {
+                    throw std::runtime_error("Invalid range");
+                }
+                res.groups.push_back({begin, end});
+                return res;
+            }
+        }
+    }
+    return {};
+}
+
+/*
+  Transforms a regex pattern to a partial match pattern that operates on a reversed input string to find partial final matches of the original pattern.
+
+  Ideally we'd like to use boost::match_partial (https://beta.boost.org/doc/libs/1_59_0/libs/regex/doc/html/boost_regex/partial_matches.html)
+  to see if a string ends with a partial regex match, but but it's not in std::regex yet.
+  Instead, we'll the regex into a partial match regex operating as a full match on the reverse iterators of the input.
+
+  - /abcd/ -> (dcba|cba|ba|a).* -> ((?:(?:(?:(?:d)?c)?b)?a).*
+  - /a|b/ -> (a|b).*
+  - /a*?/ -> error, could match ""
+  - /a*b/ -> ((?:b)?a*+).* (final repetitions become eager)
+  - /.*?ab/ -> ((?:b)?a).* (merge .*)
+  - /a.*?b/ -> ((?:b)?.*?a).* (keep reluctant matches)
+  - /a(bc)d/ -> ((?:(?:d)?(?:(?:c)?b))?a).*
+  - /a(bc|de)/ -> ((?:(?:(?:e)?d)?|(?:(?:c)?b)?)?a).*
+  - /ab{2,4}c/ -> abbb?b?c -> ((?:(?:(?:(?:(?:c)?b)?b)?b?)?b?)?a).*
+
+  The regex will match a reversed string fully, and the end of the first (And only) capturing group will indicate the reversed start of the original partial pattern
+  (i.e. just where the final .* starts in the inverted pattern; all other groups are turned into non-capturing groups, and reluctant quantifiers are ignored)
+*/
+std::string regex_to_reversed_partial_regex(const std::string & pattern) {
+    auto it = pattern.begin();
+    const auto end = pattern.end();
+
+    std::function<std::string()> process = [&]() {
+        std::vector<std::vector<std::string>> alternatives(1);
+        std::vector<std::string> * sequence = &alternatives.back();
+
+        while (it != end) {
+            if (*it == '[') {
+                auto start = it;
+                ++it;
+                while (it != end) {
+                    if ((*it == '\\') && (++it != end)) {
+                        ++it;
+                    } else if ((it != end) && (*it == ']')) {
+                        break;
+                    } else {
+                        ++it;
+                    }
+                }
+                if (it == end) {
+                    throw std::runtime_error("Unmatched '[' in pattern");
+                }
+                ++it;
+                sequence->push_back(std::string(start, it));
+            } else if (*it == '*' || *it == '?' || *it == '+') {
+                if (sequence->empty()) {
+                    throw std::runtime_error("Quantifier without preceding element");
+                }
+                sequence->back() += *it;
+                auto is_star = *it == '*';
+                ++it;
+                if (is_star) {
+                    if (*it == '?') {
+                        ++it;
+                    }
+                }
+            } else if (*it == '{') {
+                if (sequence->empty()) {
+                    throw std::runtime_error("Repetition without preceding element");
+                }
+                ++it;
+                auto start = it;
+                while (it != end && *it != '}') {
+                    ++it;
+                }
+                if (it == end) {
+                    throw std::runtime_error("Unmatched '{' in pattern");
+                }
+                auto parts = string_split(std::string(start, it), ",");
+                ++it;
+                if (parts.size() > 2) {
+                    throw std::runtime_error("Invalid repetition range in pattern");
+                }
+
+                auto parseOptInt = [&](const std::string & s, const std::optional<int> & def = std::nullopt) -> std::optional<int> {
+                    if (s.empty()) {
+                        return def;
+                    }
+                    return std::stoi(s);
+                };
+                auto min = parseOptInt(parts[0], 0);
+                auto max = parts.size() == 1 ? min : parseOptInt(parts[1]);
+                if (min && max && *max < *min) {
+                    throw std::runtime_error("Invalid repetition range in pattern");
+                }
+                // Brutal but... let's repeat at least min times, then ? for the delta between min & max (or * for unbounded)
+                auto part = sequence->back();
+                sequence->pop_back();
+                for (int i = 0; i < *min; i++) {
+                    sequence->push_back(part);
+                }
+                if (max) {
+                    for (int i = *min; i < *max; i++) {
+                        sequence->push_back(part + "?");
+                    }
+                } else {
+                    sequence->push_back(part + "*");
+                }
+            } else if (*it == '(') {
+                ++it;
+                if (it != end && *it == '?' && (it + 1 != end) && *(it + 1) == ':') {
+                    it += 2;
+                }
+                auto sub = process();
+                if (*it != ')') {
+                    throw std::runtime_error("Unmatched '(' in pattern");
+                }
+                ++it;
+                auto & part = sequence->emplace_back("(?:");
+                part += sub;
+                part += ")";
+            } else if (*it == ')') {
+                break;
+            } else if (*it == '|') {
+                ++it;
+                alternatives.emplace_back();
+                sequence = &alternatives.back();
+            } else if (*it == '\\' && (++it != end)) {
+                auto str = std::string("\\") + *it;
+                sequence->push_back(str);
+                ++it;
+            } else if (it != end) {
+                sequence->push_back(std::string(1, *it));
+                ++it;
+            }
+        }
+
+        // /abcd/ -> (dcba|cba|ba|a).* -> ((?:(?:(?:d)?c)?b)?a).*
+        // if n(=4) parts, opening n-1(=3) non-capturing groups after the 1 capturing group
+        // We'll do the outermost capturing group and final .* in the enclosing function.
+        std::vector<std::string> res_alts;
+        for (const auto & parts : alternatives) {
+            auto & res = res_alts.emplace_back();
+            for (size_t i = 0; i < parts.size() - 1; i++) {
+                res += "(?:";
+            }
+            for (auto it = parts.rbegin(); it != parts.rend(); ++it) {
+                res += *it;
+                if (it != parts.rend() - 1) {
+                    res += ")?";
+                }
+            }
+        }
+        return string_join(res_alts, "|");
+    };
+    auto res = process();
+    if (it != end) {
+        throw std::runtime_error("Unmatched '(' in pattern");
+    }
+
+    return "(" + res + ")[\\s\\S]*";
+}

llama/llama.cpp/common/regex-partial.h

@@ -0,0 +1,56 @@
+#pragma once
+
+#include <regex>
+#include <string>
+
+enum common_regex_match_type {
+    COMMON_REGEX_MATCH_TYPE_NONE,
+    COMMON_REGEX_MATCH_TYPE_PARTIAL,
+    COMMON_REGEX_MATCH_TYPE_FULL,
+};
+
+struct common_string_range {
+    size_t begin;
+    size_t end;
+    common_string_range(size_t begin, size_t end) : begin(begin), end(end) {
+        if (begin > end) {
+            throw std::runtime_error("Invalid range");
+        }
+    }
+    // prevent default ctor
+    common_string_range() = delete;
+    bool empty() const {
+        return begin == end;
+    }
+    bool operator==(const common_string_range & other) const {
+        return begin == other.begin && end == other.end;
+    }
+};
+
+struct common_regex_match {
+    common_regex_match_type type = COMMON_REGEX_MATCH_TYPE_NONE;
+    std::vector<common_string_range> groups;
+
+    bool operator==(const common_regex_match & other) const {
+        return type == other.type && groups == other.groups;
+    }
+    bool operator!=(const common_regex_match & other) const {
+        return !(*this == other);
+    }
+};
+
+class common_regex {
+    std::string pattern;
+    std::regex rx;
+    std::regex rx_reversed_partial;
+
+  public:
+    explicit common_regex(const std::string & pattern);
+
+    common_regex_match search(const std::string & input, size_t pos, bool as_match = false) const;
+
+    const std::string & str() const { return pattern; }
+};
+
+// For testing only (pretty print of failures).
+std::string regex_to_reversed_partial_regex(const std::string & pattern);

llama/llama.cpp/common/sampling.cpp

@@ -1,6 +1,7 @@
 #include "sampling.h"
 
 #include "common.h"
+#include "log.h"
 
 #include <cmath>
 #include <unordered_map>
@@ -160,7 +161,7 @@ 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<std::string> patterns_at_start;
+        std::vector<std::string> trigger_patterns;
         std::vector<std::string> patterns_anywhere;
         std::vector<llama_token> trigger_tokens;
         for (const auto & trigger : params.grammar_triggers) {
@@ -172,10 +173,13 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                     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);
+                    patterns_anywhere.push_back(trigger.value);
+                    break;
+                }
+                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL:
+                {
+                    trigger_patterns.push_back(trigger.value);
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN:
@@ -189,10 +193,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
             }
         }
 
-        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]*");
         }
@@ -229,51 +229,48 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                 params.logit_bias.data()));
 
     if (params.mirostat == 0) {
-        if (params.top_n_sigma >= 0) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k        (params.top_k));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp         (params.temp));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma  (params.top_n_sigma));
-        } else {
-            for (const auto & cnstr : params.samplers) {
-                switch (cnstr) {
-                    case COMMON_SAMPLER_TYPE_DRY:
-                        {
-                            std::vector<const char *> c_breakers;
-                            c_breakers.reserve(params.dry_sequence_breakers.size());
-                            for (const auto & str : params.dry_sequence_breakers) {
-                                c_breakers.push_back(str.c_str());
-                            }
-
-                            llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+        for (const auto & cnstr : params.samplers) {
+            switch (cnstr) {
+                case COMMON_SAMPLER_TYPE_DRY:
+                    {
+                        std::vector<const char *> c_breakers;
+                        c_breakers.reserve(params.dry_sequence_breakers.size());
+                        for (const auto & str : params.dry_sequence_breakers) {
+                            c_breakers.push_back(str.c_str());
                         }
-                        break;
-                    case COMMON_SAMPLER_TYPE_TOP_K:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TOP_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_MIN_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_XTC:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                        break;
-                    case COMMON_SAMPLER_TYPE_INFILL:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
-                        break;
-                    case COMMON_SAMPLER_TYPE_PENALTIES:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
-                        break;
-                    default:
-                        GGML_ASSERT(false && "unknown sampler type");
-                }
+
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                    }
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_K:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k       (params.top_k));
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p       (params.top_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_N_SIGMA:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma));
+                    break;
+                case COMMON_SAMPLER_TYPE_MIN_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p       (params.min_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_XTC:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc         (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                    break;
+                case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical     (params.typ_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext    (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                    break;
+                case COMMON_SAMPLER_TYPE_INFILL:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill      (vocab));
+                    break;
+                case COMMON_SAMPLER_TYPE_PENALTIES:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_penalties   (params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    break;
+                default:
+                    GGML_ASSERT(false && "unknown sampler type");
             }
         }
         llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
@@ -475,6 +472,7 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
         case COMMON_SAMPLER_TYPE_TOP_K:       return 'k';
         case COMMON_SAMPLER_TYPE_TYPICAL_P:   return 'y';
         case COMMON_SAMPLER_TYPE_TOP_P:       return 'p';
+        case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return 's';
         case COMMON_SAMPLER_TYPE_MIN_P:       return 'm';
         case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
         case COMMON_SAMPLER_TYPE_XTC:         return 'x';
@@ -490,6 +488,7 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
         case COMMON_SAMPLER_TYPE_TOP_K:       return "top_k";
         case COMMON_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
         case COMMON_SAMPLER_TYPE_TOP_P:       return "top_p";
+        case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return "top_n_sigma";
         case COMMON_SAMPLER_TYPE_MIN_P:       return "min_p";
         case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
         case COMMON_SAMPLER_TYPE_XTC:         return "xtc";
@@ -504,6 +503,7 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
         { "dry",         COMMON_SAMPLER_TYPE_DRY },
         { "top_k",       COMMON_SAMPLER_TYPE_TOP_K },
         { "top_p",       COMMON_SAMPLER_TYPE_TOP_P },
+        { "top_n_sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
         { "typ_p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "min_p",       COMMON_SAMPLER_TYPE_MIN_P },
         { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
@@ -517,6 +517,7 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
     std::unordered_map<std::string, common_sampler_type> sampler_alt_name_map {
         { "top-k",       COMMON_SAMPLER_TYPE_TOP_K },
         { "top-p",       COMMON_SAMPLER_TYPE_TOP_P },
+        { "top-n-sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
         { "nucleus",     COMMON_SAMPLER_TYPE_TOP_P },
         { "typical-p",   COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "typical",     COMMON_SAMPLER_TYPE_TYPICAL_P },
@@ -533,14 +534,16 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
         auto sampler = sampler_canonical_name_map.find(name);
         if (sampler != sampler_canonical_name_map.end()) {
             samplers.push_back(sampler->second);
-        } else {
-            if (allow_alt_names) {
-                sampler = sampler_alt_name_map.find(name);
-                if (sampler != sampler_alt_name_map.end()) {
-                    samplers.push_back(sampler->second);
-                }
+            continue;
+        }
+        if (allow_alt_names) {
+            sampler = sampler_alt_name_map.find(name);
+            if (sampler != sampler_alt_name_map.end()) {
+                samplers.push_back(sampler->second);
+                continue;
             }
         }
+        LOG_WRN("%s: unable to match sampler by name '%s'\n", __func__, name.c_str());
     }
 
     return samplers;
@@ -552,6 +555,7 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K),       COMMON_SAMPLER_TYPE_TOP_K },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P),   COMMON_SAMPLER_TYPE_TYPICAL_P },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P),       COMMON_SAMPLER_TYPE_TOP_P },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_N_SIGMA), COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P),       COMMON_SAMPLER_TYPE_MIN_P },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC),         COMMON_SAMPLER_TYPE_XTC },
@@ -566,6 +570,8 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
         const auto sampler = sampler_name_map.find(c);
         if (sampler != sampler_name_map.end()) {
             samplers.push_back(sampler->second);
+        } else {
+            LOG_WRN("%s: unable to match sampler by char '%c'\n", __func__, c);
         }
     }
 

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

@@ -1,126 +0,0 @@
-#ifndef CLIP_H
-#define CLIP_H
-
-#include "ggml.h"
-#include <stddef.h>
-#include <stdint.h>
-
-#ifdef LLAMA_SHARED
-#    if defined(_WIN32) && !defined(__MINGW32__)
-#        ifdef LLAMA_BUILD
-#            define CLIP_API __declspec(dllexport)
-#        else
-#            define CLIP_API __declspec(dllimport)
-#        endif
-#    else
-#        define CLIP_API __attribute__ ((visibility ("default")))
-#    endif
-#else
-#    define CLIP_API
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-struct clip_ctx;
-
-struct clip_image_size {
-    int width;
-    int height;
-};
-
-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;
-};
-
-// deprecated, use clip_init
-CLIP_API struct clip_ctx * clip_model_load(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_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);
-
-CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
-CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx);
-
-CLIP_API int clip_n_patches        (const struct clip_ctx * ctx);
-CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
-CLIP_API int clip_n_mmproj_embd    (const struct clip_ctx * ctx);
-
-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_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);
-
-// 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);
-
-/** interpret bytes as an image file with length bytes_length, and use the result to populate img */
-CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
-
-/** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
-CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
-
-CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
-
-CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
-CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
-
-CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
-
-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);
-
-CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
-
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // CLIP_H

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

@@ -1,586 +0,0 @@
-#include "clip.h"
-#include "llava.h"
-
-#include "llama.h"
-
-#include <algorithm>
-#include <cerrno>
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <limits>
-#include <vector>
-#include <memory>
-
-#if defined(LLAVA_LOG_OFF)
-#   define LOG_INF(...)
-#   define LOG_WRN(...)
-#   define LOG_ERR(...)
-#   define LOG_DBG(...)
-#else // defined(LLAVA_LOG_OFF)
-#   define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
-#   define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
-#   define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
-#   define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
-#endif // defined(LLAVA_LOG_OFF)
-
-// 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;
-};
-
-struct clip_image_grid_shape {
-    int first;
-    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.
- *
- * @param original_size The original size of the image in the format (width, height).
- * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
- * @return The best fit resolution in the format (width, height).
- */
-static std::pair<int, int> select_best_resolution(const std::pair<int, int>& original_size, const std::vector<std::pair<int, int>>& possible_resolutions) {
-    int original_width  = original_size.first;
-    int original_height = original_size.second;
-
-    std::pair<int, int> best_fit;
-    int max_effective_resolution = 0;
-    int min_wasted_resolution = std::numeric_limits<int>::max();
-
-    for (const auto& resolution : possible_resolutions) {
-        int width = resolution.first;
-        int height = resolution.second;
-        float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
-        int downscaled_width  = static_cast<int>(original_width * scale);
-        int downscaled_height = static_cast<int>(original_height * scale);
-        int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
-        int wasted_resolution = (width * height) - effective_resolution;
-        // LOG_DBG("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
-        if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
-            max_effective_resolution = effective_resolution;
-            min_wasted_resolution = wasted_resolution;
-            best_fit = resolution;
-        }
-    }
-
-    return best_fit;
-}
-
-/**
- * @brief Get the anyres image grid shape object
- *
- * @param image_size
- * @param grid_pinpoints
- * @param image_patch_size
- * @return <int, int>
- */
-static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<int, int> & image_size, const std::vector<std::pair<int, int>> & grid_pinpoints, int image_patch_size) {
-    /**
-        Conversion from gguf flat array to vector:
-        std::vector<std::pair<int, int>> possible_resolutions;
-        for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
-            possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
-        }
-     */
-    auto best_resolution = select_best_resolution(image_size, grid_pinpoints);
-    return {best_resolution.first / image_patch_size, best_resolution.second / image_patch_size};
-}
-
-// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
-static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
-    struct {
-        struct ggml_context * ctx;
-    } model;
-
-    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)
-
-    int num_patches_width  = grid_shape.first;  // grid 1-4
-    int num_patches_height = grid_shape.second; // grid 1-4
-
-    const size_t num_images = num_patches_width * num_patches_height + 1;
-
-    // TODO: size calculation is not calculated - it's only tens of MB
-    size_t ctx_size = 0;
-
-    {
-        ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
-        ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
-    }
-
-    struct ggml_init_params params {
-        /*.mem_size   =*/ ctx_size,
-        /*.mem_buffer =*/ NULL,
-        /*.no_alloc   =*/ false, // NOTE: this should be false when using the legacy API
-    };
-
-    // Python reference code for full unpad:
-    /*
-        base_image_feature = image_feature[0]
-        image_feature = image_feature[1:]
-        image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
-        image_feature = image_feature.flatten(1, 2).flatten(2, 3)
-        image_feature = unpad_image(image_feature, image_sizes[image_idx])
-        image_feature = torch.cat((
-            image_feature,
-            self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1)
-        ), dim=-1)
-        image_feature = image_feature.flatten(1, 2).transpose(0, 1)
-        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
-    */
-    // We now have two options: unpad or no unpad. Unpad removes tokens for faster llm eval.
-    // In terms of result quality it appears to make no difference, so we'll start with the easier approach given 5D tensors are not supported in ggml yet.
-    // Without unpad we have to split the sub-image embeddings into patches of 24 features each and permute them.
-    // Once all images are processed to prepended the base_image_features without any changes.
-
-    // Pytorch reference simplified, modified for ggml compatibility - confirmed identical output in python (for a 2x2 grid image (676x676 scaling))
-    /*
-        image_feature = image_feature.view(2, 2, 24, 24, 4096)
-        image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
-        image_feature = image_feature.view(2, 24, 2, 24, 4096)
-        image_feature = image_feature.flatten(0, 3)
-
-        // Reshape to 4D tensor by merging the last two dimensions
-        image_feature = image_feature.view(2, 2, 24, 24*4096)
-        image_feature = image_feature.permute(0, 2, 1, 3).contiguous()
-        image_feature = image_feature.view(-1, 4096)
-    */
-
-    model.ctx = ggml_init(params);
-
-    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
-    // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
-    // fill it with the image embeddings, ignoring the base
-    for (size_t i = 1; i < num_images; i++) {
-        size_t offset = (i-1) * clip_embd_nbytes(ctx_clip);
-        memcpy((uint8_t *)(image_features->data) + offset, image_embd_v[i], clip_embd_nbytes(ctx_clip));
-    }
-
-    struct ggml_cgraph  * gf = ggml_new_graph(model.ctx);
-    size_t size_ele = ggml_type_size(GGML_TYPE_F32);
-
-    struct ggml_tensor *image_features_patchview = ggml_view_4d(model.ctx, image_features,
-                                                                num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
-                                                                num_patches_per_side,
-                                                                num_patches_width,
-                                                                num_patches_height,
-                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
-                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side,
-                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side * num_patches_width, 0);
-    // ggml_tensor_printf(image_features_patchview,"image_features_patchview",__LINE__,false,false);
-    struct ggml_tensor *permuted_cont = ggml_cont(model.ctx, ggml_permute(model.ctx, image_features_patchview, 0, 2, 1, 3));
-    /**
-     At the end of each row we have to add the row_end embeddings, which are the same as the newline embeddings
-         image_feature = torch.cat((
-        image_feature,
-        self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
-    ), dim=-1)
-     *
-     */
-
-    // ggml_tensor_printf(permuted_cont,"permuted_cont",__LINE__,false,false);
-    struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side,  size_ele * clip_n_mmproj_embd(ctx_clip), 0);
-    // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
-    ggml_build_forward_expand(gf, flatten);
-    ggml_graph_compute_with_ctx(model.ctx, gf, 1);
-    struct ggml_tensor* result = ggml_graph_node(gf, -1);
-
-    memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
-    // append without newline tokens (default behavior in llava_arch when not using unpad ):
-    memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
-    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
-
-    // Debug: Test single segments
-    // Current findings: sending base image, sending a segment embedding all works similar to python
-    // However, permuted embeddings do not work yet (stride issue?)
-    // memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as context
-    // memcpy(image_embd_out, (float*)prepared_cont->data, clip_embd_nbytes(ctx_clip)); // main image as context
-    // *n_img_pos_out=576;
-
-    ggml_free(model.ctx);
-    return true;
-}
-
-static clip_image_f32 * reshape_by_patch(clip_image_f32 * image, int patch_size) {
-    int width = image->nx;
-    int height = image->ny;
-    int num_patches = (height / patch_size) * (width / patch_size);
-    clip_image_f32 * patch = clip_image_f32_init();
-    patch->nx = patch_size * num_patches;
-    patch->ny = patch_size;
-    patch->buf.resize(3 * patch->nx * patch->ny);
-
-    int patch_index = 0;
-
-    for (int i = 0; i < height; i += patch_size) {
-        for (int j = 0; j < width; j += patch_size) {
-            for (int pi = 0; pi < patch_size; ++pi) {
-                for (int pj = 0; pj < patch_size; ++pj) {
-                    int input_index = ((i + pi) * width + (j + pj)) * 3;
-                    int output_index = (pi * patch_size * num_patches + patch_index * patch_size + pj) * 3;
-                    patch->buf[output_index] = image->buf[input_index];
-                    patch->buf[output_index+1] = image->buf[input_index+1];
-                    patch->buf[output_index+2] = image->buf[input_index+2];
-                }
-            }
-            patch_index++;
-        }
-    }
-    return patch;
-}
-
-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_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__);
-        return false;
-    }
-
-    const int64_t t_img_enc_start_us = ggml_time_us();
-
-    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(n_imgs);
-        clip_image_size load_image_size;
-
-        for (size_t i = 0; i < n_imgs; i++) {
-            const int64_t t_img_enc_step_start_us = ggml_time_us();
-            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, image_embd_v[i]);
-            }
-            else {
-                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) 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)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)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, 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);
-    }
-    else if (clip_is_glm(ctx_clip)){
-        struct clip_image_size * load_image_size = clip_image_size_init();
-        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);
-
-        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");
-            return false;
-        }
-    }
-    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);
-        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");
-
-            return false;
-        }
-    }
-    else {
-        // 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(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, 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) 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)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);
-
-        std::vector<std::pair<int, int>> grid_pinpoints;
-        for (size_t i = 0; i < num_gridpoints; i += 2) {
-            grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
-        }
-
-        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);
-
-        int n_img_pos_out;
-        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
-        *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();
-
-        // debug image/segment/normalization content:
-        // clip_image_u8 * tmp = clip_image_u8_init();
-        // clip_image_convert_f32_to_u8(*image_feature, *tmp);
-        // clip_image_save_to_bmp(*tmp, "image_feature.bmp");
-    }
-
-    LOG_INF("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
-
-    const int64_t t_img_enc_end_us = ggml_time_us();
-    float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
-
-    LOG_INF("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
-
-    return true;
-}
-
-bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip) {
-        // make sure that the correct mmproj was used, i.e., compare apples to apples
-    int n_llama_embd = llama_model_n_embd(llama_get_model(ctx_llama));
-    auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
-    if (n_image_embd != n_llama_embd) {
-        LOG_ERR("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
-        return false;
-    }
-    return true;
-}
-
-bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
-    // Granite vision uses up to 10 patches + base patch
-    int num_max_patches = 11;
-    if (clip_is_minicpmv(ctx_clip)) {
-        num_max_patches = 10;
-    }
-    if (clip_is_glm(ctx_clip)) {
-        num_max_patches = 1;
-    }
-    float * image_embd;
-    if (clip_is_qwen2vl(ctx_clip)) {
-        // qwen2vl don't split image into chunks, so `num_max_patches` is not needed.
-        image_embd = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img->nx, img->ny));
-    } else {
-        image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
-    }
-    if (!image_embd) {
-        LOG_ERR("Unable to allocate memory for image embeddings\n");
-        return false;
-    }
-
-    int n_img_pos;
-    if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
-        LOG_ERR("%s: cannot encode image, aborting\n", __func__);
-        free(image_embd);
-        return false;
-    }
-    *image_embd_out = image_embd;
-    *n_img_pos_out = n_img_pos;
-
-    return true;
-}
-
-struct llava_embd_batch {
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id>   seq_id_0;
-    std::vector<llama_seq_id *> seq_ids;
-    std::vector<int8_t>         logits;
-    llama_batch batch;
-    llava_embd_batch(float * embd, int32_t n_embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-        pos     .resize(n_tokens);
-        n_seq_id.resize(n_tokens);
-        seq_ids .resize(n_tokens + 1);
-        logits  .resize(n_tokens);
-        seq_id_0.resize(1);
-        seq_id_0[0] = seq_id;
-        seq_ids [n_tokens] = nullptr;
-        batch = {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ embd,
-            /*n_embd         =*/ n_embd,
-            /*pos            =*/ pos.data(),
-            /*n_seq_id       =*/ n_seq_id.data(),
-            /*seq_id         =*/ seq_ids.data(),
-            /*logits         =*/ logits.data(),
-        };
-        for (int i = 0; i < n_tokens; i++) {
-            batch.pos     [i] = pos_0 + i;
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
-        }
-    }
-};
-
-bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
-    int n_embd  = llama_model_n_embd(llama_get_model(ctx_llama));
-
-    for (int i = 0; i < image_embed->n_image_pos; i += n_batch) {
-        int n_eval = image_embed->n_image_pos - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        float * embd = image_embed->embed+i*n_embd;
-        llava_embd_batch llava_batch = llava_embd_batch(embd, n_embd, n_eval, *n_past, 0);
-        if (llama_decode(ctx_llama, llava_batch.batch)) {
-            LOG_ERR("%s : failed to eval\n", __func__);
-            return false;
-        }
-        *n_past += n_eval;
-    }
-    return true;
-}
-
-struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
-    clip_image_u8 * img = clip_image_u8_init();
-    if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
-        clip_image_u8_free(img);
-        LOG_ERR("%s: can't load image from bytes, is it a valid image?", __func__);
-        return NULL;
-    }
-
-    float* image_embed = NULL;
-    int n_image_pos = 0;
-    bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
-    if (!image_embed_result) {
-        clip_image_u8_free(img);
-        LOG_ERR("%s: couldn't embed the image\n", __func__);
-        return NULL;
-    }
-
-    clip_image_u8_free(img);
-    auto result = (llava_image_embed*)malloc(sizeof(llava_image_embed));
-    result->embed = image_embed;
-    result->n_image_pos = n_image_pos;
-    return result;
-}
-
-static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
-    auto file = fopen(path, "rb");
-    if (file == NULL) {
-        LOG_ERR("%s: can't read file %s\n", __func__, path);
-        return false;
-    }
-
-    fseek(file, 0, SEEK_END);
-    auto fileSize = ftell(file);
-    fseek(file, 0, SEEK_SET);
-
-    auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
-    if (buffer == NULL) {
-        LOG_ERR("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
-        perror("Memory allocation error");
-        fclose(file);
-        return false;
-    }
-    errno = 0;
-    size_t ret = fread(buffer, 1, fileSize, file); // Read the file into the buffer
-    if (ferror(file)) {
-        LOG_ERR("read error: %s", strerror(errno));
-        free(buffer);
-        fclose(file);
-        return false;
-    }
-    if (ret != (size_t) fileSize) {
-        LOG_ERR("unexpectedly reached end of file");
-        free(buffer);
-        fclose(file);
-        return false;
-    }
-    fclose(file); // Close the file
-
-    *bytesOut = buffer;
-    *sizeOut = fileSize;
-    return true;
-}
-
-struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
-    unsigned char* image_bytes;
-    long image_bytes_length;
-    auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
-    if (!loaded) {
-        LOG_ERR("%s: failed to load %s\n", __func__, image_path);
-        return NULL;
-    }
-
-    llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
-    free(image_bytes);
-
-    return embed;
-}
-
-void llava_image_embed_free(struct llava_image_embed * embed) {
-    free(embed->embed);
-    free(embed);
-}

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

@@ -1,6 +0,0 @@
-package llava
-
-// #cgo CXXFLAGS: -std=c++11
-// #cgo CPPFLAGS: -I${SRCDIR}/../../include -I${SRCDIR}/../../common
-// #cgo CPPFLAGS: -I${SRCDIR}/../../../../ml/backend/ggml/ggml/include
-import "C"

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

@@ -1,49 +0,0 @@
-#ifndef LLAVA_H
-#define LLAVA_H
-
-#include "ggml.h"
-
-#ifdef LLAMA_SHARED
-#    if defined(_WIN32) && !defined(__MINGW32__)
-#        ifdef LLAMA_BUILD
-#            define LLAVA_API __declspec(dllexport)
-#        else
-#            define LLAVA_API __declspec(dllimport)
-#        endif
-#    else
-#        define LLAVA_API __attribute__ ((visibility ("default")))
-#    endif
-#else
-#    define LLAVA_API
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-struct clip_ctx;
-struct llava_image_embed {
-    float * embed;
-    int n_image_pos;
-};
-
-/** sanity check for clip <-> llava embed size match */
-LLAVA_API bool llava_validate_embed_size(const struct llama_context * ctx_llama, const struct clip_ctx * ctx_clip);
-
-LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip, int n_threads, const struct clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
-
-/** build an image embed from image file bytes */
-LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
-/** build an image embed from a path to an image filename */
-LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path);
-/** free an embedding made with llava_image_embed_make_* */
-LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
-
-/** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
-LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif

llama/llama.cpp/include/llama.h

@@ -4,6 +4,7 @@
 #include "ggml.h"
 #include "ggml-cpu.h"
 #include "ggml-backend.h"
+#include "ggml-opt.h"
 
 #include <stddef.h>
 #include <stdint.h>
@@ -60,7 +61,10 @@ extern "C" {
     struct llama_model;
     struct llama_context;
     struct llama_sampler;
-    struct llama_kv_cache;
+
+    typedef struct llama_memory_i * llama_memory_t;
+
+    struct llama_kv_cache; // DEPRECATED (use llama_memory instead)
 
     typedef int32_t llama_pos;
     typedef int32_t llama_token;
@@ -111,6 +115,8 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_TRILLION       = 31,
         LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
         LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
+        LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
+        LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 35,
     };
 
     enum llama_rope_type {
@@ -255,11 +261,10 @@ extern "C" {
 
         llama_token  *  token;
         float        *  embd;
-        int32_t         n_embd;
         llama_pos    *  pos;
-        int32_t      *  n_seq_id;
-        llama_seq_id ** seq_id;
-        int8_t       *  logits; // TODO: rename this to "output"
+        int32_t      *  n_seq_id; // TODO: remove, should belong to only 1 sequence
+        llama_seq_id ** seq_id;   // TODO: become llama_seq_id * seq_id;
+        int8_t       *  logits;   // TODO: rename this to "output"
     } llama_batch;
 
     enum llama_model_kv_override_type {
@@ -343,7 +348,7 @@ extern "C" {
         float    yarn_beta_fast;   // YaRN low correction dim
         float    yarn_beta_slow;   // YaRN high correction dim
         uint32_t yarn_orig_ctx;    // YaRN original context size
-        float    defrag_thold;     // defragment the KV cache if holes/size > thold, < 0 disabled (default)
+        float    defrag_thold;     // defragment the KV cache if holes/size > thold, <= 0 disabled (default)
 
         ggml_backend_sched_eval_callback cb_eval;
         void * cb_eval_user_data;
@@ -351,20 +356,21 @@ extern "C" {
         enum ggml_type type_k; // data type for K cache [EXPERIMENTAL]
         enum ggml_type type_v; // data type for V cache [EXPERIMENTAL]
 
-        // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
-        // TODO: move at the end of the struct
-        bool logits_all;  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
-        bool embeddings;  // if true, extract embeddings (together with logits)
-        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
-        bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
-        bool no_perf;     // whether to measure performance timings
-        bool cross_attn;  // whether to use cross attention
-
         // Abort callback
         // if it returns true, execution of llama_decode() will be aborted
         // currently works only with CPU execution
         ggml_abort_callback abort_callback;
         void *              abort_callback_data;
+
+        // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
+        bool embeddings;  // if true, extract embeddings (together with logits)
+        bool offload_kqv; // offload the KQV ops (including the KV cache) to GPU
+        bool flash_attn;  // use flash attention [EXPERIMENTAL]
+        bool no_perf;     // measure performance timings
+        bool op_offload;  // offload host tensor operations to device
+        bool swa_full;    // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+                          // NOTE: setting to false when n_seq_max > 1 can cause bad performance in some cases
+                          //       ref: https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573
     };
 
     // model quantization parameters
@@ -446,6 +452,10 @@ extern "C" {
                                  size_t    n_paths,
               struct llama_model_params    params);
 
+    LLAMA_API void llama_model_save_to_file(
+            const struct llama_model * model,
+                        const char * path_model);
+
     DEPRECATED(LLAMA_API void llama_free_model(struct llama_model * model),
             "use llama_model_free instead");
 
@@ -460,16 +470,13 @@ extern "C" {
             struct llama_context_params   params),
             "use llama_init_from_model instead");
 
-    // TODO (jmorganca): this should most likely be passed in as part of a batch
-    // and not set on the context for all batches.
-    LLAMA_API void llama_set_cross_attention(struct llama_context * ctx, bool cross_attn_state);
-
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 
     LLAMA_API int64_t llama_time_us(void);
 
     LLAMA_API size_t llama_max_devices(void);
+    LLAMA_API size_t llama_max_parallel_sequences(void);
 
     LLAMA_API bool llama_supports_mmap       (void);
     LLAMA_API bool llama_supports_mlock      (void);
@@ -489,9 +496,11 @@ 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    struct llama_kv_cache * llama_get_kv_self (      struct llama_context * ctx);
+    LLAMA_API           llama_memory_t   llama_get_memory  (const 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
 
+    DEPRECATED(LLAMA_API struct llama_kv_cache * llama_get_kv_self(struct llama_context * ctx), "use llama_get_memory instead");
+
     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);
 
@@ -500,6 +509,7 @@ extern "C" {
     LLAMA_API int32_t llama_model_n_layer    (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head_kv  (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_swa      (const struct llama_model * model);
 
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model);
@@ -604,78 +614,92 @@ extern "C" {
                          int32_t   il_end);
 
     //
-    // KV cache
+    // Memory
     //
 
-    // TODO: start using struct llama_kv_cache
+    // Clear the memory contents
+    LLAMA_API void llama_memory_clear(llama_memory_t mem);
 
-    // Information associated with an individual cell in the KV cache view.
-    struct llama_kv_cache_view_cell {
-        // The position for this cell. Takes KV cache shifts into account.
-        // May be negative if the cell is not populated.
-        llama_pos pos;
-    };
+    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
+    // seq_id < 0 : match any sequence
+    // p0 < 0     : [0,  p1]
+    // p1 < 0     : [p0, inf)
+    LLAMA_API bool llama_memory_seq_rm(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1);
 
-    // An updateable view of the KV cache.
-    struct llama_kv_cache_view {
-        // Number of KV cache cells. This will be the same as the context size.
-        int32_t n_cells;
+    // Copy all tokens that belong to the specified sequence to another sequence
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_cp(
+            llama_memory_t mem,
+              llama_seq_id seq_id_src,
+              llama_seq_id seq_id_dst,
+                 llama_pos p0,
+                 llama_pos p1);
 
-        // Maximum number of sequences that can exist in a cell. It's not an error
-        // if there are more sequences in a cell than this value, however they will
-        // not be visible in the view cells_sequences.
-        int32_t n_seq_max;
+    // Removes all tokens that do not belong to the specified sequence
+    LLAMA_API void llama_memory_seq_keep(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
 
-        // Number of tokens in the cache. For example, if there are two populated
-        // cells, the first with 1 sequence id in it and the second with 2 sequence
-        // ids then you'll have 3 tokens.
-        int32_t token_count;
+    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_add(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1,
+                 llama_pos delta);
 
-        // Number of populated cache cells.
-        int32_t used_cells;
+    // Integer division of the positions by factor of `d > 1`
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_div(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1,
+                       int d);
 
-        // Maximum contiguous empty slots in the cache.
-        int32_t max_contiguous;
+    // Returns the smallest position present in the memory for the specified sequence
+    // This is typically non-zero only for SWA caches
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_memory_seq_pos_min(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
 
-        // Index to the start of the max_contiguous slot range. Can be negative
-        // when cache is full.
-        int32_t max_contiguous_idx;
+    // Returns the largest position present in the memory for the specified sequence
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_memory_seq_pos_max(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
 
-        // Information for an individual cell.
-        struct llama_kv_cache_view_cell * cells;
+    // Check if the memory supports shifting
+    LLAMA_API bool llama_memory_can_shift(llama_memory_t mem);
 
-        // The sequences for each cell. There will be n_seq_max items per cell.
-        llama_seq_id * cells_sequences;
-    };
-
-    // Create an empty KV cache view. (use only for debugging purposes)
-    LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max);
-
-    // Free a KV cache view. (use only for debugging purposes)
-    LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view);
-
-    // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)
-    // TODO: change signature to llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct llama_context * ctx)
-    LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view);
-
-    ///
+    //
+    // KV cache for self-attention (TODO: deprecate in favor of llama_memory)
+    //
 
     // 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_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");
+    DEPRECATED(LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx),
+               "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)");
 
     // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
-    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");
+    DEPRECATED(LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx),
+               "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)");
 
     // Clear the KV cache - both cell info is erased and KV data is zeroed
     LLAMA_API void llama_kv_self_clear(
-            struct llama_context * ctx);
+                struct llama_context * ctx);
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
     // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
@@ -707,7 +731,6 @@ extern "C" {
     // 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_self_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
     LLAMA_API void llama_kv_self_seq_add(
@@ -720,7 +743,6 @@ 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_self_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
     LLAMA_API void llama_kv_self_seq_div(
@@ -730,84 +752,40 @@ extern "C" {
                        llama_pos   p1,
                              int   d);
 
+    // Returns the smallest position present in the KV cache for the specified sequence
+    // This is typically non-zero only for SWA caches
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_kv_self_seq_pos_min(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id);
+
     // Returns the largest position present in the KV cache for the specified sequence
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
+    // Return -1 if the sequence is empty
     LLAMA_API llama_pos llama_kv_self_seq_pos_max(
             struct llama_context * ctx,
-                     llama_seq_id   seq_id);
+                    llama_seq_id   seq_id);
 
     // Defragment the KV cache
     // This will be applied:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_self_update()
-    LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx);
+    DEPRECATED(LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx),
+            "simply remove this call, the context will automatically decide when to do a defragmentation based on 'defrag_thold'");
 
     // Check if the context supports KV cache shifting
     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");
-
+    DEPRECATED(LLAMA_API void llama_kv_self_update(struct llama_context * ctx),
+            "simply remove this call, updates are applied lazily on the next llama_decode()");
 
     //
     // State / sessions
     //
 
     // Returns the *actual* size in bytes of the state
-    // (logits, embedding and kv_cache)
+    // (logits, embedding and memory)
     // Only use when saving the state, not when restoring it, otherwise the size may be too small.
     LLAMA_API size_t llama_state_get_size(struct llama_context * ctx);
     LLAMA_API DEPRECATED(size_t llama_get_state_size(struct llama_context * ctx),
@@ -863,12 +841,12 @@ extern "C" {
                           size_t   n_token_count),
         "use llama_state_save_file instead");
 
-    // Get the exact size needed to copy the KV cache of a single sequence
+    // Get the exact size needed to copy the state of a single sequence
     LLAMA_API size_t llama_state_seq_get_size(
             struct llama_context * ctx,
                     llama_seq_id   seq_id);
 
-    // Copy the KV cache of a single sequence into the specified buffer
+    // Copy the state of a single sequence into the specified buffer
     LLAMA_API size_t llama_state_seq_get_data(
             struct llama_context * ctx,
                          uint8_t * dst,
@@ -929,18 +907,26 @@ extern "C" {
     // Frees a batch of tokens allocated with llama_batch_init()
     LLAMA_API void llama_batch_free(struct llama_batch batch);
 
-    // Processes a batch of tokens with the ecoder part of the encoder-decoder model.
-    // Stores the encoder output internally for later use by the decoder cross-attention layers.
+    // Process a batch of tokens.
+    // In contrast to llama_decode() - this call does not use KV cache.
+    // For encode-decoder contexts, processes the batch using the encoder.
+    // Can store the encoder output internally for later use by the decoder's cross-attention layers.
     //   0 - success
-    // < 0 - error. the KV cache state is restored to the state before this call
+    // < 0 - error. the memory state is restored to the state before this call
     LLAMA_API int32_t llama_encode(
             struct llama_context * ctx,
               struct llama_batch   batch);
 
+    // Process a batch of tokens.
+    // Requires the context to have a memory.
+    // For encode-decoder contexts, processes the batch using the decoder.
     // Positive return values does not mean a fatal error, but rather a warning.
-    //   0 - success
-    //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
-    // < 0 - error. the KV cache state is restored to the state before this call
+    // Upon non-zero return values, the memory state is restored to the state before this call
+    //    0 - success
+    //    1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
+    //    2 - aborted
+    //   -1 - invalid input batch
+    // < -1 - error
     LLAMA_API int32_t llama_decode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -1237,6 +1223,7 @@ extern "C" {
         "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    /// Setting k <= 0 makes this a noop
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
@@ -1432,6 +1419,37 @@ extern "C" {
     LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
     LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
 
+    //
+    // training
+    //
+
+    // function that returns whether or not a given tensor contains trainable parameters
+    typedef bool (*llama_opt_param_filter)(const struct ggml_tensor * tensor, void * userdata);
+
+    // always returns true
+    LLAMA_API bool llama_opt_param_filter_all(const struct ggml_tensor * tensor, void * userdata);
+
+    struct llama_opt_params {
+        uint32_t n_ctx_train; // assumed context size post training, use context size specified in llama_context if 0
+
+        llama_opt_param_filter param_filter; // callback for determining which tensors contain trainable parameters
+        void * param_filter_ud;              // userdata for determining which tensors contain trainable parameters
+
+        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
+        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+    };
+
+    LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params);
+
+    LLAMA_API void llama_opt_epoch(
+            struct llama_context    * lctx,
+            ggml_opt_dataset_t        dataset,
+            ggml_opt_result_t         result_train,
+            ggml_opt_result_t         result_eval,
+            int64_t                   idata_split,
+            ggml_opt_epoch_callback   callback_train,
+            ggml_opt_epoch_callback   callback_eval);
+
 #ifdef __cplusplus
 }
 #endif

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

@@ -253,6 +253,9 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
     std::vector<ggml_backend_buffer_type_t> buft_extra;
     {
         auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (!cpu_dev) {
+            throw std::runtime_error(format("%s: no CPU backend found", __func__));
+        }
         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)
@@ -291,6 +294,9 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
                 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);
+                if (!cpu_dev) {
+                    throw std::runtime_error(format("%s: no CPU backend found", __func__));
+                }
                 buft = ggml_backend_dev_buffer_type(cpu_dev);
 
                 break;

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

@@ -6,7 +6,6 @@
 
 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"           },
@@ -20,6 +19,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_REFACT,           "refact"           },
     { LLM_ARCH_BERT,             "bert"             },
     { LLM_ARCH_NOMIC_BERT,       "nomic-bert"       },
+    { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
     { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
     { LLM_ARCH_BLOOM,            "bloom"            },
     { LLM_ARCH_STABLELM,         "stablelm"         },
@@ -73,7 +73,6 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
     { LLM_ARCH_PLM,              "plm"              },
     { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
-    { LLM_ARCH_MISTRAL3,         "mistral3"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -109,6 +108,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
     { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
     { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
+    { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
@@ -144,7 +144,6 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { 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"             },
 
@@ -177,6 +176,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_CONVNEXT_EMBEDDING_LENGTH, "%s.convnext.embedding_length" },
     { LLM_KV_CONVNEXT_BLOCK_COUNT,      "%s.convnext.block_count"      },
 
+    { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
+
     { LLM_KV_TOKENIZER_MODEL,                "tokenizer.ggml.model"                    },
     { LLM_KV_TOKENIZER_PRE,                  "tokenizer.ggml.pre"                      },
     { LLM_KV_TOKENIZER_LIST,                 "tokenizer.ggml.tokens"                   },
@@ -274,40 +275,6 @@ 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_MLLAMA,
-        {
-            { 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_CROSS_ATTN_K_NORM,    "blk.%d.cross_attn_k_norm" },
-            { LLM_TENSOR_CROSS_ATTN_K_PROJ,    "blk.%d.cross_attn_k_proj" },
-            { LLM_TENSOR_CROSS_ATTN_O_PROJ,    "blk.%d.cross_attn_o_proj" },
-            { LLM_TENSOR_CROSS_ATTN_Q_NORM,    "blk.%d.cross_attn_q_norm" },
-            { LLM_TENSOR_CROSS_ATTN_Q_PROJ,    "blk.%d.cross_attn_q_proj" },
-            { LLM_TENSOR_CROSS_ATTN_V_PROJ,    "blk.%d.cross_attn_v_proj" },
-            { LLM_TENSOR_CROSS_ATTN_ATTN_GATE, "blk.%d.cross_attn_attn_gate" },
-            { LLM_TENSOR_CROSS_ATTN_MLP_GATE,  "blk.%d.cross_attn_mlp_gate" },
-        },
-    },
     {
         LLM_ARCH_DECI,
         {
@@ -485,6 +452,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_TOKEN_TYPES,     "token_types" },
             { LLM_TENSOR_POS_EMBD,        "position_embd" },
             { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
             { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
             { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
             { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
@@ -511,6 +479,24 @@ 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_NOMIC_BERT_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
+            { LLM_TENSOR_TOKEN_TYPES,     "token_types" },
+            { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_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_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_JINA_BERT_V2,
         {
@@ -1500,6 +1486,9 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { 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" },
         },
     },
     {
@@ -1587,22 +1576,6 @@ 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_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,
         {
@@ -1734,14 +1707,6 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     // this tensor is loaded for T5, but never used
     {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
     {LLM_TENSOR_BSKCN_TV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
-    {LLM_TENSOR_CROSS_ATTN_K_NORM,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
-    {LLM_TENSOR_CROSS_ATTN_K_PROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_CROSS_ATTN_O_PROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_CROSS_ATTN_Q_NORM,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
-    {LLM_TENSOR_CROSS_ATTN_Q_PROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_CROSS_ATTN_V_PROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_CROSS_ATTN_ATTN_GATE,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
-    {LLM_TENSOR_CROSS_ATTN_MLP_GATE,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_CONV1D,                     {LLM_TENSOR_LAYER_INPUT,     GGML_OP_IM2COL}},
     {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},

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

@@ -11,7 +11,6 @@
 enum llm_arch {
     LLM_ARCH_LLAMA,
     LLM_ARCH_LLAMA4,
-    LLM_ARCH_MLLAMA,
     LLM_ARCH_DECI,
     LLM_ARCH_FALCON,
     LLM_ARCH_BAICHUAN,
@@ -24,6 +23,7 @@ enum llm_arch {
     LLM_ARCH_REFACT,
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
+    LLM_ARCH_NOMIC_BERT_MOE,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
@@ -75,7 +75,6 @@ enum llm_arch {
     LLM_ARCH_CHAMELEON,
     LLM_ARCH_SOLAR,
     LLM_ARCH_WAVTOKENIZER_DEC,
-    LLM_ARCH_MISTRAL3,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
     LLM_ARCH_UNKNOWN,
@@ -113,6 +112,7 @@ enum llm_kv {
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,
+    LLM_KV_MOE_EVERY_N_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
@@ -148,7 +148,6 @@ enum llm_kv {
     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,
 
@@ -216,6 +215,8 @@ enum llm_kv {
     LLM_KV_CONVNEXT_EMBEDDING_LENGTH,
     LLM_KV_CONVNEXT_BLOCK_COUNT,
 
+    LLM_KV_CLASSIFIER_OUTPUT_LABELS,
+
     // deprecated:
     LLM_KV_TOKENIZER_PREFIX_ID,
     LLM_KV_TOKENIZER_SUFFIX_ID,
@@ -350,14 +351,6 @@ enum llm_tensor {
     LLM_TENSOR_CLS,
     LLM_TENSOR_CLS_OUT,
     LLM_TENSOR_BSKCN_TV,
-    LLM_TENSOR_CROSS_ATTN_K_NORM,
-    LLM_TENSOR_CROSS_ATTN_K_PROJ,
-    LLM_TENSOR_CROSS_ATTN_O_PROJ,
-    LLM_TENSOR_CROSS_ATTN_Q_NORM,
-    LLM_TENSOR_CROSS_ATTN_Q_PROJ,
-    LLM_TENSOR_CROSS_ATTN_V_PROJ,
-    LLM_TENSOR_CROSS_ATTN_ATTN_GATE,
-    LLM_TENSOR_CROSS_ATTN_MLP_GATE,
     LLM_TENSOR_CONV1D,
     LLM_TENSOR_CONVNEXT_DW,
     LLM_TENSOR_CONVNEXT_NORM,

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

@@ -1,5 +1,6 @@
 #include "llama-batch.h"
 
+#include <cassert>
 #include <cstring>
 #include <algorithm>
 
@@ -14,24 +15,31 @@ llama_ubatch llama_sbatch::reserve_ubatch(size_t n_ubatch, bool has_embd) {
             break;
         }
     }
-    ubatch_token.resize(!has_embd ? n_ubatch : 0);
-    ubatch_embd.resize(has_embd ? n_embd * n_ubatch : 0);
-    ubatch_pos.resize(n_ubatch);
-    ubatch_n_seq_id.resize(n_ubatch);
-    ubatch_seq_id.resize(n_ubatch);
-    ubatch_output.resize(n_ubatch);
+
+    udatas.push_back({});
+
+    auto & udata = udatas.back();
+
+    udata.token.resize(!has_embd ? n_ubatch : 0);
+    udata.embd.resize(has_embd ? n_embd * n_ubatch : 0);
+    udata.pos.resize(n_ubatch);
+    udata.n_seq_id.resize(n_ubatch);
+    udata.seq_id.resize(n_ubatch);
+    udata.output.resize(n_ubatch);
+
     llama_ubatch ubatch = {
         /*equal_seqs   =*/ true,
         /*n_tokens     =*/ 0,
         /*n_seq_tokens =*/ 0,
         /*n_seqs       =*/ 0,
-        /*token        =*/ !has_embd ? ubatch_token.data() : nullptr,
-        /*embd         =*/ has_embd  ? ubatch_embd.data()  : nullptr,
-        /*pos          =*/ ubatch_pos.data(),
-        /*n_seq_id     =*/ ubatch_n_seq_id.data(),
-        /*seq_id       =*/ ubatch_seq_id.data(),
-        /*output       =*/ ubatch_output.data(),
+        /*token        =*/ !has_embd ? udata.token.data() : nullptr,
+        /*embd         =*/ has_embd  ? udata.embd.data()  : nullptr,
+        /*pos          =*/ udata.pos.data(),
+        /*n_seq_id     =*/ udata.n_seq_id.data(),
+        /*seq_id       =*/ udata.seq_id.data(),
+        /*output       =*/ udata.output.data(),
     };
+
     return ubatch;
 }
 
@@ -189,7 +197,7 @@ llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
     return ubatch;
 }
 
-void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
+llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
     GGML_ASSERT(batch.n_tokens >= 0);
     this->batch = &batch;
     this->n_embd = n_embd;
@@ -203,6 +211,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
     for (size_t i = 0; i < n_tokens; ++i) {
         ids[i] = i;
     }
+
     if (simple_split) {
         seq.resize(1);
         llama_sbatch_seq & s = seq[0];
@@ -212,6 +221,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
         s.length = n_tokens;
         return;
     }
+
     std::sort(ids.begin(), ids.end(),
             [&batch](size_t a, size_t b) {
                 int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1;
@@ -239,6 +249,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
                 return n_seq_a > n_seq_b;
             }
     );
+
     // init seq
     llama_sbatch_seq * last_seq = nullptr;
 
@@ -262,6 +273,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
         seq.push_back(new_seq);
         last_seq = &seq.back();
     }
+
     // keep shared prompts first at the end, then sort by length descending.
     std::sort(seq.begin(), seq.end(),
             [](llama_sbatch_seq & a, llama_sbatch_seq & b) {
@@ -277,9 +289,10 @@ llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0
     batch = in_batch;
     GGML_ASSERT(batch.n_tokens > 0);
     if (!batch.pos) {
+        assert(p0 >= 0);
         pos.resize(batch.n_tokens);
         for (int32_t i = 0; i < batch.n_tokens; i++) {
-            pos[i] = i + p0;
+            pos[i] = p0 + i;
         }
         batch.pos = pos.data();
     }
@@ -316,7 +329,6 @@ struct llama_batch llama_batch_get_one(
         /*n_tokens       =*/ n_tokens,
         /*tokens         =*/ tokens,
         /*embd           =*/ nullptr,
-        /*n_embd         =*/ 0,
         /*pos            =*/ nullptr,
         /*n_seq_id       =*/ nullptr,
         /*seq_id         =*/ nullptr,
@@ -329,7 +341,6 @@ struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_
         /*n_tokens       =*/ 0,
         /*tokens         =*/ nullptr,
         /*embd           =*/ nullptr,
-        /*n_embd         =*/ 0,
         /*pos            =*/ nullptr,
         /*n_seq_id       =*/ nullptr,
         /*seq_id         =*/ nullptr,
@@ -338,7 +349,6 @@ struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_
 
     if (embd) {
         batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd);
-        batch.n_embd = embd;
     } else {
         batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
     }

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

@@ -11,15 +11,15 @@ struct llama_ubatch {
     bool equal_seqs;
     // TODO: whole_seqs for embeddings?
 
-    uint32_t n_tokens; // total tokens (n_seq_tokens * n_seqs)
+    uint32_t n_tokens;     // total tokens (n_seq_tokens * n_seqs)
     uint32_t n_seq_tokens; // tokens per sequence
     uint32_t n_seqs;
 
     llama_token  *  token;    // [n_tokens]
     float        *  embd;     // [n_embd, n_tokens]
     llama_pos    *  pos;      // [n_tokens]
-    int32_t      *  n_seq_id; // [n_seqs]
-    llama_seq_id ** seq_id;   // [n_seqs]
+    int32_t      *  n_seq_id; // [n_seqs] // TODO: remove, should belong to only 1 sequence
+    llama_seq_id ** seq_id;   // [n_seqs] // TODO: become llama_seq_id * seq_id;
     int8_t       *  output;   // [n_tokens]
 };
 
@@ -49,13 +49,18 @@ struct llama_sbatch {
 
     const llama_batch * batch = nullptr;
 
-    // buffers for the ubatch
-    std::vector<llama_token>    ubatch_token;
-    std::vector<float>          ubatch_embd;
-    std::vector<llama_pos>      ubatch_pos;
-    std::vector<int32_t>        ubatch_n_seq_id;
-    std::vector<llama_seq_id *> ubatch_seq_id;
-    std::vector<int8_t>         ubatch_output;
+    // buffers for the ubatches
+    // TODO: very hacky, this needs a complete rework
+    struct ubatch_data {
+        std::vector<llama_token>    token;
+        std::vector<float>          embd;
+        std::vector<llama_pos>      pos;
+        std::vector<int32_t>        n_seq_id;
+        std::vector<llama_seq_id *> seq_id;
+        std::vector<int8_t>         output;
+    };
+
+    std::vector<ubatch_data> udatas;
 
     llama_ubatch reserve_ubatch(size_t n_ubatch, bool has_embd = false);
 
@@ -70,7 +75,8 @@ struct llama_sbatch {
     // sequence-wise split
     llama_ubatch split_seq(size_t n_ubatch);
 
-    void from_batch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
+    llama_sbatch() = default;
+    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
 };
 
 // temporary allocate memory for the input batch if needed

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

@@ -35,6 +35,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "mistral-v3",        LLM_CHAT_TEMPLATE_MISTRAL_V3        },
     { "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN },
     { "mistral-v7",        LLM_CHAT_TEMPLATE_MISTRAL_V7        },
+    { "mistral-v7-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN },
     { "phi3",              LLM_CHAT_TEMPLATE_PHI_3             },
     { "phi4",              LLM_CHAT_TEMPLATE_PHI_4             },
     { "falcon3",           LLM_CHAT_TEMPLATE_FALCON_3          },
@@ -50,8 +51,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "deepseek3",         LLM_CHAT_TEMPLATE_DEEPSEEK_3        },
     { "command-r",         LLM_CHAT_TEMPLATE_COMMAND_R         },
     { "llama3",            LLM_CHAT_TEMPLATE_LLAMA_3           },
-    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGML_3         },
-    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGML_4         },
+    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGLM_3         },
+    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGLM_4         },
     { "glmedge",           LLM_CHAT_TEMPLATE_GLMEDGE           },
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
@@ -62,6 +63,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
     { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
+    { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -81,7 +83,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     if (tmpl_contains("<|im_start|>")) {
         return tmpl_contains("<|im_sep|>")
             ? LLM_CHAT_TEMPLATE_PHI_4
-            : LLM_CHAT_TEMPLATE_CHATML;
+            : tmpl_contains("<end_of_utterance>")
+                ? LLM_CHAT_TEMPLATE_SMOLVLM // SmolVLM uses <|im_start|> as BOS, but it is NOT chatml
+                : LLM_CHAT_TEMPLATE_CHATML;
     } else if (tmpl.find("mistral") == 0 || tmpl_contains("[INST]")) {
         if (tmpl_contains("[SYSTEM_PROMPT]")) {
             return LLM_CHAT_TEMPLATE_MISTRAL_V7;
@@ -119,8 +123,12 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         }
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
         return LLM_CHAT_TEMPLATE_PHI_3;
+    } else if (tmpl_contains("[gMASK]<sop>")) {
+        return LLM_CHAT_TEMPLATE_CHATGLM_4;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
         return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
+    } else if (tmpl_contains("<|{{ item['role'] }}|>") && tmpl_contains("<|begin_of_image|>")) {
+        return LLM_CHAT_TEMPLATE_GLMEDGE;
     } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
         return LLM_CHAT_TEMPLATE_ZEPHYR;
     } else if (tmpl_contains("bos_token + message['role']")) {
@@ -149,9 +157,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_LLAMA_3;
     } else if (tmpl_contains("[gMASK]sop")) {
         // chatglm3-6b
-        return LLM_CHAT_TEMPLATE_CHATGML_3;
-    } else if (tmpl_contains("[gMASK]<sop>")) {
-        return LLM_CHAT_TEMPLATE_CHATGML_4;
+        return LLM_CHAT_TEMPLATE_CHATGLM_3;
     } else if (tmpl_contains(LU8("<用户>"))) {
         // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
         return LLM_CHAT_TEMPLATE_MINICPM;
@@ -197,19 +203,20 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|im_start|>assistant\n";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7 || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN) {
         // Official mistral 'v7' template
         // See: https://huggingface.co/mistralai/Mistral-Large-Instruct-2411#basic-instruct-template-v7
+        //      https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503#basic-instruct-template-v7-tekken
+        const char * trailing_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7 ? " " : "";
         for (auto message : chat) {
             std::string role(message->role);
             std::string content(message->content);
             if (role == "system") {
-                ss << "[SYSTEM_PROMPT] " << content << "[/SYSTEM_PROMPT]";
+                ss << "[SYSTEM_PROMPT]" << trailing_space << content << "[/SYSTEM_PROMPT]";
             } else if (role == "user") {
-                ss << "[INST] " << content << "[/INST]";
-            }
-            else {
-                ss << " " << content << "</s>";
+                ss << "[INST]" << trailing_space << content << "[/INST]";
+            } else {
+                ss << trailing_space << content << "</s>";
             }
         }
     } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1
@@ -432,7 +439,7 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_3) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_3) {
         // chatglm3-6b
         ss << "[gMASK]" << "sop";
         for (auto message : chat) {
@@ -442,14 +449,14 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_4) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4) {
         ss << "[gMASK]" << "<sop>";
         for (auto message : chat) {
             std::string role(message->role);
             ss << "<|" << role << "|>" << "\n" << message->content;
         }
         if (add_ass) {
-            ss << "<|assistant|>";
+            ss << "<|assistant|>\n";
         }
     } else if (tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
         for (auto message : chat) {
@@ -620,7 +627,23 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|header_start|>assistant<|header_end|>\n\n";
         }
-    }  else {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_SMOLVLM) {
+        // SmolVLM
+        ss << "<|im_start|>"; // uses <|im_start|> as BOS, but the actual content is NOT chatml
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content << "\n\n";
+            } else if (role == "user") {
+                ss << "User: " << message->content << "<end_of_utterance>\n";
+            } else {
+                ss << "Assistant: " << message->content << "<end_of_utterance>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "Assistant:";
+        }
+    } else {
         // template not supported
         return -1;
     }

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

@@ -14,6 +14,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_MISTRAL_V3,
     LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN,
     LLM_CHAT_TEMPLATE_MISTRAL_V7,
+    LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN,
     LLM_CHAT_TEMPLATE_PHI_3,
     LLM_CHAT_TEMPLATE_PHI_4,
     LLM_CHAT_TEMPLATE_FALCON_3,
@@ -29,8 +30,8 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_DEEPSEEK_3,
     LLM_CHAT_TEMPLATE_COMMAND_R,
     LLM_CHAT_TEMPLATE_LLAMA_3,
-    LLM_CHAT_TEMPLATE_CHATGML_3,
-    LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_CHATGLM_3,
+    LLM_CHAT_TEMPLATE_CHATGLM_4,
     LLM_CHAT_TEMPLATE_GLMEDGE,
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
@@ -41,6 +42,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_YANDEX,
     LLM_CHAT_TEMPLATE_BAILING,
     LLM_CHAT_TEMPLATE_LLAMA4,
+    LLM_CHAT_TEMPLATE_SMOLVLM,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

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

@@ -5,19 +5,21 @@
 #include "llama-cparams.h"
 #include "llama-graph.h"
 #include "llama-adapter.h"
-#include "llama-kv-cache.h"
 
 #include "ggml-cpp.h"
+#include "ggml-opt.h"
 
 #include <map>
 #include <vector>
 
 struct llama_model;
-struct llama_kv_cache;
 
 class llama_io_read_i;
 class llama_io_write_i;
 
+struct llama_memory_i;
+struct llama_memory_state_i;
+
 struct llama_context {
     // init scheduler and compute buffers, reserve worst-case graphs
     llama_context(
@@ -28,7 +30,12 @@ struct llama_context {
 
     void synchronize();
 
-    const llama_model & get_model() const;
+    const llama_model   & get_model()   const;
+    const llama_cparams & get_cparams() const;
+
+    ggml_backend_sched_t get_sched() const;
+
+    ggml_context * get_ctx_compute() const;
 
     uint32_t n_ctx()         const;
     uint32_t n_ctx_per_seq() const;
@@ -39,10 +46,12 @@ struct llama_context {
     uint32_t n_threads()       const;
     uint32_t n_threads_batch() const;
 
-          llama_kv_cache * get_kv_self();
-    const llama_kv_cache * get_kv_self() const;
+    llama_memory_t get_memory() const;
 
-    void kv_self_update();
+    // return true of the KV cache was updated
+    // TODO: remove
+    bool kv_self_update(bool optimize);
+    void kv_self_defrag_sched();
 
     enum llama_pooling_type pooling_type() const;
 
@@ -66,7 +75,6 @@ struct llama_context {
     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,
@@ -84,6 +92,16 @@ struct llama_context {
                 int32_t   il_start,
                 int32_t   il_end);
 
+    // process a single ubatch with a specific graph type
+    // if memory_state is provided, it will be applied first to the context's memory
+    // ret contains the status of the graph computation
+    // returns nullptr only if ret != GGML_STATUS_SUCCESS
+    llm_graph_result_ptr process_ubatch(
+              const llama_ubatch & ubatch,
+                  llm_graph_type   gtype,
+            llama_memory_state_i * mstate,
+                     ggml_status & ret);
+
     int encode(llama_batch & inp_batch);
     int decode(llama_batch & inp_batch);
 
@@ -130,6 +148,32 @@ struct llama_context {
     llama_perf_context_data perf_get_data() const;
     void perf_reset();
 
+    //
+    // training
+    //
+
+    void opt_init(struct llama_model * model, struct llama_opt_params lopt_params);
+
+    void opt_epoch(
+            ggml_opt_dataset_t      dataset,
+            ggml_opt_result_t       result_train,
+            ggml_opt_result_t       result_eval,
+            int64_t                 idata_split,
+            ggml_opt_epoch_callback callback_train,
+            ggml_opt_epoch_callback callback_eval);
+
+    void opt_epoch_iter(
+            ggml_opt_dataset_t               dataset,
+            ggml_opt_result_t                result,
+            const std::vector<llama_token> & tokens,
+            const std::vector<llama_token> & labels_sparse,
+            llama_batch                    & batch,
+            ggml_opt_epoch_callback          callback,
+            bool                             train,
+            int64_t                          idata_in_loop,
+            int64_t                          ndata_in_loop,
+            int64_t                          t_loop_start);
+
 private:
     //
     // output
@@ -139,51 +183,32 @@ private:
     // 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
     //
 
+public:
     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);
+    ggml_status graph_compute(ggml_cgraph * gf, bool batched);
+
+    // reserve a graph with a dummy ubatch of the specified size
+    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_state_i * mstate);
+
+private:
+    llm_graph_result_ptr graph_build(
+                    ggml_context * ctx,
+                     ggml_cgraph * gf,
+              const llama_ubatch & ubatch,
+                  llm_graph_type   gtype,
+      const llama_memory_state_i * mstate);
 
     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);
@@ -200,14 +225,13 @@ private:
     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;
+    std::unique_ptr<llama_memory_i> memory;
 
-    // TODO: remove
-    bool logits_all = false;
+    // TODO: temporary, until the llama_kv_self_defrag() API is removed
+    bool memory_force_optimize = false;
 
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     size_t  logits_size = 0; // capacity (of floats) for logits
@@ -234,6 +258,9 @@ private:
 
     ggml_context_ptr ctx_compute;
 
+    // training
+    ggml_opt_context_t opt_ctx = nullptr;
+
     ggml_threadpool_t threadpool       = nullptr;
     ggml_threadpool_t threadpool_batch = nullptr;
 

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

@@ -1 +1,5 @@
 #include "llama-cparams.h"
+
+size_t llama_max_parallel_sequences(void) {
+    return LLAMA_MAX_PARALLEL_SEQUENCES;
+}

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

@@ -4,6 +4,8 @@
 
 #include <cstdint>
 
+#define LLAMA_MAX_PARALLEL_SEQUENCES 64
+
 struct llama_cparams {
     uint32_t n_ctx;           // context size used during inference
     uint32_t n_batch;
@@ -29,8 +31,8 @@ struct llama_cparams {
     bool offload_kqv;
     bool flash_attn;
     bool no_perf;
-    bool cross_attn;
     bool warmup;
+    bool op_offload;
 
     enum llama_pooling_type pooling_type;
 

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

@@ -1186,8 +1186,18 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
             for (const auto & trigger_pattern : grammar.trigger_patterns) {
                 if (std::regex_match(grammar.trigger_buffer, match, trigger_pattern.regex)) {
                     grammar.awaiting_trigger = false;
-                    // get from the first match to the end of the string
-                    auto constrained_str = grammar.trigger_buffer.substr(match.position(1));
+                    // get from the first matched capturing group to the end of the string
+                    size_t start = std::string::npos;
+                    for (auto i = 1u; i < match.size(); i++) {
+                        if (match.length(i) > 0) {
+                            start = match.position(i);
+                            break;
+                        }
+                    }
+                    if (start == std::string::npos) {
+                        start = match.position(0);
+                    }
+                    auto constrained_str = grammar.trigger_buffer.substr(start);
                     // std::string constrained_str(match[1].first, grammar.trigger_buffer.end());
                     grammar.trigger_buffer.clear();
                     llama_grammar_accept_str(grammar, constrained_str);

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

@@ -3,39 +3,15 @@
 #include "llama-impl.h"
 #include "llama-batch.h"
 #include "llama-cparams.h"
-#include "llama-kv-cache.h"
+
+#include "llama-kv-cache-unified.h"
+#include "llama-kv-cache-unified-iswa.h"
+#include "llama-kv-cache-recurrent.h"
 
 #include <cassert>
 #include <cmath>
 #include <cstring>
 
-static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) {
-    // TODO move to hparams if a T5 variant appears that uses a different value
-    const int64_t max_distance = 128;
-
-    if (bidirectional) {
-        n_buckets >>= 1;
-    }
-
-    const int64_t max_exact = n_buckets >> 1;
-
-    int32_t relative_position = x - y;
-    int32_t relative_bucket = 0;
-
-    if (bidirectional) {
-        relative_bucket += (relative_position > 0) * n_buckets;
-        relative_position = abs(relative_position);
-    } else {
-        relative_position = -std::min<int32_t>(relative_position, 0);
-    }
-
-    int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact));
-    relative_position_if_large = std::min<int32_t>(relative_position_if_large, n_buckets - 1);
-    relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large);
-
-    return relative_bucket;
-}
-
 void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
     if (ubatch->token) {
         const int64_t n_tokens = ubatch->n_tokens;
@@ -55,7 +31,21 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && pos) {
         const int64_t n_tokens = ubatch->n_tokens;
 
-        ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_token*ggml_element_size(pos));
+        if (ubatch->token && n_pos_per_embd == 4) {
+            // in case we're using M-RoPE with text tokens, convert the 1D positions to 4D
+            // the 3 first dims are the same, and 4th dim is all 0
+            std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd);
+            // copy the first dimension
+            for (int i = 0; i < n_tokens; ++i) {
+                pos_data[               i] = ubatch->pos[i];
+                pos_data[    n_tokens + i] = ubatch->pos[i];
+                pos_data[2 * n_tokens + i] = ubatch->pos[i];
+                pos_data[3 * n_tokens + i] = 0; // 4th dim is 0
+            }
+            ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos));
+        } else {
+            ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_embd*ggml_element_size(pos));
+        }
     }
 }
 
@@ -71,7 +61,7 @@ void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
             ) * f_attn_temp_scale + 1.0;
         }
 
-        ggml_backend_tensor_set(attn_scale, attn_scale_data.data(), 0, n_tokens*n_pos_per_token*ggml_element_size(attn_scale));
+        ggml_backend_tensor_set(attn_scale, attn_scale_data.data(), 0, n_tokens*ggml_element_size(attn_scale));
     }
 }
 
@@ -96,22 +86,7 @@ void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) {
 
 void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) {
     if (pos_bucket) {
-        const int64_t n_tokens = ubatch->n_tokens;
-
-        GGML_ASSERT(ggml_backend_buffer_is_host(pos_bucket->buffer));
-        GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
-
-        int32_t * data = (int32_t *) pos_bucket->data;
-
-        const int64_t n_kv = kv_self->n;
-
-        for (int h = 0; h < 1; ++h) {
-            for (int j = 0; j < n_tokens; ++j) {
-                for (int i = 0; i < n_kv; ++i) {
-                    data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(kv_self->cells[i].pos, ubatch->pos[j], hparams.n_rel_attn_bkts, false);
-                }
-            }
-        }
+        kv_state->set_input_pos_bucket(pos_bucket, ubatch);
     }
 }
 
@@ -262,7 +237,7 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
 void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
-    const int64_t n_kv = kv_self->n;
+    const int64_t n_kv = kv_state->get_n_kv();
 
     if (s_copy) {
         GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer));
@@ -270,24 +245,7 @@ void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
 
         // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
         for (uint32_t i = 0; i < n_kv; ++i) {
-            const uint32_t  cell_id = i + kv_self->head;
-
-            //////////////////////////////////////////////
-            // TODO: this should not mutate the KV cache !
-            llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
-
-            // prevent out-of-bound sources
-            if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self->size) {
-                kv_cell.src = cell_id;
-            }
-
-            data[i] = kv_cell.src;
-
-            // TODO: do not mutate the KV cache
-            // ensure copy only happens once
-            if (kv_cell.src != (int32_t) cell_id) {
-                kv_cell.src = cell_id;
-            }
+            data[i] = kv_state->s_copy(i);
         }
     }
 }
@@ -295,7 +253,7 @@ void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
 void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
-    const int64_t n_kv = kv_self->n;
+    const int64_t n_kv = kv_state->get_n_kv();
 
     if (s_mask) {
         GGML_ASSERT(ggml_backend_buffer_is_host(s_mask->buffer));
@@ -303,18 +261,7 @@ void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) {
 
         // clear unused states
         for (int i = 0; i < n_kv; ++i) {
-            const uint32_t  cell_id = i + kv_self->head;
-
-            //////////////////////////////////////////////
-            // TODO: this should not mutate the KV cache !
-            llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
-
-            data[i] = (float) (kv_cell.src >= 0);
-
-            // only clear once
-            if (kv_cell.src < 0) {
-                kv_cell.src = cell_id;
-            }
+            data[i] = kv_state->s_mask(i);
         }
     }
 }
@@ -417,99 +364,18 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask || self_kq_mask_swa) {
-        const int64_t n_kv         = kv_self->n;
-        const int64_t n_tokens     = ubatch->n_tokens;
-        const int64_t n_seq_tokens = ubatch->n_seq_tokens;
-        const int64_t n_seqs       = ubatch->n_seqs;
+    if (self_kq_mask) {
+        kv_state->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+    }
+}
 
-        float * data     = nullptr;
-        float * data_swa = nullptr;
+void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
+    if (self_kq_mask) {
+        kv_state->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+    }
 
-        if (self_kq_mask) {
-            GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer));
-            data = (float *) self_kq_mask->data;
-        }
-
-        if (self_kq_mask_swa) {
-            GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer));
-            data_swa = (float *) self_kq_mask_swa->data;
-        }
-
-        // Use only the previous KV cells of the correct sequence for each token of the ubatch.
-        // It's assumed that if a token in the batch has multiple sequences, they are equivalent.
-        // Example with a cache of 10 tokens, 2 tokens populated in cache and 3 tokens in batch:
-        //   Causal mask:
-        //      xxx-------
-        //      xxxx------
-        //      xxxxx-----
-        //   Non-causal mask:
-        //      xxxxx-----
-        //      xxxxx-----
-        //      xxxxx-----
-        // To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615
-        for (int h = 0; h < 1; ++h) {
-            for (int s = 0; s < n_seqs; ++s) {
-                const llama_seq_id seq_id = ubatch->seq_id[s][0];
-
-                for (int j = 0; j < n_seq_tokens; ++j) {
-                    const llama_pos pos = ubatch->pos[s*n_seq_tokens + j];
-                    for (int i = 0; i < n_kv; ++i) {
-                        float f;
-                        // mask the token if:
-                        if (!kv_self->cells[i].has_seq_id(seq_id) // not the correct sequence
-                            || (cparams.causal_attn && kv_self->cells[i].pos > pos) // for causal, mask future tokens
-                        ) {
-                            f = -INFINITY;
-                        } else {
-                            if (hparams.use_alibi) {
-                                f = -std::abs(kv_self->cells[i].pos - pos);
-                            } else {
-                                f = 0.0f;
-                            }
-                        }
-
-                        if (data) {
-                            data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f;
-                        }
-
-                        // may need to cut off old tokens for sliding window
-                        // TODO @ngxson : we are currently re-using the swa logic to store the chunked mask, we should rename SWA to something more generic like "aux mask"
-                        if (data_swa) {
-                            if (hparams.n_attn_chunk) {
-                                llama_pos pos_chunk_start = (pos / hparams.n_attn_chunk) * hparams.n_attn_chunk;
-                                if (kv_self->cells[i].pos < pos_chunk_start || pos < pos_chunk_start) {
-                                    f = -INFINITY;
-                                }
-                            } else {
-                                if (pos - kv_self->cells[i].pos >= (int32_t)hparams.n_swa) {
-                                    f = -INFINITY;
-                                }
-                            }
-                            data_swa[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f;
-                        }
-                    }
-                }
-            }
-
-            // mask padded tokens
-            if (data) {
-                for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
-                    for (int j = 0; j < n_kv; ++j) {
-                        data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
-                    }
-                }
-            }
-
-            // mask padded tokens
-            if (data_swa) {
-                for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
-                    for (int j = 0; j < n_kv; ++j) {
-                        data_swa[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
-                    }
-                }
-            }
-        }
+    if (self_kq_mask_swa) {
+        kv_state->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
     }
 }
 
@@ -546,12 +412,6 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
     }
 }
 
-void llm_graph_input_cross_attn_state::set_input(const llama_ubatch * ubatch) {
-    if (ubatch->embd) {
-        ggml_backend_tensor_set(cross_attn_state, ubatch->embd, 0, ggml_nbytes(cross_attn_state));
-    }
-}
-
 //
 // llm_graph_context
 //
@@ -565,7 +425,6 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     n_layer          (hparams.n_layer),
     n_rot            (hparams.n_rot),
     n_ctx            (cparams.n_ctx),
-    n_ctx_per_seq    (cparams.n_ctx / cparams.n_seq_max),
     n_head           (hparams.n_head()),
     n_head_kv        (hparams.n_head_kv()),
     n_embd_head_k    (hparams.n_embd_head_k),
@@ -592,14 +451,14 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     backend_cpu      (params.backend_cpu),
     cvec             (params.cvec),
     loras            (params.loras),
-    memory           (params.memory),
+    mstate           (params.mstate),
     cross            (params.cross),
     cb_func          (params.cb),
     res              (std::make_unique<llm_graph_result>()) {
     }
 
-int64_t llm_graph_context::n_pos_per_token() const {
-    return arch == LLM_ARCH_QWEN2VL ? 4 : 1;
+int64_t llm_graph_context::n_pos_per_embd() const {
+    return hparams.rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
 }
 
 void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const {
@@ -802,13 +661,17 @@ ggml_tensor * llm_graph_context::build_ffn(
             } break;
     }
 
-    if (type_gate == LLM_FFN_PAR) {
+    if (gate && type_gate == LLM_FFN_PAR) {
         cur = ggml_mul(ctx0, cur, tmp);
         cb(cur, "ffn_gate_par", il);
     }
 
     if (down) {
         cur = build_lora_mm(down, cur);
+        if (arch == LLM_ARCH_GLM4) {
+            // GLM4 seems to have numerical issues with half-precision accumulators
+            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+        }
     }
 
     if (down_b) {
@@ -906,9 +769,8 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
 
     if (weight_before_ffn) {
-        // TODO: this is a workaround as we don't yet have a repeat op that takes custom dim (ggml_repeat_4d)
-        ggml_tensor * repeated = ggml_new_tensor_3d(ctx0, cur->type, n_embd, n_expert_used, n_tokens);
-        repeated = ggml_repeat(ctx0, cur, repeated); // [n_embd, n_expert_used, n_tokens]
+        // repeat cur to [n_embd, n_expert_used, n_tokens]
+        ggml_tensor * repeated = ggml_repeat_4d(ctx0, cur, n_embd, n_expert_used, n_tokens, 1);
         cur = ggml_mul(ctx0, repeated, weights);
         cb(cur, "ffn_moe_weighted", il);
     }
@@ -916,28 +778,35 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
     cb(up, "ffn_moe_up", il);
 
-    ggml_tensor * gate = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
-    cb(gate, "ffn_moe_gate", il);
+    ggml_tensor * experts = nullptr;
+    if (gate_exps) {
+        cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+        cb(cur, "ffn_moe_gate", il);
+    } else {
+        cur = up;
+    }
 
     switch (type_op) {
         case LLM_FFN_SILU:
             {
-                gate = ggml_silu(ctx0, gate);
-                cb(gate, "ffn_moe_silu", il);
+                cur = ggml_silu(ctx0, cur);
+                cb(cur, "ffn_moe_silu", il);
             } break;
         case LLM_FFN_GELU:
             {
-                gate = ggml_gelu(ctx0, gate);
-                cb(gate, "ffn_moe_gelu", il);
+                cur = ggml_gelu(ctx0, cur);
+                cb(cur, "ffn_moe_gelu", il);
             } break;
         default:
             GGML_ABORT("fatal error");
     }
 
-    ggml_tensor * par = ggml_mul(ctx0, up, gate); // [n_ff, n_expert_used, n_tokens]
-    cb(par, "ffn_moe_gate_par", il);
+    if (gate_exps) {
+        cur = ggml_mul(ctx0, cur, up); // [n_ff, n_expert_used, n_tokens]
+        cb(cur, "ffn_moe_gate_par", il);
+    }
 
-    ggml_tensor * experts = build_lora_mm_id(down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens]
+    experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 
     if (!weight_before_ffn) {
@@ -980,6 +849,7 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
         inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
         //cb(inp->tokens, "inp_tokens", -1);
         ggml_set_input(inp->tokens);
+        res->t_tokens = inp->tokens;
 
         cur = ggml_get_rows(ctx0, tok_embd, inp->tokens);
 
@@ -1020,11 +890,11 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_pos() const {
-    auto inp = std::make_unique<llm_graph_input_pos>(n_pos_per_token());
+    auto inp = std::make_unique<llm_graph_input_pos>(n_pos_per_embd());
 
     auto & cur = inp->pos;
 
-    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_token());
+    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_embd());
     ggml_set_input(cur);
 
     res->add_input(std::move(inp));
@@ -1033,11 +903,12 @@ ggml_tensor * llm_graph_context::build_inp_pos() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
-    auto inp = std::make_unique<llm_graph_input_attn_temp>(n_pos_per_token(), hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
+    auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
 
     auto & cur = inp->attn_scale;
 
-    cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens*n_pos_per_token());
+    // this need to be 1x1xN for broadcasting
+    cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens);
     ggml_set_input(cur);
 
     res->add_input(std::move(inp));
@@ -1085,11 +956,11 @@ ggml_tensor * llm_graph_context::build_inp_cls() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_s_copy() const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
 
-    auto inp = std::make_unique<llm_graph_input_s_copy>(kv_self);
+    auto inp = std::make_unique<llm_graph_input_s_copy>(kv_state);
 
-    const auto n_kv = kv_self->n;
+    const auto n_kv = kv_state->get_n_kv();
 
     auto & cur = inp->s_copy;
 
@@ -1102,11 +973,11 @@ ggml_tensor * llm_graph_context::build_inp_s_copy() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_s_mask() const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
 
-    auto inp = std::make_unique<llm_graph_input_s_mask>(kv_self);
+    auto inp = std::make_unique<llm_graph_input_s_mask>(kv_state);
 
-    const auto n_kv = kv_self->n;
+    const auto n_kv = kv_state->get_n_kv();
 
     auto & cur = inp->s_mask;
 
@@ -1156,11 +1027,11 @@ ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const auto * kv_state = static_cast<const llama_kv_cache_unified_state *>(mstate);
 
-    auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, kv_self);
+    auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, kv_state);
 
-    const auto n_kv = kv_self->n;
+    const auto n_kv = kv_state->get_n_kv();
 
     auto & cur = inp->pos_bucket;
 
@@ -1195,16 +1066,12 @@ ggml_tensor * llm_graph_context::build_attn_mha(
          ggml_tensor * kq_b,
          ggml_tensor * kq_mask,
          ggml_tensor * v_mla,
-             bool      v_trans,
              float     kq_scale) const {
-  //const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
-  //const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
+    const bool v_trans = v->nb[1] > v->nb[2];
 
-  //const int64_t n_head    = hparams.n_head(il);
-  //const int64_t n_head_kv = hparams.n_head_kv(il);
-
-  //const auto & n_embd_head_k = hparams.n_embd_head_k;
-  //const auto & n_embd_head_v = hparams.n_embd_head_v;
+    q = ggml_permute(ctx0, q, 0, 2, 1, 3);
+    k = ggml_permute(ctx0, k, 0, 2, 1, 3);
+    v = ggml_permute(ctx0, v, 0, 2, 1, 3);
 
     const auto n_tokens = q->ne[1];
     const auto n_head   = q->ne[2];
@@ -1235,8 +1102,19 @@ ggml_tensor * llm_graph_context::build_attn_mha(
         ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
 
         if (v_mla) {
+#if 0
+            // v_mla can be applied as a matrix-vector multiplication with broadcasting across dimension 3 == n_tokens.
+            // However, the code is optimized for dimensions 0 and 1 being large, so this is ineffient.
             cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens);
             cur = ggml_mul_mat(ctx0, v_mla, cur);
+#else
+            // It's preferable to do the calculation as a matrix-matrix multiplication with n_tokens in dimension 1.
+            // The permutations are noops and only change how the tensor data is interpreted.
+            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+            cur = ggml_mul_mat(ctx0, v_mla, cur);
+            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+            cur = ggml_cont(ctx0, cur); // Needed because ggml_reshape_2d expects contiguous inputs.
+#endif
         }
 
         cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
@@ -1332,17 +1210,11 @@ ggml_tensor * llm_graph_context::build_attn(
 
     const auto & kq_mask = inp->get_kq_mask();
 
-    ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
-    //cb(q, "q", il);
-
-    ggml_tensor * k = ggml_permute(ctx0, k_cur, 0, 2, 1, 3);
-    //cb(k, "k", il);
-
-    ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
-    //cb(k, "v", il);
-
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale);
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = k_cur;
+    ggml_tensor * v = v_cur;
 
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1361,26 +1233,20 @@ ggml_tensor * llm_graph_context::build_attn(
 }
 
 llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const auto * kv_state = static_cast<const llama_kv_cache_unified_state *>(mstate);
 
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, kv_self);
+    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, kv_state);
 
-    const auto n_kv = kv_self->n;
+    {
+        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA");
 
-    inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-    //cb(inp->self_kq_mask, "KQ_mask", -1);
-    ggml_set_input(inp->self_kq_mask);
+        const auto n_kv = kv_state->get_n_kv();
 
-    inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
+        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        ggml_set_input(inp->self_kq_mask);
 
-    if (hparams.n_swa_pattern > 1) {
-        GGML_ASSERT(hparams.n_swa > 0);
-
-        inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
-        ggml_set_input(inp->self_kq_mask_swa);
-
-        inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
+        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
     }
 
     return (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp));
@@ -1404,84 +1270,105 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_build_forward_expand(gf, k_cur);
     ggml_build_forward_expand(gf, v_cur);
 
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
-    const auto & n_ctx = cparams.n_ctx;
-
-    const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
-    const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
-
-    const auto n_tokens = q_cur->ne[2];
-
-    const bool v_trans = !cparams.flash_attn;
+    const auto * kv_state = static_cast<const llama_kv_cache_unified_state *>(mstate);
 
     // store to KV cache
     {
-        GGML_ASSERT(!kv_self->recurrent);
-
-        const auto kv_head = kv_self->head;
-
-        GGML_ASSERT(kv_self->size == n_ctx);
-
-        ggml_tensor * k_cache_view = ggml_view_1d(ctx0, kv_self->k_l[il], n_tokens*n_embd_k_gqa, ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa)*kv_head);
-        //cb(k_cache_view, "k_cache_view", il);
-
-        // note: storing RoPE-ed version of K in the KV cache
-        ggml_build_forward_expand(gf, ggml_cpy(ctx0, k_cur, k_cache_view));
-
-        v_cur = ggml_reshape_2d(ctx0, v_cur, n_embd_v_gqa, n_tokens);
-
-        ggml_tensor * v_cache_view = nullptr;
-
-        if (!v_trans) {
-            v_cache_view = ggml_view_1d(ctx0, kv_self->v_l[il], n_tokens*n_embd_v_gqa, ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa)*kv_head);
-        } else {
-            // note: the V cache is transposed when not using flash attention
-            v_cache_view = ggml_view_2d(ctx0, kv_self->v_l[il], n_tokens, n_embd_v_gqa,
-                    (  n_ctx)*ggml_element_size(kv_self->v_l[il]),
-                    (kv_head)*ggml_element_size(kv_self->v_l[il]));
-
-            v_cur = ggml_transpose(ctx0, v_cur);
-        }
-        //cb(v_cache_view, "v_cache_view", il);
-
-        ggml_build_forward_expand(gf, ggml_cpy(ctx0, v_cur, v_cache_view));
+        ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il));
+        ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il));
     }
 
+    const auto & kq_mask = inp->get_kq_mask();
+
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = kv_state->get_k(ctx0, il);
+    ggml_tensor * v = kv_state->get_v(ctx0, il);
+
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    cb(cur, "kqv_out", il);
+
+    if (wo) {
+        cur = build_lora_mm(wo, cur);
+        if (arch == LLM_ARCH_GLM4) {
+            // GLM4 seems to have numerical issues with half-precision accumulators
+            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+        }
+    }
+
+    if (wo_b) {
+        cur = ggml_add(ctx0, cur, wo_b);
+    }
+
+    return cur;
+}
+
+llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
+    const auto * kv_state = static_cast<const llama_kv_cache_unified_iswa_state *>(mstate);
+
+    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, kv_state);
+
+    {
+        const auto n_kv = kv_state->get_base()->get_n_kv();
+
+        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        ggml_set_input(inp->self_kq_mask);
+
+        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
+    }
+
+    {
+        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified for non-SWA");
+
+        const auto n_kv = kv_state->get_swa()->get_n_kv();
+
+        inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
+        ggml_set_input(inp->self_kq_mask_swa);
+
+        inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
+    }
+
+    return (llm_graph_input_attn_kv_unified_iswa *) res->add_input(std::move(inp));
+}
+
+ggml_tensor * llm_graph_context::build_attn(
+        llm_graph_input_attn_kv_unified_iswa * inp,
+        ggml_cgraph * gf,
+        ggml_tensor * wo,
+        ggml_tensor * wo_b,
+        ggml_tensor * q_cur,
+        ggml_tensor * k_cur,
+        ggml_tensor * v_cur,
+        ggml_tensor * kq_b,
+        ggml_tensor * v_mla,
+            float     kq_scale,
+            int       il) const {
+    // these nodes are added to the graph together so that they are not reordered
+    // by doing so, the number of splits in the graph is reduced
+    ggml_build_forward_expand(gf, q_cur);
+    ggml_build_forward_expand(gf, k_cur);
+    ggml_build_forward_expand(gf, v_cur);
+
+    const auto * kv_state_iswa = static_cast<const llama_kv_cache_unified_iswa_state *>(mstate);
+
     const bool is_swa = hparams.is_swa(il);
 
+    const auto * kv_state = is_swa ? kv_state_iswa->get_swa() : kv_state_iswa->get_base();
+
+    // store to KV cache
+    {
+        ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il));
+        ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il));
+    }
+
     const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
 
-    const auto n_kv = kv_self->n;
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = kv_state->get_k(ctx0, il);
+    ggml_tensor * v = kv_state->get_v(ctx0, il);
 
-    const int64_t n_head_kv = hparams.n_head_kv(il);
-
-    const auto & n_embd_head_k = hparams.n_embd_head_k;
-    const auto & n_embd_head_v = hparams.n_embd_head_v;
-
-    ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
-    //cb(q, "q", il);
-
-    ggml_tensor * k =
-        ggml_view_3d(ctx0, kv_self->k_l[il],
-                n_embd_head_k, n_kv, n_head_kv,
-                ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
-                ggml_row_size(kv_self->k_l[il]->type, n_embd_head_k),
-                0);
-    //cb(k, "k", il);
-
-    ggml_tensor * v = !v_trans ?
-        ggml_view_3d(ctx0, kv_self->v_l[il],
-                n_embd_head_v, n_kv, n_head_kv,
-                ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa),
-                ggml_row_size(kv_self->v_l[il]->type, n_embd_head_v),
-                0) :
-        ggml_view_3d(ctx0, kv_self->v_l[il],
-                n_kv, n_embd_head_v, n_head_kv,
-                ggml_element_size(kv_self->v_l[il])*n_ctx,
-                ggml_element_size(kv_self->v_l[il])*n_ctx*n_embd_head_v,
-                0);
-
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, v_trans, kq_scale);
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1512,25 +1399,6 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
     return (llm_graph_input_attn_cross *) res->add_input(std::move(inp));
 }
 
-ggml_tensor * llm_graph_context::build_inp_cross_attn_state() const {
-    const int64_t n_embd = hparams.n_embd;
-
-    auto inp = std::make_unique<llm_graph_input_cross_attn_state>();
-
-    ggml_tensor * cur = nullptr;
-
-    inp->cross_attn_state = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd, 1601, 4);
-    ggml_set_input(inp->cross_attn_state);
-
-    cur = inp->cross_attn_state;
-
-    cb(cur, "inp_cross_attn_state", -1);
-
-    res->add_input(std::move(inp));
-
-    return cur;
-}
-
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_cross * inp,
         ggml_cgraph * gf,
@@ -1551,17 +1419,11 @@ ggml_tensor * llm_graph_context::build_attn(
 
     const auto & kq_mask = inp->get_kq_mask_cross();
 
-    ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
-    //cb(q, "q", il);
-
-    ggml_tensor * k = ggml_permute(ctx0, k_cur, 0, 2, 1, 3);
-    //cb(k, "k", il);
-
-    ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
-    //cb(k, "v", il);
-
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale);
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = k_cur;
+    ggml_tensor * v = v_cur;
 
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1586,12 +1448,12 @@ ggml_tensor * llm_graph_context::build_copy_mask_state(
          ggml_tensor * state_mask,
              int32_t   n_state,
              int32_t   n_seqs) const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
 
-    const auto n_kv    = kv_self->n;
-    const auto kv_head = kv_self->head;
+    const auto n_kv    = kv_state->get_n_kv();
+    const auto kv_head = kv_state->get_head();
 
-    ggml_tensor * states = ggml_reshape_2d(ctx0, s, n_state, kv_self->size);
+    ggml_tensor * states = ggml_reshape_2d(ctx0, s, n_state, kv_state->get_size());
 
     // copy states
     // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
@@ -1618,13 +1480,13 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
          ggml_tensor * state_mask,
   const llama_ubatch & ubatch,
                  int   il) const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
 
     const auto token_shift_count = hparams.token_shift_count;
 
     const int64_t n_seqs  = ubatch.n_seqs;
 
-    ggml_tensor * token_shift_all = kv_self->k_l[il];
+    ggml_tensor * token_shift_all = kv_state->get_k_l(il);
 
     ggml_tensor * token_shift = build_copy_mask_state(
             gf, token_shift_all, state_copy, state_mask,
@@ -1639,19 +1501,19 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
          ggml_tensor * token_shift,
   const llama_ubatch & ubatch,
                  int   il) const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
 
     const auto token_shift_count = hparams.token_shift_count;
     const auto n_embd = hparams.n_embd;
 
     const int64_t n_seqs = ubatch.n_seqs;
 
-    const auto kv_head = kv_self->head;
+    const auto kv_head = kv_state->get_head();
 
     return ggml_cpy(
         ctx0,
         ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * token_shift_count, 0),
-        ggml_view_1d(ctx0, kv_self->k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * ggml_element_size(kv_self->k_l[il]))
+        ggml_view_1d(ctx0, kv_state->get_k_l(il), hparams.n_embd_k_s()*n_seqs, hparams.n_embd_k_s()*kv_head*ggml_element_size(kv_state->get_k_l(il)))
     );
 }
 
@@ -1702,20 +1564,25 @@ void llm_graph_context::build_pooling(
                 ggml_tensor * inp_cls = build_inp_cls();
                 inp = ggml_get_rows(ctx0, inp, inp_cls);
 
-                // classification head
-                // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
-                GGML_ASSERT(cls   != nullptr);
-                GGML_ASSERT(cls_b != nullptr);
+                if (cls != nullptr && cls_b != nullptr) {
+                    // classification head
+                    // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
+                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b);
+                    cur = ggml_tanh(ctx0, cur);
 
-                cur = ggml_add (ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b);
-                cur = ggml_tanh(ctx0, cur);
-
-                // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
-                // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
-                if (cls_out) {
+                    // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
+                    // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
+                    if (cls_out) {
+                        GGML_ASSERT(cls_out_b != nullptr);
+                        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b);
+                    }
+                } else if (cls_out) {
+                    // Single layer classification head (direct projection)
+                    // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
                     GGML_ASSERT(cls_out_b != nullptr);
-
-                    cur = ggml_add (ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b);
+                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, inp), cls_out_b);
+                } else {
+                    GGML_ABORT("RANK pooling requires either cls+cls_b or cls_out+cls_out_b");
                 }
             } break;
         default:
@@ -1729,3 +1596,30 @@ void llm_graph_context::build_pooling(
 
     ggml_build_forward_expand(gf, cur);
 }
+
+int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) {
+    // TODO move to hparams if a T5 variant appears that uses a different value
+    const int64_t max_distance = 128;
+
+    if (bidirectional) {
+        n_buckets >>= 1;
+    }
+
+    const int64_t max_exact = n_buckets >> 1;
+
+    int32_t relative_position = x - y;
+    int32_t relative_bucket = 0;
+
+    if (bidirectional) {
+        relative_bucket += (relative_position > 0) * n_buckets;
+        relative_position = abs(relative_position);
+    } else {
+        relative_position = -std::min<int32_t>(relative_position, 0);
+    }
+
+    int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact));
+    relative_position_if_large = std::min<int32_t>(relative_position_if_large, n_buckets - 1);
+    relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large);
+
+    return relative_bucket;
+}

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

@@ -17,8 +17,11 @@ struct ggml_tensor;
 struct llama_ubatch;
 struct llama_cparams;
 
-class llama_memory_i;
-class llama_kv_cache_unified;
+struct llama_memory_state_i;
+
+class llama_kv_cache_unified_state;
+class llama_kv_cache_unified_iswa_state;
+class llama_kv_cache_recurrent_state;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -86,34 +89,31 @@ public:
 
     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) {}
+    llm_graph_input_pos(int64_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}
     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;
+    const int64_t n_pos_per_embd = 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) {}
+    llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
+        : 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;
 };
@@ -134,7 +134,7 @@ 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) {}
+            const llama_kv_cache_unified_state * kv_state) : hparams(hparams), kv_state(kv_state) {}
     virtual ~llm_graph_input_pos_bucket_kv() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -142,7 +142,7 @@ public:
     ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch]
 
     const llama_hparams & hparams;
-    const llama_kv_cache_unified * kv_self;
+    const llama_kv_cache_unified_state * kv_state;
 };
 
 class llm_graph_input_out_ids : public llm_graph_input_i {
@@ -189,26 +189,26 @@ public:
 
 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) {}
+    llm_graph_input_s_copy(const llama_kv_cache_recurrent_state * kv_state) : kv_state(kv_state) {}
     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;
+    const llama_kv_cache_recurrent_state * kv_state;
 };
 
 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) {}
+    llm_graph_input_s_mask(const llama_kv_cache_recurrent_state * kv_state) : kv_state(kv_state) {}
     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;
+    const llama_kv_cache_recurrent_state * kv_state;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -248,15 +248,40 @@ public:
     llm_graph_input_attn_kv_unified(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified * kv_self) :
+            const llama_kv_cache_unified_state * kv_state) :
         hparams(hparams),
         cparams(cparams),
-        kv_self(kv_self) {
+        kv_state(kv_state) {
     }
     ~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 * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+
+    const llama_kv_cache_unified_state * kv_state;
+};
+
+class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_kv_unified_iswa(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_kv_cache_unified_iswa_state * kv_state) :
+        hparams(hparams),
+        cparams(cparams),
+        kv_state(kv_state) {
+    }
+    ~llm_graph_input_attn_kv_unified_iswa() = 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; }
 
@@ -268,7 +293,7 @@ public:
     const llama_hparams & hparams;
     const llama_cparams & cparams;
 
-    const llama_kv_cache_unified * kv_self;
+    const llama_kv_cache_unified_iswa_state * kv_state;
 };
 
 class llm_graph_input_attn_cross : public llm_graph_input_i {
@@ -286,16 +311,6 @@ public:
     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
 //
@@ -310,6 +325,7 @@ class llm_graph_result_i {
 public:
     virtual ~llm_graph_result_i() = default;
 
+    virtual ggml_tensor * get_tokens()      = 0;
     virtual ggml_tensor * get_logits()      = 0;
     virtual ggml_tensor * get_embd()        = 0;
     virtual ggml_tensor * get_embd_pooled() = 0;
@@ -324,6 +340,7 @@ class llm_graph_result : public llm_graph_result_i {
 public:
     virtual ~llm_graph_result() = default;
 
+    ggml_tensor * get_tokens()      override { return t_tokens; }
     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; }
@@ -340,6 +357,7 @@ public:
     }
 
     // important graph nodes
+    ggml_tensor * t_tokens      = nullptr;
     ggml_tensor * t_logits      = nullptr;
     ggml_tensor * t_embd        = nullptr;
     ggml_tensor * t_embd_pooled = nullptr;
@@ -363,13 +381,13 @@ struct llm_graph_params {
     const llama_cparams & cparams;
     const llama_ubatch  & ubatch;
 
-    ggml_backend_sched * sched;
-    ggml_backend * backend_cpu;
+    ggml_backend_sched_t sched;
+    ggml_backend_t backend_cpu;
 
-    const llama_adapter_cvec  * cvec;
-    const llama_adapter_loras * loras;
-    const llama_memory_i      * memory;
-    const llama_cross         * cross;
+    const llama_adapter_cvec   * cvec;
+    const llama_adapter_loras  * loras;
+    const llama_memory_state_i * mstate;
+    const llama_cross          * cross;
 
     int32_t n_outputs;
 
@@ -387,7 +405,6 @@ struct llm_graph_context {
     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;
@@ -415,14 +432,14 @@ struct llm_graph_context {
 
     ggml_context * ctx0 = nullptr;
 
-    ggml_backend_sched * sched;
+    ggml_backend_sched_t sched;
 
-    ggml_backend * backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
+    ggml_backend_t 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 llama_adapter_cvec   * cvec;
+    const llama_adapter_loras  * loras;
+    const llama_memory_state_i * mstate;
+    const llama_cross          * cross;
 
     const llm_graph_cb & cb_func;
 
@@ -430,7 +447,7 @@ struct llm_graph_context {
 
     llm_graph_context(const llm_graph_params & params);
 
-    int64_t n_pos_per_token() const;
+    int64_t n_pos_per_embd() const;
 
     void cb(ggml_tensor * cur, const char * name, int il) const;
 
@@ -504,7 +521,6 @@ struct llm_graph_context {
     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;
@@ -517,13 +533,12 @@ struct llm_graph_context {
 
     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 * q,       // [n_embd_head_q, n_head_q, n_tokens]
+             ggml_tensor * k,       // [n_embd_head_k, n_head_k, n_tokens]
+             ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (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,
+             ggml_tensor * v_mla,   // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                    float   kq_scale) const;
 
     llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const;
@@ -556,6 +571,21 @@ struct llm_graph_context {
                   float   kq_scale,
                     int   il) const;
 
+    llm_graph_input_attn_kv_unified_iswa * build_attn_inp_kv_unified_iswa() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_kv_unified_iswa * 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(
@@ -606,3 +636,6 @@ struct llm_graph_context {
             ggml_tensor * cls_out,
             ggml_tensor * cls_out_b) const;
 };
+
+// TODO: better name
+int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional);

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

@@ -2,6 +2,22 @@
 
 #include "ggml.h"
 
+void llama_hparams::set_swa_pattern(uint32_t n_pattern) {
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        swa_layers[il] = n_pattern == 0 || (il % n_pattern < (n_pattern - 1));
+    }
+}
+
+bool llama_hparams::is_swa_any() const {
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (swa_layers[il]) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
 uint32_t llama_hparams::n_head(uint32_t il) const {
     if (il < n_layer) {
         return n_head_arr[il];
@@ -80,12 +96,8 @@ bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {
 
 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);
+        return swa_layers[il];
     }
 
     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

@@ -2,8 +2,6 @@
 
 #include "llama.h"
 
-#include <algorithm>
-
 #include <array>
 
 // bump if necessary
@@ -16,6 +14,12 @@ enum llama_expert_gating_func_type {
     LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID = 2,
 };
 
+enum llama_swa_type {
+    LLAMA_SWA_TYPE_NONE     = 0,
+    LLAMA_SWA_TYPE_STANDARD = 1,
+    LLAMA_SWA_TYPE_CHUNKED  = 2,
+};
+
 struct llama_hparams_posnet {
     uint32_t n_embd;
     uint32_t n_layer;
@@ -37,14 +41,11 @@ struct llama_hparams {
     uint32_t n_embd_features = 0;
     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;
     uint32_t n_expert_used = 0;
     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;
@@ -59,7 +60,6 @@ struct llama_hparams {
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
 
     std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr = {};
-    std::array<uint32_t, LLAMA_MAX_LAYERS> cross_attn_layers;
 
     uint32_t n_layer_dense_lead = 0;
     uint32_t n_lora_q           = 0;
@@ -72,6 +72,7 @@ struct llama_hparams {
     float    expert_weights_scale = 0.0;
     bool     expert_weights_norm  = false;
     uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
+    uint32_t moe_every_n_layers   = 0;
 
     float f_norm_eps;
     float f_norm_rms_eps;
@@ -101,6 +102,15 @@ struct llama_hparams {
 
     std::array<int, 4> rope_sections;
 
+    // Sliding Window Attention (SWA)
+    llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
+    // the size of the sliding window (0 - no SWA)
+    uint32_t n_swa = 0;
+    // if swa_layers[il] == true, then layer il is SWA
+    // if swa_layers[il] == false, then layer il is dense (i.e. non-SWA)
+    // by default, all layers are dense
+    std::array<bool, LLAMA_MAX_LAYERS> swa_layers;
+
     // for State Space Models
     uint32_t ssm_d_conv  = 0;
     uint32_t ssm_d_inner = 0;
@@ -121,11 +131,13 @@ struct llama_hparams {
     bool causal_attn   = true;
     bool use_alibi     = false;
     bool attn_soft_cap = false;
+    bool use_kq_norm   = true;
 
+    // for Classifiers
+    uint32_t n_cls_out = 1;
+
+    // llama4
     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;
@@ -138,6 +150,23 @@ struct llama_hparams {
     enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
     enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
 
+    // this value n_pattern means that every nth layer is dense (i.e. non-SWA)
+    // note that if n_pattern == 0, all layers are SWA
+    //           if n_pattern == 1, all layers are dense
+    // example: n_pattern = 3
+    //   il == 0: swa
+    //   il == 1: swa
+    //   il == 2: dense
+    //   il == 3: swa
+    //   il == 4: swa
+    //   il == 5: dense
+    //   il == 6: swa
+    //   etc ...
+    void set_swa_pattern(uint32_t n_pattern);
+
+    // return true if one of the layers is SWA
+    bool is_swa_any() const;
+
     uint32_t n_head(uint32_t il = 0) const;
 
     uint32_t n_head_kv(uint32_t il = 0) const;
@@ -162,9 +191,6 @@ struct llama_hparams {
     // Block skip connection
     bool n_bskcn(uint32_t n, uint32_t il) const;
 
-    // cross attention layers
-    bool cross_attention_layers(uint32_t il) const;
-
     bool is_swa(uint32_t il) const;
 };