Merge branch 'main' into drifkin/array-head-count-simple

.github/workflows/release.yaml

@@ -103,6 +103,11 @@ jobs:
         arch: [amd64]
         preset: ['CPU']
         include:
+          - os: windows
+            arch: amd64
+            preset: 'CUDA 11'
+            install: https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe
+            cuda-version: '11.3'
           - os: windows
             arch: amd64
             preset: 'CUDA 12'
@@ -319,6 +324,7 @@ jobs:
             case "$COMPONENT" in
               bin/ollama)               echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/*.so)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_v11)      echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/cuda_v12)      echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/cuda_jetpack5) echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
               lib/ollama/cuda_jetpack6) echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;

.github/workflows/test.yaml

@@ -46,7 +46,7 @@ jobs:
         include:
           - preset: CPU
           - preset: CUDA
-            container: nvidia/cuda:12.8.1-devel-ubuntu22.04
+            container: nvidia/cuda:11.8.0-devel-ubuntu22.04
             flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
           - preset: ROCm
             container: rocm/dev-ubuntu-22.04:6.1.2
@@ -78,7 +78,7 @@ jobs:
         include:
           - preset: CPU
           - preset: CUDA
-            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
+            install: https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe
             flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
           - preset: ROCm
             install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
@@ -102,7 +102,7 @@ jobs:
           $ErrorActionPreference = "Stop"
           if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
             Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_12.8", "nvcc_12.8", "cublas_12.8", "cublas_dev_12.8")) -NoNewWindow -Wait
+            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_11.3", "nvcc_11.3", "cublas_11.3", "cublas_dev_11.3")) -NoNewWindow -Wait
           }
 
           $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path

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)

CMakePresets.json

@@ -17,12 +17,20 @@
       "name": "CUDA",
       "inherits": [ "Default" ]
     },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "cacheVariables": {
+        "CMAKE_CUDA_ARCHITECTURES": "50;52;53;60;61;70;75;80;86",
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
+      }
+    },
     {
       "name": "CUDA 12",
       "inherits": [ "CUDA" ],
       "cacheVariables": {
         "CMAKE_CUDA_ARCHITECTURES": "50;60;61;70;75;80;86;87;89;90;90a;120",
-        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets"
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
       }
     },
     {
@@ -50,6 +58,7 @@
       "name": "ROCm 6",
       "inherits": [ "ROCm" ],
       "cacheVariables": {
+        "CMAKE_HIP_FLAGS": "-parallel-jobs=4",
         "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
       }
     }
@@ -70,6 +79,11 @@
       "configurePreset": "CUDA",
       "targets": [ "ggml-cuda" ]
     },
+    {
+      "name": "CUDA 11",
+      "inherits": [ "CUDA" ],
+      "configurePreset": "CUDA 11"
+    },
     {
       "name": "CUDA 12",
       "inherits": [ "CUDA" ],

Dockerfile

@@ -7,10 +7,14 @@ ARG JETPACK5VERSION=r35.4.1
 ARG JETPACK6VERSION=r36.4.0
 ARG CMAKEVERSION=3.31.2
 
+# CUDA v11 requires gcc v10.  v10.3 has regressions, so the rockylinux 8.5 AppStream has the latest compatible version
 FROM --platform=linux/amd64 rocm/dev-almalinux-8:${ROCMVERSION}-complete AS base-amd64
 RUN yum install -y yum-utils \
-    && dnf install -y ccache \
+    && yum-config-manager --add-repo https://dl.rockylinux.org/vault/rocky/8.5/AppStream/\$basearch/os/ \
+    && rpm --import https://dl.rockylinux.org/pub/rocky/RPM-GPG-KEY-Rocky-8 \
+    && dnf install -y yum-utils ccache gcc-toolset-10-gcc-10.2.1-8.2.el8 gcc-toolset-10-gcc-c++-10.2.1-8.2.el8 gcc-toolset-10-binutils-2.35-11.el8 \
     && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
+ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
 
 FROM --platform=linux/arm64 almalinux:8 AS base-arm64
 # install epel-release for ccache
@@ -34,6 +38,15 @@ RUN --mount=type=cache,target=/root/.ccache \
         && cmake --build --parallel --preset 'CPU' \
         && cmake --install build --component CPU --strip --parallel 8
 
+FROM base AS cuda-11
+ARG CUDA11VERSION=11.3
+RUN dnf install -y cuda-toolkit-${CUDA11VERSION//./-}
+ENV PATH=/usr/local/cuda-11/bin:$PATH
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'CUDA 11' \
+        && cmake --build --parallel --preset 'CUDA 11' \
+        && cmake --install build --component CUDA --strip --parallel 8
+
 FROM base AS cuda-12
 ARG CUDA12VERSION=12.8
 RUN dnf install -y cuda-toolkit-${CUDA12VERSION//./-}
@@ -85,9 +98,11 @@ RUN --mount=type=cache,target=/root/.cache/go-build \
     go build -trimpath -buildmode=pie -o /bin/ollama .
 
 FROM --platform=linux/amd64 scratch AS amd64
+COPY --from=cuda-11 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_v11
 COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
 
 FROM --platform=linux/arm64 scratch AS arm64
+COPY --from=cuda-11 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_v11
 COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
 COPY --from=jetpack-5 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_jetpack5
 COPY --from=jetpack-6 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_jetpack6

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggerganov/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=e1e8e0991ffd9e99a445c6812bb519d5bac9f4b5
+FETCH_HEAD=de4c07f93783a1a96456a44dc16b9db538ee1618
 
 .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 llama/llama.cpp
-llama/build-info.cpp: llama/build-info.cpp.in
+	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' <$< >$@
-	sed -e 's|@FETCH_HEAD@|$(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

@@ -40,10 +40,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla
 
 ## Quickstart
 
-To run and chat with [Llama 3.2](https://ollama.com/library/llama3.2):
+To run and chat with [Gemma 3](https://ollama.com/library/gemma3):
 
 ```shell
-ollama run llama3.2
+ollama run gemma3
 ```
 
 ## Model library
@@ -315,6 +315,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
@@ -404,6 +405,11 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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.) 
+- [GPTranslate](https://github.com/philberndt/GPTranslate) (A fast and lightweight, AI powered desktop translation application written with Rust and Tauri. Features real-time translation with OpenAI/Azure/Ollama.)
+- [ollama launcher](https://github.com/NGC13009/ollama-launcher) (A launcher for Ollama, aiming to provide users with convenient functions such as ollama server launching, management, or configuration.)
+- [ai-hub](https://github.com/Aj-Seven/ai-hub) (AI Hub supports multiple models via API keys and Chat support via Ollama API.)
 
 ### Cloud
 
@@ -447,6 +453,8 @@ 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
+- [ollama-multirun](https://github.com/attogram/ollama-multirun) - A bash shell script to run a single prompt against any or all of your locally installed ollama models, saving the output and performance statistics as easily navigable web pages. ([Demo](https://attogram.github.io/ai_test_zone/))
 
 ### Apple Vision Pro
 
@@ -526,6 +534,7 @@ 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
 
@@ -582,6 +591,8 @@ 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/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"`
+	Role    string `json:"role"`
-	Content   string      `json:"content"`
+	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"`
 }
@@ -478,6 +491,10 @@ 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"`
 

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

@@ -1,178 +0,0 @@
-package benchmark
-
-import (
-	"context"
-	"flag"
-	"fmt"
-	"testing"
-	"time"
-
-	"github.com/ollama/ollama/api"
-)
-
-// Command line flags
-var modelFlag string
-
-func init() {
-	flag.StringVar(&modelFlag, "m", "", "Name of the model to benchmark")
-	flag.Lookup("m").DefValue = "model"
-}
-
-// modelName returns the model name from flags, failing the test if not set
-func modelName(b *testing.B) string {
-	if modelFlag == "" {
-		b.Fatal("Error: -m flag is required for benchmark tests")
-	}
-	return modelFlag
-}
-
-type TestCase struct {
-	name      string
-	prompt    string
-	maxTokens int
-}
-
-// runGenerateBenchmark contains the common generate and metrics logic
-func runGenerateBenchmark(b *testing.B, ctx context.Context, client *api.Client, req *api.GenerateRequest) {
-	start := time.Now()
-	var ttft time.Duration
-	var metrics api.Metrics
-
-	err := client.Generate(ctx, req, func(resp api.GenerateResponse) error {
-		if ttft == 0 && resp.Response != "" {
-			ttft = time.Since(start)
-		}
-		if resp.Done {
-			metrics = resp.Metrics
-		}
-		return nil
-	})
-
-	// Report custom metrics as part of the benchmark results
-	b.ReportMetric(float64(ttft.Milliseconds()), "ttft_ms")
-	b.ReportMetric(float64(metrics.LoadDuration.Milliseconds()), "load_ms")
-
-	// Token throughput metrics
-	promptThroughput := float64(metrics.PromptEvalCount) / metrics.PromptEvalDuration.Seconds()
-	genThroughput := float64(metrics.EvalCount) / metrics.EvalDuration.Seconds()
-	b.ReportMetric(promptThroughput, "prompt_tok/s")
-	b.ReportMetric(genThroughput, "gen_tok/s")
-
-	// Token counts
-	b.ReportMetric(float64(metrics.PromptEvalCount), "prompt_tokens")
-	b.ReportMetric(float64(metrics.EvalCount), "gen_tokens")
-	if err != nil {
-		b.Fatal(err)
-	}
-}
-
-// BenchmarkColdStart runs benchmarks with model loading from cold state
-func BenchmarkColdStart(b *testing.B) {
-	client := setup(b)
-	tests := []TestCase{
-		{"short_prompt", "Write a long story", 100},
-		{"medium_prompt", "Write a detailed economic analysis", 500},
-		{"long_prompt", "Write a comprehensive AI research paper", 1000},
-	}
-	m := modelName(b)
-
-	for _, tt := range tests {
-		b.Run(fmt.Sprintf("%s/cold/%s", m, tt.name), func(b *testing.B) {
-			ctx := b.Context()
-
-			// Set number of tokens as our throughput metric
-			b.SetBytes(int64(tt.maxTokens))
-
-			for b.Loop() {
-				b.StopTimer()
-				// Ensure model is unloaded before each iteration
-				unload(client, m, b)
-				b.StartTimer()
-
-				req := &api.GenerateRequest{
-					Model:   m,
-					Prompt:  tt.prompt,
-					Options: map[string]any{"num_predict": tt.maxTokens, "temperature": 0.1},
-				}
-
-				runGenerateBenchmark(b, ctx, client, req)
-			}
-		})
-	}
-}
-
-// BenchmarkWarmStart runs benchmarks with pre-loaded model
-func BenchmarkWarmStart(b *testing.B) {
-	client := setup(b)
-	tests := []TestCase{
-		{"short_prompt", "Write a long story", 100},
-		{"medium_prompt", "Write a detailed economic analysis", 500},
-		{"long_prompt", "Write a comprehensive AI research paper", 1000},
-	}
-	m := modelName(b)
-
-	for _, tt := range tests {
-		b.Run(fmt.Sprintf("%s/warm/%s", m, tt.name), func(b *testing.B) {
-			ctx := b.Context()
-
-			// Pre-warm the model
-			warmup(client, m, tt.prompt, b)
-
-			// Set number of tokens as our throughput metric
-			b.SetBytes(int64(tt.maxTokens))
-
-			for b.Loop() {
-				req := &api.GenerateRequest{
-					Model:   m,
-					Prompt:  tt.prompt,
-					Options: map[string]any{"num_predict": tt.maxTokens, "temperature": 0.1},
-				}
-
-				runGenerateBenchmark(b, ctx, client, req)
-			}
-		})
-	}
-}
-
-// setup verifies server and model availability
-func setup(b *testing.B) *api.Client {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		b.Fatal(err)
-	}
-	if _, err := client.Show(b.Context(), &api.ShowRequest{Model: modelName(b)}); err != nil {
-		b.Fatalf("Model unavailable: %v", err)
-	}
-
-	return client
-}
-
-// warmup ensures the model is loaded and warmed up
-func warmup(client *api.Client, model string, prompt string, b *testing.B) {
-	for range 3 {
-		err := client.Generate(
-			context.Background(),
-			&api.GenerateRequest{
-				Model:   model,
-				Prompt:  prompt,
-				Options: map[string]any{"num_predict": 50, "temperature": 0.1},
-			},
-			func(api.GenerateResponse) error { return nil },
-		)
-		if err != nil {
-			b.Logf("Error during model warm-up: %v", err)
-		}
-	}
-}
-
-// unload forces model unloading using KeepAlive: 0 parameter
-func unload(client *api.Client, model string, b *testing.B) {
-	req := &api.GenerateRequest{
-		Model:     model,
-		KeepAlive: &api.Duration{Duration: 0},
-	}
-	if err := client.Generate(context.Background(), req, func(api.GenerateResponse) error { return nil }); err != nil {
-		b.Logf("Unload error: %v", err)
-	}
-	time.Sleep(1 * time.Second)
-}

cmd/cmd.go

@@ -39,6 +39,7 @@ 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"
@@ -46,6 +47,23 @@ import (
 	"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) {
@@ -265,6 +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 })
@@ -299,6 +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
@@ -362,6 +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,
@@ -747,11 +789,38 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 				case float64:
 					v = fmt.Sprintf("%g", vData)
 				case []any:
-					n := 3
+					targetWidth := 10 // Small width where we are displaying the data in a column
-					if len(vData) < n {
+
-						n = len(vData)
+					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)
 				}
@@ -772,10 +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) {
+		count := 0
-			if text := scanner.Text(); text != "" {
+		for scanner.Scan() {
-				rows = append(rows, []string{"", strings.TrimSpace(text)})
+			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
 	}
@@ -887,17 +965,19 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 type generateContextKey string
 
 type runOptions struct {
-	Model       string
+	Model        string
-	ParentModel string
+	ParentModel  string
-	Prompt      string
+	Prompt       string
-	Messages    []api.Message
+	Messages     []api.Message
-	WordWrap    bool
+	WordWrap     bool
-	Format      string
+	Format       string
-	System      string
+	System       string
-	Images      []api.ImageData
+	Images       []api.ImageData
-	Options     map[string]any
+	Options      map[string]any
-	MultiModal  bool
+	MultiModal   bool
-	KeepAlive   *api.Duration
+	KeepAlive    *api.Duration
+	Think        *bool
+	HideThinking bool
 }
 
 type displayResponseState struct {
@@ -953,6 +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 {
@@ -980,14 +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)
@@ -1004,6 +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 {
@@ -1065,13 +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
@@ -1097,6 +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 {
@@ -1200,11 +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
@@ -1282,7 +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",
@@ -1312,6 +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",
@@ -1363,7 +1506,6 @@ func NewCLI() *cobra.Command {
 		PreRunE: checkServerHeartbeat,
 		RunE:    ListRunningHandler,
 	}
-
 	copyCmd := &cobra.Command{
 		Use:     "cp SOURCE DESTINATION",
 		Short:   "Copy a model",
@@ -1452,3 +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

@@ -225,6 +225,7 @@ Weigh anchor!
   System
     You are a pirate!    
     Ahoy, matey!         
+    ...                  
 
 `
 		if diff := cmp.Diff(expect, b.String()); diff != "" {

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_darwin.go

@@ -5,7 +5,7 @@ import (
 	"errors"
 	"os"
 	"os/exec"
-	"strings"
+	"regexp"
 
 	"github.com/ollama/ollama/api"
 )
@@ -19,11 +19,12 @@ func startApp(ctx context.Context, client *api.Client) error {
 	if err != nil {
 		return err
 	}
-	if !strings.Contains(link, "Ollama.app") {
+	r := regexp.MustCompile(`^.*/Ollama\s?\d*.app`)
+	m := r.FindStringSubmatch(link)
+	if len(m) != 1 {
 		return errors.New("could not find ollama app")
 	}
-	path := strings.Split(link, "Ollama.app")
+	if err := exec.Command("/usr/bin/open", "-j", "-a", m[0], "--args", "--fast-startup").Run(); err != nil {
-	if err := exec.Command("/usr/bin/open", "-a", path[0]+"Ollama.app").Run(); err != nil {
 		return err
 	}
 	return waitForServer(ctx, client)

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 {
@@ -35,14 +45,11 @@ func startApp(ctx context.Context, client *api.Client) error {
 			}
 		}
 	}
-	// log.Printf("XXX attempting to start app %s", appExe)
 
 	cmd_path := "c:\\Windows\\system32\\cmd.exe"
-	cmd := exec.Command(cmd_path, "/c", appExe)
+	cmd := exec.Command(cmd_path, "/c", appExe, "--hide", "--fast-startup")
-	// TODO - these hide flags aren't working - still pops up a command window for some reason
 	cmd.SysProcAttr = &syscall.SysProcAttr{CreationFlags: 0x08000000, HideWindow: true}
 
-	// TODO this didn't help either...
 	cmd.Stdin = strings.NewReader("")
 	cmd.Stdout = os.Stdout
 	cmd.Stderr = os.Stderr
@@ -56,3 +63,50 @@ 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
+	}
+	if ret > uint32(len(pids)) {
+		pids = make([]uint32, ret+10)
+		if err := windows.EnumProcesses(pids, &ret); err != nil || ret == 0 {
+			slog.Debug("failed to check for running installers", "error", err)
+			return nil
+		}
+	}
+	if ret < uint32(len(pids)) {
+		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

@@ -1,6 +1,7 @@
 package convert
 
 import (
+	"cmp"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -14,13 +15,12 @@ import (
 )
 
 type ModelParameters struct {
-	Architectures []string       `json:"architectures"`
+	Architectures []string `json:"architectures"`
-	VocabSize     uint32         `json:"vocab_size"`
+	VocabSize     uint32   `json:"vocab_size"`
-	TextModel     TextParameters `json:"text_config"`
-}
 
-type TextParameters struct {
+	TextModel struct {
-	VocabSize uint32 `json:"vocab_size"`
+		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
@@ -173,6 +176,8 @@ func ConvertModel(fsys fs.FS, f *os.File) 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, f *os.File) 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, f *os.File) error {
 		return err
 	}
 
-	vocabSize := int(p.VocabSize)
+	vocabSize := int(cmp.Or(p.VocabSize, p.TextModel.VocabSize))
-	if vocabSize == 0 {
-		tVocabSize := int(p.TextModel.VocabSize)
-		vocabSize = tVocabSize
-	}
 
 	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))
 	}

convert/convert_llama.go

@@ -139,7 +139,8 @@ 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)
 			}
@@ -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_mixtral.go

@@ -2,9 +2,6 @@ package convert
 
 import (
 	"fmt"
-	"io"
-	"slices"
-	"strings"
 
 	"github.com/ollama/ollama/fs/ggml"
 )
@@ -30,65 +27,38 @@ func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
 }
 
 func (p *mixtralModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	oldnew := []string{
+	merges := make([]merge, 0, p.NumHiddenLayers*6)
-		"model.layers", "blk",
+	for i := range p.NumHiddenLayers {
-		"w1", "ffn_gate_exps",
+		merges = append(merges, merge{
-		"w2", "ffn_down_exps",
+			fmt.Sprintf("blk.%d.*.w1.weight", i),
-		"w3", "ffn_up_exps",
+			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
-	}
+		}, merge{
-
+			fmt.Sprintf("blk.%d.*.w1.bias", i),
-	for i := range p.NumLocalExperts {
+			fmt.Sprintf("blk.%d.ffn_gate_exps.bias", i),
-		oldnew = append(oldnew, fmt.Sprintf(".block_sparse_moe.experts.%d.", i), ".")
+		}, merge{
-	}
+			fmt.Sprintf("blk.%d.*.w2.weight", i),
-
+			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
-	// group experts of the same layer (model.layers.%d) and type (w[123]) into a single tensor
+		}, merge{
-	namer := strings.NewReplacer(oldnew...)
+			fmt.Sprintf("blk.%d.*.w2.bias", i),
-	experts := make(map[string]experts)
+			fmt.Sprintf("blk.%d.ffn_up_exps.bias", i),
-
+		}, merge{
-	// merge experts into a single tensor while removing them from ts
+			fmt.Sprintf("blk.%d.*.w3.weight", i),
-	ts = slices.DeleteFunc(ts, func(t Tensor) bool {
+			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
-		if !strings.Contains(t.Name(), ".block_sparse_moe.experts.") {
+		}, merge{
-			return false
+			fmt.Sprintf("blk.%d.*.w3.bias", i),
-		}
+			fmt.Sprintf("blk.%d.ffn_down_exps.bias", i),
-
-		name := namer.Replace(t.Name())
-		experts[name] = append(experts[name], t)
-		return true
-	})
-
-	var out []*ggml.Tensor
-	for n, e := range experts {
-		// TODO(mxyng): sanity check experts
-		out = append(out, &ggml.Tensor{
-			Name:     n,
-			Kind:     e[0].Kind(),
-			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),
-			WriterTo: e,
 		})
 	}
 
+	out, ts := mergeTensors(ts, merges...)
 	return append(out, p.llamaModel.Tensors(ts)...)
 }
 
 func (p *mixtralModel) Replacements() []string {
 	return append(
 		p.llamaModel.Replacements(),
+		"model.layers", "blk",
 		"block_sparse_moe.gate", "ffn_gate_inp",
+		"block_sparse_moe.experts.", ".",
 	)
 }
-
-type experts []Tensor
-
-func (e experts) WriteTo(w io.Writer) (int64, error) {
-	// TODO(mxyng): experts _should_ be numerically sorted by expert but this should check
-	for _, t := range e {
-		// the canonical merged experts tensor stacks all experts along a new, 0 axis,
-		// e.g. `tensor.Stack(0, e[0], e[1:]...)`, which requires allocating temporary buffers
-		// this accomplishes the same thing by writing each expert tensor in sequence
-		if _, err := t.WriteTo(w); err != nil {
-			return 0, err
-		}
-	}
-
-	return 0, nil
-}

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_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"`
 }
@@ -39,6 +40,8 @@ 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")
 	}

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,
+				split{Replacer: strings.NewReplacer("patch_embed.proj", "patch_embd_0")},
+				split{Replacer: 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,
+				split{Replacer: strings.NewReplacer("attn.qkv", "attn_q")},
+				split{Replacer: strings.NewReplacer("attn.qkv", "attn_k")},
+				split{Replacer: 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)
 	}
@@ -332,7 +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/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 +1,129 @@
+package convert
+
+import (
+	"cmp"
+	"io"
+	"iter"
+	"path"
+	"slices"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type split struct {
+	*strings.Replacer
+	dim int
+
+	// fn is an optional function to apply to the tensor after slicing
+	fn func(tensor.Tensor) (tensor.Tensor, error)
+}
+
+// splitDim splits a tensor along a specified dimension into multiple tensors. The dimension
+// is split evenly based on the number of replacers provided unless a specific count is given.
+func splitDim(t Tensor, dim int, splits ...split) iter.Seq[*ggml.Tensor] {
+	return func(yield func(*ggml.Tensor) bool) {
+		var offset int
+		for _, split := range splits {
+			t := t.Clone()
+			shape := slices.Clone(t.Shape())
+			shape[dim] = cmp.Or(uint64(split.dim), shape[dim]/uint64(len(splits)))
+
+			slice := slices.Repeat([]tensor.Slice{nil}, len(shape))
+			slice[dim] = tensor.S(offset, offset+int(shape[dim]))
+			offset += int(shape[dim])
+
+			t.SetRepacker(func(_ string, data []float32, shape []uint64) ([]float32, error) {
+				dims := make([]int, len(shape))
+				for i := range shape {
+					dims[i] = int(shape[i])
+				}
+
+				var tt tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+				tt, err := tt.Slice(slice...)
+				if err != nil {
+					return nil, err
+				}
+
+				tt = tensor.Materialize(tt)
+
+				if split.fn != nil {
+					tt, err = split.fn(tt)
+					if err != nil {
+						return nil, err
+					}
+				}
+
+				// flatten tensor so it can be written as a vector
+				if err := tt.Reshape(tt.Shape().TotalSize()); err != nil {
+					return nil, err
+				}
+
+				return native.VectorF32(tt.(*tensor.Dense))
+			})
+
+			if !yield(&ggml.Tensor{
+				Name:     split.Replace(t.Name()),
+				Kind:     t.Kind(),
+				Shape:    shape,
+				WriterTo: t,
+			}) {
+				break
+			}
+		}
+	}
+}
+
+type merge struct {
+	pattern, name string
+}
+
+// mergeTensors merges tensors that match a given pattern into a single tensor.
+func mergeTensors(unmatched []Tensor, merges ...merge) (out []*ggml.Tensor, _ []Tensor) {
+	var matched []Tensor
+	for i := range merges {
+		matched, unmatched = slicesSplitFunc(unmatched, func(t Tensor) bool {
+			matched, _ := path.Match(merges[i].pattern, t.Name())
+			return matched
+		})
+
+		if len(matched) > 0 {
+			out = append(out, &ggml.Tensor{
+				Name:     merges[i].name,
+				Kind:     matched[0].Kind(),
+				Shape:    append([]uint64{uint64(len(matched))}, matched[0].Shape()...),
+				WriterTo: mergeGroup(matched),
+			})
+		}
+	}
+
+	return out, unmatched
+}
+
+// slicesSplitFunc splits a slice into two slices based on a predicate function.
+func slicesSplitFunc[S ~[]E, E comparable](s S, fn func(e E) bool) (matched, unmatched S) {
+	for _, e := range s {
+		if fn(e) {
+			matched = append(matched, e)
+		} else {
+			unmatched = append(unmatched, e)
+		}
+	}
+
+	return matched, unmatched
+}
+
+type mergeGroup []Tensor
+
+func (g mergeGroup) WriteTo(w io.Writer) (int64, error) {
+	for _, t := range g {
+		if _, err := t.WriteTo(w); err != nil {
+			return 0, err
+		}
+	}
+
+	return 0, nil
+}

convert/tensor_test.go

@@ -0,0 +1,402 @@
+package convert
+
+import (
+	"bytes"
+	"encoding/binary"
+	"io"
+	"iter"
+	"slices"
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+	"github.com/ollama/ollama/fs/ggml"
+	"github.com/pdevine/tensor"
+)
+
+type fakeTensor struct {
+	name  string
+	shape []uint64
+	data  []float32
+
+	repacker Repacker
+}
+
+func (f fakeTensor) Name() string {
+	return f.name
+}
+
+func (f fakeTensor) Shape() []uint64 {
+	return f.shape
+}
+
+func (f fakeTensor) Kind() uint32 {
+	return 0
+}
+
+func (f *fakeTensor) SetRepacker(fn Repacker) {
+	f.repacker = fn
+}
+
+func (f fakeTensor) Clone() Tensor {
+	return &fakeTensor{
+		name:     f.name,
+		shape:    slices.Clone(f.shape),
+		data:     slices.Clone(f.data),
+		repacker: f.repacker,
+	}
+}
+
+func (f fakeTensor) WriteTo(w io.Writer) (n int64, err error) {
+	data := f.data
+	if f.repacker != nil {
+		data, err = f.repacker(f.name, data, f.shape)
+		if err != nil {
+			return 0, err
+		}
+	}
+
+	if err := binary.Write(w, binary.LittleEndian, data); err != nil {
+		return 0, err
+	}
+
+	return int64(len(data) * 4), nil
+}
+
+func mul(shape []uint64) int {
+	n := 1
+	for _, dim := range shape {
+		n *= int(dim)
+	}
+	return n
+}
+
+func TestSplitDim(t *testing.T) {
+	r := fakeTensor{
+		name:  "a.b",
+		shape: []uint64{3, 4},
+		data:  []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11},
+	}
+
+	t.Run("no split", func(t *testing.T) {
+		for tt := range splitDim(&r, 0, split{Replacer: strings.NewReplacer("a", "x")}) {
+			if tt.Name != "x.b" {
+				t.Fatalf("expected name 'x', got '%s'", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{3, 4}) {
+				t.Fatalf("expected shape [3, 4], got %v", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}) {
+				t.Fatalf("expected data [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11], got %v", f32s)
+			}
+		}
+	})
+
+	t.Run("even split", func(t *testing.T) {
+		next, stop := iter.Pull(splitDim(&r, 1,
+			split{Replacer: strings.NewReplacer("a", "x")},
+			split{Replacer: strings.NewReplacer("b", "y")},
+		))
+		defer stop()
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "x.b" {
+				t.Fatal("expected name 'x.b', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{3, 2}) {
+				t.Fatal("expected shape [3, 2], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{0, 1, 4, 5, 8, 9}) {
+				t.Fatal("expected data [0, 1, 4, 5, 8, 9], got", f32s)
+			}
+		}
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "a.y" {
+				t.Fatal("expected name 'a.y', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{3, 2}) {
+				t.Fatal("expected shape [3, 2], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{2, 3, 6, 7, 10, 11}) {
+				t.Fatal("expected data [2, 3, 6, 7, 10, 11], got", f32s)
+			}
+		}
+	})
+
+	t.Run("uneven split", func(t *testing.T) {
+		next, stop := iter.Pull(splitDim(&r, 0,
+			split{Replacer: strings.NewReplacer("a", "x"), dim: 2},
+			split{Replacer: strings.NewReplacer("b", "y"), dim: 1},
+		))
+		defer stop()
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "x.b" {
+				t.Fatal("expected name 'x.b', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{2, 4}) {
+				t.Fatal("expected shape [2, 4], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{0, 1, 2, 3, 4, 5, 6, 7}) {
+				t.Fatal("expected data [0, 1, 2, 3, 4, 5, 6, 7], got", f32s)
+			}
+		}
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "a.y" {
+				t.Fatal("expected name 'a.y', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{1, 4}) {
+				t.Fatal("expected shape [1, 4], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{8, 9, 10, 11}) {
+				t.Fatal("expected data [8, 9, 10, 11], got", f32s)
+			}
+		}
+	})
+
+	t.Run("split with transpose", func(t *testing.T) {
+		next, stop := iter.Pull(splitDim(&r, 1,
+			split{Replacer: strings.NewReplacer("a", "x")},
+			split{Replacer: strings.NewReplacer("b", "y"), fn: func(tt tensor.Tensor) (tensor.Tensor, error) {
+				return tensor.Transpose(tt, 1, 0)
+			}},
+		))
+		defer stop()
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "x.b" {
+				t.Fatal("expected name 'x.b', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{3, 2}) {
+				t.Fatal("expected shape [3, 2], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{0, 1, 4, 5, 8, 9}) {
+				t.Fatal("expected data [0, 1, 4, 5, 8, 9], got", f32s)
+			}
+		}
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "a.y" {
+				t.Fatal("expected name 'a.y', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{3, 2}) {
+				t.Fatal("expected shape [3, 2], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{2, 6, 10, 3, 7, 11}) {
+				t.Fatal("expected data [2, 6, 10, 3, 7, 11], got", f32s)
+			}
+		}
+	})
+}
+
+func TestMerge(t *testing.T) {
+	unmatched := []Tensor{
+		&fakeTensor{
+			name:  "a.0.b",
+			shape: []uint64{5, 2},
+			data:  []float32{10, 11, 12, 13, 14, 15, 16, 17, 18, 19},
+		},
+		&fakeTensor{
+			name:  "a.1.b",
+			shape: []uint64{5, 2},
+			data:  []float32{20, 21, 22, 23, 24, 25, 26, 27, 28, 29},
+		},
+		&fakeTensor{
+			name:  "c.0.d",
+			shape: []uint64{5, 2},
+			data:  []float32{30, 31, 32, 33, 34, 35, 36, 37, 38, 39},
+		},
+		&fakeTensor{
+			name:  "c.1.d",
+			shape: []uint64{5, 2},
+			data:  []float32{40, 41, 42, 43, 44, 45, 46, 47, 48, 49},
+		},
+		&fakeTensor{
+			name:  "e.0.f",
+			shape: []uint64{5, 2},
+			data:  []float32{50, 51, 52, 53, 54, 55, 56, 57, 58, 59},
+		},
+	}
+
+	checkMatched := func(t *testing.T, n int, matched []*ggml.Tensor) {
+		for i := range n {
+			got := matched[i]
+			if diff := cmp.Diff([]uint64{2, 5, 2}, got.Shape); diff != "" {
+				t.Errorf("unexpected (-want +got):\n%s", diff)
+			}
+
+			var b bytes.Buffer
+			if _, err := got.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, 20)
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			offset := 10 + (i * 20)
+			want := make([]float32, 20)
+			for j := range 20 {
+				want[j] = float32(offset + j)
+			}
+
+			if diff := cmp.Diff(want, f32s); diff != "" {
+				t.Errorf("unexpected data (-want +got):\n%s", diff)
+			}
+		}
+	}
+
+	t.Run("single merge", func(t *testing.T) {
+		matched, unmatched := mergeTensors(unmatched, merge{"a.*.b", "a.b"})
+		if len(unmatched) != 3 {
+			t.Error("expected 3 remaining tensors, got", len(unmatched))
+		}
+
+		if len(matched) != 1 {
+			t.Error("expected 1 merged tensor, got", len(matched))
+		}
+
+		checkMatched(t, 1, matched)
+	})
+
+	t.Run("multiple merges", func(t *testing.T) {
+		matched, unmatched := mergeTensors(unmatched, merge{"a.*.b", "a.b"}, merge{"c.*.d", "c.d"})
+		if len(unmatched) != 1 {
+			t.Error("expected 1 remaining tensors, got", len(unmatched))
+		}
+
+		if len(matched) != 2 {
+			t.Error("expected 2 merged tensor, got", len(matched))
+		}
+
+		checkMatched(t, 2, matched)
+	})
+
+	t.Run("no match", func(t *testing.T) {
+		matched, unmatched := mergeTensors(unmatched, merge{"x.*.y", "x.y"})
+		if len(unmatched) != 5 {
+			t.Error("expected 5 remaining tensors, got", len(unmatched))
+		}
+
+		if len(matched) != 0 {
+			t.Error("expected no merged tensors, got", len(matched))
+		}
+	})
+}

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/cuda_common.go

@@ -3,7 +3,6 @@
 package discover
 
 import (
-	"fmt"
 	"log/slog"
 	"os"
 	"regexp"
@@ -60,8 +59,6 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 
 	// driver 12.0 has problems with the cuda v12 library, so run v11 on those older drivers
 	if gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
-		// The detected driver is older than Feb 2023
-		slog.Warn("old CUDA driver detected - please upgrade to a newer driver", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
 		return "v11"
 	}
 	return "v12"

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/path.go

@@ -12,7 +12,7 @@ import (
 // '../lib/ollama' on Linux and the executable's directory on macOS
 // note: distribution builds, additional GPU-specific libraries are
 // found in subdirectories of the returned path, such as
-// 'cuda_v12', 'rocm', etc.
+// 'cuda_v11', 'cuda_v12', 'rocm', etc.
 var LibOllamaPath string = func() string {
 	exe, err := os.Executable()
 	if err != nil {

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):
 
@@ -490,11 +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
 
@@ -952,19 +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
@@ -1009,8 +1001,8 @@ Quantize a non-quantized model.
 
 ```shell
 curl http://localhost:11434/api/create -d '{
-  "model": "llama3.1:quantized",
+  "model": "llama3.2:quantized",
-  "from": "llama3.1:8b-instruct-fp16",
+  "from": "llama3.2:3b-instruct-fp16",
   "quantize": "q4_K_M"
 }'
 ```
@@ -1020,12 +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":"quantizing F16 model to Q4_K_M","digest":"0","total":6433687776,"completed":12302}
-{"status":"creating new layer sha256:667b0c1932bc6ffc593ed1d03f895bf2dc8dc6df21db3042284a6f4416b06a29"}
+{"status":"quantizing F16 model to Q4_K_M","digest":"0","total":6433687776,"completed":6433687552}
-{"status":"using existing layer sha256:11ce4ee3e170f6adebac9a991c22e22ab3f8530e154ee669954c4bc73061c258"}
+{"status":"verifying conversion"}
-{"status":"using existing layer sha256:0ba8f0e314b4264dfd19df045cde9d4c394a52474bf92ed6a3de22a4ca31a177"}
+{"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"}
 ```
@@ -1163,29 +1157,37 @@ A single JSON object will be returned.
 {
   "models": [
     {
-      "name": "codellama:13b",
+      "name": "deepseek-r1:latest",
-      "modified_at": "2023-11-04T14:56:49.277302595-07:00",
+      "model": "deepseek-r1:latest",
-      "size": 7365960935,
+      "modified_at": "2025-05-10T08:06:48.639712648-07:00",
-      "digest": "9f438cb9cd581fc025612d27f7c1a6669ff83a8bb0ed86c94fcf4c5440555697",
+      "size": 4683075271,
+      "digest": "0a8c266910232fd3291e71e5ba1e058cc5af9d411192cf88b6d30e92b6e73163",
       "details": {
+        "parent_model": "",
         "format": "gguf",
-        "family": "llama",
+        "family": "qwen2",
-        "families": null,
+        "families": [
-        "parameter_size": "13B",
+          "qwen2"
-        "quantization_level": "Q4_0"
+        ],
+        "parameter_size": "7.6B",
+        "quantization_level": "Q4_K_M"
       }
     },
     {
-      "name": "llama3:latest",
+      "name": "llama3.2:latest",
-      "modified_at": "2023-12-07T09:32:18.757212583-08:00",
+      "model": "llama3.2:latest",
-      "size": 3825819519,
+      "modified_at": "2025-05-04T17:37:44.706015396-07:00",
-      "digest": "fe938a131f40e6f6d40083c9f0f430a515233eb2edaa6d72eb85c50d64f2300e",
+      "size": 2019393189,
+      "digest": "a80c4f17acd55265feec403c7aef86be0c25983ab279d83f3bcd3abbcb5b8b72",
       "details": {
+        "parent_model": "",
         "format": "gguf",
         "family": "llama",
-        "families": null,
+        "families": [
-        "parameter_size": "7B",
+          "llama"
-        "quantization_level": "Q4_0"
+        ],
+        "parameter_size": "3.2B",
+        "quantization_level": "Q4_K_M"
       }
     }
   ]

docs/benchmark.md

@@ -1,59 +0,0 @@
-# Benchmark
-
-Go benchmark tests that measure end-to-end performance of a running Ollama server. Run these tests to evaluate model inference performance on your hardware and measure the impact of code changes.
-
-## When to use
-
-Run these benchmarks when:
-- Making changes to the model inference engine
-- Modifying model loading/unloading logic
-- Changing prompt processing or token generation code
-- Implementing a new model architecture
-- Testing performance across different hardware setups
-
-## Prerequisites
-- Ollama server running locally with `ollama serve` on `127.0.0.1:11434`
-## Usage and Examples
-
->[!NOTE]
->All commands must be run from the root directory of the Ollama project.
-
-Basic syntax:
-```bash
-go test -bench=. ./benchmark/... -m $MODEL_NAME
-```
-
-Required flags:
-- `-bench=.`: Run all benchmarks
-- `-m`: Model name to benchmark
-
-Optional flags:
-- `-count N`: Number of times to run the benchmark (useful for statistical analysis)
-- `-timeout T`: Maximum time for the benchmark to run (e.g. "10m" for 10 minutes)
-
-Common usage patterns:
-
-Single benchmark run with a model specified:
-```bash
-go test -bench=. ./benchmark/... -m llama3.3
-```
-
-## Output metrics
-
-The benchmark reports several key metrics:
-
-- `gen_tok/s`: Generated tokens per second
-- `prompt_tok/s`: Prompt processing tokens per second
-- `ttft_ms`: Time to first token in milliseconds
-- `load_ms`: Model load time in milliseconds
-- `gen_tokens`: Total tokens generated
-- `prompt_tokens`: Total prompt tokens processed
-
-Each benchmark runs two scenarios:
-- Cold start: Model is loaded from disk for each test
-- Warm start: Model is pre-loaded in memory
-
-Three prompt lengths are tested for each scenario:
-- Short prompt (100 tokens)
-- Medium prompt (500 tokens)
-- Long prompt (1000 tokens)

docs/development.md

@@ -118,7 +118,7 @@ To run tests, use `go test`:
 go test ./...
 ```
 
-> NOTE: In rare cirumstances, you may nedd to change a package using the new
+> NOTE: In rare cirumstances, you may need to change a package using the new
 > "synctest" package in go1.24.
 >
 > If you do not have the "synctest" package enabled, you will not see build or

docs/gpu.md

@@ -1,6 +1,6 @@
 # GPU
 ## Nvidia
-Ollama supports Nvidia GPUs with compute capability 5.0+ and driver version 531 and newer.
+Ollama supports Nvidia GPUs with compute capability 5.0+.
 
 Check your compute compatibility to see if your card is supported:
 [https://developer.nvidia.com/cuda-gpus](https://developer.nvidia.com/cuda-gpus)

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/linux.md

@@ -112,8 +112,8 @@ sudo systemctl status ollama
 > While AMD has contributed the `amdgpu` driver upstream to the official linux
 > kernel source, the version is older and may not support all ROCm features. We
 > recommend you install the latest driver from
-> https://www.amd.com/en/support/linux-drivers for best support of your Radeon
+> [AMD](https://www.amd.com/en/support/download/linux-drivers.html) for best support
-> GPU.
+> of your Radeon GPU.
 
 ## Customizing
 

docs/troubleshooting.md

@@ -43,7 +43,7 @@ Ollama includes multiple LLM libraries compiled for different GPUs and CPU vecto
 In the server log, you will see a message that looks something like this (varies from release to release):
 
 ```
-Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v12 rocm_v5]
+Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v11 rocm_v5]
 ```
 
 **Experimental LLM Library Override**

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 {
@@ -170,6 +171,8 @@ func (kv KV) OllamaEngineRequired() bool {
 		"gemma3",
 		"mistral3",
 		"llama4",
+		"mllama",
+		"qwen25vl",
 	}, kv.Architecture())
 }
 
@@ -429,12 +432,12 @@ 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
@@ -444,24 +447,25 @@ 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) {
@@ -693,6 +697,20 @@ 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

@@ -527,23 +527,17 @@ func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 		return err
 	}
 
-	keys := slices.Collect(maps.Keys(kv))
+	for _, key := range slices.Sorted(maps.Keys(kv)) {
-	slices.Sort(keys)
-
-	for _, key := range keys {
 		if err := ggufWriteKV(f, key, kv[key]); err != nil {
 			return err
 		}
 	}
 
 	slices.SortStableFunc(ts, func(a, b *Tensor) int {
-		if i, j := a.block(), b.block(); i < 0 && j > 0 {
+		if i, j := a.block(), b.block(); i > 0 && j > 0 {
-			return 1
-		} else if i > 0 && j < 0 {
-			return -1
-		} else {
 			return cmp.Compare(i, j)
 		}
+		return cmp.Compare(a.Name, b.Name)
 	})
 
 	var s uint64

fs/ggml/gguf_test.go

@@ -2,62 +2,82 @@ package ggml
 
 import (
 	"bytes"
+	"math/rand/v2"
 	"os"
-	"slices"
+	"strings"
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
 )
 
 func TestWriteGGUF(t *testing.T) {
-	w, err := os.CreateTemp(t.TempDir(), "*.bin")
+	r := rand.New(rand.NewPCG(0, 0))
-	if err != nil {
+	for range 8 {
-		t.Fatal(err)
+		t.Run("shuffle", func(t *testing.T) {
-	}
+			t.Parallel()
-	defer w.Close()
 
-	if err := WriteGGUF(w, KV{
+			ts := []*Tensor{
-		"general.alignment": uint32(16),
+				{Name: "token_embd.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-	}, []*Tensor{
+				{Name: "blk.0.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-		{Name: "test.0", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+				{Name: "blk.1.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-		{Name: "test.1", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+				{Name: "blk.2.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-		{Name: "test.2", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+				{Name: "blk.3.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-		{Name: "test.3", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+				{Name: "blk.4.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-		{Name: "test.4", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+				{Name: "blk.5.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
-		{Name: "test.5", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+				{Name: "output_norm.weight", Shape: []uint64{3, 2}, WriterTo: bytes.NewBuffer(make([]byte, 3*2))},
-	}); err != nil {
+				{Name: "output.weight", Shape: []uint64{3, 2}, WriterTo: bytes.NewBuffer(make([]byte, 3*2))},
-		t.Fatal(err)
+			}
-	}
 
-	r, err := os.Open(w.Name())
+			r.Shuffle(len(ts), func(i, j int) {
-	if err != nil {
+				ts[i], ts[j] = ts[j], ts[i]
-		t.Fatal(err)
+			})
-	}
-	defer r.Close()
 
-	ff, _, err := Decode(r, 0)
+			w, err := os.CreateTemp(t.TempDir(), strings.ReplaceAll(t.Name(), "/", "_")+"*.bin")
-	if err != nil {
+			if err != nil {
-		t.Fatal(err)
+				t.Fatal(err)
-	}
+			}
+			defer w.Close()
 
-	if diff := cmp.Diff(ff.KV(), KV{
+			if err := WriteGGUF(w, KV{
-		"general.alignment":       uint32(16),
+				"general.alignment": uint32(16),
-		"general.parameter_count": uint64(36),
+			}, ts); err != nil {
-	}); diff != "" {
+				t.Fatal(err)
-		t.Errorf("Mismatch (-want +got):\n%s", diff)
+			}
-	}
 
-	if diff := cmp.Diff(ff.Tensors(), Tensors{
+			r, err := os.Open(w.Name())
-		Offset: 336,
+			if err != nil {
-		items: []*Tensor{
+				t.Fatal(err)
-			{Name: "test.0", Offset: 0, Shape: []uint64{2, 3}},
+			}
-			{Name: "test.1", Offset: 32, Shape: []uint64{2, 3}},
+			defer r.Close()
-			{Name: "test.2", Offset: 64, Shape: []uint64{2, 3}},
+
-			{Name: "test.3", Offset: 96, Shape: []uint64{2, 3}},
+			ff, err := Decode(r, 0)
-			{Name: "test.4", Offset: 128, Shape: []uint64{2, 3}},
+			if err != nil {
-			{Name: "test.5", Offset: 160, Shape: []uint64{2, 3}},
+				t.Fatal(err)
-		},
+			}
-	}, cmp.AllowUnexported(Tensors{})); diff != "" {
+
-		t.Errorf("Mismatch (-want +got):\n%s", diff)
+			if diff := cmp.Diff(KV{
+				"general.alignment":       uint32(16),
+				"general.parameter_count": uint64(54),
+			}, ff.KV()); diff != "" {
+				t.Errorf("Mismatch (-want +got):\n%s", diff)
+			}
+
+			if diff := cmp.Diff(Tensors{
+				Offset: 608,
+				items: []*Tensor{
+					{Name: "blk.0.attn_norm.weight", Offset: 0, Shape: []uint64{2, 3}},
+					{Name: "blk.1.attn_norm.weight", Offset: 32, Shape: []uint64{2, 3}},
+					{Name: "blk.2.attn_norm.weight", Offset: 64, Shape: []uint64{2, 3}},
+					{Name: "blk.3.attn_norm.weight", Offset: 96, Shape: []uint64{2, 3}},
+					{Name: "blk.4.attn_norm.weight", Offset: 128, Shape: []uint64{2, 3}},
+					{Name: "blk.5.attn_norm.weight", Offset: 160, Shape: []uint64{2, 3}},
+					{Name: "output.weight", Offset: 192, Shape: []uint64{3, 2}},
+					{Name: "output_norm.weight", Offset: 224, Shape: []uint64{3, 2}},
+					{Name: "token_embd.weight", Offset: 256, Shape: []uint64{2, 3}},
+				},
+			}, ff.Tensors(), cmp.AllowUnexported(Tensors{})); diff != "" {
+				t.Errorf("Mismatch (-want +got):\n%s", diff)
+			}
+		})
 	}
 }

fs/ggml/type.go

@@ -12,42 +12,42 @@ type FileType uint32
 const (
 	FileTypeF32 FileType = iota
 	FileTypeF16
-	FileTypeQ4_0
+	fileTypeQ4_0
-	FileTypeQ4_1
+	fileTypeQ4_1
 	fileTypeQ4_1_F16 // unused by GGML
 	fileTypeQ4_2     // unused by GGML
 	fileTypeQ4_3     // unused by GGML
 	FileTypeQ8_0
-	FileTypeQ5_0
+	fileTypeQ5_0
-	FileTypeQ5_1
+	fileTypeQ5_1
-	FileTypeQ2_K
+	fileTypeQ2_K
-	FileTypeQ3_K_S
+	fileTypeQ3_K_S
-	FileTypeQ3_K_M
+	fileTypeQ3_K_M
-	FileTypeQ3_K_L
+	fileTypeQ3_K_L
 	FileTypeQ4_K_S
 	FileTypeQ4_K_M
-	FileTypeQ5_K_S
+	fileTypeQ5_K_S
-	FileTypeQ5_K_M
+	fileTypeQ5_K_M
-	FileTypeQ6_K
+	fileTypeQ6_K
-	fileTypeIQ2_XXS // not supported by ollama
+	fileTypeIQ2_XXS
-	fileTypeIQ2_XS  // not supported by ollama
+	fileTypeIQ2_XS
-	FileTypeQ2_K_S
+	fileTypeQ2_K_S
-	fileTypeIQ3_XS  // not supported by ollama
+	fileTypeIQ3_XS
-	fileTypeIQ3_XXS // not supported by ollama
+	fileTypeIQ3_XXS
-	fileTypeIQ1_S   // not supported by ollama
+	fileTypeIQ1_S
-	fileTypeIQ4_NL  // not supported by ollama
+	fileTypeIQ4_NL
-	fileTypeIQ3_S   // not supported by ollama
+	fileTypeIQ3_S
-	fileTypeIQ3_M   // not supported by ollama
+	fileTypeIQ3_M
-	fileTypeIQ2_S   // not supported by ollama
+	fileTypeIQ2_S
-	fileTypeIQ2_M   // not supported by ollama
+	fileTypeIQ2_M
-	fileTypeIQ4_XS  // not supported by ollama
+	fileTypeIQ4_XS
-	fileTypeIQ1_M   // not supported by ollama
+	fileTypeIQ1_M
 	FileTypeBF16
 	fileTypeQ4_0_4_4 // unused by GGML
 	fileTypeQ4_0_4_8 // unused by GGML
 	fileTypeQ4_0_8_8 // unused by GGML
-	fileTypeTQ1_0    // not supported by ollama
+	fileTypeTQ1_0
-	fileTypeTQ2_0    // not supported by ollama
+	fileTypeTQ2_0
 
 	FileTypeUnknown = 1024
 )
@@ -60,36 +60,12 @@ func ParseFileType(s string) (FileType, error) {
 		return FileTypeF32, nil
 	case "F16":
 		return FileTypeF16, nil
-	case "Q4_0":
-		return FileTypeQ4_0, nil
-	case "Q4_1":
-		return FileTypeQ4_1, 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
 	case "Q4_K_S":
 		return FileTypeQ4_K_S, nil
 	case "Q4_K_M", "Q4_K":
 		return FileTypeQ4_K_M, nil
-	case "Q5_K_S":
-		return FileTypeQ5_K_S, nil
-	case "Q5_K_M", "Q5_K":
-		return FileTypeQ5_K_M, nil
-	case "Q6_K":
-		return FileTypeQ6_K, nil
-	case "Q2_K_S":
-		return FileTypeQ2_K_S, nil
 	case "BF16":
 		return FileTypeBF16, nil
 	default:
@@ -111,40 +87,41 @@ func ParseFileType(s string) (FileType, error) {
 }
 
 func (t FileType) String() string {
+	// Note: this routine will return a broader set of file types for existing models
 	switch t {
 	case FileTypeF32:
 		return "F32"
 	case FileTypeF16:
 		return "F16"
-	case FileTypeQ4_0:
+	case fileTypeQ4_0:
 		return "Q4_0"
-	case FileTypeQ4_1:
+	case fileTypeQ4_1:
 		return "Q4_1"
 	case FileTypeQ8_0:
 		return "Q8_0"
-	case FileTypeQ5_0:
+	case fileTypeQ5_0:
 		return "Q5_0"
-	case FileTypeQ5_1:
+	case fileTypeQ5_1:
 		return "Q5_1"
-	case FileTypeQ2_K:
+	case fileTypeQ2_K:
 		return "Q2_K"
-	case FileTypeQ3_K_S:
+	case fileTypeQ3_K_S:
 		return "Q3_K_S"
-	case FileTypeQ3_K_M:
+	case fileTypeQ3_K_M:
 		return "Q3_K_M"
-	case FileTypeQ3_K_L:
+	case fileTypeQ3_K_L:
 		return "Q3_K_L"
 	case FileTypeQ4_K_S:
 		return "Q4_K_S"
 	case FileTypeQ4_K_M:
 		return "Q4_K_M"
-	case FileTypeQ5_K_S:
+	case fileTypeQ5_K_S:
 		return "Q5_K_S"
-	case FileTypeQ5_K_M:
+	case fileTypeQ5_K_M:
 		return "Q5_K_M"
-	case FileTypeQ6_K:
+	case fileTypeQ6_K:
 		return "Q6_K"
-	case FileTypeQ2_K_S:
+	case fileTypeQ2_K_S:
 		return "Q2_K_S"
 	case FileTypeBF16:
 		return "BF16"
@@ -163,35 +140,35 @@ func (ftype FileType) ToTensorType() TensorType {
 		return TensorTypeF32
 	case FileTypeF16:
 		return TensorTypeF16
-	case FileTypeQ4_0:
+	case fileTypeQ4_0:
 		return TensorTypeQ4_0
-	case FileTypeQ4_1:
+	case fileTypeQ4_1:
 		return TensorTypeQ4_1
 	case FileTypeQ8_0:
 		return TensorTypeQ8_0
-	case FileTypeQ5_0:
+	case fileTypeQ5_0:
 		return TensorTypeQ5_0
-	case FileTypeQ5_1:
+	case fileTypeQ5_1:
 		return TensorTypeQ5_1
-	case FileTypeQ2_K:
+	case fileTypeQ2_K:
 		return TensorTypeQ2_K
-	case FileTypeQ3_K_S:
+	case fileTypeQ3_K_S:
 		return TensorTypeQ3_K
-	case FileTypeQ3_K_M:
+	case fileTypeQ3_K_M:
 		return TensorTypeQ3_K
-	case FileTypeQ3_K_L:
+	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:
+	case fileTypeQ5_K_S:
 		return TensorTypeQ5_K
-	case FileTypeQ5_K_M:
+	case fileTypeQ5_K_M:
 		return TensorTypeQ5_K
-	case FileTypeQ6_K:
+	case fileTypeQ6_K:
 		return TensorTypeQ6_K
-	case FileTypeQ2_K_S:
+	case fileTypeQ2_K_S:
 		return TensorTypeQ2_K
 	case FileTypeBF16:
 		return TensorTypeBF16

fs/gguf/gguf.go

@@ -0,0 +1,347 @@
+package gguf
+
+import (
+	"bytes"
+	"cmp"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"iter"
+	"os"
+	"slices"
+	"strings"
+)
+
+const (
+	typeUint8 uint32 = iota
+	typeInt8
+	typeUint16
+	typeInt16
+	typeUint32
+	typeInt32
+	typeFloat32
+	typeBool
+	typeString
+	typeArray
+	typeUint64
+	typeInt64
+	typeFloat64
+)
+
+var ErrUnsupported = errors.New("unsupported")
+
+type File struct {
+	Magic   [4]byte
+	Version uint32
+
+	keyValues *lazy[KeyValue]
+	tensors   *lazy[TensorInfo]
+	offset    int64
+
+	file   *os.File
+	reader *bufferedReader
+	bts    []byte
+}
+
+func Open(path string) (f *File, err error) {
+	f = &File{bts: make([]byte, 4096)}
+	f.file, err = os.Open(path)
+	if err != nil {
+		return nil, err
+	}
+
+	f.reader = newBufferedReader(f.file, 32<<10)
+
+	if err := binary.Read(f.reader, binary.LittleEndian, &f.Magic); err != nil {
+		return nil, err
+	}
+
+	if bytes.Equal(f.Magic[:], []byte("gguf")) {
+		return nil, fmt.Errorf("%w file type %v", ErrUnsupported, f.Magic)
+	}
+
+	if err := binary.Read(f.reader, binary.LittleEndian, &f.Version); err != nil {
+		return nil, err
+	}
+
+	if f.Version < 2 {
+		return nil, fmt.Errorf("%w version %v", ErrUnsupported, f.Version)
+	}
+
+	f.tensors, err = newLazy(f, f.readTensor)
+	if err != nil {
+		return nil, err
+	}
+
+	f.tensors.successFunc = func() error {
+		offset := f.reader.offset
+
+		alignment := cmp.Or(f.KeyValue("general.alignment").Int(), 32)
+		f.offset = offset + (alignment-offset%alignment)%alignment
+		return nil
+	}
+
+	f.keyValues, err = newLazy(f, f.readKeyValue)
+	if err != nil {
+		return nil, err
+	}
+
+	return f, nil
+}
+
+func (f *File) readTensor() (TensorInfo, error) {
+	name, err := readString(f)
+	if err != nil {
+		return TensorInfo{}, err
+	}
+
+	dims, err := read[uint32](f)
+	if err != nil {
+		return TensorInfo{}, err
+	}
+
+	shape := make([]uint64, dims)
+	for i := range dims {
+		shape[i], err = read[uint64](f)
+		if err != nil {
+			return TensorInfo{}, err
+		}
+	}
+
+	type_, err := read[uint32](f)
+	if err != nil {
+		return TensorInfo{}, err
+	}
+
+	offset, err := read[uint64](f)
+	if err != nil {
+		return TensorInfo{}, err
+	}
+
+	return TensorInfo{
+		Name:   name,
+		Offset: offset,
+		Shape:  shape,
+		Type:   TensorType(type_),
+	}, nil
+}
+
+func (f *File) readKeyValue() (KeyValue, error) {
+	key, err := readString(f)
+	if err != nil {
+		return KeyValue{}, err
+	}
+
+	t, err := read[uint32](f)
+	if err != nil {
+		return KeyValue{}, err
+	}
+
+	value, err := func() (any, error) {
+		switch t {
+		case typeUint8:
+			return read[uint8](f)
+		case typeInt8:
+			return read[int8](f)
+		case typeUint16:
+			return read[uint16](f)
+		case typeInt16:
+			return read[int16](f)
+		case typeUint32:
+			return read[uint32](f)
+		case typeInt32:
+			return read[int32](f)
+		case typeUint64:
+			return read[uint64](f)
+		case typeInt64:
+			return read[int64](f)
+		case typeFloat32:
+			return read[float32](f)
+		case typeFloat64:
+			return read[float64](f)
+		case typeBool:
+			return read[bool](f)
+		case typeString:
+			return readString(f)
+		case typeArray:
+			return readArray(f)
+		default:
+			return nil, fmt.Errorf("%w type %d", ErrUnsupported, t)
+		}
+	}()
+	if err != nil {
+		return KeyValue{}, err
+	}
+
+	return KeyValue{
+		Key:   key,
+		Value: Value{value},
+	}, nil
+}
+
+func read[T any](f *File) (t T, err error) {
+	err = binary.Read(f.reader, binary.LittleEndian, &t)
+	return t, err
+}
+
+func readString(f *File) (string, error) {
+	n, err := read[uint64](f)
+	if err != nil {
+		return "", err
+	}
+
+	if int(n) > len(f.bts) {
+		f.bts = make([]byte, n)
+	}
+
+	bts := f.bts[:n]
+	if _, err := io.ReadFull(f.reader, bts); err != nil {
+		return "", err
+	}
+	defer clear(bts)
+
+	return string(bts), nil
+}
+
+func readArray(f *File) (any, error) {
+	t, err := read[uint32](f)
+	if err != nil {
+		return nil, err
+	}
+
+	n, err := read[uint64](f)
+	if err != nil {
+		return nil, err
+	}
+
+	switch t {
+	case typeUint8:
+		return readArrayData[uint8](f, n)
+	case typeInt8:
+		return readArrayData[int8](f, n)
+	case typeUint16:
+		return readArrayData[uint16](f, n)
+	case typeInt16:
+		return readArrayData[int16](f, n)
+	case typeUint32:
+		return readArrayData[uint32](f, n)
+	case typeInt32:
+		return readArrayData[int32](f, n)
+	case typeUint64:
+		return readArrayData[uint64](f, n)
+	case typeInt64:
+		return readArrayData[int64](f, n)
+	case typeFloat32:
+		return readArrayData[float32](f, n)
+	case typeFloat64:
+		return readArrayData[float64](f, n)
+	case typeBool:
+		return readArrayData[bool](f, n)
+	case typeString:
+		return readArrayString(f, n)
+	default:
+		return nil, fmt.Errorf("%w type %d", ErrUnsupported, t)
+	}
+}
+
+func readArrayData[T any](f *File, n uint64) (s []T, err error) {
+	s = make([]T, n)
+	for i := range n {
+		e, err := read[T](f)
+		if err != nil {
+			return nil, err
+		}
+
+		s[i] = e
+	}
+
+	return s, nil
+}
+
+func readArrayString(f *File, n uint64) (s []string, err error) {
+	s = make([]string, n)
+	for i := range n {
+		e, err := readString(f)
+		if err != nil {
+			return nil, err
+		}
+
+		s[i] = e
+	}
+
+	return s, nil
+}
+
+func (f *File) Close() error {
+	f.keyValues.stop()
+	f.tensors.stop()
+	return f.file.Close()
+}
+
+func (f *File) KeyValue(key string) KeyValue {
+	if !strings.HasPrefix(key, "general.") && !strings.HasPrefix(key, "tokenizer.") {
+		key = f.KeyValue("general.architecture").String() + "." + key
+	}
+
+	if index := slices.IndexFunc(f.keyValues.values, func(kv KeyValue) bool {
+		return kv.Key == key
+	}); index >= 0 {
+		return f.keyValues.values[index]
+	}
+
+	for keyValue, ok := f.keyValues.next(); ok; keyValue, ok = f.keyValues.next() {
+		if keyValue.Key == key {
+			return keyValue
+		}
+	}
+
+	return KeyValue{}
+}
+
+func (f *File) NumKeyValues() int {
+	return int(f.keyValues.count)
+}
+
+func (f *File) KeyValues() iter.Seq2[int, KeyValue] {
+	return f.keyValues.All()
+}
+
+func (f *File) TensorInfo(name string) TensorInfo {
+	if index := slices.IndexFunc(f.tensors.values, func(t TensorInfo) bool {
+		return t.Name == name
+	}); index >= 0 {
+		return f.tensors.values[index]
+	}
+
+	// fast-forward through key values if we haven't already
+	_ = f.keyValues.rest()
+	for tensor, ok := f.tensors.next(); ok; tensor, ok = f.tensors.next() {
+		if tensor.Name == name {
+			return tensor
+		}
+	}
+
+	return TensorInfo{}
+}
+
+func (f *File) NumTensors() int {
+	return int(f.tensors.count)
+}
+
+func (f *File) TensorInfos() iter.Seq2[int, TensorInfo] {
+	// fast forward through key values if we haven't already
+	f.keyValues.rest()
+	return f.tensors.All()
+}
+
+func (f *File) TensorReader(name string) (TensorInfo, io.Reader, error) {
+	t := f.TensorInfo(name)
+	if t.NumBytes() == 0 {
+		return TensorInfo{}, nil, fmt.Errorf("tensor %s not found", name)
+	}
+
+	// fast forward through tensor info if we haven't already
+	_ = f.tensors.rest()
+	return t, io.NewSectionReader(f.file, f.offset+int64(t.Offset), t.NumBytes()), nil
+}

fs/gguf/gguf_test.go

@@ -0,0 +1,249 @@
+package gguf_test
+
+import (
+	"bytes"
+	"os"
+	"strconv"
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+	"github.com/google/go-cmp/cmp/cmpopts"
+	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/fs/gguf"
+)
+
+func createBinFile(tb testing.TB) string {
+	tb.Helper()
+	f, err := os.CreateTemp(tb.TempDir(), "")
+	if err != nil {
+		tb.Fatal(err)
+	}
+	defer f.Close()
+
+	kv := ggml.KV{
+		"general.architecture":                   "llama",
+		"llama.block_count":                      uint32(8),
+		"llama.embedding_length":                 uint32(3),
+		"llama.attention.head_count":             uint32(2),
+		"llama.attention.head_count_kv":          uint32(2),
+		"llama.attention.key_length":             uint32(3),
+		"llama.rope.dimension_count":             uint32(4),
+		"llama.rope.freq_base":                   float32(10000.0),
+		"llama.rope.freq_scale":                  float32(1.0),
+		"llama.attention.layer_norm_rms_epsilon": float32(1e-6),
+		"tokenizer.ggml.eos_token_id":            uint32(0),
+		"tokenizer.ggml.eos_token_ids":           []int32{1, 2, 3},
+		"tokenizer.ggml.tokens":                  []string{"hello", "world"},
+		"tokenizer.ggml.scores":                  []float32{0, 1},
+	}
+
+	tensors := []*ggml.Tensor{
+		{
+			Name:     "token_embd.weight",
+			Kind:     0,
+			Shape:    []uint64{2, 3},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*2*3)),
+		},
+		{
+			Name:     "output.weight",
+			Kind:     0,
+			Shape:    []uint64{3, 2},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*3*2)),
+		},
+	}
+
+	for i := range 8 {
+		tensors = append(tensors, &ggml.Tensor{
+			Name:     "blk." + strconv.Itoa(i) + ".attn_q.weight",
+			Kind:     0,
+			Shape:    []uint64{3, 3},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
+		}, &ggml.Tensor{
+			Name:     "blk." + strconv.Itoa(i) + ".attn_k.weight",
+			Kind:     0,
+			Shape:    []uint64{3, 3},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
+		}, &ggml.Tensor{
+			Name:     "blk." + strconv.Itoa(i) + ".attn_v.weight",
+			Kind:     0,
+			Shape:    []uint64{3, 3},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
+		}, &ggml.Tensor{
+			Name:     "blk." + strconv.Itoa(i) + ".attn_output.weight",
+			Kind:     0,
+			Shape:    []uint64{3, 3},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
+		})
+	}
+
+	if err := ggml.WriteGGUF(f, kv, tensors); err != nil {
+		tb.Fatal(err)
+	}
+
+	return f.Name()
+}
+
+func TestRead(t *testing.T) {
+	f, err := gguf.Open(createBinFile(t))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	if got := f.KeyValue("does.not.exist").Valid(); got {
+		t.Errorf(`KeyValue("does.not.exist").Exists() = %v, want false`, got)
+	}
+
+	if got := f.KeyValue("general.architecture").String(); got != "llama" {
+		t.Errorf(`KeyValue("general.architecture").String() = %q, want %q`, got, "llama")
+	}
+
+	if got := f.TensorInfo("token_embd.weight"); got.Name != "token_embd.weight" {
+		t.Errorf(`TensorInfo("token_embd.weight").Name = %q, want %q`, got.Name, "token_embd.weight")
+	} else if diff := cmp.Diff(got.Shape, []uint64{2, 3}); diff != "" {
+		t.Errorf(`TensorInfo("token_embd.weight").Shape mismatch (-got +want):\n%s`, diff)
+	} else if got.Type != gguf.TensorTypeF32 {
+		t.Errorf(`TensorInfo("token_embd.weight").Type = %d, want %d`, got.Type, gguf.TensorTypeF32)
+	}
+
+	if got := f.KeyValue("block_count").Uint(); got != 8 {
+		t.Errorf(`KeyValue("block_count").Uint() = %d, want %d`, got, 8)
+	}
+
+	if diff := cmp.Diff(f.KeyValue("tokenizer.ggml.tokens").Strings(), []string{"hello", "world"}); diff != "" {
+		t.Errorf("KeyValue(\"tokenizer.ggml.tokens\").Strings() mismatch (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(f.KeyValue("tokenizer.ggml.scores").Floats(), []float64{0, 1}); diff != "" {
+		t.Errorf("KeyValue(\"tokenizer.ggml.scores\").Ints() mismatch (-got +want):\n%s", diff)
+	}
+
+	var kvs []string
+	for _, kv := range f.KeyValues() {
+		if !kv.Valid() {
+			t.Error("found invalid key-value pair:", kv)
+		}
+
+		kvs = append(kvs, kv.Key)
+	}
+
+	if len(kvs) != f.NumKeyValues() {
+		t.Errorf("iterated key count = %d, want %d", len(kvs), f.NumKeyValues())
+	}
+
+	if diff := cmp.Diff(kvs, []string{
+		"general.architecture",
+		"llama.block_count",
+		"llama.embedding_length",
+		"llama.attention.head_count",
+		"llama.attention.head_count_kv",
+		"llama.attention.key_length",
+		"llama.rope.dimension_count",
+		"llama.rope.freq_base",
+		"llama.rope.freq_scale",
+		"llama.attention.layer_norm_rms_epsilon",
+		"tokenizer.ggml.eos_token_id",
+		"tokenizer.ggml.eos_token_ids",
+		"tokenizer.ggml.tokens",
+		"tokenizer.ggml.scores",
+	}, cmpopts.SortSlices(strings.Compare)); diff != "" {
+		t.Errorf("KeyValues() mismatch (-got +want):\n%s", diff)
+	}
+
+	var tis []string
+	for _, ti := range f.TensorInfos() {
+		if !ti.Valid() {
+			t.Error("found invalid tensor info:", ti)
+		}
+
+		tis = append(tis, ti.Name)
+	}
+
+	if len(tis) != f.NumTensors() {
+		t.Errorf("iterated tensor count = %d, want %d", len(tis), f.NumTensors())
+	}
+
+	if diff := cmp.Diff(tis, []string{
+		"token_embd.weight",
+		"output.weight",
+		"blk.0.attn_q.weight",
+		"blk.0.attn_k.weight",
+		"blk.0.attn_v.weight",
+		"blk.0.attn_output.weight",
+		"blk.1.attn_q.weight",
+		"blk.1.attn_k.weight",
+		"blk.1.attn_v.weight",
+		"blk.1.attn_output.weight",
+		"blk.2.attn_q.weight",
+		"blk.2.attn_k.weight",
+		"blk.2.attn_v.weight",
+		"blk.2.attn_output.weight",
+		"blk.3.attn_q.weight",
+		"blk.3.attn_k.weight",
+		"blk.3.attn_v.weight",
+		"blk.3.attn_output.weight",
+		"blk.4.attn_q.weight",
+		"blk.4.attn_k.weight",
+		"blk.4.attn_v.weight",
+		"blk.4.attn_output.weight",
+		"blk.5.attn_q.weight",
+		"blk.5.attn_k.weight",
+		"blk.5.attn_v.weight",
+		"blk.5.attn_output.weight",
+		"blk.6.attn_q.weight",
+		"blk.6.attn_k.weight",
+		"blk.6.attn_v.weight",
+		"blk.6.attn_output.weight",
+		"blk.7.attn_q.weight",
+		"blk.7.attn_k.weight",
+		"blk.7.attn_v.weight",
+		"blk.7.attn_output.weight",
+	}, cmpopts.SortSlices(strings.Compare)); diff != "" {
+		t.Errorf("TensorInfos() mismatch (-got +want):\n%s", diff)
+	}
+
+	ti, r, err := f.TensorReader("output.weight")
+	if err != nil {
+		t.Fatalf(`TensorReader("output.weight") error: %v`, err)
+	}
+
+	if ti.Name != "output.weight" {
+		t.Errorf(`TensorReader("output.weight").Name = %q, want %q`, ti.Name, "output.weight")
+	} else if diff := cmp.Diff(ti.Shape, []uint64{3, 2}); diff != "" {
+		t.Errorf(`TensorReader("output.weight").Shape mismatch (-got +want):\n%s`, diff)
+	} else if ti.Type != gguf.TensorTypeF32 {
+		t.Errorf(`TensorReader("output.weight").Type = %d, want %d`, ti.Type, gguf.TensorTypeF32)
+	}
+
+	var b bytes.Buffer
+	if _, err := b.ReadFrom(r); err != nil {
+		t.Fatalf(`ReadFrom TensorReader("output.weight") error: %v`, err)
+	}
+
+	if b.Len() != int(ti.NumBytes()) {
+		t.Errorf(`ReadFrom TensorReader("output.weight") length = %d, want %d`, b.Len(), ti.NumBytes())
+	}
+}
+
+func BenchmarkRead(b *testing.B) {
+	b.ReportAllocs()
+
+	p := createBinFile(b)
+	for b.Loop() {
+		f, err := gguf.Open(p)
+		if err != nil {
+			b.Fatal(err)
+		}
+
+		if got := f.KeyValue("general.architecture").String(); got != "llama" {
+			b.Errorf("got = %q, want %q", got, "llama")
+		}
+
+		// Iterate through some tensors
+		for range f.TensorInfos() {
+		}
+
+		f.Close()
+	}
+}

fs/gguf/keyvalue.go

@@ -0,0 +1,90 @@
+package gguf
+
+import (
+	"reflect"
+	"slices"
+)
+
+type KeyValue struct {
+	Key string
+	Value
+}
+
+func (kv KeyValue) Valid() bool {
+	return kv.Key != "" && kv.Value.value != nil
+}
+
+type Value struct {
+	value any
+}
+
+func value[T any](v Value, kinds ...reflect.Kind) (t T) {
+	vv := reflect.ValueOf(v.value)
+	if slices.Contains(kinds, vv.Kind()) {
+		t = vv.Convert(reflect.TypeOf(t)).Interface().(T)
+	}
+	return
+}
+
+func values[T any](v Value, kinds ...reflect.Kind) (ts []T) {
+	switch vv := reflect.ValueOf(v.value); vv.Kind() {
+	case reflect.Slice:
+		if slices.Contains(kinds, vv.Type().Elem().Kind()) {
+			ts = make([]T, vv.Len())
+			for i := range vv.Len() {
+				ts[i] = vv.Index(i).Convert(reflect.TypeOf(ts[i])).Interface().(T)
+			}
+		}
+	}
+	return
+}
+
+// Int returns Value as a signed integer. If it is not a signed integer, it returns 0.
+func (v Value) Int() int64 {
+	return value[int64](v, reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64)
+}
+
+// Ints returns Value as a signed integer slice. If it is not a signed integer slice, it returns nil.
+func (v Value) Ints() (i64s []int64) {
+	return values[int64](v, reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64)
+}
+
+// Uint converts an unsigned integer value to uint64. If the value is not a unsigned integer, it returns 0.
+func (v Value) Uint() uint64 {
+	return value[uint64](v, reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64)
+}
+
+// Uints returns Value as a unsigned integer slice. If it is not a unsigned integer slice, it returns nil.
+func (v Value) Uints() (u64s []uint64) {
+	return values[uint64](v, reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64)
+}
+
+// Float returns Value as a float. If it is not a float, it returns 0.
+func (v Value) Float() float64 {
+	return value[float64](v, reflect.Float32, reflect.Float64)
+}
+
+// Floats returns Value as a float slice. If it is not a float slice, it returns nil.
+func (v Value) Floats() (f64s []float64) {
+	return values[float64](v, reflect.Float32, reflect.Float64)
+}
+
+// Bool returns Value as a boolean. If it is not a boolean, it returns false.
+func (v Value) Bool() bool {
+	return value[bool](v, reflect.Bool)
+}
+
+// Bools returns Value as a boolean slice. If it is not a boolean slice, it returns nil.
+func (v Value) Bools() (bools []bool) {
+	return values[bool](v, reflect.Bool)
+}
+
+// String returns Value as a string. If it is not a string, it returns an empty string.
+func (v Value) String() string {
+	return value[string](v, reflect.String)
+}
+
+// Strings returns Value as a string slice. If it is not a string slice, it returns nil.
+func (v Value) Strings() (strings []string) {
+	return values[string](v, reflect.String)
+}

fs/gguf/keyvalue_test.go

@@ -0,0 +1,208 @@
+package gguf
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+)
+
+func split(name string, values map[string][]any) (matched []any, unmatched []any) {
+	for key, value := range values {
+		if key == name {
+			matched = value
+		} else {
+			unmatched = append(unmatched, value...)
+		}
+	}
+	return
+}
+
+func TestValue(t *testing.T) {
+	values := map[string][]any{
+		"int64":   {int(42), int8(42), int16(42), int32(42), int64(42)},
+		"uint64":  {uint(42), uint8(42), uint16(42), uint32(42), uint64(42)},
+		"float64": {float32(42), float64(42)},
+		"string":  {"42", "hello"},
+		"bool":    {true, false},
+	}
+
+	t.Run("int64", func(t *testing.T) {
+		matched, unmatched := split("int64", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if i64 := kv.Int(); i64 != 42 {
+				t.Errorf("expected 42, got %d", i64)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if i64 := kv.Int(); i64 != 0 {
+				t.Errorf("expected 42, got %d", i64)
+			}
+		}
+	})
+
+	t.Run("uint64", func(t *testing.T) {
+		matched, unmatched := split("uint64", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if u64 := kv.Uint(); u64 != 42 {
+				t.Errorf("expected 42, got %d", u64)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if u64 := kv.Uint(); u64 != 0 {
+				t.Errorf("expected 42, got %d", u64)
+			}
+		}
+	})
+
+	t.Run("float64", func(t *testing.T) {
+		matched, unmatched := split("float64", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if f64 := kv.Float(); f64 != 42 {
+				t.Errorf("expected 42, got %f", f64)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if f64 := kv.Float(); f64 != 0 {
+				t.Errorf("expected 42, got %f", f64)
+			}
+		}
+	})
+
+	t.Run("string", func(t *testing.T) {
+		matched, unmatched := split("string", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if s := kv.String(); s != v {
+				t.Errorf("expected 42, got %s", s)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if s := kv.String(); s != "" {
+				t.Errorf("expected 42, got %s", s)
+			}
+		}
+	})
+
+	t.Run("bool", func(t *testing.T) {
+		matched, unmatched := split("bool", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if b := kv.Bool(); b != v {
+				t.Errorf("expected true, got %v", b)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if b := kv.Bool(); b != false {
+				t.Errorf("expected false, got %v", b)
+			}
+		}
+	})
+}
+
+func TestValues(t *testing.T) {
+	values := map[string][]any{
+		"int64s":   {[]int{42}, []int8{42}, []int16{42}, []int32{42}, []int64{42}},
+		"uint64s":  {[]uint{42}, []uint8{42}, []uint16{42}, []uint32{42}, []uint64{42}},
+		"float64s": {[]float32{42}, []float64{42}},
+		"strings":  {[]string{"42"}, []string{"hello"}},
+		"bools":    {[]bool{true}, []bool{false}},
+	}
+
+	t.Run("int64s", func(t *testing.T) {
+		matched, unmatched := split("int64s", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if diff := cmp.Diff(kv.Ints(), []int64{42}); diff != "" {
+				t.Errorf("diff: %s", diff)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if i64s := kv.Ints(); i64s != nil {
+				t.Errorf("expected nil, got %v", i64s)
+			}
+		}
+	})
+
+	t.Run("uint64s", func(t *testing.T) {
+		matched, unmatched := split("uint64s", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if diff := cmp.Diff(kv.Uints(), []uint64{42}); diff != "" {
+				t.Errorf("diff: %s", diff)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if u64s := kv.Uints(); u64s != nil {
+				t.Errorf("expected nil, got %v", u64s)
+			}
+		}
+	})
+
+	t.Run("float64s", func(t *testing.T) {
+		matched, unmatched := split("float64s", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if diff := cmp.Diff(kv.Floats(), []float64{42}); diff != "" {
+				t.Errorf("diff: %s", diff)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if f64s := kv.Floats(); f64s != nil {
+				t.Errorf("expected nil, got %v", f64s)
+			}
+		}
+	})
+
+	t.Run("strings", func(t *testing.T) {
+		matched, unmatched := split("strings", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if diff := cmp.Diff(kv.Strings(), v); diff != "" {
+				t.Errorf("diff: %s", diff)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if s := kv.Strings(); s != nil {
+				t.Errorf("expected nil, got %v", s)
+			}
+		}
+	})
+
+	t.Run("bools", func(t *testing.T) {
+		matched, unmatched := split("bools", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if diff := cmp.Diff(kv.Bools(), v); diff != "" {
+				t.Errorf("diff: %s", diff)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if b := kv.Bools(); b != nil {
+				t.Errorf("expected nil, got %v", b)
+			}
+		}
+	})
+}

fs/gguf/lazy.go

@@ -0,0 +1,89 @@
+package gguf
+
+import (
+	"encoding/binary"
+	"iter"
+	"log/slog"
+)
+
+type lazy[T any] struct {
+	count  uint64
+	next   func() (T, bool)
+	stop   func()
+	values []T
+
+	// successFunc is called when all values have been successfully read.
+	successFunc func() error
+}
+
+func newLazy[T any](f *File, fn func() (T, error)) (*lazy[T], error) {
+	it := lazy[T]{}
+	if err := binary.Read(f.reader, binary.LittleEndian, &it.count); err != nil {
+		return nil, err
+	}
+
+	it.values = make([]T, 0)
+	it.next, it.stop = iter.Pull(func(yield func(T) bool) {
+		for i := range it.count {
+			t, err := fn()
+			if err != nil {
+				slog.Error("error reading tensor", "index", i, "error", err)
+				return
+			}
+
+			it.values = append(it.values, t)
+			if !yield(t) {
+				break
+			}
+		}
+
+		if it.successFunc != nil {
+			it.successFunc()
+		}
+	})
+
+	return &it, nil
+}
+
+func (g *lazy[T]) Values() iter.Seq[T] {
+	return func(yield func(T) bool) {
+		for _, v := range g.All() {
+			if !yield(v) {
+				break
+			}
+		}
+	}
+}
+
+func (g *lazy[T]) All() iter.Seq2[int, T] {
+	return func(yield func(int, T) bool) {
+		for i := range int(g.count) {
+			if i < len(g.values) {
+				if !yield(i, g.values[i]) {
+					break
+				}
+			} else {
+				t, ok := g.next()
+				if !ok {
+					break
+				}
+
+				if !yield(i, t) {
+					break
+				}
+			}
+		}
+	}
+}
+
+func (g *lazy[T]) rest() (collected bool) {
+	for {
+		_, ok := g.next()
+		collected = collected || ok
+		if !ok {
+			break
+		}
+	}
+
+	return collected
+}

fs/gguf/reader.go

@@ -0,0 +1,23 @@
+package gguf
+
+import (
+	"bufio"
+	"io"
+)
+
+type bufferedReader struct {
+	offset int64
+	*bufio.Reader
+}
+
+func newBufferedReader(rs io.ReadSeeker, size int) *bufferedReader {
+	return &bufferedReader{
+		Reader: bufio.NewReaderSize(rs, size),
+	}
+}
+
+func (rs *bufferedReader) Read(p []byte) (n int, err error) {
+	n, err = rs.Reader.Read(p)
+	rs.offset += int64(n)
+	return n, err
+}

fs/gguf/tensor.go

@@ -0,0 +1,288 @@
+package gguf
+
+import (
+	"log/slog"
+	"strings"
+)
+
+type TensorInfo struct {
+	Name   string
+	Offset uint64
+	Shape  []uint64
+	Type   TensorType
+}
+
+func (ti TensorInfo) Valid() bool {
+	return ti.Name != "" && ti.NumBytes() > 0
+}
+
+func (ti TensorInfo) NumValues() int64 {
+	var numItems int64 = 1
+	for _, dim := range ti.Shape {
+		numItems *= int64(dim)
+	}
+	return numItems
+}
+
+// NumBytes returns the number of bytes in the tensor.
+func (ti TensorInfo) NumBytes() int64 {
+	return int64(float64(ti.NumValues()) * ti.Type.NumBytes())
+}
+
+func (ti TensorInfo) LogValue() slog.Value {
+	return slog.GroupValue(
+		slog.String("name", ti.Name),
+		slog.Int64("offset", int64(ti.Offset)),
+		slog.Any("shape", ti.Shape),
+		slog.Int64("num_values", ti.NumValues()),
+		slog.Int64("num_bytes", ti.NumBytes()),
+		slog.Any("type", ti.Type),
+	)
+}
+
+type TensorType uint32
+
+const (
+	TensorTypeF32 TensorType = iota
+	TensorTypeF16
+	TensorTypeQ4_0
+	TensorTypeQ4_1
+
+	// unexported // unused in gguf
+	tensorTypeQ4_2
+	tensorTypeQ4_3
+
+	TensorTypeQ5_0
+	TensorTypeQ5_1
+	TensorTypeQ8_0
+	TensorTypeQ8_1
+	TensorTypeQ2_K
+	TensorTypeQ3_K
+	TensorTypeQ4_K
+	TensorTypeQ5_K
+	TensorTypeQ6_K
+	TensorTypeQ8_K
+
+	// unexported // unquantizable by ollama
+	tensorTypeIQ2_XXS
+	tensorTypeIQ2_XS
+	tensorTypeIQ3_XXS
+	tensorTypeIQ1_S
+	tensorTypeIQ4_NL
+	tensorTypeIQ3_S
+	tensorTypeIQ2_S
+	tensorTypeIQ4_XS
+
+	TensorTypeI8
+	TensorTypeI16
+	TensorTypeI32
+	TensorTypeI64
+	TensorTypeF64
+
+	// unexported // unquantizable by ollama
+	tensorTypeIQ1_M
+
+	TensorTypeBF16
+
+	// unexported // unused in gguf
+	tensorTypeQ4_0_4_4
+	tensorTypeQ4_0_4_8
+	tensorTypeQ4_0_8_8
+
+	// unexported // unquantizable by ollama
+	tensorTypeTQ1_0
+	tensorTypeTQ2_0
+
+	// unexported // unused in gguf
+	tensorTypeIQ4_NL_4_4
+	tensorTypeIQ4_NL_4_8
+	tensorTypeIQ4_NL_8_8
+)
+
+func (tt TensorType) NumBytes() float64 {
+	return float64(tt.typeSize()) / float64(tt.blockSize())
+}
+
+func (tt TensorType) typeSize() int64 {
+	switch tt {
+	case TensorTypeF32:
+		return 4
+	case TensorTypeF16:
+		return 2
+	case TensorTypeQ4_0:
+		return 2 + tt.blockSize()/2
+	case TensorTypeQ4_1:
+		return 2 + 2 + tt.blockSize()/2
+	case TensorTypeQ5_0:
+		return 2 + 4 + tt.blockSize()/2
+	case TensorTypeQ5_1:
+		return 2 + 2 + 4 + tt.blockSize()/2
+	case TensorTypeQ8_0:
+		return 2 + tt.blockSize()
+	case TensorTypeQ8_1:
+		return 2 + 2 + tt.blockSize()
+	case TensorTypeQ2_K:
+		return tt.blockSize()/16 + tt.blockSize()/4 + 2 + 2
+	case TensorTypeQ3_K:
+		return tt.blockSize()/8 + tt.blockSize()/4 + 12 + 2
+	case TensorTypeQ4_K:
+		return 2 + 2 + 12 + tt.blockSize()/2
+	case TensorTypeQ5_K:
+		return 2 + 2 + 12 + tt.blockSize()/8 + tt.blockSize()/2
+	case TensorTypeQ6_K:
+		return tt.blockSize()/2 + tt.blockSize()/4 + tt.blockSize()/16 + 2
+	case TensorTypeQ8_K:
+		return 4 + tt.blockSize() + 2*tt.blockSize()/16
+	case tensorTypeIQ2_XXS:
+		return 2 + 2*tt.blockSize()/8
+	case tensorTypeIQ2_XS:
+		return 2 + 2*tt.blockSize()/8 + tt.blockSize()/32
+	case tensorTypeIQ3_XXS:
+		return 2 + tt.blockSize()/4 + tt.blockSize()/8
+	case tensorTypeIQ1_S:
+		return 2 + tt.blockSize()/8 + tt.blockSize()/16
+	case tensorTypeIQ4_NL:
+		return 2 + tt.blockSize()/2
+	case tensorTypeIQ3_S:
+		return 2 + tt.blockSize()/4 + tt.blockSize()/8 + tt.blockSize()/32 + 4
+	case tensorTypeIQ2_S:
+		return 2 + tt.blockSize()/4 + tt.blockSize()/16
+	case tensorTypeIQ4_XS:
+		return 2 + 2 + tt.blockSize()/2 + tt.blockSize()/64
+	case TensorTypeI8:
+		return 1
+	case TensorTypeI16:
+		return 2
+	case TensorTypeI32:
+		return 4
+	case TensorTypeI64:
+		return 8
+	case TensorTypeF64:
+		return 8
+	case tensorTypeIQ1_M:
+		return tt.blockSize()/8 + tt.blockSize()/16 + tt.blockSize()/32
+	case TensorTypeBF16:
+		return 2
+	default:
+		return 0
+	}
+}
+
+func (tt TensorType) blockSize() int64 {
+	switch tt {
+	case TensorTypeF32,
+		TensorTypeF16,
+		TensorTypeI8,
+		TensorTypeI16,
+		TensorTypeI32,
+		TensorTypeI64,
+		TensorTypeF64,
+		TensorTypeBF16:
+		return 1
+	case TensorTypeQ4_0,
+		TensorTypeQ4_1,
+		TensorTypeQ5_0,
+		TensorTypeQ5_1,
+		TensorTypeQ8_0,
+		TensorTypeQ8_1,
+		tensorTypeIQ4_NL:
+		return 32
+	default:
+		return 256
+	}
+}
+
+func (tt TensorType) String() string {
+	switch tt {
+	case TensorTypeF32:
+		return "f32"
+	case TensorTypeF16:
+		return "f16"
+	case TensorTypeQ4_0:
+		return "q4_0"
+	case TensorTypeQ4_1:
+		return "q4_1"
+	case tensorTypeQ4_2:
+		return "q4_2"
+	case tensorTypeQ4_3:
+		return "q4_3"
+	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 tensorTypeIQ2_XXS:
+		return "iq2_xxs"
+	case tensorTypeIQ2_XS:
+		return "iq2_xs"
+	case tensorTypeIQ3_XXS:
+		return "iq3_xxs"
+	case tensorTypeIQ1_S:
+		return "iq1_s"
+	case tensorTypeIQ4_NL:
+		return "iq4_nl"
+	case tensorTypeIQ3_S:
+		return "iq3_s"
+	case tensorTypeIQ2_S:
+		return "iq2_s"
+	case tensorTypeIQ4_XS:
+		return "iq4_xs"
+	case TensorTypeI8:
+		return "i8"
+	case TensorTypeI16:
+		return "i16"
+	case TensorTypeI32:
+		return "i32"
+	case TensorTypeI64:
+		return "i64"
+	case TensorTypeF64:
+		return "f64"
+	case tensorTypeIQ1_M:
+		return "iq1_m"
+	case TensorTypeBF16:
+		return "bf16"
+	case tensorTypeQ4_0_4_4:
+		return "q4_0_4_4"
+	case tensorTypeQ4_0_4_8:
+		return "q4_0_4_8"
+	case tensorTypeQ4_0_8_8:
+		return "q4_0_8_8"
+	case tensorTypeTQ1_0:
+		return "tq1_0"
+	case tensorTypeTQ2_0:
+		return "tq2_0"
+	case tensorTypeIQ4_NL_4_4:
+		return "iq4_nl_4_4"
+	case tensorTypeIQ4_NL_4_8:
+		return "iq4_nl_4_8"
+	case tensorTypeIQ4_NL_8_8:
+		return "iq4_nl_8_8"
+	default:
+		return "unknown"
+	}
+}
+
+func (tt TensorType) LogValue() slog.Value {
+	return slog.GroupValue(
+		slog.Uint64("value", uint64(tt)),
+		slog.String("name", strings.ToUpper(tt.String())),
+		slog.Int64("size", tt.typeSize()),
+		slog.Int64("block_size", tt.blockSize()),
+		slog.Float64("num_bytes", tt.NumBytes()),
+	)
+}

go.mod

@@ -19,7 +19,7 @@ require (
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
 	github.com/dlclark/regexp2 v1.11.4
 	github.com/emirpasic/gods/v2 v2.0.0-alpha
-	github.com/google/go-cmp v0.6.0
+	github.com/google/go-cmp v0.7.0
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c

go.sum

@@ -112,8 +112,8 @@ github.com/google/go-cmp v0.4.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/
 github.com/google/go-cmp v0.5.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.5/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.6/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
-github.com/google/go-cmp v0.6.0 h1:ofyhxvXcZhMsU5ulbFiLKl/XBFqE1GSq7atu8tAmTRI=
+github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
-github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
+github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=
 github.com/google/gofuzz v1.0.0/go.mod h1:dBl0BpW6vV/+mYPU4Po3pmUjxk6FQPldtuIdl/M65Eg=
 github.com/google/uuid v1.1.2/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=

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

@@ -45,6 +45,8 @@ var (
 		"qwen2.5-coder:latest",
 		"qwen:latest",
 		"solar-pro:latest",
+		"codellama:latest",
+		"nous-hermes:latest",
 	}
 )
 

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 = c.buildMask(ctx)
-	c.curMask, err = c.buildMask(ctx)
 
-	return err
+	return nil
 }
 
 func newRange() cellRange {
@@ -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)
+	maskTensor := ctx.Input().FromFloatSlice(mask, length, batchSize)
-	if err != nil {
-		return nil, err
-	}
 
 	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 = c.buildMask(ctx)
-			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))
-			}
 		}
 	}
 }
@@ -652,10 +654,7 @@ func (c *Causal) shift(seq int, beginIndex, offset int32) error {
 		}
 	}
 
-	kShift, err := ctx.Input().FromIntSlice(offsets, len(offsets))
+	kShift := ctx.Input().FromIntSlice(offsets, len(offsets))
-	if err != nil {
-		return err
-	}
 
 	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 = "e1e8e0991ffd9e99a445c6812bb519d5bac9f4b5";
+char const *LLAMA_COMMIT = "de4c07f93783a1a96456a44dc16b9db538ee1618";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/.rsync-filter

@@ -10,11 +10,11 @@ include common/stb_image.*
 include include/
 include include/llama.*
 include include/llama-*.*
-include examples/
+include tools/
-include examples/llava/
+include tools/mtmd/
-include examples/llava/clip.*
+include tools/mtmd/clip.*
-include examples/llava/clip-impl.*
+include tools/mtmd/clip-impl.*
-include examples/llava/llava.*
+include tools/mtmd/llava.*
 include src/
 include src/llama.*
 include src/llama-*.*

llama/llama.cpp/common/common.cpp

@@ -1096,7 +1096,6 @@ 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 +1113,7 @@ 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;
 
     if (params.reranking) {
         cparams.embeddings    = true;
@@ -1565,3 +1565,20 @@ 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.h

@@ -66,7 +66,6 @@ enum llama_example {
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
     LLAMA_EXAMPLE_MAIN,
-    LLAMA_EXAMPLE_INFILL,
     LLAMA_EXAMPLE_EMBEDDING,
     LLAMA_EXAMPLE_PERPLEXITY,
     LLAMA_EXAMPLE_RETRIEVAL,
@@ -96,6 +95,7 @@ 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
@@ -161,6 +161,7 @@ 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,
@@ -323,7 +324,6 @@ struct common_params {
     bool ctx_shift         = true;  // context shift on inifinite text generation
 
     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
@@ -332,6 +332,7 @@ struct common_params {
     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
 
@@ -340,7 +341,7 @@ 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
@@ -409,13 +410,14 @@ 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_positive_file = "tools/cvector-generator/positive.txt";
-    std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
+    std::string cvector_negative_file = "tools/cvector-generator/negative.txt";
 
     bool spm_infill = false; // suffix/prefix/middle pattern for infill
 
@@ -664,3 +666,9 @@ 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/sampling.cpp

@@ -1,6 +1,7 @@
 #include "sampling.h"
 
 #include "common.h"
+#include "log.h"
 
 #include <cmath>
 #include <unordered_map>
@@ -229,51 +230,48 @@ 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) {
+        for (const auto & cnstr : params.samplers) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k        (params.top_k));
+            switch (cnstr) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp         (params.temp));
+                case COMMON_SAMPLER_TYPE_DRY:
-            llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma  (params.top_n_sigma));
+                    {
-        } else {
+                        std::vector<const char *> c_breakers;
-            for (const auto & cnstr : params.samplers) {
+                        c_breakers.reserve(params.dry_sequence_breakers.size());
-                switch (cnstr) {
+                        for (const auto & str : params.dry_sequence_breakers) {
-                    case COMMON_SAMPLER_TYPE_DRY:
+                            c_breakers.push_back(str.c_str());
-                        {
-                            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()));
                         }
-                        break;
+
-                    case COMMON_SAMPLER_TYPE_TOP_K:
+                        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()));
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+                    }
-                        break;
+                    break;
-                    case COMMON_SAMPLER_TYPE_TOP_P:
+                case COMMON_SAMPLER_TYPE_TOP_K:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k       (params.top_k));
-                        break;
+                    break;
-                    case COMMON_SAMPLER_TYPE_MIN_P:
+                case COMMON_SAMPLER_TYPE_TOP_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p       (params.top_p, params.min_keep));
-                        break;
+                    break;
-                    case COMMON_SAMPLER_TYPE_XTC:
+                case COMMON_SAMPLER_TYPE_TOP_N_SIGMA:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma));
-                        break;
+                    break;
-                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                case COMMON_SAMPLER_TYPE_MIN_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p       (params.min_p, params.min_keep));
-                        break;
+                    break;
-                    case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                case COMMON_SAMPLER_TYPE_XTC:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc         (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
-                        break;
+                    break;
-                    case COMMON_SAMPLER_TYPE_INFILL:
+                case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical     (params.typ_p, params.min_keep));
-                        break;
+                    break;
-                    case COMMON_SAMPLER_TYPE_PENALTIES:
+                case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext    (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                        break;
+                    break;
-                    default:
+                case COMMON_SAMPLER_TYPE_INFILL:
-                        GGML_ASSERT(false && "unknown sampler type");
+                    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 +473,7 @@ 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 +489,7 @@ 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 +504,7 @@ 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 +518,7 @@ 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 +535,16 @@ 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 {
+            continue;
-            if (allow_alt_names) {
+        }
-                sampler = sampler_alt_name_map.find(name);
+        if (allow_alt_names) {
-                if (sampler != sampler_alt_name_map.end()) {
+            sampler = sampler_alt_name_map.find(name);
-                    samplers.push_back(sampler->second);
+            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 +556,7 @@ 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 +571,8 @@ 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/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>
@@ -112,6 +113,7 @@ extern "C" {
         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 {
@@ -256,7 +258,6 @@ extern "C" {
 
         llama_token  *  token;
         float        *  embd;
-        int32_t         n_embd;
         llama_pos    *  pos;
         int32_t      *  n_seq_id;
         llama_seq_id ** seq_id;
@@ -352,20 +353,18 @@ 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; // 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 op_offload;  // whether to offload host tensor operations to device
     };
 
     // model quantization parameters
@@ -447,6 +446,10 @@ 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");
 
@@ -461,10 +464,6 @@ 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);
 
@@ -930,14 +929,19 @@ 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.
+    // Process a batch of tokens.
-    // Stores the encoder output internally for later use by the decoder cross-attention layers.
+    // 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
     LLAMA_API int32_t llama_encode(
             struct llama_context * ctx,
               struct llama_batch   batch);
 
+    // Process a batch of tokens.
+    // Requires KV cache.
+    // 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)
@@ -1434,6 +1438,37 @@ 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"           },
@@ -145,7 +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"             },
 
@@ -275,40 +273,6 @@ 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,
         {
@@ -1737,14 +1701,6 @@ 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,
@@ -149,7 +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,
 
@@ -351,14 +349,6 @@ 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

@@ -189,7 +189,7 @@ 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 +203,7 @@ 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 +213,7 @@ 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 +241,7 @@ 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 +265,7 @@ 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) {
@@ -316,7 +320,6 @@ 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 +332,6 @@ 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 +340,6 @@ 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

@@ -70,7 +70,8 @@ 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          },
@@ -202,19 +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]";
+                ss << "[INST]" << trailing_space << content << "[/INST]";
-            }
+            } else {
-            else {
+                ss << trailing_space << content << "</s>";
-                ss << " " << content << "</s>";
             }
         }
     } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1
@@ -447,8 +449,16 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4 || tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
+    } 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|>\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
         for (auto message : chat) {
             std::string role(message->role);
             ss << "<|" << role << "|>" << "\n" << message->content;

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,

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

@@ -8,6 +8,7 @@
 #include "llama-kv-cache.h"
 
 #include "ggml-cpp.h"
+#include "ggml-opt.h"
 
 #include <map>
 #include <vector>
@@ -28,7 +29,12 @@ 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;
@@ -66,7 +72,6 @@ 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,
@@ -130,6 +135,32 @@ 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
@@ -139,50 +170,30 @@ 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);
 
+private:
+    llm_graph_result_ptr graph_build(
+            ggml_context * ctx,
+             ggml_cgraph * gf,
+      const llama_ubatch & ubatch,
+          llm_graph_type   gtype);
+
     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) 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);
@@ -199,14 +210,10 @@ 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;
 
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     size_t  logits_size = 0; // capacity (of floats) for logits
@@ -233,6 +240,9 @@ 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.h

@@ -29,8 +29,8 @@ 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-graph.cpp

@@ -284,24 +284,7 @@ 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;
+            data[i] = kv_self->s_copy(i);
-
-            //////////////////////////////////////////////
-            // 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;
-            }
         }
     }
 }
@@ -317,18 +300,7 @@ 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;
+            data[i] = kv_self->s_mask(i);
-
-            //////////////////////////////////////////////
-            // 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;
-            }
         }
     }
 }
@@ -560,12 +532,6 @@ 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
 //
@@ -816,7 +782,7 @@ 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);
     }
@@ -1005,6 +971,7 @@ 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);
 
@@ -1111,7 +1078,7 @@ 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 llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     auto inp = std::make_unique<llm_graph_input_s_copy>(kv_self);
 
@@ -1128,7 +1095,7 @@ 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 llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     auto inp = std::make_unique<llm_graph_input_s_mask>(kv_self);
 
@@ -1261,8 +1228,19 @@ 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);
@@ -1442,8 +1420,6 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // store to KV cache
     {
-        GGML_ASSERT(!kv_self->recurrent);
-
         const auto kv_head = kv_self->head;
 
         GGML_ASSERT(kv_self->size == n_ctx);
@@ -1538,25 +1514,6 @@ 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,
@@ -1612,7 +1569,7 @@ 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 llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     const auto n_kv    = kv_self->n;
     const auto kv_head = kv_self->head;
@@ -1644,7 +1601,7 @@ 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 llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     const auto token_shift_count = hparams.token_shift_count;
 
@@ -1665,7 +1622,7 @@ 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 llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     const auto token_shift_count = hparams.token_shift_count;
     const auto n_embd = hparams.n_embd;

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

@@ -19,6 +19,7 @@ struct llama_cparams;
 
 class llama_memory_i;
 class llama_kv_cache_unified;
+class llama_kv_cache_recurrent;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -86,7 +87,6 @@ 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 {
@@ -187,26 +187,26 @@ 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 * kv_self) : kv_self(kv_self) {}
     virtual ~llm_graph_input_s_copy() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * s_copy; // I32 [kv_size]
 
-    const llama_kv_cache_unified * kv_self;
+    const llama_kv_cache_recurrent * kv_self;
 };
 
 class llm_graph_input_s_mask : public llm_graph_input_i {
 public:
-    llm_graph_input_s_mask(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    llm_graph_input_s_mask(const llama_kv_cache_recurrent * kv_self) : kv_self(kv_self) {}
     virtual ~llm_graph_input_s_mask() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * s_mask; // F32 [1, n_kv]
 
-    const llama_kv_cache_unified * kv_self;
+    const llama_kv_cache_recurrent * kv_self;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -284,16 +284,6 @@ 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
 //
@@ -308,6 +298,7 @@ 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;
@@ -322,6 +313,7 @@ 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; }
@@ -338,6 +330,7 @@ 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;
@@ -361,8 +354,8 @@ struct llm_graph_params {
     const llama_cparams & cparams;
     const llama_ubatch  & ubatch;
 
-    ggml_backend_sched * sched;
+    ggml_backend_sched_t sched;
-    ggml_backend * backend_cpu;
+    ggml_backend_t backend_cpu;
 
     const llama_adapter_cvec  * cvec;
     const llama_adapter_loras * loras;
@@ -413,9 +406,9 @@ 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;
@@ -502,7 +495,6 @@ 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;

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

@@ -85,7 +85,3 @@ bool llama_hparams::is_swa(uint32_t il) const {
 
     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
@@ -44,7 +42,6 @@ struct llama_hparams {
     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 +56,6 @@ 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;
@@ -163,9 +159,6 @@ 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;
 };
 

llama/llama.cpp/src/llama-kv-cache.h

@@ -2,32 +2,72 @@
 
 #include "llama.h"
 #include "llama-io.h"
+#include "llama-graph.h"
 #include "llama-memory.h"
 
 #include "ggml-cpp.h"
 
-#include <functional>
 #include <set>
 #include <vector>
 
 struct llama_cparams;
 struct llama_hparams;
 struct llama_ubatch;
+struct llama_sbatch;
+struct llama_model;
+struct llama_context;
 
 struct llama_kv_cache : public llama_memory_i {
-    using llama_memory_i::llama_memory_i;
+    virtual ~llama_kv_cache() = default;
 
-    virtual void restore() = 0; // call if batch processing fails - restores the cache state
+    // call if batch processing fails - restores the cache state
-    virtual void commit() = 0;  // call after successful batch processing - clears any pending state
+    virtual void restore() = 0;
 
-    virtual int32_t get_n_tokens()   const = 0;
+    // call after successful batch processing - clears any pending state
-    virtual int32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
+    virtual void commit()  = 0;
 
-    virtual bool get_can_shift() const = 0;
+    // process any pending defrag/shift/etc. operations
+    // optionally call once before processing a new batch
+    virtual bool update(llama_context & lctx) = 0;
+
+    // schedule a defrag if the fragmentation threshold is exceeded. otherwise, do nothing
+    virtual void defrag_sched(float thold) = 0;
+
+    // simulate full cache, used for allocating worst-case compute buffers
+    virtual void set_full() = 0;
+
+    //
+    // batch processing
+    //
+
+    virtual llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) = 0;
+
+    // different KV caches require different batch splitting strategies
+    virtual llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const = 0;
+
+    // find an empty slot of size "n_tokens" in the cache
+    virtual bool find_slot(const llama_ubatch & batch) = 0;
+
+    // getters
+    virtual int32_t   get_n_tokens()   const = 0;
+    virtual int32_t   get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
+    virtual llama_pos get_pos_max()    const = 0;
+    virtual bool      get_can_shift()  const = 0;
 
     bool get_can_edit() const override { return get_can_shift(); }
+
+    //
+    // state write/read
+    //
+
+    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
+    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) = 0;
 };
 
+//
+// llama_kv_cache_guard
+//
+
 struct llama_kv_cache_guard {
     llama_kv_cache_guard(llama_kv_cache * kv) : kv(kv) {}
 
@@ -42,7 +82,7 @@ struct llama_kv_cache_guard {
 private:
     llama_kv_cache * kv;
 };
-
+ 
 // block of KV slots to move when defragging
 struct llama_kv_defrag_move {
     uint32_t src;
@@ -50,65 +90,50 @@ struct llama_kv_defrag_move {
     uint32_t len;
 };
 
-struct llama_kv_cell {
+//
-    llama_pos pos   = -1;
+// llama_kv_cache_unified
-    llama_pos delta =  0;
+//
-    int32_t   src   = -1; // used by recurrent state models to copy states
-    int32_t   tail  = -1;
 
-    std::set<llama_seq_id> seq_id;
-
-    bool has_seq_id(const llama_seq_id & id) const {
-        return seq_id.find(id) != seq_id.end();
-    }
-
-    bool is_empty() const {
-        return seq_id.empty();
-    }
-
-    bool is_same_seq(const llama_kv_cell & other) const {
-        return seq_id == other.seq_id;
-    }
-};
-
-// ring-buffer of cached KV data
-// TODO: pimpl
 // TODO: add notion of max sequences
 class llama_kv_cache_unified : public llama_kv_cache {
 public:
-    // can be used to query data from the model if needed
+    struct kv_cell {
-    struct callbacks {
+        llama_pos pos   = -1;
-        std::function<ggml_tensor * (uint32_t n_ctx_per_seq, int il)> get_rope_factors;
+        llama_pos delta =  0;
+
+        std::set<llama_seq_id> seq_id;
+
+        bool has_seq_id(const llama_seq_id & id) const {
+            return seq_id.find(id) != seq_id.end();
+        }
+
+        bool is_empty() const {
+            return seq_id.empty();
+        }
+
+        bool is_same_seq(const kv_cell & other) const {
+            return seq_id == other.seq_id;
+        }
     };
 
+    static uint32_t get_padding(const llama_cparams & cparams);
+
     llama_kv_cache_unified(
-            const llama_hparams & hparams,
+            const llama_model & model,
-            callbacks             cbs);
-
-    virtual ~llama_kv_cache_unified() = default;
-
-    // TODO: become constructor
-    bool init(
-            const llama_model & model,   // TODO: do not reference the model
-          const llama_cparams & cparams,
                     ggml_type   type_k,
                     ggml_type   type_v,
+                         bool   v_trans,
+                         bool   offload,
                      uint32_t   kv_size,
-                         bool   offload);
+                     uint32_t   padding);
 
-    int32_t get_n_tokens()   const override;
+    ~llama_kv_cache_unified() = default;
-    int32_t get_used_cells() const override;
 
-    size_t total_size() const;
+    //
-
+    // llama_memory_i
-    // TODO: better data structures to reduce the cost of this operation
+    //
-    llama_pos pos_max() const;
 
     void clear() override;
-    void defrag() override;
-
-    virtual void restore() override;
-    virtual void commit() override;
 
     bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
     void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
@@ -118,63 +143,40 @@ public:
 
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
-    bool get_can_shift() const override;
+    //
+    // llama_kv_cache
+    //
+
+    void restore() override;
+    void commit()  override;
+
+    bool update(llama_context & ctx) override;
+
+    void defrag_sched(float thold) override;
+
+    void set_full() override;
+
+    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
+
+    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
 
-    // find an empty slot of size "n_tokens" in the cache
     // updates the cache head
     // Note: On success, it's important that cache.head points
     // to the first cell of the slot.
-    bool find_slot(const llama_ubatch & batch);
+    bool find_slot(const llama_ubatch & batch) override;
 
-    // TODO: maybe not needed
+    int32_t get_n_tokens()   const override;
-    uint32_t get_padding(const llama_cparams & cparams) const;
+    int32_t get_used_cells() const override;
 
-    // find how many cells are currently in use
+    // TODO: better data structures to reduce the cost of this operation
-    uint32_t cell_max() const;
+    llama_pos get_pos_max() const override;
 
-    size_t size_k_bytes() const;
+    bool get_can_shift() const override;
-    size_t size_v_bytes() const;
-
-    // defrag
-
-    struct {
-        std::vector<llama_kv_defrag_move> moves;
-    } defrag_info;
-
-    // return true if cells have been moved
-    bool defrag_prepare(int32_t n_max_nodes);
-
-    // commit/restore cache
-
-    struct slot_range {
-        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
-        uint32_t c1 = 0;
-    };
-
-    // pending cell updates that are not yet committed
-    struct {
-        std::vector<slot_range> ranges;
-    } pending;
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1);
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
-
-    // members
-
-    const llama_hparams & hparams;
-
-    callbacks cbs;
-
-    bool has_shift = false;
-    bool do_defrag = false;
-
-    // TODO: remove this and implement llama_kv_cache_recurrent instead
-    bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token
-
-    bool v_trans   = true;  // the value tensor is transposed
-    bool can_shift = false;
 
     // Note: The value of head isn't only used to optimize searching
     // for a free KV slot. llama_decode_impl also uses it, so it
@@ -186,18 +188,214 @@ public:
     // computed before each graph build
     uint32_t n = 0;
 
-    std::vector<llama_kv_cell> cells;
+    std::vector<kv_cell> cells;
 
     std::vector<ggml_tensor *> k_l; // per layer
     std::vector<ggml_tensor *> v_l;
 
 private:
+    const llama_model & model;
+    const llama_hparams & hparams;
+
+    bool has_shift = false;
+    bool do_defrag = false;
+
+    bool v_trans   = true;  // the value tensor is transposed
+    bool can_shift = false;
+
+    // required padding
+    uint32_t padding = 1;
+
     ggml_type type_k = GGML_TYPE_F16;
     ggml_type type_v = GGML_TYPE_F16;
 
     std::vector<ggml_context_ptr>        ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
 
+    // defrag
+    struct {
+        std::vector<llama_kv_defrag_move> moves;
+    } defrag_info;
+
+    // return true if cells have been moved
+    bool defrag_prepare(int32_t n_max_nodes);
+
+    // commit/restore cache
+    struct slot_range {
+        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
+        uint32_t c1 = 0;
+    };
+
+    // pending cell updates that are not yet committed
+    struct {
+        std::vector<slot_range> ranges;
+    } pending;
+
+    // find how many cells are currently in use
+    uint32_t cell_max() const;
+
+    size_t total_size() const;
+
+    size_t size_k_bytes() const;
+    size_t size_v_bytes() const;
+
+    ggml_tensor * build_rope_shift(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_tensor * cur,
+                    ggml_tensor * shift,
+                    ggml_tensor * factors,
+                          float   freq_base,
+                          float   freq_scale) const;
+
+    llm_graph_result_ptr build_graph_shift(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_cgraph * gf) const;
+
+    llm_graph_result_ptr build_graph_defrag(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_cgraph * gf,
+                    const std::vector<llama_kv_defrag_move> & moves) const;
+
+    void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
+    void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
+
+    bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
+    bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
+};
+
+//
+// llama_kv_cache_recurrent
+//
+
+class llama_kv_cache_recurrent : public llama_kv_cache {
+public:
+    struct kv_cell {
+        llama_pos pos  = -1;
+        int32_t   src  = -1; // used to copy states
+        int32_t   tail = -1;
+
+        std::set<llama_seq_id> seq_id;
+
+        bool has_seq_id(const llama_seq_id & id) const {
+            return seq_id.find(id) != seq_id.end();
+        }
+
+        bool is_empty() const {
+            return seq_id.empty();
+        }
+
+        bool is_same_seq(const kv_cell & other) const {
+            return seq_id == other.seq_id;
+        }
+    };
+
+    llama_kv_cache_recurrent(
+            const llama_model & model,
+                    ggml_type   type_k,
+                    ggml_type   type_v,
+                         bool   offload,
+                     uint32_t   kv_size);
+
+    ~llama_kv_cache_recurrent() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    void clear() override;
+
+    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
+    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
+    void seq_keep(llama_seq_id seq_id) override;
+    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos delta) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    //
+    // llama_kv_cache
+    //
+
+    void restore() override;
+    void commit()  override;
+
+    bool update(llama_context & lctx) override;
+
+    void defrag_sched(float thold) override;
+
+    void set_full() override;
+
+    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
+
+    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
+
+    bool find_slot(const llama_ubatch & batch) override;
+
+    int32_t get_n_tokens()   const override;
+    int32_t get_used_cells() const override;
+
+    // TODO: better data structures to reduce the cost of this operation
+    llama_pos get_pos_max() const override;
+
+    bool get_can_shift() const override;
+
+    // TODO: temporary methods - they are not really const as they do const_cast<>, fix this
+    int32_t s_copy(int i) const;
+    float   s_mask(int i) const;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
+
+    // Note: The value of head isn't only used to optimize searching
+    // for a free KV slot. llama_decode_impl also uses it, so it
+    // cannot be freely changed after a slot has been allocated.
+    uint32_t head = 0;
+    uint32_t size = 0;
+    uint32_t used = 0; // used cells (i.e. at least one seq_id)
+
+    // computed before each graph build
+    uint32_t n = 0;
+
+    std::vector<kv_cell> cells;
+
+    std::vector<ggml_tensor *> k_l; // per layer
+    std::vector<ggml_tensor *> v_l;
+
+private:
+    //const llama_model & model;
+    const llama_hparams & hparams;
+
+    // commit/restore cache
+    // TODO: rework for recurrent cache
+    struct slot_range {
+        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
+        uint32_t c1 = 0;
+    };
+
+    // pending cell updates that are not yet committed
+    struct {
+        std::vector<slot_range> ranges;
+    } pending;
+
+    ggml_type type_k = GGML_TYPE_F16;
+    ggml_type type_v = GGML_TYPE_F16;
+
+    std::vector<ggml_context_ptr>        ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    // find how many cells are currently in use
+    uint32_t cell_max() const;
+
+    size_t total_size() const;
+
+    size_t size_k_bytes() const;
+    size_t size_v_bytes() const;
+
     void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
     void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
 
@@ -205,11 +403,6 @@ private:
     bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
 };
 
-// TODO: temporary reusing llama_kv_cache_unified -- implement recurrent cache and simplify llama_kv_cache_unified
-//class llama_kv_cache_recurrent : public llama_kv_cache_unified {
-//public:
-//    using llama_kv_cache_unified::llama_kv_cache_unified;
-//};
 
 //
 // kv cache view

llama/llama.cpp/src/llama-memory.h

@@ -2,12 +2,22 @@
 
 #include "llama.h"
 
+struct llama_memory_params {
+    // kv cache
+    ggml_type type_k;
+    ggml_type type_v;
+
+    // parameters for other types of memory
+    // ...
+};
+
 // general concept of LLM memory
 // the KV cache is a type of LLM memory, but there can be other types
 class llama_memory_i {
 public:
+    virtual ~llama_memory_i() = default;
+
     virtual void clear() = 0;
-    virtual void defrag() = 0;
 
     virtual bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) = 0;
     virtual void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0;

llama/llama.cpp/src/llama-model-loader.cpp

@@ -301,12 +301,12 @@ namespace GGUFMeta {
             GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
 
         switch (arr_info.gt) {
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T, float>::value)); break;
+            case GGUF_TYPE_UINT32:
-            case GGUF_TYPE_INT32:   GGML_ASSERT(
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
-                                            (std::is_same<T,  int32_t>::value) ||
+                                                (std::is_same<T, uint32_t>::value)); break;
-                                            (std::is_same<T, uint32_t>::value));  break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
         }
 
         result.resize(arr_info.length);
@@ -315,8 +315,6 @@ namespace GGUFMeta {
         return true;
     }
 
-    template bool llama_model_loader::get_arr<std::array<unsigned int, 512>>(enum llm_kv kid, std::array<unsigned int, 512>& result, bool required);
-
     template<typename T, size_t N_MAX>
     bool llama_model_loader::get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required) {
         const int kid = gguf_find_key(meta.get(), key.c_str());
@@ -332,12 +330,12 @@ namespace GGUFMeta {
             GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
 
         switch (arr_info.gt) {
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T, float>::value)); break;
+            case GGUF_TYPE_UINT32:
-            case GGUF_TYPE_INT32:   GGML_ASSERT(
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
-                                            (std::is_same<T,  int32_t>::value) ||
+                                                (std::is_same<T, uint32_t>::value)); break;
-                                            (std::is_same<T, uint32_t>::value));  break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
         }
 
         if (arr_info.length > N_MAX) {
@@ -826,6 +824,10 @@ void llama_model_loader::init_mappings(bool prefetch, llama_mlocks * mlock_mmaps
         mmaps_used.reserve(files.size());
         for (const auto & file : files) {
             auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
+            if (!reg) {
+                throw std::runtime_error(format("%s: no CPU backend found", __func__));
+            }
+
             auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
             std::unique_ptr<llama_mmap> mapping = std::make_unique<llama_mmap>(file.get(), prefetch ? -1 : 0, is_numa_fn());
             mmaps_used.emplace_back(mapping->size(), 0);

llama/llama.cpp/src/llama-model-saver.cpp

@@ -0,0 +1,281 @@
+#include "llama-model-saver.h"
+
+#include "gguf.h"
+
+#include "llama.h"
+#include "llama-hparams.h"
+#include "llama-model.h"
+#include "llama-vocab.h"
+
+#include <string>
+
+llama_model_saver::llama_model_saver(const struct llama_model & model) : model(model), llm_kv(model.arch) {
+    gguf_ctx = gguf_init_empty();
+}
+
+llama_model_saver::~llama_model_saver() {
+    gguf_free(gguf_ctx);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const uint32_t value) {
+    gguf_set_val_u32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const int32_t value) {
+    gguf_set_val_i32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const float value) {
+    gguf_set_val_f32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const bool value) {
+    gguf_set_val_bool(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const char * value) {
+    gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+[[noreturn]]
+void llama_model_saver::add_kv(const enum llm_kv key, const char value) {
+    GGML_UNUSED(key);
+    GGML_UNUSED(value);
+    GGML_ABORT("fatal error"); // this should never be called, only needed to make the template below compile
+}
+
+template <typename Container>
+void llama_model_saver::add_kv(const enum llm_kv key, const Container & value, const bool per_layer) {
+    const size_t n_values = per_layer ? size_t(model.hparams.n_layer) : value.size();
+    GGML_ASSERT(n_values <= value.size());
+
+    if (n_values == 0) {
+        return;
+    }
+
+    if (per_layer) {
+        bool all_values_the_same = true;
+        for (size_t i = 1; i < n_values; ++i) {
+            if (value[i] != value[0]) {
+                all_values_the_same = false;
+                break;
+            }
+        }
+        if (all_values_the_same) {
+            add_kv(key, value[0]);
+            return;
+        }
+    }
+
+    if (std::is_same<typename Container::value_type, uint8_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT8, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, int8_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT8, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, uint32_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT32, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, int32_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT32, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, float>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_FLOAT32, value.data(), n_values);
+    } else if (std::is_same<Container, std::string>::value) {
+        gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), reinterpret_cast<const char *>(value.data()));
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const std::vector<std::string> & value) {
+    std::vector<const char *> tmp(value.size());
+    for (size_t i = 0; i < value.size(); ++i) {
+        tmp[i] = value[i].c_str();
+    }
+    gguf_set_arr_str(gguf_ctx, llm_kv(key).c_str(), tmp.data(), tmp.size());
+}
+
+void llama_model_saver::add_tensor(const struct ggml_tensor * tensor) {
+    if (!tensor) {
+        return;
+    }
+    if (gguf_find_tensor(gguf_ctx, tensor->name) >= 0) {
+        GGML_ASSERT(std::string(tensor->name) == "rope_freqs.weight"); // FIXME
+        return;
+    }
+    gguf_add_tensor(gguf_ctx, tensor);
+}
+
+void llama_model_saver::add_kv_from_model() {
+    const llama_hparams & hparams = model.hparams;
+    const llama_vocab   & vocab   = model.vocab;
+
+    const int32_t n_vocab = vocab.n_tokens();
+    std::vector<std::string> tokens(n_vocab);
+    std::vector<float>       scores(n_vocab);
+    std::vector<int32_t>     token_types(n_vocab);
+
+    for (int32_t id = 0; id < n_vocab; ++id) {
+        const llama_vocab::token_data & token_data = vocab.get_token_data(id);
+
+        tokens[id] = token_data.text;
+        scores[id] = token_data.score;
+
+        switch(token_data.attr) {
+            case LLAMA_TOKEN_ATTR_UNKNOWN:      token_types[id] = LLAMA_TOKEN_TYPE_UNKNOWN;      break;
+            case LLAMA_TOKEN_ATTR_UNUSED:       token_types[id] = LLAMA_TOKEN_TYPE_UNUSED;       break;
+            case LLAMA_TOKEN_ATTR_NORMAL:       token_types[id] = LLAMA_TOKEN_TYPE_NORMAL;       break;
+            case LLAMA_TOKEN_ATTR_CONTROL:      token_types[id] = LLAMA_TOKEN_TYPE_CONTROL;      break;
+            case LLAMA_TOKEN_ATTR_USER_DEFINED: token_types[id] = LLAMA_TOKEN_TYPE_USER_DEFINED; break;
+            case LLAMA_TOKEN_ATTR_BYTE:         token_types[id] = LLAMA_TOKEN_TYPE_BYTE;         break;
+            case LLAMA_TOKEN_ATTR_UNDEFINED:
+            default:                            token_types[id] = LLAMA_TOKEN_TYPE_UNDEFINED;    break;
+        }
+    }
+
+    // add_kv(LLM_KV_GENERAL_TYPE,                      ???);
+    add_kv(LLM_KV_GENERAL_ARCHITECTURE,              model.arch_name());
+    // add_kv(LLM_KV_GENERAL_QUANTIZATION_VERSION,      ???);
+    // add_kv(LLM_KV_GENERAL_ALIGNMENT,                 ???);
+    add_kv(LLM_KV_GENERAL_NAME,                      model.name);
+    // add_kv(LLM_KV_GENERAL_AUTHOR,                    ???);
+    // add_kv(LLM_KV_GENERAL_VERSION,                   ???);
+    // add_kv(LLM_KV_GENERAL_URL,                       ???);
+    // add_kv(LLM_KV_GENERAL_DESCRIPTION,               ???);
+    // add_kv(LLM_KV_GENERAL_LICENSE,                   ???);
+    // add_kv(LLM_KV_GENERAL_SOURCE_URL,                ???);
+    // add_kv(LLM_KV_GENERAL_SOURCE_HF_REPO,            ???);
+
+    add_kv(LLM_KV_VOCAB_SIZE,                        vocab.n_tokens());
+    add_kv(LLM_KV_CONTEXT_LENGTH,                    hparams.n_ctx_train);
+    add_kv(LLM_KV_EMBEDDING_LENGTH,                  hparams.n_embd);
+    add_kv(LLM_KV_BLOCK_COUNT,                       hparams.n_layer);
+    add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead);
+    add_kv(LLM_KV_FEED_FORWARD_LENGTH,               hparams.n_ff_arr, true);
+    add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    add_kv(LLM_KV_USE_PARALLEL_RESIDUAL,             hparams.use_par_res);
+    // add_kv(LLM_KV_TENSOR_DATA_LAYOUT,                ???);
+    add_kv(LLM_KV_EXPERT_COUNT,                      hparams.n_expert);
+    add_kv(LLM_KV_EXPERT_USED_COUNT,                 hparams.n_expert_used);
+    add_kv(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);
+    add_kv(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale);
+    add_kv(LLM_KV_POOLING_TYPE,                      uint32_t(hparams.pooling_type));
+    add_kv(LLM_KV_LOGIT_SCALE,                       hparams.f_logit_scale);
+    add_kv(LLM_KV_DECODER_START_TOKEN_ID,            hparams.dec_start_token_id);
+    add_kv(LLM_KV_ATTN_LOGIT_SOFTCAPPING,            hparams.f_attn_logit_softcapping);
+    add_kv(LLM_KV_FINAL_LOGIT_SOFTCAPPING,           hparams.f_final_logit_softcapping);
+    add_kv(LLM_KV_SWIN_NORM,                         hparams.swin_norm);
+    add_kv(LLM_KV_RESCALE_EVERY_N_LAYERS,            hparams.rescale_every_n_layers);
+    add_kv(LLM_KV_TIME_MIX_EXTRA_DIM,                hparams.time_mix_extra_dim);
+    add_kv(LLM_KV_TIME_DECAY_EXTRA_DIM,              hparams.time_decay_extra_dim);
+    add_kv(LLM_KV_RESIDUAL_SCALE,                    hparams.f_residual_scale);
+    add_kv(LLM_KV_EMBEDDING_SCALE,                   hparams.f_embedding_scale);
+
+    add_kv(LLM_KV_ATTENTION_HEAD_COUNT,              hparams.n_head_arr, true);
+    add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV,           hparams.n_head_kv_arr, true);
+    add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS,          hparams.f_max_alibi_bias);
+    add_kv(LLM_KV_ATTENTION_CLAMP_KQV,               hparams.f_clamp_kqv);
+    add_kv(LLM_KV_ATTENTION_KEY_LENGTH,              hparams.n_embd_head_k);
+    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,            hparams.n_embd_head_v);
+    add_kv(LLM_KV_ATTENTION_LAYERNORM_EPS,           hparams.f_norm_eps);
+    add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
+    add_kv(LLM_KV_ATTENTION_CAUSAL,                  hparams.causal_attn);
+    add_kv(LLM_KV_ATTENTION_Q_LORA_RANK,             hparams.n_lora_q);
+    add_kv(LLM_KV_ATTENTION_KV_LORA_RANK,            hparams.n_lora_kv);
+    add_kv(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,  hparams.n_rel_attn_bkts);
+    add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW,          hparams.n_swa);
+    add_kv(LLM_KV_ATTENTION_SCALE,                   hparams.f_attention_scale);
+
+    const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;
+
+    add_kv(LLM_KV_ROPE_DIMENSION_COUNT,              hparams.n_rot);
+    add_kv(LLM_KV_ROPE_FREQ_BASE,                    hparams.rope_freq_base_train);
+    // add_kv(LLM_KV_ROPE_SCALE_LINEAR,                 rope_scaling_factor); // old name
+    add_kv(LLM_KV_ROPE_SCALING_TYPE,                 llama_rope_scaling_type_name(hparams.rope_scaling_type_train));
+    add_kv(LLM_KV_ROPE_SCALING_FACTOR,               rope_scaling_factor);
+    add_kv(LLM_KV_ROPE_SCALING_ATTN_FACTOR,          hparams.rope_attn_factor);
+    add_kv(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,         hparams.n_ctx_orig_yarn);
+    add_kv(LLM_KV_ROPE_SCALING_FINETUNED,            hparams.rope_finetuned);
+    add_kv(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,         hparams.rope_yarn_log_mul);
+
+    // TODO: implement split file support
+    // add_kv(LLM_KV_SPLIT_NO,                          ???);
+    // add_kv(LLM_KV_SPLIT_COUNT,                       ???);
+    // add_kv(LLM_KV_SPLIT_TENSORS_COUNT,               ???);
+
+    add_kv(LLM_KV_SSM_INNER_SIZE,                    hparams.ssm_d_inner);
+    add_kv(LLM_KV_SSM_CONV_KERNEL,                   hparams.ssm_d_conv);
+    add_kv(LLM_KV_SSM_STATE_SIZE,                    hparams.ssm_d_state);
+    add_kv(LLM_KV_SSM_TIME_STEP_RANK,                hparams.ssm_dt_rank);
+    add_kv(LLM_KV_SSM_DT_B_C_RMS,                    hparams.ssm_dt_b_c_rms);
+
+    add_kv(LLM_KV_WKV_HEAD_SIZE,                     hparams.wkv_head_size);
+
+    add_kv(LLM_KV_TOKENIZER_MODEL,                   vocab.get_tokenizer_model());
+    add_kv(LLM_KV_TOKENIZER_PRE,                     vocab.get_tokenizer_pre());
+    add_kv(LLM_KV_TOKENIZER_LIST,                    tokens);
+    add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE,              token_types);
+    add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,        vocab.n_token_types());
+    add_kv(LLM_KV_TOKENIZER_SCORES,                  scores);
+    add_kv(LLM_KV_TOKENIZER_MERGES,                  vocab.get_bpe_merges());
+    // FIXME llama_token is type i32 but when reading in a GGUF file u32 is expected, not an issue for writing though
+    add_kv(LLM_KV_TOKENIZER_BOS_ID,                  uint32_t(vocab.token_bos()));
+    add_kv(LLM_KV_TOKENIZER_EOS_ID,                  uint32_t(vocab.token_eos()));
+    add_kv(LLM_KV_TOKENIZER_EOT_ID,                  uint32_t(vocab.token_eot()));
+    add_kv(LLM_KV_TOKENIZER_EOM_ID,                  uint32_t(vocab.token_eom()));
+    add_kv(LLM_KV_TOKENIZER_UNK_ID,                  uint32_t(vocab.token_unk()));
+    add_kv(LLM_KV_TOKENIZER_SEP_ID,                  uint32_t(vocab.token_sep()));
+    add_kv(LLM_KV_TOKENIZER_PAD_ID,                  uint32_t(vocab.token_pad()));
+    // add_kv(LLM_KV_TOKENIZER_CLS_ID,                  uint32_t(vocab.token_bos())); // deprecated
+    // add_kv(LLM_KV_TOKENIZER_MASK_ID,                 ???);
+    add_kv(LLM_KV_TOKENIZER_ADD_BOS,                 vocab.get_add_bos());
+    add_kv(LLM_KV_TOKENIZER_ADD_EOS,                 vocab.get_add_eos());
+    add_kv(LLM_KV_TOKENIZER_ADD_PREFIX,              vocab.get_add_space_prefix());
+    add_kv(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,         vocab.get_remove_extra_whitespaces());
+    add_kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,    vocab.get_precompiled_charsmap());
+    // add_kv(LLM_KV_TOKENIZER_HF_JSON,                 ???);
+    // add_kv(LLM_KV_TOKENIZER_RWKV,                    ???);
+    add_kv(LLM_KV_TOKENIZER_FIM_PRE_ID,              uint32_t(vocab.token_fim_pre()));
+    add_kv(LLM_KV_TOKENIZER_FIM_SUF_ID,              uint32_t(vocab.token_fim_suf()));
+    add_kv(LLM_KV_TOKENIZER_FIM_MID_ID,              uint32_t(vocab.token_fim_mid()));
+    add_kv(LLM_KV_TOKENIZER_FIM_PAD_ID,              uint32_t(vocab.token_fim_pad()));
+    add_kv(LLM_KV_TOKENIZER_FIM_REP_ID,              uint32_t(vocab.token_fim_rep()));
+    add_kv(LLM_KV_TOKENIZER_FIM_SEP_ID,              uint32_t(vocab.token_fim_sep()));
+
+    // TODO: implement LoRA support
+    // add_kv(LLM_KV_ADAPTER_TYPE,                      ???);
+    // add_kv(LLM_KV_ADAPTER_LORA_ALPHA,                ???);
+
+    // deprecated
+    // add_kv(LLM_KV_TOKENIZER_PREFIX_ID,               ???);
+    // add_kv(LLM_KV_TOKENIZER_SUFFIX_ID,               ???);
+    // add_kv(LLM_KV_TOKENIZER_MIDDLE_ID,               ???);
+}
+
+void llama_model_saver::add_tensors_from_model() {
+    if (std::string(model.output->name) != std::string(model.tok_embd->name)) {
+        add_tensor(model.tok_embd); // some models use the same tensor for tok_embd and output
+    }
+    add_tensor(model.type_embd);
+    add_tensor(model.pos_embd);
+    add_tensor(model.tok_norm);
+    add_tensor(model.tok_norm_b);
+    add_tensor(model.output_norm);
+    add_tensor(model.output_norm_b);
+    add_tensor(model.output);
+    add_tensor(model.output_b);
+    add_tensor(model.output_norm_enc);
+    add_tensor(model.cls);
+    add_tensor(model.cls_b);
+    add_tensor(model.cls_out);
+    add_tensor(model.cls_out_b);
+
+    for (const struct llama_layer & layer : model.layers) {
+        for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {
+            add_tensor(reinterpret_cast<const struct ggml_tensor * const *>(&layer)[i]);
+        }
+    }
+}
+
+void llama_model_saver::save(const std::string & path_model) {
+    gguf_write_to_file(gguf_ctx, path_model.c_str(), false);
+}
+

llama/llama.cpp/src/llama-model-saver.h

@@ -0,0 +1,37 @@
+#pragma once
+
+#include "llama.h"
+#include "llama-arch.h"
+
+#include <vector>
+
+struct llama_model_saver {
+    struct gguf_context * gguf_ctx = nullptr;
+    const struct llama_model & model;
+    const struct LLM_KV llm_kv;
+
+    llama_model_saver(const struct llama_model & model);
+    ~llama_model_saver();
+
+    void add_kv(enum llm_kv key, uint32_t     value);
+    void add_kv(enum llm_kv key, int32_t      value);
+    void add_kv(enum llm_kv key, float        value);
+    void add_kv(enum llm_kv key, bool         value);
+    void add_kv(enum llm_kv key, const char * value);
+
+    [[noreturn]]
+    void add_kv(enum llm_kv key, char value); // needed to make the template below compile
+
+    template <typename Container>
+    void add_kv(enum llm_kv key, const Container & value, bool per_layer = false);
+
+    void add_kv(enum llm_kv key, const std::vector<std::string> & value);
+
+    void add_tensor(const struct ggml_tensor * tensor);
+
+    void add_kv_from_model();
+
+    void add_tensors_from_model();
+
+    void save(const std::string & path_model);
+};

llama/llama.cpp/src/llama-model.cpp

@@ -40,6 +40,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_335M:          return "335M";
         case LLM_TYPE_410M:          return "410M";
         case LLM_TYPE_450M:          return "450M";
+        case LLM_TYPE_475M:          return "475M";
         case LLM_TYPE_770M:          return "770M";
         case LLM_TYPE_780M:          return "780M";
         case LLM_TYPE_0_5B:          return "0.5B";
@@ -79,6 +80,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_236B:          return "236B";
         case LLM_TYPE_290B:          return "290B";
         case LLM_TYPE_314B:          return "314B";
+        case LLM_TYPE_405B:          return "405B";
         case LLM_TYPE_671B:          return "671B";
         case LLM_TYPE_SMALL:         return "0.1B";
         case LLM_TYPE_MEDIUM:        return "0.4B";
@@ -115,6 +117,10 @@ static const std::map<llama_rope_scaling_type, const char *> LLAMA_ROPE_SCALING_
     { LLAMA_ROPE_SCALING_TYPE_LONGROPE,   "longrope"   },
 };
 
+std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type) {
+    return LLAMA_ROPE_SCALING_TYPES.at(rope_scaling_type);
+}
+
 static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) {
     for (const auto & kv : LLAMA_ROPE_SCALING_TYPES) {
         if (kv.second == name) {
@@ -297,6 +303,10 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
     // add extra buffer types, only if no GPU device is present
     // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094
     auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    if (cpu_dev == nullptr) {
+        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)
         ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
@@ -423,7 +433,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
     // get general kv
     ml.get_key(LLM_KV_GENERAL_NAME, name, false);
-    ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab, false);
 
     // everything past this point is not vocab-related
     if (hparams.vocab_only) {
@@ -435,7 +444,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_BLOCK_COUNT,       hparams.n_layer);
     ml.get_key(LLM_KV_EXPERT_COUNT,      hparams.n_expert,      false);
     ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false);
-    ml.get_key(LLM_KV_VOCAB_SIZE,        hparams.n_vocab,       false);
 
     if (arch == LLM_ARCH_WAVTOKENIZER_DEC) {
         ml.get_key(LLM_KV_FEATURES_LENGTH, hparams.n_embd_features);
@@ -459,11 +467,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
     std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
     std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);
-    std::fill(hparams.cross_attn_layers.begin(), hparams.cross_attn_layers.end(), -1);
 
     ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff_arr,   hparams.n_layer, false);
     ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer, false);
-    ml.get_arr(LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, hparams.cross_attn_layers, false);
 
     // n_head_kv is optional, default to n_head
     hparams.n_head_kv_arr = hparams.n_head_arr;
@@ -516,7 +522,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
         ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
 
-        if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_MLLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON) {
+        if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON) {
             if (hparams.n_rot != hparams.n_embd_head_k) {
                 throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
             }
@@ -579,22 +585,13 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.use_kq_norm = false;
                 }
             } break;
-        case LLM_ARCH_MLLAMA:
-            {
-                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-
-                switch (hparams.n_layer) {
-                    case 40: type = LLM_TYPE_11B; break;
-                    case 100: type = LLM_TYPE_90B; break;
-                    default: type = LLM_TYPE_UNKNOWN;
-                }
-            } break;
         case LLM_ARCH_DECI:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
                     case 32: type = LLM_TYPE_7B; break;
                     case 80: type = LLM_TYPE_70B; break;
+                    case 162: type = LLM_TYPE_405B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -721,7 +718,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_MOE_EVERY_N_LAYERS,         hparams.moe_every_n_layers, 0);
 
                 if (hparams.n_layer == 12 && hparams.n_embd == 768) {
-                    type = LLM_TYPE_137M;
+                    if (arch == LLM_ARCH_NOMIC_BERT) {
+                        type = LLM_TYPE_137M;
+                    } else if (arch == LLM_ARCH_NOMIC_BERT_MOE && hparams.moe_every_n_layers == 2) {
+                        type = LLM_TYPE_475M;
+                    }
                 }
             } break;
         case LLM_ARCH_BLOOM:
@@ -782,6 +783,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             // fall through
         case LLM_ARCH_QWEN2:
             {
+                ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type, false);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
                     case 24: type = hparams.n_embd == 1024 ? LLM_TYPE_0_5B : LLM_TYPE_1B; break;
@@ -1505,6 +1507,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     }
 
     ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    if (cpu_dev == nullptr) {
+        throw std::runtime_error(format("%s: no CPU backend found", __func__));
+    }
     const int i_gpu_start = std::max((int) hparams.n_layer - n_gpu_layers, (int) 0);
     const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, (int)n_layer + 1);
     auto get_layer_buft_list = [&](int il) -> llama_model::impl::layer_dev {
@@ -1576,7 +1581,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         const int64_t n_embd_head_v = hparams.n_embd_head_v;
         const int64_t n_ff          = hparams.n_ff();
         const int64_t n_embd_gqa    = n_embd_v_gqa;
-        const int64_t n_vocab       = hparams.n_vocab;
+        const int64_t n_vocab       = vocab.n_tokens();
         const int64_t n_token_types = vocab.n_token_types();
         const int64_t n_rot         = hparams.n_rot;
         const int64_t n_expert      = hparams.n_expert;
@@ -1672,8 +1677,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 for (const auto * overrides = ml.tensor_buft_overrides; overrides->pattern != nullptr; ++overrides) {
                     std::regex pattern(overrides->pattern);
                     if (std::regex_search(tensor_name, pattern)) {
-                        LLAMA_LOG_DEBUG("tensor %s buffer type overriden to %s\n", tensor_name.c_str(), ggml_backend_buft_name(overrides->buft));
                         buft = overrides->buft;
+                        LLAMA_LOG_DEBUG("tensor %s (%zu MiB %s) buffer type overridden to %s\n",
+                                tensor_name.c_str(),
+                                ggml_nbytes(t_meta) / 1024 / 1024, ggml_type_name(t_meta->type),
+                                ggml_backend_buft_name(buft));
                         break;
                     }
                 }
@@ -1690,6 +1698,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             auto * buft_dev = ggml_backend_buft_get_device(buft);
             if (ml.use_mmap && buft_dev && buft == ggml_backend_dev_host_buffer_type(buft_dev)) {
                 auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+                if (!cpu_dev) {
+                    throw std::runtime_error("no CPU backend found");
+                }
                 buft = ggml_backend_dev_buffer_type(cpu_dev);
             }
 
@@ -1829,52 +1840,6 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         }
                     }
                 } break;
-            case LLM_ARCH_MLLAMA:
-                {
-                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab+8}, 0);
-
-                    // output
-                    {
-                        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                        output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
-
-                        // if output is NULL, init from the input tok embed
-                        if (output == NULL) {
-                            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
-                        }
-                    }
-
-                    for (int i = 0; i < n_layer; ++i) {
-                        auto & layer = layers[i];
-
-                        if (hparams.cross_attention_layers(i)) {
-                            layer.cross_attn_k_norm = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_K_NORM,   "weight", i), {128}, 0);
-                            layer.cross_attn_k_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_K_PROJ,   "weight", i), {n_embd, 1024}, 0);
-                            layer.cross_attn_o_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_O_PROJ,   "weight", i), {n_embd, n_embd}, 0);
-                            layer.cross_attn_q_norm = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_Q_NORM, "weight", i), {128}, 0);
-                            layer.cross_attn_q_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_Q_PROJ, "weight", i), {n_embd, n_embd}, 0);
-                            layer.cross_attn_v_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_V_PROJ, "weight", i), {n_embd, 1024}, 0);
-                            layer.cross_attn_attn_gate = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_ATTN_GATE, i), {1}, 0);
-                            layer.cross_attn_mlp_gate = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_MLP_GATE, i), {1}, 0);
-                            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
-                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
-                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
-                            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
-                        } else {
-                            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
-                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
-                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
-                            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
-                            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
-                            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
-                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
-                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
-                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
-                        }
-                    }
-                } break;
             case LLM_ARCH_DECI:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -1917,7 +1882,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
                         layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd},     TENSOR_NOT_REQUIRED);
 
-                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        if (n_ff > 0) {
+                            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        }
 
                         if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
                             layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
@@ -1927,9 +1894,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
                         }
 
-                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        if (n_ff > 0) {
-                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
-                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        }
 
                         // optional MLP bias
                         layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
@@ -3573,7 +3542,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                     // output
                     output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -4206,6 +4179,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         if (!dev) {
             // FIXME: workaround for CPU backend buft having a NULL device
             dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+            if (!dev) {
+                throw std::runtime_error(format("%s: no CPU backend found", __func__));
+            }
         }
         ggml_backend_dev_props props;
         ggml_backend_dev_get_props(dev, &props);
@@ -4335,7 +4311,7 @@ uint64_t llama_model::n_elements() const {
 }
 
 void llama_model::print_info() const {
-    const char * rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train);
+    const std::string rope_scaling_type = llama_rope_scaling_type_name(hparams.rope_scaling_type_train);
 
     auto print_f = [](const std::function<uint32_t(uint32_t)> & f, uint32_t n) {
         bool is_var = false;
@@ -4396,7 +4372,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: causal attn      = %d\n",     __func__, hparams.causal_attn);
         LLAMA_LOG_INFO("%s: pooling type     = %d\n",     __func__, hparams.pooling_type);
         LLAMA_LOG_INFO("%s: rope type        = %d\n",     __func__, hparams.rope_type);
-        LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type);
+        LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type.c_str());
         LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
         LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
@@ -4543,6 +4519,19 @@ const ggml_tensor * llama_model::get_tensor(const char * name) const {
     return it->second;
 }
 
+ggml_tensor * llama_model::get_rope_factors(uint32_t n_ctx_per_seq, int il) const {
+    // choose long/short freq factors based on the context size
+    if (layers[il].rope_freqs != nullptr) {
+        return layers[il].rope_freqs;
+    }
+
+    if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
+        return layers[il].rope_long;
+    }
+
+    return layers[il].rope_short;
+}
+
 struct llm_build_llama : public llm_graph_context {
     llm_build_llama(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -4583,7 +4572,7 @@ struct llm_build_llama : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -4767,246 +4756,6 @@ struct llm_build_llama : public llm_graph_context {
     }
 };
 
-struct llm_build_mllama: public llm_graph_context {
-    llm_build_mllama(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        ggml_tensor * cur;
-        ggml_tensor * inpL;
-        ggml_tensor * inpCAS;
-
-        inpL = build_inp_embd(model.tok_embd);
-        inpCAS = build_inp_cross_attn_state();
-
-          // inp_pos - contains the positions
-        ggml_tensor * inp_pos = build_inp_pos();
-
-        auto * inp_attn = build_attn_inp_kv_unified();
-        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
-
-        for (int il = 0; il < n_layer; ++il) {
-            ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            if (hparams.cross_attention_layers(il)) {
-                if (!ubatch.embd && !cparams.cross_attn) {
-                    continue;
-                }
-
-                // cross attention layer
-                ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_q_proj, cur);
-                cb(Qcur, "Qcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                cb(Qcur, "Qcur", il);
-
-                Qcur = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 0, 2, 1, 3));
-                cb(Qcur, "Qcur", il);
-
-                Qcur = build_norm(Qcur, model.layers[il].cross_attn_q_norm, NULL, LLM_NORM_RMS, il);
-                cb(Qcur, "Qcur", il);
-
-                ggml_tensor * Kcur, * Vcur;
-                if (ubatch.embd) {
-                    Kcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_k_proj, inpCAS);
-                    cb(Kcur, "Kcur", il);
-
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, 6404);
-                    cb(Kcur, "Kcur", il);
-
-                    Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
-                    cb(Kcur, "Kcur", il);
-
-                    Kcur = build_norm(Kcur, model.layers[il].cross_attn_k_norm, NULL, LLM_NORM_RMS, il);
-                    cb(Kcur, "Kcur", il);
-
-                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, kv_self->k_l[il]));
-
-                    Vcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_v_proj, inpCAS);
-                    cb(Vcur, "Vcur", il);
-
-                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, 6404);
-                    cb(Vcur, "Vcur", il);
-
-                    Vcur = ggml_permute(ctx0, Vcur, 0, 2, 1, 3);
-                    cb(Vcur, "Vcur", il);
-
-                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, kv_self->v_l[il]));
-                } else {
-                    Kcur = ggml_view_tensor(ctx0, kv_self->k_l[il]);
-                    cb(Kcur, "Kcur (view)", il);
-
-                    Vcur = ggml_view_tensor(ctx0, kv_self->v_l[il]);
-                    cb(Vcur, "Vcur (view)", il);
-                }
-
-                struct ggml_tensor * kq = ggml_mul_mat(ctx0, Kcur, Qcur);
-                cb(kq, "kq", il);
-
-                // TODO: apply causal masks
-                struct ggml_tensor * kq_soft_max = ggml_soft_max_ext(ctx0, kq, nullptr, 1.f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias);
-                cb(kq_soft_max, "kq_soft_max", il);
-
-                Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, Vcur));
-                cb(Vcur, "Vcur", il);
-
-                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, Vcur, kq_soft_max);
-                cb(kqv, "kqv", il);
-
-                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-                cb(kqv_merged, "kqv_merged", il);
-
-                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
-                cb(cur, "kqv_merged_cont", il);
-
-                cur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_o_proj, cur);
-                cb(cur, "cur", il);
-
-                // TODO: do this in place once?
-                cur = ggml_mul(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_attn_gate));
-
-                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-                cb(ffn_inp, "ffn_inp", il);
-
-                // feed-forward network
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-
-                // TODO: do this inplace once?
-                cur = ggml_add_inplace(ctx0, ggml_mul_inplace(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_mlp_gate)), ffn_inp);
-                cb(cur, "ffn_out", il);
-
-                cur = build_cvec(cur, il);
-                cb(cur, "l_out", il);
-
-                // input for next layer
-                inpL = cur;
-            } else {
-                // self attention layer
-
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
-
-                // compute Q and K and RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn, gf,
-                    model.layers[il].wo, model.layers[il].bo,
-                    Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
-
-                if (il == n_layer - 1) {
-                    // skip computing output for unused tokens
-                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                    n_tokens = n_outputs;
-                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-                }
-
-                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-                cb(ffn_inp, "ffn_inp", il);
-
-                // feed-forward network
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-
-                cur = ggml_add(ctx0, cur, ffn_inp);
-                cb(cur, "ffn_out", il);
-
-                cur = build_cvec(cur, il);
-                cb(cur, "l_out", il);
-
-                // input for next layer
-                inpL = cur;
-            }
-        }
-
-        cur = inpL;
-
-        cur = build_norm(cur,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, -1);
-        cb(cur, "result_norm", -1);
-        res->t_embd = cur;
-
-        // lm_head
-        cur = build_lora_mm(model.output, cur);
-
-        cb(cur, "result_output", -1);
-        res->t_logits = cur;
-
-        ggml_build_forward_expand(gf, cur);
-    }
-};
-
 struct llm_build_deci : public llm_graph_context {
     llm_build_deci(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -5029,6 +4778,7 @@ struct llm_build_deci : public llm_graph_context {
             ggml_tensor * inpSA = inpL;
             const int64_t n_head_kv = hparams.n_head_kv(il);
             const int64_t n_head    = hparams.n_head(il);
+            const int64_t n_ff      = hparams.n_ff(il);
 
             if (n_head == 0) {
                 // attention-free layer of Llama-3_1-Nemotron-51B
@@ -5048,7 +4798,7 @@ struct llm_build_deci : public llm_graph_context {
             } else if (n_head > 0) {
                 // self-attention
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -5104,6 +4854,11 @@ struct llm_build_deci : public llm_graph_context {
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
 
+            // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B
+            if (n_ff == 0) {
+                continue;
+            }
+
             // For Granite architecture
             if (hparams.f_residual_scale) {
                 cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
@@ -7530,7 +7285,7 @@ struct llm_build_phi3 : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for 128k context
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 ggml_tensor* attn_norm_output = build_norm(inpL,
                         model.layers[il].attn_norm,
@@ -8282,7 +8037,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
-            ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+            ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
             // norm
             cur = build_norm(inpL,
@@ -9049,7 +8804,7 @@ struct llm_build_mamba : public llm_graph_context {
              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 llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
         const auto kv_head = kv_self->head;
 
@@ -9350,7 +9105,7 @@ struct llm_build_cohere2 : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for 128k context
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -10288,7 +10043,7 @@ struct llm_build_deepseek : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -11652,7 +11407,7 @@ struct llm_build_exaone : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -11797,7 +11552,7 @@ struct llm_build_rwkv6_base : public llm_graph_context {
             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 llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
         const auto n_tokens = ubatch.n_tokens;
         const auto n_seqs = ubatch.n_seqs;
@@ -12193,7 +11948,7 @@ struct llm_build_rwkv7_base : public llm_graph_context {
             ggml_tensor *& first_layer_value,
             const llama_ubatch & ubatch,
             int   il) const {
-        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+        const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
         const auto n_tokens = ubatch.n_tokens;
         const auto n_seqs = ubatch.n_seqs;
@@ -12741,7 +12496,7 @@ struct llm_build_solar : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -13192,7 +12947,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -13312,36 +13067,46 @@ struct llm_build_bailingmoe : public llm_graph_context {
     }
 };
 
-llama_memory_i * llama_model::create_memory() const {
+llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
     llama_memory_i * res;
 
     switch (arch) {
+        case LLM_ARCH_BERT:
+        case LLM_ARCH_JINA_BERT_V2:
+        case LLM_ARCH_NOMIC_BERT:
+        case LLM_ARCH_NOMIC_BERT_MOE:
+            {
+                res = nullptr;
+            } break;
         case LLM_ARCH_MAMBA:
         case LLM_ARCH_RWKV6:
         case LLM_ARCH_RWKV6QWEN2:
         case LLM_ARCH_RWKV7:
         case LLM_ARCH_ARWKV7:
             {
-                res = new llama_kv_cache_unified(hparams, {
+                res = new llama_kv_cache_recurrent(
-                    /*.get_rope_factors =*/ nullptr
+                        *this,
-                });
+                        GGML_TYPE_F32,
+                        GGML_TYPE_F32,
+                        cparams.offload_kqv,
+                        std::max((uint32_t) 1, cparams.n_seq_max));
             } break;
         default:
             {
-                res = new llama_kv_cache_unified(hparams, {
+                const auto padding = llama_kv_cache_unified::get_padding(cparams);
-                    /*.get_rope_factors =*/ [this](uint32_t n_ctx_per_seq, int il) {
-                        // choose long/short freq factors based on the context size
-                        if (layers[il].rope_freqs != nullptr) {
-                            return layers[il].rope_freqs;
-                        }
 
-                        if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
+                cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
-                            return layers[il].rope_long;
-                        }
 
-                        return layers[il].rope_short;
+                LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
-                    }
+
-                });
+                res = new llama_kv_cache_unified(
+                        *this,
+                        params.type_k,
+                        params.type_v,
+                        !cparams.flash_attn,
+                        cparams.offload_kqv,
+                        cparams.n_ctx,
+                        padding);
             }
     }
 
@@ -13363,10 +13128,6 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_llama>(*this, params, gf);
             } break;
-        case LLM_ARCH_MLLAMA:
-            {
-                llm = std::make_unique<llm_build_mllama>(*this, params, gf);
-            } break;
         case LLM_ARCH_DECI:
             {
                 llm = std::make_unique<llm_build_deci>(*this, params, gf);
@@ -13728,12 +13489,9 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         // use what we call a normal RoPE, operating on pairs of consecutive head values
         case LLM_ARCH_LLAMA:
         case LLM_ARCH_LLAMA4:
-        case LLM_ARCH_MLLAMA:
         case LLM_ARCH_DECI:
         case LLM_ARCH_BAICHUAN:
         case LLM_ARCH_STARCODER:
-        case LLM_ARCH_PLAMO:
-        case LLM_ARCH_ORION:
         case LLM_ARCH_INTERNLM2:
         case LLM_ARCH_MINICPM:
         case LLM_ARCH_XVERSE:
@@ -13772,6 +13530,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_PHI2:
         case LLM_ARCH_PHI3:
         case LLM_ARCH_PHIMOE:
+        case LLM_ARCH_PLAMO:
         case LLM_ARCH_GEMMA:
         case LLM_ARCH_GEMMA2:
         case LLM_ARCH_GEMMA3:
@@ -13779,6 +13538,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_OPENELM:
         case LLM_ARCH_GPTNEOX:
         case LLM_ARCH_CODESHELL:
+        case LLM_ARCH_ORION:
         case LLM_ARCH_NEMOTRON:
         case LLM_ARCH_EXAONE:
         case LLM_ARCH_MINICPM3:
@@ -13851,6 +13611,14 @@ const char * llama_model_chat_template(const llama_model * model, const char * n
         : LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE);
     const auto & it = model->gguf_kv.find(key);
     if (it == model->gguf_kv.end()) {
+        // one-off fix for very popular models (so we are not flooded with issues)
+        // do not extend this list unless absolutely necessary
+        // Mistral-Small-2503 does not have built-in chat template
+        llama_vocab_pre_type pre_type = model->vocab.get_pre_type();
+        if (pre_type == LLAMA_VOCAB_PRE_TYPE_TEKKEN && model->layers.size() == 40) {
+            return "mistral-v7-tekken";
+        }
+
         return nullptr;
     }
 

llama/llama.cpp/src/llama-model.h

@@ -11,7 +11,6 @@
 #include <string>
 #include <unordered_map>
 #include <vector>
-#include <stdexcept>
 
 struct llama_cparams;
 struct llama_ubatch;
@@ -37,6 +36,7 @@ enum llm_type {
     LLM_TYPE_335M,
     LLM_TYPE_410M,
     LLM_TYPE_450M,
+    LLM_TYPE_475M,
     LLM_TYPE_770M,
     LLM_TYPE_780M,
     LLM_TYPE_0_5B,
@@ -74,10 +74,10 @@ enum llm_type {
     LLM_TYPE_40B,
     LLM_TYPE_65B,
     LLM_TYPE_70B,
-    LLM_TYPE_90B,
     LLM_TYPE_236B,
     LLM_TYPE_290B,
     LLM_TYPE_314B,
+    LLM_TYPE_405B,
     LLM_TYPE_671B,
     LLM_TYPE_SMALL,
     LLM_TYPE_MEDIUM,
@@ -97,6 +97,8 @@ enum llm_type {
     LLM_TYPE_235B_A22B,
 };
 
+std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type);
+
 struct llama_layer_posnet {
     // resnet
     struct ggml_tensor * norm1   = nullptr;
@@ -316,16 +318,6 @@ struct llama_layer {
 
     struct ggml_tensor * bskcn_tv = nullptr;
 
-    // cross attention
-    struct ggml_tensor * cross_attn_k_norm = nullptr;
-    struct ggml_tensor * cross_attn_k_proj = nullptr;
-    struct ggml_tensor * cross_attn_o_proj = nullptr;
-    struct ggml_tensor * cross_attn_q_norm = nullptr;
-    struct ggml_tensor * cross_attn_q_proj = nullptr;
-    struct ggml_tensor * cross_attn_v_proj = nullptr;
-    struct ggml_tensor * cross_attn_attn_gate = nullptr;
-    struct ggml_tensor * cross_attn_mlp_gate = nullptr;
-
     struct llama_layer_posnet posnet;
 
     struct llama_layer_convnext convnext;
@@ -409,8 +401,11 @@ struct llama_model {
 
     const struct ggml_tensor * get_tensor(const char * name) const;
 
+    ggml_tensor * get_rope_factors(uint32_t n_ctx_per_seq, int il) const;
+
+    // note: can mutate `cparams`
     // TODO: move this to new llm_arch_model_i interface
-    llama_memory_i * create_memory() const; // TODO: params
+    llama_memory_i * create_memory(const llama_memory_params & params, llama_cparams & cparams) const;
 
     // TODO: move this to new llm_arch_model_i interface
     llm_graph_result_ptr build_graph(

llama/llama.cpp/src/llama-quant.cpp

@@ -519,7 +519,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         nthread = std::thread::hardware_concurrency();
     }
 
-    // mmap consistently increases speed Linux, and also increases speed on Windows with
+    // mmap consistently increases speed on Linux, and also increases speed on Windows with
     // hot cache. It may cause a slowdown on macOS, possibly related to free memory.
 #if defined(__linux__) || defined(_WIN32)
     constexpr bool use_mmap = true;
@@ -529,7 +529,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
     llama_model_kv_override * kv_overrides = nullptr;
     if (params->kv_overrides) {
-        auto v = (std::vector<llama_model_kv_override>*)params->kv_overrides;
+        auto * v = (std::vector<llama_model_kv_override>*)params->kv_overrides;
         kv_overrides = v->data();
     }
 
@@ -639,9 +639,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         if (llama_model_has_encoder(&model)) {
             n_attn_layer *= 3;
         }
-        if (qs.n_attention_wv != n_attn_layer) {
+        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
-            LLAMA_LOG_WARN("%s: n_attention_wv is unexpected, expected: %d, found: %d\n", __func__, n_attn_layer, qs.n_attention_wv);
-        }
     }
 
     size_t total_size_org = 0;

llama/llama.cpp/src/llama-sampling.cpp

@@ -1750,23 +1750,35 @@ static const char * llama_sampler_top_n_sigma_name(const struct llama_sampler *
 static void llama_sampler_top_n_sigma_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     const auto * ctx = (llama_sampler_top_n_sigma *) smpl->ctx;
 
+    if (ctx->n <= 0.0f || cur_p->size <= 1) {
+        return;
+    }
+
     // find max logit and calculate mean
     float max = cur_p->data[0].logit;
     float logits_sum = 0;
+    size_t valid_count = 0;
     for (size_t i = 0; i < cur_p->size; ++i) {
-        if (cur_p->data[i].logit > max) {
+        // Only count non-negative infinity values
-            max = cur_p->data[i].logit;
+        if (cur_p->data[i].logit != -INFINITY) {
+            if (cur_p->data[i].logit > max) {
+                max = cur_p->data[i].logit;
+            }
+            logits_sum += cur_p->data[i].logit;
+            valid_count++;
         }
-        logits_sum += cur_p->data[i].logit;
     }
-    float mean = logits_sum/cur_p->size;
+    float mean = valid_count > 0 ? logits_sum/valid_count : 0;
 
     // calculate standard deviation
     float acc = 0;
     for (size_t i = 0; i < cur_p->size; ++i) {
-        acc += pow(cur_p->data[i].logit - mean, 2);
+        // Skip -infinity in std calculation
+        if (cur_p->data[i].logit != -INFINITY) {
+            acc += pow(cur_p->data[i].logit - mean, 2);
+        }
     }
-    float std = sqrt(acc/cur_p->size);
+    float std = valid_count > 0 ? sqrt(acc/valid_count) : 0;
 
     //apply mask
     for (size_t i = 0; i < cur_p->size; ++i) {

llama/llama.cpp/src/llama-vocab.cpp

@@ -1,5 +1,7 @@
 #include "llama-vocab.h"
 
+#include "ggml.h"
+#include "gguf.h"
 #include "llama-impl.h"
 #include "llama-model-loader.h"
 
@@ -415,6 +417,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "'(?:[sSdDmMtT]|[lL][lL]|[vV][eE]|[rR][eE])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_SEED_CODER:
+                regex_exprs = {
+                    // original regex from tokenizer.json
+                    // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1}| ?[^\\s\\p{L}\\p{N}\r\n]+|\\s*[\r\n]+|\\s+(?!\\S)|\\s+"
+                    "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1}| ?[^\\s\\p{L}\\p{N}\\r\\n]+|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -1227,6 +1236,9 @@ struct fragment_buffer_variant {
 struct llama_vocab::impl {
     uint32_t n_token_types = 0; // for BERT-style token types
 
+    std::string tokenizer_model;
+    std::string tokenizer_pre;
+
     enum llama_vocab_type     type     = LLAMA_VOCAB_TYPE_SPM;
     enum llama_vocab_pre_type pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
 
@@ -1362,9 +1374,6 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
     // determine vocab type
     {
-        std::string tokenizer_model;
-        std::string tokenizer_pre;
-
         ml.get_key(LLM_KV_TOKENIZER_MODEL, tokenizer_model);
         ml.get_key(LLM_KV_TOKENIZER_PRE,   tokenizer_pre, false);
 
@@ -1459,7 +1468,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
             const int precompiled_charsmap_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP).c_str());
             if (precompiled_charsmap_keyidx != -1) {
-                size_t n_precompiled_charsmap = gguf_get_arr_data_n(ctx, precompiled_charsmap_keyidx);
+                const gguf_type pc_type = gguf_get_arr_type(ctx, precompiled_charsmap_keyidx);
+                const size_t n_precompiled_charsmap = gguf_get_arr_data_n(ctx, precompiled_charsmap_keyidx);
                 const char * pc = (const char *) gguf_get_arr_data(ctx, precompiled_charsmap_keyidx);
                 precompiled_charsmap.assign(pc, pc + n_precompiled_charsmap);
 #ifdef IS_BIG_ENDIAN
@@ -1625,6 +1635,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "bailingmoe") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "seed-coder") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_SEED_CODER;
+                clean_spaces = false;
             } else {
                 LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
                 pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
@@ -2770,6 +2784,14 @@ void llama_vocab::load(llama_model_loader & ml, const LLM_KV & kv) {
     pimpl->load(ml, kv);
 }
 
+std::string llama_vocab::get_tokenizer_model() const {
+    return pimpl->tokenizer_model;
+}
+
+std::string llama_vocab::get_tokenizer_pre() const {
+    return pimpl->tokenizer_pre;
+}
+
 enum llama_vocab_type llama_vocab::get_type() const {
     return pimpl->type;
 }
@@ -2992,6 +3014,20 @@ int llama_vocab::find_bpe_rank(const std::string & token_left, const std::string
     return it->second;
 }
 
+std::vector<std::string> llama_vocab::get_bpe_merges() const {
+    std::vector<std::string> result(pimpl->bpe_ranks.size());
+
+    for (const auto & pair : pimpl->bpe_ranks) {
+        result[pair.second] = pair.first.first + " " + pair.first.second;
+    }
+
+    return result;
+}
+
+std::vector<char> llama_vocab::get_precompiled_charsmap() const {
+    return pimpl->precompiled_charsmap;
+}
+
 int32_t llama_vocab::tokenize(
                   const char * text,
                      int32_t   text_len,

llama/llama.cpp/src/llama-vocab.h

@@ -21,6 +21,9 @@ struct llama_vocab {
 
     void load(llama_model_loader & ml, const LLM_KV & kv);
 
+    std::string get_tokenizer_model() const;
+    std::string get_tokenizer_pre() const;
+
     enum llama_vocab_type     get_type()     const;
     enum llama_vocab_pre_type get_pre_type() const;
 
@@ -80,6 +83,9 @@ struct llama_vocab {
     int max_token_len() const;
 
     int find_bpe_rank(const std::string & token_left, const std::string & token_right) const;
+    std::vector<std::string> get_bpe_merges() const;
+
+    std::vector<char> get_precompiled_charsmap() const;
 
     int32_t tokenize(
                    const char * text,

llama/llama.cpp/src/llama.cpp

@@ -4,6 +4,7 @@
 #include "llama-mmap.h"
 #include "llama-vocab.h"
 #include "llama-model-loader.h"
+#include "llama-model-saver.h"
 #include "llama-model.h"
 
 #include "ggml.h"
@@ -253,6 +254,13 @@ struct llama_model * llama_model_load_from_splits(
     return llama_model_load_from_file_impl(splits.front(), splits, params);
 }
 
+void llama_model_save_to_file(const struct llama_model * model, const char * path_model) {
+    llama_model_saver ms(*model);
+    ms.add_kv_from_model();
+    ms.add_tensors_from_model();
+    ms.save(path_model);
+}
+
 //
 // chat templates
 //
@@ -338,3 +346,4 @@ const char * llama_print_system_info(void) {
 
     return s.c_str();
 }
+

llama/llama.cpp/tools/mtmd/clip-impl.h

@@ -31,9 +31,7 @@
 #define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
 #define KEY_PROJ_SCALE_FACTOR   "clip.vision.projector.scale_factor"
 #define KEY_PROJ_TYPE           "clip.projector_type"
-
+#define KEY_SPATIAL_MERGE_SIZE  "clip.vision.spatial_merge_size"
-#define KEY_USE_GLU_MLP         "clip.use_glu_mlp"  // for qwen2.5vl
-#define KEY_USE_RMS_NORM        "clip.use_rms_norm" // for qwen2.5vl
 
 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
@@ -55,12 +53,16 @@
 #define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
 #define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
 #define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
+#define TN_ATTN_K_NORM     "%s.blk.%d.attn_k_norm.%s"
+#define TN_ATTN_Q_NORM     "%s.blk.%d.attn_q_norm.%s"
 #define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
 #define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
 #define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
 #define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
-#define TN_LN_1            "%s.blk.%d.ln1.%s"
+#define TN_LN_1            "%s.blk.%d.ln1.%s" // layer norm
-#define TN_LN_2            "%s.blk.%d.ln2.%s"
+#define TN_LN_2            "%s.blk.%d.ln2.%s" // layer norm
+#define TN_LS_1            "%s.blk.%d.ls1.%s" // layer scale
+#define TN_LS_2            "%s.blk.%d.ls2.%s" // layer scale
 #define TN_LN_PRE          "%s.pre_ln.%s"
 #define TN_LN_POST         "%s.post_ln.%s"
 #define TN_LLAVA_PROJ      "mm.%d.%s"
@@ -68,10 +70,14 @@
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
+#define TN_MM_INP_NORM     "mm.input_norm.weight"
 #define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
 #define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
 #define TN_MM_PROJECTOR    "mm.model.fc.weight"         // idefics3
+#define TN_MM_PATCH_MERGER "mm.patch_merger.weight"     // mistral small 3.1
 #define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
+#define TN_TOK_GLM_BOI     "adapter.boi"                // glm-edge (these embeddings are not in text model)
+#define TN_TOK_GLM_EOI     "adapter.eoi"                // glm-edge (these embeddings are not in text model)
 
 // mimicpmv
 #define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
@@ -88,6 +94,9 @@
 #define TN_GLM_ADAPTER_GATE     "adapter.linear.gate.%s"
 #define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
 
+// align x to upper multiple of n
+#define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
+
 enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
@@ -100,6 +109,7 @@ enum projector_type {
     PROJECTOR_TYPE_IDEFICS3,
     PROJECTOR_TYPE_PIXTRAL,
     PROJECTOR_TYPE_QWEN25VL,
+    PROJECTOR_TYPE_INTERNVL,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -114,6 +124,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
     { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
     { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
+    { PROJECTOR_TYPE_INTERNVL,  "internvl"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {
@@ -228,6 +239,15 @@ struct clip_image_u8_batch {
 
 struct clip_image_f32_batch {
     std::vector<clip_image_f32_ptr> entries;
+
+    clip_image_f32_batch clone() const {
+        clip_image_f32_batch new_batch;
+        new_batch.entries.reserve(entries.size());
+        for (const auto & entry : entries) {
+            new_batch.entries.emplace_back(new clip_image_f32(*entry));
+        }
+        return new_batch;
+    }
 };
 
 //

llama/llama.cpp/tools/mtmd/clip.h

@@ -78,10 +78,10 @@ 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_size      * clip_image_size_init(void);
-CLIP_API struct clip_image_u8        * clip_image_u8_init ();
+CLIP_API struct clip_image_u8        * clip_image_u8_init (void);
-CLIP_API struct clip_image_f32       * clip_image_f32_init();
+CLIP_API struct clip_image_f32       * clip_image_f32_init(void);
-CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(); // only used by libllava
+CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(void); // 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);

llama/llama.cpp/tools/mtmd/llava.cpp

@@ -2,6 +2,7 @@
 #include "llava.h"
 
 #include "llama.h"
+#include "ggml-cpp.h"
 
 #include <algorithm>
 #include <cerrno>
@@ -209,7 +210,11 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     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);
+
+    ggml_backend_ptr backend { ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr) };
+    GGML_ASSERT(backend != nullptr && "failed to initialize CPU backend");
+    ggml_backend_graph_compute(backend.get(), gf);
+
     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
@@ -457,7 +462,7 @@ struct llava_embd_batch {
     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) {
+    llava_embd_batch(float * 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);
@@ -469,7 +474,6 @@ struct llava_embd_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(),
@@ -493,7 +497,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
             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);
+        llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, *n_past, 0);
         if (llama_decode(ctx_llama, llava_batch.batch)) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return false;

llama/llama.cpp/tools/mtmd/mtmd.go

@@ -1,4 +1,4 @@
-package llava
+package mtmd
 
 // #cgo CXXFLAGS: -std=c++11
 // #cgo CPPFLAGS: -I${SRCDIR}/../../include -I${SRCDIR}/../../common