feat: qwen3 dense

2025-05-07 14:03:27 -07:00
212 changed files with 9802 additions and 11094 deletions
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@ -103,11 +103,6 @@ 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'
@ -324,7 +319,6 @@ 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 ;;
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@ -46,7 +46,7 @@ jobs:
        include:
          - preset: CPU
          - preset: CUDA
-            container: nvidia/cuda:11.8.0-devel-ubuntu22.04
+            container: nvidia/cuda:12.8.1-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/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe
+            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.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_11.3", "nvcc_11.3", "cublas_11.3", "cublas_dev_11.3")) -NoNewWindow -Wait
+            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_12.8", "nvcc_12.8", "cublas_12.8", "cublas_dev_12.8")) -NoNewWindow -Wait
          }

          $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path
--- a/.golangci.yaml
+++ b/.golangci.yaml
@ -19,8 +19,8 @@ linters:
    - nolintlint
    - nosprintfhostport
    - staticcheck
+    - tenv
    - unconvert
-    - usetesting
    - wastedassign
    - whitespace
  disable:
--- a/CMakePresets.json
+++ b/CMakePresets.json
@ -17,14 +17,6 @@
      "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"
-      }
-    },
    {
      "name": "CUDA 12",
      "inherits": [ "CUDA" ],
@ -78,11 +70,6 @@
      "configurePreset": "CUDA",
      "targets": [ "ggml-cuda" ]
    },
-    {
-      "name": "CUDA 11",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "CUDA 11"
-    },
    {
      "name": "CUDA 12",
      "inherits": [ "CUDA" ],
--- a/17
+++ b/17
@ -7,14 +7,10 @@ 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:${ROCMVERSION}-complete AS base-amd64
 RUN yum install -y yum-utils \
-    && yum-config-manager --add-repo https://dl.rockylinux.org/vault/rocky/8.5/AppStream/\$basearch/os/ \
-    && rpm --import https://dl.rockylinux.org/pub/rocky/RPM-GPG-KEY-Rocky-8 \
-    && dnf install -y yum-utils ccache gcc-toolset-10-gcc-10.2.1-8.2.el8 gcc-toolset-10-gcc-c++-10.2.1-8.2.el8 gcc-toolset-10-binutils-2.35-11.el8 \
+    && dnf install -y ccache \
    && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
-ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH

 FROM --platform=linux/arm64 almalinux:8 AS base-arm64
 # install epel-release for ccache
@ -38,15 +34,6 @@ 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//./-}
@ -98,11 +85,9 @@ 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
--- a/Makefile.sync
+++ b/Makefile.sync
@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggerganov/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=de4c07f93783a1a96456a44dc16b9db538ee1618
+FETCH_HEAD=e1e8e0991ffd9e99a445c6812bb519d5bac9f4b5

 .PHONY: help
 help:
@ -15,13 +15,11 @@ help:
 	@echo "    make -f $(lastword $(MAKEFILE_LIST)) clean sync"

 .PHONY: sync
-sync: llama/build-info.cpp ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal
+sync: llama/build-info.cpp llama/llama.cpp ml/backend/ggml/ggml

-llama/build-info.cpp: llama/build-info.cpp.in llama/llama.cpp
-	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' <$< >$@
-
-ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal: ml/backend/ggml/ggml
-	go generate ./$(@D)
+.PHONY: llama/build-info.cpp
+llama/build-info.cpp: llama/build-info.cpp.in
+	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' $< > $@

 .PHONY: llama/llama.cpp
 llama/llama.cpp: llama/vendor/
@ -32,13 +30,12 @@ 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: $(PATCHED)
+apply-patches: $(addsuffix ed, $(PATCHES))

-llama/patches/.%.patched: llama/patches/%.patch
+%.patched: %.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
@ -60,4 +57,4 @@ format-patches: llama/patches

 .PHONE: clean
 clean: checkout
-	$(RM) llama/patches/.*.patched
+	$(RM) $(addsuffix ed, $(PATCHES))
--- a/README.md
+++ b/README.md
@ -315,7 +315,6 @@ 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)
@ -527,7 +526,6 @@ 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

@ -584,7 +582,6 @@ 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)
- [UnityCodeLama](https://github.com/HardCodeDev777/UnityCodeLama) (Unity Edtior tool to analyze scripts via Ollama)

 ### Supported backends

--- a/api/client_test.go
+++ b/api/client_test.go
@ -1,6 +1,7 @@
 package api

 import (
+	"context"
 	"encoding/json"
 	"fmt"
 	"net/http"
@ -136,7 +137,7 @@ func TestClientStream(t *testing.T) {
 			client := NewClient(&url.URL{Scheme: "http", Host: ts.Listener.Addr().String()}, http.DefaultClient)

 			var receivedChunks []ChatResponse
-			err := client.stream(t.Context(), http.MethodPost, "/v1/chat", nil, func(chunk []byte) error {
+			err := client.stream(context.Background(), http.MethodPost, "/v1/chat", nil, func(chunk []byte) error {
 				var resp ChatResponse
 				if err := json.Unmarshal(chunk, &resp); err != nil {
 					return fmt.Errorf("failed to unmarshal chunk: %w", err)
@ -222,7 +223,7 @@ func TestClientDo(t *testing.T) {
 				ID      string `json:"id"`
 				Success bool   `json:"success"`
 			}
-			err := client.do(t.Context(), http.MethodPost, "/v1/messages", nil, &resp)
+			err := client.do(context.Background(), http.MethodPost, "/v1/messages", nil, &resp)

 			if tc.wantErr != "" {
 				if err == nil {
--- a/api/types.go
+++ b/api/types.go
@ -271,6 +271,9 @@ type Options struct {
 	RepeatPenalty    float32  `json:"repeat_penalty,omitempty"`
 	PresencePenalty  float32  `json:"presence_penalty,omitempty"`
 	FrequencyPenalty float32  `json:"frequency_penalty,omitempty"`
+	Mirostat         int      `json:"mirostat,omitempty"`
+	MirostatTau      float32  `json:"mirostat_tau,omitempty"`
+	MirostatEta      float32  `json:"mirostat_eta,omitempty"`
 	Stop             []string `json:"stop,omitempty"`
 }

@ -645,6 +648,9 @@ func DefaultOptions() Options {
 		RepeatPenalty:    1.1,
 		PresencePenalty:  0.0,
 		FrequencyPenalty: 0.0,
+		Mirostat:         0,
+		MirostatTau:      5.0,
+		MirostatEta:      0.1,
 		Seed:             -1,

 		Runner: Runner{
--- a/app/lifecycle/logging.go
+++ b/app/lifecycle/logging.go
@ -4,14 +4,20 @@ 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
@ -27,8 +33,20 @@ 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")
 }

--- a/benchmark/server_benchmark_test.go
+++ b/benchmark/server_benchmark_test.go
@ -78,7 +78,7 @@ func BenchmarkColdStart(b *testing.B) {

 	for _, tt := range tests {
 		b.Run(fmt.Sprintf("%s/cold/%s", m, tt.name), func(b *testing.B) {
-			ctx := b.Context()
+			ctx := context.Background()

 			// Set number of tokens as our throughput metric
 			b.SetBytes(int64(tt.maxTokens))
@ -113,7 +113,7 @@ func BenchmarkWarmStart(b *testing.B) {

 	for _, tt := range tests {
 		b.Run(fmt.Sprintf("%s/warm/%s", m, tt.name), func(b *testing.B) {
-			ctx := b.Context()
+			ctx := context.Background()

 			// Pre-warm the model
 			warmup(client, m, tt.prompt, b)
@ -140,7 +140,7 @@ func setup(b *testing.B) *api.Client {
 	if err != nil {
 		b.Fatal(err)
 	}
-	if _, err := client.Show(b.Context(), &api.ShowRequest{Model: modelName(b)}); err != nil {
+	if _, err := client.Show(context.Background(), &api.ShowRequest{Model: modelName(b)}); err != nil {
 		b.Fatalf("Model unavailable: %v", err)
 	}

--- a/cmd/cmd_test.go
+++ b/cmd/cmd_test.go
@ -2,6 +2,7 @@ package cmd

 import (
 	"bytes"
+	"context"
 	"encoding/json"
 	"io"
 	"net/http"
@ -336,7 +337,7 @@ func TestDeleteHandler(t *testing.T) {
 	t.Cleanup(mockServer.Close)

 	cmd := &cobra.Command{}
-	cmd.SetContext(t.Context())
+	cmd.SetContext(context.TODO())
 	if err := DeleteHandler(cmd, []string{"test-model"}); err != nil {
 		t.Fatalf("DeleteHandler failed: %v", err)
 	}
@ -398,6 +399,11 @@ func TestGetModelfileName(t *testing.T) {
 			var expectedFilename string

 			if tt.fileExists {
+				tempDir, err := os.MkdirTemp("", "modelfiledir")
+				defer os.RemoveAll(tempDir)
+				if err != nil {
+					t.Fatalf("temp modelfile dir creation failed: %v", err)
+				}
 				var fn string
 				if tt.modelfileName != "" {
 					fn = tt.modelfileName
@ -405,7 +411,7 @@ func TestGetModelfileName(t *testing.T) {
 					fn = "Modelfile"
 				}

-				tempFile, err := os.CreateTemp(t.TempDir(), fn)
+				tempFile, err := os.CreateTemp(tempDir, fn)
 				if err != nil {
 					t.Fatalf("temp modelfile creation failed: %v", err)
 				}
@ -524,7 +530,7 @@ func TestPushHandler(t *testing.T) {

 			cmd := &cobra.Command{}
 			cmd.Flags().Bool("insecure", false, "")
-			cmd.SetContext(t.Context())
+			cmd.SetContext(context.TODO())

 			// Redirect stderr to capture progress output
 			oldStderr := os.Stderr
@ -629,7 +635,7 @@ func TestListHandler(t *testing.T) {
 			t.Setenv("OLLAMA_HOST", mockServer.URL)

 			cmd := &cobra.Command{}
-			cmd.SetContext(t.Context())
+			cmd.SetContext(context.TODO())

 			// Capture stdout
 			oldStdout := os.Stdout
@ -724,7 +730,7 @@ func TestCreateHandler(t *testing.T) {
 			}))
 			t.Setenv("OLLAMA_HOST", mockServer.URL)
 			t.Cleanup(mockServer.Close)
-			tempFile, err := os.CreateTemp(t.TempDir(), "modelfile")
+			tempFile, err := os.CreateTemp("", "modelfile")
 			if err != nil {
 				t.Fatal(err)
 			}
@ -744,7 +750,7 @@ func TestCreateHandler(t *testing.T) {
 			}

 			cmd.Flags().Bool("insecure", false, "")
-			cmd.SetContext(t.Context())
+			cmd.SetContext(context.TODO())

 			// Redirect stderr to capture progress output
 			oldStderr := os.Stderr
--- a/cmd/interactive.go
+++ b/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, .png, or .webp images.\n", filepath.FromSlash("/path/to/file"))
+			fmt.Fprintf(os.Stderr, "Use %s to include .jpg or .png images.\n", filepath.FromSlash("/path/to/file"))
 		}

 		fmt.Fprintln(os.Stderr, "")
@ -511,7 +511,7 @@ 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|webp)\b`
+	regexPattern := `(?:[a-zA-Z]:)?(?:\./|/|\\)[\S\\ ]+?\.(?i:jpg|jpeg|png)\b`
 	re := regexp.MustCompile(regexPattern)

 	return re.FindAllString(input, -1)
@ -531,8 +531,6 @@ 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)
 	}
@ -553,7 +551,7 @@ func getImageData(filePath string) ([]byte, error) {
 	}

 	contentType := http.DetectContentType(buf)
-	allowedTypes := []string{"image/jpeg", "image/jpg", "image/png", "image/webp"}
+	allowedTypes := []string{"image/jpeg", "image/jpg", "image/png"}
 	if !slices.Contains(allowedTypes, contentType) {
 		return nil, fmt.Errorf("invalid image type: %s", contentType)
 	}
--- a/cmd/interactive_test.go
+++ b/cmd/interactive_test.go
@ -1,8 +1,6 @@
 package cmd

 import (
-	"os"
-	"path/filepath"
 	"testing"

 	"github.com/stretchr/testify/assert"
@ -12,17 +10,14 @@ 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 /six.webp inbetween6 /valid\ path/dir/seven.WEBP`
+/unescaped space /three.jpeg inbetween3 /valid\ path/dir/four.png "./quoted with spaces/five.JPG`
 	res := extractFileNames(input)
-	assert.Len(t, res, 7)
+	assert.Len(t, res, 5)
 	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")
@ -33,12 +28,10 @@ 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
-c:/users/jdoe/eleven.webp inbetween11 c:/program files/someplace/twelve.WebP inbetween12
-d:\path with\spaces\thirteen.WEBP some ending
+ d:\program files\someplace\nine.png inbetween9 "E:\program files\someplace\ten.PNG some ending
 `
 	res = extractFileNames(input)
-	assert.Len(t, res, 13)
+	assert.Len(t, res, 10)
 	assert.NotContains(t, res, "inbetween2")
 	assert.Contains(t, res[0], "one.png")
 	assert.Contains(t, res[0], "c:")
@ -56,31 +49,4 @@ d:\path with\spaces\thirteen.WEBP some ending
 	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")
 }
--- a/convert/convert.go
+++ b/convert/convert.go
@ -1,7 +1,6 @@
 package convert

 import (
-	"cmp"
 	"encoding/json"
 	"errors"
 	"fmt"
@ -15,12 +14,13 @@ import (
 )

 type ModelParameters struct {
-	Architectures []string `json:"architectures"`
-	VocabSize     uint32   `json:"vocab_size"`
+	Architectures []string       `json:"architectures"`
+	VocabSize     uint32         `json:"vocab_size"`
+	TextModel     TextParameters `json:"text_config"`
+}

-	TextModel struct {
-		VocabSize uint32 `json:"vocab_size"`
-	} `json:"text_config"`
+type TextParameters struct {
+	VocabSize uint32 `json:"vocab_size"`
 }

 type AdapterParameters struct {
@ -173,8 +173,6 @@ 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":
@ -191,8 +189,6 @@ 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":
@ -216,22 +212,24 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		return err
 	}

-	vocabSize := int(cmp.Or(p.VocabSize, p.TextModel.VocabSize))
+	vocabSize := int(p.VocabSize)
+	if vocabSize == 0 {
+		tVocabSize := int(p.TextModel.VocabSize)
+		vocabSize = tVocabSize
+	}

 	switch {
 	case vocabSize == 0:
-		slog.Debug("vocabulary size was not explicitly set by the model", "default size", len(t.Vocabulary.Tokens))
+		slog.Warn("vocabulary size was not explicitly set by the model", "default size", len(t.Vocabulary.Tokens))
 	case vocabSize > len(t.Vocabulary.Tokens):
-		slog.Debug("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
+		slog.Warn("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):
-		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))
+		return fmt.Errorf("vocabulary is larger than expected '%d' instead of '%d'", len(t.Vocabulary.Tokens), vocabSize)
 	default:
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}
--- a/convert/convert_mllama.go
+++ b/convert/convert_mllama.go
@ -1,160 +0,0 @@
-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 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()))
-			out = append(out, &ggml.Tensor{
-				Name:     t.Name(),
-				Kind:     t.Kind(),
-				Shape:    t.Shape(),
-				WriterTo: t,
-			})
-		} else if strings.HasPrefix(t.Name(), "v.") || strings.HasPrefix(t.Name(), "mm.") {
-			out = append(out, &ggml.Tensor{
-				Name:     t.Name(),
-				Kind:     t.Kind(),
-				Shape:    t.Shape(),
-				WriterTo: t,
-			})
-		} else {
-			text = append(text, 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))
-
-		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))
-	}
-}
--- a/convert/convert_qwen2.go
+++ b/convert/convert_qwen2.go
@ -15,7 +15,6 @@ 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"`
 }
@ -40,8 +39,6 @@ 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")
 	}
--- a/convert/convert_qwen25vl.go
+++ b/convert/convert_qwen25vl.go
@ -1,102 +0,0 @@
-package convert
-
-import (
-	"cmp"
-	"slices"
-	"strings"
-
-	"github.com/ollama/ollama/fs/ggml"
-)
-
-type qwen25VLModel struct {
-	qwen2Model
-
-	VisionModel struct {
-		Depth               uint32  `json:"depth"`
-		HiddenSize          uint32  `json:"hidden_size"`
-		NumHeads            uint32  `json:"num_heads"`
-		InChannels          uint32  `json:"in_chans"`
-		PatchSize           uint32  `json:"patch_size"`
-		SpatialMergeSize    uint32  `json:"spatial_merge_size"`
-		SpatialPatchSize    uint32  `json:"spatial_patch_size"`
-		WindowSize          uint32  `json:"window_size"`
-		RMSNormEps          float32 `json:"layer_norm_epsilon"`
-		RopeTheta           float32 `json:"rope_theta"`
-		FullAttentionBlocks []int32 `json:"fullatt_block_indexes"`
-		TemporalPatchSize   uint32  `json:"temporal_patch_size"`
-	} `json:"vision_config"`
-}
-
-var _ ModelConverter = (*qwen25VLModel)(nil)
-
-func (q *qwen25VLModel) KV(t *Tokenizer) ggml.KV {
-	kv := q.ModelParameters.KV(t)
-	kv["general.architecture"] = "qwen25vl"
-
-	for k, v := range q.qwen2Model.KV(t) {
-		if strings.HasPrefix(k, "qwen2.") {
-			kv[strings.Replace(k, "qwen2.", "qwen25vl.", 1)] = v
-		}
-	}
-
-	if q.VisionModel.FullAttentionBlocks == nil {
-		kv["qwen25vl.vision.fullatt_block_indexes"] = []int32{7, 15, 23, 31}
-	}
-
-	kv["qwen25vl.vision.block_count"] = cmp.Or(q.VisionModel.Depth, 32)
-	kv["qwen25vl.vision.embedding_length"] = q.VisionModel.HiddenSize
-	kv["qwen25vl.vision.attention.head_count"] = cmp.Or(q.VisionModel.NumHeads, 16)
-	kv["qwen25vl.vision.num_channels"] = q.VisionModel.InChannels
-	kv["qwen25vl.vision.patch_size"] = cmp.Or(q.VisionModel.PatchSize, 14)
-	kv["qwen25vl.vision.spatial_merge_size"] = cmp.Or(q.VisionModel.SpatialMergeSize, 2)
-	kv["qwen25vl.vision.spatial_patch_size"] = q.VisionModel.SpatialPatchSize
-	kv["qwen25vl.vision.window_size"] = cmp.Or(q.VisionModel.WindowSize, 112)
-	kv["qwen25vl.vision.attention.layer_norm_epsilon"] = cmp.Or(q.VisionModel.RMSNormEps, 1e-6)
-	kv["qwen25vl.vision.rope.freq_base"] = cmp.Or(q.VisionModel.RopeTheta, 1e4)
-	kv["qwen25vl.vision.fullatt_block_indexes"] = q.VisionModel.FullAttentionBlocks
-	kv["qwen25vl.vision.temporal_patch_size"] = cmp.Or(q.VisionModel.TemporalPatchSize, 2)
-
-	return kv
-}
-
-func (q *qwen25VLModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	var out []*ggml.Tensor
-
-	for _, t := range ts {
-		if strings.Contains(t.Name(), "patch_embed.proj") {
-			for t := range splitDim(t, 2,
-				strings.NewReplacer("patch_embed.proj", "patch_embd_0"),
-				strings.NewReplacer("patch_embed.proj", "patch_embd_1"),
-			) {
-				t.Shape = slices.DeleteFunc(t.Shape, func(i uint64) bool { return i == 1 })
-				out = append(out, t)
-			}
-		} else if strings.Contains(t.Name(), "attn.qkv") {
-			out = append(out, slices.Collect(splitDim(t, 0,
-				strings.NewReplacer("attn.qkv", "attn_q"),
-				strings.NewReplacer("attn.qkv", "attn_k"),
-				strings.NewReplacer("attn.qkv", "attn_v"),
-			))...)
-		} else {
-			out = append(out, &ggml.Tensor{
-				Name:     t.Name(),
-				Kind:     t.Kind(),
-				Shape:    t.Shape(),
-				WriterTo: t,
-			})
-		}
-	}
-
-	return out
-}
-
-func (p *qwen25VLModel) Replacements() []string {
-	return append(
-		p.qwen2Model.Replacements(),
-		"visual", "v",
-		"blocks", "blk",
-		"attn.proj", "attn_out",
-		"norm1", "ln1",
-		"norm2", "ln2",
-	)
-}
--- a/convert/fs.go
+++ b/convert/fs.go
@ -0,0 +1,58 @@
+package convert
+
+import (
+	"archive/zip"
+	"errors"
+	"io"
+	"io/fs"
+	"os"
+	"path/filepath"
+)
+
+type ZipReader struct {
+	r *zip.Reader
+	p string
+
+	// limit is the maximum size of a file that can be read directly
+	// from the zip archive. Files larger than this size will be extracted
+	limit int64
+}
+
+func NewZipReader(r *zip.Reader, p string, limit int64) fs.FS {
+	return &ZipReader{r, p, limit}
+}
+
+func (z *ZipReader) Open(name string) (fs.File, error) {
+	r, err := z.r.Open(name)
+	if err != nil {
+		return nil, err
+	}
+	defer r.Close()
+
+	if fi, err := r.Stat(); err != nil {
+		return nil, err
+	} else if fi.Size() < z.limit {
+		return r, nil
+	}
+
+	if !filepath.IsLocal(name) {
+		return nil, zip.ErrInsecurePath
+	}
+
+	n := filepath.Join(z.p, name)
+	if _, err := os.Stat(n); errors.Is(err, os.ErrNotExist) {
+		w, err := os.Create(n)
+		if err != nil {
+			return nil, err
+		}
+		defer w.Close()
+
+		if _, err := io.Copy(w, r); err != nil {
+			return nil, err
+		}
+	} else if err != nil {
+		return nil, err
+	}
+
+	return os.Open(n)
+}
--- a/convert/reader.go
+++ b/convert/reader.go
@ -38,10 +38,7 @@ 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.tile_position_embd.weight" ||
-		t.name == "v.pre_tile_position_embd.weight" ||
-		t.name == "v.post_tile_position_embd.weight" {
+		t.name == "v.positional_embedding_vlm" {
 		// these tensors are always F32
 		return 0
 	}
--- a/convert/tensor.go
+++ b/convert/tensor.go
@ -1,56 +0,0 @@
-package convert
-
-import (
-	"iter"
-	"slices"
-	"strings"
-
-	"github.com/ollama/ollama/fs/ggml"
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-)
-
-// splitDim splits a tensor along a specified dimension into multiple tensors. The dimension
-// is split evenly based on the number of replacers provided.
-func splitDim(t Tensor, dim int, replacers ...*strings.Replacer) iter.Seq[*ggml.Tensor] {
-	return func(yield func(*ggml.Tensor) bool) {
-		for i, replacer := range replacers {
-			shape := slices.Clone(t.Shape())
-			shape[dim] = shape[dim] / uint64(len(replacers))
-
-			slice := slices.Repeat([]tensor.Slice{nil}, len(shape))
-			slice[dim] = tensor.S(i*int(shape[dim]), (i+1)*int(shape[dim]))
-
-			tt := t.Clone()
-			tt.SetRepacker(func(_ string, data []float32, shape []uint64) ([]float32, error) {
-				dims := make([]int, len(shape))
-				for i := range shape {
-					dims[i] = int(shape[i])
-				}
-
-				var t tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
-				t, err := t.Slice(slice...)
-				if err != nil {
-					return nil, err
-				}
-
-				t = tensor.Materialize(t)
-				// flatten tensor so it can be written as a vector
-				if err := t.Reshape(t.Shape().TotalSize()); err != nil {
-					return nil, err
-				}
-
-				return native.VectorF32(t.(*tensor.Dense))
-			})
-
-			if !yield(&ggml.Tensor{
-				Name:     replacer.Replace(t.Name()),
-				Kind:     t.Kind(),
-				Shape:    shape,
-				WriterTo: tt,
-			}) {
-				break
-			}
-		}
-	}
-}
--- a/discover/cuda_common.go
+++ b/discover/cuda_common.go
@ -3,6 +3,7 @@
 package discover

 import (
+	"fmt"
 	"log/slog"
 	"os"
 	"regexp"
@ -59,6 +60,8 @@ 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"
--- a/discover/gpu.go
+++ b/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.LogLevel() < slog.LevelInfo {
+	if envconfig.Debug() {
 		return C.uint16_t(1)
 	}
 	return C.uint16_t(0)
--- a/discover/path.go
+++ b/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_v11', 'cuda_v12', 'rocm', etc.
+// 'cuda_v12', 'rocm', etc.
 var LibOllamaPath string = func() string {
 	exe, err := os.Executable()
 	if err != nil {
--- a/docs/api.md
+++ b/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-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.
+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.

 ### Durations

@ -394,6 +394,9 @@ curl http://localhost:11434/api/generate -d '{
    "repeat_penalty": 1.2,
    "presence_penalty": 1.5,
    "frequency_penalty": 1.0,
+    "mirostat": 1,
+    "mirostat_tau": 0.8,
+    "mirostat_eta": 0.6,
    "penalize_newline": true,
    "stop": ["\n", "user:"],
    "numa": false,
@ -952,8 +955,19 @@ 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
@ -998,8 +1012,8 @@ Quantize a non-quantized model.

 ```shell
 curl http://localhost:11434/api/create -d '{
-  "model": "llama3.2:quantized",
-  "from": "llama3.2:3b-instruct-fp16",
+  "model": "llama3.1:quantized",
+  "from": "llama3.1:8b-instruct-fp16",
  "quantize": "q4_K_M"
 }'
 ```
@ -1009,14 +1023,12 @@ curl http://localhost:11434/api/create -d '{
 A stream of JSON objects is returned:

 ```json
-{"status":"quantizing F16 model to Q4_K_M","digest":"0","total":6433687776,"completed":12302}
-{"status":"quantizing F16 model to Q4_K_M","digest":"0","total":6433687776,"completed":6433687552}
-{"status":"verifying conversion"}
-{"status":"creating new layer sha256:fb7f4f211b89c6c4928ff4ddb73db9f9c0cfca3e000c3e40d6cf27ddc6ca72eb"}
-{"status":"using existing layer sha256:966de95ca8a62200913e3f8bfbf84c8494536f1b94b49166851e76644e966396"}
-{"status":"using existing layer sha256:fcc5a6bec9daf9b561a68827b67ab6088e1dba9d1fa2a50d7bbcc8384e0a265d"}
-{"status":"using existing layer sha256:a70ff7e570d97baaf4e62ac6e6ad9975e04caa6d900d3742d37698494479e0cd"}
+{"status":"quantizing F16 model to Q4_K_M"}
+{"status":"creating new layer sha256:667b0c1932bc6ffc593ed1d03f895bf2dc8dc6df21db3042284a6f4416b06a29"}
+{"status":"using existing layer sha256:11ce4ee3e170f6adebac9a991c22e22ab3f8530e154ee669954c4bc73061c258"}
+{"status":"using existing layer sha256:0ba8f0e314b4264dfd19df045cde9d4c394a52474bf92ed6a3de22a4ca31a177"}
 {"status":"using existing layer sha256:56bb8bd477a519ffa694fc449c2413c6f0e1d3b1c88fa7e3c9d88d3ae49d4dcb"}
+{"status":"creating new layer sha256:455f34728c9b5dd3376378bfb809ee166c145b0b4c1f1a6feca069055066ef9a"}
 {"status":"writing manifest"}
 {"status":"success"}
 ```
@ -1154,37 +1166,29 @@ A single JSON object will be returned.
 {
  "models": [
    {
-      "name": "deepseek-r1:latest",
-      "model": "deepseek-r1:latest",
-      "modified_at": "2025-05-10T08:06:48.639712648-07:00",
-      "size": 4683075271,
-      "digest": "0a8c266910232fd3291e71e5ba1e058cc5af9d411192cf88b6d30e92b6e73163",
+      "name": "codellama:13b",
+      "modified_at": "2023-11-04T14:56:49.277302595-07:00",
+      "size": 7365960935,
+      "digest": "9f438cb9cd581fc025612d27f7c1a6669ff83a8bb0ed86c94fcf4c5440555697",
      "details": {
-        "parent_model": "",
        "format": "gguf",
-        "family": "qwen2",
-        "families": [
-          "qwen2"
-        ],
-        "parameter_size": "7.6B",
-        "quantization_level": "Q4_K_M"
+        "family": "llama",
+        "families": null,
+        "parameter_size": "13B",
+        "quantization_level": "Q4_0"
      }
    },
    {
-      "name": "llama3.2:latest",
-      "model": "llama3.2:latest",
-      "modified_at": "2025-05-04T17:37:44.706015396-07:00",
-      "size": 2019393189,
-      "digest": "a80c4f17acd55265feec403c7aef86be0c25983ab279d83f3bcd3abbcb5b8b72",
+      "name": "llama3:latest",
+      "modified_at": "2023-12-07T09:32:18.757212583-08:00",
+      "size": 3825819519,
+      "digest": "fe938a131f40e6f6d40083c9f0f430a515233eb2edaa6d72eb85c50d64f2300e",
      "details": {
-        "parent_model": "",
        "format": "gguf",
        "family": "llama",
-        "families": [
-          "llama"
-        ],
-        "parameter_size": "3.2B",
-        "quantization_level": "Q4_K_M"
+        "families": null,
+        "parameter_size": "7B",
+        "quantization_level": "Q4_0"
      }
    }
  ]
--- a/docs/gpu.md
+++ b/docs/gpu.md
@ -1,6 +1,6 @@
 # GPU
 ## Nvidia
-Ollama supports Nvidia GPUs with compute capability 5.0+.
+Ollama supports Nvidia GPUs with compute capability 5.0+ and driver version 531 and newer.

 Check your compute compatibility to see if your card is supported:
 [https://developer.nvidia.com/cuda-gpus](https://developer.nvidia.com/cuda-gpus)
--- a/docs/modelfile.md
+++ b/docs/modelfile.md
@ -150,6 +150,9 @@ PARAMETER <parameter> <parametervalue>

 | Parameter      | Description                                                                                                                                                                                                                                             | Value Type | Example Usage        |
 | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------- |
+| mirostat       | Enable Mirostat sampling for controlling perplexity. (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)                                                                                                                                         | int        | mirostat 0           |
+| mirostat_eta   | Influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. (Default: 0.1)                        | float      | mirostat_eta 0.1     |
+| mirostat_tau   | Controls the balance between coherence and diversity of the output. A lower value will result in more focused and coherent text. (Default: 5.0)                                                                                                         | float      | mirostat_tau 5.0     |
 | num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                    | int        | num_ctx 4096         |
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                           | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                     | float      | repeat_penalty 1.1   |
--- a/docs/troubleshooting.md
+++ b/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_v11 rocm_v5]
+Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v12 rocm_v5]
 ```

 **Experimental LLM Library Override**
--- a/envconfig/config.go
+++ b/envconfig/config.go
@ -149,22 +149,9 @@ 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.
@ -222,6 +209,8 @@ 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 {
@ -249,7 +238,7 @@ type EnvVar struct {

 func AsMap() map[string]EnvVar {
 	ret := map[string]EnvVar{
-		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", LogLevel(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
+		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug(), "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)"},
--- a/envconfig/config_test.go
+++ b/envconfig/config_test.go
@ -1,13 +1,11 @@
 package envconfig

 import (
-	"log/slog"
 	"math"
 	"testing"
 	"time"

 	"github.com/google/go-cmp/cmp"
-	"github.com/ollama/ollama/logutil"
 )

 func TestHost(t *testing.T) {
@ -294,34 +292,3 @@ 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)
-			}
-		})
-	}
-}
--- a/fs/ggml/ggml.go
+++ b/fs/ggml/ggml.go
@ -6,7 +6,6 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
-	"math"
 	"slices"
 	"strings"

@ -126,8 +125,6 @@ func (kv KV) OllamaEngineRequired() bool {
 		"gemma3",
 		"mistral3",
 		"llama4",
-		"mllama",
-		"qwen25vl",
 	}, kv.Architecture())
 }

@ -651,29 +648,6 @@ 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))
-		mergeSize := uint64(llm.KV().Uint("vision.spatial_merge_size", 2))
-		temporalPatchSize := uint64(2)
-
-		// Calculate max possible patches based on max_pixels
-		maxHeight := uint64(math.Sqrt(float64(maxPixels)))
-		maxWidth := maxPixels / maxHeight
-		maxGridHeight := maxHeight / patchSize
-		maxGridWidth := maxWidth / patchSize
-		// Account for merged patches (2x2 grid)
-		numPatches := (maxGridHeight * maxGridWidth) / (mergeSize * mergeSize)
-
-		// Calculate graph size based on typical operations in ProcessImage and createPatches
-		graphSize = 4 * (maxPixels*numChannels + // Original image storage
-			// Normalized pixels
-			maxPixels*numChannels +
-			// Patches storage (numPatches * channels * temporalPatchSize * patchSize^2)
-			numPatches*numChannels*temporalPatchSize*patchSize*patchSize +
-			// Self-attention calculations (similar to other architectures)
-			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
--- a/fs/ggml/type.go
+++ b/fs/ggml/type.go
@ -12,42 +12,42 @@ type FileType uint32
 const (
 	FileTypeF32 FileType = iota
 	FileTypeF16
-	fileTypeQ4_0
-	fileTypeQ4_1
+	FileTypeQ4_0
+	FileTypeQ4_1
 	fileTypeQ4_1_F16 // unused by GGML
 	fileTypeQ4_2     // unused by GGML
 	fileTypeQ4_3     // unused by GGML
 	FileTypeQ8_0
-	fileTypeQ5_0
-	fileTypeQ5_1
-	fileTypeQ2_K
-	fileTypeQ3_K_S
-	fileTypeQ3_K_M
-	fileTypeQ3_K_L
+	FileTypeQ5_0
+	FileTypeQ5_1
+	FileTypeQ2_K
+	FileTypeQ3_K_S
+	FileTypeQ3_K_M
+	FileTypeQ3_K_L
 	FileTypeQ4_K_S
 	FileTypeQ4_K_M
-	fileTypeQ5_K_S
-	fileTypeQ5_K_M
-	fileTypeQ6_K
-	fileTypeIQ2_XXS
-	fileTypeIQ2_XS
-	fileTypeQ2_K_S
-	fileTypeIQ3_XS
-	fileTypeIQ3_XXS
-	fileTypeIQ1_S
-	fileTypeIQ4_NL
-	fileTypeIQ3_S
-	fileTypeIQ3_M
-	fileTypeIQ2_S
-	fileTypeIQ2_M
-	fileTypeIQ4_XS
-	fileTypeIQ1_M
+	FileTypeQ5_K_S
+	FileTypeQ5_K_M
+	FileTypeQ6_K
+	fileTypeIQ2_XXS // not supported by ollama
+	fileTypeIQ2_XS  // not supported by ollama
+	FileTypeQ2_K_S
+	fileTypeIQ3_XS  // not supported by ollama
+	fileTypeIQ3_XXS // not supported by ollama
+	fileTypeIQ1_S   // not supported by ollama
+	fileTypeIQ4_NL  // not supported by ollama
+	fileTypeIQ3_S   // not supported by ollama
+	fileTypeIQ3_M   // not supported by ollama
+	fileTypeIQ2_S   // not supported by ollama
+	fileTypeIQ2_M   // not supported by ollama
+	fileTypeIQ4_XS  // not supported by ollama
+	fileTypeIQ1_M   // not supported by ollama
 	FileTypeBF16
 	fileTypeQ4_0_4_4 // unused by GGML
 	fileTypeQ4_0_4_8 // unused by GGML
 	fileTypeQ4_0_8_8 // unused by GGML
-	fileTypeTQ1_0
-	fileTypeTQ2_0
+	fileTypeTQ1_0    // not supported by ollama
+	fileTypeTQ2_0    // not supported by ollama

 	FileTypeUnknown = 1024
 )
@ -60,12 +60,36 @@ 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:
@ -87,41 +111,40 @@ 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"
@ -140,35 +163,35 @@ 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
--- a/llama/build-info.cpp
+++ b/llama/build-info.cpp
@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "de4c07f93783a1a96456a44dc16b9db538ee1618";
+char const *LLAMA_COMMIT = "e1e8e0991ffd9e99a445c6812bb519d5bac9f4b5";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";
--- a/llama/llama.cpp/.rsync-filter
+++ b/llama/llama.cpp/.rsync-filter
@ -10,11 +10,11 @@ include common/stb_image.*
 include include/
 include include/llama.*
 include include/llama-*.*
-include tools/
-include tools/mtmd/
-include tools/mtmd/clip.*
-include tools/mtmd/clip-impl.*
-include tools/mtmd/llava.*
+include examples/
+include examples/llava/
+include examples/llava/clip.*
+include examples/llava/clip-impl.*
+include examples/llava/llava.*
 include src/
 include src/llama.*
 include src/llama-*.*
--- a/llama/llama.cpp/common/common.cpp
+++ b/llama/llama.cpp/common/common.cpp
@ -1096,6 +1096,7 @@ 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;
@ -1113,7 +1114,6 @@ 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;
@ -1565,20 +1565,3 @@ 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;
-}
--- a/llama/llama.cpp/common/common.h
+++ b/llama/llama.cpp/common/common.h
@ -66,6 +66,7 @@ enum llama_example {
    LLAMA_EXAMPLE_COMMON,
    LLAMA_EXAMPLE_SPECULATIVE,
    LLAMA_EXAMPLE_MAIN,
+    LLAMA_EXAMPLE_INFILL,
    LLAMA_EXAMPLE_EMBEDDING,
    LLAMA_EXAMPLE_PERPLEXITY,
    LLAMA_EXAMPLE_RETRIEVAL,
@ -95,7 +96,6 @@ 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
@ -161,7 +161,6 @@ 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,
@ -324,6 +323,7 @@ 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
@ -332,7 +332,6 @@ 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

@ -341,7 +340,7 @@ struct common_params {

    common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;

-    // multimodal models (see tools/mtmd)
+    // multimodal models (see examples/llava)
    struct common_params_model mmproj;
    bool mmproj_use_gpu = true;     // use GPU for multimodal model
    bool no_mmproj = false;         // explicitly disable multimodal model
@ -410,14 +409,13 @@ 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 = "tools/cvector-generator/positive.txt";
-    std::string cvector_negative_file = "tools/cvector-generator/negative.txt";
+    std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
+    std::string cvector_negative_file = "examples/cvector-generator/negative.txt";

    bool spm_infill = false; // suffix/prefix/middle pattern for infill

@ -666,9 +664,3 @@ 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);
--- a/llama/llama.cpp/common/sampling.cpp
+++ b/llama/llama.cpp/common/sampling.cpp
@ -1,7 +1,6 @@
 #include "sampling.h"

 #include "common.h"
-#include "log.h"

 #include <cmath>
 #include <unordered_map>
@ -230,48 +229,51 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                params.logit_bias.data()));

    if (params.mirostat == 0) {
-        for (const auto & cnstr : params.samplers) {
-            switch (cnstr) {
-                case COMMON_SAMPLER_TYPE_DRY:
-                    {
-                        std::vector<const char *> c_breakers;
-                        c_breakers.reserve(params.dry_sequence_breakers.size());
-                        for (const auto & str : params.dry_sequence_breakers) {
-                            c_breakers.push_back(str.c_str());
-                        }
+        if (params.top_n_sigma >= 0) {
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k        (params.top_k));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_temp         (params.temp));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma  (params.top_n_sigma));
+        } else {
+            for (const auto & cnstr : params.samplers) {
+                switch (cnstr) {
+                    case COMMON_SAMPLER_TYPE_DRY:
+                        {
+                            std::vector<const char *> c_breakers;
+                            c_breakers.reserve(params.dry_sequence_breakers.size());
+                            for (const auto & str : params.dry_sequence_breakers) {
+                                c_breakers.push_back(str.c_str());
+                            }

-                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
-                    }
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_K:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k       (params.top_k));
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p       (params.top_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_N_SIGMA:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma));
-                    break;
-                case COMMON_SAMPLER_TYPE_MIN_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p       (params.min_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_XTC:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc         (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
-                    break;
-                case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical     (params.typ_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext    (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                    break;
-                case COMMON_SAMPLER_TYPE_INFILL:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill      (vocab));
-                    break;
-                case COMMON_SAMPLER_TYPE_PENALTIES:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_penalties   (params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
-                    break;
-                default:
-                    GGML_ASSERT(false && "unknown sampler type");
+                            llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                        }
+                        break;
+                    case COMMON_SAMPLER_TYPE_TOP_K:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TOP_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_MIN_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_XTC:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                        break;
+                    case COMMON_SAMPLER_TYPE_INFILL:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
+                        break;
+                    case COMMON_SAMPLER_TYPE_PENALTIES:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                        break;
+                    default:
+                        GGML_ASSERT(false && "unknown sampler type");
+                }
            }
        }
        llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
@ -473,7 +475,6 @@ 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';
@ -489,7 +490,6 @@ 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";
@ -504,7 +504,6 @@ 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 },
@ -518,7 +517,6 @@ 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 },
@ -535,16 +533,14 @@ 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);
-            continue;
-        }
-        if (allow_alt_names) {
-            sampler = sampler_alt_name_map.find(name);
-            if (sampler != sampler_alt_name_map.end()) {
-                samplers.push_back(sampler->second);
-                continue;
+        } else {
+            if (allow_alt_names) {
+                sampler = sampler_alt_name_map.find(name);
+                if (sampler != sampler_alt_name_map.end()) {
+                    samplers.push_back(sampler->second);
+                }
            }
        }
-        LOG_WRN("%s: unable to match sampler by name '%s'\n", __func__, name.c_str());
    }

    return samplers;
@ -556,7 +552,6 @@ 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 },
@ -571,8 +566,6 @@ 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);
        }
    }

--- a/llama/llama.cpp/examples/llava/clip-impl.h
+++ b/llama/llama.cpp/examples/llava/clip-impl.h
@ -31,7 +31,9 @@
 #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"
@ -53,16 +55,12 @@
 #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" // layer norm
-#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_1            "%s.blk.%d.ln1.%s"
+#define TN_LN_2            "%s.blk.%d.ln2.%s"
 #define TN_LN_PRE          "%s.pre_ln.%s"
 #define TN_LN_POST         "%s.post_ln.%s"
 #define TN_LLAVA_PROJ      "mm.%d.%s"
@ -70,14 +68,10 @@
 #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"
@ -94,9 +88,6 @@
 #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,
@ -109,7 +100,6 @@ enum projector_type {
    PROJECTOR_TYPE_IDEFICS3,
    PROJECTOR_TYPE_PIXTRAL,
    PROJECTOR_TYPE_QWEN25VL,
-    PROJECTOR_TYPE_INTERNVL,
    PROJECTOR_TYPE_UNKNOWN,
 };

@ -124,7 +114,6 @@ 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) {
@ -239,15 +228,6 @@ 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;
-    }
 };

 //
--- a/llama/llama.cpp/examples/llava/clip.cpp
+++ b/llama/llama.cpp/examples/llava/clip.cpp
--- a/llama/llama.cpp/examples/llava/clip.h
+++ b/llama/llama.cpp/examples/llava/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(void);
-CLIP_API struct clip_image_u8        * clip_image_u8_init (void);
-CLIP_API struct clip_image_f32       * clip_image_f32_init(void);
-CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(void); // only used by libllava
+CLIP_API struct clip_image_size      * clip_image_size_init();
+CLIP_API struct clip_image_u8        * clip_image_u8_init ();
+CLIP_API struct clip_image_f32       * clip_image_f32_init();
+CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(); // only used by libllava

 // nx, ny are the output image dimensions
 CLIP_API unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny);
--- a/llama/llama.cpp/examples/llava/llava.cpp
+++ b/llama/llama.cpp/examples/llava/llava.cpp
@ -2,7 +2,6 @@
 #include "llava.h"

 #include "llama.h"
-#include "ggml-cpp.h"

 #include <algorithm>
 #include <cerrno>
@ -210,11 +209,7 @@ 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_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);
-
+    ggml_graph_compute_with_ctx(model.ctx, gf, 1);
    struct ggml_tensor* result = ggml_graph_node(gf, -1);

    memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
@ -462,7 +457,7 @@ 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_tokens, llama_pos pos_0, llama_seq_id seq_id) {
+    llava_embd_batch(float * embd, int32_t n_embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
        pos     .resize(n_tokens);
        n_seq_id.resize(n_tokens);
        seq_ids .resize(n_tokens + 1);
@ -474,6 +469,7 @@ 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(),
@ -497,7 +493,7 @@ 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_eval, *n_past, 0);
+        llava_embd_batch llava_batch = llava_embd_batch(embd, n_embd, n_eval, *n_past, 0);
        if (llama_decode(ctx_llama, llava_batch.batch)) {
            LOG_ERR("%s : failed to eval\n", __func__);
            return false;
--- a/llama/llama.cpp/examples/llava/llava.go
+++ b/llama/llama.cpp/examples/llava/llava.go
@ -1,4 +1,4 @@
-package mtmd
+package llava

 // #cgo CXXFLAGS: -std=c++11
 // #cgo CPPFLAGS: -I${SRCDIR}/../../include -I${SRCDIR}/../../common
--- a/llama/llama.cpp/examples/llava/llava.h
+++ b/llama/llama.cpp/examples/llava/llava.h
--- a/llama/llama.cpp/include/llama.h
+++ b/llama/llama.cpp/include/llama.h
@ -4,7 +4,6 @@
 #include "ggml.h"
 #include "ggml-cpu.h"
 #include "ggml-backend.h"
-#include "ggml-opt.h"

 #include <stddef.h>
 #include <stdint.h>
@ -113,7 +112,6 @@ 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 {
@ -258,6 +256,7 @@ extern "C" {

        llama_token  *  token;
        float        *  embd;
+        int32_t         n_embd;
        llama_pos    *  pos;
        int32_t      *  n_seq_id;
        llama_seq_id ** seq_id;
@ -353,18 +352,20 @@ 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
@ -446,10 +447,6 @@ 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");

@ -464,6 +461,10 @@ 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);

@ -929,19 +930,14 @@ extern "C" {
    // Frees a batch of tokens allocated with llama_batch_init()
    LLAMA_API void llama_batch_free(struct llama_batch batch);

-    // Process a batch of tokens.
-    // In contrast to llama_decode() - this call does not use KV cache.
-    // For encode-decoder contexts, processes the batch using the encoder.
-    // Can store the encoder output internally for later use by the decoder's cross-attention layers.
+    // Processes a batch of tokens with the ecoder part of the encoder-decoder model.
+    // Stores the encoder output internally for later use by the decoder cross-attention layers.
    //   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)
@ -1438,37 +1434,6 @@ 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
--- a/llama/llama.cpp/src/llama-adapter.cpp
+++ b/llama/llama.cpp/src/llama-adapter.cpp
@ -253,9 +253,6 @@ 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)
@ -294,9 +291,6 @@ 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;
--- a/llama/llama.cpp/src/llama-arch.cpp
+++ b/llama/llama.cpp/src/llama-arch.cpp
@ -6,6 +6,7 @@

 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_LLAMA,            "llama"            },
+    { LLM_ARCH_MLLAMA,           "mllama"           },
    { LLM_ARCH_LLAMA4,           "llama4"           },
    { LLM_ARCH_DECI,             "deci"             },
    { LLM_ARCH_FALCON,           "falcon"           },
@ -144,6 +145,7 @@ 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"             },

@ -273,6 +275,40 @@ 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,
        {
@ -1701,6 +1737,14 @@ 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}},
--- a/llama/llama.cpp/src/llama-arch.h
+++ b/llama/llama.cpp/src/llama-arch.h
@ -11,6 +11,7 @@
 enum llm_arch {
    LLM_ARCH_LLAMA,
    LLM_ARCH_LLAMA4,
+    LLM_ARCH_MLLAMA,
    LLM_ARCH_DECI,
    LLM_ARCH_FALCON,
    LLM_ARCH_BAICHUAN,
@ -148,6 +149,7 @@ 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,

@ -349,6 +351,14 @@ 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,
--- a/llama/llama.cpp/src/llama-batch.cpp
+++ b/llama/llama.cpp/src/llama-batch.cpp
@ -189,7 +189,7 @@ llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
    return ubatch;
 }

-llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
+void llama_sbatch::from_batch(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;
@ -203,7 +203,6 @@ llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple
    for (size_t i = 0; i < n_tokens; ++i) {
        ids[i] = i;
    }
-
    if (simple_split) {
        seq.resize(1);
        llama_sbatch_seq & s = seq[0];
@ -213,7 +212,6 @@ llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple
        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;
@ -241,7 +239,6 @@ llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple
                return n_seq_a > n_seq_b;
            }
    );
-
    // init seq
    llama_sbatch_seq * last_seq = nullptr;

@ -265,7 +262,6 @@ llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple
        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) {
@ -320,6 +316,7 @@ 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,
@ -332,6 +329,7 @@ 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,
@ -340,6 +338,7 @@ 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);
    }
--- a/llama/llama.cpp/src/llama-batch.h
+++ b/llama/llama.cpp/src/llama-batch.h
@ -70,8 +70,7 @@ struct llama_sbatch {
    // sequence-wise split
    llama_ubatch split_seq(size_t n_ubatch);

-    llama_sbatch() = default;
-    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
+    void from_batch(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
--- a/llama/llama.cpp/src/llama-chat.cpp
+++ b/llama/llama.cpp/src/llama-chat.cpp
@ -35,7 +35,6 @@ 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          },
@ -203,20 +202,19 @@ int32_t llm_chat_apply_template(
        if (add_ass) {
            ss << "<|im_start|>assistant\n";
        }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7 || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7) {
        // 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]" << trailing_space << content << "[/SYSTEM_PROMPT]";
+                ss << "[SYSTEM_PROMPT] " << content << "[/SYSTEM_PROMPT]";
            } else if (role == "user") {
-                ss << "[INST]" << trailing_space << content << "[/INST]";
-            } else {
-                ss << trailing_space << content << "</s>";
+                ss << "[INST] " << content << "[/INST]";
+            }
+            else {
+                ss << " " << content << "</s>";
            }
        }
    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1
@ -449,16 +447,8 @@ int32_t llm_chat_apply_template(
        if (add_ass) {
            ss << "<|assistant|>";
        }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4 || tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
        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;
--- a/llama/llama.cpp/src/llama-chat.h
+++ b/llama/llama.cpp/src/llama-chat.h
@ -14,7 +14,6 @@ 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,
--- a/llama/llama.cpp/src/llama-context.cpp
+++ b/llama/llama.cpp/src/llama-context.cpp
--- a/llama/llama.cpp/src/llama-context.h
+++ b/llama/llama.cpp/src/llama-context.h
@ -8,7 +8,6 @@
 #include "llama-kv-cache.h"

 #include "ggml-cpp.h"
-#include "ggml-opt.h"

 #include <map>
 #include <vector>
@ -29,12 +28,7 @@ struct llama_context {

    void synchronize();

-    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;
+    const llama_model & get_model() const;

    uint32_t n_ctx()         const;
    uint32_t n_ctx_per_seq() const;
@ -72,6 +66,7 @@ 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,
@ -135,32 +130,6 @@ 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
@ -170,30 +139,50 @@ 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();

-    // 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);

+    // returns the result of ggml_backend_sched_graph_compute_async execution
+    ggml_status graph_compute(
+            ggml_cgraph * gf,
+                   bool   batched);
+
    llm_graph_cb graph_get_cb() const;

+    // used by kv_self_update()
+    ggml_tensor * build_rope_shift(
+        ggml_context * ctx0,
+        ggml_tensor * cur,
+        ggml_tensor * shift,
+        ggml_tensor * factors,
+              float   freq_base,
+              float   freq_scale) 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);
@ -210,10 +199,14 @@ 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_memory_i> memory;
+    std::unique_ptr<llama_kv_cache_unified> kv_self;
+
+    // TODO: remove
+    bool logits_all = false;

    // decode output (2-dimensional array: [n_outputs][n_vocab])
    size_t  logits_size = 0; // capacity (of floats) for logits
@ -240,9 +233,6 @@ 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;

--- a/llama/llama.cpp/src/llama-cparams.h
+++ b/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;

--- a/llama/llama.cpp/src/llama-graph.cpp
+++ b/llama/llama.cpp/src/llama-graph.cpp
@ -284,7 +284,24 @@ 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) {
-            data[i] = kv_self->s_copy(i);
+            const uint32_t  cell_id = i + kv_self->head;
+
+            //////////////////////////////////////////////
+            // TODO: this should not mutate the KV cache !
+            llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
+
+            // prevent out-of-bound sources
+            if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self->size) {
+                kv_cell.src = cell_id;
+            }
+
+            data[i] = kv_cell.src;
+
+            // TODO: do not mutate the KV cache
+            // ensure copy only happens once
+            if (kv_cell.src != (int32_t) cell_id) {
+                kv_cell.src = cell_id;
+            }
        }
    }
 }
@ -300,7 +317,18 @@ void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) {

        // clear unused states
        for (int i = 0; i < n_kv; ++i) {
-            data[i] = kv_self->s_mask(i);
+            const uint32_t  cell_id = i + kv_self->head;
+
+            //////////////////////////////////////////////
+            // TODO: this should not mutate the KV cache !
+            llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
+
+            data[i] = (float) (kv_cell.src >= 0);
+
+            // only clear once
+            if (kv_cell.src < 0) {
+                kv_cell.src = cell_id;
+            }
        }
    }
 }
@ -532,6 +560,12 @@ 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
 //
@ -782,7 +816,7 @@ ggml_tensor * llm_graph_context::build_ffn(
            } break;
    }

-    if (gate && type_gate == LLM_FFN_PAR) {
+    if (type_gate == LLM_FFN_PAR) {
        cur = ggml_mul(ctx0, cur, tmp);
        cb(cur, "ffn_gate_par", il);
    }
@ -971,7 +1005,6 @@ 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);

@ -1078,7 +1111,7 @@ ggml_tensor * llm_graph_context::build_inp_cls() const {
 }

 ggml_tensor * llm_graph_context::build_inp_s_copy() const {
-    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);

    auto inp = std::make_unique<llm_graph_input_s_copy>(kv_self);

@ -1095,7 +1128,7 @@ ggml_tensor * llm_graph_context::build_inp_s_copy() const {
 }

 ggml_tensor * llm_graph_context::build_inp_s_mask() const {
-    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);

    auto inp = std::make_unique<llm_graph_input_s_mask>(kv_self);

@ -1228,19 +1261,8 @@ 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);
@ -1420,6 +1442,8 @@ 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);
@ -1514,6 +1538,25 @@ 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,
@ -1569,7 +1612,7 @@ 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_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);

    const auto n_kv    = kv_self->n;
    const auto kv_head = kv_self->head;
@ -1601,7 +1644,7 @@ 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_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);

    const auto token_shift_count = hparams.token_shift_count;

@ -1622,7 +1665,7 @@ 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_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);

    const auto token_shift_count = hparams.token_shift_count;
    const auto n_embd = hparams.n_embd;
--- a/llama/llama.cpp/src/llama-graph.h
+++ b/llama/llama.cpp/src/llama-graph.h
@ -19,7 +19,6 @@ 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 {
@ -87,6 +86,7 @@ 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_recurrent * kv_self) : kv_self(kv_self) {}
+    llm_graph_input_s_copy(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
    virtual ~llm_graph_input_s_copy() = default;

    void set_input(const llama_ubatch * ubatch) override;

    ggml_tensor * s_copy; // I32 [kv_size]

-    const llama_kv_cache_recurrent * kv_self;
+    const llama_kv_cache_unified * kv_self;
 };

 class llm_graph_input_s_mask : public llm_graph_input_i {
 public:
-    llm_graph_input_s_mask(const llama_kv_cache_recurrent * kv_self) : kv_self(kv_self) {}
+    llm_graph_input_s_mask(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
    virtual ~llm_graph_input_s_mask() = default;

    void set_input(const llama_ubatch * ubatch) override;

    ggml_tensor * s_mask; // F32 [1, n_kv]

-    const llama_kv_cache_recurrent * kv_self;
+    const llama_kv_cache_unified * kv_self;
 };

 class llm_graph_input_cross_embd : public llm_graph_input_i {
@ -284,6 +284,16 @@ 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
 //
@ -298,7 +308,6 @@ 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;
@ -313,7 +322,6 @@ 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; }
@ -330,7 +338,6 @@ 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;
@ -354,8 +361,8 @@ struct llm_graph_params {
    const llama_cparams & cparams;
    const llama_ubatch  & ubatch;

-    ggml_backend_sched_t sched;
-    ggml_backend_t backend_cpu;
+    ggml_backend_sched * sched;
+    ggml_backend * backend_cpu;

    const llama_adapter_cvec  * cvec;
    const llama_adapter_loras * loras;
@ -406,9 +413,9 @@ struct llm_graph_context {

    ggml_context * ctx0 = nullptr;

-    ggml_backend_sched_t sched;
+    ggml_backend_sched * sched;

-    ggml_backend_t backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
+    ggml_backend * backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?

    const llama_adapter_cvec  * cvec;
    const llama_adapter_loras * loras;
@ -495,6 +502,7 @@ 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;
--- a/llama/llama.cpp/src/llama-hparams.cpp
+++ b/llama/llama.cpp/src/llama-hparams.cpp
@ -85,3 +85,7 @@ 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();
+}
--- a/llama/llama.cpp/src/llama-hparams.h
+++ b/llama/llama.cpp/src/llama-hparams.h
@ -2,6 +2,8 @@

 #include "llama.h"

+#include <algorithm>
+
 #include <array>

 // bump if necessary
@ -42,6 +44,7 @@ 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;
@ -56,6 +59,7 @@ 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;
@ -159,6 +163,9 @@ 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;
 };

--- a/llama/llama.cpp/src/llama-kv-cache.cpp
+++ b/llama/llama.cpp/src/llama-kv-cache.cpp
--- a/llama/llama.cpp/src/llama-kv-cache.h
+++ b/llama/llama.cpp/src/llama-kv-cache.h
@ -2,72 +2,32 @@

 #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 {
-    virtual ~llama_kv_cache() = default;
+    using llama_memory_i::llama_memory_i;

-    // call if batch processing fails - restores the cache state
-    virtual void restore() = 0;
+    virtual void restore() = 0; // call if batch processing fails - restores the cache state
+    virtual void commit() = 0;  // call after successful batch processing - clears any pending state

-    // call after successful batch processing - clears any pending state
-    virtual void commit()  = 0;
+    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

-    // 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;
+    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) {}

@ -82,7 +42,7 @@ 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;
@ -90,50 +50,65 @@ struct llama_kv_defrag_move {
    uint32_t len;
 };

-//
-// llama_kv_cache_unified
-//
+struct llama_kv_cell {
+    llama_pos pos   = -1;
+    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:
-    struct kv_cell {
-        llama_pos pos   = -1;
-        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;
-        }
+    // can be used to query data from the model if needed
+    struct callbacks {
+        std::function<ggml_tensor * (uint32_t n_ctx_per_seq, int il)> get_rope_factors;
    };

-    static uint32_t get_padding(const llama_cparams & cparams);
-
    llama_kv_cache_unified(
-            const llama_model & model,
+            const llama_hparams & hparams,
+            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,
-                     uint32_t   padding);
+                         bool   offload);

-    ~llama_kv_cache_unified() = default;
+    int32_t get_n_tokens()   const override;
+    int32_t get_used_cells() const override;

-    //
-    // llama_memory_i
-    //
+    size_t total_size() const;
+
+    // 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;
@ -143,76 +118,25 @@ public:

    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 & 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;
+    bool get_can_shift() 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) override;
+    bool find_slot(const llama_ubatch & batch);

-    int32_t get_n_tokens()   const override;
-    int32_t get_used_cells() const override;
+    // TODO: maybe not needed
+    uint32_t get_padding(const llama_cparams & cparams) const;

-    // TODO: better data structures to reduce the cost of this operation
-    llama_pos get_pos_max() const override;
+    // find how many cells are currently in use
+    uint32_t cell_max() const;

-    bool get_can_shift() const override;
-
-    // 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;
-
-    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;
+    size_t size_k_bytes() const;
+    size_t size_v_bytes() const;

    // defrag
+
    struct {
        std::vector<llama_kv_defrag_move> moves;
    } defrag_info;
@ -221,6 +145,7 @@ private:
    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;
@ -231,125 +156,25 @@ private:
        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;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1);
+
+    // 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
@ -361,41 +186,18 @@ public:
    // computed before each graph build
    uint32_t n = 0;

-    std::vector<kv_cell> cells;
+    std::vector<llama_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;

@ -403,6 +205,11 @@ 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
--- a/llama/llama.cpp/src/llama-memory.h
+++ b/llama/llama.cpp/src/llama-memory.h
@ -2,22 +2,12 @@

 #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;
--- a/llama/llama.cpp/src/llama-model-loader.cpp
+++ b/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_UINT32:
-            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
-                                                (std::is_same<T, uint32_t>::value)); break;
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T, float>::value)); break;
+            case GGUF_TYPE_INT32:   GGML_ASSERT(
+                                            (std::is_same<T,  int32_t>::value) ||
+                                            (std::is_same<T, uint32_t>::value));  break;
            default:
-                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str()));
        }

        result.resize(arr_info.length);
@ -315,6 +315,8 @@ 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());
@ -330,12 +332,12 @@ namespace GGUFMeta {
            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);

        switch (arr_info.gt) {
-            case GGUF_TYPE_UINT32:
-            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
-                                                (std::is_same<T, uint32_t>::value)); break;
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T, float>::value)); break;
+            case GGUF_TYPE_INT32:   GGML_ASSERT(
+                                            (std::is_same<T,  int32_t>::value) ||
+                                            (std::is_same<T, uint32_t>::value));  break;
            default:
-                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str()));
        }

        if (arr_info.length > N_MAX) {
@ -824,10 +826,6 @@ 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);
--- a/llama/llama.cpp/src/llama-model-saver.cpp
+++ b/llama/llama.cpp/src/llama-model-saver.cpp
@ -1,281 +0,0 @@
-#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);
-}
-
--- a/llama/llama.cpp/src/llama-model-saver.h
+++ b/llama/llama.cpp/src/llama-model-saver.h
@ -1,37 +0,0 @@
-#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);
-};
--- a/llama/llama.cpp/src/llama-model.cpp
+++ b/llama/llama.cpp/src/llama-model.cpp
@ -40,7 +40,6 @@ 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";
@ -80,7 +79,6 @@ 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";
@ -117,10 +115,6 @@ 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) {
@ -303,10 +297,6 @@ 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");
@ -433,6 +423,7 @@ 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) {
@ -444,6 +435,7 @@ 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);
@ -467,9 +459,11 @@ 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;
@ -522,7 +516,7 @@ 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_DECI || arch == LLM_ARCH_FALCON) {
+        if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_MLLAMA || 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));
            }
@ -585,13 +579,22 @@ 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;
@ -718,11 +721,7 @@ 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) {
-                    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;
-                    }
+                    type = LLM_TYPE_137M;
                }
            } break;
        case LLM_ARCH_BLOOM:
@ -783,7 +782,6 @@ 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;
@ -1507,9 +1505,6 @@ 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 {
@ -1581,7 +1576,7 @@ 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       = vocab.n_tokens();
+        const int64_t n_vocab       = hparams.n_vocab;
        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;
@ -1677,11 +1672,8 @@ 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;
                    }
                }
@ -1698,9 +1690,6 @@ 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);
            }

@ -1840,6 +1829,52 @@ 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);
@ -1882,9 +1917,7 @@ 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);

-                        if (n_ff > 0) {
-                            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 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));
@ -1894,11 +1927,9 @@ 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));
                        }

-                        if (n_ff > 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);
-                        }
+                        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);

                        // optional MLP bias
                        layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
@ -3542,11 +3573,7 @@ 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}, 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);
-                    }
+                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);

                    for (int i = 0; i < n_layer; ++i) {
                        auto & layer = layers[i];
@ -4179,9 +4206,6 @@ 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);
@ -4311,7 +4335,7 @@ uint64_t llama_model::n_elements() const {
 }

 void llama_model::print_info() const {
-    const std::string rope_scaling_type = llama_rope_scaling_type_name(hparams.rope_scaling_type_train);
+    const char * rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train);

    auto print_f = [](const std::function<uint32_t(uint32_t)> & f, uint32_t n) {
        bool is_var = false;
@ -4372,7 +4396,7 @@ 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.c_str());
+        LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type);
        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);
@ -4519,19 +4543,6 @@ 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;
@ -4572,7 +4583,7 @@ 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 = model.get_rope_factors(n_ctx_per_seq, il);
+                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);
@ -4756,6 +4767,246 @@ 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;
@ -4778,7 +5029,6 @@ 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
@ -4798,7 +5048,7 @@ 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 = model.get_rope_factors(n_ctx_per_seq, il);
+                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);
@ -4854,11 +5104,6 @@ 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);
@ -7285,7 +7530,7 @@ struct llm_build_phi3 : public llm_graph_context {
            // self-attention
            {
                // rope freq factors for 128k context
-                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);

                ggml_tensor* attn_norm_output = build_norm(inpL,
                        model.layers[il].attn_norm,
@ -8037,7 +8282,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
        for (int il = 0; il < n_layer; ++il) {
            ggml_tensor * inpSA = inpL;

-            ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+            ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);

            // norm
            cur = build_norm(inpL,
@ -8804,7 +9049,7 @@ struct llm_build_mamba : public llm_graph_context {
             ggml_tensor * state_mask,
      const llama_ubatch & ubatch,
                     int   il) const {
-        const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
+        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);

        const auto kv_head = kv_self->head;

@ -9105,7 +9350,7 @@ struct llm_build_cohere2 : public llm_graph_context {
            // self-attention
            {
                // rope freq factors for 128k context
-                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                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);
@ -10043,7 +10288,7 @@ 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 = model.get_rope_factors(n_ctx_per_seq, il);
+                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);
@ -11407,7 +11652,7 @@ 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 = model.get_rope_factors(n_ctx_per_seq, il);
+                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);
@ -11552,7 +11797,7 @@ struct llm_build_rwkv6_base : public llm_graph_context {
            ggml_tensor * state_mask,
            const llama_ubatch & ubatch,
            int   il) const {
-        const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
+        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);

        const auto n_tokens = ubatch.n_tokens;
        const auto n_seqs = ubatch.n_seqs;
@ -11948,7 +12193,7 @@ 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_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
+        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);

        const auto n_tokens = ubatch.n_tokens;
        const auto n_seqs = ubatch.n_seqs;
@ -12496,7 +12741,7 @@ 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 = model.get_rope_factors(n_ctx_per_seq, il);
+                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);
@ -12947,7 +13192,7 @@ 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 = model.get_rope_factors(n_ctx_per_seq, il);
+                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);
@ -13067,46 +13312,36 @@ struct llm_build_bailingmoe : public llm_graph_context {
    }
 };

-llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
+llama_memory_i * llama_model::create_memory() 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_recurrent(
-                        *this,
-                        GGML_TYPE_F32,
-                        GGML_TYPE_F32,
-                        cparams.offload_kqv,
-                        std::max((uint32_t) 1, cparams.n_seq_max));
+                res = new llama_kv_cache_unified(hparams, {
+                    /*.get_rope_factors =*/ nullptr
+                });
            } break;
        default:
            {
-                const auto padding = llama_kv_cache_unified::get_padding(cparams);
+                res = new llama_kv_cache_unified(hparams, {
+                    /*.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;
+                        }

-                cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
+                        if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
+                            return layers[il].rope_long;
+                        }

-                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);
+                        return layers[il].rope_short;
+                    }
+                });
            }
    }

@ -13128,6 +13363,10 @@ 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);
@ -13489,9 +13728,12 @@ 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:
@ -13530,7 +13772,6 @@ 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:
@ -13538,7 +13779,6 @@ 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:
@ -13611,14 +13851,6 @@ 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;
    }

--- a/llama/llama.cpp/src/llama-model.h
+++ b/llama/llama.cpp/src/llama-model.h
@ -11,6 +11,7 @@
 #include <string>
 #include <unordered_map>
 #include <vector>
+#include <stdexcept>

 struct llama_cparams;
 struct llama_ubatch;
@ -36,7 +37,6 @@ 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,
@ -97,8 +97,6 @@ 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;
@ -318,6 +316,16 @@ 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;
@ -401,11 +409,8 @@ 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 llama_memory_params & params, llama_cparams & cparams) const;
+    llama_memory_i * create_memory() const; // TODO: params

    // TODO: move this to new llm_arch_model_i interface
    llm_graph_result_ptr build_graph(
--- a/llama/llama.cpp/src/llama-quant.cpp
+++ b/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 on Linux, and also increases speed on Windows with
+    // mmap consistently increases speed 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();
    }

@ -639,7 +639,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
        if (llama_model_has_encoder(&model)) {
            n_attn_layer *= 3;
        }
-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
+        if (qs.n_attention_wv != n_attn_layer) {
+            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;
--- a/llama/llama.cpp/src/llama-sampling.cpp
+++ b/llama/llama.cpp/src/llama-sampling.cpp
@ -1750,35 +1750,23 @@ 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) {
-        // Only count non-negative infinity values
-        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++;
+        if (cur_p->data[i].logit > max) {
+            max = cur_p->data[i].logit;
        }
+        logits_sum += cur_p->data[i].logit;
    }
-    float mean = valid_count > 0 ? logits_sum/valid_count : 0;
+    float mean = logits_sum/cur_p->size;

    // calculate standard deviation
    float acc = 0;
    for (size_t i = 0; i < cur_p->size; ++i) {
-        // Skip -infinity in std calculation
-        if (cur_p->data[i].logit != -INFINITY) {
-            acc += pow(cur_p->data[i].logit - mean, 2);
-        }
+        acc += pow(cur_p->data[i].logit - mean, 2);
    }
-    float std = valid_count > 0 ? sqrt(acc/valid_count) : 0;
+    float std = sqrt(acc/cur_p->size);

    //apply mask
    for (size_t i = 0; i < cur_p->size; ++i) {
--- a/llama/llama.cpp/src/llama-vocab.cpp
+++ b/llama/llama.cpp/src/llama-vocab.cpp
@ -1,7 +1,5 @@
 #include "llama-vocab.h"

-#include "ggml.h"
-#include "gguf.h"
 #include "llama-impl.h"
 #include "llama-model-loader.h"

@ -417,13 +415,6 @@ 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 = {
@ -1236,9 +1227,6 @@ 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;

@ -1374,6 +1362,9 @@ 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);

@ -1468,8 +1459,7 @@ 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) {
-                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);
+                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
@ -1635,10 +1625,6 @@ 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;
@ -2784,14 +2770,6 @@ 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;
 }
@ -3014,20 +2992,6 @@ 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,
--- a/llama/llama.cpp/src/llama-vocab.h
+++ b/llama/llama.cpp/src/llama-vocab.h
@ -21,9 +21,6 @@ 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;

@ -83,9 +80,6 @@ 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,
--- a/llama/llama.cpp/src/llama.cpp
+++ b/llama/llama.cpp/src/llama.cpp
@ -4,7 +4,6 @@
 #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"
@ -254,13 +253,6 @@ 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
 //
@ -346,4 +338,3 @@ const char * llama_print_system_info(void) {

    return s.c_str();
 }
-
--- a/llama/llama.go
+++ b/llama/llama.go
@ -6,7 +6,7 @@ package llama
 #cgo CXXFLAGS: -std=c++17
 #cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/include
 #cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/common
-#cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/tools/mtmd
+#cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/examples/llava
 #cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/src
 #cgo CPPFLAGS: -I${SRCDIR}/../ml/backend/ggml/ggml/include

@ -17,6 +17,7 @@ package llama
 #include "llava.h"
 #include "gguf.h"

+#include "mllama.h"
 #include "sampling_ext.h"

 extern bool llamaProgressCallback(float progress, void *user_data);
@ -39,8 +40,8 @@ import (
 	"unsafe"

 	_ "github.com/ollama/ollama/llama/llama.cpp/common"
+	_ "github.com/ollama/ollama/llama/llama.cpp/examples/llava"
 	_ "github.com/ollama/ollama/llama/llama.cpp/src"
-	_ "github.com/ollama/ollama/llama/llama.cpp/tools/mtmd"
 	ggml "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
 )

@ -509,6 +510,63 @@ func (c *ClipContext) NewEmbed(llamaContext *Context, data []byte) ([][]float32,
 	return embed, nil
 }

+type MllamaContext struct {
+	c *C.struct_mllama_ctx
+}
+
+func NewMllamaContext(llamaContext *Context, modelPath string) (*MllamaContext, error) {
+	mp := C.CString(modelPath)
+	defer C.free(unsafe.Pointer(mp))
+	c := C.mllama_model_load(mp, 1)
+	if c == nil {
+		return nil, fmt.Errorf("unable to load mllama model: %v", modelPath)
+	}
+
+	projEmbedSize := int(C.mllama_n_embd(c))
+	modelEmbedSize := llamaContext.Model().NEmbd()
+	if projEmbedSize != modelEmbedSize {
+		return nil, fmt.Errorf("projector embedding size (%d) does not match model (%d)", projEmbedSize, modelEmbedSize)
+	}
+
+	return &MllamaContext{c: c}, nil
+}
+
+func (m *MllamaContext) Free() {
+	C.mllama_free(m.c)
+}
+
+func (m *MllamaContext) NewEmbed(llamaContext *Context, data []byte, aspectRatioId int) ([][]float32, error) {
+	img := C.mllama_image_init()
+	defer C.mllama_image_free(img)
+
+	ok := bool(C.mllama_image_load_from_data(unsafe.Pointer(&data[0]), C.int(len(data)), 560, 560, 3, 4, C.int(aspectRatioId), img))
+	if !ok {
+		return nil, errors.New("unable to load mllama image data")
+	}
+
+	rows := make([]float32, m.EmbedSize(llamaContext))
+	ok = bool(C.mllama_image_encode(m.c, C.int(llamaContext.numThreads), img, (*C.float)(unsafe.Pointer(&rows[0]))))
+	if !ok {
+		return nil, errors.New("unable to make mllama embedding from image")
+	}
+
+	embed := make([][]float32, 1)
+	embed[0] = rows
+
+	return embed, nil
+}
+
+func (m *MllamaContext) EmbedSize(llamaContext *Context) int {
+	numTokens := int(C.mllama_n_positions(m.c) * C.mllama_n_tiles(m.c))
+	numEmbed := llamaContext.Model().NEmbd()
+
+	return numTokens * numEmbed
+}
+
+func (c *Context) SetCrossAttention(state bool) {
+	C.llama_set_cross_attention(c.c, C.bool(state))
+}
+
 func (c *Context) Synchronize() {
 	C.llama_synchronize(c.c)
 }
@ -529,6 +587,9 @@ type SamplingParams struct {
 	PenaltyRepeat  float32
 	PenaltyFreq    float32
 	PenaltyPresent float32
+	Mirostat       int
+	MirostatTau    float32
+	MirostatEta    float32
 	PenalizeNl     bool
 	Seed           uint32
 	Grammar        string
@ -545,6 +606,9 @@ func NewSamplingContext(model *Model, params SamplingParams) (*SamplingContext,
 	cparams.penalty_repeat = C.float(params.PenaltyRepeat)
 	cparams.penalty_freq = C.float(params.PenaltyFreq)
 	cparams.penalty_present = C.float(params.PenaltyFreq)
+	cparams.mirostat = C.int32_t(params.Mirostat)
+	cparams.mirostat_tau = C.float(params.MirostatTau)
+	cparams.mirostat_eta = C.float(params.MirostatEta)
 	cparams.seed = C.uint32_t(params.Seed)

 	grammar := C.CString(params.Grammar)
@ -579,8 +643,8 @@ func SchemaToGrammar(schema []byte) []byte {
 	cStr := C.CString(string(schema))
 	defer C.free(unsafe.Pointer(cStr))

-	// Allocate buffer for grammar based on schema length but with upper bound
-	maxLen := min(1024*1024, len(schema)*4)
+	// Allocate buffer for grammar output with reasonable size
+	const maxLen = 32768 // 32KB
 	buf := make([]byte, maxLen)

 	// Call C function to convert schema to grammar
--- a/llama/mllama.cpp
+++ b/llama/mllama.cpp
@ -0,0 +1,887 @@
+// NOTE: This is modified from clip.cpp for Mllama only
+#include "mllama.h"
+
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include "ggml-cpu.h"
+#include "ggml.h"
+#include "gguf.h"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
+#include <algorithm>
+#include <cmath>
+#include <cstdarg>
+#include <cstdlib>
+#include <cstring>
+#include <fstream>
+#include <stdexcept>
+#include <vector>
+
+#define REQUIRE(x)                                           \
+    do {                                                     \
+        if (!(x)) {                                          \
+            throw std::runtime_error("REQUIRE failed: " #x); \
+        }                                                    \
+    } while (0)
+
+#define LOG(fmt, ...) fprintf(stderr, "%s: " fmt "\n", __func__, ##__VA_ARGS__)
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
+#include <windows.h>
+#if __GLIBCXX__
+#include <cstdio>
+#include <ext/stdio_filebuf.h>
+#include <fcntl.h>
+#endif
+#endif
+
+struct mllama_image {
+    int width;
+    int height;
+
+    int num_channels = 3;
+    int num_tiles = 4;
+
+    int aspect_ratio_id;
+
+    std::vector<float> data;
+};
+
+static std::string format(const char *fmt, ...) {
+    va_list args;
+    va_start(args, fmt);
+    std::vector<char> b(128);
+    int n = vsnprintf(b.data(), b.size(), fmt, args);
+    REQUIRE(n >= 0 && n < b.size());
+    va_end(args);
+    return std::string(b.data(), b.size());
+}
+
+//
+// utilities to get data from a gguf file
+//
+
+static int get_key_index(const gguf_context *ctx, const char *key) {
+    int key_index = gguf_find_key(ctx, key);
+    REQUIRE(key_index != -1);
+    return key_index;
+}
+
+static std::vector<uint32_t> get_u32_array(const gguf_context *ctx, const std::string &key) {
+    const int i = get_key_index(ctx, key.c_str());
+    const int n = gguf_get_arr_n(ctx, i);
+    const uint32_t *data = (uint32_t *)gguf_get_arr_data(ctx, i);
+
+    std::vector<uint32_t> s(n);
+    for (size_t j = 0; j < s.size(); j++) {
+        s[j] = data[j];
+    }
+
+    return s;
+}
+
+static uint32_t get_u32(const gguf_context *ctx, const std::string &key) {
+    return gguf_get_val_u32(ctx, get_key_index(ctx, key.c_str()));
+}
+
+static float get_f32(const gguf_context *ctx, const std::string &key) {
+    return gguf_get_val_f32(ctx, get_key_index(ctx, key.c_str()));
+}
+
+static std::string get_ftype(int ftype) {
+    return ggml_type_name(static_cast<ggml_type>(ftype));
+}
+
+//
+// mllama layers
+//
+
+struct mllama_hparams {
+    uint32_t image_size;
+    uint32_t patch_size;
+    uint32_t hidden_size;
+    uint32_t n_intermediate;
+    uint32_t projection_dim;
+    uint32_t n_head;
+    uint32_t n_layer;
+    uint32_t n_global_layer;
+    uint32_t n_tiles;
+
+    float eps;
+
+    std::vector<bool> intermediate_layers;
+};
+
+struct mllama_layer {
+    // attention
+    struct ggml_tensor *k_w;
+    struct ggml_tensor *k_b;
+    struct ggml_tensor *q_w;
+    struct ggml_tensor *q_b;
+    struct ggml_tensor *v_w;
+    struct ggml_tensor *v_b;
+
+    struct ggml_tensor *o_w;
+    struct ggml_tensor *o_b;
+
+    struct ggml_tensor *attn_gate;
+
+    // layernorm 1
+    struct ggml_tensor *ln_1_w;
+    struct ggml_tensor *ln_1_b;
+
+    // ff
+    struct ggml_tensor *ff_i_w;
+    struct ggml_tensor *ff_i_b;
+
+    struct ggml_tensor *ff_o_w;
+    struct ggml_tensor *ff_o_b;
+
+    struct ggml_tensor *ff_gate;
+
+    // layernorm 2
+    struct ggml_tensor *ln_2_w;
+    struct ggml_tensor *ln_2_b;
+};
+
+struct mllama_vision_model {
+    struct mllama_hparams hparams;
+
+    // embeddings
+    struct ggml_tensor *class_embedding;
+    struct ggml_tensor *patch_embeddings;
+    struct ggml_tensor *position_embeddings;
+    struct ggml_tensor *position_embeddings_gate;
+    struct ggml_tensor *tile_position_embeddings;
+    struct ggml_tensor *tile_position_embeddings_gate;
+    struct ggml_tensor *pre_tile_position_embeddings;
+    struct ggml_tensor *pre_tile_position_embeddings_gate;
+    struct ggml_tensor *post_tile_position_embeddings;
+    struct ggml_tensor *post_tile_position_embeddings_gate;
+
+    struct ggml_tensor *pre_ln_w;
+    struct ggml_tensor *pre_ln_b;
+
+    std::vector<mllama_layer> layers;
+    std::vector<mllama_layer> global_layers;
+
+    struct ggml_tensor *post_ln_w;
+    struct ggml_tensor *post_ln_b;
+
+    struct ggml_tensor *mm_0_w;
+    struct ggml_tensor *mm_0_b;
+};
+
+struct mllama_ctx {
+    struct mllama_vision_model vision_model;
+
+    uint32_t ftype = 1;
+
+    struct gguf_context *ctx_gguf;
+    struct ggml_context *ctx_data;
+
+    std::vector<uint8_t> buf_compute_meta;
+
+    // memory buffers to evaluate the model
+    ggml_backend_buffer_t params_buffer = nullptr;
+
+    ggml_backend_t backend = nullptr;
+    ggml_gallocr_t compute_alloc = nullptr;
+};
+
+static ggml_tensor *mllama_image_build_encoder_layer(
+    struct ggml_context *ctx0, const size_t il, const struct mllama_layer &layer, struct ggml_tensor *embeddings,
+    const float eps, const int hidden_size, const int batch_size, const int n_head, const int d_head) {
+    struct ggml_tensor *cur = embeddings;
+
+    {
+        // layernorm1
+        cur = ggml_norm(ctx0, cur, eps);
+        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ln_1_w), layer.ln_1_b);
+        ggml_set_name(cur, format("%d pre layernorm", il).c_str());
+    }
+
+    {
+        // self-attention
+        struct ggml_tensor *Q = ggml_mul_mat(ctx0, layer.q_w, cur);
+        if (layer.q_b != nullptr) {
+            Q = ggml_add(ctx0, Q, layer.q_b);
+        }
+
+        Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, Q->ne[1], batch_size);
+        Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+        ggml_set_name(Q, format("%d query", il).c_str());
+
+        struct ggml_tensor *K = ggml_mul_mat(ctx0, layer.k_w, cur);
+        if (layer.k_b != nullptr) {
+            K = ggml_add(ctx0, K, layer.k_b);
+        }
+
+        K = ggml_reshape_4d(ctx0, K, d_head, n_head, K->ne[1], batch_size);
+        K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+        ggml_set_name(K, format("%d key", il).c_str());
+
+        struct ggml_tensor *V = ggml_mul_mat(ctx0, layer.v_w, cur);
+        if (layer.v_b != nullptr) {
+            V = ggml_add(ctx0, V, layer.v_b);
+        }
+
+        V = ggml_reshape_4d(ctx0, V, d_head, n_head, V->ne[1], batch_size);
+        V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+        ggml_set_name(V, format("%d value", il).c_str());
+
+        struct ggml_tensor *KQ = ggml_mul_mat(ctx0, K, Q);
+        KQ = ggml_scale_inplace(ctx0, KQ, 1.0f / sqrtf((float)d_head));
+        KQ = ggml_soft_max_inplace(ctx0, KQ);
+        ggml_set_name(KQ, format("%d KQ", il).c_str());
+
+        struct ggml_tensor *KQV = ggml_mul_mat(ctx0, V, KQ);
+        KQV = ggml_reshape_4d(ctx0, KQV, d_head, KQV->ne[1], n_head, batch_size);
+        KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+        KQV = ggml_cont_3d(ctx0, KQV, hidden_size, KQV->ne[2], batch_size);
+        ggml_set_name(KQV, format("%d KQV", il).c_str());
+
+        cur = ggml_mul_mat(ctx0, layer.o_w, KQV);
+        if (layer.o_b != nullptr) {
+            cur = ggml_add(ctx0, cur, layer.o_b);
+        }
+        ggml_set_name(cur, format("%d self attention", il).c_str());
+
+        if (layer.attn_gate != nullptr) {
+            cur = ggml_mul_inplace(ctx0, cur, layer.attn_gate);
+            ggml_set_name(cur, format("%d self attention gate", il).c_str());
+        }
+    }
+
+    cur = ggml_add(ctx0, cur, embeddings);
+    ggml_set_name(cur, format("%d residual", il).c_str());
+
+    embeddings = cur;
+
+    {
+        // layernorm2
+        cur = ggml_norm(ctx0, cur, eps);
+        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ln_2_w), layer.ln_2_b);
+        ggml_set_name(cur, format("%d post layernorm", il).c_str());
+    }
+
+    {
+        // feed forward
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, layer.ff_i_w, cur), layer.ff_i_b);
+        cur = ggml_gelu_inplace(ctx0, cur);
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, layer.ff_o_w, cur), layer.ff_o_b);
+        ggml_set_name(cur, format("%d feed forward", il).c_str());
+
+        if (layer.ff_gate != nullptr) {
+            cur = ggml_mul_inplace(ctx0, cur, layer.ff_gate);
+            ggml_set_name(cur, format("%d feed forward gate", il).c_str());
+        }
+    }
+
+    // residual 2
+    cur = ggml_add(ctx0, cur, embeddings);
+    ggml_set_name(cur, format("%d residual", il).c_str());
+
+    embeddings = cur;
+
+    return embeddings;
+}
+
+static ggml_cgraph *mllama_image_build_graph(mllama_ctx *ctx, const mllama_image_batch *imgs) {
+    const auto &model = ctx->vision_model;
+    const auto &hparams = model.hparams;
+
+    const int image_size = hparams.image_size;
+    const int image_size_width = image_size;
+    const int image_size_height = image_size;
+
+    const int patch_size = hparams.patch_size;
+    const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int num_positions = num_patches + (model.class_embedding == nullptr ? 0 : 1);
+    const int hidden_size = hparams.hidden_size;
+    const int n_head = hparams.n_head;
+    const int d_head = hidden_size / n_head;
+
+    const int batch_size = imgs->size;
+    REQUIRE(batch_size == 1);
+
+    int num_tiles = 4;
+    int num_channels = 3;
+    if (imgs->data != nullptr) {
+        num_tiles = imgs->data[0].num_tiles > 0 ? imgs->data[0].num_tiles : num_tiles;
+        num_channels = imgs->data[0].num_channels > 0 ? imgs->data[0].num_channels : num_channels;
+    }
+
+    struct ggml_init_params params = {
+        ctx->buf_compute_meta.size(), // mem_size
+        ctx->buf_compute_meta.data(), // mem_buffer
+        true,                         // no_alloc
+    };
+
+    struct ggml_context *ctx0 = ggml_init(params);
+    struct ggml_cgraph *gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor *inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, num_channels, num_tiles);
+    ggml_set_name(inp_raw, "inp_raw");
+    ggml_set_input(inp_raw);
+
+    struct ggml_tensor *inp = ggml_conv_2d(ctx0, model.patch_embeddings, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, num_tiles);
+    inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
+
+    struct ggml_tensor *aspect_ratios = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, imgs->size);
+    ggml_set_name(aspect_ratios, "aspect_ratios");
+    ggml_set_input(aspect_ratios);
+
+    if (model.pre_tile_position_embeddings != nullptr) {
+        struct ggml_tensor *pre_tile_position_embeddings = ggml_get_rows(ctx0, model.pre_tile_position_embeddings, aspect_ratios);
+        ggml_set_name(pre_tile_position_embeddings, "pre_tile_position_embeddings");
+
+        pre_tile_position_embeddings = ggml_reshape_3d(ctx0, pre_tile_position_embeddings, hidden_size, 1, num_tiles);
+        if (model.pre_tile_position_embeddings_gate != nullptr) {
+            pre_tile_position_embeddings = ggml_mul_inplace(ctx0, pre_tile_position_embeddings, model.pre_tile_position_embeddings_gate);
+        }
+
+        inp = ggml_add(ctx0, inp, pre_tile_position_embeddings);
+    }
+
+    struct ggml_tensor *embeddings = inp;
+
+    if (model.class_embedding != nullptr) {
+        // concat class_embeddings and patch_embeddings
+        embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, num_tiles);
+        ggml_set_name(embeddings, "embeddings");
+        ggml_set_input(embeddings);
+        for (int i = 0; i < num_tiles; ++i) {
+            // repeat class embeddings for each tile
+            embeddings = ggml_acc_inplace(ctx0, embeddings, model.class_embedding, embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], i * embeddings->nb[2]);
+        }
+
+        embeddings = ggml_acc_inplace(ctx0, embeddings, inp, embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+    }
+
+    struct ggml_tensor *positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    struct ggml_tensor *position_embd = ggml_get_rows(ctx0, model.position_embeddings, positions);
+    if (model.position_embeddings_gate != nullptr) {
+        position_embd = ggml_mul_inplace(ctx0, position_embd, model.position_embeddings_gate);
+    }
+
+    embeddings = ggml_add(ctx0, embeddings, position_embd);
+
+    if (model.tile_position_embeddings != nullptr) {
+        struct ggml_tensor *tile_position_embeddings = ggml_get_rows(ctx0, model.tile_position_embeddings, aspect_ratios);
+        ggml_set_name(tile_position_embeddings, "tile_position_embeddings");
+
+        tile_position_embeddings = ggml_reshape_3d(ctx0, tile_position_embeddings, hidden_size, num_positions, num_tiles);
+        if (model.tile_position_embeddings_gate != nullptr) {
+            tile_position_embeddings = ggml_mul_inplace(ctx0, tile_position_embeddings, model.tile_position_embeddings_gate);
+        }
+
+        embeddings = ggml_add(ctx0, embeddings, tile_position_embeddings);
+    }
+
+    // pre-layernorm
+    if (model.pre_ln_w != nullptr) {
+        embeddings = ggml_mul(ctx0, ggml_norm(ctx0, embeddings, hparams.eps), model.pre_ln_w);
+        if (model.pre_ln_b != nullptr) {
+            embeddings = ggml_add(ctx0, embeddings, model.pre_ln_b);
+        }
+
+        ggml_set_name(embeddings, "pre layernorm");
+    }
+
+    const int num_padding_patches = 8 - (embeddings->ne[1] % 8) % 8;
+
+    embeddings = ggml_pad(ctx0, embeddings, 0, num_padding_patches, 0, 0);
+    embeddings = ggml_view_3d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1] * embeddings->ne[2], batch_size, embeddings->nb[1], embeddings->nb[2] * embeddings->ne[3], 0);
+
+    std::vector<struct ggml_tensor *> intermediate_embeddings;
+
+    // encoder
+    for (size_t il = 0; il < model.layers.size(); il++) {
+        if (hparams.intermediate_layers[il]) {
+            intermediate_embeddings.push_back(embeddings);
+        }
+
+        embeddings = mllama_image_build_encoder_layer(
+            ctx0, il, model.layers[il], embeddings,
+            hparams.eps, hidden_size, batch_size, n_head, d_head);
+    }
+
+    // post-layernorm
+    if (model.post_ln_w != nullptr) {
+        embeddings = ggml_mul(ctx0, ggml_norm(ctx0, embeddings, hparams.eps), model.post_ln_w);
+        if (model.post_ln_b != nullptr) {
+            embeddings = ggml_add(ctx0, embeddings, model.post_ln_b);
+        }
+
+        ggml_set_name(embeddings, "post layernorm");
+    }
+
+    embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size, num_positions + num_padding_patches, num_tiles);
+
+    if (model.post_tile_position_embeddings != nullptr) {
+        struct ggml_tensor *post_tile_position_embeddings = ggml_get_rows(ctx0, model.post_tile_position_embeddings, aspect_ratios);
+        ggml_set_name(post_tile_position_embeddings, "post_tile_position_embeddings");
+
+        post_tile_position_embeddings = ggml_reshape_3d(ctx0, post_tile_position_embeddings, hidden_size, 1, num_tiles);
+        if (model.post_tile_position_embeddings_gate != nullptr) {
+            post_tile_position_embeddings = ggml_mul(ctx0, post_tile_position_embeddings, model.post_tile_position_embeddings_gate);
+        }
+
+        embeddings = ggml_add(ctx0, embeddings, post_tile_position_embeddings);
+    }
+
+    embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size, num_tiles * (num_positions + num_padding_patches), 1);
+
+    // global encoder
+    for (size_t il = 0; il < model.global_layers.size(); il++) {
+        embeddings = mllama_image_build_encoder_layer(
+            ctx0, il, model.global_layers[il], embeddings,
+            hparams.eps, hidden_size, batch_size, n_head, d_head);
+    }
+
+    struct ggml_tensor *stacked_embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 0, hidden_size, (num_positions + num_padding_patches) * num_tiles);
+    for (size_t i = 0; i < intermediate_embeddings.size(); ++i) {
+        stacked_embeddings = ggml_concat(ctx0, stacked_embeddings, ggml_reshape_3d(ctx0, intermediate_embeddings[i], 1, intermediate_embeddings[i]->ne[0], intermediate_embeddings[i]->ne[1]), 0);
+    }
+
+    stacked_embeddings = ggml_reshape_4d(ctx0, stacked_embeddings, intermediate_embeddings.size() * hidden_size, num_positions + num_padding_patches, num_tiles, batch_size);
+    stacked_embeddings = ggml_unpad(ctx0, stacked_embeddings, 0, num_padding_patches, 0, 0);
+
+    embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size, num_positions + num_padding_patches, num_tiles);
+    embeddings = ggml_unpad(ctx0, embeddings, 0, num_padding_patches, 0, 0);
+    embeddings = ggml_concat(ctx0, embeddings, stacked_embeddings, 0);
+
+    // mllama projector
+    embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_0_w, embeddings), model.mm_0_b);
+    ggml_set_name(embeddings, "multi modal projector");
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+static struct ggml_tensor *mllama_tensor_load(struct ggml_context *ctx, const char *name, const bool optional) {
+    struct ggml_tensor *cur = ggml_get_tensor(ctx, name);
+    REQUIRE(cur != nullptr || optional);
+    return cur;
+}
+
+static std::vector<struct mllama_layer> mllama_layers_load(struct ggml_context *ctx, const char *prefix, const int n) {
+    std::vector<struct mllama_layer> layers(n);
+    for (size_t i = 0; i < layers.size(); i++) {
+        auto &layer = layers[i];
+        layer.ln_1_w = mllama_tensor_load(ctx, format("%s.blk.%d.ln1.weight", prefix, i).c_str(), false);
+        layer.ln_1_b = mllama_tensor_load(ctx, format("%s.blk.%d.ln1.bias", prefix, i).c_str(), false);
+        layer.ln_2_w = mllama_tensor_load(ctx, format("%s.blk.%d.ln2.weight", prefix, i).c_str(), false);
+        layer.ln_2_b = mllama_tensor_load(ctx, format("%s.blk.%d.ln2.bias", prefix, i).c_str(), false);
+
+        layer.k_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_k.weight", prefix, i).c_str(), false);
+        layer.k_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_k.bias", prefix, i).c_str(), true);
+        layer.q_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_q.weight", prefix, i).c_str(), false);
+        layer.q_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_q.bias", prefix, i).c_str(), true);
+        layer.v_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_v.weight", prefix, i).c_str(), false);
+        layer.v_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_v.bias", prefix, i).c_str(), true);
+        layer.o_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_out.weight", prefix, i).c_str(), false);
+        layer.o_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_out.bias", prefix, i).c_str(), true);
+
+        layer.ff_i_w = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_down.weight", prefix, i).c_str(), false);
+        layer.ff_i_b = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_down.bias", prefix, i).c_str(), false);
+        layer.ff_o_w = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_up.weight", prefix, i).c_str(), false);
+        layer.ff_o_b = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_up.bias", prefix, i).c_str(), false);
+
+        layer.attn_gate = mllama_tensor_load(ctx, format("%s.blk.%d.attn_gate", prefix, i).c_str(), true);
+        layer.ff_gate = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_gate", prefix, i).c_str(), true);
+    }
+
+    return layers;
+}
+
+// read and create ggml_context containing the tensors and their data
+struct mllama_ctx *mllama_model_load(const char *fname, const int verbosity = 1) {
+    struct ggml_context *meta = nullptr;
+
+    struct gguf_init_params params = {
+        true,  // no_alloc
+        &meta, // ctx
+    };
+
+    struct gguf_context *ctx = gguf_init_from_file(fname, params);
+    REQUIRE(ctx != nullptr);
+
+    if (verbosity >= 1) {
+        const int n_tensors = gguf_get_n_tensors(ctx);
+        const int n_kv = gguf_get_n_kv(ctx);
+        const std::string ftype = get_ftype(get_u32(ctx, "general.file_type"));
+        const int idx_desc = get_key_index(ctx, "general.description");
+        const std::string description = gguf_get_val_str(ctx, idx_desc);
+        const int idx_name = gguf_find_key(ctx, "general.name");
+        if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
+            const std::string name = gguf_get_val_str(ctx, idx_name);
+            LOG("model name:   %s", name.c_str());
+        }
+        LOG("description:  %s", description.c_str());
+        LOG("GGUF version: %d", gguf_get_version(ctx));
+        LOG("alignment:    %zu", gguf_get_alignment(ctx));
+        LOG("n_tensors:    %d", n_tensors);
+        LOG("n_kv:         %d", n_kv);
+        LOG("ftype:        %s", ftype.c_str());
+        LOG("");
+    }
+    const int n_tensors = gguf_get_n_tensors(ctx);
+
+    mllama_ctx *new_mllama = new mllama_ctx{};
+
+    ggml_backend_t backend = ggml_backend_init_best();
+    if (backend == nullptr) {
+        LOG("%s: failed to initialize backend\n", __func__);
+        mllama_free(new_mllama);
+        gguf_free(ctx);
+        return nullptr;
+    }
+    LOG("%s: using %s backend\n", __func__, ggml_backend_name(backend));
+    new_mllama->backend = backend;
+
+    // load tensors
+    {
+        std::vector<uint8_t> read_buf;
+        struct ggml_init_params params = {
+            (n_tensors + 1) * ggml_tensor_overhead(), // mem_size
+            nullptr,                                  // mem_buffer
+            true,                                     // no_alloc
+        };
+
+        new_mllama->ctx_data = ggml_init(params);
+        if (!new_mllama->ctx_data) {
+            LOG("ggml_init() failed");
+            mllama_free(new_mllama);
+            gguf_free(ctx);
+            return nullptr;
+        }
+
+#ifdef _WIN32
+        int wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, NULL, 0);
+        if (!wlen) {
+            return NULL;
+        }
+        wchar_t * wbuf = (wchar_t *) malloc(wlen * sizeof(wchar_t));
+        wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, wbuf, wlen);
+        if (!wlen) {
+            free(wbuf);
+            return NULL;
+        }
+#if __GLIBCXX__
+        int fd = _wopen(wbuf, _O_RDONLY | _O_BINARY);
+        __gnu_cxx::stdio_filebuf<char> buffer(fd, std::ios_base::in);
+        std::istream fin(&buffer);
+#else // MSVC
+        // unused in our current build
+        auto fin = std::ifstream(wbuf, std::ios::binary);
+#endif
+        free(wbuf);
+#else
+        auto fin = std::ifstream(fname, std::ios::binary);
+#endif
+        if (!fin) {
+            LOG("cannot open model file for loading tensors\n");
+            mllama_free(new_mllama);
+            gguf_free(ctx);
+            return nullptr;
+        }
+
+        // add tensors to context
+        for (int i = 0; i < n_tensors; ++i) {
+            const char *name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor *t = ggml_get_tensor(meta, name);
+            struct ggml_tensor *cur = ggml_dup_tensor(new_mllama->ctx_data, t);
+            ggml_set_name(cur, name);
+        }
+
+        // alloc memory and offload data
+        new_mllama->params_buffer = ggml_backend_alloc_ctx_tensors(new_mllama->ctx_data, new_mllama->backend);
+        for (int i = 0; i < n_tensors; ++i) {
+            const char *name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor *cur = ggml_get_tensor(new_mllama->ctx_data, name);
+            const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
+            fin.seekg(offset, std::ios::beg);
+            if (!fin) {
+                LOG("failed to seek for tensor %s\n", name);
+                mllama_free(new_mllama);
+                gguf_free(ctx);
+                return nullptr;
+            }
+            int num_bytes = ggml_nbytes(cur);
+            if (ggml_backend_buffer_is_host(new_mllama->params_buffer)) {
+                // for the CPU and Metal backend, we can read directly into the tensor
+                fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
+            } else {
+                // read into a temporary buffer first, then copy to device memory
+                read_buf.resize(num_bytes);
+                fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
+                ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
+            }
+        }
+
+#if defined(_WIN32) && defined(__GLIBCXX__)
+        close(fd);
+#else
+        fin.close();
+#endif
+    }
+
+    // vision model
+    // load vision model
+    auto &vision_model = new_mllama->vision_model;
+    auto &hparams = vision_model.hparams;
+    hparams.hidden_size = get_u32(ctx, "mllama.vision.embedding_length");
+    hparams.n_head = get_u32(ctx, "mllama.vision.attention.head_count");
+    hparams.n_intermediate = get_u32(ctx, "mllama.vision.feed_forward_length");
+    hparams.n_layer = get_u32(ctx, "mllama.vision.block_count");
+    hparams.n_global_layer = get_u32(ctx, "mllama.vision.global.block_count");
+    hparams.n_tiles = get_u32(ctx, "mllama.vision.max_num_tiles");
+    hparams.image_size = get_u32(ctx, "mllama.vision.image_size");
+    hparams.patch_size = get_u32(ctx, "mllama.vision.patch_size");
+    hparams.projection_dim = get_u32(ctx, "mllama.vision.projection_dim");
+    hparams.eps = get_f32(ctx, "mllama.vision.attention.layer_norm_epsilon");
+
+    std::vector<uint32_t> intermediate_layers_indices = get_u32_array(ctx, "mllama.vision.intermediate_layers_indices");
+    hparams.intermediate_layers.resize(hparams.n_layer);
+    for (size_t i = 0; i < intermediate_layers_indices.size(); i++) {
+        hparams.intermediate_layers[intermediate_layers_indices[i]] = true;
+    }
+
+    if (verbosity >= 2) {
+        LOG("");
+        LOG("vision model hparams");
+        LOG("image_size         %d", hparams.image_size);
+        LOG("patch_size         %d", hparams.patch_size);
+        LOG("v_hidden_size      %d", hparams.hidden_size);
+        LOG("v_n_intermediate   %d", hparams.n_intermediate);
+        LOG("v_projection_dim   %d", hparams.projection_dim);
+        LOG("v_n_head           %d", hparams.n_head);
+        LOG("v_n_layer          %d", hparams.n_layer);
+        LOG("v_n_global_layer   %d", hparams.n_global_layer);
+        LOG("v_eps              %f", hparams.eps);
+    }
+
+    vision_model.class_embedding = mllama_tensor_load(new_mllama->ctx_data, "v.class_embd", true);
+    vision_model.patch_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.patch_embd.weight", true);
+
+    vision_model.position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.position_embd.weight", true);
+    vision_model.position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.position_embd.gate", true);
+
+    vision_model.pre_ln_w = mllama_tensor_load(new_mllama->ctx_data, "v.pre_ln.weight", true);
+    vision_model.pre_ln_b = mllama_tensor_load(new_mllama->ctx_data, "v.pre_ln.bias", true);
+    vision_model.post_ln_w = mllama_tensor_load(new_mllama->ctx_data, "v.post_ln.weight", true);
+    vision_model.post_ln_b = mllama_tensor_load(new_mllama->ctx_data, "v.post_ln.bias", true);
+
+    vision_model.tile_position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.tile_position_embd.weight", true);
+    vision_model.tile_position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.tile_position_embd.gate", true);
+
+    vision_model.pre_tile_position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.pre_tile_position_embd.weight", true);
+    vision_model.pre_tile_position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.pre_tile_position_embd.gate", true);
+
+    vision_model.post_tile_position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.post_tile_position_embd.weight", true);
+    vision_model.post_tile_position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.post_tile_position_embd.gate", true);
+
+    vision_model.mm_0_w = mllama_tensor_load(new_mllama->ctx_data, "mm.0.weight", false);
+    vision_model.mm_0_b = mllama_tensor_load(new_mllama->ctx_data, "mm.0.bias", false);
+
+    vision_model.layers = mllama_layers_load(new_mllama->ctx_data, "v", hparams.n_layer);
+    vision_model.global_layers = mllama_layers_load(new_mllama->ctx_data, "v.global", hparams.n_global_layer);
+
+    ggml_free(meta);
+
+    new_mllama->ctx_gguf = ctx;
+
+    {
+        // measure mem requirement and allocate
+        new_mllama->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead());
+        new_mllama->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_mllama->backend));
+        struct mllama_image_batch batch;
+        batch.size = 1;
+        ggml_cgraph *gf = mllama_image_build_graph(new_mllama, &batch);
+        ggml_gallocr_reserve(new_mllama->compute_alloc, gf);
+        size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_mllama->compute_alloc, 0);
+        LOG("compute allocated memory: %.2f MB", compute_memory_buffer_size / 1024.0 / 1024.0);
+    }
+
+    return new_mllama;
+}
+
+struct mllama_image *mllama_image_init() {
+    return new mllama_image();
+}
+
+void mllama_image_free(struct mllama_image *img) { delete img; }
+void mllama_image_batch_free(struct mllama_image_batch *batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
+    }
+}
+
+bool mllama_image_load_from_data(const void *data, const int n, const int width, const int height, const int num_channels, const int num_tiles, const int aspect_ratio_id, struct mllama_image *img) {
+    img->width = width;
+    img->height = height;
+    img->num_channels = num_channels;
+    img->num_tiles = num_tiles;
+    img->aspect_ratio_id = aspect_ratio_id;
+    img->data.resize(n);
+
+    memcpy(img->data.data(), data, n);
+    return true;
+}
+
+inline int mllama(int x, int lower, int upper) {
+    return std::max(lower, std::min(x, upper));
+}
+
+void mllama_free(mllama_ctx *ctx) {
+    ggml_free(ctx->ctx_data);
+    gguf_free(ctx->ctx_gguf);
+
+    ggml_backend_buffer_free(ctx->params_buffer);
+    ggml_backend_free(ctx->backend);
+    ggml_gallocr_free(ctx->compute_alloc);
+    delete ctx;
+}
+
+bool mllama_image_encode(struct mllama_ctx *ctx, const int n_threads, mllama_image *img, float *vec) {
+    mllama_image_batch imgs{};
+    imgs.size = 1;
+    imgs.data = img;
+    return mllama_image_batch_encode(ctx, n_threads, &imgs, vec);
+}
+
+bool mllama_image_batch_encode(mllama_ctx *ctx, const int n_threads, const mllama_image_batch *imgs, float *vec) {
+    int batch_size = imgs->size;
+    REQUIRE(batch_size == 1);
+
+    // build the inference graph
+    ggml_cgraph *gf = mllama_image_build_graph(ctx, imgs);
+    ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
+
+    // set inputs
+    const auto &model = ctx->vision_model;
+    const auto &hparams = model.hparams;
+
+    const int image_size = hparams.image_size;
+    int image_size_width = image_size;
+    int image_size_height = image_size;
+
+    const int patch_size = hparams.patch_size;
+    const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int num_positions = num_patches + (model.class_embedding == nullptr ? 0 : 1);
+
+    {
+        struct ggml_tensor *inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
+        ggml_backend_tensor_set(inp_raw, imgs->data[0].data.data(), 0, ggml_nbytes(inp_raw));
+    }
+
+    {
+        struct ggml_tensor *embeddings = ggml_graph_get_tensor(gf, "embeddings");
+        if (embeddings != nullptr) {
+            void *zeros = malloc(ggml_nbytes(embeddings));
+            memset(zeros, 0, ggml_nbytes(embeddings));
+            ggml_backend_tensor_set(embeddings, zeros, 0, ggml_nbytes(embeddings));
+            free(zeros);
+        }
+    }
+
+    {
+        struct ggml_tensor *positions = ggml_graph_get_tensor(gf, "positions");
+        if (positions != nullptr) {
+            int *positions_data = (int *)malloc(ggml_nbytes(positions));
+            for (int i = 0; i < num_positions; i++) {
+                positions_data[i] = i;
+            }
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
+        }
+    }
+
+    {
+        struct ggml_tensor *aspect_ratios = ggml_graph_get_tensor(gf, "aspect_ratios");
+        if (aspect_ratios != nullptr) {
+            int *aspect_ratios_data = (int *)malloc(ggml_nbytes(aspect_ratios));
+            aspect_ratios_data[0] = imgs->data[0].aspect_ratio_id;
+            ggml_backend_tensor_set(aspect_ratios, aspect_ratios_data, 0, ggml_nbytes(aspect_ratios));
+            free(aspect_ratios_data);
+        }
+    }
+
+    if (ggml_backend_is_cpu(ctx->backend)) {
+        ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
+    }
+
+    ggml_backend_graph_compute(ctx->backend, gf);
+
+    // the last node is the embedding tensor
+    struct ggml_tensor *embeddings = ggml_graph_node(gf, ggml_graph_n_nodes(gf) - 1);
+
+    // copy the embeddings to the location passed by the user
+    ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
+
+    return true;
+}
+
+int32_t mllama_image_size(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.image_size;
+}
+
+int32_t mllama_patch_size(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.patch_size;
+}
+
+int32_t mllama_hidden_size(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.hidden_size;
+}
+
+int mllama_n_patches(const struct mllama_ctx *ctx) {
+    const auto &hparams = ctx->vision_model.hparams;
+    return (hparams.image_size / hparams.patch_size) * (hparams.image_size / hparams.patch_size);
+}
+
+int mllama_n_positions(const struct mllama_ctx *ctx) {
+    return mllama_n_patches(ctx) + (ctx->vision_model.class_embedding == nullptr ? 0 : 1);
+}
+
+int mllama_n_tiles(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.n_tiles;
+}
+
+int mllama_n_embd(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.projection_dim;
+}
+
+size_t mllama_n_embd_bytes(const struct mllama_ctx *ctx) {
+    return mllama_n_positions(ctx) * mllama_n_embd(ctx) * mllama_n_tiles(ctx) * sizeof(float);
+}
--- a/llama/mllama.h
+++ b/llama/mllama.h
@ -0,0 +1,61 @@
+#ifndef MLLAMA_H
+#define MLLAMA_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef LLAMA_SHARED
+#if defined(_WIN32) && !defined(__MINGW32__)
+#ifdef LLAMA_BUILD
+#define MLLAMA_API __declspec(dllexport)
+#else
+#define MLLAMA_API __declspec(dllimport)
+#endif
+#else
+#define MLLAMA_API __attribute__((visibility("default")))
+#endif
+#else
+#define MLLAMA_API
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct mllama_ctx;
+
+struct mllama_image_batch {
+    struct mllama_image *data;
+    size_t size;
+};
+
+MLLAMA_API struct mllama_ctx *mllama_model_load(const char *fname, int verbosity);
+MLLAMA_API struct mllama_ctx *mllama_model_load_cpu(const char *fname, int verbosity);
+
+MLLAMA_API void mllama_free(struct mllama_ctx *ctx);
+
+MLLAMA_API int32_t mllama_image_size(const struct mllama_ctx *ctx);
+MLLAMA_API int32_t mllama_patch_size(const struct mllama_ctx *ctx);
+MLLAMA_API int32_t mllama_hidden_size(const struct mllama_ctx *ctx);
+
+MLLAMA_API int mllama_n_patches(const struct mllama_ctx *ctx);
+MLLAMA_API int mllama_n_positions(const struct mllama_ctx *ctx);
+MLLAMA_API int mllama_n_tiles(const struct mllama_ctx *ctx);
+MLLAMA_API int mllama_n_embd(const struct mllama_ctx *ctx);
+MLLAMA_API size_t mllama_n_embd_bytes(const struct mllama_ctx *ctx);
+
+MLLAMA_API struct mllama_image *mllama_image_init();
+
+MLLAMA_API void mllama_image_free(struct mllama_image *img);
+MLLAMA_API void mllama_image_batch_free(struct mllama_image_batch *batch);
+
+MLLAMA_API bool mllama_image_load_from_data(const void *data, const int n, const int nx, const int ny, const int nc, const int nt, const int aspect_ratio_id, struct mllama_image *img);
+
+MLLAMA_API bool mllama_image_encode(struct mllama_ctx *ctx, int n_threads, struct mllama_image *img, float *vec);
+MLLAMA_API bool mllama_image_batch_encode(struct mllama_ctx *ctx, int n_threads, const struct mllama_image_batch *imgs, float *vec);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // MLLAMA_H
--- a/llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch
+++ b/llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch
@ -24,7 +24,7 @@ problem.
 9 files changed, 21 insertions(+), 2 deletions(-)

 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index b30b4cb3..0ce73a99 100644
+index 273075f4..dd11f304 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
@@ -107,7 +107,6 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
@ -43,7 +43,7 @@ index b30b4cb3..0ce73a99 100644
 }
 
 static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-@@ -1871,6 +1871,11 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+@@ -1867,6 +1867,11 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
 
 static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_aligned_free(buffer->context, buffer->size);
@ -55,7 +55,7 @@ index b30b4cb3..0ce73a99 100644
 }
 
 static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
-@@ -1918,7 +1923,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
+@@ -1914,7 +1919,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
 };
 
 static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
@ -85,7 +85,7 @@ index e2617b06..242e50a7 100644
 
 /**
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index b4b85abc..cb0d8528 100644
+index 9fb2134f..04ce764e 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -534,6 +534,7 @@ struct ggml_backend_cuda_buffer_context {
@ -96,7 +96,7 @@ index b4b85abc..cb0d8528 100644
 }
 
 static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
-@@ -790,6 +791,7 @@ struct ggml_backend_cuda_split_buffer_context {
+@@ -789,6 +790,7 @@ struct ggml_backend_cuda_split_buffer_context {
 static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
     delete ctx;
@ -104,7 +104,7 @@ index b4b85abc..cb0d8528 100644
 }
 
 static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1067,6 +1069,7 @@ static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_
+@@ -1062,6 +1064,7 @@ static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_
 
 static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     CUDA_CHECK(cudaFreeHost(buffer->context));
@ -125,10 +125,10 @@ index 50579227..2799a0a5 100644
 
 static void * ggml_backend_kompute_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
-index 576f9581..1b56f858 100644
+index d92392ed..425524d0 100644
 --- a/ggml/src/ggml-metal/ggml-metal.m
 +++ b/ggml/src/ggml-metal/ggml-metal.m
-@@ -5214,6 +5214,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
+@@ -5077,6 +5077,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
     }
 
     free(ctx);
@ -149,10 +149,10 @@ index 05a2f4e6..392cc18d 100644
 
 static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
-index 4f0abb5a..de1ec184 100644
+index 140a775f..e33c4ba0 100644
 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp
 +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
-@@ -483,6 +483,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -477,6 +477,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
     GGML_ASSERT(status);
     delete ctx;
@ -161,10 +161,10 @@ index 4f0abb5a..de1ec184 100644
 
 static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
-index 0ea72994..ae3a3c33 100644
+index 66b6f2cc..e3e6deae 100644
 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp
 +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
-@@ -320,6 +320,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
+@@ -317,6 +317,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
     ggml_sycl_set_device(ctx->device);
 
     delete ctx;
@ -172,7 +172,7 @@ index 0ea72994..ae3a3c33 100644
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-@@ -765,6 +766,7 @@ struct ggml_backend_sycl_split_buffer_context {
+@@ -762,6 +763,7 @@ struct ggml_backend_sycl_split_buffer_context {
 static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
     delete ctx;
@ -180,7 +180,7 @@ index 0ea72994..ae3a3c33 100644
 }
 
 static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1099,6 +1101,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
+@@ -1096,6 +1098,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
 
 static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_sycl_host_free(buffer->context);
@ -189,10 +189,10 @@ index 0ea72994..ae3a3c33 100644
 
 static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index e2b357fd..68768029 100644
+index c0bdb9e1..03d03064 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -8962,6 +8962,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -8660,6 +8660,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
     ggml_vk_destroy_buffer(ctx->dev_buffer);
     delete ctx;
@ -200,7 +200,7 @@ index e2b357fd..68768029 100644
 }
 
 static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -9105,6 +9106,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
+@@ -8803,6 +8804,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
 static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     VK_LOG_MEMORY("ggml_backend_vk_host_buffer_free_buffer()");
     ggml_vk_host_free(vk_instance.devices[0], buffer->context);
--- a/llama/patches/0002-pretokenizer.patch
+++ b/llama/patches/0002-pretokenizer.patch
@ -10,10 +10,10 @@ logs instead of throwing an error
 1 file changed, 3 insertions(+), 11 deletions(-)

 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 9389ca80..806c1b3d 100644
+index 50ded286..a9ee9f03 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
-@@ -1503,16 +1503,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+@@ -1491,16 +1491,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
         if (type == LLAMA_VOCAB_TYPE_BPE) {
             add_space_prefix = false;
             clean_spaces = true;
@ -31,8 +31,8 @@ index 9389ca80..806c1b3d 100644
                 pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
             } else if (
                     tokenizer_pre == "llama3"   ||
-@@ -1651,7 +1642,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
-                 pre_type = LLAMA_VOCAB_PRE_TYPE_SEED_CODER;
+@@ -1635,7 +1626,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+                 pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE;
                 clean_spaces = false;
             } else {
 -                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
--- a/llama/patches/0003-embeddings.patch
+++ b/llama/patches/0003-embeddings.patch
@ -11,10 +11,10 @@ instead of forcing one or the error
 1 file changed, 3 insertions(+), 3 deletions(-)

 diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index 62246c10..dca22d8b 100644
+index 5a2eef9b..9c1fe93f 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
-@@ -901,7 +901,7 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1225,7 +1225,7 @@ int llama_context::decode(llama_batch & inp_batch) {
     int64_t n_outputs_all = 0;
 
     // count outputs
@ -23,7 +23,7 @@ index 62246c10..dca22d8b 100644
         for (uint32_t i = 0; i < n_tokens_all; ++i) {
             n_outputs_all += batch.logits[i] != 0;
         }
-@@ -982,7 +982,7 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1337,7 +1337,7 @@ int llama_context::decode(llama_batch & inp_batch) {
         //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
         //}
 
@ -32,7 +32,7 @@ index 62246c10..dca22d8b 100644
         auto * t_embd   = cparams.embeddings ? res->get_embd() : nullptr;
 
         if (t_embd && res->get_embd_pooled()) {
-@@ -1151,7 +1151,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
+@@ -1481,7 +1481,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto n_embd  = hparams.n_embd;
 
     // TODO: use a per-batch flag for logits presence instead
--- a/llama/patches/0004-clip-unicode.patch
+++ b/llama/patches/0004-clip-unicode.patch
@ -6,16 +6,16 @@ Subject: [PATCH] clip-unicode
 fixes loading vision models in llama.cpp on windows
 filesystems for paths that include wide characters
 ---
- tools/mtmd/clip.cpp | 39 +++++++++++++++++++++++++++++++++++++++
+ examples/llava/clip.cpp | 39 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 39 insertions(+)

-diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
-index 41ba45a7..cdd8ca44 100644
--- a/tools/mtmd/clip.cpp
-+++ b/tools/mtmd/clip.cpp
-@@ -31,6 +31,19 @@
+diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
+index ad3e7df1..b3218c78 100644
+--- a/examples/llava/clip.cpp
+++ b/examples/llava/clip.cpp
+@@ -30,6 +30,19 @@
+ #include <array>
 #include <numeric>
- #include <functional>
 
 +#if defined(_WIN32)
 +#define WIN32_LEAN_AND_MEAN
@ -32,8 +32,8 @@ index 41ba45a7..cdd8ca44 100644
 +
 struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL};
 
- enum ffn_op_type {
-@@ -2190,7 +2203,29 @@ struct clip_model_loader {
+ //#define CLIP_DEBUG_FUNCTIONS
+@@ -1971,7 +1984,29 @@ struct clip_model_loader {
         {
             std::vector<uint8_t> read_buf;
 
@ -63,7 +63,7 @@ index 41ba45a7..cdd8ca44 100644
             if (!fin) {
                 throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
             }
-@@ -2217,7 +2252,11 @@ struct clip_model_loader {
+@@ -1998,7 +2033,11 @@ struct clip_model_loader {
                     ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
                 }
             }
--- a/llama/patches/0005-solar-pro.patch
+++ b/llama/patches/0005-solar-pro.patch
@ -138,7 +138,7 @@ index 7ee6a5b7..48dce407 100644
 };
 
 diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
-index 4cce5166..7f6617fa 100644
+index ea73a8a7..a012aeae 100644
 --- a/src/llama-model-loader.cpp
 +++ b/src/llama-model-loader.cpp
@@ -439,6 +439,7 @@ namespace GGUFMeta {
@ -150,10 +150,10 @@ index 4cce5166..7f6617fa 100644
 llama_model_loader::llama_model_loader(
         const std::string & fname,
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index 3a4e72a3..831b68c0 100644
+index 822e2bb2..572378c9 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1402,6 +1402,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -1386,6 +1386,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                }
             } break;
@ -175,7 +175,7 @@ index 3a4e72a3..831b68c0 100644
         case LLM_ARCH_WAVTOKENIZER_DEC:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
-@@ -3774,6 +3789,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -3741,6 +3756,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
@ -210,7 +210,7 @@ index 3a4e72a3..831b68c0 100644
                         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);
-@@ -12397,6 +12440,165 @@ struct llm_build_chameleon : public llm_graph_context {
+@@ -12342,6 +12385,165 @@ struct llm_build_chameleon : public llm_graph_context {
     }
 };
 
@ -270,7 +270,7 @@ index 3a4e72a3..831b68c0 100644
 +            // self-attention
 +            {
 +                // rope freq factors for llama3; may return nullptr for llama2 and other models
-+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                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);
@ -376,7 +376,7 @@ index 3a4e72a3..831b68c0 100644
 struct llm_build_wavtokenizer_dec : public llm_graph_context {
     llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         ggml_tensor * cur;
-@@ -13157,6 +13359,10 @@ llm_graph_result_ptr llama_model::build_graph(
+@@ -13092,6 +13294,10 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_chameleon>(*this, params, gf);
             } break;
@ -387,7 +387,7 @@ index 3a4e72a3..831b68c0 100644
         case LLM_ARCH_WAVTOKENIZER_DEC:
             {
                 llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params, gf);
-@@ -13301,6 +13507,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+@@ -13238,6 +13444,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
         case LLM_ARCH_CHAMELEON:
@ -396,10 +396,10 @@ index 3a4e72a3..831b68c0 100644
             return LLAMA_ROPE_TYPE_NORM;
 
 diff --git a/src/llama-model.h b/src/llama-model.h
-index 6bdec263..43746c7d 100644
+index 95eca002..856e6042 100644
 --- a/src/llama-model.h
 +++ b/src/llama-model.h
-@@ -65,6 +65,7 @@ enum llm_type {
+@@ -64,6 +64,7 @@ enum llm_type {
     LLM_TYPE_15B,
     LLM_TYPE_16B,
     LLM_TYPE_20B,
@ -407,7 +407,7 @@ index 6bdec263..43746c7d 100644
     LLM_TYPE_27B,
     LLM_TYPE_30B,
     LLM_TYPE_32B,
-@@ -315,6 +316,8 @@ struct llama_layer {
+@@ -311,6 +312,8 @@ struct llama_layer {
     struct ggml_tensor * ffn_up_scale   = nullptr;
     struct ggml_tensor * ffn_down_scale = nullptr;
 
--- a/llama/patches/0006-add-mllama-support.patch
+++ b/llama/patches/0006-add-mllama-support.patch
--- a/llama/patches/0007-add-unpad-operator.patch
+++ b/llama/patches/0007-add-unpad-operator.patch
@ -0,0 +1,419 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: jmorganca <jmorganca@gmail.com>
+Date: Sun, 13 Apr 2025 22:10:06 -0400
+Subject: [PATCH] add unpad operator
+
+adds the unpad operator to GGML
+---
+ ggml/include/ggml.h                  | 10 +++++
+ ggml/src/ggml-cpu/ggml-cpu.c         |  5 +++
+ ggml/src/ggml-cpu/ops.cpp            | 55 ++++++++++++++++++++++++++++
+ ggml/src/ggml-cpu/ops.h              |  1 +
+ ggml/src/ggml-cuda/ggml-cuda.cu      |  4 ++
+ ggml/src/ggml-cuda/pad.cu            | 46 +++++++++++++++++++++++
+ ggml/src/ggml-cuda/pad.cuh           |  1 +
+ ggml/src/ggml-metal/ggml-metal.m     | 33 +++++++++++++++++
+ ggml/src/ggml-metal/ggml-metal.metal | 45 +++++++++++++++++++++++
+ ggml/src/ggml.c                      | 25 ++++++++++++-
+ 10 files changed, 223 insertions(+), 2 deletions(-)
+
+diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
+index 1b8603e7..53ef31b2 100644
+--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
+@@ -489,6 +489,7 @@ extern "C" {
+         GGML_OP_UPSCALE, // nearest interpolate
+         GGML_OP_PAD,
+         GGML_OP_PAD_REFLECT_1D,
+        GGML_OP_UNPAD,
+         GGML_OP_ARANGE,
+         GGML_OP_TIMESTEP_EMBEDDING,
+         GGML_OP_ARGSORT,
+@@ -1777,6 +1778,15 @@ extern "C" {
+             int                   p0,
+             int                   p1);
+ 
+    // unpad each dimension: [x, ..., x, y, ..., y] -> [x, ..., x]
+    GGML_API struct ggml_tensor * ggml_unpad(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                  p0,
+            int                  p1,
+            int                  p2,
+            int                  p3);
+
+     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
+     // timesteps: [N,]
+     // return: [N, dim]
+diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
+index 64405449..34624cca 100644
+--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
+@@ -1964,6 +1964,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
+             {
+                 ggml_compute_forward_pad_reflect_1d(params, tensor);
+             } break;
+        case GGML_OP_UNPAD:
+            {
+                ggml_compute_forward_unpad(params, tensor);
+            } break;
+         case GGML_OP_ARANGE:
+             {
+                 ggml_compute_forward_arange(params, tensor);
+@@ -2287,6 +2291,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
+         case GGML_OP_UPSCALE:
+         case GGML_OP_PAD:
+         case GGML_OP_PAD_REFLECT_1D:
+        case GGML_OP_UNPAD:
+         case GGML_OP_ARANGE:
+         case GGML_OP_TIMESTEP_EMBEDDING:
+         case GGML_OP_ARGSORT:
+diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
+index 7413192b..becdae07 100644
+--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
+@@ -6703,6 +6703,61 @@ void ggml_compute_forward_pad_reflect_1d(
+     }
+ }
+ 
+// ggml_compute_forward_unpad
+
+static void ggml_compute_forward_unpad_f32(
+    const struct ggml_compute_params *params,
+    struct ggml_tensor *dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT( dst->nb[0] == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    float * dst_ptr = (float *) dst->data;
+
+    // TODO: optimize
+
+    for (int64_t i2 = 0; i2 < ne2; ++i2) {
+        for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
+            for (int64_t i0 = 0; i0 < ne0; ++i0) {
+                for (int64_t i3 = 0; i3 < ne3; ++i3) {
+                    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+
+                    const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+
+                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                        dst_ptr[dst_idx] = *src_ptr;
+                    }
+                }
+            }
+        }
+    }
+}
+
+void ggml_compute_forward_unpad(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_unpad_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+ // ggml_compute_forward_arange
+ 
+ static void ggml_compute_forward_arange_f32(
+diff --git a/ggml/src/ggml-cpu/ops.h b/ggml/src/ggml-cpu/ops.h
+index dc081b9e..a7125555 100644
+--- a/ggml/src/ggml-cpu/ops.h
+++ b/ggml/src/ggml-cpu/ops.h
+@@ -72,6 +72,7 @@ void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params
+ void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+ void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+ void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_unpad(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+ void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+ void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+ void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
+index 04ce764e..491acccb 100644
+--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
+@@ -2223,6 +2223,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
+         case GGML_OP_PAD:
+             ggml_cuda_op_pad(ctx, dst);
+             break;
+        case GGML_OP_UNPAD:
+            ggml_cuda_op_unpad(ctx, dst);
+            break;
+         case GGML_OP_ARANGE:
+             ggml_cuda_op_arange(ctx, dst);
+             break;
+@@ -3197,6 +3200,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
+         case GGML_OP_UPSCALE:
+             return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
+         case GGML_OP_PAD:
+        case GGML_OP_UNPAD:
+         case GGML_OP_ARANGE:
+         case GGML_OP_TIMESTEP_EMBEDDING:
+         case GGML_OP_LEAKY_RELU:
+diff --git a/ggml/src/ggml-cuda/pad.cu b/ggml/src/ggml-cuda/pad.cu
+index 77432b04..7d45a7e1 100644
+--- a/ggml/src/ggml-cuda/pad.cu
+++ b/ggml/src/ggml-cuda/pad.cu
+@@ -47,3 +47,49 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+         src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+ }
+
+static __global__ void unpad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) {
+    // blockIdx.z: idx of ne2*ne3, aka ne02*ne03
+    // blockIdx.y: idx of ne1
+    // blockIDx.x: idx of ne0 / BLOCK_SIZE
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne00 +
+            blockIdx.z * ne00 * ne01;
+        dst[offset_dst] = x[offset_src];
+    }
+}
+
+static void unpad_f32_cuda(const float * x, float * dst,
+    const int ne00, const int ne01, const int ne02, const int ne03,
+    const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2*ne3);
+    unpad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
+}
+
+void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    unpad_f32_cuda(src0_d, dst_d,
+        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+}
+\ No newline at end of file
+diff --git a/ggml/src/ggml-cuda/pad.cuh b/ggml/src/ggml-cuda/pad.cuh
+index 8fd386b0..e2ededc3 100644
+--- a/ggml/src/ggml-cuda/pad.cuh
+++ b/ggml/src/ggml-cuda/pad.cuh
+@@ -3,3 +3,4 @@
+ #define CUDA_PAD_BLOCK_SIZE 256
+ 
+ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
+index 425524d0..112abef6 100644
+--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
+@@ -341,6 +341,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
+     GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
+     GGML_METAL_KERNEL_TYPE_PAD_F32,
+     GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32,
+    GGML_METAL_KERNEL_TYPE_UNPAD_F32,
+     GGML_METAL_KERNEL_TYPE_ARANGE_F32,
+     GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
+     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
+@@ -1277,6 +1278,7 @@ @implementation GGMLMetalClass
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                     upscale_f32,                     true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                         pad_f32,                         true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32,              pad_reflect_1d_f32,              true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UNPAD_F32,                       unpad_f32,                       true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,          timestep_embedding_f32,          true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                      arange_f32,                      true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,             argsort_f32_i32_asc,             true);
+@@ -1647,6 +1649,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
+         case GGML_OP_POOL_2D:
+         case GGML_OP_PAD:
+         case GGML_OP_PAD_REFLECT_1D:
+        case GGML_OP_UNPAD:
+         case GGML_OP_TIMESTEP_EMBEDDING:
+         case GGML_OP_ARGSORT:
+         case GGML_OP_LEAKY_RELU:
+@@ -4047,6 +4050,36 @@ static bool ggml_metal_encode_node(
+ 
+                 const int nth = MIN(1024, ne0);
+ 
+                [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_UNPAD:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UNPAD_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
+
+                const int nth = MIN(1024, ne0);
+
+                 [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+             } break;
+         case GGML_OP_ARANGE:
+diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
+index 9f4147e9..6ceb3cef 100644
+--- a/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
+@@ -2975,6 +2975,51 @@ kernel void kernel_pad_reflect_1d_f32(
+     }
+ }
+ 
+kernel void kernel_unpad_f32(
+    device  const char * src0,
+    device        char * dst,
+    constant   int64_t & ne00,
+    constant   int64_t & ne01,
+    constant   int64_t & ne02,
+    constant   int64_t & ne03,
+    constant  uint64_t & nb00,
+    constant  uint64_t & nb01,
+    constant  uint64_t & nb02,
+    constant  uint64_t & nb03,
+    constant   int64_t & ne0,
+    constant   int64_t & ne1,
+    constant   int64_t & ne2,
+    constant   int64_t & ne3,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    constant  uint64_t & nb2,
+    constant  uint64_t & nb3,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
+
+    if (i1 < ne01 && i2 < ne02 && i3 < ne03) {
+        for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+            if (i0 < ne00) {
+                dst_ptr[i0] = src0_ptr[i0];
+            }
+        }
+
+        return;
+    }
+}
+
+ kernel void kernel_arange_f32(
+     device        char * dst,
+     constant   ggml_metal_kargs_arange & args,
+diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
+index 7654ae17..3c57aff8 100644
+--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
+@@ -923,6 +923,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
+     "UPSCALE",
+     "PAD",
+     "PAD_REFLECT_1D",
+    "UNPAD",
+     "ARANGE",
+     "TIMESTEP_EMBEDDING",
+     "ARGSORT",
+@@ -953,7 +954,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
+     "OPT_STEP_ADAMW",
+ };
+ 
+-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
+ 
+ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+     "none",
+@@ -1018,6 +1019,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+     "upscale(x)",
+     "pad(x)",
+     "pad_reflect_1d(x)",
+    "unpad(x)",
+     "arange(start, stop, step)",
+     "timestep_embedding(timesteps, dim, max_period)",
+     "argsort(x)",
+@@ -1048,7 +1050,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+     "adamw(x)",
+ };
+ 
+-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
+ 
+ static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
+ 
+@@ -4270,6 +4272,25 @@ struct ggml_tensor * ggml_pad_reflect_1d(
+     return result;
+ }
+ 
+// ggml_unpad
+
+struct ggml_tensor * ggml_unpad(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    int p0, int p1, int p2, int p3) {
+
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
+            a->ne[0] - p0,
+            a->ne[1] - p1,
+            a->ne[2] - p2,
+            a->ne[3] - p3);
+
+    result->op = GGML_OP_UNPAD;
+    result->src[0] = a;
+
+    return result;
+}
+
+ // ggml_arange
+ 
+ struct ggml_tensor * ggml_arange(
--- a/llama/patches/0008-ensure-KV-cache-is-fully-defragmented.patch
+++ b/llama/patches/0008-ensure-KV-cache-is-fully-defragmented.patch
@ -1,352 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: jmorganca <jmorganca@gmail.com>
-Date: Tue, 15 Apr 2025 14:27:40 -0400
-Subject: [PATCH] ensure KV cache is fully defragmented
-
-Sometimes the KV cache requires defragmentation even without
-triggering the threshold heuristic. In this case, decoding
-will not being able to find a KV cache slot. This is particularly
-difficult for the caller to handle if it happens in between
-ubatches. To avoid this, we should immediately trigger a defrag.
-
-In addition, a heavily fragmented cache can require more than
-max_moves to defragment. Currently, we stop when we hit the limit
-but this can leave a cache that still does not have adequate space
-even after defragmentation is triggered. Instead, we should do
-multiple batches of processing until everything is complete.
---
- src/llama-context.cpp  |  18 ++++---
- src/llama-context.h    |   1 +
- src/llama-kv-cache.cpp | 107 ++++++++++++++---------------------------
- src/llama-kv-cache.h   |  12 ++++-
- 4 files changed, 59 insertions(+), 79 deletions(-)
-
-diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index c22687e4..c5948e8f 100644
--- a/src/llama-context.cpp
-+++ b/src/llama-context.cpp
-@@ -950,9 +950,12 @@ int llama_context::decode(llama_batch & inp_batch) {
- 
-         // find KV slot
-         if (!kv_self->find_slot(ubatch)) {
-            LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
-
-            return 1;
-+            kv_self->defrag_sched(-1.0f);
-+            kv_self->update(*this);
-+            if (!kv_self->find_slot(ubatch)) {
-+                LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
-+                return 1;
-+            }
-         }
- 
-         ggml_backend_sched_reset(sched.get());
-@@ -1967,9 +1970,12 @@ void llama_context::opt_epoch_iter(
- 
-             // TODO: not sure if this is needed
-             if (!kv_self->find_slot(ubatch)) {
-                LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
-
-                GGML_ABORT("TODO: handle this error");
-+                kv_self->defrag_sched(-1.0f);
-+                kv_self->update(*this);
-+                if (!kv_self->find_slot(ubatch)) {
-+                    LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
-+                    GGML_ABORT("TODO: handle this error");
-+                }
-             }
- 
-             auto * gf = graph_init();
-diff --git a/src/llama-context.h b/src/llama-context.h
-index c0ceacb1..0264e937 100644
--- a/src/llama-context.h
-+++ b/src/llama-context.h
-@@ -5,6 +5,7 @@
- #include "llama-cparams.h"
- #include "llama-graph.h"
- #include "llama-adapter.h"
-+#include "llama-kv-cache.h"
- 
- #include "ggml-cpp.h"
- #include "ggml-opt.h"
-diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
-index 3dcad65b..60e67b03 100644
--- a/src/llama-kv-cache.cpp
-+++ b/src/llama-kv-cache.cpp
-@@ -364,8 +364,6 @@ void llama_kv_cache_unified::commit() {
- }
- 
- bool llama_kv_cache_unified::update(llama_context & lctx) {
-    bool need_reserve = false;
-
-     auto * sched = lctx.get_sched();
- 
-     if (has_shift) {
-@@ -388,8 +386,6 @@ bool llama_kv_cache_unified::update(llama_context & lctx) {
-             res->set_inputs(nullptr);
- 
-             lctx.graph_compute(gf, false);
-
-            need_reserve = true;
-         }
- 
-         {
-@@ -403,27 +399,36 @@ bool llama_kv_cache_unified::update(llama_context & lctx) {
- 
-     if (do_defrag) {
-         LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__);
-+        const uint32_t n_max_nodes = lctx.graph_max_nodes();
-+        const uint32_t max_moves = (n_max_nodes - 2*model.hparams.n_layer)/(6*model.hparams.n_layer);
-+        if (!defrag_prepare(n_max_nodes)) {
-+            LLAMA_LOG_ERROR("%s: failed to prepare defragmentation\n", __func__);
-+            return false;
-+        }
-+
-+        for (std::size_t i = 0; i < defrag_info.moves.size(); i += max_moves) {
-+            std::vector<struct llama_kv_defrag_move> chunk;
-+            auto end = std::min(i + max_moves, defrag_info.moves.size());
-+            chunk.assign(defrag_info.moves.begin() + i, defrag_info.moves.begin() + end);
- 
-        if (defrag_prepare(lctx.graph_max_nodes())) {
-             ggml_backend_sched_reset(sched);
- 
-             auto * gf = lctx.graph_init();
- 
-            auto res = build_graph_defrag(lctx.get_cparams(), lctx.get_ctx_compute(), gf);
-+            auto res = build_graph_defrag(lctx.get_cparams(), lctx.get_ctx_compute(), gf, chunk);
- 
-             ggml_backend_sched_alloc_graph(sched, gf);
- 
-             res->set_inputs(nullptr);
- 
-             lctx.graph_compute(gf, false);
-
-            need_reserve = true;
-         }
- 
-         do_defrag = false;
-     }
- 
-    return need_reserve;
-+    // we never need to reserve a worst case graph
-+    return false;
- }
- 
- void llama_kv_cache_unified::defrag_sched(float thold) {
-@@ -707,11 +712,10 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_shift(
- llm_graph_result_ptr llama_kv_cache_unified::build_graph_defrag(
-         const llama_cparams & cparams,
-                ggml_context * ctx,
-                ggml_cgraph * gf) const {
-+                ggml_cgraph * gf,
-+                const std::vector<struct llama_kv_defrag_move> & moves) const {
-     auto res = std::make_unique<llm_graph_result>();
- 
-    const auto & ids = defrag_info.ids;
-
- #if 0
-     // CPU defrag
-     //
-@@ -783,32 +787,20 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_defrag(
-         ggml_backend_tensor_set(v_l[il], buf_v.data(), 0, buf_v.size());
-     }
- #else
-    for (uint32_t i = 0; i < ids.size(); ++i) {
-        const uint32_t id = ids[i];
-
-        if (i == id || id == ids.size()) {
-            continue;
-        }
-
-        uint32_t nm = 1;
-
-        while (i + nm < ids.size() && ids[i + nm] == id + nm) {
-            nm++;
-        }
-
-+    for (const auto & move : moves) {
-         for (uint32_t il = 0; il < hparams.n_layer; ++il) { // NOLINT
-             const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
-             const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
- 
-             ggml_tensor * view_k_src = ggml_view_2d(ctx, k_l[il],
-                    n_embd_k_gqa, nm,
-+                    n_embd_k_gqa, move.len,
-                     ggml_row_size(k_l[il]->type, n_embd_k_gqa),
-                    ggml_row_size(k_l[il]->type, n_embd_k_gqa*i));
-+                    ggml_row_size(k_l[il]->type, n_embd_k_gqa*move.src));
- 
-             ggml_tensor * view_k_dst = ggml_view_2d(ctx, k_l[il],
-                    n_embd_k_gqa, nm,
-+                    n_embd_k_gqa, move.len,
-                     ggml_row_size(k_l[il]->type, n_embd_k_gqa),
-                    ggml_row_size(k_l[il]->type, n_embd_k_gqa*id));
-+                    ggml_row_size(k_l[il]->type, n_embd_k_gqa*move.dst));
- 
-             ggml_tensor * view_v_src;
-             ggml_tensor * view_v_dst;
-@@ -816,31 +808,29 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_defrag(
-             if (cparams.flash_attn) {
-                 // NOTE: the V cache is not transposed when using flash attention
-                 view_v_src = ggml_view_2d(ctx, v_l[il],
-                        n_embd_v_gqa, nm,
-+                        n_embd_v_gqa, move.len,
-                         ggml_row_size(v_l[il]->type, n_embd_v_gqa),
-                        ggml_row_size(v_l[il]->type, n_embd_v_gqa*i));
-+                        ggml_row_size(v_l[il]->type, n_embd_v_gqa*move.dst));
- 
-                 view_v_dst = ggml_view_2d(ctx, v_l[il],
-                        n_embd_v_gqa, nm,
-+                        move.len, n_embd_v_gqa,
-                         ggml_row_size(v_l[il]->type, n_embd_v_gqa),
-                        ggml_row_size(v_l[il]->type, n_embd_v_gqa*id));
-+                        ggml_row_size(v_l[il]->type, move.src));
-             } else {
-                 view_v_src = ggml_view_2d(ctx, v_l[il],
-                        nm, n_embd_v_gqa,
-+                        move.len, n_embd_v_gqa,
-                         ggml_row_size(v_l[il]->type, size),
-                        ggml_row_size(v_l[il]->type, i));
-+                        ggml_row_size(v_l[il]->type, move.src));
- 
-                 view_v_dst = ggml_view_2d(ctx, v_l[il],
-                        nm, n_embd_v_gqa,
-+                        move.len, n_embd_v_gqa,
-                         ggml_row_size(v_l[il]->type, size),
-                        ggml_row_size(v_l[il]->type, id));
-+                        ggml_row_size(v_l[il]->type, move.dst));
-             }
- 
-             ggml_build_forward_expand(gf, ggml_cpy(ctx, view_k_src, view_k_dst));
-             ggml_build_forward_expand(gf, ggml_cpy(ctx, view_v_src, view_v_dst));
-         }
-
-        i += nm - 1;
-     }
- 
-     //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes);
-@@ -857,17 +847,7 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
- 
-     assert(n_used <= n_kv);
- 
-    //const int64_t t_start = ggml_time_us();
-
-    // number of cells moved
-    uint32_t n_moves = 0;
-
-    // each move requires 6*n_layer tensors (see graph_build_kv_self_defrag)
-    //   - source view, destination view, copy operation
-    //   - x2 for keys and values
-    //const uint32_t max_moves = max_nodes()/(6*n_layer);
-    // TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516
-    const uint32_t max_moves = (n_max_nodes - 2*n_layer)/(6*n_layer);
-+    defrag_info.moves.clear();
- 
-     // determine which KV cells to move where
-     //
-@@ -875,10 +855,7 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
-     //
-     //  if ids[i] == i || ids[i] == n_kv, then cell i is not moved
-     //
-    auto & ids = defrag_info.ids;
-
-    ids.clear();
-    ids.resize(n_kv, n_kv);
-+    std::vector<uint32_t> ids(n_kv, n_kv);
- 
-     for (uint32_t i0 = 0; i0 < n_used; ++i0) {
-         const auto & cell0 = cells[i0];
-@@ -927,19 +904,11 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
-         // are we moving a continuous block of memory?
-         bool cont = false;
- 
-        // should we stop searching for the next move?
-        bool stop = false;
-
-         // go back and move the nf cells to the hole
-         for (; i1 < n_kv; ++i1) {
-             auto & cell1 = cells[i1];
- 
-             if (cell1.is_empty() || ids[i1] != n_kv) {
-                if (n_moves == max_moves) {
-                    stop = true;
-                    break;
-                }
-
-                 cont = false;
-                 continue;
-             }
-@@ -955,8 +924,10 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
-             head = n_used;
- 
-             if (!cont) {
-                n_moves++;
-+                defrag_info.moves.push_back({i1, i0 + nf, 1});
-                 cont = true;
-+            } else {
-+                defrag_info.moves.back().len++;
-             }
- 
-             nf++;
-@@ -966,22 +937,16 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
-             }
-         }
- 
-        if (stop || n_moves == max_moves) {
-            break;
-        }
-
-         //LLAMA_LOG_INFO("(tmp log) KV defrag: move [%u, %u) to [%u, %u)\n", is, i1 + 1, i0, i0 + nh);
- 
-         i0 += nh - 1;
-     }
- 
-    if (n_moves == 0) {
-+    if (defrag_info.moves.size() == 0) {
-         return false;
-     }
- 
-    LLAMA_LOG_DEBUG("%s: (tmp log) KV defrag cell moves: %u\n", __func__, n_moves);
-
-    LLAMA_LOG_DEBUG("%s: expected gf nodes: %u\n", __func__, 6*n_moves*n_layer);
-+    // LLAMA_LOG_DEBUG("(tmp log) KV defrag cell moves: %u\n", n_moves);
- 
-     return true;
- }
-diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h
-index bf3b4b6a..928b9712 100644
--- a/src/llama-kv-cache.h
-+++ b/src/llama-kv-cache.h
-@@ -82,6 +82,13 @@ 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;
-+    uint32_t dst;
-+    uint32_t len;
-+};
- 
- //
- // llama_kv_cache_unified
-@@ -207,7 +214,7 @@ private:
- 
-     // defrag
-     struct {
-        std::vector<uint32_t> ids;
-+        std::vector<llama_kv_defrag_move> moves;
-     } defrag_info;
- 
-     // return true if cells have been moved
-@@ -249,7 +256,8 @@ private:
-     llm_graph_result_ptr build_graph_defrag(
-             const llama_cparams & cparams,
-                    ggml_context * ctx,
-                    ggml_cgraph * gf) const;
-+                    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;
--- a/llama/patches/0008-fix-deepseek-deseret-regex.patch
+++ b/llama/patches/0008-fix-deepseek-deseret-regex.patch
@ -12,10 +12,10 @@ regex
 2 files changed, 22 insertions(+), 1 deletion(-)

 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 806c1b3d..10f34d33 100644
+index a9ee9f03..1306864e 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
-@@ -298,7 +298,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
+@@ -296,7 +296,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
             case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM:
                 regex_exprs = {
                     "[\r\n]",
--- a/llama/patches/0009-maintain-ordering-for-rules-for-grammar.patch
+++ b/llama/patches/0009-maintain-ordering-for-rules-for-grammar.patch
--- a/llama/patches/0010-ensure-KV-cache-is-fully-defragmented.patch
+++ b/llama/patches/0010-ensure-KV-cache-is-fully-defragmented.patch
@ -0,0 +1,361 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: jmorganca <jmorganca@gmail.com>
+Date: Tue, 15 Apr 2025 14:27:40 -0400
+Subject: [PATCH] ensure KV cache is fully defragmented
+
+Sometimes the KV cache requires defragmentation even without
+triggering the threshold heuristic. In this case, decoding
+will not being able to find a KV cache slot. This is particularly
+difficult for the caller to handle if it happens in between
+ubatches. To avoid this, we should immediately trigger a defrag.
+
+In addition, a heavily fragmented cache can require more than
+max_moves to defragment. Currently, we stop when we hit the limit
+but this can leave a cache that still does not have adequate space
+even after defragmentation is triggered. Instead, we should do
+multiple batches of processing until everything is complete.
+---
+ src/llama-context.cpp  | 105 +++++++++++++----------------------------
+ src/llama-context.h    |   4 +-
+ src/llama-kv-cache.cpp |  39 +++------------
+ src/llama-kv-cache.h   |   9 +++-
+ 4 files changed, 51 insertions(+), 106 deletions(-)
+
+diff --git a/src/llama-context.cpp b/src/llama-context.cpp
+index cd06ad91..77177c5e 100644
+--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
+@@ -583,13 +583,12 @@ llm_graph_result_ptr llama_context::build_kv_self_shift(
+ 
+ llm_graph_result_ptr llama_context::build_kv_self_defrag(
+         ggml_context * ctx0,
+-        ggml_cgraph * gf) const {
+        ggml_cgraph * gf,
+        const std::vector<struct llama_kv_defrag_move> & moves) const {
+     auto res = std::make_unique<llm_graph_result>();
+ 
+     const auto & hparams = model.hparams;
+ 
+-    const auto & ids = kv_self->defrag_info.ids;
+-
+ #if 0
+     // CPU defrag
+     //
+@@ -661,32 +660,20 @@ llm_graph_result_ptr llama_context::build_kv_self_defrag(
+         ggml_backend_tensor_set(v_l[il], buf_v.data(), 0, buf_v.size());
+     }
+ #else
+-    for (uint32_t i = 0; i < ids.size(); ++i) {
+-        const uint32_t id = ids[i];
+-
+-        if (i == id || id == ids.size()) {
+-            continue;
+-        }
+-
+-        uint32_t nm = 1;
+-
+-        while (i + nm < ids.size() && ids[i + nm] == id + nm) {
+-            nm++;
+-        }
+-
+    for (const auto & move : moves) {
+         for (uint32_t il = 0; il < hparams.n_layer; ++il) { // NOLINT
+             const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
+             const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
+ 
+             ggml_tensor * view_k_src = ggml_view_2d(ctx0, kv_self->k_l[il],
+-                    n_embd_k_gqa, nm,
+                    n_embd_k_gqa, move.len,
+                     ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
+-                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*i));
+                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*move.src));
+ 
+             ggml_tensor * view_k_dst = ggml_view_2d(ctx0, kv_self->k_l[il],
+-                    n_embd_k_gqa, nm,
+                    n_embd_k_gqa, move.len,
+                     ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
+-                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*id));
+                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*move.dst));
+ 
+             ggml_tensor * view_v_src;
+             ggml_tensor * view_v_dst;
+@@ -694,34 +681,30 @@ llm_graph_result_ptr llama_context::build_kv_self_defrag(
+             if (cparams.flash_attn) {
+                 // NOTE: the V cache is not transposed when using flash attention
+                 view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il],
+-                        n_embd_v_gqa, nm,
+                        n_embd_v_gqa, move.len,
+                         ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa),
+-                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*i));
+                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*move.src));
+ 
+                 view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il],
+-                        n_embd_v_gqa, nm,
+                        n_embd_v_gqa, move.len,
+                         ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa),
+-                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*id));
+                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*move.dst));
+             } else {
+                 view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il],
+-                        nm, n_embd_v_gqa,
+                        move.len, n_embd_v_gqa,
+                         ggml_row_size(kv_self->v_l[il]->type, kv_self->size),
+-                        ggml_row_size(kv_self->v_l[il]->type, i));
+                        ggml_row_size(kv_self->v_l[il]->type, move.src));
+ 
+                 view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il],
+-                        nm, n_embd_v_gqa,
+                        move.len, n_embd_v_gqa,
+                         ggml_row_size(kv_self->v_l[il]->type, kv_self->size),
+-                        ggml_row_size(kv_self->v_l[il]->type, id));
+                        ggml_row_size(kv_self->v_l[il]->type, move.dst));
+             }
+ 
+             ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_k_src, view_k_dst));
+             ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_v_src, view_v_dst));
+         }
+-
+-        i += nm - 1;
+     }
+-
+-    //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes);
+ #endif
+ 
+     return res;
+@@ -730,8 +713,6 @@ llm_graph_result_ptr llama_context::build_kv_self_defrag(
+ void llama_context::kv_self_update() {
+     auto & kv = kv_self;
+ 
+-    bool need_reserve = false;
+-
+     if (kv->has_shift) {
+         if (!kv->get_can_shift()) {
+             GGML_ABORT("The current context does not support K-shift");
+@@ -752,8 +733,6 @@ void llama_context::kv_self_update() {
+             res->set_inputs(nullptr);
+ 
+             graph_compute(gf, false);
+-
+-            need_reserve = true;
+         }
+ 
+         {
+@@ -768,49 +747,28 @@ void llama_context::kv_self_update() {
+     // defragment the KV cache if needed
+     if (kv->do_defrag) {
+         LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__);
+        const uint32_t n_max_nodes = graph_max_nodes();
+        const uint32_t max_moves = (n_max_nodes - 2*model.hparams.n_layer)/(6*model.hparams.n_layer);
+        if (!kv->defrag_prepare(n_max_nodes)) {
+            LLAMA_LOG_ERROR("%s: failed to prepare defragmentation\n", __func__);
+            return;
+        }
+ 
+-        if (kv->defrag_prepare(graph_max_nodes())) {
+-            ggml_backend_sched_reset(sched.get());
+        for (std::size_t i = 0; i < kv_self->defrag_info.moves.size(); i += max_moves) {
+            std::vector<struct llama_kv_defrag_move> chunk;
+            auto end = std::min(i + max_moves, kv_self->defrag_info.moves.size());
+            chunk.assign(kv_self->defrag_info.moves.begin() + i, kv_self->defrag_info.moves.begin() + end);
+ 
+            ggml_backend_sched_reset(sched.get());
+             auto * gf = graph_init();
+-
+-            auto res = build_kv_self_defrag(ctx_compute.get(), gf);
+-
+            auto res = build_kv_self_defrag(ctx_compute.get(), gf, chunk);
+             ggml_backend_sched_alloc_graph(sched.get(), gf);
+-
+             res->set_inputs(nullptr);
+-
+             graph_compute(gf, false);
+-
+-            need_reserve = true;
+         }
+ 
+         kv->do_defrag = false;
+     }
+-
+-    // reserve a worst case graph if needed
+-    if (need_reserve) {
+-        LLAMA_LOG_DEBUG("%s: reserving a worst case graph\n", __func__);
+-
+-        // build worst-case graph
+-        uint32_t n_seqs = 1; // TODO: worst-case number of sequences
+-        uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
+-
+-        // simulate full KV cache
+-        kv_self->n = kv_self->size;
+-
+-        llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
+-        llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};
+-
+-        auto * gf = graph_init();
+-        graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT);
+-
+-        // initialize scheduler with the worst-case graph
+-        ggml_backend_sched_reset(sched.get());
+-        if (!ggml_backend_sched_reserve(sched.get(), gf)) {
+-            LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
+-        }
+-    }
+ }
+ 
+ enum llama_pooling_type llama_context::pooling_type() const {
+@@ -1294,9 +1252,12 @@ int llama_context::decode(llama_batch & inp_batch) {
+         // find KV slot
+         {
+             if (!kv_self->find_slot(ubatch)) {
+-                LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
+-
+-                return 1;
+                kv_self->defrag();
+                kv_self_update();
+                if (!kv_self->find_slot(ubatch)) {
+                    LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
+                    return 1;
+                }
+             }
+ 
+             if (!kv_self->recurrent) {
+diff --git a/src/llama-context.h b/src/llama-context.h
+index a50c4afa..30f84bfd 100644
+--- a/src/llama-context.h
+++ b/src/llama-context.h
+@@ -5,6 +5,7 @@
+ #include "llama-cparams.h"
+ #include "llama-graph.h"
+ #include "llama-adapter.h"
+#include "llama-kv-cache.h"
+ 
+ #include "ggml-cpp.h"
+ 
+@@ -179,7 +180,8 @@ private:
+ 
+     llm_graph_result_ptr build_kv_self_defrag(
+             ggml_context * ctx0,
+-            ggml_cgraph * gf) const;
+            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);
+diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
+index 69f8d35a..35a750d3 100644
+--- a/src/llama-kv-cache.cpp
+++ b/src/llama-kv-cache.cpp
+@@ -781,17 +781,7 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
+ 
+     assert(n_used <= n_kv);
+ 
+-    //const int64_t t_start = ggml_time_us();
+-
+-    // number of cells moved
+-    uint32_t n_moves = 0;
+-
+-    // each move requires 6*n_layer tensors (see graph_build_kv_self_defrag)
+-    //   - source view, destination view, copy operation
+-    //   - x2 for keys and values
+-    //const uint32_t max_moves = max_nodes()/(6*n_layer);
+-    // TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516
+-    const uint32_t max_moves = (n_max_nodes - 2*n_layer)/(6*n_layer);
+    defrag_info.moves.clear();
+ 
+     // determine which KV cells to move where
+     //
+@@ -799,10 +789,7 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
+     //
+     //  if ids[i] == i || ids[i] == n_kv, then cell i is not moved
+     //
+-    auto & ids = defrag_info.ids;
+-
+-    ids.clear();
+-    ids.resize(n_kv, n_kv);
+    std::vector<uint32_t> ids(n_kv, n_kv);
+ 
+     for (uint32_t i0 = 0; i0 < n_used; ++i0) {
+         const auto & cell0 = cells[i0];
+@@ -851,19 +838,11 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
+         // are we moving a continuous block of memory?
+         bool cont = false;
+ 
+-        // should we stop searching for the next move?
+-        bool stop = false;
+-
+         // go back and move the nf cells to the hole
+         for (; i1 < n_kv; ++i1) {
+             auto & cell1 = cells[i1];
+ 
+             if (cell1.is_empty() || ids[i1] != n_kv) {
+-                if (n_moves == max_moves) {
+-                    stop = true;
+-                    break;
+-                }
+-
+                 cont = false;
+                 continue;
+             }
+@@ -879,8 +858,10 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
+             head = n_used;
+ 
+             if (!cont) {
+-                n_moves++;
+                defrag_info.moves.push_back({i1, i0 + nf, 1});
+                 cont = true;
+            } else {
+                defrag_info.moves.back().len++;
+             }
+ 
+             nf++;
+@@ -890,22 +871,16 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
+             }
+         }
+ 
+-        if (stop || n_moves == max_moves) {
+-            break;
+-        }
+-
+         //LLAMA_LOG_INFO("(tmp log) KV defrag: move [%u, %u) to [%u, %u)\n", is, i1 + 1, i0, i0 + nh);
+ 
+         i0 += nh - 1;
+     }
+ 
+-    if (n_moves == 0) {
+    if (defrag_info.moves.size() == 0) {
+         return false;
+     }
+ 
+-    LLAMA_LOG_DEBUG("(tmp log) KV defrag cell moves: %u\n", n_moves);
+-
+-    LLAMA_LOG_DEBUG("expected gf nodes: %u\n", 6*n_moves*n_layer);
+    // LLAMA_LOG_DEBUG("(tmp log) KV defrag cell moves: %u\n", n_moves);
+ 
+     return true;
+ }
+diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h
+index 56c74035..25cbcb56 100644
+--- a/src/llama-kv-cache.h
+++ b/src/llama-kv-cache.h
+@@ -43,6 +43,13 @@ private:
+     llama_kv_cache * kv;
+ };
+ 
+// block of KV slots to move when defragging
+struct llama_kv_defrag_move {
+    uint32_t src;
+    uint32_t dst;
+    uint32_t len;
+};
+
+ struct llama_kv_cell {
+     llama_pos pos   = -1;
+     llama_pos delta =  0;
+@@ -131,7 +138,7 @@ public:
+     // defrag
+ 
+     struct {
+-        std::vector<uint32_t> ids;
+        std::vector<llama_kv_defrag_move> moves;
+     } defrag_info;
+ 
+     // return true if cells have been moved
--- a/llama/patches/0011-sort-devices-by-score.patch
+++ b/llama/patches/0011-sort-devices-by-score.patch
@ -11,7 +11,7 @@ with the fastest acceleration is loaded
 1 file changed, 13 insertions(+), 8 deletions(-)

 diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
-index 405d8e31..4e67d243 100644
+index 82ae1b5b..1487f322 100644
 --- a/ggml/src/ggml-backend-reg.cpp
 +++ b/ggml/src/ggml-backend-reg.cpp
@@ -157,7 +157,7 @@ struct ggml_backend_reg_entry {
--- a/llama/patches/0012-add-phony-target-ggml-cpu-for-all-cpu-variants.patch
+++ b/llama/patches/0012-add-phony-target-ggml-cpu-for-all-cpu-variants.patch
@ -8,7 +8,7 @@ Subject: [PATCH] add phony target ggml-cpu for all cpu variants
 1 file changed, 2 insertions(+)

 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index ddea5ad3..45918bf6 100644
+index 43d9fc4f..4c0d3824 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
@@ -279,6 +279,7 @@ function(ggml_add_cpu_backend_variant tag_name)
--- a/llama/patches/0013-remove-amx.patch
+++ b/llama/patches/0013-remove-amx.patch
@ -9,7 +9,7 @@ disable amx as it reduces performance on some systems
 1 file changed, 4 deletions(-)

 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index 45918bf6..0beaed86 100644
+index 4c0d3824..79c26312 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
@@ -296,10 +296,6 @@ if (GGML_CPU_ALL_VARIANTS)
--- a/llama/patches/0014-fix-string-arr-kv-loading.patch
+++ b/llama/patches/0014-fix-string-arr-kv-loading.patch
@ -9,8 +9,8 @@ such as vocab fields
 ---
 ggml/include/gguf.h | 1 +
 ggml/src/gguf.cpp   | 7 +++++--
- src/llama-vocab.cpp | 4 +---
- 3 files changed, 7 insertions(+), 5 deletions(-)
+ src/llama-vocab.cpp | 2 +-
+ 3 files changed, 7 insertions(+), 3 deletions(-)

 diff --git a/ggml/include/gguf.h b/ggml/include/gguf.h
 index 79ee2020..3efb22f0 100644
@ -53,17 +53,15 @@ index 381a9c7d..e45b453d 100644
 }
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 10f34d33..9f5fd57b 100644
+index 1306864e..d6515ff6 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
-@@ -1469,9 +1469,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+@@ -1459,7 +1459,7 @@ 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) {
-                 const gguf_type pc_type = gguf_get_arr_type(ctx, precompiled_charsmap_keyidx);
-                GGML_ASSERT(pc_type == GGUF_TYPE_INT8 || pc_type == GGUF_TYPE_UINT8);
-
-                const size_t n_precompiled_charsmap = gguf_get_arr_n(ctx, precompiled_charsmap_keyidx);
-+                const size_t n_precompiled_charsmap = gguf_get_arr_data_n(ctx, precompiled_charsmap_keyidx);
+-                size_t n_precompiled_charsmap = gguf_get_arr_n(ctx, precompiled_charsmap_keyidx);
+                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
--- a/llama/patches/0015-add-argsort-and-cuda-copy-for-i32.patch
+++ b/llama/patches/0015-add-argsort-and-cuda-copy-for-i32.patch
@ -1,277 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Michael Yang <git@mxy.ng>
-Date: Thu, 1 May 2025 13:45:12 -0700
-Subject: [PATCH] add argsort and cuda copy for i32
-
---
- ggml/src/ggml-cpu/ops.cpp     |  43 ++++++++++++++
- ggml/src/ggml-cuda/argsort.cu | 102 +++++++++++++++++++++++++++++++++-
- ggml/src/ggml-cuda/cpy.cu     |  49 ++++++++++++++++
- 3 files changed, 192 insertions(+), 2 deletions(-)
-
-diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
-index becdae07..7a44b6cf 100644
--- a/ggml/src/ggml-cpu/ops.cpp
-+++ b/ggml/src/ggml-cpu/ops.cpp
-@@ -6890,6 +6890,45 @@ static void ggml_compute_forward_argsort_f32(
-     }
- }
- 
-+static void ggml_compute_forward_argsort_i32(
-+    const ggml_compute_params * params,
-+    ggml_tensor * dst) {
-+
-+    const ggml_tensor * src0 = dst->src[0];
-+
-+    GGML_TENSOR_UNARY_OP_LOCALS
-+
-+    GGML_ASSERT(nb0 == sizeof(int32_t));
-+
-+    const int ith = params->ith;
-+    const int nth = params->nth;
-+
-+    const int64_t nr = ggml_nrows(src0);
-+
-+    ggml_sort_order order = (ggml_sort_order) ggml_get_op_params_i32(dst, 0);
-+
-+    for (int64_t i = ith; i < nr; i += nth) {
-+        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
-+        const int32_t * src_data = (int32_t *)((char *) src0->data + i*nb01);
-+
-+        for (int64_t j = 0; j < ne0; j++) {
-+            dst_data[j] = j;
-+        }
-+
-+        // C doesn't have a functional sort, so we do a bubble sort instead
-+        for (int64_t j = 0; j < ne0; j++) {
-+            for (int64_t k = j + 1; k < ne0; k++) {
-+                if ((order == GGML_SORT_ORDER_ASC  && src_data[dst_data[j]] > src_data[dst_data[k]]) ||
-+                    (order == GGML_SORT_ORDER_DESC && src_data[dst_data[j]] < src_data[dst_data[k]])) {
-+                    int32_t tmp = dst_data[j];
-+                    dst_data[j] = dst_data[k];
-+                    dst_data[k] = tmp;
-+                }
-+            }
-+        }
-+    }
-+}
-+
- void ggml_compute_forward_argsort(
-     const ggml_compute_params * params,
-     ggml_tensor * dst) {
-@@ -6901,6 +6940,10 @@ void ggml_compute_forward_argsort(
-             {
-                 ggml_compute_forward_argsort_f32(params, dst);
-             } break;
-+        case GGML_TYPE_I32:
-+            {
-+                ggml_compute_forward_argsort_i32(params, dst);
-+            } break;
-         default:
-             {
-                 GGML_ABORT("fatal error");
-diff --git a/ggml/src/ggml-cuda/argsort.cu b/ggml/src/ggml-cuda/argsort.cu
-index 607ded85..53b02634 100644
--- a/ggml/src/ggml-cuda/argsort.cu
-+++ b/ggml/src/ggml-cuda/argsort.cu
-@@ -85,13 +85,107 @@ static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, co
-     }
- }
- 
-+
-+template<ggml_sort_order order>
-+static __global__ void k_argsort_i32_i32(const int32_t * x, int * dst, const int ncols, const int ncols_pad) {
-+    extern __shared__ int shared_mem[];
-+    int * indices = shared_mem;
-+
-+    const int tid = threadIdx.x;
-+    const int row = blockIdx.y;
-+
-+    // Initialize all indices, handling the case where threads < ncols_pad
-+    for (int i = tid; i < ncols_pad; i += blockDim.x) {
-+        indices[i] = i < ncols ? i : 0; // Use 0 for padding indices
-+    }
-+    __syncthreads();
-+
-+    // Bitonic sort
-+    for (int k = 2; k <= ncols_pad; k *= 2) {
-+        for (int j = k/2; j > 0; j /= 2) {
-+            for (int i = tid; i < ncols_pad; i += blockDim.x) {
-+                const int ij = i ^ j;
-+                if (ij > i) {
-+                    // Only compare values within the actual data range
-+                    if (i < ncols && ij < ncols) {
-+                        if ((i & k) == 0) {
-+                            if (order == GGML_SORT_ORDER_ASC) {
-+                                if (x[row * ncols + indices[i]] > x[row * ncols + indices[ij]]) {
-+                                    int tmp = indices[i];
-+                                    indices[i] = indices[ij];
-+                                    indices[ij] = tmp;
-+                                }
-+                            } else {
-+                                if (x[row * ncols + indices[i]] < x[row * ncols + indices[ij]]) {
-+                                    int tmp = indices[i];
-+                                    indices[i] = indices[ij];
-+                                    indices[ij] = tmp;
-+                                }
-+                            }
-+                        } else {
-+                            if (order == GGML_SORT_ORDER_ASC) {
-+                                if (x[row * ncols + indices[i]] < x[row * ncols + indices[ij]]) {
-+                                    int tmp = indices[i];
-+                                    indices[i] = indices[ij];
-+                                    indices[ij] = tmp;
-+                                }
-+                            } else {
-+                                if (x[row * ncols + indices[i]] > x[row * ncols + indices[ij]]) {
-+                                    int tmp = indices[i];
-+                                    indices[i] = indices[ij];
-+                                    indices[ij] = tmp;
-+                                }
-+                            }
-+                        }
-+                    }
-+                }
-+            }
-+            __syncthreads();
-+        }
-+    }
-+
-+    // Write sorted indices to output, only threads handling valid data
-+    for (int i = tid; i < ncols; i += blockDim.x) {
-+        dst[row * ncols + i] = indices[i];
-+    }
-+}
-+
-+static void argsort_i32_i32_cuda(const int32_t * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
-+    // Bitonic sort requires ncols to be power of 2
-+    const int ncols_pad = next_power_of_2(ncols);
-+
-+    // Ensure thread count doesn't exceed maximum (typically 1024)
-+    const int max_threads = 1024;  // This is the typical max for most GPUs
-+    const int threads_per_block = ncols_pad > max_threads ? max_threads : ncols_pad;
-+
-+    const dim3 block_dims(threads_per_block, 1, 1);
-+    const dim3 block_nums(1, nrows, 1);
-+    const size_t shared_mem = ncols_pad * sizeof(int);
-+
-+    // Check if shared memory size is within limits
-+    const size_t max_shared_mem = ggml_cuda_info().devices[ggml_cuda_get_device()].smpb;
-+
-+    // Instead of logging an error, use GGML_ASSERT with a descriptive message
-+    GGML_ASSERT(shared_mem <= max_shared_mem && "argsort: required shared memory exceeds device limit");
-+
-+    // Launch kernels with the updated thread configuration
-+    if (order == GGML_SORT_ORDER_ASC) {
-+        k_argsort_i32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
-+    } else if (order == GGML_SORT_ORDER_DESC) {
-+        k_argsort_i32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
-+    } else {
-+        GGML_ABORT("fatal error");
-+    }
-+}
-+
-+
- void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-     const ggml_tensor * src0 = dst->src[0];
-     const float * src0_d = (const float *)src0->data;
-     float * dst_d = (float *)dst->data;
-     cudaStream_t stream = ctx.stream();
- 
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32);
-     GGML_ASSERT( dst->type == GGML_TYPE_I32);
-     GGML_ASSERT(ggml_is_contiguous(src0));
- 
-@@ -100,5 +194,9 @@ void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
- 
-     enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
- 
-    argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream);
-+    if (src0->type == GGML_TYPE_I32) {
-+        argsort_i32_i32_cuda((const int32_t *)src0_d, (int *)dst_d, ncols, nrows, order, stream);
-+    } else {
-+        argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream);
-+    }
- }
-diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu
-index 2d46176e..47383486 100644
--- a/ggml/src/ggml-cuda/cpy.cu
-+++ b/ggml/src/ggml-cuda/cpy.cu
-@@ -38,6 +38,13 @@ static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
-     *dsti = *xi;
- }
- 
-+static __device__ void cpy_1_i32_i32(const char * cxi, char * cdsti) {
-+    const int32_t * xi = (const int32_t *) cxi;
-+    int32_t * dsti = (int32_t *) cdsti;
-+
-+    *dsti = *xi;
-+}
-+
- template <cpy_kernel_t cpy_1>
- static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const int ne,
-                                    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-@@ -68,6 +75,44 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const in
-     cpy_1(cx + x_offset, cdst + dst_offset);
- }
- 
-+// First, add this template function after the other template functions
-+template <cpy_kernel_t cpy_1>
-+static __global__ void cpy_i32_i32(const char * cx, char * cdst, const int ne,
-+                                 const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-+                                 const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-+                                 const int nb12, const int nb13) {
-+    const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
-+
-+    if (i >= ne) {
-+        return;
-+    }
-+
-+    const int64_t i03 = i/(ne00 * ne01 * ne02);
-+    const int64_t i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
-+    const int64_t i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
-+    const int64_t i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
-+    const int64_t x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
-+
-+    const int64_t i13 = i/(ne10 * ne11 * ne12);
-+    const int64_t i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
-+    const int64_t i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
-+    const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
-+    const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13;
-+
-+    cpy_1(cx + x_offset, cdst + dst_offset);
-+}
-+
-+// Then modify the ggml_cpy_i32_i32_cuda function to use the new template
-+static void ggml_cpy_i32_i32_cuda(
-+    const char * cx, char * cdst, const int ne,
-+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int graph_cpynode_index) {
-+
-+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-+    cpy_i32_i32<cpy_1_i32_i32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
-+}
-+
- static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
-     const float * xi = (const float *) cxi;
-     block_q8_0 * dsti = (block_q8_0 *) cdsti;
-@@ -631,6 +676,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
-         ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
-     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-         ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
-+    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) {
-+        ggml_cpy_i32_i32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
-     } else {
-         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
-                 ggml_type_name(src0->type), ggml_type_name(src1->type));
-@@ -686,6 +733,8 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
-         return (void*) cpy_f32_f16<cpy_1_f32_f16>;
-     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-         return (void*) cpy_f32_f16<cpy_1_f16_f32>;
-+    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) {
-+        return (void*) cpy_i32_i32<cpy_1_i32_i32>;
-     } else {
-         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
-                 ggml_type_name(src0->type), ggml_type_name(src1->type));
--- a/llama/patches/0015-ollama-debug-tensor.patch
+++ b/llama/patches/0015-ollama-debug-tensor.patch
@ -8,7 +8,7 @@ Subject: [PATCH] ollama debug tensor
 1 file changed, 6 insertions(+)

 diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
-index a30e67f2..2462d2b8 100644
+index 34624cca..59bd3c62 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.c
 +++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -15,6 +15,8 @@
@ -20,7 +20,7 @@ index a30e67f2..2462d2b8 100644
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
 #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
-@@ -2841,6 +2843,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
+@@ -2859,6 +2861,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
 
         ggml_compute_forward(&params, node);
 
--- a/llama/patches/0016-add-ollama-vocab-for-grammar-support.patch
+++ b/llama/patches/0016-add-ollama-vocab-for-grammar-support.patch
@ -184,7 +184,7 @@ index f8c291de..2a3a62db 100644
                       const char * grammar_root,
                               bool lazy,
 diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp
-index 804b11e0..15a10ca8 100644
+index c0a5f934..75731053 100644
 --- a/src/llama-sampling.cpp
 +++ b/src/llama-sampling.cpp
@@ -1466,7 +1466,7 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
--- a/llama/patches/0017-ggml-Don-t-assert-fail-when-tensor-data-changes-1322.patch
+++ b/llama/patches/0017-ggml-Don-t-assert-fail-when-tensor-data-changes-1322.patch
@ -0,0 +1,38 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Jesse Gross <jesse@kernel.org>
+Date: Thu, 1 May 2025 13:46:10 -0700
+Subject: [PATCH] ggml: Don't assert fail when tensor data changes (#13222)
+
+The following scenario will cause an assertion failure in the graph
+allocator:
+ - Build and allocate a graph containing a tensor with a non-NULL data
+   pointer
+ - Build and allocate a new graph where that data is NULL
+
+Result:
+ggml-alloc.c:819: GGML_ASSERT(talloc->buffer_id >= 0) failed
+
+This happens during revalidation because we think that memory should
+have been previously allocated based on the current graph but in
+reality the previous graph was different. In this situation, we
+should do a full reallocation pass.
+---
+ ggml/src/ggml-alloc.c | 5 ++++-
+ 1 file changed, 4 insertions(+), 1 deletion(-)
+
+diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
+index a3d3f690..5fd379f6 100644
+--- a/ggml/src/ggml-alloc.c
+++ b/ggml/src/ggml-alloc.c
+@@ -816,7 +816,10 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
+ static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
+     size_t node_size = 0;
+     if (!node->data && !node->view_src) {
+-        GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        // If we previously had data but don't now then reallocate
+        if (talloc->buffer_id < 0) {
+            return false;
+        }
+         node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+     }
+     return talloc->size_max >= node_size;
--- a/llama/sampling_ext.cpp
+++ b/llama/sampling_ext.cpp
@ -19,6 +19,9 @@ struct common_sampler *common_sampler_cinit(const struct llama_model *model, str
        sparams.penalty_repeat = params->penalty_repeat;
        sparams.penalty_freq = params->penalty_freq;
        sparams.penalty_present = params->penalty_present;
+        sparams.mirostat = params->mirostat;
+        sparams.mirostat_tau = params->mirostat_tau;
+        sparams.mirostat_eta = params->mirostat_eta;
        sparams.seed = params->seed;
        sparams.grammar = params->grammar;
        sparams.xtc_probability = 0.0;
@ -114,9 +117,6 @@ void grammar_free(struct llama_grammar *g) {
        if (g->vocab != nullptr) {
            delete g->vocab;
        }
-        if (g->o_vocab != nullptr) {
-                delete g->o_vocab;
-        }
        llama_grammar_free_impl(g);
    }
 }
--- a/llama/sampling_ext.h
+++ b/llama/sampling_ext.h
@ -20,6 +20,9 @@ extern "C"
        float penalty_repeat;
        float penalty_freq;
        float penalty_present;
+        int32_t mirostat;
+        float mirostat_tau;
+        float mirostat_eta;
        uint32_t seed;
        char *grammar;
    };
--- a/llm/memory.go
+++ b/llm/memory.go
@ -1,12 +1,9 @@
 package llm

 import (
-	"cmp"
 	"fmt"
 	"log/slog"
-	"maps"
 	"os"
-	"slices"
 	"strconv"
 	"strings"

@ -111,8 +108,9 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 	slog.Debug("evaluating", "library", gpus[0].Library, "gpu_count", len(gpus), "available", availableList)

 	for _, projector := range projectors {
-		weight := projectorMemoryRequirements(projector)
+		weight, graph := projectorMemoryRequirements(projector)
 		projectorWeights += weight
+		projectorGraph += graph

 		// multimodal models require at least 2048 context
 		opts.NumCtx = max(opts.NumCtx, 2048)
@ -122,10 +120,12 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 	}

 	layers := f.Tensors().GroupLayers()
-	// add one layer (chosing the max layer) worth of memory as a buffer
-	layerSize = slices.MaxFunc(slices.Collect(maps.Values(layers)), func(a, b ggml.Layer) int {
-		return cmp.Compare(a.Size(), b.Size())
-	}).Size()
+	// add one layer worth of memory as a buffer
+	if blk0, ok := layers["blk.0"]; ok {
+		layerSize = blk0.Size()
+	} else {
+		slog.Warn("model missing blk.0 layer size")
+	}

 	var kvct string
 	if envconfig.FlashAttention() &&
@ -219,7 +219,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 	}

 	// For all the layers, find where they can fit on the GPU(s)
-	for i := int(f.KV().BlockCount()) - 1; i >= 0; i-- {
+	for i := range int(f.KV().BlockCount()) {
 		// Some models have inconsistent layer sizes
 		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
 			layerSize = blk.Size()
@ -229,7 +229,6 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin

 		if opts.NumGPU >= 0 && layerCount >= opts.NumGPU {
 			// Stop allocating on GPU(s) once we hit the users target NumGPU
-			overflow += layerSize
 			continue
 		}

@ -246,13 +245,13 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 				gpusWithSpace = append(gpusWithSpace[:i%j], gpusWithSpace[i%j+1:]...)
 			}
 		}
-
-		if len(gpusWithSpace) == 0 {
-			overflow += layerSize
-		}
 	}
 	if layerCount >= int(f.KV().BlockCount()) {
 		fullyLoaded = true
+	} else {
+		for i := layerCount; i < int(f.KV().BlockCount()); i++ {
+			overflow += layerSize
+		}
 	}

 	// Determine if we need to consider output then find where it fits
@ -408,21 +407,51 @@ func (m MemoryEstimate) LogValue() slog.Value {
 	return slog.GroupValue(attrs...)
 }

-func projectorMemoryRequirements(filename string) (weights uint64) {
+func projectorMemoryRequirements(filename string) (weights, graphSize uint64) {
 	file, err := os.Open(filename)
 	if err != nil {
-		return 0
+		return 0, 0
 	}
 	defer file.Close()

 	ggml, _, err := ggml.Decode(file, 1024)
 	if err != nil {
-		return 0
+		return 0, 0
 	}

 	for _, layer := range ggml.Tensors().GroupLayers() {
 		weights += layer.Size()
 	}

-	return weights
+	switch arch := ggml.KV().Architecture(); arch {
+	case "mllama":
+		kv := func(n string) uint64 {
+			if v, ok := ggml.KV()[arch+".vision."+n].(uint32); ok {
+				return uint64(v)
+			}
+
+			return 0
+		}
+
+		imageSize := kv("image_size")
+
+		maxNumTiles := kv("max_num_tiles")
+		embeddingLength := kv("embedding_length")
+		headCount := kv("attention.head_count")
+
+		numPatches := (imageSize / kv("patch_size")) * (imageSize / kv("patch_size"))
+		if _, ok := ggml.Tensors().GroupLayers()["v"]["class_embd"]; ok {
+			numPatches++
+		}
+
+		numPaddedPatches := numPatches + 8 - (numPatches%8)%8
+
+		graphSize = 4 * (8 +
+			imageSize*imageSize*kv("num_channels")*maxNumTiles +
+			embeddingLength*numPatches*maxNumTiles +
+			9*embeddingLength*numPaddedPatches*maxNumTiles +
+			numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount)
+	}
+
+	return weights, graphSize
 }
--- a/llm/server.go
+++ b/llm/server.go
@ -17,7 +17,6 @@ import (
 	"os/exec"
 	"path/filepath"
 	"runtime"
-	"slices"
 	"strconv"
 	"strings"
 	"sync"
@ -31,37 +30,9 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/llama"
-	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/model"
 )

-type filteredEnv []string
-
-func (e filteredEnv) LogValue() slog.Value {
-	var attrs []slog.Attr
-	for _, env := range e {
-		if key, value, ok := strings.Cut(env, "="); ok {
-			switch {
-			case strings.HasPrefix(key, "OLLAMA_"),
-				strings.HasPrefix(key, "CUDA_"),
-				strings.HasPrefix(key, "ROCR_"),
-				strings.HasPrefix(key, "ROCM_"),
-				strings.HasPrefix(key, "HIP_"),
-				strings.HasPrefix(key, "GPU_"),
-				strings.HasPrefix(key, "HSA_"),
-				strings.HasPrefix(key, "GGML_"),
-				slices.Contains([]string{
-					"PATH",
-					"LD_LIBRARY_PATH",
-					"DYLD_LIBRARY_PATH",
-				}, key):
-				attrs = append(attrs, slog.String(key, value))
-			}
-		}
-	}
-	return slog.GroupValue(attrs...)
-}
-
 type LlamaServer interface {
 	Ping(ctx context.Context) error
 	WaitUntilRunning(ctx context.Context) error
@ -177,6 +148,10 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		params = append(params, "--n-gpu-layers", strconv.Itoa(opts.NumGPU))
 	}

+	if envconfig.Debug() {
+		params = append(params, "--verbose")
+	}
+
 	if opts.MainGPU > 0 {
 		params = append(params, "--main-gpu", strconv.Itoa(opts.MainGPU))
 	}
@ -311,7 +286,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		params = append(params, "--mmproj", projectors[0])
 	}

-	// iterate through compatible GPU libraries such as 'cuda_v12', 'cuda_v11', 'rocm', etc.
+	// iterate through compatible GPU libraries such as 'cuda_v12', 'rocm', etc.
 	// adding each library's respective path to the LD_LIBRARY_PATH, until finally running
 	// without any LD_LIBRARY_PATH flags
 	for {
@ -429,7 +404,26 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		}

 		slog.Info("starting llama server", "cmd", s.cmd)
-		slog.Debug("subprocess", "", filteredEnv(s.cmd.Env))
+		if envconfig.Debug() {
+			filteredEnv := []string{}
+			for _, ev := range s.cmd.Env {
+				if strings.HasPrefix(ev, "OLLAMA_") ||
+					strings.HasPrefix(ev, "CUDA_") ||
+					strings.HasPrefix(ev, "ROCR_") ||
+					strings.HasPrefix(ev, "ROCM_") ||
+					strings.HasPrefix(ev, "HIP_") ||
+					strings.HasPrefix(ev, "GPU_") ||
+					strings.HasPrefix(ev, "HSA_") ||
+					strings.HasPrefix(ev, "GGML_") ||
+					strings.HasPrefix(ev, "PATH=") ||
+					strings.HasPrefix(ev, "LD_LIBRARY_PATH=") ||
+					strings.HasPrefix(ev, "DYLD_LIBRARY_PATH=") {
+					filteredEnv = append(filteredEnv, ev)
+				}
+			}
+			// Log at debug as the environment is inherited and might contain sensitive information
+			slog.Debug("subprocess", "environment", filteredEnv)
+		}

 		if err = s.cmd.Start(); err != nil {
 			var msg string
@ -679,8 +673,9 @@ ws ::= ([ \t\n] ws)?
 const maxBufferSize = 512 * format.KiloByte

 type ImageData struct {
-	Data []byte `json:"data"`
-	ID   int    `json:"id"`
+	Data          []byte `json:"data"`
+	ID            int    `json:"id"`
+	AspectRatioID int    `json:"aspect_ratio_id"`
 }

 type CompletionRequest struct {
@ -726,9 +721,6 @@ type CompletionResponse struct {
 }

 func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error {
-	slog.Debug("completion request", "images", len(req.Images), "prompt", len(req.Prompt), "format", string(req.Format))
-	slog.Log(ctx, logutil.LevelTrace, "completion request", "prompt", req.Prompt)
-
 	if len(req.Format) > 0 {
 		switch string(req.Format) {
 		case `null`, `""`:
@ -892,8 +884,6 @@ type EmbeddingResponse struct {
 }

 func (s *llmServer) Embedding(ctx context.Context, input string) ([]float32, error) {
-	slog.Log(ctx, logutil.LevelTrace, "embedding request", "input", input)
-
 	if err := s.sem.Acquire(ctx, 1); err != nil {
 		if errors.Is(err, context.Canceled) {
 			slog.Info("aborting embedding request due to client closing the connection")
--- a/llm/server_test.go
+++ b/llm/server_test.go
@ -16,7 +16,7 @@ func TestLLMServerCompletionFormat(t *testing.T) {
 	// of a mess, and but it's good enough, until we can refactoring the
 	// Completion method to be more testable.

-	ctx, cancel := context.WithCancel(t.Context())
+	ctx, cancel := context.WithCancel(context.Background())
 	s := &llmServer{
 		sem: semaphore.NewWeighted(1), // required to prevent nil panic
 	}
--- a/logutil/logutil.go
+++ b/logutil/logutil.go
@ -1,29 +0,0 @@
-package logutil
-
-import (
-	"io"
-	"log/slog"
-	"path/filepath"
-)
-
-const LevelTrace slog.Level = -8
-
-func NewLogger(w io.Writer, level slog.Level) *slog.Logger {
-	return slog.New(slog.NewTextHandler(w, &slog.HandlerOptions{
-		Level:     level,
-		AddSource: true,
-		ReplaceAttr: func(_ []string, attr slog.Attr) slog.Attr {
-			switch attr.Key {
-			case slog.LevelKey:
-				switch attr.Value.Any().(slog.Level) {
-				case LevelTrace:
-					attr.Value = slog.StringValue("TRACE")
-				}
-			case slog.SourceKey:
-				source := attr.Value.Any().(*slog.Source)
-				source.File = filepath.Base(source.File)
-			}
-			return attr
-		},
-	}))
-}
--- a/ml/backend.go
+++ b/ml/backend.go
@ -5,7 +5,6 @@ import (
 	"context"
 	"encoding/binary"
 	"fmt"
-	"math"
 	"os"
 	"slices"
 	"strconv"
@ -119,21 +118,6 @@ type Context interface {
 	Layer(int) Context
 }

-// RopeOptions contains optional parameters for RoPE function
-type RopeOptions struct {
-	OriginalContextLen uint32
-}
-
-// RopeOption defines a function that modifies RopeOpts
-type RopeOption func(*RopeOptions)
-
-// WithContextLen sets a custom context length
-func WithContextLen(len uint32) RopeOption {
-	return func(opts *RopeOptions) {
-		opts.OriginalContextLen = len
-	}
-}
-
 type Tensor interface {
 	Dim(n int) int
 	Stride(n int) int
@ -159,7 +143,7 @@ type Tensor interface {
 	AvgPool2D(ctx Context, k, s int, p float32) Tensor
 	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor

-	RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim, ropeType uint32, base, scale float32, options ...RopeOption) Tensor
+	RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim, ropeType uint32, base, scale float32) Tensor
 	IM2Col(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor

 	Sin(ctx Context) Tensor
@ -176,6 +160,7 @@ type Tensor interface {
 	Set(ctx Context, t2 Tensor, offset int, strides ...int) Tensor

 	Pad(ctx Context, shape ...int) Tensor
+	Unpad(ctx Context, shape ...int) Tensor

 	Stack(ctx Context, dim int, s ...Tensor) Tensor

@ -187,7 +172,6 @@ type Tensor interface {
 	Duplicate(ctx Context) Tensor

 	TopK(ctx Context, k int) Tensor
-	Argsort(ctx Context) Tensor
 }

 // ScaledDotProductAttention implements a fused attention
@ -230,58 +214,35 @@ func mul[T number](s ...T) T {
 	return p
 }

-type DumpOptions func(*dumpOptions)
+type DumpOptions struct {
+	// Items is the number of elements to print at the beginning and end of each dimension.
+	Items int

-// DumpWithPrecision sets the number of decimal places to print. Applies to float32 and float64.
-func DumpWithPrecision(n int) DumpOptions {
-	return func(opts *dumpOptions) {
-		opts.Precision = n
-	}
+	// Precision is the number of decimal places to print. Applies to float32 and float64.
+	Precision int
 }

-// DumpWithThreshold sets the threshold for printing the entire tensor. If the number of elements
-// is less than or equal to this value, the entire tensor will be printed. Otherwise, only the
-// beginning and end of each dimension will be printed.
-func DumpWithThreshold(n int) DumpOptions {
-	return func(opts *dumpOptions) {
-		opts.Threshold = n
-	}
-}
-
-// DumpWithEdgeItems sets the number of elements to print at the beginning and end of each dimension.
-func DumpWithEdgeItems(n int) DumpOptions {
-	return func(opts *dumpOptions) {
-		opts.EdgeItems = n
-	}
-}
-
-type dumpOptions struct {
-	Precision, Threshold, EdgeItems int
-}
-
-func Dump(ctx Context, t Tensor, optsFuncs ...DumpOptions) string {
-	opts := dumpOptions{Precision: 4, Threshold: 1000, EdgeItems: 3}
-	for _, optsFunc := range optsFuncs {
-		optsFunc(&opts)
-	}
-
-	if mul(t.Shape()...) <= opts.Threshold {
-		opts.EdgeItems = math.MaxInt
+func Dump(ctx Context, t Tensor, opts ...DumpOptions) string {
+	if len(opts) < 1 {
+		opts = append(opts, DumpOptions{
+			Items:     3,
+			Precision: 4,
+		})
 	}

 	switch t.DType() {
 	case DTypeF32:
-		return dump[[]float32](ctx, t, opts.EdgeItems, func(f float32) string {
-			return strconv.FormatFloat(float64(f), 'f', opts.Precision, 32)
+		return dump[[]float32](ctx, t, opts[0].Items, func(f float32) string {
+			return strconv.FormatFloat(float64(f), 'f', opts[0].Precision, 32)
 		})
 	case DTypeF16, DTypeQ80, DTypeQ40:
 		f32 := ctx.Input().Empty(DTypeF32, t.Shape()...)
 		f32 = t.Copy(ctx, f32)
-		return dump[[]float32](ctx, f32, opts.EdgeItems, func(f float32) string {
-			return strconv.FormatFloat(float64(f), 'f', opts.Precision, 32)
+		return dump[[]float32](ctx, f32, opts[0].Items, func(f float32) string {
+			return strconv.FormatFloat(float64(f), 'f', opts[0].Precision, 32)
 		})
 	case DTypeI32:
-		return dump[[]int32](ctx, t, opts.EdgeItems, func(i int32) string {
+		return dump[[]int32](ctx, t, opts[0].Items, func(i int32) string {
 			return strconv.FormatInt(int64(i), 10)
 		})
 	default:
--- a/Show More
+++ b/Show More