Merge remote-tracking branch 'upstream/main' into VulkanV3Update

.github/workflows/release.yaml

@@ -94,7 +94,7 @@ jobs:
             install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
             rocm-version: '6.2'
             flags: '-DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
-            runner_dir: ''
+            runner_dir: 'rocm'
     runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
     environment: release
     env:
@@ -163,6 +163,7 @@ jobs:
           cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }} -DOLLAMA_RUNNER_DIR="${{ matrix.runner_dir }}"
           cmake --build --parallel --preset "${{ matrix.preset }}"
           cmake --install build --component "${{ startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || 'CPU' }}" --strip --parallel 8
+          Remove-Item -Path dist\lib\ollama\rocm\rocblas\library\*gfx906* -ErrorAction SilentlyContinue
         env:
           CMAKE_GENERATOR: Ninja
       - uses: actions/upload-artifact@v4
@@ -185,7 +186,8 @@ jobs:
         if: matrix.arch == 'amd64'
         run: |
           $ErrorActionPreference = "Stop"
-          Start-Process "C:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
+          # TODO mingw-w64-clang-x86_64-clang v21 is currently broken for cgo build, but v20 + gcc-compat isn't available
+          # Start-Process "C:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
           echo "C:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
       - name: Install ARM64 system dependencies
@@ -209,6 +211,7 @@ jobs:
         with:
           go-version-file: go.mod
       - run: |
+          go clean -cache
           go build -o dist/${{ matrix.os }}-${{ matrix.arch }}/ .
       - uses: actions/upload-artifact@v4
         with:

.gitignore

@@ -6,6 +6,7 @@
 dist
 build
 .cache
+.gocache
 *.exe
 .idea
 test_data

CMakeLists.txt

@@ -89,23 +89,26 @@ if(CMAKE_CUDA_COMPILER)
     )
 endif()
 
-set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a|1200|1201):xnack[+-]$"
+set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(908|90a|1200|1201):xnack[+-]$"
     CACHE STRING
-    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a|1200|1201):xnack[+-]$\"."
+    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(908|90a|1200|1201):xnack[+-]$\"."
 )
 
 check_language(HIP)
 if(CMAKE_HIP_COMPILER)
     set(HIP_PLATFORM "amd")
 
-    find_package(hip REQUIRED)
     if(NOT AMDGPU_TARGETS)
-        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012]|120[01])$")
-    elseif(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
+        find_package(hip REQUIRED)
+        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(94[012]|101[02]|1030|110[012]|120[01])$")
+    endif()
+
+    if(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
         list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
     endif()
 
     if(AMDGPU_TARGETS)
+        find_package(hip REQUIRED)
         add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-hip)
 
         if (WIN32)

CMakePresets.json

@@ -68,7 +68,7 @@
       "inherits": [ "ROCm" ],
       "cacheVariables": {
         "CMAKE_HIP_FLAGS": "-parallel-jobs=4",
-        "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
+        "AMDGPU_TARGETS": "gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
       }
     },
     {

Dockerfile

@@ -91,6 +91,7 @@ RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'ROCm 6' -DOLLAMA_RUNNER_DIR="rocm" \
         && cmake --build --parallel ${PARALLEL} --preset 'ROCm 6' \
         && cmake --install build --component HIP --strip --parallel ${PARALLEL}
+RUN rm -f dist/lib/ollama/rocm/rocblas/library/*gfx90[06]*
 
 FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK5VERSION} AS jetpack-5
 ARG CMAKEVERSION
@@ -146,9 +147,9 @@ COPY --from=vulkan  dist/lib/ollama  /lib/ollama/
 FROM --platform=linux/arm64 scratch AS arm64
 # COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
 COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
-COPY --from=cuda-13 dist/lib/ollama /lib/ollama/
-COPY --from=jetpack-5 dist/lib/ollama /lib/ollama/
-COPY --from=jetpack-6 dist/lib/ollama /lib/ollama/
+COPY --from=cuda-13 dist/lib/ollama/ /lib/ollama/
+COPY --from=jetpack-5 dist/lib/ollama/ /lib/ollama/
+COPY --from=jetpack-6 dist/lib/ollama/ /lib/ollama/
 
 FROM scratch AS rocm
 COPY --from=rocm-6 dist/lib/ollama /lib/ollama

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggml-org/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=e54d41befcc1575f4c898c5ff4ef43970cead75f
+FETCH_HEAD=364a7a6d4a786e98947c8a90430ea581213c0ba9
 
 .PHONY: help
 help:

api/client.go

@@ -45,6 +45,12 @@ func checkError(resp *http.Response, body []byte) error {
 		return nil
 	}
 
+	if resp.StatusCode == http.StatusUnauthorized {
+		authError := AuthorizationError{StatusCode: resp.StatusCode}
+		json.Unmarshal(body, &authError)
+		return authError
+	}
+
 	apiError := StatusError{StatusCode: resp.StatusCode}
 
 	err := json.Unmarshal(body, &apiError)
@@ -215,6 +221,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 	for scanner.Scan() {
 		var errorResponse struct {
 			Error     string `json:"error,omitempty"`
+			SigninURL string `json:"signin_url,omitempty"`
 		}
 
 		bts := scanner.Bytes()
@@ -223,14 +230,10 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		}
 
 		if response.StatusCode == http.StatusUnauthorized {
-			pubKey, pkErr := auth.GetPublicKey()
-			if pkErr != nil {
-				return pkErr
-			}
 			return AuthorizationError{
 				StatusCode: response.StatusCode,
 				Status:     response.Status,
-				PublicKey:  pubKey,
+				SigninURL:  errorResponse.SigninURL,
 			}
 		} else if response.StatusCode >= http.StatusBadRequest {
 			return StatusError{
@@ -439,8 +442,13 @@ func (c *Client) Version(ctx context.Context) (string, error) {
 	return version.Version, nil
 }
 
-// Signout will disconnect an ollama instance from ollama.com
-func (c *Client) Signout(ctx context.Context, encodedKey string) error {
+// Signout will signout a client for a local ollama server.
+func (c *Client) Signout(ctx context.Context) error {
+	return c.do(ctx, http.MethodPost, "/api/signout", nil, nil)
+}
+
+// Disconnect will disconnect an ollama instance from ollama.com.
+func (c *Client) Disconnect(ctx context.Context, encodedKey string) error {
 	return c.do(ctx, http.MethodDelete, fmt.Sprintf("/api/user/keys/%s", encodedKey), nil, nil)
 }
 

api/types.go

@@ -41,7 +41,7 @@ func (e StatusError) Error() string {
 type AuthorizationError struct {
 	StatusCode int
 	Status     string
-	PublicKey  string `json:"public_key"`
+	SigninURL  string `json:"signin_url"`
 }
 
 func (e AuthorizationError) Error() string {

auth/auth.go

@@ -18,46 +18,13 @@ import (
 
 const defaultPrivateKey = "id_ed25519"
 
-func keyPath() (string, error) {
-	fileIsReadable := func(fp string) bool {
-		info, err := os.Stat(fp)
-		if err != nil {
-			return false
-		}
-
-		// Check that it's a regular file, not a directory or other file type
-		if !info.Mode().IsRegular() {
-			return false
-		}
-
-		// Try to open it to check readability
-		file, err := os.Open(fp)
-		if err != nil {
-			return false
-		}
-		file.Close()
-		return true
-	}
-
-	systemPath := filepath.Join("/usr/share/ollama/.ollama", defaultPrivateKey)
-	if fileIsReadable(systemPath) {
-		return systemPath, nil
-	}
-
+func GetPublicKey() (string, error) {
 	home, err := os.UserHomeDir()
 	if err != nil {
 		return "", err
 	}
 
-	return filepath.Join(home, ".ollama", defaultPrivateKey), nil
-}
-
-func GetPublicKey() (string, error) {
-	keyPath, err := keyPath()
-	if err != nil {
-		return "", err
-	}
-
+	keyPath := filepath.Join(home, ".ollama", defaultPrivateKey)
 	privateKeyFile, err := os.ReadFile(keyPath)
 	if err != nil {
 		slog.Info(fmt.Sprintf("Failed to load private key: %v", err))
@@ -84,11 +51,12 @@ func NewNonce(r io.Reader, length int) (string, error) {
 }
 
 func Sign(ctx context.Context, bts []byte) (string, error) {
-	keyPath, err := keyPath()
+	home, err := os.UserHomeDir()
 	if err != nil {
 		return "", err
 	}
 
+	keyPath := filepath.Join(home, ".ollama", defaultPrivateKey)
 	privateKeyFile, err := os.ReadFile(keyPath)
 	if err != nil {
 		slog.Info(fmt.Sprintf("Failed to load private key: %v", err))

cmd/cmd.go

@@ -5,7 +5,6 @@ import (
 	"context"
 	"crypto/ed25519"
 	"crypto/rand"
-	"encoding/base64"
 	"encoding/json"
 	"encoding/pem"
 	"errors"
@@ -15,7 +14,6 @@ import (
 	"math"
 	"net"
 	"net/http"
-	"net/url"
 	"os"
 	"os/signal"
 	"path/filepath"
@@ -37,7 +35,6 @@ import (
 	"golang.org/x/term"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/auth"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/parser"
@@ -50,7 +47,7 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
-const ConnectInstructions = "To sign in, navigate to:\n    https://ollama.com/connect?name=%s&key=%s\n\n"
+const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
 
 // ensureThinkingSupport emits a warning if the model does not advertise thinking support
 func ensureThinkingSupport(ctx context.Context, client *api.Client, name string) {
@@ -452,16 +449,10 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
 			var sErr api.AuthorizationError
 			if errors.As(err, &sErr) && sErr.StatusCode == http.StatusUnauthorized {
-				pubKey, pkErr := auth.GetPublicKey()
-				if pkErr != nil {
-					return pkErr
-				}
-				// the server and the client both have the same public key
-				if pubKey == sErr.PublicKey {
-					h, _ := os.Hostname()
-					encKey := base64.RawURLEncoding.EncodeToString([]byte(pubKey))
 				fmt.Printf("You need to be signed in to Ollama to run Cloud models.\n\n")
-					fmt.Printf(ConnectInstructions, url.PathEscape(h), encKey)
+
+				if sErr.SigninURL != "" {
+					fmt.Printf(ConnectInstructions, sErr.SigninURL)
 				}
 				return nil
 			}
@@ -493,6 +484,16 @@ func SigninHandler(cmd *cobra.Command, args []string) error {
 
 	user, err := client.Whoami(cmd.Context())
 	if err != nil {
+		var aErr api.AuthorizationError
+		if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
+			fmt.Println("You need to be signed in to Ollama to run Cloud models.")
+			fmt.Println()
+
+			if aErr.SigninURL != "" {
+				fmt.Printf(ConnectInstructions, aErr.SigninURL)
+			}
+			return nil
+		}
 		return err
 	}
 
@@ -502,34 +503,27 @@ func SigninHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}
 
-	pubKey, pkErr := auth.GetPublicKey()
-	if pkErr != nil {
-		return pkErr
-	}
-	encKey := base64.RawURLEncoding.EncodeToString([]byte(pubKey))
-
-	h, _ := os.Hostname()
-	fmt.Printf(ConnectInstructions, url.PathEscape(h), encKey)
-
 	return nil
 }
 
 func SignoutHandler(cmd *cobra.Command, args []string) error {
-	pubKey, pkErr := auth.GetPublicKey()
-	if pkErr != nil {
-		return pkErr
-	}
-	encKey := base64.RawURLEncoding.EncodeToString([]byte(pubKey))
-
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
 
-	err = client.Signout(cmd.Context(), encKey)
+	err = client.Signout(cmd.Context())
 	if err != nil {
+		var aErr api.AuthorizationError
+		if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
+			fmt.Println("You are not signed in to ollama.com")
+			fmt.Println()
+			return nil
+		} else {
 			return err
 		}
+	}
+
 	fmt.Println("You have signed out of ollama.com")
 	fmt.Println()
 	return nil
@@ -546,6 +540,25 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
+	n := model.ParseName(args[0])
+	if strings.HasSuffix(n.Host, ".ollama.ai") || strings.HasSuffix(n.Host, ".ollama.com") {
+		_, err := client.Whoami(cmd.Context())
+		if err != nil {
+			var aErr api.AuthorizationError
+			if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
+				fmt.Println("You need to be signed in to push models to ollama.com.")
+				fmt.Println()
+
+				if aErr.SigninURL != "" {
+					fmt.Printf(ConnectInstructions, aErr.SigninURL)
+				}
+				return nil
+			}
+
+			return err
+		}
+	}
+
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
@@ -582,7 +595,6 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 
 	request := api.PushRequest{Name: args[0], Insecure: insecure}
 
-	n := model.ParseName(args[0])
 	if err := client.Push(cmd.Context(), &request, fn); err != nil {
 		if spinner != nil {
 			spinner.Stop()
@@ -1106,6 +1118,51 @@ type runOptions struct {
 	ShowConnect  bool
 }
 
+func (r runOptions) Copy() runOptions {
+	var messages []api.Message
+	if r.Messages != nil {
+		messages = make([]api.Message, len(r.Messages))
+		copy(messages, r.Messages)
+	}
+
+	var images []api.ImageData
+	if r.Images != nil {
+		images = make([]api.ImageData, len(r.Images))
+		copy(images, r.Images)
+	}
+
+	var opts map[string]any
+	if r.Options != nil {
+		opts = make(map[string]any, len(r.Options))
+		for k, v := range r.Options {
+			opts[k] = v
+		}
+	}
+
+	var think *api.ThinkValue
+	if r.Think != nil {
+		cThink := *r.Think
+		think = &cThink
+	}
+
+	return runOptions{
+		Model:        r.Model,
+		ParentModel:  r.ParentModel,
+		Prompt:       r.Prompt,
+		Messages:     messages,
+		WordWrap:     r.WordWrap,
+		Format:       r.Format,
+		System:       r.System,
+		Images:       images,
+		Options:      opts,
+		MultiModal:   r.MultiModal,
+		KeepAlive:    r.KeepAlive,
+		Think:        think,
+		HideThinking: r.HideThinking,
+		ShowConnect:  r.ShowConnect,
+	}
+}
+
 type displayResponseState struct {
 	lineLength int
 	wordBuffer string

cmd/cmd_test.go

@@ -8,6 +8,7 @@ import (
 	"net/http"
 	"net/http/httptest"
 	"os"
+	"reflect"
 	"strings"
 	"testing"
 	"time"
@@ -491,9 +492,35 @@ func TestPushHandler(t *testing.T) {
 						w.(http.Flusher).Flush()
 					}
 				},
+				"/api/me": func(w http.ResponseWriter, r *http.Request) {
+					if r.Method != http.MethodPost {
+						t.Errorf("expected POST request, got %s", r.Method)
+					}
+				},
 			},
 			expectedOutput: "\nYou can find your model at:\n\n\thttps://ollama.com/test-model\n",
 		},
+		{
+			name:      "not signed in push",
+			modelName: "notsignedin-model",
+			serverResponse: map[string]func(w http.ResponseWriter, r *http.Request){
+				"/api/me": func(w http.ResponseWriter, r *http.Request) {
+					if r.Method != http.MethodPost {
+						t.Errorf("expected POST request, got %s", r.Method)
+					}
+					w.Header().Set("Content-Type", "application/json")
+					w.WriteHeader(http.StatusUnauthorized)
+					err := json.NewEncoder(w).Encode(map[string]string{
+						"error":      "unauthorized",
+						"signin_url": "https://somethingsomething",
+					})
+					if err != nil {
+						t.Fatal(err)
+					}
+				},
+			},
+			expectedOutput: "You need to be signed in to push",
+		},
 		{
 			name:      "unauthorized push",
 			modelName: "unauthorized-model",
@@ -508,6 +535,11 @@ func TestPushHandler(t *testing.T) {
 						t.Fatal(err)
 					}
 				},
+				"/api/me": func(w http.ResponseWriter, r *http.Request) {
+					if r.Method != http.MethodPost {
+						t.Errorf("expected POST request, got %s", r.Method)
+					}
+				},
 			},
 			expectedError: "you are not authorized to push to this namespace, create the model under a namespace you own",
 		},
@@ -525,6 +557,9 @@ func TestPushHandler(t *testing.T) {
 			defer mockServer.Close()
 
 			t.Setenv("OLLAMA_HOST", mockServer.URL)
+			tmpDir := t.TempDir()
+			t.Setenv("HOME", tmpDir)
+			t.Setenv("USERPROFILE", tmpDir)
 			initializeKeypair()
 
 			cmd := &cobra.Command{}
@@ -561,7 +596,7 @@ func TestPushHandler(t *testing.T) {
 					t.Errorf("expected no error, got %v", err)
 				}
 				if tt.expectedOutput != "" {
-					if got := string(stdout); got != tt.expectedOutput {
+					if got := string(stdout); !strings.Contains(got, tt.expectedOutput) {
 						t.Errorf("expected output %q, got %q", tt.expectedOutput, got)
 					}
 				}
@@ -919,3 +954,286 @@ func TestNewCreateRequest(t *testing.T) {
 		})
 	}
 }
+
+func TestRunOptions_Copy(t *testing.T) {
+	// Setup test data
+	originalKeepAlive := &api.Duration{Duration: 5 * time.Minute}
+	originalThink := &api.ThinkValue{Value: "test reasoning"}
+
+	original := runOptions{
+		Model:       "test-model",
+		ParentModel: "parent-model",
+		Prompt:      "test prompt",
+		Messages: []api.Message{
+			{Role: "user", Content: "hello"},
+			{Role: "assistant", Content: "hi there"},
+		},
+		WordWrap: true,
+		Format:   "json",
+		System:   "system prompt",
+		Images: []api.ImageData{
+			[]byte("image1"),
+			[]byte("image2"),
+		},
+		Options: map[string]any{
+			"temperature": 0.7,
+			"max_tokens":  1000,
+			"top_p":       0.9,
+		},
+		MultiModal:   true,
+		KeepAlive:    originalKeepAlive,
+		Think:        originalThink,
+		HideThinking: false,
+		ShowConnect:  true,
+	}
+
+	// Test the copy
+	copied := original.Copy()
+
+	// Test 1: Verify the copy is not the same instance
+	if &copied == &original {
+		t.Error("Copy should return a different instance")
+	}
+
+	// Test 2: Verify all fields are copied correctly
+	tests := []struct {
+		name string
+		got  interface{}
+		want interface{}
+	}{
+		{"Model", copied.Model, original.Model},
+		{"ParentModel", copied.ParentModel, original.ParentModel},
+		{"Prompt", copied.Prompt, original.Prompt},
+		{"WordWrap", copied.WordWrap, original.WordWrap},
+		{"Format", copied.Format, original.Format},
+		{"System", copied.System, original.System},
+		{"MultiModal", copied.MultiModal, original.MultiModal},
+		{"HideThinking", copied.HideThinking, original.HideThinking},
+		{"ShowConnect", copied.ShowConnect, original.ShowConnect},
+	}
+
+	for _, tt := range tests {
+		if !reflect.DeepEqual(tt.got, tt.want) {
+			t.Errorf("%s mismatch: got %v, want %v", tt.name, tt.got, tt.want)
+		}
+	}
+
+	// Test 3: Verify Messages slice is deeply copied
+	if len(copied.Messages) != len(original.Messages) {
+		t.Errorf("Messages length mismatch: got %d, want %d", len(copied.Messages), len(original.Messages))
+	}
+
+	if len(copied.Messages) > 0 && &copied.Messages[0] == &original.Messages[0] {
+		t.Error("Messages should be different instances")
+	}
+
+	// Modify original to verify independence
+	if len(original.Messages) > 0 {
+		originalContent := original.Messages[0].Content
+		original.Messages[0].Content = "modified"
+		if len(copied.Messages) > 0 && copied.Messages[0].Content == "modified" {
+			t.Error("Messages should be independent after copy")
+		}
+		// Restore for other tests
+		original.Messages[0].Content = originalContent
+	}
+
+	// Test 4: Verify Images slice is deeply copied
+	if len(copied.Images) != len(original.Images) {
+		t.Errorf("Images length mismatch: got %d, want %d", len(copied.Images), len(original.Images))
+	}
+
+	if len(copied.Images) > 0 && &copied.Images[0] == &original.Images[0] {
+		t.Error("Images should be different instances")
+	}
+
+	// Modify original to verify independence
+	if len(original.Images) > 0 {
+		originalImage := original.Images[0]
+		original.Images[0] = []byte("modified")
+		if len(copied.Images) > 0 && string(copied.Images[0]) == "modified" {
+			t.Error("Images should be independent after copy")
+		}
+		// Restore for other tests
+		original.Images[0] = originalImage
+	}
+
+	// Test 5: Verify Options map is deeply copied
+	if len(copied.Options) != len(original.Options) {
+		t.Errorf("Options length mismatch: got %d, want %d", len(copied.Options), len(original.Options))
+	}
+
+	if len(copied.Options) > 0 && &copied.Options == &original.Options {
+		t.Error("Options map should be different instances")
+	}
+
+	// Modify original to verify independence
+	if len(original.Options) > 0 {
+		originalTemp := original.Options["temperature"]
+		original.Options["temperature"] = 0.9
+		if copied.Options["temperature"] == 0.9 {
+			t.Error("Options should be independent after copy")
+		}
+		// Restore for other tests
+		original.Options["temperature"] = originalTemp
+	}
+
+	// Test 6: Verify KeepAlive pointer is copied (shallow copy)
+	if copied.KeepAlive != original.KeepAlive {
+		t.Error("KeepAlive pointer should be the same (shallow copy)")
+	}
+
+	// Test 7: Verify Think pointer creates a new instance
+	if original.Think != nil && copied.Think == original.Think {
+		t.Error("Think should be a different instance")
+	}
+
+	if original.Think != nil && copied.Think != nil {
+		if !reflect.DeepEqual(copied.Think.Value, original.Think.Value) {
+			t.Errorf("Think.Value mismatch: got %v, want %v", copied.Think.Value, original.Think.Value)
+		}
+	}
+
+	// Test 8: Test with zero values
+	zeroOriginal := runOptions{}
+	zeroCopy := zeroOriginal.Copy()
+
+	if !reflect.DeepEqual(zeroCopy, zeroOriginal) {
+		fmt.Printf("orig: %#v\ncopy: %#v\n", zeroOriginal, zeroCopy)
+		t.Error("Copy of zero value should equal original zero value")
+	}
+}
+
+func TestRunOptions_Copy_EmptySlicesAndMaps(t *testing.T) {
+	// Test with empty slices and maps
+	original := runOptions{
+		Messages: []api.Message{},
+		Images:   []api.ImageData{},
+		Options:  map[string]any{},
+	}
+
+	copied := original.Copy()
+
+	if copied.Messages == nil {
+		t.Error("Empty Messages slice should remain empty, not nil")
+	}
+
+	if copied.Images == nil {
+		t.Error("Empty Images slice should remain empty, not nil")
+	}
+
+	if copied.Options == nil {
+		t.Error("Empty Options map should remain empty, not nil")
+	}
+
+	if len(copied.Messages) != 0 {
+		t.Error("Empty Messages slice should remain empty")
+	}
+
+	if len(copied.Images) != 0 {
+		t.Error("Empty Images slice should remain empty")
+	}
+
+	if len(copied.Options) != 0 {
+		t.Error("Empty Options map should remain empty")
+	}
+}
+
+func TestRunOptions_Copy_NilPointers(t *testing.T) {
+	// Test with nil pointers
+	original := runOptions{
+		KeepAlive: nil,
+		Think:     nil,
+	}
+
+	copied := original.Copy()
+
+	if copied.KeepAlive != nil {
+		t.Error("Nil KeepAlive should remain nil")
+	}
+
+	if copied.Think != nil {
+		t.Error("Nil Think should remain nil")
+	}
+}
+
+func TestRunOptions_Copy_ThinkValueVariants(t *testing.T) {
+	tests := []struct {
+		name  string
+		think *api.ThinkValue
+	}{
+		{"nil Think", nil},
+		{"bool true", &api.ThinkValue{Value: true}},
+		{"bool false", &api.ThinkValue{Value: false}},
+		{"string value", &api.ThinkValue{Value: "reasoning text"}},
+		{"int value", &api.ThinkValue{Value: 42}},
+		{"nil value", &api.ThinkValue{Value: nil}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			original := runOptions{Think: tt.think}
+			copied := original.Copy()
+
+			if tt.think == nil {
+				if copied.Think != nil {
+					t.Error("Nil Think should remain nil")
+				}
+				return
+			}
+
+			if copied.Think == nil {
+				t.Error("Non-nil Think should not become nil")
+				return
+			}
+
+			if copied.Think == original.Think {
+				t.Error("Think should be a different instance")
+			}
+
+			if !reflect.DeepEqual(copied.Think.Value, original.Think.Value) {
+				t.Errorf("Think.Value mismatch: got %v, want %v", copied.Think.Value, original.Think.Value)
+			}
+		})
+	}
+}
+
+func TestRunOptions_Copy_Independence(t *testing.T) {
+	// Test that modifications to original don't affect copy
+	originalThink := &api.ThinkValue{Value: "original"}
+	original := runOptions{
+		Model:    "original-model",
+		Messages: []api.Message{{Role: "user", Content: "original"}},
+		Options:  map[string]any{"key": "value"},
+		Think:    originalThink,
+	}
+
+	copied := original.Copy()
+
+	// Modify original
+	original.Model = "modified-model"
+	if len(original.Messages) > 0 {
+		original.Messages[0].Content = "modified"
+	}
+	original.Options["key"] = "modified"
+	if original.Think != nil {
+		original.Think.Value = "modified"
+	}
+
+	// Verify copy is unchanged
+	if copied.Model == "modified-model" {
+		t.Error("Copy Model should not be affected by original modification")
+	}
+
+	if len(copied.Messages) > 0 && copied.Messages[0].Content == "modified" {
+		t.Error("Copy Messages should not be affected by original modification")
+	}
+
+	if copied.Options["key"] == "modified" {
+		t.Error("Copy Options should not be affected by original modification")
+	}
+
+	if copied.Think != nil && copied.Think.Value == "modified" {
+		t.Error("Copy Think should not be affected by original modification")
+	}
+}

cmd/interactive.go

@@ -195,16 +195,24 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 				fmt.Println("Usage:\n  /load <modelname>")
 				continue
 			}
+			origOpts := opts.Copy()
+
 			opts.Model = args[1]
 			opts.Messages = []api.Message{}
 			fmt.Printf("Loading model '%s'\n", opts.Model)
 			opts.Think, err = inferThinkingOption(nil, &opts, thinkExplicitlySet)
 			if err != nil {
+				if strings.Contains(err.Error(), "not found") {
+					fmt.Printf("Couldn't find model '%s'\n", opts.Model)
+					opts = origOpts.Copy()
+					continue
+				}
 				return err
 			}
 			if err := loadOrUnloadModel(cmd, &opts); err != nil {
 				if strings.Contains(err.Error(), "not found") {
-					fmt.Printf("error: %v\n", err)
+					fmt.Printf("Couldn't find model '%s'\n", opts.Model)
+					opts = origOpts.Copy()
 					continue
 				}
 				if strings.Contains(err.Error(), "does not support thinking") {

discover/gpu.go

@@ -25,8 +25,10 @@ func GetCPUInfo() GpuInfo {
 
 	return GpuInfo{
 		memInfo: mem,
+		DeviceID: ml.DeviceID{
 			Library: "cpu",
 			ID:      "0",
+		},
 	}
 }
 
@@ -47,23 +49,22 @@ func devInfoToInfoList(devs []ml.DeviceInfo) GpuInfoList {
 
 	for _, dev := range devs {
 		info := GpuInfo{
-			ID:       dev.ID,
+			DeviceID: dev.DeviceID,
 			filterID: dev.FilteredID,
 			Name:     dev.Description,
 			memInfo: memInfo{
 				TotalMemory: dev.TotalMemory,
 				FreeMemory:  dev.FreeMemory,
 			},
-			Library: dev.Library,
 			// TODO can we avoid variant
 			DependencyPath: dev.LibraryPath,
 			DriverMajor:    dev.DriverMajor,
 			DriverMinor:    dev.DriverMinor,
 		}
-		if dev.Library == "CUDA" || dev.Library == "HIP" {
+		if dev.Library == "CUDA" || dev.Library == "ROCm" {
 			info.MinimumMemory = 457 * format.MebiByte
 		}
-		if dev.Library == "HIP" {
+		if dev.Library == "ROCm" {
 			info.Compute = fmt.Sprintf("gfx%x%02x", dev.ComputeMajor, dev.ComputeMinor)
 			if rocmDir != "" {
 				info.DependencyPath = append(info.DependencyPath, rocmDir)
@@ -82,8 +83,10 @@ func devInfoToInfoList(devs []ml.DeviceInfo) GpuInfoList {
 
 		resp = append(resp, GpuInfo{
 			memInfo: mem,
+			DeviceID: ml.DeviceID{
 				Library: "cpu",
 				ID:      "0",
+			},
 		})
 	}
 	return resp
@@ -91,6 +94,8 @@ func devInfoToInfoList(devs []ml.DeviceInfo) GpuInfoList {
 
 // Given the list of GPUs this instantiation is targeted for,
 // figure out the visible devices environment variable
+//
+// If different libraries are detected, the first one is what we use
 func (l GpuInfoList) GetVisibleDevicesEnv() []string {
 	if len(l) == 0 {
 		return nil

discover/runner.go

@@ -115,18 +115,19 @@ func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.Dev
 		supported := make(map[string]map[string]map[string]int) // [Library][libDir][ID] = pre-deletion devices index
 		for i := range devices {
 			libDir := devices[i].LibraryPath[len(devices[i].LibraryPath)-1]
+			if devices[i].Library == "Metal" {
+				continue
+			}
 			slog.Debug("verifying GPU is supported", "library", libDir, "description", devices[i].Description, "compute", devices[i].Compute(), "pci_id", devices[i].PCIID)
 			wg.Add(1)
 			go func(i int) {
 				defer wg.Done()
 				var envVar string
-				if devices[i].Library == "HIP" {
 					if runtime.GOOS != "linux" {
 						envVar = "HIP_VISIBLE_DEVICES"
 					} else {
 						envVar = "ROCR_VISIBLE_DEVICES"
 					}
-				} else if devices[i].Library == "CUDA" {
 					envVar = "CUDA_VISIBLE_DEVICES"
 				} else if devices[i].Library == "VULKAN" {
 					envVar = "GGML_VK_VISIBLE_DEVICES"
@@ -165,7 +166,6 @@ func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.Dev
 				devices = append(devices[:i], devices[i+1:]...)
 				needsDelete = append(needsDelete[:i], needsDelete[i+1:]...)
 				i--
-			} else if devices[i].Library == "HIP" {
 				if _, err := strconv.Atoi(devices[i].ID); err == nil {
 					// Replace the numeric ID with the post-filtered IDs
 					devices[i].FilteredID = devices[i].ID
@@ -175,7 +175,6 @@ func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.Dev
 			}
 		}
 
-		// Now filter out any overlap with different libraries (favor CUDA/HIP over others)
 		for i := 0; i < len(devices); i++ {
 			for j := i + 1; j < len(devices); j++ {
 				// For this pass, we only drop exact duplicates
@@ -188,7 +187,7 @@ func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.Dev
 				case ml.DuplicateDevice:
 					// Different library, choose based on priority
 					var droppedDevice ml.DeviceInfo
-					if devices[i].Library == "CUDA" || devices[i].Library == "HIP" {
+					if devices[i].Library == "CUDA" || devices[i].Library == "ROCm" {
 						droppedDevice = devices[j]
 					} else {
 						droppedDevice = devices[i]
@@ -268,7 +267,7 @@ func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.Dev
 		devCheck:
 			for _, dev := range deviceIDs {
 				for i := range devices {
-					if dev.ID == devices[i].ID && dev.Library == devices[i].Library {
+					if dev == devices[i].DeviceID {
 						if !updated[i] {
 							skip = false
 							break devCheck
@@ -289,7 +288,7 @@ func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.Dev
 			slog.Debug("existing runner discovery took", "duration", time.Since(start))
 			for _, u := range updatedDevices {
 				for i := range devices {
-					if u.Library == devices[i].Library && u.ID == devices[i].ID {
+					if u.DeviceID == devices[i].DeviceID {
 						updated[i] = true
 						devices[i].FreeMemory = u.FreeMemory
 						break
@@ -313,7 +312,7 @@ func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.Dev
 				updatedDevices := bootstrapDevices(ctx, []string{LibOllamaPath, dir}, nil)
 				for _, u := range updatedDevices {
 					for i := range devices {
-						if u.Library == devices[i].Library && u.ID == devices[i].ID {
+						if u.DeviceID == devices[i].DeviceID {
 							updated[i] = true
 							devices[i].FreeMemory = u.FreeMemory
 							break
@@ -331,6 +330,9 @@ func GPUDevices(ctx context.Context, runners []FilteredRunnerDiscovery) []ml.Dev
 		}
 	}
 
+	// Apply any iGPU workarounds
+	iGPUWorkarounds(devices)
+
 	return devices
 }
 
@@ -435,9 +437,8 @@ func bootstrapDevices(ctx context.Context, ollamaLibDirs []string, extraEnvs []s
 
 	cmd := exec.Command(exe, params...)
 	cmd.Env = os.Environ()
+	if envconfig.LogLevel() == logutil.LevelTrace {
 		cmd.Stdout = os.Stdout
-	errBuf := &bytes.Buffer{}
-	if envconfig.LogLevel() == slog.Level(-8) {
 		cmd.Stderr = os.Stderr
 	} else {
 		cmd.Stderr = errBuf
@@ -473,7 +474,7 @@ func bootstrapDevices(ctx context.Context, ollamaLibDirs []string, extraEnvs []s
 			cmd.Env = append(cmd.Env, extraEnvs[i])
 		}
 	}
-	slog.Log(context.TODO(), logutil.LevelTrace, "starting runner for device discovery", "env", cmd.Env, "cmd", cmd)
+	logutil.Trace("starting runner for device discovery", "env", cmd.Env, "cmd", cmd)
 	if err := cmd.Start(); err != nil {
 		slog.Warn("unable to start discovery subprocess", "cmd", cmd, "error", err)
 		return nil
@@ -487,7 +488,7 @@ func bootstrapDevices(ctx context.Context, ollamaLibDirs []string, extraEnvs []s
 	if err != nil {
 		if cmd.ProcessState != nil && cmd.ProcessState.ExitCode() >= 0 {
 			// Expected during bootstrapping while we filter out unsupported AMD GPUs
-			slog.Log(context.TODO(), logutil.LevelTrace, "runner exited", "OLLAMA_LIBRARY_PATH", ollamaLibDirs, "extra_envs", extraEnvs, "code", cmd.ProcessState.ExitCode())
+			logutil.Trace("runner exited", "OLLAMA_LIBRARY_PATH", ollamaLibDirs, "extra_envs", extraEnvs, "code", cmd.ProcessState.ExitCode())
 		} else {
 			slog.Info("failure during GPU discovery", "OLLAMA_LIBRARY_PATH", ollamaLibDirs, "extra_envs", extraEnvs, "error", err)
 		}
@@ -544,3 +545,32 @@ func GetDevicesFromRunner(ctx context.Context, runner BaseRunner) ([]ml.DeviceIn
 		}
 	}
 }
+
+func iGPUWorkarounds(devices []ml.DeviceInfo) {
+	// short circuit if we have no iGPUs
+	anyiGPU := false
+	for i := range devices {
+		if devices[i].Integrated {
+			anyiGPU = true
+			break
+		}
+	}
+	if !anyiGPU {
+		return
+	}
+
+	memInfo, err := GetCPUMem()
+	if err != nil {
+		slog.Debug("failed to fetch system memory information for iGPU", "error", err)
+		return
+	}
+	for i := range devices {
+		if !devices[i].Integrated {
+			continue
+		}
+		// NVIDIA iGPUs return useless free VRAM data which ignores system buff/cache
+		if devices[i].Library == "CUDA" {
+			devices[i].FreeMemory = memInfo.FreeMemory
+		}
+	}
+}

discover/runner_test.go

@@ -0,0 +1,108 @@
+package discover
+
+import (
+	"testing"
+
+	"github.com/ollama/ollama/app/lifecycle"
+)
+
+func init() {
+	lifecycle.InitLogging()
+}
+
+func TestFilterOverlapByLibrary(t *testing.T) {
+	type testcase struct {
+		name string
+		inp  map[string]map[string]map[string]int
+		exp  []bool
+	}
+	for _, tc := range []testcase{
+		{
+			name: "empty",
+			inp:  map[string]map[string]map[string]int{},
+			exp:  []bool{}, // needs deletion
+		},
+		{
+			name: "single no overlap",
+			inp: map[string]map[string]map[string]int{
+				"CUDA": {
+					"cuda_v12": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 0,
+					},
+				},
+			},
+			exp: []bool{false},
+		},
+		{
+			name: "100% overlap pick 2nd",
+			inp: map[string]map[string]map[string]int{
+				"CUDA": {
+					"cuda_v12": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 0,
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 1,
+					},
+					"cuda_v13": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 2,
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 3,
+					},
+				},
+			},
+			exp: []bool{true, true, false, false},
+		},
+		{
+			name: "100% overlap pick 1st",
+			inp: map[string]map[string]map[string]int{
+				"CUDA": {
+					"cuda_v13": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 0,
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 1,
+					},
+					"cuda_v12": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 2,
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 3,
+					},
+				},
+			},
+			exp: []bool{false, false, true, true},
+		},
+		{
+			name: "partial overlap pick older",
+			inp: map[string]map[string]map[string]int{
+				"CUDA": {
+					"cuda_v13": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 0,
+					},
+					"cuda_v12": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 1,
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 2,
+					},
+				},
+			},
+			exp: []bool{true, false, false},
+		},
+		{
+			name: "no overlap",
+			inp: map[string]map[string]map[string]int{
+				"CUDA": {
+					"cuda_v13": {
+						"GPU-d7b00605-c0c8-152d-529d-e03726d5dc52": 0,
+					},
+					"cuda_v12": {
+						"GPU-cd6c3216-03d2-a8eb-8235-2ffbf571712e": 1,
+					},
+				},
+			},
+			exp: []bool{false, false},
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			needsDelete := make([]bool, len(tc.exp))
+			filterOverlapByLibrary(tc.inp, needsDelete)
+			for i, exp := range tc.exp {
+				if needsDelete[i] != exp {
+					t.Fatalf("expected: %v\ngot: %v", tc.exp, needsDelete)
+				}
+			}
+		})
+	}
+}

discover/types.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"log/slog"
 	"path/filepath"
+	"runtime"
 	"strings"
 
 	"github.com/ollama/ollama/format"
@@ -18,8 +19,8 @@ type memInfo struct {
 
 // Beginning of an `ollama info` command
 type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
+	ml.DeviceID
 	memInfo
-	Library string `json:"library,omitempty"`
 
 	// Optional variant to select (e.g. versions, cpu feature flags)
 	Variant string `json:"variant"`
@@ -36,11 +37,9 @@ type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
 	UnreliableFreeMemory bool
 
 	// GPU information
-	ID             string `json:"gpu_id"` // string to use for selection of this specific GPU
 	filterID string // AMD Workaround: The numeric ID of the device used to filter out other devices
 	Name     string `json:"name"`    // user friendly name if available
 	Compute  string `json:"compute"` // Compute Capability or gfx
-	FlashAttention bool   `json:"flash_attention"` // is flash attention supported
 
 	// Driver Information - TODO no need to put this on each GPU
 	DriverMajor int `json:"driver_major,omitempty"`
@@ -126,7 +125,6 @@ func LogDetails(devices []ml.DeviceInfo) {
 	// CPU inference
 	if len(devices) == 0 {
 		dev, _ := GetCPUMem()
-		// TODO more details about CPU
 		slog.Info("inference compute",
 			"id", "cpu",
 			"library", "cpu",
@@ -158,7 +156,8 @@ type SystemInfo struct {
 // Return the optimal number of threads to use for inference
 func (si SystemInfo) GetOptimalThreadCount() int {
 	if len(si.System.CPUs) == 0 {
-		return 0
+		// Fall back to Go's num CPU
+		return runtime.NumCPU()
 	}
 
 	coreCount := 0
@@ -173,10 +172,10 @@ func (si SystemInfo) GetOptimalThreadCount() int {
 func (l GpuInfoList) FlashAttentionSupported() bool {
 	for _, gpu := range l {
 		supportsFA := gpu.Library == "cpu" ||
-			gpu.Name == "Metal" ||
+			gpu.Name == "Metal" || gpu.Library == "Metal" ||
 			(gpu.Library == "CUDA" && gpu.DriverMajor >= 7) ||
-			gpu.Library == "HIP" ||
-			gpu.Library == "VULKAN"
+			gpu.Library == "ROCm"
+			gpu.Library == "Vulkan"
 
 		if !supportsFA {
 			return false

docs/cloud.md

@@ -0,0 +1,40 @@
+# Cloud
+
+| Ollama's cloud is currently in preview. For full documentation, see [Ollama's documentation](https://docs.ollama.com/cloud).
+
+## Cloud Models
+
+[Cloud models](https://ollama.com/cloud) are a new kind of model in Ollama that can run without a powerful GPU. Instead, cloud models are automatically offloaded to Ollama's cloud while offering the same capabilities as local models, making it possible to keep using your local tools while running larger models that wouldn’t fit on a personal computer.
+
+Ollama currently supports the following cloud models, with more coming soon:
+
+- `gpt-oss:20b-cloud`
+- `gpt-oss:120b-cloud`
+- `deepseek-v3.1:671b-cloud`
+- `qwen3-coder:480b-cloud`
+
+### Get started
+
+To run a cloud model, open the terminal and run:
+
+```
+ollama run gpt-oss:120b-cloud
+```
+
+To run cloud models with integrations that work with Ollama, first download the cloud model:
+
+```
+ollama pull qwen3-coder:480b-cloud
+```
+
+Then sign in to Ollama:
+
+```
+ollama signin
+```
+
+Finally, access the model using the model name `qwen3-coder:480b-cloud` via Ollama's local API or tooling.
+
+## Cloud API access
+
+Cloud models can also be accessed directly on ollama.com's API. For more information, see the [docs](https://docs.ollama.com/cloud).

docs/gpu.md

@@ -52,13 +52,13 @@ Ollama supports the following AMD GPUs:
 
 ### Linux Support
 | Family         | Cards and accelerators                                                                                                                   |
-| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
-| AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` `Vega 64` `Vega 56`    |
-| AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` `V420` `V340` `V320` `Vega II Duo` `Vega II` `VII` `SSG` |
-| AMD Instinct   | `MI300X` `MI300A` `MI300` `MI250X` `MI250` `MI210` `MI200` `MI100` `MI60` `MI50`                                                               |
+| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------- |
+| AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` `Vega 64`            |
+| AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` `V420` `V340` `V320` `Vega II Duo` `Vega II` `SSG` |
+| AMD Instinct   | `MI300X` `MI300A` `MI300` `MI250X` `MI250` `MI210` `MI200` `MI100` `MI60`                                                                |
 
 ### Windows Support
-With ROCm v6.1, the following GPUs are supported on Windows.
+With ROCm v6.2, the following GPUs are supported on Windows.
 
 | Family         | Cards and accelerators                                                                                                               |
 | -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
@@ -88,8 +88,6 @@ At this time, the known supported GPU types on linux are the following LLVM Targ
 This table shows some example GPUs that map to these LLVM targets:
 | **LLVM Target** | **An Example GPU** |
 |-----------------|---------------------|
-| gfx900 | Radeon RX Vega 56 |
-| gfx906 | Radeon Instinct MI50 |
 | gfx908 | Radeon Instinct MI100 |
 | gfx90a | Radeon Instinct MI210 |
 | gfx940 | Radeon Instinct MI300 |

docs/macos.md

@@ -2,7 +2,7 @@
 
 ## System Requirements
 
-* MacOS Monterey (v12) or newer
+* MacOS Sonoma (v14) or newer
 * Apple M series (CPU and GPU support) or x86 (CPU only)
 
 

docs/turbo.md

@@ -1,107 +0,0 @@
-# Turbo
-
-> ⚠️ Turbo is preview
-
-Ollama’s [Turbo](https://ollama.com/turbo) is a new way to run open-source models with acceleration from datacenter-grade hardware.
-
-Currently, the following models are available in Turbo:
-
-- `gpt-oss:20b`
-- `gpt-oss:120b`
-
-## Get started
-
-### Ollama for macOS & Windows
-
-Download Ollama
-
-- Select a model such as `gpt-oss:20b` or `gpt-oss:120b`
-- Click on **Turbo**. You’ll be prompted to create an account or sign in
-
-### Ollama’s CLI
-
-- [Sign up](https://ollama.com/signup) for an Ollama account
-- Add your Ollama key [to ollama.com](https://ollama.com/settings/keys).
-
-  On macOS and Linux:
-
-  ```shell
-  cat ~/.ollama/id_ed25519.pub
-  ```
-
-  On Windows:
-
-  ```
-  type "%USERPROFILE%\.ollama\id_ed25519.pub"
-  ```
-
-- Then run a model setting `OLLAMA_HOST` to `ollama.com`:
-  ```shell
-  OLLAMA_HOST=ollama.com ollama run gpt-oss:120b
-  ```
-
-### Ollama’s Python library
-
-- Download Ollama's [Python library](https://github.com/ollama/ollama-python)
-- [Sign up](https://ollama.com/signup) for an Ollama account
-- Create an API key by visiting https://ollama.com/settings/keys
-
-```python
-from ollama import Client
-
-client = Client(
-    host="https://ollama.com",
-    headers={'Authorization': '<api key>'}
-)
-
-messages = [
-  {
-    'role': 'user',
-    'content': 'Why is the sky blue?',
-  },
-]
-
-for part in client.chat('gpt-oss:120b', messages=messages, stream=True):
-  print(part['message']['content'], end='', flush=True)
-```
-
-### Ollama’s JavaScript library
-
-- Download Ollama's [JavaScript library](https://github.com/ollama/ollama-js)
-- [Sign up](https://ollama.com/signup) for an Ollama account
-- Create an API key by visiting https://ollama.com/settings/keys
-
-```typescript
-import { Ollama } from 'ollama';
-
-const ollama = new Ollama({
-  host: 'https://ollama.com',
-  headers: {
-	  Authorization: "Bearer <api key>"
-  }
-});
-
-const response = await ollama.chat({
-  model: 'gpt-oss:120b',
-  messages: [{ role: 'user', content: 'Explain quantum computing' }],
-  stream: true
-});
-
-for await (const part of response) {
-    process.stdout.write(part.message.content)
-}
-```
-
-### Community integrations
-
-Turbo mode is also compatible with several community integrations.
-
-#### Open WebUI
-
-- Go to **settings** → **Admin settings** → **Connections**
-- Under **Ollama API,** click **+**
-- For the **URL** put `https://ollama.com`
-- For the **API key,** create an API key on https://ollama.com/settings/keys and add it.
-- Click **Save**
-
-Now, if you navigate to the model selector, Turbo models should be available under **External**.

envconfig/config.go

@@ -145,8 +145,8 @@ func Remotes() []string {
 	return r
 }
 
-func Bool(k string) func() bool {
-	return func() bool {
+func BoolWithDefault(k string) func(defaultValue bool) bool {
+	return func(defaultValue bool) bool {
 		if s := Var(k); s != "" {
 			b, err := strconv.ParseBool(s)
 			if err != nil {
@@ -156,7 +156,14 @@ func Bool(k string) func() bool {
 			return b
 		}
 
-		return false
+		return defaultValue
+	}
+}
+
+func Bool(k string) func() bool {
+	withDefault := BoolWithDefault(k)
+	return func() bool {
+		return withDefault(false)
 	}
 }
 
@@ -177,7 +184,7 @@ func LogLevel() slog.Level {
 
 var (
 	// FlashAttention enables the experimental flash attention feature.
-	FlashAttention = Bool("OLLAMA_FLASH_ATTENTION")
+	FlashAttention = BoolWithDefault("OLLAMA_FLASH_ATTENTION")
 	// KvCacheType is the quantization type for the K/V cache.
 	KvCacheType = String("OLLAMA_KV_CACHE_TYPE")
 	// NoHistory disables readline history.
@@ -264,7 +271,7 @@ 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_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(), "Enabled flash attention"},
+		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(false), "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)"},
 		"OLLAMA_HOST":              {"OLLAMA_HOST", Host(), "IP Address for the ollama server (default 127.0.0.1:11434)"},

fs/ggml/ggml.go

@@ -244,6 +244,7 @@ func (kv KV) OllamaEngineRequired() bool {
 		"gemma3n",
 		"mistral3",
 		"qwen3",
+		"qwen3moe",
 		"llama4",
 		"mllama",
 		"qwen25vl",
@@ -869,11 +870,6 @@ func (f GGML) SupportsKVCacheType(cacheType string) bool {
 		return true
 	}
 
-	if arch := f.KV().Architecture(); slices.Contains([]string{"gptoss", "gpt-oss"}, arch) {
-		// gpt-oss uses attention with sinks which does not support quantized cache types
-		slog.Warn("model only supports non-quantized cache types", "model", arch)
-		return false
-	}
 	return slices.Contains([]string{"q8_0", "q4_0"}, cacheType)
 }
 
@@ -898,6 +894,8 @@ func (f GGML) SupportsFlashAttention() bool {
 func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
 		"gptoss", "gpt-oss",
+		"qwen3",
+		"qwen3moe",
 	}, f.KV().String("general.architecture"))
 }
 

harmony/harmonyparser.go

@@ -1,6 +1,7 @@
 package harmony
 
 import (
+	"encoding/json"
 	"fmt"
 	"log/slog"
 	"strings"
@@ -265,6 +266,8 @@ type HarmonyMessageHandler struct {
 	state           harmonyMessageState
 	HarmonyParser   *HarmonyParser
 	FunctionNameMap *FunctionNameMap
+	toolAccumulator *HarmonyToolCallAccumulator
+	convertedTools  map[string]struct{}
 }
 
 // NewHarmonyMessageHandler creates a new message handler
@@ -277,6 +280,7 @@ func NewHarmonyMessageHandler() *HarmonyMessageHandler {
 			HeaderEndTag:    "<|message|>",
 		},
 		FunctionNameMap: NewFunctionNameMap(),
+		convertedTools:  make(map[string]struct{}),
 	}
 }
 
@@ -384,8 +388,85 @@ func NewFunctionNameMap() *FunctionNameMap {
 	}
 }
 
+// Init initializes the handler with tools and optional last message
+// Implements the Parser interface
+func (h *HarmonyMessageHandler) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
+	// Initialize the harmony parser
+	if h.HarmonyParser == nil {
+		h.HarmonyParser = &HarmonyParser{
+			MessageStartTag: "<|start|>",
+			MessageEndTag:   "<|end|>",
+			HeaderEndTag:    "<|message|>",
+		}
+	}
+
+	// Handle prefill for chat mode
+	if lastMessage != nil {
+		h.HarmonyParser.AddImplicitStartOrPrefill(lastMessage)
+	} else {
+		h.HarmonyParser.AddImplicitStart()
+	}
+
+	// Initialize tool accumulator
+	h.toolAccumulator = h.CreateToolParser()
+
+	// Process tools and return renamed versions
+	if len(tools) == 0 {
+		return tools
+	}
+
+	processedTools := make([]api.Tool, len(tools))
+	copy(processedTools, tools)
+	for i, tool := range processedTools {
+		if tool.Function.Name != "" {
+			processedTools[i].Function.Name = h.FunctionNameMap.ConvertAndAdd(tool.Function.Name)
+			h.convertedTools[tool.Function.Name] = struct{}{}
+		}
+	}
+	return processedTools
+}
+
+// Add implements the Parser interface - processes streamed content and extracts content, thinking, and tool calls
+func (h *HarmonyMessageHandler) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
+	content, thinking, toolContent := h.AddContent(s, h.toolAccumulator)
+	if toolContent != "" {
+		h.toolAccumulator.Add(toolContent)
+	}
+
+	// tool calls always happen one at a time, and always at the end of a message,
+	// so for simplicity we defer parsing them until we know we're done
+	if done {
+		toolName, raw := h.toolAccumulator.Drain()
+		if toolName != nil {
+			name := strings.TrimPrefix(*toolName, "functions.")
+			name = h.FunctionNameMap.OriginalFromConverted(name)
+			var args api.ToolCallFunctionArguments
+			if err := json.Unmarshal([]byte(raw), &args); err != nil {
+				return "", "", nil, fmt.Errorf("error parsing tool call: raw='%s', err=%w", raw, err)
+			}
+			calls = append(calls, api.ToolCall{Function: api.ToolCallFunction{Name: name, Arguments: args}})
+		}
+	}
+
+	return content, thinking, calls, nil
+}
+
+// HasToolSupport implements the Parser interface
+func (h *HarmonyMessageHandler) HasToolSupport() bool {
+	return true
+}
+
+// HasThinkingSupport implements the Parser interface
+func (h *HarmonyMessageHandler) HasThinkingSupport() bool {
+	return true
+}
+
 func (m *FunctionNameMap) ConvertAndAdd(userFunctionName string) string {
 	harmonyFunctionName := m.deriveName(userFunctionName)
+	// built-in functions should not be renamed
+	if userFunctionName == "browser.open" || userFunctionName == "browser.search" || userFunctionName == "browser.find" || userFunctionName == "python" {
+		harmonyFunctionName = userFunctionName
+	}
 	m.userToHarmony[userFunctionName] = harmonyFunctionName
 	m.harmonyToUser[harmonyFunctionName] = userFunctionName
 	return harmonyFunctionName

harmony/harmonyparser_test.go

@@ -513,6 +513,7 @@ func TestFunctionConvertAndAdd(t *testing.T) {
 		{name: "dupes from different user-specified names", in: []string{"get weather", "get_weather", "get-weather"}, want: []string{"get_weather", "get_weather_2", "get_weather_3"}},
 		{name: "non dupes after dupes", in: []string{"get weather", "get_weather", "get-weather", "something-different"}, want: []string{"get_weather", "get_weather_2", "get_weather_3", "something_different"}},
 		{name: "multiple sets of dupes", in: []string{"a", "a", "b", "a", "a", "b", "a"}, want: []string{"a", "a_2", "b", "a_3", "a_4", "b_2", "a_5"}},
+		{name: "built-in functions should not be renamed", in: []string{"browser.open", "python", "not.a.built-in.function", "browser.not_a_real_built_in"}, want: []string{"browser.open", "python", "not_a_built_in_function", "browser_not_a_real_built_in"}},
 	}
 
 	for i, tt := range tests {

integration/README.md

@@ -12,3 +12,6 @@ The integration tests have 2 modes of operating.
 
 > [!IMPORTANT]
 > Before running the tests locally without the "test existing" setting, compile ollama from the top of the source tree  `go build .` in addition to GPU support with cmake if applicable on your platform.  The integration tests expect to find an ollama binary at the top of the tree.
+
+
+Many tests use a default small model suitable to run on many systems.  You can override this default model by setting `OLLAMA_TEST_DEFAULT_MODEL`

integration/api_test.go

@@ -22,13 +22,12 @@ func TestAPIGenerate(t *testing.T) {
 	// Set up the test data
 	req := api.GenerateRequest{
 		Model:  smol,
-		Prompt: "why is the sky blue? be brief",
+		Prompt: blueSkyPrompt,
 		Options: map[string]interface{}{
 			"temperature": 0,
 			"seed":        123,
 		},
 	}
-	anyResp := []string{"rayleigh", "scattering"}
 
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
@@ -120,14 +119,14 @@ func TestAPIGenerate(t *testing.T) {
 				// Verify the response contains the expected data
 				response := buf.String()
 				atLeastOne := false
-				for _, resp := range anyResp {
+				for _, resp := range blueSkyExpected {
 					if strings.Contains(strings.ToLower(response), resp) {
 						atLeastOne = true
 						break
 					}
 				}
 				if !atLeastOne {
-					t.Errorf("none of %v found in %s", anyResp, response)
+					t.Errorf("none of %v found in %s", blueSkyExpected, response)
 				}
 			case <-ctx.Done():
 				t.Error("outer test context done while waiting for generate")
@@ -181,7 +180,7 @@ func TestAPIChat(t *testing.T) {
 		Messages: []api.Message{
 			{
 				Role:    "user",
-				Content: "why is the sky blue?  be brief",
+				Content: blueSkyPrompt,
 			},
 		},
 		Options: map[string]interface{}{
@@ -189,7 +188,6 @@ func TestAPIChat(t *testing.T) {
 			"seed":        123,
 		},
 	}
-	anyResp := []string{"rayleigh", "scattering"}
 
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
@@ -279,14 +277,14 @@ func TestAPIChat(t *testing.T) {
 				// Verify the response contains the expected data
 				response := buf.String()
 				atLeastOne := false
-				for _, resp := range anyResp {
+				for _, resp := range blueSkyExpected {
 					if strings.Contains(strings.ToLower(response), resp) {
 						atLeastOne = true
 						break
 					}
 				}
 				if !atLeastOne {
-					t.Errorf("none of %v found in %s", anyResp, response)
+					t.Errorf("none of %v found in %s", blueSkyExpected, response)
 				}
 			case <-ctx.Done():
 				t.Error("outer test context done while waiting for chat")

integration/basic_test.go

@@ -19,14 +19,14 @@ func TestBlueSky(t *testing.T) {
 	// Set up the test data
 	req := api.GenerateRequest{
 		Model:  smol,
-		Prompt: "why is the sky blue?",
+		Prompt: blueSkyPrompt,
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
+	GenerateTestHelper(ctx, t, req, blueSkyExpected)
 }
 
 func TestUnicode(t *testing.T) {
@@ -110,12 +110,12 @@ func TestUnicodeModelDir(t *testing.T) {
 
 	req := api.GenerateRequest{
 		Model:  smol,
-		Prompt: "why is the sky blue?",
+		Prompt: blueSkyPrompt,
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
+	GenerateTestHelper(ctx, t, req, blueSkyExpected)
 }

integration/context_test.go

@@ -36,7 +36,7 @@ func TestLongInputContext(t *testing.T) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("PullIfMissing failed: %v", err)
 	}
-	DoGenerate(ctx, t, client, req, []string{"russia", "germany", "france", "england", "austria", "prussia", "europe", "individuals", "coalition", "conflict"}, 120*time.Second, 10*time.Second)
+	DoGenerate(ctx, t, client, req, []string{"russia", "german", "france", "england", "austria", "prussia", "europe", "individuals", "coalition", "conflict"}, 120*time.Second, 10*time.Second)
 }
 
 func TestContextExhaustion(t *testing.T) {
@@ -63,11 +63,11 @@ func TestContextExhaustion(t *testing.T) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("PullIfMissing failed: %v", err)
 	}
-	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived", "sunny", "cloudy", "clear", "water"}, 120*time.Second, 10*time.Second)
+	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived", "sunny", "cloudy", "clear", "water", "time", "travel", "world"}, 120*time.Second, 10*time.Second)
 }
 
 // Send multiple generate requests with prior context and ensure the response is coherant and expected
-func TestGenerateWithHistory(t *testing.T) {
+func TestParallelGenerateWithHistory(t *testing.T) {
 	modelOverride := ollamaEngineChatModels[0] // Most recent ollama engine model
 	req, resp := GenerateRequests()
 	numParallel := 2
@@ -113,8 +113,48 @@ func TestGenerateWithHistory(t *testing.T) {
 	wg.Wait()
 }
 
+// Send generate requests with prior context and ensure the response is coherant and expected
+func TestGenerateWithHistory(t *testing.T) {
+	req := api.GenerateRequest{
+		Model:     smol,
+		Prompt:    rainbowPrompt,
+		Stream:    &stream,
+		KeepAlive: &api.Duration{Duration: 10 * time.Second},
+		Options: map[string]any{
+			"num_ctx": 16384,
+		},
+	}
+
+	softTimeout, hardTimeout := getTimeouts(t)
+	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	// Get the server running (if applicable) warm the model up with a single initial request
+	slog.Info("loading", "model", req.Model)
+	err := client.Generate(ctx,
+		&api.GenerateRequest{Model: req.Model, KeepAlive: &api.Duration{Duration: 10 * time.Second}, Options: req.Options},
+		func(response api.GenerateResponse) error { return nil },
+	)
+	if err != nil {
+		t.Fatalf("failed to load model %s: %s", req.Model, err)
+	}
+
+	req.Context = DoGenerate(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
+
+	for i := 0; i < len(rainbowFollowups); i++ {
+		req.Prompt = rainbowFollowups[i]
+		if time.Now().Sub(started) > softTimeout {
+			slog.Info("exceeded soft timeout, winding down test")
+			return
+		}
+		req.Context = DoGenerate(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
+	}
+}
+
 // Send multiple chat requests with prior context and ensure the response is coherant and expected
-func TestChatWithHistory(t *testing.T) {
+func TestParallelChatWithHistory(t *testing.T) {
 	modelOverride := ollamaEngineChatModels[0] // Most recent ollama engine model
 	req, resp := ChatRequests()
 	numParallel := 2
@@ -164,3 +204,55 @@ func TestChatWithHistory(t *testing.T) {
 	}
 	wg.Wait()
 }
+
+// Send generate requests with prior context and ensure the response is coherant and expected
+func TestChatWithHistory(t *testing.T) {
+	req := api.ChatRequest{
+		Model:     smol,
+		Stream:    &stream,
+		KeepAlive: &api.Duration{Duration: 10 * time.Second},
+		Options: map[string]any{
+			"num_ctx": 16384,
+		},
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: rainbowPrompt,
+			},
+		},
+	}
+
+	softTimeout, hardTimeout := getTimeouts(t)
+	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	// Get the server running (if applicable) warm the model up with a single initial request
+	slog.Info("loading", "model", req.Model)
+	err := client.Generate(ctx,
+		&api.GenerateRequest{Model: req.Model, KeepAlive: &api.Duration{Duration: 10 * time.Second}, Options: req.Options},
+		func(response api.GenerateResponse) error { return nil },
+	)
+	if err != nil {
+		t.Fatalf("failed to load model %s: %s", req.Model, err)
+	}
+
+	assistant := DoChat(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
+
+	for i := 0; i < len(rainbowFollowups); i++ {
+		if time.Now().Sub(started) > softTimeout {
+			slog.Info("exceeded soft timeout, winding down test")
+			return
+		}
+		req.Messages = append(req.Messages,
+			*assistant,
+			api.Message{Role: "user", Content: rainbowFollowups[i]},
+		)
+
+		assistant = DoChat(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
+		if assistant == nil {
+			t.Fatalf("didn't get an assistant response for context")
+		}
+	}
+}

integration/library_models_test.go

@@ -4,7 +4,9 @@ package integration
 
 import (
 	"context"
+	"fmt"
 	"log/slog"
+	"os"
 	"testing"
 	"time"
 
@@ -20,6 +22,7 @@ func TestLibraryModelsGenerate(t *testing.T) {
 	defer cancel()
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
+	targetArch := os.Getenv("OLLAMA_TEST_ARCHITECTURE")
 
 	chatModels := libraryChatModels
 	for _, model := range chatModels {
@@ -30,16 +33,26 @@ func TestLibraryModelsGenerate(t *testing.T) {
 			if err := PullIfMissing(ctx, client, model); err != nil {
 				t.Fatalf("pull failed %s", err)
 			}
+			if targetArch != "" {
+				resp, err := client.Show(ctx, &api.ShowRequest{Name: model})
+				if err != nil {
+					t.Fatalf("unable to show model: %s", err)
+				}
+				arch := resp.ModelInfo["general.architecture"].(string)
+				if arch != targetArch {
+					t.Skip(fmt.Sprintf("Skipping %s architecture %s != %s", model, arch, targetArch))
+				}
+			}
 			req := api.GenerateRequest{
 				Model:     model,
-				Prompt:    "why is the sky blue?",
+				Prompt:    blueSkyPrompt,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"temperature": 0.1,
 					"seed":        123,
 				},
 			}
-			anyResp := []string{"rayleigh", "scatter", "atmosphere", "nitrogen", "oxygen", "wavelength"}
+			anyResp := blueSkyExpected
 			// Special cases
 			if model == "duckdb-nsql" {
 				anyResp = []string{"select", "from"}

integration/model_arch_test.go

@@ -68,14 +68,13 @@ func TestModelsGenerate(t *testing.T) {
 			// TODO - fiddle with context size
 			req := api.GenerateRequest{
 				Model:  model,
-				Prompt: "why is the sky blue?",
+				Prompt: blueSkyPrompt,
 				Options: map[string]interface{}{
 					"temperature": 0,
 					"seed":        123,
 				},
 			}
-			anyResp := []string{"rayleigh", "scattering", "atmosphere", "nitrogen", "oxygen"}
-			DoGenerate(ctx, t, client, req, anyResp, 120*time.Second, 30*time.Second)
+			DoGenerate(ctx, t, client, req, blueSkyExpected, 120*time.Second, 30*time.Second)
 		})
 	}
 }

integration/model_perf_test.go

@@ -40,6 +40,18 @@ var (
 // cat int.log | grep MODEL_PERF_HEADER | head -1| cut -f2- -d: > perf.csv
 // cat int.log | grep MODEL_PERF_DATA | cut -f2- -d: >> perf.csv
 func TestModelsPerf(t *testing.T) {
+	if s := os.Getenv("OLLAMA_NEW_ENGINE"); s != "" {
+		doModelPerfTest(t, ollamaEngineChatModels)
+	} else {
+		doModelPerfTest(t, append(ollamaEngineChatModels, llamaRunnerChatModels...))
+	}
+}
+
+func TestLibraryModelsPerf(t *testing.T) {
+	doModelPerfTest(t, libraryChatModels)
+}
+
+func doModelPerfTest(t *testing.T, chatModels []string) {
 	softTimeout, hardTimeout := getTimeouts(t)
 	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
 	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
@@ -65,14 +77,12 @@ func TestModelsPerf(t *testing.T) {
 	}
 	longPrompt := "summarize the following: " + string(data)
 
-	var chatModels []string
-	if s := os.Getenv("OLLAMA_NEW_ENGINE"); s != "" {
-		chatModels = ollamaEngineChatModels
-	} else {
-		chatModels = append(ollamaEngineChatModels, llamaRunnerChatModels...)
-	}
+	targetArch := os.Getenv("OLLAMA_TEST_ARCHITECTURE")
 
 	for _, model := range chatModels {
+		if !strings.Contains(model, ":") {
+			model = model + ":latest"
+		}
 		t.Run(model, func(t *testing.T) {
 			if time.Now().Sub(started) > softTimeout {
 				t.Skip("skipping remaining tests to avoid excessive runtime")
@@ -88,6 +98,9 @@ func TestModelsPerf(t *testing.T) {
 			}
 			arch := resp.ModelInfo["general.architecture"].(string)
 			maxContext = int(resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))
+			if targetArch != "" && arch != targetArch {
+				t.Skip(fmt.Sprintf("Skipping %s architecture %s != %s", model, arch, targetArch))
+			}
 
 			if maxVram > 0 {
 				resp, err := client.List(ctx)
@@ -151,8 +164,8 @@ func TestModelsPerf(t *testing.T) {
 					prompt  string
 					anyResp []string
 				}{
-					{"why is the sky blue?", []string{"rayleigh", "scattering", "atmosphere", "nitrogen", "oxygen"}},
-					{maxPrompt, []string{"shakespeare", "oppression", "sorrows", "gutenberg", "child", "license", "sonnet", "melancholy"}},
+					{blueSkyPrompt, blueSkyExpected},
+					{maxPrompt, []string{"shakespeare", "oppression", "sorrows", "gutenberg", "child", "license", "sonnet", "melancholy", "love", "sorrow", "beauty"}},
 				}
 				var gpuPercent int
 				for _, tc := range testCases {
@@ -241,11 +254,12 @@ func TestModelsPerf(t *testing.T) {
 							}
 						}
 					}
+					// Round the logged prompt count for comparisons across versions/configurations which can vary slightly
 					fmt.Fprintf(os.Stderr, "MODEL_PERF_HEADER:%s,%s,%s,%s,%s,%s,%s\n",
 						"MODEL",
 						"CONTEXT",
 						"GPU PERCENT",
-						"PROMPT COUNT",
+						"APPROX PROMPT COUNT",
 						"LOAD TIME",
 						"PROMPT EVAL TPS",
 						"EVAL TPS",
@@ -254,7 +268,7 @@ func TestModelsPerf(t *testing.T) {
 						model,
 						numCtx,
 						gpuPercent,
-						resp.PromptEvalCount,
+						(resp.PromptEvalCount/10)*10,
 						float64(resp.LoadDuration)/1000000000.0,
 						float64(resp.PromptEvalCount)/(float64(resp.PromptEvalDuration)/1000000000.0),
 						float64(resp.EvalCount)/(float64(resp.EvalDuration)/1000000000.0),

integration/quantization_test.go

@@ -76,7 +76,7 @@ func TestQuantization(t *testing.T) {
 				stream := true
 				genReq := api.GenerateRequest{
 					Model:     newName,
-					Prompt:    "why is the sky blue?",
+					Prompt:    blueSkyPrompt,
 					KeepAlive: &api.Duration{Duration: 3 * time.Second},
 					Options: map[string]any{
 						"seed":        42,
@@ -88,14 +88,13 @@ func TestQuantization(t *testing.T) {
 
 				// Some smaller quantizations can cause models to have poor quality
 				// or get stuck in repetition loops, so we stop as soon as we have any matches
-				anyResp := []string{"rayleigh", "scattering", "day", "sun", "moon", "color", "nitrogen", "oxygen"}
 				reqCtx, reqCancel := context.WithCancel(ctx)
 				atLeastOne := false
 				var buf bytes.Buffer
 				genfn := func(response api.GenerateResponse) error {
 					buf.Write([]byte(response.Response))
 					fullResp := strings.ToLower(buf.String())
-					for _, resp := range anyResp {
+					for _, resp := range blueSkyExpected {
 						if strings.Contains(fullResp, resp) {
 							atLeastOne = true
 							t.Log(fullResp)

integration/utils_test.go

@@ -256,13 +256,28 @@ var (
 		"snowflake-arctic-embed",
 		"snowflake-arctic-embed2",
 	}
+
+	blueSkyPrompt   = "why is the sky blue? Be brief but factual in your reply"
+	blueSkyExpected = []string{"rayleigh", "scatter", "atmosphere", "nitrogen", "oxygen", "wavelength", "interact"}
+
+	rainbowPrompt    = "how do rainbows form? Be brief but factual in your reply"
+	rainbowFollowups = []string{
+		"Explain the physics involved in them.  Be breif in your reply",
+		"Explain the chemistry involved in them.  Be breif in your reply",
+		"What are common myths related to them? Be brief in your reply",
+		"What are common fairytales related to them? Be brief in your reply",
+		"Can they form if there is no rain?  Be breif in your reply",
+		"Can they form if there are no clouds?  Be breif in your reply",
+		"Do they happen on other planets? Be brief in your reply",
+	}
+	rainbowExpected = []string{"water", "droplet", "mist", "glow", "refract", "reflect", "scatter", "wave", "color", "spectrum", "raindrop", "atmosphere", "frequency", "end", "gold", "fortune", "blessing", "prosperity", "magic", "shower", "sky", "shimmer", "light", "storm", "sunny"}
 )
 
 func init() {
 	lifecycle.InitLogging()
-	custom := os.Getenv("OLLAMA_TEST_SMOL_MODEL")
+	custom := os.Getenv("OLLAMA_TEST_DEFAULT_MODEL")
 	if custom != "" {
-		slog.Info("setting smol test model to " + custom)
+		slog.Info("setting default test model to " + custom)
 		smol = custom
 	}
 }
@@ -579,7 +594,7 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 			}, {
 				Model:     smol,
-				Prompt:    "how do rainbows form? Be brief but factual in your reply",
+				Prompt:    rainbowPrompt,
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 			}, {
@@ -595,11 +610,11 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 			},
 		},
 		[][]string{
-			{"sunlight", "scattering", "interact", "color", "surface", "depth", "red", "orange", "yellow", "absorbs", "wavelength"},
-			{"soil", "organic", "earth", "black", "tan", "chemical", "processes", "pigments", "particles", "iron oxide", "rust", "air", "water", "mixture", "mixing"},
-			{"water", "droplet", "refracted", "reflect", "color", "spectrum"},
+			{"sunlight", "scatter", "interact", "color", "surface", "depth", "red", "orange", "yellow", "absorb", "wavelength", "water", "molecule"},
+			{"soil", "organic", "earth", "black", "tan", "chemical", "processes", "pigment", "particle", "iron oxide", "rust", "air", "water", "wet", "mixture", "mixing", "mineral", "element", "decomposed", "matter", "wavelength"},
+			rainbowExpected,
 			{"fourth", "july", "declaration", "independence"},
-			{"nitrogen", "oxygen", "carbon", "dioxide", "water", "vapor"},
+			{"nitrogen", "oxygen", "carbon", "dioxide", "water", "vapor", "fluid", "particles", "gas"},
 		}
 }
 

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "e54d41befcc1575f4c898c5ff4ef43970cead75f";
+char const *LLAMA_COMMIT = "364a7a6d4a786e98947c8a90430ea581213c0ba9";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/common/common.cpp

@@ -14,6 +14,7 @@
 #include <climits>
 #include <cmath>
 #include <codecvt>
+#include <chrono>
 #include <cstdarg>
 #include <cstring>
 #include <ctime>
@@ -41,6 +42,7 @@
 #endif
 #include <locale>
 #include <windows.h>
+#include <string.h>
 #include <fcntl.h>
 #include <io.h>
 #else
@@ -49,6 +51,11 @@
 #include <unistd.h>
 #endif
 
+#if defined(__linux__)
+#include <sys/types.h>
+#include <pwd.h>
+#endif
+
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
@@ -557,13 +564,6 @@ std::string string_from(const struct llama_context * ctx, const std::vector<llam
 
         auto detokenized = common_token_to_piece(ctx, token);
 
-        detokenized.erase(
-            std::remove_if(
-                detokenized.begin(),
-                detokenized.end(),
-                [](const unsigned char c) { return !std::isprint(c); }),
-            detokenized.end());
-
         buf << "'" << detokenized << "'"
             << ":" << std::to_string(token);
     }
@@ -588,13 +588,6 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 
         auto detokenized = common_token_to_piece(ctx, batch.token[i]);
 
-        detokenized.erase(
-                std::remove_if(
-                    detokenized.begin(),
-                    detokenized.end(),
-                    [](const unsigned char c) { return !std::isprint(c); }),
-                detokenized.end());
-
         buf << "\n"          << std::to_string(i)
             << ", token '"   << detokenized << "'"
             << ", pos "      << std::to_string(batch.pos[i])
@@ -877,8 +870,20 @@ std::string fs_get_cache_directory() {
 #if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
-        } else {
+        } else if (std::getenv("HOME")) {
             cache_directory = std::getenv("HOME") + std::string("/.cache/");
+        } else {
+#if defined(__linux__)
+            /* no $HOME is defined, fallback to getpwuid */
+            struct passwd *pw = getpwuid(getuid());
+            if ((!pw) || (!pw->pw_dir)) {
+                throw std::runtime_error("Failed to find $HOME directory");
+            }
+
+            cache_directory = std::string(pw->pw_dir) + std::string("/.cache/");
+#else /* defined(__linux__) */
+            throw std::runtime_error("Failed to find $HOME directory");
+#endif /* defined(__linux__) */
         }
 #elif defined(__APPLE__)
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
@@ -914,7 +919,8 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
     if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
+        LOG_ERR("%s: failed to load model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+            __func__, params.model.path.c_str());
         return iparams;
     }
 
@@ -924,7 +930,8 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
+        LOG_ERR("%s: failed to create context with model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+            __func__, params.model.path.c_str());
         llama_model_free(model);
         return iparams;
     }
@@ -971,15 +978,13 @@ struct common_init_result common_init_from_params(common_params & params) {
 
         bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
         bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
+        bool has_rerank_prompt = llama_model_chat_template(model, "rerank") != NULL;
 
-        if (!has_eos && !has_sep) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
+        if (!has_eos && !has_sep && !has_rerank_prompt) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, SEP token, or rerank prompt. Reranking will not work\n", __func__);
             ok = false;
         } else if (!has_eos) {
             LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
-        } else if (!has_sep) {
-            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
-            ok = false;
         }
 
         if (!ok) {
@@ -1001,7 +1006,12 @@ struct common_init_result common_init_from_params(common_params & params) {
             return iparams;
         }
 
+        char buf[1024];
         la.ptr = lora.get();
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
+        la.task_name = buf;
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
+        la.prompt_prefix = buf;
         iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
     }
 
@@ -1165,11 +1175,10 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
     cparams.pooling_type      = params.pooling_type;
     cparams.attention_type    = params.attention_type;
-    cparams.defrag_thold      = params.defrag_thold;
+    cparams.flash_attn_type   = params.flash_attn_type;
     cparams.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;
-    cparams.flash_attn        = params.flash_attn;
     cparams.no_perf           = params.no_perf;
     cparams.op_offload        = !params.no_op_offload;
     cparams.swa_full          = params.swa_full;
@@ -1565,3 +1574,56 @@ ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std
 
     return result;
 }
+
+ggml_opt_optimizer_params common_opt_lr_pars(void * userdata) {
+    ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(nullptr);
+    const lr_opt &            d      = *(lr_opt *) userdata;
+    result.adamw.alpha = result.sgd.alpha = d.get_lr(d.epoch);
+    result.sgd.wd = result.adamw.wd = d.wd;
+    return result;
+}
+
+// TODO make all command line args case-insensitive
+static inline bool eq_case_insensitive(char const* a, char const* b) {
+    return !
+#if defined(_MSC_VER)
+        _stricmp
+#else
+        strcasecmp
+#endif // defined(_MSC_VER)
+        (a, b);
+}
+
+enum ggml_opt_optimizer_type common_opt_get_optimizer(const char * n) {
+    if (eq_case_insensitive("adamw", n)) {
+        return GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    }
+    if (eq_case_insensitive("sgd", n)) {
+        return GGML_OPT_OPTIMIZER_TYPE_SGD;
+    }
+    return GGML_OPT_OPTIMIZER_TYPE_COUNT;
+}
+
+// TODO simplify to use just log and exp
+static float const k_log_2 = std::log(2.f);
+
+void lr_opt::init() {
+    if (lr_min > 0 && lr_min < lr0) {
+        float nhalf = std::log(lr0 / lr_min) / k_log_2;
+        float e     = epochs;
+        if (decay_epochs > 0 && decay_epochs < e) {
+            e = decay_epochs;
+        } else {
+            decay_epochs = e;
+        }
+        scale_epoch = nhalf / e;
+    }
+}
+
+float lr_opt::get_lr(float epoch) const {
+    float r = lr_min <= 0 ? lr0 :
+        epoch >= decay_epochs ? lr_min :
+        lr0 * std::pow(0.5f, epoch * scale_epoch);
+    LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
+    return r;
+}

llama/llama.cpp/common/common.h

@@ -2,14 +2,17 @@
 
 #pragma once
 
-#include "llama-cpp.h"
-
 #include <set>
+#include <sstream>
 #include <string>
 #include <string_view>
 #include <vector>
 #include <map>
 #include <sstream>
+#include <cmath>
+
+#include "ggml-opt.h"
+#include "llama-cpp.h"
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -31,6 +34,9 @@ struct common_adapter_lora_info {
     std::string path;
     float scale;
 
+    std::string task_name;
+    std::string prompt_prefix;
+
     struct llama_adapter_lora * ptr;
 };
 
@@ -82,6 +88,7 @@ enum llama_example {
     LLAMA_EXAMPLE_PARALLEL,
     LLAMA_EXAMPLE_TTS,
     LLAMA_EXAMPLE_DIFFUSION,
+    LLAMA_EXAMPLE_FINETUNE,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -190,6 +197,7 @@ struct common_params_model {
     std::string url         = ""; // model url to download                                  // NOLINT
     std::string hf_repo     = ""; // HF repo                                                // NOLINT
     std::string hf_file     = ""; // HF file                                                // NOLINT
+    std::string docker_repo = ""; // Docker repo                                            // NOLINT
 };
 
 struct common_params_speculative {
@@ -202,6 +210,7 @@ struct common_params_speculative {
     float   p_split      =  0.1f; // speculative decoding split probability
     float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
     std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
 
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
@@ -234,14 +243,36 @@ struct common_params_diffusion {
     bool    add_gumbel_noise = false; // add gumbel noise to the logits if temp > 0.0
 };
 
+// reasoning API response format (not to be confused as chat template's reasoning format)
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_AUTO,
+    COMMON_REASONING_FORMAT_AUTO,            // Same as deepseek, using `message.reasoning_content`
     COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
     COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
-    COMMON_REASONING_FORMAT_GRANITE,         // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
+    // do not extend this enum unless you absolutely have to
+    // in most cases, use COMMON_REASONING_FORMAT_AUTO
+    // see: https://github.com/ggml-org/llama.cpp/pull/15408
 };
 
+
+struct lr_opt {
+    float    lr0          = 1e-5; // learning rate at first epoch
+    float    lr_min       = -1;
+    float    decay_epochs = -1;   // if >0, the learning rate starts at lr0 and decays to lr_min after this many epochs
+    float    scale_epoch  = 0;
+    float    wd           = 0;
+    unsigned epochs       = 2;
+
+    unsigned epoch; // set by optimizer outer (epochs) loop
+    // learning rate decay - constant LR per epoch only for now
+    float get_lr(float e) const;
+    float get_lr() const { return get_lr(epoch); }
+    // must call after arg parse, before get_lr
+    void init();
+};
+
+struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
+
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =  4096; // context size
@@ -257,11 +288,10 @@ struct common_params {
     float   rope_freq_base        =  0.0f; // RoPE base frequency
     float   rope_freq_scale       =  0.0f; // RoPE frequency scaling factor
     float   yarn_ext_factor       = -1.0f; // YaRN extrapolation mix factor
-    float   yarn_attn_factor      =  1.0f; // YaRN magnitude scaling factor
-    float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
-    float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
+    float   yarn_attn_factor      = -1.0f; // YaRN magnitude scaling factor
+    float   yarn_beta_fast        = -1.0f; // YaRN low correction dim
+    float   yarn_beta_slow        = -1.0f; // YaRN high correction dim
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
-    float   defrag_thold          =  0.1f; // KV cache defragmentation threshold
 
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
@@ -283,6 +313,7 @@ struct common_params {
     enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
+    enum llama_flash_attn_type   flash_attn_type   = LLAMA_FLASH_ATTN_TYPE_AUTO; // whether to use Flash Attention
 
     struct common_params_sampling    sampling;
     struct common_params_speculative speculative;
@@ -346,9 +377,8 @@ struct common_params {
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
-    bool flash_attn        = false; // flash attention
     bool no_perf           = false; // disable performance metrics
-    bool ctx_shift         = true;  // context shift on inifinite text generation
+    bool ctx_shift         = false;  // context shift on infinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     bool kv_unified        = false; // enable unified KV cache
 
@@ -376,6 +406,11 @@ struct common_params {
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
 
+    // finetune
+    struct lr_opt lr;
+    enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    float val_split = 0.05f; // fraction of the data used for the validation set
+
     // embedding
     bool embedding         = false; // get only sentence embedding
     int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
@@ -389,6 +424,7 @@ struct common_params {
     int32_t timeout_write     = timeout_read; // http write timeout in seconds
     int32_t n_threads_http    = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
     int32_t n_cache_reuse     = 0;            // min chunk size to reuse from the cache via KV shifting
+    int32_t n_swa_checkpoints = 3;            // max number of SWA checkpoints per slot
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
@@ -409,7 +445,7 @@ struct common_params {
 
     // "advanced" endpoints are disabled by default for better security
     bool webui            = true;
-    bool endpoint_slots   = false;
+    bool endpoint_slots   = true;
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
 
@@ -417,7 +453,7 @@ struct common_params {
 
     std::string slot_save_path;
 
-    float slot_prompt_similarity = 0.5f;
+    float slot_prompt_similarity = 0.1f;
 
     // batched-bench params
     bool is_pp_shared = false;
@@ -698,8 +734,25 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 
 }
 
+//
+// MoE utils
+//
+
+const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_(ch|)exps";
+
+static std::string llm_ffn_exps_block_regex(int idx) {
+    return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
+}
+
+static llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
+    return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
+}
+
 //
 // training utils
 //
 
 ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);
+
+// "adamw" or "sgd" (case insensitive)
+enum ggml_opt_optimizer_type common_opt_get_optimizer(const char *);

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

@@ -257,12 +257,13 @@ std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_RULES = {
 };
 
 static bool is_reserved_name(const std::string & name) {
-    static std::unordered_set<std::string> RESERVED_NAMES;
-    if (RESERVED_NAMES.empty()) {
-        RESERVED_NAMES.insert("root");
-        for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
-        for (const auto &p : STRING_FORMAT_RULES) RESERVED_NAMES.insert(p.first);
-    }
+    static const std::unordered_set<std::string> RESERVED_NAMES = [] {
+        std::unordered_set<std::string> s;
+        s.insert("root");
+        for (const auto & p : PRIMITIVE_RULES) s.insert(p.first);
+        for (const auto & p : STRING_FORMAT_RULES) s.insert(p.first);
+        return s;
+    }();
     return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
 }
 
@@ -843,9 +844,10 @@ public:
                 _build_object_rule(
                     properties, required, name,
                     schema.contains("additionalProperties") ? schema["additionalProperties"] : json()));
-        } else if ((schema_type.is_null() || schema_type == "object") && schema.contains("allOf")) {
+        } else if ((schema_type.is_null() || schema_type == "object" || schema_type == "string") && schema.contains("allOf")) {
             std::unordered_set<std::string> required;
             std::vector<std::pair<std::string, json>> properties;
+            std::map<std::string, size_t> enum_values;
             std::string hybrid_name = name;
             std::function<void(const json &, bool)> add_component = [&](const json & comp_schema, bool is_required) {
                 if (comp_schema.contains("$ref")) {
@@ -857,6 +859,14 @@ public:
                             required.insert(prop.key());
                         }
                     }
+                } else if (comp_schema.contains("enum")) {
+                    for (const auto & v : comp_schema["enum"]) {
+                        const auto rule = _generate_constant_rule(v);
+                        if (enum_values.find(rule) == enum_values.end()) {
+                            enum_values[rule] = 0;
+                        }
+                        enum_values[rule] += 1;
+                    }
                 } else {
                   // todo warning
                 }
@@ -870,6 +880,17 @@ public:
                     add_component(t, true);
                 }
             }
+            if (!enum_values.empty()) {
+                std::vector<std::string> enum_intersection;
+                for (const auto & p : enum_values) {
+                    if (p.second == schema["allOf"].size()) {
+                        enum_intersection.push_back(p.first);
+                    }
+                }
+                if (!enum_intersection.empty()) {
+                    return _add_rule(rule_name, "(" + string_join(enum_intersection, " | ") + ") space");
+                }
+            }
             return _add_rule(rule_name, _build_object_rule(properties, required, hybrid_name, json()));
         } else if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) {
             json items = schema.contains("items") ? schema["items"] : schema["prefixItems"];

llama/llama.cpp/common/log.cpp

@@ -4,17 +4,52 @@
 #include <condition_variable>
 #include <cstdarg>
 #include <cstdio>
+#include <cstdlib>
+#include <cstring>
 #include <mutex>
 #include <sstream>
 #include <thread>
 #include <vector>
 
+#if defined(_WIN32)
+#    include <io.h>
+#    include <windows.h>
+#    define isatty _isatty
+#    define fileno _fileno
+#else
+#    include <unistd.h>
+#endif // defined(_WIN32)
+
 int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;
 
 void common_log_set_verbosity_thold(int verbosity) {
     common_log_verbosity_thold = verbosity;
 }
 
+// Auto-detect if colors should be enabled based on terminal and environment
+static bool common_log_should_use_colors_auto() {
+    // Check NO_COLOR environment variable (https://no-color.org/)
+    if (const char * no_color = std::getenv("NO_COLOR")) {
+        if (no_color[0] != '\0') {
+            return false;
+        }
+    }
+
+    // Check TERM environment variable
+    if (const char * term = std::getenv("TERM")) {
+        if (std::strcmp(term, "dumb") == 0) {
+            return false;
+        }
+    }
+
+    // Check if stdout and stderr are connected to a terminal
+    // We check both because log messages can go to either
+    bool stdout_is_tty = isatty(fileno(stdout));
+    bool stderr_is_tty = isatty(fileno(stderr));
+
+    return stdout_is_tty || stderr_is_tty;
+}
+
 static int64_t t_us() {
     return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
 }
@@ -353,6 +388,11 @@ struct common_log * common_log_init() {
 
 struct common_log * common_log_main() {
     static struct common_log log;
+    static std::once_flag    init_flag;
+    std::call_once(init_flag, [&]() {
+        // Set default to auto-detect colors
+        log.set_colors(common_log_should_use_colors_auto());
+    });
 
     return &log;
 }
@@ -380,8 +420,19 @@ void common_log_set_file(struct common_log * log, const char * file) {
     log->set_file(file);
 }
 
-void common_log_set_colors(struct common_log * log, bool colors) {
-    log->set_colors(colors);
+void common_log_set_colors(struct common_log * log, log_colors colors) {
+    if (colors == LOG_COLORS_AUTO) {
+        log->set_colors(common_log_should_use_colors_auto());
+        return;
+    }
+
+    if (colors == LOG_COLORS_DISABLED) {
+        log->set_colors(false);
+        return;
+    }
+
+    GGML_ASSERT(colors == LOG_COLORS_ENABLED);
+    log->set_colors(true);
 }
 
 void common_log_set_prefix(struct common_log * log, bool prefix) {

llama/llama.cpp/common/log.h

@@ -24,6 +24,12 @@
 #define LOG_DEFAULT_DEBUG 1
 #define LOG_DEFAULT_LLAMA 0
 
+enum log_colors {
+    LOG_COLORS_AUTO     = -1,
+    LOG_COLORS_DISABLED = 0,
+    LOG_COLORS_ENABLED  = 1,
+};
+
 // needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
 // set via common_log_set_verbosity()
 extern int common_log_verbosity_thold;
@@ -66,7 +72,7 @@ void common_log_add(struct common_log * log, enum ggml_log_level level, const ch
 //
 
 void common_log_set_file      (struct common_log * log, const char * file); // not thread-safe
-void common_log_set_colors    (struct common_log * log,       bool   colors);     // not thread-safe
+void common_log_set_colors    (struct common_log * log, log_colors colors); // not thread-safe
 void common_log_set_prefix    (struct common_log * log, bool prefix);       // whether to output prefix to each log
 void common_log_set_timestamps(struct common_log * log, bool timestamps);   // whether to output timestamps in the prefix
 

llama/llama.cpp/common/sampling.cpp

@@ -332,6 +332,7 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
     }
     if (ctx) {
         llama_perf_context_print(ctx);
+        llama_memory_breakdown_print(ctx);
     }
 }
 
@@ -426,8 +427,29 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
 
 // helpers
 
-llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl) {
-    return &gsmpl->cur_p;
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort) {
+    auto * res = &gsmpl->cur_p;
+
+    if (do_sort && !res->sorted) {
+        // remember the selected token before sorting
+        const llama_token id = res->data[res->selected].id;
+
+        std::sort(res->data, res->data + res->size, [](const llama_token_data & a, const llama_token_data & b) {
+            return a.p > b.p;
+        });
+
+        // restore the selected token after sorting
+        for (size_t i = 0; i < res->size; ++i) {
+            if (res->data[i].id == id) {
+                res->selected = i;
+                break;
+            }
+        }
+
+        res->sorted = true;
+    }
+
+    return res;
 }
 
 llama_token common_sampler_last(const struct common_sampler * gsmpl) {

llama/llama.cpp/common/sampling.h

@@ -86,7 +86,9 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 // helpers
 
 // access the internal list of current candidate tokens
-llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl);
+// if do_sort == true, the candidates are guaranteed to be sorted afterwards (in descending order of probability)
+// the .sorted flag of the result indicates whether the returned candidates are sorted
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort);
 
 // get the last accepted token
 llama_token common_sampler_last(const struct common_sampler * gsmpl);

llama/llama.cpp/include/llama.h

@@ -64,8 +64,6 @@ extern "C" {
 
     typedef struct llama_memory_i * llama_memory_t;
 
-    struct llama_kv_cache; // DEPRECATED (use llama_memory instead)
-
     typedef int32_t llama_pos;
     typedef int32_t llama_token;
     typedef int32_t llama_seq_id;
@@ -181,6 +179,14 @@ extern "C" {
         LLAMA_ATTENTION_TYPE_NON_CAUSAL  = 1,
     };
 
+    enum llama_flash_attn_type {
+        LLAMA_FLASH_ATTN_TYPE_AUTO     = -1,
+        LLAMA_FLASH_ATTN_TYPE_DISABLED = 0,
+        LLAMA_FLASH_ATTN_TYPE_ENABLED  = 1,
+    };
+
+    LLAMA_API const char * llama_flash_attn_type_name(enum llama_flash_attn_type flash_attn_type);
+
     enum llama_split_mode {
         LLAMA_SPLIT_MODE_NONE  = 0, // single GPU
         LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
@@ -200,7 +206,7 @@ extern "C" {
         llama_token_data * data;
         size_t size;
         int64_t selected; // this is the index in the data array (i.e. not the token id)
-        bool sorted;
+        bool sorted;      // note: do not assume the data is sorted - always check this flag
     } llama_token_data_array;
 
     typedef bool (*llama_progress_callback)(float progress, void * user_data);
@@ -305,6 +311,7 @@ extern "C" {
         enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
         enum llama_pooling_type      pooling_type;      // whether to pool (sum) embedding results by sequence id
         enum llama_attention_type    attention_type;    // attention type to use for embeddings
+        enum llama_flash_attn_type   flash_attn_type;   // when to enable Flash Attention
 
         // ref: https://github.com/ggml-org/llama.cpp/pull/2054
         float    rope_freq_base;   // RoPE base frequency, 0 = from model
@@ -314,7 +321,7 @@ extern "C" {
         float    yarn_beta_fast;   // YaRN low correction dim
         float    yarn_beta_slow;   // YaRN high correction dim
         uint32_t yarn_orig_ctx;    // YaRN original context size
-        float    defrag_thold;     // defragment the KV cache if holes/size > thold, <= 0 disabled (default)
+        float    defrag_thold;     // [DEPRECATED] defragment the KV cache if holes/size > thold, <= 0 disabled (default)
 
         ggml_backend_sched_eval_callback cb_eval;
         void * cb_eval_user_data;
@@ -331,7 +338,6 @@ extern "C" {
         // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
         bool embeddings;  // if true, extract embeddings (together with logits)
         bool offload_kqv; // offload the KQV ops (including the KV cache) to GPU
-        bool flash_attn;  // use flash attention [EXPERIMENTAL]
         bool no_perf;     // measure performance timings
         bool op_offload;  // offload host tensor operations to device
         bool swa_full;    // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
@@ -469,8 +475,6 @@ extern "C" {
     LLAMA_API           llama_memory_t   llama_get_memory  (const struct llama_context * ctx);
     LLAMA_API  enum llama_pooling_type   llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type
 
-    DEPRECATED(LLAMA_API struct llama_kv_cache * llama_get_kv_self(struct llama_context * ctx), "use llama_get_memory instead");
-
     LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
     LLAMA_API enum llama_rope_type       llama_model_rope_type(const struct llama_model * model);
 
@@ -557,10 +561,32 @@ extern "C" {
             struct llama_model * model,
             const char * path_lora);
 
+    // Functions to access the adapter's GGUF metadata scalar values
+    // - The functions return the length of the string on success, or -1 on failure
+    // - The output string is always null-terminated and cleared on failure
+    // - When retrieving a string, an extra byte must be allocated to account for the null terminator
+    // - GGUF array values are not supported by these functions
+
+    // Get metadata value as a string by key name
+    LLAMA_API int32_t llama_adapter_meta_val_str(const struct llama_adapter_lora * adapter, const char * key, char * buf, size_t buf_size);
+
+    // Get the number of metadata key/value pairs
+    LLAMA_API int32_t llama_adapter_meta_count(const struct llama_adapter_lora * adapter);
+
+    // Get metadata key name by index
+    LLAMA_API int32_t llama_adapter_meta_key_by_index(const struct llama_adapter_lora * adapter, int32_t i, char * buf, size_t buf_size);
+
+    // Get metadata value as a string by index
+    LLAMA_API int32_t llama_adapter_meta_val_str_by_index(const struct llama_adapter_lora * adapter, int32_t i, char * buf, size_t buf_size);
+
     // Manually free a LoRA adapter
     // Note: loaded adapters will be free when the associated model is deleted
     LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter);
 
+    // Get the invocation tokens if the current lora is an alora
+    LLAMA_API uint64_t            llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter);
+    LLAMA_API const llama_token * llama_adapter_get_alora_invocation_tokens  (const struct llama_adapter_lora * adapter);
+
     // The following functions operate on a llama_context, hence the naming: llama_verb_...
 
     // Add a loaded LoRA adapter to given context
@@ -667,111 +693,6 @@ extern "C" {
     // Check if the memory supports shifting
     LLAMA_API bool llama_memory_can_shift(llama_memory_t mem);
 
-    //
-    // KV cache for self-attention (TODO: deprecate in favor of llama_memory)
-    //
-
-    // Returns the number of tokens in the KV cache (slow, use only for debug)
-    // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
-    DEPRECATED(LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx),
-               "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)");
-
-    // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
-    DEPRECATED(LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx),
-               "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)");
-
-    // Clear the KV cache - both cell info is erased and KV data is zeroed
-    DEPRECATED(LLAMA_API void llama_kv_self_clear(
-                struct llama_context * ctx),
-            "Use llama_memory_clear() instead");
-
-    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
-    // seq_id < 0 : match any sequence
-    // p0 < 0     : [0,  p1]
-    // p1 < 0     : [p0, inf)
-    DEPRECATED(LLAMA_API bool llama_kv_self_seq_rm(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1),
-            "Use llama_memory_seq_rm() instead");
-
-    // Copy all tokens that belong to the specified sequence to another sequence
-    // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_cp(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id_src,
-                    llama_seq_id   seq_id_dst,
-                       llama_pos   p0,
-                       llama_pos   p1),
-            "Use llama_memory_seq_cp() instead");
-
-    // Removes all tokens that do not belong to the specified sequence
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_keep(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id),
-            "Use llama_memory_seq_keep() instead");
-
-    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // If the KV cache is RoPEd, the KV data is updated accordingly:
-    //   - lazily on next llama_decode()
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_add(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1,
-                       llama_pos   delta),
-            "Use llama_memory_seq_add() instead");
-
-    // Integer division of the positions by factor of `d > 1`
-    // If the KV cache is RoPEd, the KV data is updated accordingly:
-    //   - lazily on next llama_decode()
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_div(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1,
-                             int   d),
-            "Use llama_memory_seq_div() instead");
-
-    // Returns the smallest position present in the KV cache for the specified sequence
-    // This is typically non-zero only for SWA caches
-    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
-    // Return -1 if the sequence is empty
-    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_min(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id),
-            "Use llama_memory_seq_pos_min() instead");
-
-    // Returns the largest position present in the KV cache for the specified sequence
-    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
-    // Return -1 if the sequence is empty
-    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_max(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id),
-            "Use llama_memory_seq_pos_max() instead");
-
-    // Defragment the KV cache
-    // This will be applied:
-    //   - lazily on next llama_decode()
-    DEPRECATED(LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx),
-            "simply remove this call, the context will automatically decide when to do a defragmentation based on 'defrag_thold'");
-
-    // Check if the context supports KV cache shifting
-    DEPRECATED(LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx),
-            "use llama_memory_can_shift() instead");
-
-    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    DEPRECATED(LLAMA_API void llama_kv_self_update(struct llama_context * ctx),
-            "simply remove this call, updates are applied lazily on the next llama_decode()");
-
     //
     // State / sessions
     //
@@ -870,6 +791,29 @@ extern "C" {
                           size_t   n_token_capacity,
                           size_t * n_token_count_out);
 
+#define LLAMA_STATE_SEQ_FLAGS_SWA_ONLY 1
+
+    typedef uint32_t llama_state_seq_flags;
+
+    LLAMA_API size_t llama_state_seq_get_size_ext(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+           llama_state_seq_flags   flags);
+
+    LLAMA_API size_t llama_state_seq_get_data_ext(
+            struct llama_context * ctx,
+                         uint8_t * dst,
+                          size_t   size,
+                    llama_seq_id   seq_id,
+           llama_state_seq_flags   flags);
+
+    LLAMA_API size_t llama_state_seq_set_data_ext(
+            struct llama_context * ctx,
+                   const uint8_t * src,
+                          size_t   size,
+                    llama_seq_id   dest_seq_id,
+           llama_state_seq_flags   flags);
+
     //
     // Decoding
     //
@@ -1216,11 +1160,6 @@ extern "C" {
     LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
     LLAMA_API struct llama_sampler * llama_sampler_init_dist  (uint32_t seed);
 
-    /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
-    /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
-    DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void),
-        "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
-
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     /// Setting k <= 0 makes this a noop
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
@@ -1390,24 +1329,25 @@ extern "C" {
     //
     // Performance utils
     //
-    // NOTE: Used by llama.cpp examples, avoid using in third-party apps. Instead, do your own performance measurements.
+    // NOTE: Used by llama.cpp examples/tools, avoid using in third-party apps. Instead, do your own performance measurements.
     //
 
     struct llama_perf_context_data {
-        double t_start_ms;
-        double t_load_ms;
-        double t_p_eval_ms;
-        double t_eval_ms;
+        // ms == milliseconds
+        double t_start_ms;  // absolute start time
+        double t_load_ms;   // time needed for loading the model
+        double t_p_eval_ms; // time needed for processing the prompt
+        double t_eval_ms;   // time needed for generating tokens
 
-        int32_t n_p_eval;
-        int32_t n_eval;
+        int32_t n_p_eval;   // number of prompt tokens
+        int32_t n_eval;     // number of generated tokens
         int32_t n_reused;   // number of times a ggml compute graph had been reused
     };
 
     struct llama_perf_sampler_data {
-        double t_sample_ms;
+        double t_sample_ms; // time needed for sampling in ms
 
-        int32_t n_sample;
+        int32_t n_sample;   // number of sampled tokens
     };
 
     LLAMA_API struct llama_perf_context_data llama_perf_context      (const struct llama_context * ctx);
@@ -1419,6 +1359,9 @@ 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);
 
+    // print a breakdown of per-device memory use via LLAMA_LOG:
+    LLAMA_API void llama_memory_breakdown_print(const struct llama_context * ctx);
+
     //
     // training
     //
@@ -1437,6 +1380,8 @@ extern "C" {
 
         ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
         void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+
+        enum ggml_opt_optimizer_type optimizer_type;
     };
 
     LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params);

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

@@ -6,6 +6,7 @@
 
 #include <map>
 #include <cassert>
+#include <sstream>
 #include <stdexcept>
 
 // vec
@@ -163,13 +164,38 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
 
     // check metadata
     {
+        const gguf_context * gguf_ctx = ctx_gguf.get();
+
+        LLAMA_LOG_INFO("%s: Dumping metadata keys/values.\n", __func__);
+
+        // get metadata as string
+        for (int i = 0; i < gguf_get_n_kv(gguf_ctx); i++) {
+            gguf_type type = gguf_get_kv_type(gguf_ctx, i);
+            const std::string type_name =
+                type == GGUF_TYPE_ARRAY
+                ? format("%s[%s,%zu]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(gguf_ctx, i)), gguf_get_arr_n(gguf_ctx, i))
+                : gguf_type_name(type);
+            const char * name = gguf_get_key(gguf_ctx, i);
+            const std::string value = gguf_kv_to_str(gguf_ctx, i);
+
+            if (type != GGUF_TYPE_ARRAY) {
+                adapter.gguf_kv.emplace(name, value);
+            }
+
+            const size_t MAX_VALUE_LEN = 40;
+            std::string print_value = value.size() > MAX_VALUE_LEN ? format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str()) : value;
+            replace_all(print_value, "\n", "\\n");
+
+            LLAMA_LOG_INFO("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), print_value.c_str());
+        }
+
         auto get_kv_str = [&](const std::string & key) -> std::string {
-            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
-            return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf.get(), id));
+            int id = gguf_find_key(gguf_ctx, key.c_str());
+            return id < 0 ? "" : std::string(gguf_get_val_str(gguf_ctx, id));
         };
         auto get_kv_f32 = [&](const std::string & key) -> float {
-            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
-            return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf.get(), id);
+            int id = gguf_find_key(gguf_ctx, key.c_str());
+            return id < 0 ? 0.0f : gguf_get_val_f32(gguf_ctx, id);
         };
         LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN);
 
@@ -190,6 +216,26 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
         }
 
         adapter.alpha = get_kv_f32(llm_kv(LLM_KV_ADAPTER_LORA_ALPHA));
+
+        // parse alora invocation sequence vector
+        const auto & key = llm_kv(LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS);
+        const int kid = gguf_find_key(ctx_gguf.get(), key.c_str());
+        if (kid >= 0) {
+            if (gguf_get_kv_type(ctx_gguf.get(), kid) != GGUF_TYPE_ARRAY) {
+                throw std::runtime_error("invalid gguf type for " + key);
+            }
+            const auto arr_type = gguf_get_arr_type(ctx_gguf.get(), kid);
+            if (arr_type != GGUF_TYPE_UINT32) {
+                throw std::runtime_error("invalid gguf element type for " + key);
+            }
+            const size_t seq_len = gguf_get_arr_n(ctx_gguf.get(), kid);
+            const void * data = gguf_get_arr_data(ctx_gguf.get(), kid);
+            adapter.alora_invocation_tokens.resize(seq_len);
+            std::copy(
+                (const llama_token *)data,
+                (const llama_token *)data + seq_len,
+                adapter.alora_invocation_tokens.begin());
+        }
     }
 
     int n_tensors = gguf_get_n_tensors(ctx_gguf.get());
@@ -383,6 +429,57 @@ llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * p
     return nullptr;
 }
 
+int32_t llama_adapter_meta_val_str(const llama_adapter_lora * adapter, const char * key, char * buf, size_t buf_size) {
+    const auto & it = adapter->gguf_kv.find(key);
+    if (it == adapter->gguf_kv.end()) {
+        if (buf_size > 0) {
+            buf[0] = '\0';
+        }
+        return -1;
+    }
+    return snprintf(buf, buf_size, "%s", it->second.c_str());
+}
+
+int32_t llama_adapter_meta_count(const llama_adapter_lora * adapter) {
+    return (int)adapter->gguf_kv.size();
+}
+
+int32_t llama_adapter_meta_key_by_index(const llama_adapter_lora * adapter, int i, char * buf, size_t buf_size) {
+    if (i < 0 || i >= (int)adapter->gguf_kv.size()) {
+        if (buf_size > 0) {
+            buf[0] = '\0';
+        }
+        return -1;
+    }
+    auto it = adapter->gguf_kv.begin();
+    std::advance(it, i);
+    return snprintf(buf, buf_size, "%s", it->first.c_str());
+}
+
+int32_t llama_adapter_meta_val_str_by_index(const llama_adapter_lora * adapter, int32_t i, char * buf, size_t buf_size) {
+    if (i < 0 || i >= (int)adapter->gguf_kv.size()) {
+        if (buf_size > 0) {
+            buf[0] = '\0';
+        }
+        return -1;
+    }
+    auto it = adapter->gguf_kv.begin();
+    std::advance(it, i);
+    return snprintf(buf, buf_size, "%s", it->second.c_str());
+}
+
 void llama_adapter_lora_free(llama_adapter_lora * adapter) {
     delete adapter;
 }
+
+uint64_t llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter) {
+    if (!adapter) {
+        return 0;
+    }
+    return adapter->alora_invocation_tokens.size();
+}
+
+const llama_token * llama_adapter_get_alora_invocation_tokens(const llama_adapter_lora * adapter) {
+    GGML_ASSERT(adapter);
+    return adapter->alora_invocation_tokens.data();
+}

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

@@ -67,6 +67,12 @@ struct llama_adapter_lora {
 
     float alpha;
 
+    // gguf metadata
+    std::unordered_map<std::string, std::string> gguf_kv;
+
+    // activated lora (aLoRA)
+    std::vector<llama_token> alora_invocation_tokens;
+
     llama_adapter_lora() = default;
     ~llama_adapter_lora() = default;
 

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

@@ -22,6 +22,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
     { LLM_ARCH_NEO_BERT,         "neo-bert"         },
     { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
+    { LLM_ARCH_JINA_BERT_V3,     "jina-bert-v3"     },
     { LLM_ARCH_BLOOM,            "bloom"            },
     { LLM_ARCH_STABLELM,         "stablelm"         },
     { LLM_ARCH_QWEN,             "qwen"             },
@@ -44,6 +45,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_GEMMA2,           "gemma2"           },
     { LLM_ARCH_GEMMA3,           "gemma3"           },
     { LLM_ARCH_GEMMA3N,          "gemma3n"          },
+    { LLM_ARCH_GEMMA_EMBEDDING,  "gemma-embedding"  },
     { LLM_ARCH_STARCODER2,       "starcoder2"       },
     { LLM_ARCH_MAMBA,            "mamba"            },
     { LLM_ARCH_MAMBA2,           "mamba2"           },
@@ -68,6 +70,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_T5ENCODER,        "t5encoder"        },
     { LLM_ARCH_JAIS,             "jais"             },
     { LLM_ARCH_NEMOTRON,         "nemotron"         },
+    { LLM_ARCH_NEMOTRON_H,       "nemotron_h"       },
     { LLM_ARCH_EXAONE,           "exaone"           },
     { LLM_ARCH_EXAONE4,          "exaone4"          },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
@@ -94,6 +97,9 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DREAM,            "dream"            },
     { LLM_ARCH_SMALLTHINKER,     "smallthinker"     },
     { LLM_ARCH_LLADA,            "llada"            },
+    { LLM_ARCH_LLADA_MOE,        "llada-moe"        },
+    { LLM_ARCH_SEED_OSS,         "seed_oss"         },
+    { LLM_ARCH_GROVEMOE,         "grovemoe"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -121,6 +127,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_FEED_FORWARD_LENGTH,               "%s.feed_forward_length"               },
     { LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        "%s.expert_feed_forward_length"        },
     { LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" },
+    { LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,  "%s.expert_chunk_feed_forward_length"  },
     { LLM_KV_USE_PARALLEL_RESIDUAL,             "%s.use_parallel_residual"             },
     { LLM_KV_TENSOR_DATA_LAYOUT,                "%s.tensor_data_layout"                },
     { LLM_KV_EXPERT_COUNT,                      "%s.expert_count"                      },
@@ -129,12 +136,16 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
     { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
     { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
+    { LLM_KV_EXPERT_GROUP_SCALE,                "%s.expert_group_scale"                },
+    { LLM_KV_EXPERTS_PER_GROUP,                 "%s.experts_per_group"                 },
     { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
     { LLM_KV_NEXTN_PREDICT_LAYERS,              "%s.nextn_predict_layers"              },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
+    { LLM_KV_DECODER_BLOCK_COUNT,               "%s.decoder_block_count"               },
     { LLM_KV_ATTN_LOGIT_SOFTCAPPING,            "%s.attn_logit_softcapping"            },
+    { LLM_KV_ROUTER_LOGIT_SOFTCAPPING,          "%s.router_logit_softcapping"          },
     { LLM_KV_FINAL_LOGIT_SOFTCAPPING,           "%s.final_logit_softcapping"           },
     { LLM_KV_SWIN_NORM,                         "%s.swin_norm"                         },
     { LLM_KV_RESCALE_EVERY_N_LAYERS,            "%s.rescale_every_n_layers"            },
@@ -165,6 +176,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
     { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
+    { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
+    { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
     { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
@@ -179,6 +192,10 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,     "%s.rope.scaling.original_context_length" },
     { LLM_KV_ROPE_SCALING_FINETUNED,        "%s.rope.scaling.finetuned"               },
     { LLM_KV_ROPE_SCALING_YARN_LOG_MUL,     "%s.rope.scaling.yarn_log_multiplier"     },
+    { LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR,  "%s.rope.scaling.yarn_ext_factor"         },
+    { LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, "%s.rope.scaling.yarn_attn_factor"        },
+    { LLM_KV_ROPE_SCALING_YARN_BETA_FAST,   "%s.rope.scaling.yarn_beta_fast"          },
+    { LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,   "%s.rope.scaling.yarn_beta_slow"          },
 
     { LLM_KV_SPLIT_NO,            "split.no"            },
     { LLM_KV_SPLIT_COUNT,         "split.count"         },
@@ -237,6 +254,9 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
 
     { LLM_KV_ADAPTER_TYPE,                    "adapter.type"               },
     { LLM_KV_ADAPTER_LORA_ALPHA,              "adapter.lora.alpha"         },
+    { LLM_KV_ADAPTER_LORA_TASK_NAME,          "adapter.lora.task_name"     },
+    { LLM_KV_ADAPTER_LORA_PROMPT_PREFIX,      "adapter.lora.prompt_prefix" },
+    { LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS, "adapter.alora.invocation_tokens" },
 
     // deprecated
     { LLM_KV_TOKENIZER_PREFIX_ID, "tokenizer.ggml.prefix_token_id" },
@@ -392,12 +412,16 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
             { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
             { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
             { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
             { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
             { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
             { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
             { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
             { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
             { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
         },
@@ -576,6 +600,20 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_CLS,             "cls" },
         },
     },
+    {
+        LLM_ARCH_JINA_BERT_V3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
+            { LLM_TENSOR_TOKEN_TYPES,     "token_types" },
+            { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
+        },
+    },
     {
         LLM_ARCH_BLOOM,
         {
@@ -689,6 +727,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
             { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
             { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_CLS_OUT,         "cls.output" },
             { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
             { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
             { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
@@ -1021,6 +1060,27 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_LAUREL_POST_NORM,     "blk.%d.laurel_post_norm" },
         },
     },
+    {
+        LLM_ARCH_GEMMA_EMBEDDING,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { 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_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
     {
         LLM_ARCH_STARCODER2,
         {
@@ -1534,6 +1594,31 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_NEMOTRON_H,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,     "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,    "output_norm" },
+            { LLM_TENSOR_OUTPUT,         "output" },
+            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm" },
+            // mamba(2) ssm layers
+            { LLM_TENSOR_SSM_IN,         "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,     "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,         "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,          "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,          "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,       "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,        "blk.%d.ssm_out" },
+            // attention layers
+            { 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" },
+            // dense FFN
+            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_EXAONE,
         {
@@ -2030,6 +2115,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
             { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
             { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT,            "output" },
         }
     },
     {
@@ -2087,6 +2173,66 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_LLADA_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
+    {
+        LLM_ARCH_SEED_OSS,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_GROVEMOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_CHEXPS,    "blk.%d.ffn_gate_chexps" },
+            { LLM_TENSOR_FFN_DOWN_CHEXPS,    "blk.%d.ffn_down_chexps" },
+            { LLM_TENSOR_FFN_UP_CHEXPS,      "blk.%d.ffn_up_chexps" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2219,6 +2365,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_DOWN_EXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_GATE_EXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_UP_EXPS,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
+    {LLM_TENSOR_FFN_DOWN_CHEXPS,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
+    {LLM_TENSOR_FFN_GATE_CHEXPS,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
+    {LLM_TENSOR_FFN_UP_CHEXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_EXP_PROBS_B,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     // altup / laurel (gemma 3n)
     {LLM_TENSOR_PER_LAYER_TOKEN_EMBD,       {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_GET_ROWS}},
@@ -2340,6 +2489,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_PLAMO2:
         case LLM_ARCH_GRANITE_HYBRID:
         case LLM_ARCH_LFM2:
+        case LLM_ARCH_NEMOTRON_H:
             return true;
         default:
             return false;
@@ -2350,6 +2500,7 @@ bool llm_arch_is_diffusion(const llm_arch & arch) {
     switch (arch) {
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
+        case LLM_ARCH_LLADA_MOE:
             return true;
         default:
             return false;

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

@@ -26,6 +26,7 @@ enum llm_arch {
     LLM_ARCH_NOMIC_BERT_MOE,
     LLM_ARCH_NEO_BERT,
     LLM_ARCH_JINA_BERT_V2,
+    LLM_ARCH_JINA_BERT_V3,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
     LLM_ARCH_QWEN,
@@ -48,6 +49,7 @@ enum llm_arch {
     LLM_ARCH_GEMMA2,
     LLM_ARCH_GEMMA3,
     LLM_ARCH_GEMMA3N,
+    LLM_ARCH_GEMMA_EMBEDDING,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
     LLM_ARCH_MAMBA2,
@@ -72,6 +74,7 @@ enum llm_arch {
     LLM_ARCH_T5ENCODER,
     LLM_ARCH_JAIS,
     LLM_ARCH_NEMOTRON,
+    LLM_ARCH_NEMOTRON_H,
     LLM_ARCH_EXAONE,
     LLM_ARCH_EXAONE4,
     LLM_ARCH_RWKV6,
@@ -98,6 +101,9 @@ enum llm_arch {
     LLM_ARCH_DREAM,
     LLM_ARCH_SMALLTHINKER,
     LLM_ARCH_LLADA,
+    LLM_ARCH_LLADA_MOE,
+    LLM_ARCH_SEED_OSS,
+    LLM_ARCH_GROVEMOE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -125,6 +131,7 @@ enum llm_kv {
     LLM_KV_FEED_FORWARD_LENGTH,
     LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
     LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
+    LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,
     LLM_KV_USE_PARALLEL_RESIDUAL,
     LLM_KV_TENSOR_DATA_LAYOUT,
     LLM_KV_EXPERT_COUNT,
@@ -133,12 +140,16 @@ enum llm_kv {
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,
+    LLM_KV_EXPERT_GROUP_SCALE,
+    LLM_KV_EXPERTS_PER_GROUP,
     LLM_KV_MOE_EVERY_N_LAYERS,
     LLM_KV_NEXTN_PREDICT_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
+    LLM_KV_DECODER_BLOCK_COUNT,
     LLM_KV_ATTN_LOGIT_SOFTCAPPING,
+    LLM_KV_ROUTER_LOGIT_SOFTCAPPING,
     LLM_KV_FINAL_LOGIT_SOFTCAPPING,
     LLM_KV_SWIN_NORM,
     LLM_KV_RESCALE_EVERY_N_LAYERS,
@@ -169,6 +180,8 @@ enum llm_kv {
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
+    LLM_KV_ATTENTION_OUTPUT_SCALE,
+    LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
     LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
@@ -183,6 +196,10 @@ enum llm_kv {
     LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,
     LLM_KV_ROPE_SCALING_FINETUNED,
     LLM_KV_ROPE_SCALING_YARN_LOG_MUL,
+    LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR,
+    LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR,
+    LLM_KV_ROPE_SCALING_YARN_BETA_FAST,
+    LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,
 
     LLM_KV_SPLIT_NO,
     LLM_KV_SPLIT_COUNT,
@@ -231,6 +248,9 @@ enum llm_kv {
 
     LLM_KV_ADAPTER_TYPE,
     LLM_KV_ADAPTER_LORA_ALPHA,
+    LLM_KV_ADAPTER_LORA_TASK_NAME,
+    LLM_KV_ADAPTER_LORA_PROMPT_PREFIX,
+    LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS,
 
     LLM_KV_POSNET_EMBEDDING_LENGTH,
     LLM_KV_POSNET_BLOCK_COUNT,
@@ -287,6 +307,9 @@ enum llm_tensor {
     LLM_TENSOR_FFN_DOWN_SHEXP,
     LLM_TENSOR_FFN_GATE_SHEXP,
     LLM_TENSOR_FFN_UP_SHEXP,
+    LLM_TENSOR_FFN_DOWN_CHEXPS,
+    LLM_TENSOR_FFN_GATE_CHEXPS,
+    LLM_TENSOR_FFN_UP_CHEXPS,
     LLM_TENSOR_FFN_EXP_PROBS_B,
     LLM_TENSOR_ATTN_Q_NORM,
     LLM_TENSOR_ATTN_K_NORM,

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

@@ -477,7 +477,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
 
 llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
     if (sequential && has_cpl) {
-        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch\n", __func__);
+        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch (you may need to use the -kvu flag)\n", __func__);
 
         return {};
     }

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

@@ -16,10 +16,10 @@
 static std::string trim(const std::string & str) {
     size_t start = 0;
     size_t end = str.size();
-    while (start < end && isspace(str[start])) {
+    while (start < end && isspace(static_cast<unsigned char>(str[start]))) {
         start += 1;
     }
-    while (end > start && isspace(str[end - 1])) {
+    while (end > start && isspace(static_cast<unsigned char>(str[end - 1]))) {
         end -= 1;
     }
     return str.substr(start, end - start);
@@ -69,6 +69,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "gpt-oss",           LLM_CHAT_TEMPLATE_OPENAI_MOE        },
     { "hunyuan-dense",     LLM_CHAT_TEMPLATE_HUNYUAN_DENSE     },
     { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
+    { "seed_oss",          LLM_CHAT_TEMPLATE_SEED_OSS          },
+    { "grok-2",            LLM_CHAT_TEMPLATE_GROK_2            },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -201,6 +203,10 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_HUNYUAN_DENSE;
     } else if (tmpl_contains("<|im_assistant|>assistant<|im_middle|>")) {
         return LLM_CHAT_TEMPLATE_KIMI_K2;
+    } else if (tmpl_contains("<seed:bos>")) {
+        return LLM_CHAT_TEMPLATE_SEED_OSS;
+    } else if (tmpl_contains("'Assistant: '  + message['content'] + '<|separator|>")) {
+        return LLM_CHAT_TEMPLATE_GROK_2;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -752,6 +758,28 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|im_assistant|>assistant<|im_middle|>";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_SEED_OSS) {
+        for (auto message: chat) {
+            std::string role(message->role);
+            ss << "<seed:bos>" << role << "\n" << (role == "assistant" ? trim(message->content) : message->content) << "<seed:eos>";
+        }
+        if (add_ass) {
+            ss << "<seed:bos>assistant\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GROK_2) {
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "System: " << trim(message->content) << "<|separator|>\n\n";
+            } else if (role == "user") {
+                ss << "Human: " << trim(message->content) << "<|separator|>\n\n";
+            } else if (role == "assistant") {
+                ss << "Assistant: " << message->content << "<|separator|>\n\n";
+            }
+        }
+        if (add_ass) {
+            ss << "Assistant:";
+        }
     } else {
         // template not supported
         return -1;

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

@@ -49,6 +49,8 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_OPENAI_MOE,
     LLM_CHAT_TEMPLATE_HUNYUAN_DENSE,
     LLM_CHAT_TEMPLATE_KIMI_K2,
+    LLM_CHAT_TEMPLATE_SEED_OSS,
+    LLM_CHAT_TEMPLATE_GROK_2,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

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

@@ -35,14 +35,12 @@ llama_context::llama_context(
 
     cparams.n_threads        = params.n_threads;
     cparams.n_threads_batch  = params.n_threads_batch;
-    cparams.yarn_ext_factor  = params.yarn_ext_factor;
-    cparams.yarn_attn_factor = params.yarn_attn_factor;
-    cparams.yarn_beta_fast   = params.yarn_beta_fast;
-    cparams.yarn_beta_slow   = params.yarn_beta_slow;
-    cparams.defrag_thold     = params.defrag_thold;
+    cparams.yarn_ext_factor  = params.yarn_ext_factor  >= 0.0f ? params.yarn_ext_factor  : hparams.yarn_ext_factor;
+    cparams.yarn_attn_factor = params.yarn_attn_factor >= 0.0f ? params.yarn_attn_factor : hparams.yarn_attn_factor;
+    cparams.yarn_beta_fast   = params.yarn_beta_fast   >= 0.0f ? params.yarn_beta_fast   : hparams.yarn_beta_fast;
+    cparams.yarn_beta_slow   = params.yarn_beta_slow   >= 0.0f ? params.yarn_beta_slow   : hparams.yarn_beta_slow;
     cparams.embeddings       = params.embeddings;
     cparams.offload_kqv      = params.offload_kqv;
-    cparams.flash_attn       = params.flash_attn;
     cparams.no_perf          = params.no_perf;
     cparams.pooling_type     = params.pooling_type;
     cparams.warmup           = false;
@@ -87,13 +85,15 @@ llama_context::llama_context(
         cparams.causal_attn = params.attention_type == LLAMA_ATTENTION_TYPE_CAUSAL;
     }
 
+    cparams.flash_attn = params.flash_attn_type != LLAMA_FLASH_ATTN_TYPE_DISABLED;
+
     // with causal attention, the batch size is limited by the context size
     cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
 
     // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask
     // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext)
     // ref: https://github.com/ggerganov/llama.cpp/pull/5021
-    // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_context_kv_self
+    // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_memory
     if (cparams.n_batch < GGML_KQ_MASK_PAD) {
         LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
         cparams.n_batch = GGML_KQ_MASK_PAD;
@@ -103,16 +103,6 @@ llama_context::llama_context(
     cparams.op_offload = params.op_offload;
     cparams.kv_unified = params.kv_unified;
 
-    {
-        const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
-        supports_set_rows = LLAMA_SET_ROWS ? (atoi(LLAMA_SET_ROWS) != 0) : supports_set_rows;
-
-        if (!supports_set_rows && !cparams.kv_unified) {
-            LLAMA_LOG_WARN("%s: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache\n", __func__);
-            cparams.kv_unified = true;
-        }
-    }
-
     {
         const char * LLAMA_GRAPH_REUSE_DISABLE = getenv("LLAMA_GRAPH_REUSE_DISABLE");
         graph_reuse_disable = LLAMA_GRAPH_REUSE_DISABLE ? (atoi(LLAMA_GRAPH_REUSE_DISABLE) != 0) : graph_reuse_disable;
@@ -130,7 +120,7 @@ llama_context::llama_context(
     LLAMA_LOG_INFO("%s: n_batch       = %u\n",   __func__, cparams.n_batch);
     LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
     LLAMA_LOG_INFO("%s: causal_attn   = %d\n",   __func__, cparams.causal_attn);
-    LLAMA_LOG_INFO("%s: flash_attn    = %d\n",   __func__, cparams.flash_attn);
+    LLAMA_LOG_INFO("%s: flash_attn    = %s\n",   __func__, llama_flash_attn_type_name(params.flash_attn_type));
     LLAMA_LOG_INFO("%s: kv_unified    = %s\n",   __func__, cparams.kv_unified ? "true" : "false");
     LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
     LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);
@@ -145,11 +135,6 @@ llama_context::llama_context(
                 __func__, n_ctx_per_seq, hparams.n_ctx_train);
     }
 
-    if (!params.swa_full && cparams.n_seq_max > 1 && hparams.is_swa_any()) {
-        LLAMA_LOG_WARN("%s: requested n_seq_max (%u) > 1, but swa_full is not enabled -- performance may be degraded: %s\n",
-                __func__, cparams.n_seq_max, "https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573");
-    }
-
     if (!hparams.vocab_only) {
         // GPU backends
         for (auto * dev : model.devices) {
@@ -196,7 +181,7 @@ llama_context::llama_context(
         // graph outputs buffer
         {
             // resized during inference when a batch uses more outputs
-            if ((uint32_t) output_reserve(params.n_seq_max) < params.n_seq_max) {
+            if (output_reserve(params.n_seq_max) < params.n_seq_max) {
                 throw std::runtime_error("failed to reserve initial output buffer");
             }
 
@@ -285,28 +270,75 @@ llama_context::llama_context(
         }
     }
 
-    // reserve worst-case graph
-    if (!hparams.vocab_only && memory) {
+    if (!hparams.vocab_only) {
+        llama_memory_context_ptr mctx;
+        if (memory) {
+            LLAMA_LOG_DEBUG("%s: reserving full memory module\n", __func__);
+            mctx = memory->init_full();
+            if (!mctx) {
+                throw std::runtime_error("failed to initialize memory module");
+            }
+        }
+
+        cross.v_embd.clear();
+
         const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
+        // avoid reserving graphs with zero outputs - assume one output per sequence
+        n_outputs = n_seqs;
+
         LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
 
+        // resolve automatic Flash Attention use
+        if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO) {
+            auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
+            if (!gf) {
+                throw std::runtime_error("failed to split graph for Flash Attention check");
+            }
+
+            const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FATTN) + 1;
+            bool fa_device_mismatch = false;
+            for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
+                ggml_tensor * n = ggml_graph_node(gf, i);
+                if (n->op != GGML_OP_FLASH_ATTN_EXT) {
+                    continue;
+                }
+                ggml_backend_dev_t device_fa = ggml_backend_get_device(
+                    ggml_backend_sched_get_tensor_backend(sched.get(), n));
+
+                // TODO: instead of the tensor names, use a map to keep track of which (FA) tensors belong to which layer
+                GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FATTN "-", prefix_len) == 0);
+                const int il = std::stoi(n->name + prefix_len);
+                ggml_backend_dev_t device_kv = model.dev_layer(il);
+                if (device_fa != device_kv) {
+                    LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the Flash Attention tensor "
+                        "is assigned to device %s (usually due to missing support)\n",
+                        __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_fa));
+                    // FIXME: fa_device_mismatch logic is wrong for --no-kv-offload, but this is broken anyways
+                    fa_device_mismatch = true;
+                    break;
+                }
+            }
+            if (fa_device_mismatch) {
+                cparams.flash_attn = false;
+                LLAMA_LOG_WARN("%s: Flash Attention was auto, set to disabled\n", __func__);
+                if (ggml_is_quantized(params.type_v)) {
+                    throw std::runtime_error("quantized V cache was requested, but this requires Flash Attention");
+                }
+            } else {
+                cparams.flash_attn = true;
+                LLAMA_LOG_INFO("%s: Flash Attention was auto, set to enabled\n", __func__);
+            }
+        }
+
+        // reserve worst-case graph
         int n_splits_pp = -1;
         int n_nodes_pp  = -1;
 
         int n_splits_tg = -1;
         int n_nodes_tg  = -1;
 
-        // simulate full KV cache
-
-        const auto mctx = memory->init_full();
-        if (!mctx) {
-            throw std::runtime_error("failed to initialize KV cache");
-        }
-
-        cross.v_embd.clear();
-
         // reserve pp (prompt processing) graph first so that buffers are only allocated once
         {
             auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
@@ -444,26 +476,12 @@ llama_memory_t llama_context::get_memory() const {
     return memory.get();
 }
 
-// deprecated
-void llama_context::kv_self_defrag_sched() {
-    if (!memory) {
-        return;
-    }
-
-    memory_force_optimize = true;
-}
-
-// deprecated
-bool llama_context::kv_self_update(bool optimize) {
+bool llama_context::memory_update(bool optimize) {
     if (!memory) {
         return false;
     }
 
     {
-        // TODO: remove in the future
-        optimize |= memory_force_optimize;
-        memory_force_optimize = false;
-
         const auto mctx = memory->init_update(this, optimize);
         switch (mctx->get_status()) {
             case LLAMA_MEMORY_STATUS_SUCCESS:
@@ -908,12 +926,6 @@ int llama_context::encode(const llama_batch & batch_inp) {
         }
     }
 
-    if (!supports_set_rows) {
-        // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
-        // overlap with device computation.
-        ggml_backend_sched_reset(sched.get());
-    }
-
     // TODO: hacky solution
     if (model.arch == LLM_ARCH_T5 && t_embd) {
         //cross.t_embd = t_embd;
@@ -996,8 +1008,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
     bool did_optimize = false;
 
-    // handle any pending defrags/shifts
-    kv_self_update(false);
+    // handle any pending shifts/copies
+    memory_update(false);
 
     llama_memory_context_ptr mctx;
 
@@ -1022,7 +1034,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
                     if (!did_optimize) {
                         did_optimize = true;
 
-                        if (kv_self_update(true)) {
+                        if (memory_update(true)) {
                             LLAMA_LOG_DEBUG("%s: retrying batch size %d after cache optimization\n", __func__, balloc->get_n_tokens());
 
                             continue;
@@ -1075,7 +1087,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
         const auto * res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mctx.get(), status);
 
         if (!res) {
-            // the last ubatch failed or was aborted -> remove all positions of that ubatch from the KV cache
+            // the last ubatch failed or was aborted -> remove all positions of that ubatch from the memory module
             llama_pos pos_min[LLAMA_MAX_SEQ];
             for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
                 pos_min[s] = std::numeric_limits<llama_pos>::max();
@@ -1092,7 +1104,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
                     continue;
                 }
 
-                LLAMA_LOG_WARN("%s: removing KV cache entries for seq_id = %d, pos = [%d, +inf)\n", __func__, s, pos_min[s]);
+                LLAMA_LOG_WARN("%s: removing memory module entries for seq_id = %d, pos = [%d, +inf)\n", __func__, s, pos_min[s]);
 
                 memory->seq_rm(s, pos_min[s], -1);
             }
@@ -1243,12 +1255,6 @@ int llama_context::decode(const llama_batch & batch_inp) {
     // wait for the computation to finish (automatically done when obtaining the model output)
     //synchronize();
 
-    if (!supports_set_rows) {
-        // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
-        // overlap with device computation.
-        ggml_backend_sched_reset(sched.get());
-    }
-
     return 0;
 }
 
@@ -1362,8 +1368,9 @@ llm_graph_result * llama_context::get_gf_res_reserve() const {
     return static_cast<llm_graph_result *>(gf_res_reserve.get());
 }
 
-ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx) {
+ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only) {
     LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
+    GGML_ASSERT(n_outputs >= 1);
 
     if (n_tokens % n_seqs != 0) {
         n_tokens = ((n_tokens + (n_seqs - 1)) / n_seqs) * n_seqs; // round to next multiple of n_seqs
@@ -1397,7 +1404,9 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
     this->n_outputs = save_n_outputs;
 
     // initialize scheduler with the specified graph
-    if (!ggml_backend_sched_reserve(sched.get(), gf)) {
+    if (split_only) {
+        ggml_backend_sched_split_graph(sched.get(), gf);
+    } else if (!ggml_backend_sched_reserve(sched.get(), gf)) {
         LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
         return nullptr;
     }
@@ -1437,8 +1446,10 @@ ggml_status llama_context::graph_compute(
     if (backend_cpu != nullptr) {
         auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_cpu));
         auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool");
+        if (set_threadpool_fn) {
             set_threadpool_fn(backend_cpu, tp);
         }
+    }
 
     // set the number of threads for all the backends
     for (const auto & set_n_threads_fn : set_n_threads_fns) {
@@ -1656,30 +1667,30 @@ size_t llama_context::state_set_data(const uint8_t * src, size_t size) {
     }
 }
 
-size_t llama_context::state_seq_get_size(llama_seq_id seq_id) {
+size_t llama_context::state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags) {
     llama_io_write_dummy io;
     try {
-        return state_seq_write_data(io, seq_id);
+        return state_seq_write_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what());
         return 0;
     }
 }
 
-size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size) {
+size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size, llama_state_seq_flags flags) {
     llama_io_write_buffer io(dst, size);
     try {
-        return state_seq_write_data(io, seq_id);
+        return state_seq_write_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what());
         return 0;
     }
 }
 
-size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size) {
+size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags) {
     llama_io_read_buffer io(src, size);
     try {
-        return state_seq_read_data(io, seq_id);
+        return state_seq_read_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what());
         return 0;
@@ -1777,7 +1788,7 @@ size_t llama_context::state_seq_load_file(llama_seq_id seq_id, const char * file
     {
         const size_t state_size = file.size() - file.tell();
         llama_io_read_file io(&file);
-        const size_t nread = state_seq_read_data(io, seq_id);
+        const size_t nread = state_seq_read_data(io, seq_id, 0);
         if (!nread) {
             LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__);
             return 0;
@@ -1801,7 +1812,7 @@ size_t llama_context::state_seq_save_file(llama_seq_id seq_id, const char * file
 
     // save the context state using stream saving
     llama_io_write_file io(&file);
-    state_seq_write_data(io, seq_id);
+    state_seq_write_data(io, seq_id, 0);
 
     const size_t res = file.tell();
     GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + io.n_bytes());
@@ -1876,7 +1887,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
     }
 
     if (memory != nullptr) {
-        LLAMA_LOG_DEBUG("%s: - writing KV self\n", __func__);
+        LLAMA_LOG_DEBUG("%s: - writing memory module\n", __func__);
         memory->state_write(io);
     }
 
@@ -1962,7 +1973,7 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
     }
 
     if (memory) {
-        LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__);
+        LLAMA_LOG_DEBUG("%s: - reading memory module\n", __func__);
 
         memory->state_read(io);
     }
@@ -1970,21 +1981,21 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
     return io.n_bytes();
 }
 
-size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) {
+size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        memory->state_write(io, seq_id);
+        memory->state_write(io, seq_id, flags);
     }
 
     return io.n_bytes();
 }
 
-size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) {
+size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        memory->state_read(io, seq_id);
+        memory->state_read(io, seq_id, flags);
     }
 
     return io.n_bytes();
@@ -2015,6 +2026,21 @@ void llama_context::perf_reset() {
     n_reused    = 0;
 }
 
+std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> llama_context::memory_breakdown() const {
+    std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> ret;
+    for (const auto & buft_size : model.memory_breakdown()) {
+        ret[buft_size.first].model += buft_size.second;
+    }
+    for (const auto & buft_size : memory->memory_breakdown()) {
+        ret[buft_size.first].context += buft_size.second;
+    }
+    for (const auto & backend_ptr : backends) {
+        ggml_backend_t backend = backend_ptr.get();
+        ret[ggml_backend_sched_get_buffer_type(sched.get(), backend)].compute += ggml_backend_sched_get_buffer_size(sched.get(), backend);
+    }
+    return ret;
+}
+
 //
 // training
 //
@@ -2047,7 +2073,7 @@ void llama_context::opt_init(struct llama_model * model, struct llama_opt_params
     opt_params.opt_period      = n_batch / n_ubatch;
     opt_params.get_opt_pars    = lopt_params.get_opt_pars;
     opt_params.get_opt_pars_ud = lopt_params.get_opt_pars_ud;
-
+    opt_params.optimizer       = lopt_params.optimizer_type;
     opt_ctx = ggml_opt_init(opt_params);
 
     llama_opt_param_filter param_filter = lopt_params.param_filter;
@@ -2247,12 +2273,13 @@ llama_context_params llama_context_default_params() {
         /*.rope_scaling_type           =*/ LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED,
         /*.pooling_type                =*/ LLAMA_POOLING_TYPE_UNSPECIFIED,
         /*.attention_type              =*/ LLAMA_ATTENTION_TYPE_UNSPECIFIED,
+        /*.flash_attn_type             =*/ LLAMA_FLASH_ATTN_TYPE_AUTO,
         /*.rope_freq_base              =*/ 0.0f,
         /*.rope_freq_scale             =*/ 0.0f,
         /*.yarn_ext_factor             =*/ -1.0f,
-        /*.yarn_attn_factor            =*/ 1.0f,
-        /*.yarn_beta_fast              =*/ 32.0f,
-        /*.yarn_beta_slow              =*/ 1.0f,
+        /*.yarn_attn_factor            =*/ -1.0f,
+        /*.yarn_beta_fast              =*/ -1.0f,
+        /*.yarn_beta_slow              =*/ -1.0f,
         /*.yarn_orig_ctx               =*/ 0,
         /*.defrag_thold                =*/ -1.0f,
         /*.cb_eval                     =*/ nullptr,
@@ -2263,7 +2290,6 @@ llama_context_params llama_context_default_params() {
         /*.abort_callback_data         =*/ nullptr,
         /*.embeddings                  =*/ false,
         /*.offload_kqv                 =*/ true,
-        /*.flash_attn                  =*/ false,
         /*.no_perf                     =*/ true,
         /*.op_offload                  =*/ true,
         /*.swa_full                    =*/ true,
@@ -2291,12 +2317,30 @@ llama_context * llama_init_from_model(
         return nullptr;
     }
 
-    if (params.flash_attn && model->arch == LLM_ARCH_GROK) {
+    if (params.flash_attn_type != LLAMA_FLASH_ATTN_TYPE_DISABLED && model->arch == LLM_ARCH_GROK) {
         LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
-        params.flash_attn = false;
+        params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_DISABLED;
     }
 
-    if (ggml_is_quantized(params.type_v) && !params.flash_attn) {
+    if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO && ggml_is_quantized(params.type_k)) {
+        const uint32_t blck_size = ggml_blck_size(params.type_k);
+        if (model->hparams.n_embd_head_k % blck_size != 0) {
+            LLAMA_LOG_ERROR("%s: K cache type %s with block size %u does not divide n_embd_head_k=%u\n",
+                __func__, ggml_type_name(params.type_k), blck_size, model->hparams.n_embd_head_k);
+            return nullptr;
+        }
+    }
+
+    if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO && ggml_is_quantized(params.type_v)) {
+        const uint32_t blck_size = ggml_blck_size(params.type_v);
+        if (model->hparams.n_embd_head_v % blck_size != 0) {
+            LLAMA_LOG_ERROR("%s: V cache type %s with block size %u does not divide n_embd_head_k=%u\n",
+                __func__, ggml_type_name(params.type_v), blck_size, model->hparams.n_embd_head_v);
+            return nullptr;
+        }
+    }
+
+    if (ggml_is_quantized(params.type_v) && params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_DISABLED) {
         LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
         return nullptr;
     }
@@ -2342,16 +2386,6 @@ const llama_model * llama_get_model(const llama_context * ctx) {
     return &ctx->get_model();
 }
 
-// deprecated
-llama_kv_cache * llama_get_kv_self(llama_context * ctx) {
-    return dynamic_cast<llama_kv_cache *>(ctx->get_memory());
-}
-
-// deprecated
-void llama_kv_self_update(llama_context * ctx) {
-    ctx->kv_self_update(false);
-}
-
 enum llama_pooling_type llama_pooling_type(const llama_context * ctx) {
     return ctx->pooling_type();
 }
@@ -2569,168 +2603,6 @@ bool llama_memory_can_shift(llama_memory_t mem) {
     return mem->get_can_shift();
 }
 
-//
-// kv cache
-//
-
-// deprecated
-int32_t llama_kv_self_n_tokens(const llama_context * ctx) {
-    const auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return 0;
-    }
-
-    int32_t res = 0;
-
-    for (uint32_t s = 0; s < ctx->get_cparams().n_seq_max; s++) {
-        const llama_pos p0 = kv->seq_pos_min(s);
-        const llama_pos p1 = kv->seq_pos_max(s);
-
-        if (p0 >= 0) {
-            res += (p1 - p0) + 1;
-        }
-    }
-
-    return res;
-}
-
-// deprecated
-// note: this is the same as above - will be removed anyway, so it's ok
-int32_t llama_kv_self_used_cells(const llama_context * ctx) {
-    const auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return 0;
-    }
-
-    int32_t res = 0;
-
-    for (uint32_t s = 0; s < ctx->get_cparams().n_seq_max; s++) {
-        const llama_pos p0 = kv->seq_pos_min(s);
-        const llama_pos p1 = kv->seq_pos_max(s);
-
-        if (p0 >= 0) {
-            res += (p1 - p0) + 1;
-        }
-    }
-
-    return res;
-}
-
-// deprecated
-void llama_kv_self_clear(llama_context * ctx) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_clear(kv, true);
-}
-
-// deprecated
-bool llama_kv_self_seq_rm(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return true;
-    }
-
-    return llama_memory_seq_rm(kv, seq_id, p0, p1);
-}
-
-// deprecated
-void llama_kv_self_seq_cp(
-        llama_context * ctx,
-         llama_seq_id   seq_id_src,
-         llama_seq_id   seq_id_dst,
-            llama_pos   p0,
-            llama_pos   p1) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_cp(kv, seq_id_src, seq_id_dst, p0, p1);
-}
-
-// deprecated
-void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_keep(kv, seq_id);
-}
-
-// deprecated
-void llama_kv_self_seq_add(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1,
-            llama_pos   delta) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_add(kv, seq_id, p0, p1, delta);
-}
-
-// deprecated
-void llama_kv_self_seq_div(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1,
-                  int   d) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_div(kv, seq_id, p0, p1, d);
-}
-
-// deprecated
-llama_pos llama_kv_self_seq_pos_min(llama_context * ctx, llama_seq_id seq_id) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return -1;
-    }
-
-    return llama_memory_seq_pos_min(kv, seq_id);
-}
-
-// deprecated
-llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return -1;
-    }
-
-    return llama_memory_seq_pos_max(kv, seq_id);
-}
-
-// deprecated
-void llama_kv_self_defrag(llama_context * ctx) {
-    // force defrag
-    ctx->kv_self_defrag_sched();
-}
-
-// deprecated
-bool llama_kv_self_can_shift(const llama_context * ctx) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return false;
-    }
-
-    return llama_memory_can_shift(kv);
-}
-
 // llama state API
 
 // deprecated
@@ -2800,19 +2672,31 @@ bool llama_state_save_file(llama_context * ctx, const char * path_session, const
 }
 
 size_t llama_state_seq_get_size(llama_context * ctx, llama_seq_id seq_id) {
-    return ctx->state_seq_get_size(seq_id);
+    return llama_state_seq_get_size_ext(ctx, seq_id, 0);
 }
 
 size_t llama_state_seq_get_data(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id) {
-    ctx->synchronize();
-
-    return ctx->state_seq_get_data(seq_id, dst, size);
+    return llama_state_seq_get_data_ext(ctx, dst, size, seq_id, 0);
 }
 
 size_t llama_state_seq_set_data(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id) {
+    return llama_state_seq_set_data_ext(ctx, src, size, seq_id, 0);
+}
+
+size_t llama_state_seq_get_size_ext(llama_context * ctx, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    return ctx->state_seq_get_size(seq_id, flags);
+}
+
+size_t llama_state_seq_get_data_ext(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id, llama_state_seq_flags flags) {
     ctx->synchronize();
 
-    return ctx->state_seq_set_data(seq_id, src, size);
+    return ctx->state_seq_get_data(seq_id, dst, size, flags);
+}
+
+size_t llama_state_seq_set_data_ext(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    ctx->synchronize();
+
+    return ctx->state_seq_set_data(seq_id, src, size, flags);
 }
 
 size_t llama_state_seq_save_file(llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) {
@@ -2895,6 +2779,142 @@ void llama_perf_context_reset(llama_context * ctx) {
     ctx->perf_reset();
 }
 
+void llama_memory_breakdown_print(const struct llama_context * ctx) {
+    const std::vector<ggml_backend_dev_t> & devices = ctx->get_model().devices;
+
+    std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> memory_breakdown = ctx->memory_breakdown();
+
+    std::vector<std::array<std::string, 9>> table_data;
+    table_data.reserve(devices.size());
+    const std::string template_header = "%s: | %s | %s   %s    %s   %s   %s   %s    %s |\n";
+    const std::string template_gpu    = "%s: | %s | %s = %s + (%s = %s + %s + %s) + %s |\n";
+    const std::string template_other  = "%s: | %s | %s   %s    %s = %s + %s + %s    %s |\n";
+
+    table_data.push_back({template_header, "memory breakdown [MiB]", "total", "free", "self", "model", "context", "compute", "unaccounted"});
+
+    constexpr size_t MiB = 1024 * 1024;
+    const std::vector<std::string> desc_prefixes_strip = {"NVIDIA ", "GeForce ", "Tesla ", "AMD ", "Radeon ", "Instinct "};
+
+    // track seen buffer types to avoid double counting:
+    std::set<ggml_backend_buffer_type_t> seen_buffer_types;
+
+    // accumulative memory breakdown for each device and for host:
+    std::vector<llama_memory_breakdown_data> mb_dev(devices.size());
+    llama_memory_breakdown_data              mb_host;
+
+    for (const auto & buft_mb : memory_breakdown) {
+        ggml_backend_buffer_type_t          buft = buft_mb.first;
+        const llama_memory_breakdown_data & mb   = buft_mb.second;
+        if (ggml_backend_buft_is_host(buft)) {
+            mb_host.model   += mb.model;
+            mb_host.context += mb.context;
+            mb_host.compute += mb.compute;
+            seen_buffer_types.insert(buft);
+            continue;
+        }
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        if (dev) {
+            int i_dev = -1;
+            for (size_t i = 0; i < devices.size(); i++) {
+                if (devices[i] == dev) {
+                    i_dev = i;
+                    break;
+                }
+            }
+            if (i_dev != -1) {
+                mb_dev[i_dev].model   += mb.model;
+                mb_dev[i_dev].context += mb.context;
+                mb_dev[i_dev].compute += mb.compute;
+                seen_buffer_types.insert(buft);
+                continue;
+            }
+        }
+    }
+
+    // print memory breakdown for each device:
+    for (size_t i = 0; i < devices.size(); i++) {
+        ggml_backend_dev_t          dev = devices[i];
+        llama_memory_breakdown_data mb  = mb_dev[i];
+
+        const std::string name = ggml_backend_dev_name(dev);
+        std::string desc = ggml_backend_dev_description(dev);
+        for (const std::string & prefix : desc_prefixes_strip) {
+            if (desc.length() >= prefix.length() && desc.substr(0, prefix.length()) == prefix) {
+                desc = desc.substr(prefix.length());
+            }
+        }
+
+        size_t free, total;
+        ggml_backend_dev_memory(dev, &free, &total);
+
+        const size_t self = mb.model + mb.context + mb.compute;
+        const size_t unaccounted = total - self - free;
+
+        table_data.push_back({
+            template_gpu,
+            "  - " + name + " (" + desc + ")",
+            std::to_string(total / MiB),
+            std::to_string(free / MiB),
+            std::to_string(self / MiB),
+            std::to_string(mb.model / MiB),
+            std::to_string(mb.context / MiB),
+            std::to_string(mb.compute / MiB),
+            std::to_string(unaccounted / MiB)});
+    }
+
+    // print memory breakdown for host:
+    {
+        const size_t self = mb_host.model + mb_host.context + mb_host.compute;
+        table_data.push_back({
+            template_other,
+            "  - Host",
+            "", // total
+            "", // free
+            std::to_string(self / MiB),
+            std::to_string(mb_host.model / MiB),
+            std::to_string(mb_host.context / MiB),
+            std::to_string(mb_host.compute / MiB),
+            ""}); // unaccounted
+    }
+
+    // print memory breakdown for all remaining buffer types:
+    for (const auto & buft_mb : memory_breakdown) {
+        ggml_backend_buffer_type_t          buft = buft_mb.first;
+        const llama_memory_breakdown_data & mb   = buft_mb.second;
+        if (seen_buffer_types.count(buft) == 1) {
+            continue;
+        }
+        const std::string name = ggml_backend_buft_name(buft);
+        const size_t self = mb.model + mb.context + mb.compute;
+        table_data.push_back({
+            template_other,
+            "  - " + name,
+            "", // total
+            "", // free
+            std::to_string(self / MiB),
+            std::to_string(mb.model / MiB),
+            std::to_string(mb.context / MiB),
+            std::to_string(mb.compute / MiB),
+            ""}); // unaccounted
+        seen_buffer_types.insert(buft);
+    }
+
+    for (size_t j = 1; j < table_data[0].size(); j++) {
+        size_t max_len = 0;
+        for (const auto & td : table_data) {
+            max_len = std::max(max_len, td[j].length());
+        }
+        for (auto & td : table_data) {
+            td[j].insert(j == 1 ? td[j].length() : 0, max_len - td[j].length(), ' ');
+        }
+    }
+    for (const auto & td : table_data) {
+        LLAMA_LOG_INFO(td[0].c_str(),
+            __func__, td[1].c_str(), td[2].c_str(), td[3].c_str(), td[4].c_str(), td[5].c_str(),
+            td[6].c_str(), td[7].c_str(), td[8].c_str());
+    }
+}
+
 //
 // training
 //

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

@@ -17,9 +17,17 @@ class llama_batch_allocr;
 class llama_io_read_i;
 class llama_io_write_i;
 
+// "memory" as in abstract memory for the context
 struct llama_memory_i;
 struct llama_memory_context_i;
 
+// "memory" as in physical memory for a buffer type, in bytes
+struct llama_memory_breakdown_data {
+    size_t model   = 0; // memory allocated for the model
+    size_t context = 0; // memory allocated for the context
+    size_t compute = 0; // memory allocated for temporary compute buffers
+};
+
 struct llama_context {
     // init scheduler and compute buffers, reserve worst-case graphs
     llama_context(
@@ -46,10 +54,8 @@ struct llama_context {
 
     llama_memory_t get_memory() const;
 
-    // return true of the KV cache was updated
-    // TODO: remove
-    bool kv_self_update(bool optimize);
-    void kv_self_defrag_sched();
+    // return true if the memory was updated
+    bool memory_update(bool optimize);
 
     enum llama_pooling_type pooling_type() const;
 
@@ -111,9 +117,9 @@ struct llama_context {
     size_t state_get_data(      uint8_t * dst, size_t size);
     size_t state_set_data(const uint8_t * src, size_t size);
 
-    size_t state_seq_get_size(llama_seq_id seq_id);
-    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size);
-    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size);
+    size_t state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags);
+    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size, llama_state_seq_flags flags);
+    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags);
 
     bool state_load_file(
             const char * filepath,
@@ -146,12 +152,15 @@ struct llama_context {
     llama_perf_context_data perf_get_data() const;
     void perf_reset();
 
+    std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> memory_breakdown() const;
+
     //
     // training
     //
 
     void opt_init(struct llama_model * model, struct llama_opt_params lopt_params);
 
+    // TODO: more flexible combinations of logical/physical batch size and context size
     void opt_epoch(
             ggml_opt_dataset_t      dataset,
             ggml_opt_result_t       result_train,
@@ -197,7 +206,7 @@ public:
     ggml_status graph_compute(ggml_cgraph * gf, bool batched);
 
     // reserve a graph with a dummy ubatch of the specified size
-    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx);
+    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false);
 
 private:
     llm_graph_params graph_params(
@@ -212,8 +221,8 @@ private:
     size_t state_write_data(llama_io_write_i & io);
     size_t state_read_data (llama_io_read_i  & io);
 
-    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id);
-    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id);
+    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags);
+    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id, llama_state_seq_flags flags);
 
     //
     // members
@@ -229,9 +238,6 @@ private:
 
     std::unique_ptr<llama_memory_i> memory;
 
-    // TODO: temporary, until the llama_kv_self_defrag() API is removed
-    bool memory_force_optimize = false;
-
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     size_t  logits_size = 0; // capacity (of floats) for logits
     float * logits      = nullptr;
@@ -287,10 +293,6 @@ private:
 
     bool has_evaluated_once = false;
 
-    // env: LLAMA_SET_ROWS (temporary)
-    // ref: https://github.com/ggml-org/llama.cpp/pull/14285
-    bool supports_set_rows = true;
-
     // env: LLAMA_GRAPH_REUSE_DISABLE
     bool graph_reuse_disable = false;
 

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

@@ -4,7 +4,7 @@
 
 #include <cstdint>
 
-#define LLAMA_MAX_SEQ 64
+#define LLAMA_MAX_SEQ 256
 
 struct llama_cparams {
     uint32_t n_ctx;           // context size used during inference
@@ -24,7 +24,6 @@ struct llama_cparams {
     float yarn_attn_factor;
     float yarn_beta_fast;
     float yarn_beta_slow;
-    float defrag_thold;
 
     bool embeddings;
     bool causal_attn;

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

@@ -4,8 +4,8 @@
 #include "llama-batch.h"
 #include "llama-cparams.h"
 
-#include "llama-kv-cache-unified.h"
-#include "llama-kv-cache-unified-iswa.h"
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-iswa.h"
 #include "llama-memory-hybrid.h"
 #include "llama-memory-recurrent.h"
 
@@ -204,7 +204,10 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
         std::vector<int> target_pos(n_seqs_unq, -1);
         std::vector<int> target_row(n_seqs_unq, -1);
 
-        bool last = cparams.pooling_type == LLAMA_POOLING_TYPE_LAST;
+        const bool last = (
+             cparams.pooling_type == LLAMA_POOLING_TYPE_LAST ||
+            (cparams.pooling_type == LLAMA_POOLING_TYPE_RANK && arch == LLM_ARCH_QWEN3) // qwen3 reranking & embedding models use last token
+        );
 
         for (int i = 0; i < n_tokens; ++i) {
             const llama_pos pos = ubatch->pos[i];
@@ -258,6 +261,36 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+static void print_mask(float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
+    LLAMA_LOG_DEBUG("%s: === Attention mask ===\n", __func__);
+    const char * swa_type_str = (swa_type == LLAMA_SWA_TYPE_NONE) ? "LLAMA_SWA_TYPE_NONE" :
+                          (swa_type == LLAMA_SWA_TYPE_STANDARD) ? "LLAMA_SWA_TYPE_STANDARD" :
+                          (swa_type == LLAMA_SWA_TYPE_CHUNKED) ? "LLAMA_SWA_TYPE_CHUNKED" :
+                          (swa_type == LLAMA_SWA_TYPE_SYMMETRIC) ? "LLAMA_SWA_TYPE_SYMMETRIC" : "unknown";
+    LLAMA_LOG_DEBUG("%s: n_swa : %d, n_kv: %d, swq_type: %s\n", __func__, (int)n_swa, (int)n_kv, swa_type_str);
+    LLAMA_LOG_DEBUG("%s: '0' = can attend, '∞' = masked\n", __func__);
+    LLAMA_LOG_DEBUG("%s: Rows = query tokens, Columns = key/value tokens\n\n", __func__);
+
+    LLAMA_LOG_DEBUG("    ");
+    for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
+        LLAMA_LOG_DEBUG("%2d", j);
+    }
+    LLAMA_LOG_DEBUG("\n");
+
+    for (int i = 0; i < std::min((int64_t)20, n_tokens); ++i) {
+        LLAMA_LOG_DEBUG(" %2d ", i);
+        for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
+            float val = data[i * n_kv + j];
+            if (val == -INFINITY) {
+                LLAMA_LOG_DEBUG(" ∞");
+            } else {
+                LLAMA_LOG_DEBUG(" 0");
+            }
+        }
+        LLAMA_LOG_DEBUG("\n");
+    }
+}
+
 void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
     const int64_t n_kv     = ubatch->n_tokens;
     const int64_t n_tokens = ubatch->n_tokens;
@@ -267,6 +300,9 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
 
     float * data = (float *) kq_mask->data;
 
+    // [TAG_NO_CACHE_ISWA]
+    GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "TODO: implement");
+
     for (int h = 0; h < 1; ++h) {
         for (int i1 = 0; i1 < n_tokens; ++i1) {
             const llama_seq_id s1 = ubatch->seq_id[i1][0];
@@ -277,32 +313,44 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
                 for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) {
                     const llama_seq_id s0 = ubatch->seq_id[i0][0];
 
+                    if (s0 != s1) {
+                        continue; // skip different sequences
+                    }
+
+                    if (cparams.causal_attn && ubatch->pos[i0] > ubatch->pos[i1]) {
+                        continue; // skip future tokens for causal attention
+                    }
+
+                    // TODO: this does not take into account that some layers are SWA and others are note (i.e. iSWA) [TAG_NO_CACHE_ISWA]
+                    //if (hparams.is_masked_swa(ubatch->pos[i0], ubatch->pos[i1])) {
+                    //    continue; // skip masked tokens for SWA
+                    //}
+
                     // TODO: reimplement this like in llama_kv_cache_unified
-                    if (s0 == s1 && (!cparams.causal_attn || ubatch->pos[i0] <= ubatch->pos[i1])) {
                     if (hparams.use_alibi) {
                         f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
                     } else {
                         f = 0.0f;
                     }
-                        break;
                 }
-                }
-
                 data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f;
             }
         }
     }
+    if (debug) {
+        print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
+    }
 }
 
-void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
+void llm_graph_input_attn_kv::set_input(const llama_ubatch * ubatch) {
     mctx->set_input_k_idxs(self_k_idxs, ubatch);
     mctx->set_input_v_idxs(self_v_idxs, ubatch);
 
     mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
 }
 
-bool llm_graph_input_attn_kv_unified::can_reuse(const llm_graph_params & params) {
-    const auto * mctx = static_cast<const llama_kv_cache_unified_context *>(params.mctx);
+bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_context *>(params.mctx);
 
     this->mctx = mctx;
 
@@ -314,12 +362,10 @@ bool llm_graph_input_attn_kv_unified::can_reuse(const llm_graph_params & params)
     res &= self_kq_mask->ne[0] == mctx->get_n_kv();
     res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
 
-    res &= mctx->get_supports_set_rows(); // TODO: tmp
-
     return res;
 }
 
-void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
+void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
     mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
     mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);
 
@@ -331,8 +377,8 @@ void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch
     mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
 }
 
-bool llm_graph_input_attn_kv_unified_iswa::can_reuse(const llm_graph_params & params) {
-    const auto * mctx = static_cast<const llama_kv_cache_unified_iswa_context *>(params.mctx);
+bool llm_graph_input_attn_kv_iswa::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_iswa_context *>(params.mctx);
 
     this->mctx = mctx;
 
@@ -350,8 +396,6 @@ bool llm_graph_input_attn_kv_unified_iswa::can_reuse(const llm_graph_params & pa
     res &= self_kq_mask_swa->ne[0] == mctx->get_swa()->get_n_kv();
     res &= self_kq_mask_swa->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
 
-    res &= mctx->get_base()->get_supports_set_rows(); // TODO: tmp
-
     return res;
 }
 
@@ -879,15 +923,29 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         selection_probs = logits;
     }
 
+    if (arch == LLM_ARCH_GROVEMOE) {
+        selection_probs = ggml_sigmoid(ctx0, logits); // [n_expert, n_tokens]
+        cb(selection_probs, "ffn_moe_probs_biased", il);
+    }
+
     // select experts
     ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
     cb(selected_experts->src[0], "ffn_moe_argsort", il);
     cb(selected_experts, "ffn_moe_topk", il);
 
-    ggml_tensor * weights = ggml_get_rows(ctx0,
-            ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens]
+    if (arch == LLM_ARCH_GROVEMOE && n_expert != hparams.n_expert) {
+        // TODO: Use scalar div instead when/if implemented
+        ggml_tensor * f_sel = ggml_cast(ctx0, selected_experts, GGML_TYPE_F32);
+        selected_experts = ggml_cast(ctx0, ggml_scale(ctx0, f_sel, 1.0f / float(hparams.n_group_experts)), GGML_TYPE_I32);
+        probs = ggml_reshape_3d(ctx0, probs, 1, hparams.n_expert, n_tokens);
+    } else {
+        probs = ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens);
+    }
+
+    ggml_tensor * weights = ggml_get_rows(ctx0, probs, selected_experts); // [1, n_expert_used, n_tokens]
     cb(weights, "ffn_moe_weights", il);
 
+
     if (gating_op == LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT) {
         weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);
         weights = ggml_soft_max(ctx0, weights); // [n_expert_used, n_tokens]
@@ -911,6 +969,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         cb(weights, "ffn_moe_weights_scaled", il);
     }
 
+    //call early so that topk-moe can be used
+    ggml_build_forward_expand(gf, weights);
+
     cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
 
     if (weight_before_ffn) {
@@ -1136,7 +1197,7 @@ ggml_tensor * llm_graph_context::build_inp_mean() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_cls() const {
-    auto inp = std::make_unique<llm_graph_input_cls>(cparams);
+    auto inp = std::make_unique<llm_graph_input_cls>(cparams, arch);
 
     auto & cur = inp->cls;
 
@@ -1186,7 +1247,7 @@ ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
+    const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx);
 
     auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, mctx_cur);
 
@@ -1223,15 +1284,16 @@ ggml_tensor * llm_graph_context::build_attn_mha(
          ggml_tensor * v,
          ggml_tensor * kq_b,
          ggml_tensor * kq_mask,
-         ggml_tensor * v_mla,
          ggml_tensor * sinks,
-             float     kq_scale) const {
+         ggml_tensor * v_mla,
+               float   kq_scale,
+                 int   il) const {
     const bool v_trans = v->nb[1] > v->nb[2];
 
     // split the batch into streams if needed
     const auto n_stream = k->ne[3];
 
-    q = ggml_reshape_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream);
+    q = ggml_view_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream, q->nb[1], q->nb[2], q->nb[3]/n_stream, 0);
 
     q = ggml_permute(ctx0, q, 0, 2, 1, 3);
     k = ggml_permute(ctx0, k, 0, 2, 1, 3);
@@ -1260,6 +1322,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
                                   hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
+        cb(cur, LLAMA_TENSOR_NAME_FATTN, il);
 
         ggml_flash_attn_ext_add_sinks(cur, sinks);
         ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);
@@ -1275,6 +1338,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
             // 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);
+            cb(cur, "fattn_mla", il);
             cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
             cur = ggml_cont(ctx0, cur); // Needed because ggml_reshape_2d expects contiguous inputs.
 #endif
@@ -1283,6 +1347,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
         cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
     } else {
         ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+        cb(kq, "kq", il);
 
         // note: this op tends to require high floating point range
         //       while for some models F16 is enough, for others it is not, so we default to F32 here
@@ -1290,38 +1355,48 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         if (arch == LLM_ARCH_GROK) {
             // need to do the following:
-            // multiply by attn_output_multiplyer of 0.08838834764831845
+            // multiply by attn_output_multiplier
             // and then :
             // kq = 30 * tanh(kq / 30)
             // before the softmax below
 
-            kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, 0.08838834764831845f/30.0f));
-            kq = ggml_scale(ctx0, kq, 30);
+            kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, hparams.f_attn_out_scale / hparams.f_attn_logit_softcapping));
+            cb(kq, "kq_tanh", il);
+            kq = ggml_scale(ctx0, kq, hparams.f_attn_logit_softcapping);
+            cb(kq, "kq_scaled", il);
         }
 
         if (hparams.attn_soft_cap) {
             kq = ggml_scale(ctx0, kq, 1.0f / hparams.f_attn_logit_softcapping);
+            cb(kq, "kq_scaled_1", il);
             kq = ggml_tanh (ctx0, kq);
+            cb(kq, "kq_tanh", il);
             kq = ggml_scale(ctx0, kq, hparams.f_attn_logit_softcapping);
+            cb(kq, "kq_scaled_2", il);
         }
 
         if (kq_b) {
             kq = ggml_add(ctx0, kq, kq_b);
+            cb(kq, "kq_plus_kq_b", il);
         }
 
         kq = ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
         ggml_soft_max_add_sinks(kq, sinks);
+        cb(kq, "kq_soft_max", il);
 
         if (!v_trans) {
             // note: avoid this branch
             v = ggml_cont(ctx0, ggml_transpose(ctx0, v));
+            cb(v, "v_cont", il);
         }
 
         ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
+        cb(kqv, "kqv", il);
 
         // for MLA with the absorption optimization, we need to "decompress" from MQA back to MHA
         if (v_mla) {
             kqv = ggml_mul_mat(ctx0, v_mla, kqv);
+            cb(kqv, "kqv_mla", il);
         }
 
         cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
@@ -1360,6 +1435,7 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * sinks,
         ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
@@ -1375,13 +1451,14 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // [TAG_NO_CACHE_PAD]
     // TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams
-    assert(!ubatch.equal_seqs());
+    //       but it might not be worth it: https://github.com/ggml-org/llama.cpp/pull/15636
+    //assert(!ubatch.equal_seqs() || (k_cur->ne[3] == 1 && k_cur->ne[3] == ubatch.n_seqs_unq));
 
     ggml_tensor * q = q_cur;
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1399,17 +1476,17 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-static std::unique_ptr<llm_graph_input_attn_kv_unified> build_attn_inp_kv_unified_impl(
+static std::unique_ptr<llm_graph_input_attn_kv> build_attn_inp_kv_impl(
            ggml_context * ctx0,
      const llama_ubatch & ubatch,
     const llama_hparams & hparams,
     const llama_cparams & cparams,
-    const llama_kv_cache_unified_context * mctx_cur) {
+    const llama_kv_cache_context * mctx_cur) {
 
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, mctx_cur);
+    auto inp = std::make_unique<llm_graph_input_attn_kv>(hparams, cparams, mctx_cur);
 
     {
-        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA");
+        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA");
 
         const auto n_kv     = mctx_cur->get_n_kv();
         const auto n_tokens = ubatch.n_tokens;
@@ -1427,22 +1504,23 @@ static std::unique_ptr<llm_graph_input_attn_kv_unified> build_attn_inp_kv_unifie
     return inp;
 }
 
-llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
+llm_graph_input_attn_kv * llm_graph_context::build_attn_inp_kv() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx);
 
-    auto inp = build_attn_inp_kv_unified_impl(ctx0, ubatch, hparams, cparams, mctx_cur);
+    auto inp = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur);
 
-    return (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp));
+    return (llm_graph_input_attn_kv *) res->add_input(std::move(inp));
 }
 
 ggml_tensor * llm_graph_context::build_attn(
-        llm_graph_input_attn_kv_unified * inp,
+        llm_graph_input_attn_kv * inp,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * sinks,
         ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
@@ -1469,7 +1547,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1488,40 +1566,15 @@ ggml_tensor * llm_graph_context::build_attn(
 }
 
 ggml_tensor * llm_graph_context::build_attn(
-        llm_graph_input_attn_kv_unified_iswa * inp,
+        llm_graph_input_attn_kv_iswa * inp,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
-        ggml_tensor * v_mla,
-            float     kq_scale,
-            int       il) const {
-    return build_attn_with_sinks(
-            inp,
-            wo,
-            wo_b,
-            q_cur,
-            k_cur,
-            v_cur,
-            kq_b,
-            v_mla,
-            nullptr,
-            kq_scale,
-            il);
-}
-
-ggml_tensor * llm_graph_context::build_attn_with_sinks(
-        llm_graph_input_attn_kv_unified_iswa * inp,
-        ggml_tensor * wo,
-        ggml_tensor * wo_b,
-        ggml_tensor * q_cur,
-        ggml_tensor * k_cur,
-        ggml_tensor * v_cur,
-        ggml_tensor * kq_b,
-        ggml_tensor * v_mla,
         ggml_tensor * sinks,
+        ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
     // these nodes are added to the graph together so that they are not reordered
@@ -1561,7 +1614,7 @@ ggml_tensor * llm_graph_context::build_attn_with_sinks(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, sinks, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1600,6 +1653,7 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * sinks,
         ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
@@ -1615,7 +1669,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1636,10 +1690,10 @@ ggml_tensor * llm_graph_context::build_attn(
 // TODO: maybe separate the inner implementation into a separate function
 //       like with the non-sliding window equivalent
 //       once sliding-window hybrid caches are a thing.
-llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_iswa_context *>(mctx);
+llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_iswa_context *>(mctx);
 
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, mctx_cur);
+    auto inp = std::make_unique<llm_graph_input_attn_kv_iswa>(hparams, cparams, mctx_cur);
 
     const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
 
@@ -1656,7 +1710,7 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
     }
 
     {
-        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified for non-SWA");
+        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache for non-SWA");
 
         const auto n_kv = mctx_cur->get_swa()->get_n_kv();
 
@@ -1669,7 +1723,7 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
     }
 
-    return (llm_graph_input_attn_kv_unified_iswa *) res->add_input(std::move(inp));
+    return (llm_graph_input_attn_kv_iswa *) res->add_input(std::move(inp));
 }
 
 ggml_tensor * llm_graph_context::build_rs(
@@ -1792,7 +1846,7 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
     const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx);
 
     auto inp_rs   = build_rs_inp_impl(ctx0, ubatch, mctx_cur->get_recr());
-    auto inp_attn = build_attn_inp_kv_unified_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn());
+    auto inp_attn = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn());
 
     auto inp = std::make_unique<llm_graph_input_mem_hybrid>(std::move(inp_attn), std::move(inp_rs), mctx_cur);
 
@@ -1843,34 +1897,32 @@ void llm_graph_context::build_pooling(
         case LLAMA_POOLING_TYPE_RANK:
             {
                 ggml_tensor * inp_cls = build_inp_cls();
-                inp = ggml_get_rows(ctx0, inp, inp_cls);
+                cur = ggml_get_rows(ctx0, inp, inp_cls);
 
-                if (cls) {
                 // classification head
                 // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
-                    cur = ggml_mul_mat(ctx0, cls, inp);
+                if (cls) {
+                    cur = ggml_mul_mat(ctx0, cls, cur);
                     if (cls_b) {
                         cur = ggml_add(ctx0, cur, cls_b);
                     }
                     cur = ggml_tanh(ctx0, cur);
+                }
 
                 // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
                 // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
+                // Single layer classification head (direct projection)
+                // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
                 if (cls_out) {
                     cur = ggml_mul_mat(ctx0, cls_out, cur);
                     if (cls_out_b) {
                         cur = ggml_add(ctx0, cur, cls_out_b);
                     }
                 }
-                } else if (cls_out) {
-                    // Single layer classification head (direct projection)
-                    // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
-                    cur = ggml_mul_mat(ctx0, cls_out, inp);
-                    if (cls_out_b) {
-                        cur = ggml_add(ctx0, cur, cls_out_b);
-                    }
-                } else {
-                    GGML_ABORT("RANK pooling requires either cls+cls_b or cls_out+cls_out_b");
+
+                // softmax for qwen3 reranker
+                if (arch == LLM_ARCH_QWEN3) {
+                    cur = ggml_soft_max(ctx0, cur);
                 }
             } break;
         default:

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

@@ -19,8 +19,8 @@ struct llama_cparams;
 
 struct llama_memory_context_i;
 
-class llama_kv_cache_unified_context;
-class llama_kv_cache_unified_iswa_context;
+class llama_kv_cache_context;
+class llama_kv_cache_iswa_context;
 class llama_memory_recurrent_context;
 class llama_memory_hybrid_context;
 
@@ -78,6 +78,11 @@ struct llm_graph_params;
 
 class llm_graph_input_i {
 public:
+    llm_graph_input_i() {
+        const char * LLAMA_GRAPH_INPUT_DEBUG = getenv("LLAMA_GRAPH_INPUT_DEBUG");
+        debug = LLAMA_GRAPH_INPUT_DEBUG ? atoi(LLAMA_GRAPH_INPUT_DEBUG) : 0;
+    }
+
     virtual ~llm_graph_input_i() = default;
 
     virtual void set_input(const llama_ubatch * ubatch) = 0;
@@ -90,6 +95,9 @@ public:
         GGML_UNUSED(params);
         return false;
     }
+protected:
+    // env: LLAMA_GRAPH_INPUT_DEBUG
+    int debug = 0;
 };
 
 using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
@@ -152,7 +160,7 @@ class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
 public:
     llm_graph_input_pos_bucket_kv(
             const llama_hparams & hparams,
-            const llama_kv_cache_unified_context * mctx) : hparams(hparams), mctx(mctx) {}
+            const llama_kv_cache_context * mctx) : hparams(hparams), mctx(mctx) {}
     virtual ~llm_graph_input_pos_bucket_kv() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -161,7 +169,7 @@ public:
 
     const llama_hparams hparams;
 
-    const llama_kv_cache_unified_context * mctx;
+    const llama_kv_cache_context * mctx;
 };
 
 class llm_graph_input_out_ids : public llm_graph_input_i {
@@ -198,7 +206,7 @@ public:
 
 class llm_graph_input_cls : public llm_graph_input_i {
 public:
-    llm_graph_input_cls(const llama_cparams & cparams) : cparams(cparams) {}
+    llm_graph_input_cls(const llama_cparams & cparams, const llm_arch arch) : cparams(cparams), arch(arch) {}
     virtual ~llm_graph_input_cls() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -206,6 +214,7 @@ public:
     ggml_tensor * cls; // I32 [n_batch]
 
     const llama_cparams cparams;
+    const llm_arch arch;
 };
 
 class llm_graph_input_rs : public llm_graph_input_i {
@@ -257,17 +266,17 @@ public:
     const llama_cparams cparams;
 };
 
-class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
+class llm_graph_input_attn_kv : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_kv_unified(
+    llm_graph_input_attn_kv(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified_context * mctx) :
+            const llama_kv_cache_context * mctx) :
         hparams(hparams),
         cparams(cparams),
         mctx(mctx) {
     }
-    ~llm_graph_input_attn_kv_unified() = default;
+    ~llm_graph_input_attn_kv() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
@@ -290,20 +299,20 @@ public:
     const llama_hparams hparams;
     const llama_cparams cparams;
 
-    const llama_kv_cache_unified_context * mctx;
+    const llama_kv_cache_context * mctx;
 };
 
-class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
+class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_kv_unified_iswa(
+    llm_graph_input_attn_kv_iswa(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified_iswa_context * mctx) :
+            const llama_kv_cache_iswa_context * mctx) :
         hparams(hparams),
         cparams(cparams),
         mctx(mctx) {
     }
-    ~llm_graph_input_attn_kv_unified_iswa() = default;
+    ~llm_graph_input_attn_kv_iswa() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
@@ -330,7 +339,7 @@ public:
     const llama_hparams hparams;
     const llama_cparams cparams;
 
-    const llama_kv_cache_unified_iswa_context * mctx;
+    const llama_kv_cache_iswa_context * mctx;
 };
 
 class llm_graph_input_attn_cross : public llm_graph_input_i {
@@ -351,7 +360,7 @@ public:
 class llm_graph_input_mem_hybrid : public llm_graph_input_i {
 public:
     llm_graph_input_mem_hybrid(
-            std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn,
+            std::unique_ptr<llm_graph_input_attn_kv> inp_attn,
             std::unique_ptr<llm_graph_input_rs>              inp_rs,
             const llama_memory_hybrid_context *              mctx) :
         inp_attn(std::move(inp_attn)),
@@ -361,10 +370,10 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
-    std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn;
+    std::unique_ptr<llm_graph_input_attn_kv> inp_attn;
     std::unique_ptr<llm_graph_input_rs>      inp_rs;
 
-    llm_graph_input_attn_kv_unified * get_attn() const { return inp_attn.get(); }
+    llm_graph_input_attn_kv * get_attn() const { return inp_attn.get(); }
     llm_graph_input_rs      * get_recr() const { return inp_rs.get(); }
 
     const llama_memory_hybrid_context * mctx;
@@ -685,9 +694,10 @@ struct llm_graph_context {
             ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
             ggml_tensor * kq_b,
             ggml_tensor * kq_mask,
-             ggml_tensor * sinks,
+            ggml_tensor * sinks,   // [n_head_q]
             ggml_tensor * v_mla,   // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
-                   float   kq_scale) const;
+                  float   kq_scale,
+                    int   il) const;
 
     llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const;
 
@@ -699,50 +709,39 @@ struct llm_graph_context {
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
-    llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified() const;
+    llm_graph_input_attn_kv * build_attn_inp_kv() const;
 
     ggml_tensor * build_attn(
-            llm_graph_input_attn_kv_unified * inp,
+            llm_graph_input_attn_kv * inp,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
-    llm_graph_input_attn_kv_unified_iswa * build_attn_inp_kv_unified_iswa() const;
+    llm_graph_input_attn_kv_iswa * build_attn_inp_kv_iswa() const;
 
     // note: if k_cur or v_cur are not provided, they will not be stored in the memory
     ggml_tensor * build_attn(
-            llm_graph_input_attn_kv_unified_iswa * inp,
+            llm_graph_input_attn_kv_iswa * inp,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
             ggml_tensor * kq_b,
-            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
-                  float   kq_scale,
-                    int   il) const;
-
-    // TODO: temporary to keep the diff small. after the code is public will refactor to simplify this
-    ggml_tensor * build_attn_with_sinks(
-            llm_graph_input_attn_kv_unified_iswa * inp,
-            ggml_tensor * wo,
-            ggml_tensor * wo_b,
-            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
-            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
-            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
-            ggml_tensor * kq_b,
-            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
             ggml_tensor * sinks, // [n_head_q]
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
@@ -756,6 +755,7 @@ struct llm_graph_context {
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
@@ -765,7 +765,7 @@ struct llm_graph_context {
     //
 
     // TODO: move this implementation to llama_memory_recurrent.
-    //       this is analogous to llama_kv_cache_unified::cpy_k / cpy_v
+    //       this is analogous to llama_kv_cache::cpy_k / cpy_v
     //       when moving, avoid passing `ggml_cgraph` - only pass `ggml_context`. would likely need to split the
     //         implementation in 2 separate methods. the goal is to avoid calling `ggml_build_forward_expand` in
     //         `llama_memory_recurrent`

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

@@ -1,6 +1,7 @@
 #include "llama-hparams.h"
 
 #include "ggml.h"
+#include <cassert>
 
 void llama_hparams::set_swa_pattern(uint32_t n_pattern, bool dense_first) {
     if (dense_first) {
@@ -161,3 +162,64 @@ bool llama_hparams::is_swa(uint32_t il) const {
 
     GGML_ABORT("fatal error");
 }
+
+bool llama_hparams::has_kv(uint32_t il) const {
+    if (n_layer_kv_from_start >= 0) {
+        if (il < (uint32_t) n_layer_kv_from_start) {
+            return true;
+        }
+
+        return false;
+    }
+
+    // by default, all layers have kv
+    return true;
+}
+
+uint32_t llama_hparams::n_layer_kv() const {
+    uint32_t res = 0;
+
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (has_kv(il)) {
+            res++;
+        }
+    }
+
+    return res;
+}
+
+bool llama_hparams::is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1) {
+    assert(p0 >= 0 && p1 >= 0);
+
+    switch (swa_type) {
+        case LLAMA_SWA_TYPE_NONE:
+            {
+            } break;
+        case LLAMA_SWA_TYPE_STANDARD:
+            {
+                if (p1 - p0 >= (int32_t) n_swa) {
+                    return true;
+                }
+            } break;
+        case LLAMA_SWA_TYPE_CHUNKED:
+            {
+                const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
+
+                if (p0 < pos_chunk_start) {
+                    return true;
+                }
+            } break;
+        case LLAMA_SWA_TYPE_SYMMETRIC:
+            {
+                const int32_t half_n_swa = (int32_t) n_swa / 2;
+                const int32_t pos_diff = p1 - p0;
+
+                // Mask if outside the symmetric window
+                if (pos_diff < -half_n_swa || pos_diff > half_n_swa) {
+                    return true;
+                }
+            } break;
+    }
+
+    return false;
+}

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

@@ -19,6 +19,7 @@ enum llama_swa_type {
     LLAMA_SWA_TYPE_NONE      = 0,
     LLAMA_SWA_TYPE_STANDARD  = 1,
     LLAMA_SWA_TYPE_CHUNKED   = 2,
+    LLAMA_SWA_TYPE_SYMMETRIC = 3,
 };
 
 struct llama_hparams_posnet {
@@ -41,6 +42,7 @@ struct llama_hparams {
     uint32_t n_embd;
     uint32_t n_embd_features = 0;
     uint32_t n_layer;
+     int32_t n_layer_kv_from_start = -1; // if non-negative, the first n_layer_kv_from_start layers have KV cache
     uint32_t n_rot;
     uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
     uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
@@ -69,10 +71,13 @@ struct llama_hparams {
     uint32_t n_lora_kv          = 0;
     uint32_t n_ff_exp           = 0;
     uint32_t n_ff_shexp         = 0;
+    uint32_t n_ff_chexp         = 0;
     uint32_t n_expert_shared    = 0;
     uint32_t n_norm_groups      = 0;
+    uint32_t n_group_experts    = 0;
 
-    float    expert_weights_scale = 0.0;
+    float    expert_group_scale   = 0.05f;
+    float    expert_weights_scale = 0.0f;
     bool     expert_weights_norm  = false;
     uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
     uint32_t moe_every_n_layers   = 0;
@@ -83,6 +88,7 @@ struct llama_hparams {
     float f_norm_group_eps;
 
     float f_attn_logit_softcapping   = 50.0f;
+    float f_router_logit_softcapping = 30.0f;
     float f_final_logit_softcapping  = 30.0f;
 
     // for RWKV
@@ -104,6 +110,11 @@ struct llama_hparams {
     uint32_t n_ctx_orig_yarn;
     float    rope_yarn_log_mul = 0.0f;
 
+    float    yarn_ext_factor  = -1.0f;
+    float    yarn_attn_factor =  1.0f;
+    float    yarn_beta_fast   = 32.0f;
+    float    yarn_beta_slow   =  1.0f;
+
     std::array<int, 4> rope_sections;
 
     // Sliding Window Attention (SWA)
@@ -136,10 +147,14 @@ struct llama_hparams {
     float f_embedding_scale = 0.0f;
     float f_attention_scale = 0.0f;
 
+    // grok-2
+    float    f_attn_out_scale = 0.0f;
+    uint32_t attn_temp_length = 0;
+
     bool causal_attn   = true;
     bool use_alibi     = false;
     bool attn_soft_cap = false;
-    bool use_kq_norm   = true;
+    bool use_kq_norm   = false;
 
     // for Classifiers
     uint32_t n_cls_out = 1;
@@ -159,6 +174,7 @@ struct llama_hparams {
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
+    uint32_t    dec_n_layer        = 0;
 
     enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
     enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
@@ -226,6 +242,16 @@ struct llama_hparams {
     bool n_bskcn(uint32_t n, uint32_t il) const;
 
     bool is_swa(uint32_t il) const;
+
+    bool has_kv(uint32_t il) const;
+
+    // number of layers for which has_kv() returns true
+    uint32_t n_layer_kv() const;
+
+    // note that this function uses different SWA parameters from those in the hparams
+    // TODO: think of a better place for this function
+    // TODO: pack the SWA params in a struct?
+    static bool is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1);
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");

llama/llama.cpp/src/llama-impl.h

@@ -59,3 +59,5 @@ std::string llama_format_tensor_shape(const std::vector<int64_t> & ne);
 std::string llama_format_tensor_shape(const struct ggml_tensor * t);
 
 std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);
+
+#define LLAMA_TENSOR_NAME_FATTN "__fattn__"

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

@@ -1,4 +1,4 @@
-#include "llama-kv-cache-unified-iswa.h"
+#include "llama-kv-cache-iswa.h"
 
 #include "llama-impl.h"
 #include "llama-batch.h"
@@ -8,10 +8,10 @@
 #include <cassert>
 
 //
-// llama_kv_cache_unified_iswa
+// llama_kv_cache_iswa
 //
 
-llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
+llama_kv_cache_iswa::llama_kv_cache_iswa(
         const llama_model & model,
                 ggml_type   type_k,
                 ggml_type   type_v,
@@ -22,9 +22,26 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
                  uint32_t   kv_size,
                  uint32_t   n_seq_max,
                  uint32_t   n_ubatch,
-                 uint32_t   n_pad) : hparams(model.hparams), unified(unified) {
-    llama_kv_cache_unified::layer_filter_cb filter_base = [&](int32_t il) { return !model.hparams.is_swa(il); };
-    llama_kv_cache_unified::layer_filter_cb filter_swa  = [&](int32_t il) { return  model.hparams.is_swa(il); };
+                 uint32_t   n_pad,
+    const layer_filter_cb & filter,
+    const  layer_reuse_cb & reuse) : hparams(model.hparams), unified(unified) {
+
+    // chain filters
+    const layer_filter_cb filter_base = [&](int32_t il) {
+        if (filter && !filter(il)) {
+            return false;
+        }
+
+        return !model.hparams.is_swa(il);
+    };
+
+    const layer_filter_cb filter_swa  = [&](int32_t il) {
+        if (filter && !filter(il)) {
+            return false;
+        }
+
+        return  model.hparams.is_swa(il);
+    };
 
     const uint32_t size_base = kv_size;
 
@@ -40,25 +57,25 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
 
     LLAMA_LOG_INFO("%s: creating non-SWA KV cache, size = %u cells\n", __func__, size_base);
 
-    kv_base = std::make_unique<llama_kv_cache_unified>(
-            model, std::move(filter_base), type_k, type_v,
+    kv_base = std::make_unique<llama_kv_cache>(
+            model, type_k, type_v,
             v_trans, offload, unified, size_base, n_seq_max, n_pad,
-            0, LLAMA_SWA_TYPE_NONE);
+            0, LLAMA_SWA_TYPE_NONE, filter_base, reuse);
 
     LLAMA_LOG_INFO("%s: creating     SWA KV cache, size = %u cells\n", __func__, size_swa);
 
-    kv_swa = std::make_unique<llama_kv_cache_unified>(
-            model, std::move(filter_swa), type_k, type_v,
+    kv_swa = std::make_unique<llama_kv_cache>(
+            model, type_k, type_v,
             v_trans, offload, unified, size_swa, n_seq_max, n_pad,
-            hparams.n_swa, hparams.swa_type);
+            hparams.n_swa, hparams.swa_type, filter_swa, reuse);
 }
 
-void llama_kv_cache_unified_iswa::clear(bool data) {
+void llama_kv_cache_iswa::clear(bool data) {
     kv_base->clear(data);
     kv_swa ->clear(data);
 }
 
-bool llama_kv_cache_unified_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+bool llama_kv_cache_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
     bool res = true;
 
     res = res & kv_base->seq_rm(seq_id, p0, p1);
@@ -67,36 +84,44 @@ bool llama_kv_cache_unified_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llam
     return res;
 }
 
-void llama_kv_cache_unified_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+void llama_kv_cache_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
     kv_base->seq_cp(seq_id_src, seq_id_dst, p0, p1);
     kv_swa ->seq_cp(seq_id_src, seq_id_dst, p0, p1);
 }
 
-void llama_kv_cache_unified_iswa::seq_keep(llama_seq_id seq_id) {
+void llama_kv_cache_iswa::seq_keep(llama_seq_id seq_id) {
     kv_base->seq_keep(seq_id);
     kv_swa ->seq_keep(seq_id);
 }
 
-void llama_kv_cache_unified_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+void llama_kv_cache_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
     kv_base->seq_add(seq_id, p0, p1, shift);
     kv_swa ->seq_add(seq_id, p0, p1, shift);
 }
 
-void llama_kv_cache_unified_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+void llama_kv_cache_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
     kv_base->seq_div(seq_id, p0, p1, d);
     kv_swa ->seq_div(seq_id, p0, p1, d);
 }
 
-llama_pos llama_kv_cache_unified_iswa::seq_pos_min(llama_seq_id seq_id) const {
+llama_pos llama_kv_cache_iswa::seq_pos_min(llama_seq_id seq_id) const {
     // the base cache is a superset of the SWA cache, so we can just check the SWA cache
     return kv_swa->seq_pos_min(seq_id);
 }
 
-llama_pos llama_kv_cache_unified_iswa::seq_pos_max(llama_seq_id seq_id) const {
+llama_pos llama_kv_cache_iswa::seq_pos_max(llama_seq_id seq_id) const {
     return kv_swa->seq_pos_max(seq_id);
 }
 
-llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache_iswa::memory_breakdown() const {
+    std::map<ggml_backend_buffer_type_t, size_t> mb = kv_base->memory_breakdown();
+    for (const auto & buft_size : kv_swa->memory_breakdown()) {
+        mb[buft_size.first] += buft_size.second;
+    }
+    return mb;
+}
+
+llama_memory_context_ptr llama_kv_cache_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
     GGML_UNUSED(embd_all);
 
     // first try simple split
@@ -136,7 +161,7 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
 
         assert(sinfos_base.size() == sinfos_swa.size());
 
-        return std::make_unique<llama_kv_cache_unified_iswa_context>(
+        return std::make_unique<llama_kv_cache_iswa_context>(
                 this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
     } while (false);
 
@@ -172,61 +197,67 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
 
         assert(sinfos_base.size() == sinfos_swa.size());
 
-        return std::make_unique<llama_kv_cache_unified_iswa_context>(
+        return std::make_unique<llama_kv_cache_iswa_context>(
                 this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
     } while (false);
 
     // TODO: if we fail again, we should attempt different splitting strategies
     //       but to do that properly, we first have to refactor the batches to be more flexible
 
-    return std::make_unique<llama_kv_cache_unified_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+    return std::make_unique<llama_kv_cache_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
-llama_memory_context_ptr llama_kv_cache_unified_iswa::init_full() {
-    return std::make_unique<llama_kv_cache_unified_iswa_context>(this);
+llama_memory_context_ptr llama_kv_cache_iswa::init_full() {
+    return std::make_unique<llama_kv_cache_iswa_context>(this);
 }
 
-llama_memory_context_ptr llama_kv_cache_unified_iswa::init_update(llama_context * lctx, bool optimize) {
-    return std::make_unique<llama_kv_cache_unified_iswa_context>(this, lctx, optimize);
+llama_memory_context_ptr llama_kv_cache_iswa::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_kv_cache_iswa_context>(this, lctx, optimize);
 }
 
-bool llama_kv_cache_unified_iswa::get_can_shift() const {
+bool llama_kv_cache_iswa::get_can_shift() const {
     return kv_base->get_size() == kv_swa->get_size();
 }
 
-void llama_kv_cache_unified_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
-    kv_base->state_write(io, seq_id);
-    kv_swa ->state_write(io, seq_id);
+void llama_kv_cache_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_SWA_ONLY) == 0) {
+        kv_base->state_write(io, seq_id, flags);
+    }
+
+    kv_swa->state_write(io, seq_id, flags);
 }
 
-void llama_kv_cache_unified_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
-    kv_base->state_read(io, seq_id);
-    kv_swa ->state_read(io, seq_id);
+void llama_kv_cache_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_SWA_ONLY) == 0) {
+        kv_base->state_read(io, seq_id, flags);
+    }
+
+    kv_swa->state_read(io, seq_id, flags);
 }
 
-llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_base() const {
+llama_kv_cache * llama_kv_cache_iswa::get_base() const {
     return kv_base.get();
 }
 
-llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_swa() const {
+llama_kv_cache * llama_kv_cache_iswa::get_swa() const {
     return kv_swa.get();
 }
 
 //
-// llama_kv_cache_unified_iswa_context
+// llama_kv_cache_iswa_context
 //
 
-llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(llama_memory_status status) : status(status) {}
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(llama_memory_status status) : status(status) {}
 
-llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
-        llama_kv_cache_unified_iswa * kv) :
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(
+        llama_kv_cache_iswa * kv) :
     ctx_base(kv->get_base()->init_full()),
     ctx_swa (kv->get_swa ()->init_full()),
     status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 
-llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
-        llama_kv_cache_unified_iswa * kv,
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(
+        llama_kv_cache_iswa * kv,
         llama_context * lctx,
         bool optimize) :
     ctx_base(kv->get_base()->init_update(lctx, optimize)),
@@ -234,21 +265,21 @@ llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
     status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 
-llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
-        llama_kv_cache_unified_iswa * kv,
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(
+        llama_kv_cache_iswa * kv,
         slot_info_vec_t sinfos_base,
         slot_info_vec_t sinfos_swa,
         std::vector<llama_ubatch> ubatches) :
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
-    ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(sinfos_base), this->ubatches)),
-    ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(sinfos_swa),  this->ubatches)),
+    ctx_base(new llama_kv_cache_context(kv->get_base(), std::move(sinfos_base), this->ubatches)),
+    ctx_swa (new llama_kv_cache_context(kv->get_swa (), std::move(sinfos_swa),  this->ubatches)),
     status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 
-llama_kv_cache_unified_iswa_context:: ~llama_kv_cache_unified_iswa_context() = default;
+llama_kv_cache_iswa_context:: ~llama_kv_cache_iswa_context() = default;
 
-bool llama_kv_cache_unified_iswa_context::next() {
+bool llama_kv_cache_iswa_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     ctx_base->next();
@@ -261,7 +292,7 @@ bool llama_kv_cache_unified_iswa_context::next() {
     return true;
 }
 
-bool llama_kv_cache_unified_iswa_context::apply() {
+bool llama_kv_cache_iswa_context::apply() {
     assert(!llama_memory_status_is_fail(status));
 
     bool res = true;
@@ -272,24 +303,24 @@ bool llama_kv_cache_unified_iswa_context::apply() {
     return res;
 }
 
-llama_memory_status llama_kv_cache_unified_iswa_context::get_status() const {
+llama_memory_status llama_kv_cache_iswa_context::get_status() const {
     return status;
 }
 
-const llama_ubatch & llama_kv_cache_unified_iswa_context::get_ubatch() const {
+const llama_ubatch & llama_kv_cache_iswa_context::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     return ubatches[i_next];
 }
 
-const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_base() const {
+const llama_kv_cache_context * llama_kv_cache_iswa_context::get_base() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return static_cast<const llama_kv_cache_unified_context *>(ctx_base.get());
+    return static_cast<const llama_kv_cache_context *>(ctx_base.get());
 }
 
-const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_swa()  const {
+const llama_kv_cache_context * llama_kv_cache_iswa_context::get_swa()  const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return static_cast<const llama_kv_cache_unified_context *>(ctx_swa.get());
+    return static_cast<const llama_kv_cache_context *>(ctx_swa.get());
 }

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

@@ -1,19 +1,19 @@
 #pragma once
 
-#include "llama-kv-cache-unified.h"
+#include "llama-kv-cache.h"
 
 #include <vector>
 
 //
-// llama_kv_cache_unified_iswa
+// llama_kv_cache_iswa
 //
 
-// utilizes two instances of llama_kv_cache_unified
+// utilizes two instances of llama_kv_cache
 //   the first instance is for the non-SWA layers of the model and the second instance is for the SWA layers
 
-class llama_kv_cache_unified_iswa : public llama_memory_i {
+class llama_kv_cache_iswa : public llama_memory_i {
 public:
-    llama_kv_cache_unified_iswa(
+    llama_kv_cache_iswa(
             const llama_model & model,
                     ggml_type   type_k,
                     ggml_type   type_v,
@@ -24,9 +24,11 @@ public:
                      uint32_t   kv_size,
                      uint32_t   n_seq_max,
                      uint32_t   n_ubatch,
-                     uint32_t   n_pad);
+                     uint32_t   n_pad,
+        const layer_filter_cb & filter,
+        const  layer_reuse_cb & reuse);
 
-    ~llama_kv_cache_unified_iswa() = default;
+    ~llama_kv_cache_iswa() = default;
 
     //
     // llama_memory_i
@@ -54,52 +56,54 @@ public:
     llama_pos seq_pos_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
+    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() 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;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     //
-    // llama_kv_cache_unified_iswa specific API
+    // llama_kv_cache_iswa specific API
     //
 
-    llama_kv_cache_unified * get_base() const;
-    llama_kv_cache_unified * get_swa () const;
+    llama_kv_cache * get_base() const;
+    llama_kv_cache * get_swa () const;
 
 private:
     const llama_hparams & hparams;
 
     const bool unified;
 
-    std::unique_ptr<llama_kv_cache_unified> kv_base;
-    std::unique_ptr<llama_kv_cache_unified> kv_swa;
+    std::unique_ptr<llama_kv_cache> kv_base;
+    std::unique_ptr<llama_kv_cache> kv_swa;
 };
 
-class llama_kv_cache_unified_iswa_context : public llama_memory_context_i {
+class llama_kv_cache_iswa_context : public llama_memory_context_i {
 public:
-    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+    using slot_info_vec_t = llama_kv_cache::slot_info_vec_t;
 
     // used for errors
-    llama_kv_cache_unified_iswa_context(llama_memory_status status);
+    llama_kv_cache_iswa_context(llama_memory_status status);
 
     // used to create a full-cache context
-    llama_kv_cache_unified_iswa_context(
-            llama_kv_cache_unified_iswa * kv);
+    llama_kv_cache_iswa_context(
+            llama_kv_cache_iswa * kv);
 
     // used to create an update context
-    llama_kv_cache_unified_iswa_context(
-            llama_kv_cache_unified_iswa * kv,
+    llama_kv_cache_iswa_context(
+            llama_kv_cache_iswa * kv,
             llama_context * lctx,
             bool optimize);
 
     // used to create a batch processing context from a batch
-    llama_kv_cache_unified_iswa_context(
-            llama_kv_cache_unified_iswa * kv,
+    llama_kv_cache_iswa_context(
+            llama_kv_cache_iswa * kv,
             slot_info_vec_t sinfos_base,
             slot_info_vec_t sinfos_swa,
             std::vector<llama_ubatch> ubatches);
 
-    virtual ~llama_kv_cache_unified_iswa_context();
+    virtual ~llama_kv_cache_iswa_context();
 
     //
     // llama_memory_context_i
@@ -112,14 +116,14 @@ public:
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_kv_cache_unified_iswa_context specific API
+    // llama_kv_cache_iswa_context specific API
     //
 
-    const llama_kv_cache_unified_context * get_base() const;
-    const llama_kv_cache_unified_context * get_swa()  const;
+    const llama_kv_cache_context * get_base() const;
+    const llama_kv_cache_context * get_swa()  const;
 
 private:
-    //llama_kv_cache_unified_iswa * kv;
+    //llama_kv_cache_iswa * kv;
 
     // the index of the next ubatch to process
     size_t i_next = 0;

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

@@ -14,27 +14,13 @@ struct llama_model;
 struct llama_context;
 
 //
-// llama_kv_cache_unified
+// llama_kv_cache
 //
 
-class llama_kv_cache_unified : public llama_memory_i {
+class llama_kv_cache : public llama_memory_i {
 public:
     static uint32_t get_padding(const llama_cparams & cparams);
 
-    // this callback is used to filter out layers that should not be included in the cache
-    using layer_filter_cb = std::function<bool(int32_t il)>;
-
-    struct defrag_info {
-        bool empty() const {
-            return ids.empty();
-        }
-
-        // contains information about which cell moves where:
-        //  - cell i moves to ids[i]
-        //  - if ids[i] == i || ids[i] == ids.size(), then cell i is not moved
-        std::vector<uint32_t> ids;
-    };
-
     struct stream_copy_info {
         bool empty() const {
             assert(ssrc.size() == sdst.size());
@@ -52,8 +38,8 @@ public:
         using idx_vec_t = std::vector<uint32_t>;
 
         // number of streams: ns = s1 - s0 + 1
-        llama_seq_id s0;
-        llama_seq_id s1;
+        uint32_t s0;
+        uint32_t s1;
 
         std::vector<llama_seq_id> strm; // [ns]
         std::vector<idx_vec_t>    idxs; // [ns]
@@ -92,9 +78,8 @@ public:
 
     using slot_info_vec_t = std::vector<slot_info>;
 
-    llama_kv_cache_unified(
+    llama_kv_cache(
             const llama_model & model,
-              layer_filter_cb && filter,
                     ggml_type   type_k,
                     ggml_type   type_v,
                          bool   v_trans,
@@ -104,9 +89,11 @@ public:
                      uint32_t   n_seq_max,
                      uint32_t   n_pad,
                      uint32_t   n_swa,
-               llama_swa_type    swa_type);
+               llama_swa_type   swa_type,
+        const layer_filter_cb & filter,
+        const  layer_reuse_cb & reuse);
 
-    ~llama_kv_cache_unified() = default;
+    ~llama_kv_cache() = default;
 
     //
     // llama_memory_i
@@ -134,13 +121,15 @@ public:
     llama_pos seq_pos_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
+    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() 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;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     //
-    // llama_kv_cache_unified specific API
+    // llama_kv_cache specific API
     //
 
     uint32_t get_size()     const;
@@ -152,10 +141,7 @@ public:
     // graph_build API
     //
 
-    uint32_t get_n_kv() const;
-
-    // TODO: temporary
-    bool get_supports_set_rows() const;
+    uint32_t get_n_kv(const slot_info & sinfo) const;
 
     // get views of the current state of the cache
     ggml_tensor * get_k(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
@@ -173,7 +159,7 @@ public:
     // return empty vector on failure
     slot_info_vec_t prepare(const std::vector<llama_ubatch> & ubatches);
 
-    bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo, const stream_copy_info & sc_info);
+    bool update(llama_context * lctx, bool do_shift, const stream_copy_info & sc_info);
 
     // find a slot of kv cells that can hold the ubatch
     // if cont == true, then the slot must be continuous
@@ -228,10 +214,7 @@ private:
     // env: LLAMA_KV_CACHE_DEBUG
     int debug = 0;
 
-    // env: LLAMA_SET_ROWS (temporary)
-    // ref: https://github.com/ggml-org/llama.cpp/pull/14285
-    bool supports_set_rows = true;
-
+    // this is the SWA type of the cache - not to be confused with the model SWA type
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
     std::vector<ggml_context_ptr>        ctxs;
@@ -241,7 +224,7 @@ private:
     // note: this is not part of the KV state and it's only used to speed-up the find_slot() method
     std::vector<uint32_t> v_heads;
 
-    std::vector<llama_kv_cells_unified> v_cells;
+    std::vector<llama_kv_cells> v_cells;
 
     // maps from a sequence id to a stream id
     std::vector<uint32_t> seq_to_stream;
@@ -254,9 +237,6 @@ private:
     // model layer id -> KV cache layer id
     std::unordered_map<int32_t, int32_t> map_layer_ids;
 
-    // return non-empty vector if cells have been moved
-    defrag_info defrag_prepare(int32_t n_max_nodes) const;
-
     size_t total_size() const;
 
     size_t size_k_bytes() const;
@@ -277,11 +257,6 @@ private:
                llm_graph_result * res,
                   llama_context * lctx) const;
 
-    ggml_cgraph * build_graph_defrag(
-               llm_graph_result * res,
-                  llama_context * lctx,
-              const defrag_info & dinfo) const;
-
     struct cell_ranges_t {
         uint32_t strm;
 
@@ -295,35 +270,33 @@ private:
     bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count);
 };
 
-class llama_kv_cache_unified_context : public llama_memory_context_i {
+class llama_kv_cache_context : public llama_memory_context_i {
 public:
     // some shorthands
-    using slot_info_vec_t  = llama_kv_cache_unified::slot_info_vec_t;
-    using defrag_info      = llama_kv_cache_unified::defrag_info;
-    using stream_copy_info = llama_kv_cache_unified::stream_copy_info;
+    using slot_info_vec_t  = llama_kv_cache::slot_info_vec_t;
+    using stream_copy_info = llama_kv_cache::stream_copy_info;
 
     // used for errors
-    llama_kv_cache_unified_context(llama_memory_status status);
+    llama_kv_cache_context(llama_memory_status status);
 
     // used to create a full-cache context
-    llama_kv_cache_unified_context(
-            llama_kv_cache_unified * kv);
+    llama_kv_cache_context(
+            llama_kv_cache * kv);
 
     // used to create an update context
-    llama_kv_cache_unified_context(
-            llama_kv_cache_unified * kv,
+    llama_kv_cache_context(
+            llama_kv_cache * kv,
             llama_context * lctx,
             bool do_shift,
-            defrag_info dinfo,
             stream_copy_info sc_info);
 
     // used to create a batch procesing context from a batch
-    llama_kv_cache_unified_context(
-            llama_kv_cache_unified * kv,
+    llama_kv_cache_context(
+            llama_kv_cache * kv,
             slot_info_vec_t sinfos,
             std::vector<llama_ubatch> ubatches);
 
-    virtual ~llama_kv_cache_unified_context();
+    virtual ~llama_kv_cache_context();
 
     //
     // llama_memory_context_i
@@ -336,22 +309,27 @@ public:
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_kv_cache_unified_context specific API
+    // llama_kv_cache_context specific API
     //
 
     uint32_t get_n_kv() const;
 
-    // TODO: temporary
-    bool get_supports_set_rows() const;
-
     // get views of the current state of the cache
     ggml_tensor * get_k(ggml_context * ctx, int32_t il) const;
     ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
+    // note: the heads in k_cur and v_cur should be layed out contiguously in memory
+    //   - k_cur  [n_embd_head_k, n_head_k, n_tokens]
+    //   - k_idxs [n_tokens]
+    //   - v_cur  [n_embd_head_v, n_head_v, n_tokens]
+    //   - v_idxs [n_tokens] or [n_tokens*n_embd_v_gqa] depending if V cache is transposed
     ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const;
     ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const;
 
+    // create destination indices for each head of the current batch for where it would be written in the KV cache
+    // the indices address the global KV cache (not per stream) - this is not relevant for the user of this API, but
+    //   helps understand the implementation logic of cpy_k and cpy_v
     ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
     ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
 
@@ -365,7 +343,7 @@ public:
 private:
     llama_memory_status status;
 
-    llama_kv_cache_unified * kv;
+    llama_kv_cache * kv;
     llama_context * lctx;
 
     //
@@ -374,8 +352,6 @@ private:
 
     bool do_shift = false;
 
-    defrag_info dinfo;
-
     stream_copy_info sc_info;
 
     //

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

@@ -11,7 +11,7 @@
 
 // meta information about KV cells that can be part of multiple sequences at the same time
 // TODO: add unit tests
-class llama_kv_cells_unified {
+class llama_kv_cells {
 public:
     void reset() {
         for (uint32_t i = 0; i < pos.size(); ++i) {
@@ -77,30 +77,30 @@ public:
     }
 
     // move cell isrc to idst (used during defrag)
-    void mv(uint32_t isrc, uint32_t idst) {
-        assert(isrc < pos.size());
-        assert(idst < pos.size());
+    //void mv(uint32_t isrc, uint32_t idst) {
+    //    assert(isrc < pos.size());
+    //    assert(idst < pos.size());
 
-        assert(pos[idst] == -1);
-        assert(pos[isrc] != -1);
+    //    assert(pos[idst] == -1);
+    //    assert(pos[isrc] != -1);
 
-        pos  [idst] = pos  [isrc];
-        shift[idst] = shift[isrc];
-        seq  [idst] = seq  [isrc];
+    //    pos  [idst] = pos  [isrc];
+    //    shift[idst] = shift[isrc];
+    //    seq  [idst] = seq  [isrc];
 
-        pos  [isrc] = -1;
-        shift[isrc] =  0;
-        seq  [isrc].reset();
+    //    pos  [isrc] = -1;
+    //    shift[isrc] =  0;
+    //    seq  [isrc].reset();
 
-        used.erase (isrc);
-        used.insert(idst);
-    }
+    //    used.erase (isrc);
+    //    used.insert(idst);
+    //}
 
     // copy the state of cells [i, i + n) (used for save/restore the state of the cells)
-    llama_kv_cells_unified cp(uint32_t i, uint32_t n) const {
+    llama_kv_cells cp(uint32_t i, uint32_t n) const {
         assert(i + n <= pos.size());
 
-        llama_kv_cells_unified res;
+        llama_kv_cells res;
 
         res.resize(n);
 
@@ -117,8 +117,8 @@ public:
     }
 
     // copy the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
-    llama_kv_cells_unified cp(const std::vector<uint32_t> & idxs) const {
-        llama_kv_cells_unified res;
+    llama_kv_cells cp(const std::vector<uint32_t> & idxs) const {
+        llama_kv_cells res;
 
         res.resize(idxs.size());
 
@@ -135,7 +135,7 @@ public:
     }
 
     // set the state of cells [i, i + other.pos.size()) (used for save/restore the state of the cells)
-    void set(uint32_t i, const llama_kv_cells_unified & other) {
+    void set(uint32_t i, const llama_kv_cells & other) {
         assert(i + other.pos.size() <= pos.size());
 
         for (uint32_t j = 0; j < other.pos.size(); ++j) {
@@ -165,7 +165,7 @@ public:
     }
 
     // set the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
-    void set(const std::vector<uint32_t> & idxs, const llama_kv_cells_unified & other) {
+    void set(const std::vector<uint32_t> & idxs, const llama_kv_cells & other) {
         assert(idxs.size() == other.pos.size());
 
         for (uint32_t j = 0; j < other.pos.size(); ++j) {

llama/llama.cpp/src/llama-memory-hybrid.cpp

@@ -27,14 +27,11 @@ llama_memory_hybrid::llama_memory_hybrid(
                      bool   offload,
                      bool   unified,
                             /* layer filters */
-      layer_filter_cb && filter_attn,
-      layer_filter_cb && filter_recr) :
+    const layer_filter_cb & filter_attn,
+    const layer_filter_cb & filter_recr) :
     hparams(model.hparams),
-    mem_attn(new llama_kv_cache_unified(
+    mem_attn(new llama_kv_cache(
         model,
-        filter_attn == nullptr ?
-            [&](int32_t il) { return !hparams.is_recurrent(il); }
-            : filter_attn,
         type_k,
         type_v,
         v_trans,
@@ -44,18 +41,22 @@ llama_memory_hybrid::llama_memory_hybrid(
         n_seq_max,
         n_pad,
         n_swa,
-        swa_type
+        swa_type,
+        filter_attn == nullptr ?
+            [&](int32_t il) { return !hparams.is_recurrent(il); }
+            : filter_attn,
+        nullptr
     )),
     mem_recr(new llama_memory_recurrent(
         model,
-        filter_recr == nullptr ?
-            [&](int32_t il) { return hparams.is_recurrent(il); }
-            : filter_recr,
         type_r,
         type_s,
         offload,
         rs_size,
-        n_seq_max
+        n_seq_max,
+        filter_recr == nullptr ?
+            [&](int32_t il) { return hparams.is_recurrent(il); }
+            : filter_recr
     )) {}
 
 llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
@@ -165,17 +166,29 @@ llama_pos llama_memory_hybrid::seq_pos_max(llama_seq_id seq_id) const {
     return std::min(mem_attn->seq_pos_max(seq_id), mem_recr->seq_pos_max(seq_id));
 }
 
-void llama_memory_hybrid::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+std::map<ggml_backend_buffer_type_t, size_t> llama_memory_hybrid::memory_breakdown() const {
+    std::map<ggml_backend_buffer_type_t, size_t> mb = mem_attn->memory_breakdown();
+    for (const auto & buft_size : mem_recr->memory_breakdown()) {
+        mb[buft_size.first] += buft_size.second;
+    }
+    return mb;
+}
+
+void llama_memory_hybrid::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    GGML_UNUSED(flags);
+
     mem_attn->state_write(io, seq_id);
     mem_recr->state_write(io, seq_id);
 }
 
-void llama_memory_hybrid::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_memory_hybrid::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    GGML_UNUSED(flags);
+
     mem_attn->state_read(io, seq_id);
     mem_recr->state_read(io, seq_id);
 }
 
-llama_kv_cache_unified * llama_memory_hybrid::get_mem_attn() const {
+llama_kv_cache * llama_memory_hybrid::get_mem_attn() const {
     return mem_attn.get();
 }
 
@@ -206,7 +219,7 @@ llama_memory_hybrid_context::llama_memory_hybrid_context(
         std::vector<llama_ubatch>   ubatches) :
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
-    ctx_attn(new llama_kv_cache_unified_context(mem->get_mem_attn(), std::move(sinfos_attn), this->ubatches)),
+    ctx_attn(new llama_kv_cache_context(mem->get_mem_attn(), std::move(sinfos_attn), this->ubatches)),
     ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(),                        this->ubatches)),
     status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
 }
@@ -244,8 +257,8 @@ const llama_ubatch & llama_memory_hybrid_context::get_ubatch() const {
     return ubatches[i_next];
 }
 
-const llama_kv_cache_unified_context * llama_memory_hybrid_context::get_attn() const {
-    return static_cast<const llama_kv_cache_unified_context *>(ctx_attn.get());
+const llama_kv_cache_context * llama_memory_hybrid_context::get_attn() const {
+    return static_cast<const llama_kv_cache_context *>(ctx_attn.get());
 }
 
 const llama_memory_recurrent_context * llama_memory_hybrid_context::get_recr() const {

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

@@ -2,7 +2,7 @@
 
 #include "llama-batch.h"
 #include "llama-graph.h"
-#include "llama-kv-cache-unified.h"
+#include "llama-kv-cache.h"
 #include "llama-memory.h"
 #include "llama-memory-recurrent.h"
 
@@ -13,15 +13,11 @@
 // llama_memory_hybrid
 //
 
-// utilizes instances of llama_memory_recurrent and llama_kv_cache_unified to
+// utilizes instances of llama_memory_recurrent and llama_kv_cache to
 //   support models where each layer may be either attention-based or recurrent
 
 class llama_memory_hybrid : public llama_memory_i {
 public:
-
-    // this callback is used to filter out layers that should not be included in the cache
-    using layer_filter_cb = std::function<bool(int32_t il)>;
-
     llama_memory_hybrid(
         const llama_model & model,
                             /* attn */
@@ -41,8 +37,8 @@ public:
                      bool   offload,
                      bool   unified,
                             /* layer filters */
-          layer_filter_cb && filter_attn = nullptr,
-          layer_filter_cb && filter_recr = nullptr);
+    const layer_filter_cb & filter_attn = nullptr,
+    const layer_filter_cb & filter_recr = nullptr);
 
     ~llama_memory_hybrid() = default;
 
@@ -72,28 +68,30 @@ public:
     llama_pos seq_pos_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
+    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() 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;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0)       override;
 
     //
     // llama_memory_hybrid specific API
     //
 
-    llama_kv_cache_unified * get_mem_attn() const;
+    llama_kv_cache * get_mem_attn() const;
     llama_memory_recurrent * get_mem_recr() const;
 
 private:
     const llama_hparams & hparams;
 
-    const std::unique_ptr<llama_kv_cache_unified> mem_attn;
+    const std::unique_ptr<llama_kv_cache> mem_attn;
     const std::unique_ptr<llama_memory_recurrent> mem_recr;
 };
 
 class llama_memory_hybrid_context : public llama_memory_context_i {
 public:
-    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+    using slot_info_vec_t = llama_kv_cache::slot_info_vec_t;
 
     // init failure
     explicit llama_memory_hybrid_context(llama_memory_status status);
@@ -125,7 +123,7 @@ public:
     // llama_memory_hybrid_context
     //
 
-    const llama_kv_cache_unified_context * get_attn() const;
+    const llama_kv_cache_context * get_attn() const;
     const llama_memory_recurrent_context * get_recr() const;
 
 private:

llama/llama.cpp/src/llama-memory-recurrent.cpp

@@ -17,12 +17,12 @@
 
 llama_memory_recurrent::llama_memory_recurrent(
         const llama_model & model,
-          layer_filter_cb && filter,
                 ggml_type   type_r,
                 ggml_type   type_s,
                      bool   offload,
                  uint32_t   mem_size,
-                 uint32_t    n_seq_max) : hparams(model.hparams), n_seq_max(n_seq_max) {
+                 uint32_t   n_seq_max,
+    const layer_filter_cb & filter) : hparams(model.hparams), n_seq_max(n_seq_max) {
     const int32_t n_layer = hparams.n_layer;
 
     head = 0;
@@ -359,6 +359,14 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const {
     return result;
 }
 
+std::map<ggml_backend_buffer_type_t, size_t> llama_memory_recurrent::memory_breakdown() const {
+    std::map<ggml_backend_buffer_type_t, size_t> ret;
+    for (const ggml_backend_buffer_ptr & buf_ptr : bufs) {
+        ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get());
+    }
+    return ret;
+}
+
 llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
     do {
         balloc.split_reset();
@@ -680,7 +688,9 @@ size_t llama_memory_recurrent::size_s_bytes() const {
     return size_s_bytes;
 }
 
-void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    GGML_UNUSED(flags);
+
     std::vector<std::pair<uint32_t, uint32_t>> cell_ranges; // ranges, from inclusive, to exclusive
     uint32_t cell_count = 0;
 
@@ -718,7 +728,9 @@ void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq
     state_write_data(io, cell_ranges);
 }
 
-void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    GGML_UNUSED(flags);
+
     uint32_t cell_count;
     io.read_to(&cell_count, sizeof(cell_count));
 

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

@@ -4,6 +4,7 @@
 #include "llama-graph.h"
 #include "llama-memory.h"
 
+#include <map>
 #include <set>
 #include <vector>
 
@@ -12,21 +13,17 @@
 //
 
 // TODO: extract the cache state used for graph computation into llama_memory_recurrent_context_i
-//       see the implementation of llama_kv_cache_unified_context_i for an example how to do it
+//       see the implementation of llama_kv_cache_context_i for an example how to do it
 class llama_memory_recurrent : public llama_memory_i {
 public:
-
-    // this callback is used to filter out layers that should not be included in the cache
-    using layer_filter_cb = std::function<bool(int32_t il)>;
-
     llama_memory_recurrent(
             const llama_model & model,
-              layer_filter_cb && filter,
                     ggml_type   type_r,
                     ggml_type   type_s,
                          bool   offload,
                      uint32_t   mem_size,
-                     uint32_t    n_seq_max);
+                     uint32_t   n_seq_max,
+        const layer_filter_cb & filter);
 
     ~llama_memory_recurrent() = default;
 
@@ -54,6 +51,8 @@ public:
     llama_pos seq_pos_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
+    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
+
     bool prepare(const std::vector<llama_ubatch> & ubatches);
 
     // find a contiguous slot of memory cells and emplace the ubatch there
@@ -63,8 +62,8 @@ public:
 
     // 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, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
     uint32_t size = 0; // total number of cells, shared across all sequences

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

@@ -2,7 +2,9 @@
 
 #include "llama.h"
 
+#include <map>
 #include <memory>
+#include <functional>
 
 struct llama_ubatch;
 
@@ -36,8 +38,8 @@ bool llama_memory_status_is_fail(llama_memory_status status);
 
 // the interface for managing the memory context during batch processing
 // this interface is implemented per memory type. see:
-//   - llama_kv_cache_unified_context
-//   - llama_kv_cache_unified_iswa_context
+//   - llama_kv_cache_context
+//   - llama_kv_cache_iswa_context
 //   ...
 //
 // the only method that should mutate the memory and the memory context is llama_memory_i::apply()
@@ -64,6 +66,13 @@ using llama_memory_context_ptr = std::unique_ptr<llama_memory_context_i>;
 // general concept of LLM memory
 // the KV cache is a type of LLM memory, but there can be other types
 struct llama_memory_i {
+    // this callback is used to filter out layers that should not be included in the cache
+    using layer_filter_cb = std::function<bool(int32_t il)>;
+
+    // this callback is used to specify which layers should reuse memory from other layers
+    // return negative value to indicate that the layer il should not reuse memory
+    using layer_reuse_cb = std::function<int32_t(int32_t il)>;
+
     virtual ~llama_memory_i() = default;
 
     // split the input batch into a set of ubatches and verify that they can fit into the cache
@@ -77,7 +86,7 @@ struct llama_memory_i {
     // simulate full cache, used for allocating worst-case compute buffers
     virtual llama_memory_context_ptr init_full() = 0;
 
-    // prepare for any pending memory updates, such as shifts, defrags, etc.
+    // prepare for any pending memory updates, such as shifts, copies, etc.
     // status == LLAMA_MEMORY_STATUS_NO_UPDATE if there is nothing to update
     virtual llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) = 0;
 
@@ -100,17 +109,14 @@ struct llama_memory_i {
     virtual llama_pos seq_pos_min(llama_seq_id seq_id) const = 0;
     virtual llama_pos seq_pos_max(llama_seq_id seq_id) const = 0;
 
+    virtual std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const = 0;
+
     //
     // 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;
+    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const = 0;
+    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) = 0;
 };
 
 using llama_memory_ptr = std::unique_ptr<llama_memory_i>;
-
-// TODO: temporary until the llama_kv_cache is removed from the public API
-struct llama_kv_cache : public llama_memory_i {
-    virtual ~llama_kv_cache() = default;
-};

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

@@ -789,6 +789,7 @@ const struct ggml_tensor * llama_model_loader::check_tensor_dims(const std::stri
 }
 
 struct ggml_tensor * llama_model_loader::create_tensor(struct ggml_context * ctx, const std::string & name, const std::initializer_list<int64_t> & ne, int flags) {
+    LLAMA_LOG_DEBUG("%s: loading tensor %s\n", __func__, name.c_str());
     const struct ggml_tensor * cur = check_tensor_dims(name, ne, !(flags & TENSOR_NOT_REQUIRED));
 
     if (cur == NULL) {

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

@@ -7,6 +7,7 @@
 #include "llama-memory.h"
 #include "llama-vocab.h"
 
+#include <map>
 #include <memory>
 #include <string>
 #include <unordered_map>
@@ -28,6 +29,7 @@ enum llm_type {
     LLM_TYPE_80M,
     LLM_TYPE_109M,
     LLM_TYPE_137M,
+    LLM_TYPE_140M,
     LLM_TYPE_160M,
     LLM_TYPE_190M,
     LLM_TYPE_220M,
@@ -36,12 +38,15 @@ enum llm_type {
     LLM_TYPE_270M,
     LLM_TYPE_335M,
     LLM_TYPE_350M,
+    LLM_TYPE_360M,
     LLM_TYPE_410M,
     LLM_TYPE_450M,
     LLM_TYPE_475M,
+    LLM_TYPE_558M,
     LLM_TYPE_700M,
     LLM_TYPE_770M,
     LLM_TYPE_780M,
+    LLM_TYPE_950M,
     LLM_TYPE_0_3B,
     LLM_TYPE_0_5B,
     LLM_TYPE_0_6B,
@@ -54,6 +59,7 @@ enum llm_type {
     LLM_TYPE_1_7B,
     LLM_TYPE_1_8B,
     LLM_TYPE_2B,
+    LLM_TYPE_2_6B,
     LLM_TYPE_2_8B,
     LLM_TYPE_2_9B,
     LLM_TYPE_3B,
@@ -76,9 +82,11 @@ enum llm_type {
     LLM_TYPE_32B,
     LLM_TYPE_34B,
     LLM_TYPE_35B,
+    LLM_TYPE_36B,
     LLM_TYPE_40B,
     LLM_TYPE_65B,
     LLM_TYPE_70B,
+    LLM_TYPE_120B,
     LLM_TYPE_142B,
     LLM_TYPE_236B,
     LLM_TYPE_290B,
@@ -268,6 +276,11 @@ struct llama_layer {
     struct ggml_tensor * ffn_down_shexp     = nullptr;
     struct ggml_tensor * ffn_up_shexp       = nullptr;
 
+    // ff adjugate experts (chexps)
+    struct ggml_tensor * ffn_gate_chexps     = nullptr;
+    struct ggml_tensor * ffn_down_chexps     = nullptr;
+    struct ggml_tensor * ffn_up_chexps       = nullptr;
+
     // ff bias
     struct ggml_tensor * ffn_gate_b = nullptr;
     struct ggml_tensor * ffn_down_b = nullptr; // b2
@@ -449,10 +462,12 @@ struct llama_model {
 
     std::string desc() const;
 
-    size_t size() const;
+    size_t size() const; // file size
     size_t n_tensors() const;
     size_t n_devices() const;
 
+    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const;
+
     // total number of parameters in the model
     uint64_t n_elements() const;
 

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

@@ -725,7 +725,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         // attention layers have a non-zero number of kv heads
         int32_t n_attn_layer = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
         if (llama_model_has_encoder(&model)) {
-            n_attn_layer *= 3;
+            // now n_attn_layer is the number of attention layers in the encoder
+            // for each decoder block, there are 2 attention layers
+            n_attn_layer += 2 * model.hparams.dec_n_layer;
         }
         GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
     }
@@ -920,7 +922,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             new_type = tensor->type;
             new_data = tensor->data;
             new_size = ggml_nbytes(tensor);
-            LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0);
+            LLAMA_LOG_INFO("size = %8.3f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0);
         } else {
             const int64_t nelements = ggml_nelements(tensor);
 
@@ -1037,8 +1039,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     }
     close_ofstream();
 
-    LLAMA_LOG_INFO("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
-    LLAMA_LOG_INFO("%s: quant size  = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
+    LLAMA_LOG_INFO("%s: model size  = %8.2f MiB\n", __func__, total_size_org/1024.0/1024.0);
+    LLAMA_LOG_INFO("%s: quant size  = %8.2f MiB\n", __func__, total_size_new/1024.0/1024.0);
 
     if (qs.n_fallback > 0) {
         LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) required fallback quantization\n",

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

@@ -128,6 +128,89 @@ struct ring_buffer {
     std::vector<T> data;
 };
 
+// writes result in res, does not mutate cur
+static void llama_token_data_array_partial_sort(const llama_token_data_array & cur, int npartial, std::vector<llama_token_data> & res) {
+    static const auto comp = [](const llama_token_data & a, const llama_token_data & b) {
+        return a.logit > b.logit;
+    };
+
+    constexpr int   nbuckets     = 128;
+    constexpr float bucket_low   = -10.0f;
+    constexpr float bucket_high  =  10.0f;
+    constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
+    constexpr float bucket_inter = -bucket_low * bucket_scale;
+
+    std::vector<int> bucket_idx;
+    std::vector<int> histo(nbuckets, 0);
+
+    std::vector<llama_token_data*> bucket_ptrs;
+
+    bucket_idx.reserve(cur.size);
+
+    for (int i = 0; i < (int)cur.size; ++i) {
+        const float val = cur.data[i].logit;
+        int ib = int(bucket_scale * val + bucket_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
+        ib = std::max(0, std::min(nbuckets - 1, ib));
+        bucket_idx.push_back(ib);
+        ++histo[ib];
+    }
+    int nhave = 0;
+    int ib = nbuckets - 1;
+    for ( ; ib >= 0; --ib) {
+        nhave += histo[ib];
+        if (nhave >= npartial) {
+            break;
+        }
+    }
+    res.resize(nhave);
+    auto * ptr = res.data();
+    bucket_ptrs.reserve(nbuckets - ib);
+    for (int j = nbuckets - 1; j >= ib; --j) {
+        bucket_ptrs.push_back(ptr);
+        ptr += histo[j];
+    }
+    for (int i = 0; i < (int)cur.size; ++i) {
+        int j = bucket_idx[i];
+        if (j >= ib) {
+            *bucket_ptrs[nbuckets - 1 - j]++ = cur.data[i];
+        }
+    }
+
+    ptr = res.data();
+    int ndone = 0;
+    for (int j = nbuckets - 1; j > ib; --j) {
+        std::sort(ptr, ptr + histo[j], comp);
+        ptr += histo[j];
+        ndone += histo[j];
+    }
+    std::partial_sort(ptr, ptr + npartial - ndone, ptr + histo[ib], comp);
+}
+
+// reduces the size of cur_p to npartial, keeping only the top npartial elements
+static void llama_token_data_array_partial_sort_inplace(llama_token_data_array * cur_p, int npartial) {
+    static const auto comp = [](const llama_token_data & a, const llama_token_data & b) {
+        return a.logit > b.logit;
+    };
+
+    if (npartial <= 128) {
+        std::partial_sort(cur_p->data, cur_p->data + npartial, cur_p->data + cur_p->size, comp);
+
+        cur_p->size = npartial;
+        cur_p->sorted = true;
+
+        return;
+    }
+
+    std::vector<llama_token_data> tmp;
+
+    llama_token_data_array_partial_sort(*cur_p, npartial, tmp);
+
+    std::copy(tmp.data(), tmp.data() + npartial, cur_p->data);
+
+    cur_p->size = npartial;
+    cur_p->sorted = true;
+}
+
 static int llama_sample_dist(llama_token_data_array * cur_p, std::mt19937 & rng) {
     // iterator for the probabilities
 #ifdef __GNUC__
@@ -200,18 +283,21 @@ static void llama_sampler_temp_impl(llama_token_data_array * cur_p, float temp)
     }
 }
 
-static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
+static void llama_sampler_softmax_impl(llama_token_data_array * cur_p, bool do_sort) {
     GGML_ASSERT(cur_p->size > 0);
 
-    // Sort the logits in descending order
-    if (!cur_p->sorted) {
-        std::sort(cur_p->data, cur_p->data + cur_p->size, [](const llama_token_data & a, const llama_token_data & b) {
-            return a.logit > b.logit;
-        });
-        cur_p->sorted = true;
+    // Sort the logits in descending order if requested
+    if (do_sort && !cur_p->sorted) {
+        llama_token_data_array_partial_sort_inplace(cur_p, cur_p->size);
     }
 
     float max_l = cur_p->data[0].logit;
+    if (!cur_p->sorted) {
+        for (size_t i = 1; i < cur_p->size; ++i) {
+            max_l = std::max(max_l, cur_p->data[i].logit);
+        }
+    }
+
     float cum_sum = 0.0f;
 
     for (size_t i = 0; i < cur_p->size; ++i) {
@@ -226,7 +312,6 @@ static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
 }
 
 static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k) {
-    // TODO: move bucket sort to separate function so that top_p/typical/softmax first is equally fast
     // if (k >= (int32_t)cur_p->size) {
     //     return;
     // }
@@ -239,64 +324,7 @@ static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k)
 
     // Sort scores in descending order
     if (!cur_p->sorted) {
-        auto comp = [](const llama_token_data & a, const llama_token_data & b) {
-            return a.logit > b.logit;
-        };
-        if (k <= 128) {
-            std::partial_sort(cur_p->data, cur_p->data + k, cur_p->data + cur_p->size, comp);
-        } else {
-            constexpr int   nbuckets     = 128;
-            constexpr float bucket_low   = -10.0f;
-            constexpr float bucket_high  =  10.0f;
-            constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
-            constexpr float bucket_inter = -bucket_low * bucket_scale;
-
-            std::vector<int> bucket_idx(cur_p->size);
-            std::vector<int> histo(nbuckets, 0);
-
-            for (int i = 0; i < (int)cur_p->size; ++i) {
-                const float val = cur_p->data[i].logit;
-                int ib = int(bucket_scale * val + bucket_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
-                ib = std::max(0, std::min(nbuckets - 1, ib));
-                bucket_idx[i] = ib;
-                ++histo[ib];
-            }
-            int nhave = 0;
-            int ib = nbuckets - 1;
-            for ( ; ib >= 0; --ib) {
-                nhave += histo[ib];
-                if (nhave >= k) {
-                    break;
-                }
-            }
-            std::vector<llama_token_data> tmp_tokens(nhave);
-            auto * ptr = tmp_tokens.data();
-            std::vector<llama_token_data*> bucket_ptrs;
-            bucket_ptrs.reserve(nbuckets - ib);
-            for (int j = nbuckets - 1; j >= ib; --j) {
-                bucket_ptrs.push_back(ptr);
-                ptr += histo[j];
-            }
-            for (int i = 0; i < (int)cur_p->size; ++i) {
-                int j = bucket_idx[i];
-                if (j >= ib) {
-                    *bucket_ptrs[nbuckets - 1 - j]++ = cur_p->data[i];
-                }
-            }
-
-            ptr = tmp_tokens.data();
-            int ndone = 0;
-            for (int j = nbuckets - 1; j > ib; --j) {
-                std::sort(ptr, ptr + histo[j], comp);
-                ptr += histo[j];
-                ndone += histo[j];
-            }
-            std::partial_sort(ptr, ptr + k - ndone, ptr + histo[ib], comp);
-
-            std::memcpy(cur_p->data, tmp_tokens.data(), k*sizeof(llama_token_data));
-
-        }
-        cur_p->sorted = true;
+        llama_token_data_array_partial_sort_inplace(cur_p, k);
     }
 
     cur_p->size = k;
@@ -576,9 +604,73 @@ static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl*
 static void llama_sampler_dist_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_dist *) smpl->ctx;
 
-    llama_sampler_softmax_impl(cur_p);
+    // edge cases
+    if (cur_p->size == 0) {
+        cur_p->selected = -1;
+        return;
+    }
+
+    cur_p->selected = 0;
+
+    if (cur_p->size == 1) {
+        cur_p->data[0].p = 1.0f;
+        return;
+    }
+
+    // max logit for numerical stability
+    float max_l = cur_p->data[0].logit;
+    if (!cur_p->sorted) {
+        for (size_t i = 1; i < cur_p->size; ++i) {
+            max_l = std::max(max_l, cur_p->data[i].logit);
+        }
+    }
+
+    // apply softmax to obtain the probabilities
+    double sum_cum = 0.0f;
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        float p = expf(cur_p->data[i].logit - max_l);
+        cur_p->data[i].p = p;
+        sum_cum += p;
+    }
+
+#if 1
+    // sample from the obtained probabilities and normalize the probs in a single pass
+    // this is ~3x faster on Mac with full gpt-oss vocab than the version below
+    //
+    std::uniform_real_distribution<double> dist(0.0f, 1.0f);
+    const double rnd = dist(ctx->rng);
+
+          double sum_run = 0.0f;
+    const double sum_tgt = sum_cum*rnd;
+
+    bool found = false;
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (!found) {
+            // accumulate probs until we reach the target sum
+            sum_run += cur_p->data[i].p;
+            if (sum_run >= sum_tgt) {
+                cur_p->selected = i;
+                found = true;
+            }
+        }
+
+        // normalize probs
+        cur_p->data[i].p /= sum_cum;
+    }
+
+    // fallback to the last token (don't think this can happen)
+    assert(found);
+    if (!found) {
+        cur_p->selected = cur_p->size - 1;
+    }
+#else
+    // for clarity, this is the same as above but does one pass for normalization and one extra pass for sampling
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].p /= sum_cum;
+    }
 
     cur_p->selected = llama_sample_dist(cur_p, ctx->rng);
+#endif
 }
 
 static struct llama_sampler * llama_sampler_dist_clone(const struct llama_sampler * smpl) {
@@ -626,32 +718,6 @@ struct llama_sampler * llama_sampler_init_dist(uint32_t seed) {
     );
 }
 
-// softmax
-
-static const char * llama_sampler_softmax_name(const struct llama_sampler * /*smpl*/) {
-    return "softmax";
-}
-
-static void llama_sampler_softmax_apply(struct llama_sampler * /*smpl*/, llama_token_data_array * cur_p) {
-    llama_sampler_softmax_impl(cur_p);
-}
-
-static struct llama_sampler_i llama_sampler_softmax_i = {
-    /* .name   = */ llama_sampler_softmax_name,
-    /* .accept = */ nullptr,
-    /* .apply  = */ llama_sampler_softmax_apply,
-    /* .reset  = */ nullptr,
-    /* .clone  = */ nullptr,
-    /* .free   = */ nullptr,
-};
-
-struct llama_sampler * llama_sampler_init_softmax() {
-    return llama_sampler_init(
-        /* .iface = */ &llama_sampler_softmax_i,
-        /* .ctx   = */ nullptr
-    );
-}
-
 // top-k
 
 struct llama_sampler_top_k {
@@ -663,7 +729,7 @@ static const char * llama_sampler_top_k_name(const struct llama_sampler * /*smpl
 }
 
 static void llama_sampler_top_k_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_top_k *) smpl->ctx;
+    auto * ctx = (llama_sampler_top_k *) smpl->ctx;
     llama_sampler_top_k_impl(cur_p, ctx->k);
 }
 
@@ -699,6 +765,8 @@ struct llama_sampler * llama_sampler_init_top_k(int32_t k) {
 struct llama_sampler_top_p {
     const float  p;
     const size_t min_keep;
+
+    std::vector<llama_token_data> buf_sort;
 };
 
 static const char * llama_sampler_top_p_name(const struct llama_sampler * /*smpl*/) {
@@ -706,20 +774,35 @@ static const char * llama_sampler_top_p_name(const struct llama_sampler * /*smpl
 }
 
 static void llama_sampler_top_p_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_top_p *) smpl->ctx;
+    auto * ctx = (llama_sampler_top_p *) smpl->ctx;
 
     if (ctx->p >= 1.0f) {
         return;
     }
 
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, false);
+
+    size_t k = cur_p->size;
+    auto * pdata = cur_p->data;
+
+    auto & buf_sort = ctx->buf_sort;
+
+    // if not sorted, try adaptive top-k sorting
+    if (!cur_p->sorted && cur_p->size > 1024) {
+        k = std::min<size_t>(256, cur_p->size);
+        llama_token_data_array_partial_sort(*cur_p, k, buf_sort);
+        pdata = buf_sort.data();
+    } else if (!cur_p->sorted) {
+        // small candidates -> sort inplace
+        llama_token_data_array_partial_sort_inplace(cur_p, k);
+    }
 
     // Compute the cumulative probabilities
     float cum_sum = 0.0f;
     size_t last_idx = cur_p->size;
 
     for (size_t i = 0; i < cur_p->size; ++i) {
-        cum_sum += cur_p->data[i].p;
+        cum_sum += pdata[i].p;
 
         // Check if the running sum is at least p or if we have kept at least min_keep tokens
         // we set the last index to i+1 to indicate that the current iterate should be included in the set
@@ -727,9 +810,21 @@ static void llama_sampler_top_p_apply(struct llama_sampler * smpl, llama_token_d
             last_idx = i + 1;
             break;
         }
+
+        // we exceeded the current top-k heuristic -> increase k and continue
+        if (!cur_p->sorted && i == k - 1) {
+            k = cur_p->size;
+            llama_token_data_array_partial_sort(*cur_p, k, buf_sort);
+            pdata = buf_sort.data();
+        }
     }
 
     // Resize the output vector to keep only the top-p tokens
+    if (!cur_p->sorted) {
+        std::copy(buf_sort.data(), buf_sort.data() + last_idx, cur_p->data);
+        cur_p->sorted = true;
+    }
+
     cur_p->size = last_idx;
 }
 
@@ -757,6 +852,7 @@ struct llama_sampler * llama_sampler_init_top_p(float p, size_t min_keep) {
         /* .ctx   = */ new llama_sampler_top_p {
             /* .p        = */ p,
             /* .min_keep = */ min_keep,
+            /* .buf_sort = */ {},
         }
     );
 }
@@ -773,7 +869,7 @@ static const char * llama_sampler_min_p_name(const struct llama_sampler * /*smpl
 }
 
 static void llama_sampler_min_p_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_min_p *) smpl->ctx;
+    auto * ctx = (llama_sampler_min_p *) smpl->ctx;
 
     if (ctx->p <= 0.0f || !cur_p->size) {
         return;
@@ -799,7 +895,7 @@ static void llama_sampler_min_p_apply(struct llama_sampler * smpl, llama_token_d
 
         // if we have enough values the operation was a success
         if (!filtered_tokens.empty() && filtered_tokens.size() >= ctx->min_keep) {
-            memcpy(cur_p->data, filtered_tokens.data(), filtered_tokens.size()*sizeof(llama_token_data));
+            std::copy(filtered_tokens.begin(), filtered_tokens.end(), cur_p->data);
             cur_p->size = filtered_tokens.size();
             min_p_applied = true;
         }
@@ -809,10 +905,7 @@ static void llama_sampler_min_p_apply(struct llama_sampler * smpl, llama_token_d
     if (!min_p_applied) {
         // Sort the logits in descending order
         if (!cur_p->sorted) {
-            std::sort(cur_p->data, cur_p->data + cur_p->size, [](const llama_token_data & a, const llama_token_data & b) {
-                return a.logit > b.logit;
-            });
-            cur_p->sorted = true;
+            llama_token_data_array_partial_sort_inplace(cur_p, cur_p->size);
         }
 
         const float min_logit = cur_p->data[0].logit + logf(ctx->p); // min logit for p_i >= p * p_max
@@ -869,7 +962,7 @@ static const char * llama_sampler_typical_name(const struct llama_sampler * /*sm
 }
 
 static void llama_sampler_typical_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_typical *) smpl->ctx;
+    auto * ctx = (llama_sampler_typical *) smpl->ctx;
 
     // Reference implementation:
     // https://github.com/huggingface/transformers/compare/main...cimeister:typical-sampling:typical-pr
@@ -878,7 +971,7 @@ static void llama_sampler_typical_apply(struct llama_sampler * smpl, llama_token
     }
 
     // Compute the softmax of logits and calculate entropy
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
     float entropy = 0.0f;
     for (size_t i = 0; i < cur_p->size; ++i) {
@@ -1012,7 +1105,7 @@ static const char * llama_sampler_temp_ext_name(const struct llama_sampler * /*s
 }
 
 static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_temp_ext *) smpl->ctx;
+    auto * ctx = (llama_sampler_temp_ext *) smpl->ctx;
     if (ctx->delta > 0) {
         const float min_temp = std::max(0.0f, ctx->temp - ctx->delta);
         const float max_temp = ctx->temp + ctx->delta;
@@ -1027,7 +1120,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
         // Calculate maximum possible entropy
         float max_entropy = -logf(1.0f / cur_p->size);
 
-        llama_sampler_softmax_impl(cur_p);
+        llama_sampler_softmax_impl(cur_p, true);
 
         // Calculate entropy of the softmax probabilities
         float entropy = 0.0f;
@@ -1139,17 +1232,20 @@ static void llama_sample_xtc_apply(struct llama_sampler * smpl, llama_token_data
 
     std::uniform_real_distribution<float> distribution(0.0f, 1.0f);
     float chance = distribution(ctx->rng);
-    if (chance > ctx->probability) return;
+    if (chance > ctx->probability) {
+        return;
+    }
 
-    // in case it's not sorted/recalculated yet
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
     int pos_last = 0;
 
     for (size_t i = 0; i < cur_p->size; ++i) {
         if (cur_p->data[i].p >= ctx->threshold) {
             pos_last = i;
-        } else break;
+        } else {
+            break;
+        }
     }
 
     if (cur_p->size - pos_last >= ctx->min_keep && pos_last > 0) {
@@ -1231,7 +1327,7 @@ static const char * llama_sampler_mirostat_name(const struct llama_sampler * /*s
 static void llama_sampler_mirostat_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_mirostat *) smpl->ctx;
 
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
     // Estimate s_hat using the most probable m tokens
     float s_hat = 0.0;
@@ -1250,7 +1346,8 @@ static void llama_sampler_mirostat_apply(struct llama_sampler * smpl, llama_toke
     float k = powf((epsilon_hat * powf(2, ctx->mu)) / (1 - powf(ctx->n_vocab, -epsilon_hat)), 1 / s_hat);
 
     llama_sampler_top_k_impl(cur_p, std::max(int(k), 1));
-    llama_sampler_softmax_impl(cur_p);
+
+    llama_sampler_softmax_impl(cur_p, true);
 
     const int idx = llama_sample_dist(cur_p, ctx->rng);
 
@@ -1336,7 +1433,7 @@ static const char * llama_sampler_mirostat_v2_name(const struct llama_sampler *
 static void llama_sampler_mirostat_v2_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_mirostat_v2 *) smpl->ctx;
 
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
     // Truncate the words with surprise values greater than mu
     cur_p->size = std::distance(cur_p->data, std::find_if(cur_p->data, cur_p->data + cur_p->size, [&](const llama_token_data & candidate) {
@@ -1348,7 +1445,7 @@ static void llama_sampler_mirostat_v2_apply(struct llama_sampler * smpl, llama_t
     }
 
     // Normalize the probabilities of the remaining words
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
     const int idx = llama_sample_dist(cur_p, ctx->rng);
 
@@ -1540,7 +1637,7 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
                 trigger_pattern += std::regex_replace(trigger_words[i], special_chars, "\\$0");
             }
             trigger_pattern += ")[\\s\\S]*";
-            auto trigger_pattern_c = trigger_pattern.c_str();
+            const auto * trigger_pattern_c = trigger_pattern.c_str();
             trigger_patterns = &trigger_pattern_c;
             num_trigger_patterns = 1;
         }
@@ -1748,7 +1845,7 @@ 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;
+    auto * ctx = (llama_sampler_top_n_sigma *) smpl->ctx;
 
     if (ctx->n <= 0.0f || cur_p->size <= 1) {
         return;
@@ -1780,13 +1877,14 @@ static void llama_sampler_top_n_sigma_apply(struct llama_sampler * smpl, llama_t
     }
     float std = valid_count > 0 ? sqrt(acc/valid_count) : 0;
 
-    //apply mask
+    // apply mask
     for (size_t i = 0; i < cur_p->size; ++i) {
         if (cur_p->data[i].logit < max - (ctx->n * std)) {
             cur_p->data[i].logit = -INFINITY;
         }
     }
-    llama_sampler_softmax_impl(cur_p);
+
+    llama_sampler_softmax_impl(cur_p, true);
 }
 
 static struct llama_sampler * llama_sampler_top_n_sigma_clone(const struct llama_sampler * smpl) {
@@ -1991,7 +2089,9 @@ static void llama_sampler_dry_apply(struct llama_sampler * smpl, llama_token_dat
 
     {
         const int last = last_n_repeat - 1;
-        int rt = 0, lt = 0;
+
+        int rt = 0;
+        int lt = 0;
 
         for (int k = 1; k < last_n_repeat; ++k) {
             if (k > rt) {
@@ -2135,8 +2235,8 @@ static struct llama_sampler_i llama_sampler_dry_i = {
     /* .free   = */ llama_sampler_dry_free,
 };
 
-struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab, int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
-    int32_t effective_dry_penalty_last_n = (dry_penalty_last_n == -1) ? context_size : std::max(dry_penalty_last_n, 0);
+struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab, int32_t n_ctx_train, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
+    int32_t effective_dry_penalty_last_n = (dry_penalty_last_n == -1) ? n_ctx_train : std::max(dry_penalty_last_n, 0);
     std::unordered_multimap<llama_token, std::vector<llama_token>> processed_breakers;
     const int MAX_CHAR_LEN = 40;
     const int MAX_SEQ_LEN = 20;
@@ -2169,7 +2269,7 @@ struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab,
     return llama_sampler_init(
         /* .iface = */ &llama_sampler_dry_i,
         /* .ctx   = */ new llama_sampler_dry {
-            /* .total_context_size     = */ context_size,
+            /* .total_context_size     = */ n_ctx_train,
             /* .dry_multiplier         = */ dry_multiplier,
             /* .dry_base               = */ dry_base,
             /* .dry_allowed_length     = */ dry_allowed_length,
@@ -2308,7 +2408,7 @@ static const char * llama_sampler_infill_name(const struct llama_sampler * /*smp
 static void llama_sampler_infill_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_infill *) smpl->ctx;
 
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
 #if defined(GGML_DEBUG_SAMPLER_INFILL)
 #define LOG_DBG_CUR LLAMA_LOG_DEBUG

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

@@ -434,6 +434,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "(?:'[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;
+            case LLAMA_VOCAB_PRE_TYPE_GROK_2:
+                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}| ?[^\\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}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -1763,7 +1770,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 const size_t n_precompiled_charsmap = gguf_get_arr_data_n(ctx, precompiled_charsmap_keyidx);
                 const char * pc = (const char *) gguf_get_arr_data(ctx, precompiled_charsmap_keyidx);
                 precompiled_charsmap.assign(pc, pc + n_precompiled_charsmap);
-#ifdef IS_BIG_ENDIAN
+#if defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
                 // correct endiannes of data in precompiled_charsmap binary blob
                 uint32_t * xcda_blob_size = (uint32_t *) &precompiled_charsmap[0];
                 *xcda_blob_size = __builtin_bswap32(*xcda_blob_size);
@@ -1944,7 +1951,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_TRILLION;
                 clean_spaces = false;
             } else if (
-                tokenizer_pre == "bailingmoe") {
+                tokenizer_pre == "bailingmoe" ||
+                tokenizer_pre == "llada-moe") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE;
                 clean_spaces = false;
             } else if (
@@ -1963,6 +1971,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "kimi-k2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_KIMI_K2;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "grok-2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_GROK_2;
+                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;
@@ -2331,7 +2343,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
         // @ngxson : quick hack for gpt-oss, always render these tokens
         for (const auto & t : token_to_id) {
-            if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>") {
+            if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>" || t.first == "<|constrain|>") {
                 id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
             }
         }
@@ -2378,6 +2390,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
             if (has_return && has_call && has_end) {
                 special_eog_ids.erase(end_id);
+                id_to_token[end_id].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
                 LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
             }
         }
@@ -2459,7 +2472,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
         // set attributes by model/tokenizer/architecture name
         if (false
                 || _contains_any(tokenizer_pre, {"jina-v2-de", "jina-v2-es", "jina-v2-code"})
-                || _contains_any(general_arch, {"nomic-bert-moe"})
+                || _contains_any(general_arch, {"nomic-bert-moe", "jina-bert-v3"})
            ) {
             if (token_to_id.count("<mask>") == 0) {
                 LLAMA_LOG_WARN("%s: Mask token is missing in vocab, please reconvert model!\n", __func__);

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

@@ -47,6 +47,7 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_HUNYUAN        = 36,
     LLAMA_VOCAB_PRE_TYPE_KIMI_K2        = 37,
     LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE  = 38,
+    LLAMA_VOCAB_PRE_TYPE_GROK_2         = 39,
 };
 
 struct LLM_KV;

llama/llama.cpp/src/llama.cpp

@@ -25,6 +25,18 @@
 // interface implementation
 //
 
+const char * llama_flash_attn_type_name(enum llama_flash_attn_type flash_attn_type) {
+    switch (flash_attn_type) {
+        case LLAMA_FLASH_ATTN_TYPE_AUTO:
+            return "auto";
+        case LLAMA_FLASH_ATTN_TYPE_DISABLED:
+            return "disabled";
+        case LLAMA_FLASH_ATTN_TYPE_ENABLED:
+            return "enabled";
+    }
+    GGML_ABORT("fatal error");
+}
+
 struct llama_sampler_chain_params llama_sampler_chain_default_params() {
     struct llama_sampler_chain_params result = {
         /*.no_perf                     =*/ true,
@@ -47,6 +59,7 @@ bool llama_supports_mlock(void) {
 
 bool llama_supports_gpu_offload(void) {
     return ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU) != nullptr ||
+           ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU) != nullptr ||
            llama_supports_rpc();
 }
 
@@ -71,8 +84,10 @@ void llama_numa_init(enum ggml_numa_strategy numa) {
         GGML_ASSERT(dev && "CPU backend is not loaded");
         auto * reg = ggml_backend_dev_backend_reg(dev);
         auto * numa_init_fn = (decltype(ggml_numa_init) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_numa_init");
+        if (numa_init_fn) {
             numa_init_fn(numa);
         }
+    }
 }
 
 void llama_backend_free(void) {
@@ -170,8 +185,13 @@ static struct llama_model * llama_model_load_from_file_impl(
             model->devices.push_back(*dev);
         }
     } else {
+        // default device selection
+
+        // build list of available devices
+        std::vector<ggml_backend_dev_t> gpus;
+        std::vector<ggml_backend_dev_t> igpus;
         std::vector<ggml_backend_dev_t> rpc_servers;
-        // use all available devices
+
         for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
             ggml_backend_dev_t dev = ggml_backend_dev_get(i);
             switch (ggml_backend_dev_type(dev)) {
@@ -180,19 +200,51 @@ static struct llama_model * llama_model_load_from_file_impl(
                     // skip CPU backends since they are handled separately
                     break;
 
-                case GGML_BACKEND_DEVICE_TYPE_GPU:
+                case GGML_BACKEND_DEVICE_TYPE_GPU: {
                     ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
                     if (ggml_backend_reg_name(reg) == std::string("RPC")) {
                         rpc_servers.push_back(dev);
                     } else {
-                        model->devices.push_back(dev);
+                        // check if there is already a GPU with the same device id
+                        ggml_backend_dev_props props;
+                        ggml_backend_dev_get_props(dev, &props);
+                        auto it = std::find_if(gpus.begin(), gpus.end(), [&props](ggml_backend_dev_t d) {
+                            ggml_backend_dev_props d_props;
+                            ggml_backend_dev_get_props(d, &d_props);
+                            if (props.device_id && d_props.device_id) {
+                                return strcmp(props.device_id, d_props.device_id) == 0;
+                            }
+                            return false;
+                        });
+
+                        if (it != gpus.end()) {
+                            LLAMA_LOG_INFO("%s: skipping device %s (%s) with id %s - already using device %s (%s) with the same id\n",
+                                    __func__,
+                                    ggml_backend_dev_name(dev), ggml_backend_dev_description(dev),
+                                    props.device_id ? props.device_id : "unknown id",
+                                    ggml_backend_dev_name(*it), ggml_backend_dev_description(*it));
+                        } else {
+                            gpus.push_back(dev);
+                        }
                     }
                     break;
                 }
+
+                case GGML_BACKEND_DEVICE_TYPE_IGPU:
+                    igpus.push_back(dev);
+                    break;
             }
-        // add RPC servers at the front of the list
-        if (!rpc_servers.empty()) {
+        }
+
+        // add RPC servers at the front of the list to minimize network transfers
         model->devices.insert(model->devices.begin(), rpc_servers.begin(), rpc_servers.end());
+
+        // add GPUs
+        model->devices.insert(model->devices.end(), gpus.begin(), gpus.end());
+
+        // add integrated GPUs only if no other devices were found
+        if (model->devices.empty()) {
+            model->devices.insert(model->devices.end(), igpus.begin(), igpus.end());
         }
     }
 
@@ -213,9 +265,12 @@ static struct llama_model * llama_model_load_from_file_impl(
     }
 
     for (auto * dev : model->devices) {
-        size_t free, total; // NOLINT
-        ggml_backend_dev_memory(dev, &free, &total);
-        LLAMA_LOG_INFO("%s: using device %s (%s) - %zu MiB free\n", __func__, ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), free/1024/1024);
+        ggml_backend_dev_props props;
+        ggml_backend_dev_get_props(dev, &props);
+        LLAMA_LOG_INFO("%s: using device %s (%s) (%s) - %zu MiB free\n", __func__,
+                ggml_backend_dev_name(dev), ggml_backend_dev_description(dev),
+                props.device_id ? props.device_id : "unknown id",
+                props.memory_free/1024/1024);
     }
 
     const int status = llama_model_load(path_model, splits, *model, params);

llama/llama.cpp/src/unicode.h

@@ -4,6 +4,7 @@
 #include <string>
 #include <vector>
 
+// TODO: reimplement this structure in endian-independent way
 struct unicode_cpt_flags {
     enum {
         UNDEFINED       = 0x0001,
@@ -15,6 +16,10 @@ struct unicode_cpt_flags {
         SYMBOL          = 0x0040,  // regex: \p{S}
         CONTROL         = 0x0080,  // regex: \p{C}
         MASK_CATEGORIES = 0x00FF,
+        WHITESPACE      = 0x0100,
+        LOWERCASE       = 0x0200,
+        UPPERCASE       = 0x0400,
+        NFD             = 0x0800,
     };
 
     // codepoint type
@@ -34,11 +39,49 @@ struct unicode_cpt_flags {
 
     // decode from uint16
     inline unicode_cpt_flags(const uint16_t flags = 0) {
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
         *reinterpret_cast<uint16_t*>(this) = flags;
+#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+        is_undefined   = (flags & UNDEFINED)   ? 1 : 0;
+        is_number      = (flags & NUMBER)      ? 1 : 0;
+        is_letter      = (flags & LETTER)      ? 1 : 0;
+        is_separator   = (flags & SEPARATOR)   ? 1 : 0;
+        is_accent_mark = (flags & ACCENT_MARK) ? 1 : 0;
+        is_punctuation = (flags & PUNCTUATION) ? 1 : 0;
+        is_symbol      = (flags & SYMBOL)      ? 1 : 0;
+        is_control     = (flags & CONTROL)     ? 1 : 0;
+        is_whitespace  = (flags & WHITESPACE)  ? 1 : 0;
+        is_lowercase   = (flags & LOWERCASE)   ? 1 : 0;
+        is_uppercase   = (flags & UPPERCASE)   ? 1 : 0;
+        is_nfd         = (flags & NFD)         ? 1 : 0;
+#else
+#error Unexpected or undefined __BYTE_ORDER__
+#endif
     }
 
     inline uint16_t as_uint() const {
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
         return *reinterpret_cast<const uint16_t*>(this);
+#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+        uint16_t result =
+              is_undefined   * UNDEFINED
+            + is_number      * NUMBER
+            + is_letter      * LETTER
+            + is_separator   * SEPARATOR
+            + is_accent_mark * ACCENT_MARK
+            + is_punctuation * PUNCTUATION
+            + is_symbol      * SYMBOL
+            + is_control     * CONTROL
+            + is_whitespace  * WHITESPACE
+            + is_lowercase   * LOWERCASE
+            + is_uppercase   * UPPERCASE
+            + is_nfd         * NFD
+            ;
+
+        return result;
+#else
+#error Unexpected or undefined __BYTE_ORDER__
+#endif
     }
 
     inline uint16_t category_flag() const {

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

@@ -44,6 +44,7 @@
 #define KEY_WIN_ATTN_PATTERN      "clip.vision.n_wa_pattern"
 #define KEY_ATTN_WINDOW_SIZE      "clip.vision.window_size"
 #define KEY_MINICPMV_VERSION      "clip.minicpmv_version"
+#define KEY_MINICPMV_QUERY_NUM    "clip.minicpmv_query_num"
 
 // audio-specific
 #define KEY_A_NUM_MEL_BINS      "clip.audio.num_mel_bins"
@@ -81,6 +82,7 @@
 #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_NORM_B   "mm.input_norm.bias"
 #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
@@ -132,6 +134,8 @@ enum projector_type {
     PROJECTOR_TYPE_QWEN2A,
     PROJECTOR_TYPE_QWEN25O, // will be replaced by QWEN2A or QWEN25VL depending on clip_ctx
     PROJECTOR_TYPE_VOXTRAL,
+    PROJECTOR_TYPE_LFM2,
+    PROJECTOR_TYPE_KIMIVL,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -152,6 +156,8 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_QWEN2A,    "qwen2a"},
     { PROJECTOR_TYPE_QWEN25O,   "qwen2.5o"},
     { PROJECTOR_TYPE_VOXTRAL,   "voxtral"},
+    { PROJECTOR_TYPE_LFM2,      "lfm2"},
+    { PROJECTOR_TYPE_KIMIVL,    "kimivl"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {

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

@@ -214,6 +214,7 @@ struct clip_hparams {
     // legacy
     bool has_llava_projector = false;
     int minicpmv_version = 0;
+    int32_t minicpmv_query_num = 0;         // MiniCPM-V query number
 };
 
 struct clip_layer {
@@ -277,6 +278,7 @@ struct clip_model {
 
     // LLaVA projection
     ggml_tensor * mm_input_norm_w = nullptr;
+    ggml_tensor * mm_input_norm_b = nullptr;
     ggml_tensor * mm_0_w = nullptr;
     ggml_tensor * mm_0_b = nullptr;
     ggml_tensor * mm_2_w = nullptr;
@@ -417,6 +419,7 @@ struct clip_ctx {
             }
             if (!backend) {
                 backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr);
+                backend = backend ? backend : ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU, nullptr);
             }
         }
 
@@ -500,11 +503,17 @@ struct clip_graph {
 
     ggml_cgraph * build_siglip() {
         ggml_tensor * inp = build_inp();
+
+        ggml_tensor * learned_pos_embd = model.position_embeddings;
+        if (ctx->proj_type() == PROJECTOR_TYPE_LFM2) {
+            learned_pos_embd = resize_position_embeddings();
+        }
+
         ggml_tensor * cur = build_vit(
                                 inp, n_patches,
                                 NORM_TYPE_NORMAL,
                                 hparams.ffn_op,
-                                model.position_embeddings,
+                                learned_pos_embd,
                                 nullptr);
 
         if (ctx->proj_type() == PROJECTOR_TYPE_GEMMA3) {
@@ -513,8 +522,8 @@ struct clip_graph {
             const int patches_per_image = n_patches_x;
             const int kernel_size = hparams.proj_scale_factor;
 
-            cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-            cur = ggml_reshape_4d(ctx0, cur, patches_per_image, patches_per_image, n_embd, batch_size);
+            cur = ggml_transpose(ctx0, cur);
+            cur = ggml_cont_4d(ctx0, cur, patches_per_image, patches_per_image, n_embd, batch_size);
 
             // doing a pool2d to reduce the number of output tokens
             cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0);
@@ -531,29 +540,27 @@ struct clip_graph {
                 cur);
 
         } else if (ctx->proj_type() == PROJECTOR_TYPE_IDEFICS3) {
+            // pixel_shuffle
             // https://github.com/huggingface/transformers/blob/0a950e0bbe1ed58d5401a6b547af19f15f0c195e/src/transformers/models/idefics3/modeling_idefics3.py#L578
-
             const int scale_factor = model.hparams.proj_scale_factor;
-            const int n_embd = cur->ne[0];
-            const int seq    = cur->ne[1];
-            const int bsz    = 1; // batch size, always 1 for now since we don't support batching
-            const int height = std::sqrt(seq);
-            const int width  = std::sqrt(seq);
-            GGML_ASSERT(scale_factor != 0);
-            cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, width / scale_factor, height, bsz);
-            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
-            cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
-                n_embd * scale_factor * scale_factor,
-                height / scale_factor,
-                width / scale_factor,
-                bsz);
-            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
-            cur = ggml_reshape_3d(ctx0, ggml_cont(ctx0, cur),
-                n_embd * scale_factor * scale_factor,
-                seq / (scale_factor * scale_factor),
-                bsz);
-
+            cur = build_patch_merge_permute(cur, scale_factor);
             cur = ggml_mul_mat(ctx0, model.projection, cur);
+
+        } else if (ctx->proj_type() == PROJECTOR_TYPE_LFM2) {
+            // pixel unshuffle block
+            const int scale_factor = model.hparams.proj_scale_factor;
+            cur = build_patch_merge_permute(cur, scale_factor);
+
+            // projection
+            cur = ggml_norm(ctx0, cur, 1e-5); // default nn.LayerNorm
+            cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);
+            cur = ggml_add(ctx0, cur, model.mm_input_norm_b);
+
+            cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
+            cur = ggml_add(ctx0, cur, model.mm_1_b);
+            cur = ggml_gelu(ctx0, cur);
+            cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);
+            cur = ggml_add(ctx0, cur, model.mm_2_b);
         } else {
             GGML_ABORT("SigLIP: Unsupported projector type");
         }
@@ -681,15 +688,15 @@ struct clip_graph {
             auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
             inp = ggml_add(ctx0, inp, inp_1);
 
-            inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 2, 0, 3));  // [w, h, c, b] -> [c, w, h, b]
-            inp = ggml_reshape_4d(
+            inp = ggml_permute(ctx0, inp, 1, 2, 0, 3);  // [w, h, c, b] -> [c, w, h, b]
+            inp = ggml_cont_4d(
                 ctx0, inp,
                 n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
             inp = ggml_reshape_4d(
                 ctx0, inp,
                 n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
-            inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 0, 2, 1, 3));
-            inp = ggml_reshape_3d(
+            inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
+            inp = ggml_cont_3d(
                 ctx0, inp,
                 n_embd, n_patches_x * n_patches_y, batch_size);
         }
@@ -879,21 +886,8 @@ struct clip_graph {
             int n_embd = clip_n_mmproj_embd(ctx);
             const int d_head = 128;
             int n_head = n_embd/d_head;
-            int num_query = 96;
-            if (ctx->model.hparams.minicpmv_version == 2) {
-                // MiniCPM-V 2.5
-                num_query = 96;
-            } else if (ctx->model.hparams.minicpmv_version == 3) {
-                // MiniCPM-V 2.6
-                num_query = 64;
-            } else if (ctx->model.hparams.minicpmv_version == 4) {
-                // MiniCPM-o 2.6
-                num_query = 64;
-            } else if (ctx->model.hparams.minicpmv_version == 5) {
-                // MiniCPM-V 4.0
-                num_query = 64;
-            }
-
+            // Use actual config value if available, otherwise fall back to hardcoded values
+            int num_query = ctx->model.hparams.minicpmv_query_num;
             ggml_tensor * Q = ggml_add(ctx0,
                 ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q),
                 model.mm_model_attn_q_b);
@@ -967,14 +961,14 @@ struct clip_graph {
             GGML_ASSERT(scale_factor > 0);
             cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, height / scale_factor, width, bsz);
             cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
-            cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
+            cur = ggml_cont_4d(ctx0, cur,
                 n_embd * scale_factor * scale_factor,
                 height / scale_factor,
                 width / scale_factor,
                 bsz);
             cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
             // flatten to 2D
-            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, cur),
+            cur = ggml_cont_2d(ctx0, cur,
                 n_embd * scale_factor * scale_factor,
                 cur->ne[1] * cur->ne[2]);
         }
@@ -1060,14 +1054,14 @@ struct clip_graph {
                 n_patches_y,
                 bsz);
             cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
-            cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
+            cur = ggml_cont_4d(ctx0, cur,
                 n_embd * scale_factor * scale_factor,
                 n_patches_x / scale_factor,
                 n_patches_y / scale_factor,
                 bsz);
-            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+            //cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
             // flatten to 2D
-            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, cur),
+            cur = ggml_cont_2d(ctx0, cur,
                 n_embd * scale_factor * scale_factor,
                 n_patches / scale_factor / scale_factor);
             cb(cur, "pixel_shuffle", -1);
@@ -1092,6 +1086,67 @@ struct clip_graph {
         return gf;
     }
 
+    ggml_cgraph * build_kimivl() {
+        // 2D input positions
+        ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+        ggml_set_name(pos_h, "pos_h");
+        ggml_set_input(pos_h);
+
+        ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+        ggml_set_name(pos_w, "pos_w");
+        ggml_set_input(pos_w);
+
+        ggml_tensor * learned_pos_embd = resize_position_embeddings();
+
+        // build ViT with 2D position embeddings
+        auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+            // first half is X axis and second half is Y axis
+            return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);
+        };
+
+        ggml_tensor * inp = build_inp();
+        ggml_tensor * cur = build_vit(
+                                inp, n_patches,
+                                NORM_TYPE_NORMAL,
+                                hparams.ffn_op,
+                                learned_pos_embd,
+                                add_pos);
+
+        cb(cur, "vit_out", -1);
+
+        {
+            // patch_merger
+            const int scale_factor = model.hparams.proj_scale_factor;
+            cur = build_patch_merge_permute(cur, scale_factor);
+
+            // projection norm
+            int proj_inp_dim = cur->ne[0];
+            cur = ggml_view_2d(ctx0, cur,
+                n_embd, cur->ne[1] * scale_factor * scale_factor,
+                ggml_row_size(cur->type, n_embd), 0);
+            cur = ggml_norm(ctx0, cur, 1e-5); // default nn.LayerNorm
+            cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);
+            cur = ggml_add(ctx0, cur, model.mm_input_norm_b);
+            cur = ggml_view_2d(ctx0, cur,
+                proj_inp_dim, cur->ne[1] / scale_factor / scale_factor,
+                ggml_row_size(cur->type, proj_inp_dim), 0);
+            cb(cur, "proj_inp_normed", -1);
+
+            // projection mlp
+            cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
+            cur = ggml_add(ctx0, cur, model.mm_1_b);
+            cur = ggml_gelu(ctx0, cur);
+            cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);
+            cur = ggml_add(ctx0, cur, model.mm_2_b);
+            cb(cur, "proj_out", -1);
+        }
+
+        // build the graph
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     // this graph is used by llava, granite and glm
     // due to having embedding_stack (used by granite), we cannot reuse build_vit
     ggml_cgraph * build_llava() {
@@ -1300,8 +1355,8 @@ struct clip_graph {
                 ggml_tensor * block_1 = nullptr;
                 {
                     // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24]
-                    mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3));
-                    mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
+                    mlp_3 = ggml_permute(ctx0, mlp_3, 1, 0, 2, 3);
+                    mlp_3 = ggml_cont_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
                     // stride = 1, padding = 1, bias is nullptr
                     block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);
 
@@ -1406,9 +1461,9 @@ struct clip_graph {
                 mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b);
                 // mlp_2 ne = [2048, 576, 1, 1]
                 // // AVG Pool Layer 2*2, strides = 2
-                mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3));
+                mlp_2 = ggml_permute(ctx0, mlp_2, 1, 0, 2, 3);
                 // mlp_2 ne = [576, 2048, 1, 1]
-                mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
+                mlp_2 = ggml_cont_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
                 // mlp_2 ne [24, 24, 2048, 1]
                 mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
                 // weight ne = [3, 3, 2048, 1]
@@ -1428,8 +1483,8 @@ struct clip_graph {
         // glm projector
         else if (ctx->proj_type() == PROJECTOR_TYPE_GLM_EDGE) {
             size_t gridsz = (size_t)sqrt(embeddings->ne[1]);
-            embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3));
-            embeddings = ggml_reshape_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]);
+            embeddings = ggml_permute(ctx0,embeddings,1,0,2,3);
+            embeddings = ggml_cont_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]);
             embeddings = ggml_conv_2d(ctx0, model.mm_model_adapter_conv_w, embeddings, 2, 2, 0, 0, 1, 1);
             embeddings = ggml_reshape_3d(ctx0, embeddings,embeddings->ne[0]*embeddings->ne[1] , embeddings->ne[2], batch_size);
             embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings, 1, 0, 2, 3));
@@ -1585,6 +1640,29 @@ private:
         }
     }
 
+    // siglip2 naflex
+    ggml_tensor * resize_position_embeddings() {
+        ggml_tensor * pos_embd = model.position_embeddings;
+        const int height       = img.ny / patch_size;
+        const int width        = img.nx / patch_size;
+        const uint32_t mode    = GGML_SCALE_MODE_BILINEAR;
+        const int n_per_side   = (int)std::sqrt(pos_embd->ne[1]);
+
+        GGML_ASSERT(pos_embd);
+
+        if (height == n_per_side && width == n_per_side) {
+            return pos_embd;
+        }
+
+        pos_embd = ggml_reshape_3d(ctx0, pos_embd, n_embd, n_per_side, n_per_side);  // -> (n_embd, n_per_side, n_per_side)
+        pos_embd = ggml_permute(ctx0, pos_embd, 2, 0, 1, 3);                         // -> (n_per_side, n_per_side, n_embd)
+        pos_embd = ggml_interpolate(ctx0, pos_embd, width, height, n_embd, 1, mode); // -> (width, height, n_embd)
+        pos_embd = ggml_permute(ctx0, pos_embd, 1, 2, 0, 3);                         // -> (n_embd, width, height)
+        pos_embd = ggml_cont_2d(ctx0, pos_embd, n_embd, width * height);             // -> (n_embd, width * height)
+
+        return pos_embd;
+    }
+
     // build vision transformer (ViT) cgraph
     // this function should cover most of the models
     // if your model has specific features, you should probably duplicate this function
@@ -1963,7 +2041,6 @@ private:
                 ggml_row_size(cur->type, n_dim),
                 ggml_row_size(cur->type, n_dim*n_head),
                 n_dim/2 * ggml_element_size(cur));
-            second = ggml_cont(ctx0, second); // copy, because ggml_rope don't play well with non-contiguous tensors
             second = ggml_rope_ext(
                 ctx0,
                 second,
@@ -1980,6 +2057,39 @@ private:
         return cur;
     }
 
+    // aka pixel_shuffle / pixel_unshuffle / patch_merger (Kimi-VL)
+    // support dynamic resolution
+    ggml_tensor * build_patch_merge_permute(ggml_tensor * cur, int scale_factor) {
+        GGML_ASSERT(scale_factor > 1);
+
+        const int n_embd = cur->ne[0];
+        int width  = img.nx / patch_size;
+        int height = img.ny / patch_size;
+
+        // pad width and height to factor
+        const int64_t pad_width  = CLIP_ALIGN(width,  scale_factor) - width;
+        const int64_t pad_height = CLIP_ALIGN(height, scale_factor) - height;
+        cur = ggml_reshape_3d(ctx0, cur, n_embd, width, height);
+        if (pad_width || pad_height) {
+            cur     = ggml_pad(ctx0, cur, 0, pad_width, pad_height, 0);
+            width  += pad_width;
+            height += pad_height;
+        }
+
+        // unshuffle h
+        cur = ggml_reshape_3d(ctx0, cur, n_embd * scale_factor, width / scale_factor, height);
+        cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+
+        // unshuffle w
+        cur = ggml_cont_3d(ctx0, cur, n_embd * scale_factor * scale_factor, height / scale_factor, width / scale_factor);
+        cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+
+        cur = ggml_cont_2d(ctx0, cur, cur->ne[0], cur->ne[1] * cur->ne[2]);
+        cb(cur, "pixel_shuffle", -1);
+
+        return cur;
+    }
+
 };
 
 static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch & imgs) {
@@ -1991,6 +2101,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     switch (ctx->proj_type()) {
         case PROJECTOR_TYPE_GEMMA3:
         case PROJECTOR_TYPE_IDEFICS3:
+        case PROJECTOR_TYPE_LFM2:
             {
                 res = graph.build_siglip();
             } break;
@@ -2021,6 +2132,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             {
                 res = graph.build_whisper_enc();
             } break;
+        case PROJECTOR_TYPE_KIMIVL:
+            {
+                res = graph.build_kimivl();
+            } break;
         default:
             {
                 res = graph.build_llava();
@@ -2151,7 +2266,21 @@ struct clip_model_loader {
                 get_u32(KEY_PATCH_SIZE, hparams.patch_size);
                 get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false);
                 get_i32(KEY_MINICPMV_VERSION, hparams.minicpmv_version, false); // legacy
-
+                get_u32(KEY_MINICPMV_QUERY_NUM, hparams.minicpmv_query_num, false);
+                if (hparams.minicpmv_query_num == 0) {
+                    // Fallback to hardcoded values for legacy models
+                    if (hparams.minicpmv_version == 3) {
+                        hparams.minicpmv_query_num = 64;
+                    } else if (hparams.minicpmv_version == 4) {
+                        hparams.minicpmv_query_num = 64;
+                    } else if (hparams.minicpmv_version == 5) {
+                        hparams.minicpmv_query_num = 64;
+                    } else if (hparams.minicpmv_version == 6) {
+                        hparams.minicpmv_query_num = 64;
+                    } else {
+                        hparams.minicpmv_query_num = 96;
+                    }
+                }
             } else if (is_audio) {
                 get_u32(KEY_A_NUM_MEL_BINS, hparams.n_mel_bins);
 
@@ -2243,6 +2372,7 @@ struct clip_model_loader {
                         }
                     } break;
                 case PROJECTOR_TYPE_IDEFICS3:
+                case PROJECTOR_TYPE_LFM2:
                 case PROJECTOR_TYPE_INTERNVL:
                     {
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
@@ -2256,6 +2386,12 @@ struct clip_model_loader {
                         hparams.image_size = 1024;
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.spatial_merge_size, false);
                     } break;
+                case PROJECTOR_TYPE_KIMIVL:
+                    {
+                        hparams.rope_theta = 10000.0f;
+                        hparams.warmup_image_size = hparams.patch_size * 8;
+                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
+                    } break;
                 case PROJECTOR_TYPE_GEMMA3:
                     {
                         // default value (used by all model sizes in gemma 3 family)
@@ -2420,7 +2556,20 @@ struct clip_model_loader {
 
             // some models already exported with legacy (incorrect) naming which is quite messy, let's fix it here
             // note: Qwen model converted from the old surgery script has n_ff = 0, so we cannot use n_ff to check!
-            if (layer.ff_up_w && layer.ff_down_w && layer.ff_down_w->ne[0] == hparams.n_embd) {
+            bool is_ffn_swapped = (
+                    // only old models need this fix
+                    model.proj_type == PROJECTOR_TYPE_MLP
+                    || model.proj_type == PROJECTOR_TYPE_MLP_NORM
+                    || model.proj_type == PROJECTOR_TYPE_LDP
+                    || model.proj_type == PROJECTOR_TYPE_LDPV2
+                    || model.proj_type == PROJECTOR_TYPE_QWEN2VL
+                    || model.proj_type == PROJECTOR_TYPE_QWEN25VL
+                    || model.proj_type == PROJECTOR_TYPE_GLM_EDGE
+                    || model.proj_type == PROJECTOR_TYPE_GEMMA3
+                    || model.proj_type == PROJECTOR_TYPE_IDEFICS3
+                    || model.proj_type == PROJECTOR_TYPE_MINICPMV
+                ) && layer.ff_up_w && layer.ff_down_w && layer.ff_down_w->ne[0] == hparams.n_embd;
+            if (is_ffn_swapped) {
                 // swap up and down weights
                 ggml_tensor * tmp = layer.ff_up_w;
                 layer.ff_up_w = layer.ff_down_w;
@@ -2429,6 +2578,9 @@ struct clip_model_loader {
                 tmp = layer.ff_up_b;
                 layer.ff_up_b = layer.ff_down_b;
                 layer.ff_down_b = tmp;
+                if (il == 0) {
+                    LOG_WRN("%s: ffn up/down are swapped\n", __func__);
+                }
             }
         }
 
@@ -2546,6 +2698,16 @@ struct clip_model_loader {
                 {
                     model.projection = get_tensor(TN_MM_PROJECTOR);
                 } break;
+            case PROJECTOR_TYPE_LFM2:
+            case PROJECTOR_TYPE_KIMIVL:
+                {
+                    model.mm_input_norm_w = get_tensor(TN_MM_INP_NORM);
+                    model.mm_input_norm_b = get_tensor(TN_MM_INP_NORM_B);
+                    model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
+                    model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
+                    model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
+                    model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
+                } break;
             case PROJECTOR_TYPE_PIXTRAL:
                 {
                     model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
@@ -2944,7 +3106,7 @@ struct image_manipulation {
         dst.buf.resize(3 * target_width * target_height);
 
         float Cc;
-        float C[5];
+        float C[5] = {};
         float d0, d2, d3, a0, a1, a2, a3;
         int i, j, k, jj;
         int x, y;
@@ -3467,6 +3629,45 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
         res_imgs->grid_y = inst.grid_size.height;
         return true;
 
+    } else if ( ctx->proj_type() == PROJECTOR_TYPE_LFM2
+             || ctx->proj_type() == PROJECTOR_TYPE_KIMIVL
+    ) {
+        GGML_ASSERT(params.proj_scale_factor);
+
+        // smart resize
+        const int width = img->nx;
+        const int height = img->ny;
+        const int total_factor = params.patch_size * params.proj_scale_factor;
+        constexpr int min_image_tokens = 64;
+        constexpr int max_image_tokens = 1024;
+        const float min_pixels = min_image_tokens * total_factor * total_factor;
+        const float max_pixels = max_image_tokens * total_factor * total_factor;
+
+        auto round_by_factor = [f = total_factor](float x) { return static_cast<int>(std::nearbyintf(x / static_cast<float>(f))) * f; };
+        auto ceil_by_factor  = [f = total_factor](float x) { return static_cast<int>(std::ceil(x / static_cast<float>(f))) * f; };
+        auto floor_by_factor = [f = total_factor](float x) { return static_cast<int>(std::floor(x / static_cast<float>(f))) * f; };
+
+        int h_bar = std::max(total_factor, round_by_factor(height));
+        int w_bar = std::max(total_factor, round_by_factor(width));
+
+        if (h_bar * w_bar > max_pixels) {
+            const auto beta = std::sqrt((height * width) / max_pixels);
+            h_bar = std::max(total_factor, floor_by_factor(height / beta));
+            w_bar = std::max(total_factor, floor_by_factor(width / beta));
+        } else if (h_bar * w_bar < min_pixels) {
+            const auto beta = std::sqrt(min_pixels / (height * width));
+            h_bar = ceil_by_factor(height * beta);
+            w_bar = ceil_by_factor(width * beta);
+        }
+
+        const std::array<uint8_t, 3> pad_color = {122, 116, 104};
+
+        clip_image_u8 resized_img;
+        image_manipulation::resize_and_pad_image(*img, resized_img, clip_image_size{w_bar, h_bar}, pad_color);
+        clip_image_f32_ptr res(clip_image_f32_init());
+        normalize_image_u8_to_f32(resized_img, *res, params.image_mean, params.image_std);
+        res_imgs->entries.push_back(std::move(res));
+        return true;
     }
 
     // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
@@ -3506,10 +3707,10 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
         }
 
         return true;
-
+    } else {
+        GGML_ABORT("Unknown image preprocessing type");
     }
 
-    GGML_ASSERT(false && "Unknown image preprocessing type");
 }
 
 ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
@@ -3573,8 +3774,9 @@ int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 *
 int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
     const auto & params = ctx->model.hparams;
 
-    // only for models using fixed size square images
-    int n_patches_sq = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
+    // for models with fixed size image, the input image is already pre-processed and resized to square
+    int patch_size = params.patch_size;
+    int n_patches = (img->nx / patch_size) * (img->ny / patch_size);
 
     projector_type proj = ctx->proj_type();
 
@@ -3588,89 +3790,97 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_LDPV2:
         case PROJECTOR_TYPE_GLM_EDGE:
             {
-                n_patches_sq /= 4;
+                n_patches /= 4;
                 if (ctx->model.mm_glm_tok_boi) {
-                    n_patches_sq += 2; // for BOI and EOI token embeddings
+                    n_patches += 2; // for BOI and EOI token embeddings
                 }
             } break;
         case PROJECTOR_TYPE_MINICPMV:
             {
+                // Use actual config value if available, otherwise fall back to hardcoded values
+                if (params.minicpmv_query_num > 0) {
+                    n_patches = params.minicpmv_query_num;
+                } else {
+                    // Fallback to hardcoded values for legacy models
                     if (params.minicpmv_version == 2) {
-                    // MiniCPM-V 2.5
-                    n_patches_sq = 96;
+                        n_patches = 96;
                     } else if (params.minicpmv_version == 3) {
-                    // MiniCPM-V 2.6
-                    n_patches_sq = 64;
+                        n_patches = 64;
                     } else if (params.minicpmv_version == 4) {
-                    // MiniCPM-o 2.6
-                    n_patches_sq = 64;
+                        n_patches = 64;
                     } else if (params.minicpmv_version == 5) {
                         // MiniCPM-V 4.0
-                    n_patches_sq = 64;
+                        n_patches = 64;
+                    } else if (params.minicpmv_version == 6) {
+                        // MiniCPM-V 4.5
+                        n_patches = 64;
                     } else {
                         GGML_ABORT("Unknown minicpmv version");
                     }
+                }
             } break;
         case PROJECTOR_TYPE_QWEN2VL:
         case PROJECTOR_TYPE_QWEN25VL:
             {
-                // dynamic size
+                // dynamic size (2 conv, so double patch size)
                 int patch_size = params.patch_size * 2;
                 int x_patch = img->nx / patch_size + (int)(img->nx % patch_size > 0);
                 int y_patch = img->ny / patch_size + (int)(img->ny % patch_size > 0);
-                n_patches_sq = x_patch * y_patch;
+                n_patches = x_patch * y_patch;
             } break;
         case PROJECTOR_TYPE_GEMMA3:
-            {
-                int n_per_side = params.image_size / params.patch_size;
-                int n_per_side_2d_pool = n_per_side / params.proj_scale_factor;
-                n_patches_sq = n_per_side_2d_pool * n_per_side_2d_pool;
-            } break;
         case PROJECTOR_TYPE_IDEFICS3:
         case PROJECTOR_TYPE_INTERNVL:
+        case PROJECTOR_TYPE_LLAMA4:
             {
-                // both W and H are divided by proj_scale_factor
-                n_patches_sq /= (params.proj_scale_factor * params.proj_scale_factor);
+                // both X and Y are downscaled by the scale factor
+                int scale_factor = ctx->model.hparams.proj_scale_factor;
+                n_patches /= (scale_factor * scale_factor);
+            } break;
+        case PROJECTOR_TYPE_LFM2:
+        case PROJECTOR_TYPE_KIMIVL:
+            {
+                // dynamic size
+                int scale_factor = ctx->model.hparams.proj_scale_factor;
+                int out_patch_size = params.patch_size * scale_factor;
+                int x_patch = CLIP_ALIGN(img->nx, out_patch_size) / out_patch_size;
+                int y_patch = CLIP_ALIGN(img->ny, out_patch_size) / out_patch_size;
+                n_patches = x_patch * y_patch;
             } break;
         case PROJECTOR_TYPE_PIXTRAL:
             {
                 // dynamic size
                 int n_merge = params.spatial_merge_size;
-                int n_patches_x = img->nx / params.patch_size / (n_merge > 0 ? n_merge : 1);
-                int n_patches_y = img->ny / params.patch_size / (n_merge > 0 ? n_merge : 1);
-                n_patches_sq = n_patches_y * n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
-            } break;
-        case PROJECTOR_TYPE_LLAMA4:
-            {
-                int scale_factor = ctx->model.hparams.proj_scale_factor;
-                n_patches_sq /= (scale_factor * scale_factor);
+                int n_patches_x = img->nx / patch_size / (n_merge > 0 ? n_merge : 1);
+                int n_patches_y = img->ny / patch_size / (n_merge > 0 ? n_merge : 1);
+                n_patches = n_patches_y * n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
             } break;
         case PROJECTOR_TYPE_VOXTRAL:
         case PROJECTOR_TYPE_ULTRAVOX:
         case PROJECTOR_TYPE_QWEN2A:
             {
-                n_patches_sq = img->nx;
+                n_patches = img->nx;
 
                 const int proj_stack_factor = ctx->model.hparams.proj_stack_factor;
                 if (ctx->model.audio_has_stack_frames()) {
                     GGML_ASSERT(proj_stack_factor > 0);
-                    const int n_len = CLIP_ALIGN(n_patches_sq, proj_stack_factor);
-                    n_patches_sq = n_len / proj_stack_factor;
+                    const int n_len = CLIP_ALIGN(n_patches, proj_stack_factor);
+                    n_patches = n_len / proj_stack_factor;
                 }
 
                 // whisper downscales input token by half after conv1d
-                n_patches_sq /= 2;
+                n_patches /= 2;
 
                 if (ctx->model.audio_has_avgpool()) {
                     // divide by 2 because of nn.AvgPool1d(2, stride=2)
-                    n_patches_sq /= 2;
+                    n_patches /= 2;
                 }
             } break;
         default:
             GGML_ABORT("unsupported projector type");
     }
 
-    return n_patches_sq;
+    return n_patches;
 }
 
 static std::vector<std::vector<std::vector<float>>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector<std::vector<float>> & pos) {
@@ -4019,6 +4229,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                 set_input_i32("positions", positions);
             } break;
         case PROJECTOR_TYPE_PIXTRAL:
+        case PROJECTOR_TYPE_KIMIVL:
             {
                 // set the 2D positions
                 int n_patches_per_col = image_size_width / patch_size;
@@ -4070,6 +4281,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         case PROJECTOR_TYPE_INTERNVL:
         case PROJECTOR_TYPE_QWEN2A:
         case PROJECTOR_TYPE_ULTRAVOX:
+        case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_VOXTRAL:
             {
                 // do nothing
@@ -4141,7 +4353,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
 }
 
 int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
-    const auto & hparams = ctx->model.hparams;
     switch (ctx->model.proj_type) {
         case PROJECTOR_TYPE_LDP:
             return ctx->model.mm_model_block_1_block_2_1_b->ne[0];
@@ -4153,20 +4364,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_MLP_NORM:
             return ctx->model.mm_3_b->ne[0];
         case PROJECTOR_TYPE_MINICPMV:
-            if (hparams.minicpmv_version == 2) {
-                // MiniCPM-V 2.5
-                return 4096;
-            } else if (hparams.minicpmv_version == 3) {
-                // MiniCPM-V 2.6
-                return 3584;
-            } else if (hparams.minicpmv_version == 4) {
-                // MiniCPM-o 2.6
-                return 3584;
-            } else if (hparams.minicpmv_version == 5) {
-                // MiniCPM-V 4.0
-                return 2560;
-            }
-            GGML_ABORT("Unknown minicpmv version");
+            return ctx->model.mm_model_proj->ne[0];
         case PROJECTOR_TYPE_GLM_EDGE:
             return ctx->model.mm_model_mlp_3_w->ne[1];
         case PROJECTOR_TYPE_QWEN2VL:
@@ -4185,6 +4383,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->model.mm_model_proj->ne[1];
         case PROJECTOR_TYPE_QWEN2A:
             return ctx->model.mm_fc_w->ne[1];
+        case PROJECTOR_TYPE_LFM2:
+        case PROJECTOR_TYPE_KIMIVL:
+            return ctx->model.mm_2_w->ne[1];
         default:
             GGML_ABORT("Unknown projector type");
     }

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

@@ -82,11 +82,6 @@ struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch
  */
 void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img);
 
-bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
-
-/** interpret bytes as an image file with length bytes_length, and use the result to populate img */
-bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
-
 /** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
 bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
 

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

@@ -217,7 +217,7 @@ struct mtmd_context {
             tok_row_end_trail = false; // no trailing end-of-row token
             ov_img_first      = true;
 
-        } else if (minicpmv_version == 3 || minicpmv_version == 4 || minicpmv_version == 5) {
+        } else if (minicpmv_version == 3 || minicpmv_version == 4 || minicpmv_version == 5 || minicpmv_version == 6) {
             // minicpmv 2.6 format:
             // <image> (overview) </image><slice> (slice) </slice><slice> (slice) </slice>\n ...
             slice_tmpl        = MTMD_SLICE_TMPL_MINICPMV_2_6;

llama/llama.go

@@ -42,6 +42,7 @@ import (
 	_ "github.com/ollama/ollama/llama/llama.cpp/common"
 	_ "github.com/ollama/ollama/llama/llama.cpp/src"
 	_ "github.com/ollama/ollama/llama/llama.cpp/tools/mtmd"
+	"github.com/ollama/ollama/ml"
 	ggml "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
 )
 
@@ -62,8 +63,8 @@ func BackendInit() {
 	C.llama_backend_init()
 }
 
-func EnumerateGPUs() []string {
-	var ids []string
+func EnumerateGPUs() []ml.DeviceID {
+	var ids []ml.DeviceID
 
 	for i := range C.ggml_backend_dev_count() {
 		device := C.ggml_backend_dev_get(i)
@@ -71,7 +72,10 @@ func EnumerateGPUs() []string {
 		if C.ggml_backend_dev_type(device) == C.GGML_BACKEND_DEVICE_TYPE_GPU {
 			var props C.struct_ggml_backend_dev_props
 			C.ggml_backend_dev_get_props(device, &props)
-			ids = append(ids, C.GoString(props.id))
+			ids = append(ids, ml.DeviceID{
+				ID:      C.GoString(props.id),
+				Library: C.GoString(props.library),
+			})
 		}
 	}
 
@@ -112,7 +116,11 @@ func NewContextParams(numCtx int, batchSize int, numSeqMax int, threads int, fla
 	params.n_threads = C.int(threads)
 	params.n_threads_batch = params.n_threads
 	params.embeddings = C.bool(true)
-	params.flash_attn = C.bool(flashAttention)
+	if flashAttention {
+		params.flash_attn_type = C.LLAMA_FLASH_ATTN_TYPE_ENABLED
+	} else {
+		params.flash_attn_type = C.LLAMA_FLASH_ATTN_TYPE_DISABLED
+	}
 	params.type_k = kvCacheTypeFromStr(strings.ToLower(kvCacheType))
 	params.type_v = kvCacheTypeFromStr(strings.ToLower(kvCacheType))
 

llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch

@@ -15,18 +15,18 @@ problem.
  ggml/src/ggml-backend.cpp            | 9 +++++++--
  ggml/src/ggml-cann/ggml-cann.cpp     | 2 ++
  ggml/src/ggml-cuda/ggml-cuda.cu      | 3 +++
- ggml/src/ggml-metal/ggml-metal.m     | 1 +
+ ggml/src/ggml-metal/ggml-metal.cpp   | 2 ++
  ggml/src/ggml-opencl/ggml-opencl.cpp | 1 +
  ggml/src/ggml-rpc/ggml-rpc.cpp       | 1 +
  ggml/src/ggml-sycl/ggml-sycl.cpp     | 3 +++
  ggml/src/ggml-vulkan/ggml-vulkan.cpp | 2 ++
- 8 files changed, 20 insertions(+), 2 deletions(-)
+ 8 files changed, 21 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index 1b9d29e9..97f47abd 100644
+index ff9135fe..8ba86f82 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) {
+@@ -113,7 +113,6 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
      if (buffer->iface.free_buffer != NULL) {
          buffer->iface.free_buffer(buffer);
      }
@@ -34,7 +34,7 @@ index 1b9d29e9..97f47abd 100644
  }
  
  size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
-@@ -529,6 +528,7 @@ static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer)
+@@ -586,6 +585,7 @@ static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer)
  
      free(ctx->buffers);
      free(ctx);
@@ -42,9 +42,9 @@ index 1b9d29e9..97f47abd 100644
  }
  
  static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-@@ -1890,6 +1890,11 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
- 
+@@ -2075,6 +2075,11 @@ 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_ASSERT(buffer);
      ggml_aligned_free(buffer->context, buffer->size);
 +    delete buffer;
 +}
@@ -54,7 +54,7 @@ index 1b9d29e9..97f47abd 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) {
-@@ -1937,7 +1942,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
+@@ -2127,7 +2132,7 @@ 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 = {
@@ -64,10 +64,10 @@ index 1b9d29e9..97f47abd 100644
      /* .init_tensor     = */ NULL, // no initialization required
      /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
 diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
-index cf575b36..ca1addfa 100755
+index b51b554e..3ba0f5a6 100755
 --- a/ggml/src/ggml-cann/ggml-cann.cpp
 +++ b/ggml/src/ggml-cann/ggml-cann.cpp
-@@ -826,6 +826,7 @@ static void ggml_backend_cann_buffer_free_buffer(
+@@ -843,6 +843,7 @@ static void ggml_backend_cann_buffer_free_buffer(
      ggml_backend_cann_buffer_context* ctx =
          (ggml_backend_cann_buffer_context*)buffer->context;
      delete ctx;
@@ -75,7 +75,7 @@ index cf575b36..ca1addfa 100755
  }
  
  /**
-@@ -1572,6 +1573,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
+@@ -1630,6 +1631,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
   */
  static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
      ACL_CHECK(aclrtFreeHost(buffer->context));
@@ -84,7 +84,7 @@ index cf575b36..ca1addfa 100755
  
  /**
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index d9110491..37ee2a6d 100644
+index b7e81b21..fdf8c63d 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -567,6 +567,7 @@ struct ggml_backend_cuda_buffer_context {
@@ -111,23 +111,31 @@ index d9110491..37ee2a6d 100644
  }
  
  static void * ggml_cuda_host_malloc(size_t size) {
-diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
-index cb8eff4a..7bccc7bf 100644
---- a/ggml/src/ggml-metal/ggml-metal.m
-+++ b/ggml/src/ggml-metal/ggml-metal.m
-@@ -6032,6 +6032,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
+diff --git a/ggml/src/ggml-metal/ggml-metal.cpp b/ggml/src/ggml-metal/ggml-metal.cpp
+index e11555a7..909e17de 100644
+--- a/ggml/src/ggml-metal/ggml-metal.cpp
++++ b/ggml/src/ggml-metal/ggml-metal.cpp
+@@ -25,6 +25,7 @@ static void ggml_backend_metal_buffer_shared_free_buffer(ggml_backend_buffer_t b
+     GGML_ASSERT(ggml_metal_buffer_is_shared(ctx));
+ 
+     ggml_metal_buffer_free(ctx);
++    delete buffer;
  }
  
-     free(ctx);
-+    free(buffer);
+ static void * ggml_backend_metal_buffer_shared_get_base(ggml_backend_buffer_t buffer) {
+@@ -99,6 +100,7 @@ static void ggml_backend_metal_buffer_private_free_buffer(ggml_backend_buffer_t
+     GGML_ASSERT(!ggml_metal_buffer_is_shared(ctx));
+ 
+     ggml_metal_buffer_free(ctx);
++    delete buffer;
  }
  
- static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
+ static void * ggml_backend_metal_buffer_private_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
-index 8ba1e00d..8163e8dc 100644
+index 0cf3b924..09d706b5 100644
 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp
 +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
-@@ -2745,6 +2745,7 @@ struct ggml_backend_opencl_buffer_context {
+@@ -3215,6 +3215,7 @@ struct ggml_backend_opencl_buffer_context {
  static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
      delete ctx;
@@ -136,10 +144,10 @@ index 8ba1e00d..8163e8dc 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 df6ba540..2e395968 100644
+index f99681c8..59591770 100644
 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp
 +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
-@@ -486,6 +486,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -505,6 +505,7 @@ 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);
      RPC_STATUS_ASSERT(status);
      delete ctx;
@@ -148,7 +156,7 @@ index df6ba540..2e395968 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 3992dad0..67503951 100644
+index 4ac919ea..447ea3c4 100644
 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp
 +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
 @@ -331,6 +331,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
@@ -176,10 +184,10 @@ index 3992dad0..67503951 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 4070e248..394a2839 100644
+index 2608cbd0..061cd078 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -10209,6 +10209,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -11603,6 +11603,7 @@ 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;
@@ -187,7 +195,7 @@ index 4070e248..394a2839 100644
  }
  
  static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -10352,6 +10353,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
+@@ -11746,6 +11747,7 @@ 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);

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 f7e03e70..8ebe11cf 100644
+index da938af0..2a38abf4 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
-@@ -1804,16 +1804,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+@@ -1811,16 +1811,7 @@ 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 f7e03e70..8ebe11cf 100644
                  pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -1975,7 +1966,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
-                 pre_type = LLAMA_VOCAB_PRE_TYPE_KIMI_K2;
+@@ -1987,7 +1978,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+                 pre_type = LLAMA_VOCAB_PRE_TYPE_GROK_2;
                  clean_spaces = false;
              } else {
 -                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));

llama/patches/0003-clip-unicode.patch

@@ -10,7 +10,7 @@ filesystems for paths that include wide characters
  1 file changed, 39 insertions(+)
 
 diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
-index 20c21733..f4f69cfc 100644
+index 210ecc88..355219a9 100644
 --- a/tools/mtmd/clip.cpp
 +++ b/tools/mtmd/clip.cpp
 @@ -28,6 +28,19 @@
@@ -33,7 +33,7 @@ index 20c21733..f4f69cfc 100644
  struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL};
  
  enum ffn_op_type {
-@@ -2597,7 +2610,29 @@ struct clip_model_loader {
+@@ -2759,7 +2772,29 @@ struct clip_model_loader {
          {
              std::vector<uint8_t> read_buf;
  
@@ -63,7 +63,7 @@ index 20c21733..f4f69cfc 100644
              if (!fin) {
                  throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
              }
-@@ -2624,7 +2659,11 @@ struct clip_model_loader {
+@@ -2786,7 +2821,11 @@ struct clip_model_loader {
                      ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
                  }
              }

llama/patches/0004-solar-pro.patch

@@ -15,10 +15,10 @@ adds support for the Solar Pro architecture
  7 files changed, 248 insertions(+)
 
 diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
-index 18dcc6dd..4b285646 100644
+index 4e8d54c4..f98a3574 100644
 --- a/src/llama-arch.cpp
 +++ b/src/llama-arch.cpp
-@@ -78,6 +78,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+@@ -81,6 +81,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
      { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
      { LLM_ARCH_GRANITE_HYBRID,   "granitehybrid"    },
      { LLM_ARCH_CHAMELEON,        "chameleon"        },
@@ -26,15 +26,15 @@ index 18dcc6dd..4b285646 100644
      { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
      { LLM_ARCH_PLM,              "plm"              },
      { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
-@@ -164,6 +165,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
-     { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
-     { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
+@@ -177,6 +178,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
      { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
+     { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
+     { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
 +    { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
      { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
      { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
  
-@@ -1794,6 +1796,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+@@ -1879,6 +1881,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
              { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
          },
      },
@@ -59,7 +59,7 @@ index 18dcc6dd..4b285646 100644
      {
          LLM_ARCH_WAVTOKENIZER_DEC,
          {
-@@ -2219,6 +2239,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
+@@ -2368,6 +2388,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
      {LLM_TENSOR_LAUREL_POST_NORM,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      // this tensor is loaded for T5, but never used
      {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
@@ -68,10 +68,10 @@ index 18dcc6dd..4b285646 100644
      {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 diff --git a/src/llama-arch.h b/src/llama-arch.h
-index 7af587e7..3ea994c7 100644
+index b5c6f3d7..aa8e0e7b 100644
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
-@@ -82,6 +82,7 @@ enum llm_arch {
+@@ -85,6 +85,7 @@ enum llm_arch {
      LLM_ARCH_GRANITE_MOE,
      LLM_ARCH_GRANITE_HYBRID,
      LLM_ARCH_CHAMELEON,
@@ -79,15 +79,15 @@ index 7af587e7..3ea994c7 100644
      LLM_ARCH_WAVTOKENIZER_DEC,
      LLM_ARCH_PLM,
      LLM_ARCH_BAILINGMOE,
-@@ -168,6 +169,7 @@ enum llm_kv {
-     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
-     LLM_KV_ATTENTION_SLIDING_WINDOW,
+@@ -181,6 +182,7 @@ enum llm_kv {
      LLM_KV_ATTENTION_SCALE,
+     LLM_KV_ATTENTION_OUTPUT_SCALE,
+     LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
 +    LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
      LLM_KV_ATTENTION_KEY_LENGTH_MLA,
      LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
  
-@@ -394,6 +396,7 @@ enum llm_tensor {
+@@ -417,6 +419,7 @@ enum llm_tensor {
      LLM_TENSOR_ENC_OUTPUT_NORM,
      LLM_TENSOR_CLS,
      LLM_TENSOR_CLS_OUT,
@@ -96,10 +96,10 @@ index 7af587e7..3ea994c7 100644
      LLM_TENSOR_CONVNEXT_DW,
      LLM_TENSOR_CONVNEXT_NORM,
 diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
-index 7a06368d..35fc054f 100644
+index c04ac58f..24a515a0 100644
 --- a/src/llama-hparams.cpp
 +++ b/src/llama-hparams.cpp
-@@ -146,6 +146,14 @@ uint32_t llama_hparams::n_pos_per_embd() const {
+@@ -147,6 +147,14 @@ uint32_t llama_hparams::n_pos_per_embd() const {
      return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
  }
  
@@ -115,10 +115,10 @@ index 7a06368d..35fc054f 100644
      if (il < n_layer) {
          return swa_layers[il];
 diff --git a/src/llama-hparams.h b/src/llama-hparams.h
-index bd231224..29bd9056 100644
+index 0fe4b569..eb13709f 100644
 --- a/src/llama-hparams.h
 +++ b/src/llama-hparams.h
-@@ -62,6 +62,8 @@ struct llama_hparams {
+@@ -64,6 +64,8 @@ struct llama_hparams {
      std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
      std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
  
@@ -127,7 +127,7 @@ index bd231224..29bd9056 100644
      uint32_t n_layer_dense_lead = 0;
      uint32_t n_lora_q           = 0;
      uint32_t n_lora_kv          = 0;
-@@ -220,6 +222,9 @@ struct llama_hparams {
+@@ -236,6 +238,9 @@ struct llama_hparams {
  
      uint32_t n_pos_per_embd() const;
  
@@ -135,10 +135,10 @@ index bd231224..29bd9056 100644
 +    bool n_bskcn(uint32_t n, uint32_t il) const;
 +
      bool is_swa(uint32_t il) const;
- };
  
+     bool has_kv(uint32_t il) const;
 diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
-index f71c40f8..7eab9b68 100644
+index 8182a9ad..daef900c 100644
 --- a/src/llama-model-loader.cpp
 +++ b/src/llama-model-loader.cpp
 @@ -465,6 +465,7 @@ namespace GGUFMeta {
@@ -150,10 +150,10 @@ index f71c40f8..7eab9b68 100644
  llama_model_loader::llama_model_loader(
          const std::string & fname,
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index 58ca7df7..280129e1 100644
+index 2470f878..0398b553 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1706,6 +1706,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -1845,6 +1845,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                      default: type = LLM_TYPE_UNKNOWN;
                 }
              } break;
@@ -175,7 +175,7 @@ index 58ca7df7..280129e1 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
-@@ -4793,6 +4808,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -5113,6 +5128,34 @@ 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 58ca7df7..280129e1 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);
-@@ -15495,6 +15538,165 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
+@@ -16273,6 +16316,165 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
      }
  };
  
@@ -229,7 +229,7 @@ index 58ca7df7..280129e1 100644
 +        struct ggml_tensor * inp_pos = build_inp_pos();
 +
 +        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-+        auto * inp_attn = build_attn_inp_kv_unified();
++        auto * inp_attn = build_attn_inp_kv();
 +
 +        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 +
@@ -316,7 +316,7 @@ index 58ca7df7..280129e1 100644
 +
 +                cur = build_attn(inp_attn,
 +                        model.layers[il].wo, model.layers[il].bo,
-+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
++                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
 +                cb(cur, "attn_out", il);
 +            }
 +
@@ -376,7 +376,7 @@ index 58ca7df7..280129e1 100644
  // ref: https://github.com/facebookresearch/chameleon
  // based on the original build_llama() function, changes:
  //   * qk-norm
-@@ -18439,6 +18641,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
+@@ -19552,6 +19754,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
              {
                  llm = std::make_unique<llm_build_chameleon>(*this, params);
              } break;
@@ -387,7 +387,7 @@ index 58ca7df7..280129e1 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params);
-@@ -18652,6 +18858,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+@@ -19770,6 +19976,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
          case LLM_ARCH_GRANITE_MOE:
          case LLM_ARCH_GRANITE_HYBRID:
          case LLM_ARCH_CHAMELEON:
@@ -396,10 +396,10 @@ index 58ca7df7..280129e1 100644
          case LLM_ARCH_NEO_BERT:
          case LLM_ARCH_SMOLLM3:
 diff --git a/src/llama-model.h b/src/llama-model.h
-index 6fcd74d5..09964533 100644
+index d73ce969..c086f94e 100644
 --- a/src/llama-model.h
 +++ b/src/llama-model.h
-@@ -70,6 +70,7 @@ enum llm_type {
+@@ -76,6 +76,7 @@ enum llm_type {
      LLM_TYPE_15B,
      LLM_TYPE_16B,
      LLM_TYPE_20B,
@@ -407,7 +407,7 @@ index 6fcd74d5..09964533 100644
      LLM_TYPE_27B,
      LLM_TYPE_30B,
      LLM_TYPE_32B,
-@@ -367,6 +368,8 @@ struct llama_layer {
+@@ -380,6 +381,8 @@ struct llama_layer {
      // openai-moe
      struct ggml_tensor * attn_sinks = nullptr;
  

llama/patches/0005-fix-deepseek-deseret-regex.patch

@@ -12,7 +12,7 @@ regex
  2 files changed, 22 insertions(+), 1 deletion(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 8ebe11cf..c011008f 100644
+index 2a38abf4..26fa9fad 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -299,7 +299,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {

llama/patches/0006-maintain-ordering-for-rules-for-grammar.patch

@@ -8,10 +8,10 @@ Subject: [PATCH] maintain ordering for rules for grammar
  1 file changed, 1 insertion(+), 1 deletion(-)
 
 diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp
-index 637891f5..98b8280f 100644
+index db1f0b23..f4de7e34 100644
 --- a/common/json-schema-to-grammar.cpp
 +++ b/common/json-schema-to-grammar.cpp
-@@ -307,7 +307,7 @@ private:
+@@ -308,7 +308,7 @@ private:
      friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
      std::function<json(const std::string &)> _fetch_json;
      bool _dotall;

llama/patches/0007-sort-devices-by-score.patch

@@ -11,10 +11,10 @@ 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 6c315137..3040b2aa 100644
+index 136afec7..f794d9cf 100644
 --- a/ggml/src/ggml-backend-reg.cpp
 +++ b/ggml/src/ggml-backend-reg.cpp
-@@ -162,7 +162,7 @@ struct ggml_backend_reg_entry {
+@@ -175,7 +175,7 @@ struct ggml_backend_reg_entry {
  
  struct ggml_backend_registry {
      std::vector<ggml_backend_reg_entry> backends;
@@ -23,7 +23,7 @@ index 6c315137..3040b2aa 100644
  
      ggml_backend_registry() {
  #ifdef GGML_USE_CUDA
-@@ -207,7 +207,7 @@ struct ggml_backend_registry {
+@@ -223,7 +223,7 @@ struct ggml_backend_registry {
          }
      }
  
@@ -32,7 +32,7 @@ index 6c315137..3040b2aa 100644
          if (!reg) {
              return;
          }
-@@ -218,15 +218,20 @@ struct ggml_backend_registry {
+@@ -234,15 +234,20 @@ struct ggml_backend_registry {
  #endif
          backends.push_back({ reg, std::move(handle) });
          for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
@@ -56,7 +56,7 @@ index 6c315137..3040b2aa 100644
      }
  
      ggml_backend_reg_t load_backend(const fs::path & path, bool silent) {
-@@ -270,7 +275,7 @@ struct ggml_backend_registry {
+@@ -286,7 +291,7 @@ struct ggml_backend_registry {
  
          GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path_str(path).c_str());
  
@@ -65,7 +65,7 @@ index 6c315137..3040b2aa 100644
  
          return reg;
      }
-@@ -293,7 +298,7 @@ struct ggml_backend_registry {
+@@ -309,7 +314,7 @@ struct ggml_backend_registry {
          // remove devices
          devices.erase(
              std::remove_if(devices.begin(), devices.end(),
@@ -74,7 +74,7 @@ index 6c315137..3040b2aa 100644
              devices.end());
  
          // remove backend
-@@ -351,7 +356,7 @@ size_t ggml_backend_dev_count() {
+@@ -367,7 +372,7 @@ size_t ggml_backend_dev_count() {
  
  ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
      GGML_ASSERT(index < ggml_backend_dev_count());

llama/patches/0008-add-phony-target-ggml-cpu-for-all-cpu-variants.patch

@@ -8,10 +8,10 @@ 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 177fb282..f5a5079a 100644
+index c8f3d859..ff6229a0 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
-@@ -304,6 +304,7 @@ function(ggml_add_cpu_backend_variant tag_name)
+@@ -307,6 +307,7 @@ function(ggml_add_cpu_backend_variant tag_name)
      endif()
  
      ggml_add_cpu_backend_variant_impl(${tag_name})
@@ -19,7 +19,7 @@ index 177fb282..f5a5079a 100644
  endfunction()
  
  ggml_add_backend(CPU)
-@@ -314,6 +315,7 @@ if (GGML_CPU_ALL_VARIANTS)
+@@ -317,6 +318,7 @@ if (GGML_CPU_ALL_VARIANTS)
      elseif (GGML_CPU_ARM_ARCH)
          message(FATAL_ERROR "Cannot use both GGML_CPU_ARM_ARCH and GGML_CPU_ALL_VARIANTS")
      endif()

llama/patches/0009-remove-amx.patch

@@ -9,10 +9,10 @@ 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 f5a5079a..5158acd6 100644
+index ff6229a0..33b3a15f 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
-@@ -324,10 +324,6 @@ if (GGML_CPU_ALL_VARIANTS)
+@@ -327,10 +327,6 @@ if (GGML_CPU_ALL_VARIANTS)
          ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
          ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
          ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)

llama/patches/0010-fix-string-arr-kv-loading.patch

@@ -25,7 +25,7 @@ index 79ee2020..3efb22f0 100644
      // get ith C string from array with given key_id
      GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int64_t key_id, size_t i);
 diff --git a/ggml/src/gguf.cpp b/ggml/src/gguf.cpp
-index 53504399..0f71d5f3 100644
+index 8cc4ef1c..d950dbdf 100644
 --- a/ggml/src/gguf.cpp
 +++ b/ggml/src/gguf.cpp
 @@ -805,10 +805,14 @@ enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int64_t key_id
@@ -53,10 +53,10 @@ index 53504399..0f71d5f3 100644
  }
  
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index c011008f..fa388b03 100644
+index 26fa9fad..64c78a16 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
-@@ -1760,9 +1760,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+@@ -1767,9 +1767,7 @@ 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);
@@ -66,4 +66,4 @@ index c011008f..fa388b03 100644
 +                const size_t n_precompiled_charsmap = gguf_get_arr_data_n(ctx, precompiled_charsmap_keyidx);
                  const char * pc = (const char *) gguf_get_arr_data(ctx, precompiled_charsmap_keyidx);
                  precompiled_charsmap.assign(pc, pc + n_precompiled_charsmap);
- #ifdef IS_BIG_ENDIAN
+ #if defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__

llama/patches/0011-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 d89cd8f4..a5689c18 100644
+index dbc07301..f8574d01 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 d89cd8f4..a5689c18 100644
  #if defined(_MSC_VER) || defined(__MINGW32__)
  #include <malloc.h> // using malloc.h with MSC/MINGW
  #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
-@@ -2858,6 +2860,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
+@@ -2881,6 +2883,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
  
          ggml_compute_forward(&params, node);
  

llama/patches/0012-add-ollama-vocab-for-grammar-support.patch

@@ -184,10 +184,10 @@ index f8c291de..2a3a62db 100644
                        const char * grammar_root,
                                bool lazy,
 diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp
-index bfbf5fa2..11f93f42 100644
+index 2186f827..8fb86009 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) {
+@@ -1563,7 +1563,7 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
          trigger_patterns_c.push_back(trigger_pattern.pattern.c_str());
      }
  
@@ -196,7 +196,7 @@ index bfbf5fa2..11f93f42 100644
                                                   ctx->grammar->lazy, trigger_patterns_c.data(), trigger_patterns_c.size(),
                                                   ctx->grammar->trigger_tokens.data(), ctx->grammar->trigger_tokens.size());
  
-@@ -1548,7 +1548,7 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
+@@ -1645,7 +1645,7 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
              /* .vocab        = */ vocab,
              /* .grammar_str  = */ grammar_str,
              /* .grammar_root = */ grammar_root,

llama/patches/0013-add-argsort-and-cuda-copy-for-i32.patch

@@ -4,17 +4,18 @@ 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-cpu/ops.cpp            |  43 +++++++++++
+ ggml/src/ggml-cuda/argsort.cu        | 102 ++++++++++++++++++++++++++-
  ggml/src/ggml-cuda/cpy-utils.cuh     |   6 ++
- ggml/src/ggml-cuda/cpy.cu        |  43 +++++++++++++
- 4 files changed, 192 insertions(+), 2 deletions(-)
+ ggml/src/ggml-cuda/cpy.cu            |  43 +++++++++++
+ ggml/src/ggml-metal/ggml-metal.metal |  64 +++++++++++++++++
+ 5 files changed, 256 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
-index 854f1c2b..a2924757 100644
+index 14f7dcf4..f7f8da35 100644
 --- a/ggml/src/ggml-cpu/ops.cpp
 +++ b/ggml/src/ggml-cpu/ops.cpp
-@@ -8146,6 +8146,45 @@ static void ggml_compute_forward_argsort_f32(
+@@ -7893,6 +7893,45 @@ static void ggml_compute_forward_argsort_f32(
      }
  }
  
@@ -60,7 +61,7 @@ index 854f1c2b..a2924757 100644
  void ggml_compute_forward_argsort(
      const ggml_compute_params * params,
      ggml_tensor * dst) {
-@@ -8157,6 +8196,10 @@ void ggml_compute_forward_argsort(
+@@ -7904,6 +7943,10 @@ void ggml_compute_forward_argsort(
              {
                  ggml_compute_forward_argsort_f32(params, dst);
              } break;
@@ -196,12 +197,12 @@ index 607ded85..53b02634 100644
 +    }
  }
 diff --git a/ggml/src/ggml-cuda/cpy-utils.cuh b/ggml/src/ggml-cuda/cpy-utils.cuh
-index 410c12b7..b8e9e107 100644
+index e621cb98..597c0c8b 100644
 --- a/ggml/src/ggml-cuda/cpy-utils.cuh
 +++ b/ggml/src/ggml-cuda/cpy-utils.cuh
-@@ -223,3 +223,9 @@ template<typename src_t, typename dst_t>
+@@ -215,3 +215,9 @@ template<typename src_t, typename dst_t>
  static __device__ void cpy_1_flt(const char * cxi, char * cdsti) {
-     convert_flt((const src_t *)cxi, (dst_t *)cdsti);
+     *(dst_t *) cdsti = ggml_cuda_cast<dst_t>(*(const src_t *) cxi);
  }
 +
 +static __device__ void cpy_1_i32_i32(const char * cxi, char * cdsti) {
@@ -210,10 +211,10 @@ index 410c12b7..b8e9e107 100644
 +    *dst = *src;
 +}
 diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu
-index f9bb0256..9c3774e5 100644
+index 746f4396..911220e9 100644
 --- a/ggml/src/ggml-cuda/cpy.cu
 +++ b/ggml/src/ggml-cuda/cpy.cu
-@@ -278,6 +278,47 @@ static void ggml_cpy_f32_iq4_nl_cuda(
+@@ -277,6 +277,47 @@ static void ggml_cpy_f32_iq4_nl_cuda(
          (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
  }
  
@@ -261,7 +262,7 @@ index f9bb0256..9c3774e5 100644
  void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
      const int64_t ne = ggml_nelements(src0);
      GGML_ASSERT(ne == ggml_nelements(src1));
-@@ -369,6 +410,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
+@@ -372,6 +413,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
          ggml_cpy_flt_cuda<half, nv_bfloat16> (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_flt_cuda<half, float> (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);
@@ -270,3 +271,80 @@ index f9bb0256..9c3774e5 100644
      } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
          ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (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_BF16 && src1->type == GGML_TYPE_F16) {
+diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
+index 96df6f0c..44dc31c0 100644
+--- a/ggml/src/ggml-metal/ggml-metal.metal
++++ b/ggml/src/ggml-metal/ggml-metal.metal
+@@ -4428,8 +4428,72 @@ kernel void kernel_argsort_f32_i32(
+     }
+ }
+ 
++typedef void (i32_argsort_t)(
++        constant   ggml_metal_kargs_argsort & args,
++        device  const int32_t * x,
++        device        int32_t * dst,
++        threadgroup   int32_t * shared_values [[threadgroup(0)]],
++        uint3 tgpig[[threadgroup_position_in_grid]],
++        uint3 tpitg[[thread_position_in_threadgroup]]);
++
++template<ggml_sort_order order>
++kernel void kernel_argsort_i32_i32(
++        constant   ggml_metal_kargs_argsort & args,
++        device const int32_t * x,
++        device       int32_t * dst,
++        threadgroup int32_t  * shared_values [[threadgroup(0)]],
++        uint3 tgpig[[threadgroup_position_in_grid]],
++        uint3 tpitg[[thread_position_in_threadgroup]]) {
++    // bitonic sort
++    int col = tpitg[0];
++    int row = tgpig[1];
++
++    if (col >= args.ncols_pad) return;
++
++    device const int32_t * x_row   = x + row * args.ncols;
++    threadgroup int32_t  * dst_row = shared_values;
++
++    // initialize indices
++    dst_row[col] = col;
++
++    threadgroup_barrier(mem_flags::mem_threadgroup);
++
++    for (int k = 2; k <= args.ncols_pad; k *= 2) {
++        for (int j = k / 2; j > 0; j /= 2) {
++            int ixj = col ^ j;
++            if (ixj > col) {
++                if ((col & k) == 0) {
++                    if (dst_row[col] >= args.ncols ||
++                        (dst_row[ixj] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
++                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
++                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
++                    ) {
++                        SWAP(dst_row[col], dst_row[ixj]);
++                    }
++                } else {
++                    if (dst_row[ixj] >= args.ncols ||
++                        (dst_row[col] < args.ncols && (order == GGML_SORT_ORDER_ASC ?
++                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
++                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
++                    ) {
++                        SWAP(dst_row[col], dst_row[ixj]);
++                    }
++                }
++            }
++            threadgroup_barrier(mem_flags::mem_threadgroup);
++        }
++    }
++
++    // copy the result to dst without the padding
++    if (col < args.ncols) {
++        dst[row * args.ncols + col] = dst_row[col];
++    }
++}
++
+ template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
+ template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
++template [[host_name("kernel_argsort_i32_i32_asc")]] kernel i32_argsort_t kernel_argsort_i32_i32<GGML_SORT_ORDER_ASC>;
++template [[host_name("kernel_argsort_i32_i32_desc")]] kernel i32_argsort_t kernel_argsort_i32_i32<GGML_SORT_ORDER_DESC>;
+ 
+ kernel void kernel_leaky_relu_f32(
+         constant     ggml_metal_kargs_leaky_relu & args,

llama/patches/0014-graph-memory-reporting-on-failure.patch

@@ -4,60 +4,50 @@ Date: Fri, 18 Apr 2025 15:58:19 -0700
 Subject: [PATCH] graph memory reporting on failure
 
 ---
- ggml/include/ggml-alloc.h   |  6 ++++++
- ggml/include/ggml-backend.h |  6 ++++++
- ggml/src/ggml-alloc.c       | 38 +++++++++++++++++++++++++++++++++----
- ggml/src/ggml-backend.cpp   | 10 ++++++++++
- 4 files changed, 56 insertions(+), 4 deletions(-)
+ ggml/include/ggml-alloc.h   |  1 +
+ ggml/include/ggml-backend.h |  1 +
+ ggml/src/ggml-alloc.c       | 34 +++++++++++++++++++++++++++++++---
+ ggml/src/ggml-backend.cpp   |  7 +++++++
+ 4 files changed, 40 insertions(+), 3 deletions(-)
 
 diff --git a/ggml/include/ggml-alloc.h b/ggml/include/ggml-alloc.h
-index 2cb150fd..781b1e10 100644
+index 2cb150fd..7ab3f019 100644
 --- a/ggml/include/ggml-alloc.h
 +++ b/ggml/include/ggml-alloc.h
-@@ -66,6 +66,12 @@ GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph
+@@ -65,6 +65,7 @@ GGML_API bool ggml_gallocr_reserve_n(
+ GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
  
  GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
++GGML_API size_t ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, int buffer_id);
  
-+struct ggml_allocr_buffer_status {
-+    size_t size;
-+    bool allocated;
-+};
-+GGML_API struct ggml_allocr_buffer_status ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, int buffer_id);
-+
  // Utils
  // Create a buffer and allocate all the tensors in a ggml_context
- GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
 diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
-index a2977ea2..8a91b381 100644
+index 62b6d65e..fe20dca3 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
-@@ -304,6 +304,12 @@ extern "C" {
+@@ -316,6 +316,7 @@ extern "C" {
  
+     GGML_API ggml_backend_buffer_type_t ggml_backend_sched_get_buffer_type(ggml_backend_sched_t sched, ggml_backend_t backend);
      GGML_API size_t                     ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
++    GGML_API size_t                     ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
  
-+    struct ggml_backend_buffer_status {
-+        size_t size;
-+        bool allocated;
-+    };
-+    GGML_API struct ggml_backend_buffer_status ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
-+
      GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
      GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
- 
 diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
-index 8b6e6028..41c8c4a2 100644
+index fa46f3b4..421ff7c7 100644
 --- a/ggml/src/ggml-alloc.c
 +++ b/ggml/src/ggml-alloc.c
-@@ -350,6 +350,7 @@ struct node_alloc {
+@@ -492,6 +492,7 @@ struct node_alloc {
  struct ggml_gallocr {
      ggml_backend_buffer_type_t * bufts; // [n_buffers]
-     ggml_backend_buffer_t * buffers; // [n_buffers]
+     struct vbuffer ** buffers; // [n_buffers]
 +    size_t *buffer_sizes; // [n_buffers]
      struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
      int n_buffers;
  
-@@ -373,6 +374,9 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
-     galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t));
+@@ -515,6 +516,9 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
+     galloc->buffers = calloc(n_bufs, sizeof(struct vbuffer *));
      GGML_ASSERT(galloc->buffers != NULL);
  
 +    galloc->buffer_sizes = calloc(n_bufs, sizeof(size_t));
@@ -66,7 +56,7 @@ index 8b6e6028..41c8c4a2 100644
      galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
      GGML_ASSERT(galloc->buf_tallocs != NULL);
  
-@@ -439,6 +443,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
+@@ -582,6 +586,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
      ggml_hash_set_free(&galloc->hash_set);
      free(galloc->hash_values);
      free(galloc->bufts);
@@ -74,7 +64,7 @@ index 8b6e6028..41c8c4a2 100644
      free(galloc->buffers);
      free(galloc->buf_tallocs);
      free(galloc->node_allocs);
-@@ -734,6 +739,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
+@@ -875,6 +880,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
          }
      }
  
@@ -83,23 +73,21 @@ index 8b6e6028..41c8c4a2 100644
      // reallocate buffers if needed
      for (int i = 0; i < galloc->n_buffers; i++) {
          // if the buffer type is used multiple times, we reuse the same buffer
-@@ -755,15 +762,20 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
+@@ -896,14 +903,19 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
  
-             ggml_backend_buffer_free(galloc->buffers[i]);
-             galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
+             ggml_vbuffer_free(galloc->buffers[i]);
+             galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
 -            if (galloc->buffers[i] == NULL) {
 +            if (galloc->buffers[i]) {
-+                galloc->buffer_sizes[i] = ggml_backend_buffer_get_size(galloc->buffers[i]);
-+                ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
++                galloc->buffer_sizes[i] = ggml_vbuffer_size(galloc->buffers[i]);
 +            } else {
                  GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
 -                return false;
 +                galloc->buffer_sizes[i] = new_size;
 +                success = false;
              }
--            ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
 +        } else {
-+            galloc->buffer_sizes[i] = ggml_backend_buffer_get_size(galloc->buffers[i]);
++            galloc->buffer_sizes[i] = ggml_vbuffer_size(galloc->buffers[i]);
          }
      }
  
@@ -108,11 +96,11 @@ index 8b6e6028..41c8c4a2 100644
  }
  
  bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
-@@ -920,6 +932,24 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
-     return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
+@@ -1058,6 +1070,22 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
+     return ggml_vbuffer_size(galloc->buffers[buffer_id]);
  }
  
-+struct ggml_allocr_buffer_status ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
++size_t ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
 +    GGML_ASSERT(buffer_id >= 0 && buffer_id < galloc->n_buffers);
 +
 +    for (int i = 0; i < buffer_id; i++) {
@@ -121,36 +109,31 @@ index 8b6e6028..41c8c4a2 100644
 +            // (See above.) However, we need a different check because multiple buffers might be NULL in our
 +            // case and we still want to know the attempted size.
 +
-+            struct ggml_allocr_buffer_status status = {0, true};
-+            return status;
++            return 0;
 +        }
 +    }
 +
-+    struct ggml_allocr_buffer_status status = {galloc->buffer_sizes[buffer_id], galloc->buffers[buffer_id] != NULL};
-+    return status;
++    return galloc->buffer_sizes[buffer_id];
 +}
 +
  // utils
  
  static void free_buffers(ggml_backend_buffer_t ** buffers, const size_t * n_buffers) {
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index 97f47abd..eded0291 100644
+index 8ba86f82..cb2b9956 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
-@@ -1631,6 +1631,16 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
+@@ -1809,6 +1809,13 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
      return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
  }
  
-+struct ggml_backend_buffer_status ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
++size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
 +    int backend_index = ggml_backend_sched_backend_id(sched, backend);
 +    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
 +
-+    struct ggml_allocr_buffer_status allocr_status = ggml_gallocr_get_attempted_buffer_size(sched->galloc, backend_index);
-+    struct ggml_backend_buffer_status status = {allocr_status.size, allocr_status.allocated};
-+
-+    return status;
++    return ggml_gallocr_get_attempted_buffer_size(sched->galloc, backend_index);
 +}
 +
  void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
+     GGML_ASSERT(sched);
      int backend_index = ggml_backend_sched_backend_id(sched, backend);
-     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);

llama/patches/0015-ggml-Export-GPU-UUIDs.patch

@@ -7,27 +7,27 @@ This enables matching up devices and information reported by the backend
 with tools (e.g. nvidia-smi) and system management libraries (e.g. nvml).
 ---
  ggml/include/ggml-backend.h        |  1 +
- ggml/src/ggml-cuda/ggml-cuda.cu  | 67 +++++++++++++++++++++++++++++---
- ggml/src/ggml-metal/ggml-metal.m |  1 +
+ ggml/src/ggml-cuda/ggml-cuda.cu    | 67 +++++++++++++++++++++++++++---
+ ggml/src/ggml-metal/ggml-metal.cpp |  1 +
  3 files changed, 63 insertions(+), 6 deletions(-)
 
 diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
-index 8a91b381..9424394e 100644
+index fe20dca3..48777212 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
-@@ -152,6 +152,7 @@ extern "C" {
-     struct ggml_backend_dev_props {
-         const char * name;
+@@ -158,6 +158,7 @@ extern "C" {
          const char * description;
-+        const char * id;
+         // device free memory in bytes
          size_t memory_free;
++        const char * id;
+         // device total memory in bytes
          size_t memory_total;
-         enum ggml_backend_dev_type type;
+         // device type
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 37ee2a6d..57eae461 100644
+index fdf8c63d..ad389ece 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -179,6 +179,51 @@ static int ggml_cuda_parse_id(char devName[]) {
+@@ -183,6 +183,51 @@ static int ggml_cuda_parse_id(char devName[]) {
  }
  #endif // defined(GGML_USE_HIP)
  
@@ -77,9 +77,9 @@ index 37ee2a6d..57eae461 100644
 +}
 +
  static ggml_cuda_device_info ggml_cuda_init() {
- #if defined(GGML_USE_HIP)
-     // Workaround for a rocBLAS bug when using multiple graphics cards:
-@@ -267,22 +312,24 @@ static ggml_cuda_device_info ggml_cuda_init() {
+     ggml_cuda_device_info info = {};
+ 
+@@ -249,22 +294,24 @@ static ggml_cuda_device_info ggml_cuda_init() {
                  info.devices[id].cc += prop.minor * 0x10;
              }
          }
@@ -107,18 +107,18 @@ index 37ee2a6d..57eae461 100644
 +        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s, ID: %s\n",
 +                        id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
 +                        ggml_cuda_parse_uuid(prop, id).c_str());
- #endif // defined(GGML_USE_HIP)
-     }
- 
-@@ -3144,6 +3191,7 @@ struct ggml_backend_cuda_device_context {
-     int device;
+         std::string device_name(prop.name);
+         if (device_name == "NVIDIA GeForce MX450") {
+             turing_devices_without_mma.push_back({ id, device_name });
+@@ -3273,6 +3320,7 @@ struct ggml_backend_cuda_device_context {
      std::string name;
      std::string description;
+     std::string pci_bus_id;
 +    std::string id;
  };
  
  static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
-@@ -3156,6 +3204,11 @@ static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t
+@@ -3285,6 +3333,11 @@ static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t
      return ctx->description.c_str();
  }
  
@@ -130,31 +130,31 @@ index 37ee2a6d..57eae461 100644
  static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
      ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
      ggml_cuda_set_device(ctx->device);
-@@ -3170,6 +3223,7 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
- static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+@@ -3301,6 +3354,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
+ 
      props->name        = ggml_backend_cuda_device_get_name(dev);
      props->description = ggml_backend_cuda_device_get_description(dev);
 +    props->id          = ggml_backend_cuda_device_get_id(dev);
      props->type        = ggml_backend_cuda_device_get_type(dev);
+     props->device_id   = ctx->pci_bus_id.empty() ? nullptr : ctx->pci_bus_id.c_str();
      ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
- 
-@@ -3767,6 +3821,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+@@ -3871,6 +3925,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                  cudaDeviceProp prop;
                  CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                  dev_ctx->description = prop.name;
 +                dev_ctx->id = ggml_cuda_parse_uuid(prop, i);
  
-                 ggml_backend_dev_t dev = new ggml_backend_device {
-                     /* .iface   = */ ggml_backend_cuda_device_interface,
-diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
-index 7bccc7bf..fe7b2f0a 100644
---- a/ggml/src/ggml-metal/ggml-metal.m
-+++ b/ggml/src/ggml-metal/ggml-metal.m
-@@ -6522,6 +6522,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
- static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+                 char pci_bus_id[16] = {};
+                 snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.0", prop.pciDomainID, prop.pciBusID, prop.pciDeviceID);
+diff --git a/ggml/src/ggml-metal/ggml-metal.cpp b/ggml/src/ggml-metal/ggml-metal.cpp
+index 909e17de..08ab4fc9 100644
+--- a/ggml/src/ggml-metal/ggml-metal.cpp
++++ b/ggml/src/ggml-metal/ggml-metal.cpp
+@@ -538,6 +538,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
+ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
      props->name        = ggml_backend_metal_device_get_name(dev);
      props->description = ggml_backend_metal_device_get_description(dev);
 +    props->id          = "0";
      props->type        = ggml_backend_metal_device_get_type(dev);
+ 
      ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total);
-     props->caps = (struct ggml_backend_dev_caps) {

llama/patches/0016-add-C-API-for-mtmd_input_text.patch

@@ -10,7 +10,7 @@ Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
  2 files changed, 13 insertions(+)
 
 diff --git a/tools/mtmd/mtmd.cpp b/tools/mtmd/mtmd.cpp
-index a05373d5..6f70f7f4 100644
+index cd022c5e..3d680945 100644
 --- a/tools/mtmd/mtmd.cpp
 +++ b/tools/mtmd/mtmd.cpp
 @@ -79,6 +79,16 @@ enum mtmd_slice_tmpl {

llama/patches/0017-no-power-throttling-win32-with-gnuc.patch

@@ -8,10 +8,10 @@ Subject: [PATCH] no power throttling win32 with gnuc
  1 file changed, 1 insertion(+), 1 deletion(-)
 
 diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
-index a5689c18..85af19a3 100644
+index f8574d01..530efce0 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.c
 +++ b/ggml/src/ggml-cpu/ggml-cpu.c
-@@ -2412,7 +2412,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
+@@ -2431,7 +2431,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
          // Newer Windows 11 versions aggresively park (offline) CPU cores and often place
          // all our threads onto the first 4 cores which results in terrible performance with
          // n_threads > 4

llama/patches/0018-BF16-macos-version-guard.patch

@@ -5,23 +5,24 @@ Subject: [PATCH] BF16 macos version guard
 
 Only enable BF16 on supported MacOS versions (v14+)
 ---
- ggml/src/ggml-metal/ggml-metal.m | 6 +++++-
- 1 file changed, 5 insertions(+), 1 deletion(-)
+ ggml/src/ggml-metal/ggml-metal-context.m | 7 ++++++-
+ 1 file changed, 6 insertions(+), 1 deletion(-)
 
-diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
-index fe7b2f0a..e4c31268 100644
---- a/ggml/src/ggml-metal/ggml-metal.m
-+++ b/ggml/src/ggml-metal/ggml-metal.m
-@@ -106,7 +106,11 @@ static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_dev
-         ctx->has_bfloat |= [ctx->mtl_device supportsFamily:MTLGPUFamilyApple6];
+diff --git a/ggml/src/ggml-metal/ggml-metal-context.m b/ggml/src/ggml-metal/ggml-metal-context.m
+index 052efb7a..b47dc787 100644
+--- a/ggml/src/ggml-metal/ggml-metal-context.m
++++ b/ggml/src/ggml-metal/ggml-metal-context.m
+@@ -125,7 +125,12 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
  
- #if defined(GGML_METAL_USE_BF16)
--        ctx->use_bfloat = ctx->has_bfloat;
+     res->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
+ 
+-    res->use_bfloat      = props_dev->has_bfloat;
 +    if (@available(macOS 14.0, *)) {
-+            ctx->use_bfloat = ctx->has_bfloat;
++        res->use_bfloat = props_dev->has_bfloat;
 +    } else {
-+            ctx->use_bfloat = false;
++        res->use_bfloat = false;
 +    }
- #else
-         ctx->use_bfloat = false;
- #endif
++
+     res->use_fusion      = getenv("GGML_METAL_FUSION_DISABLE") == nil;
+     res->use_concurrency = getenv("GGML_METAL_CONCURRENCY_DISABLE") == nil;
+