GGML update to ec98e2002 (#13451)

* Revert "add support for NVIDIA Nemotron 3 Nano"

This reverts commit e7d2ae9d69.

* GGML update to 380b4c984

Remove MaskBatchPadding as GGML_KQ_MASK_PAD is no longer present (no
padding required)

* update to c45f89d55

* ec98e2002

solar pro needed more adjusting - needs verification

* review comments

.github/workflows/release.yaml

@@ -16,13 +16,15 @@ jobs:
     outputs:
       GOFLAGS: ${{ steps.goflags.outputs.GOFLAGS }}
       VERSION: ${{ steps.goflags.outputs.VERSION }}
+      vendorsha: ${{ steps.changes.outputs.vendorsha }}
     steps:
       - uses: actions/checkout@v4
       - name: Set environment
         id: goflags
         run: |
-          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${GITHUB_REF_NAME#v}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_OUTPUT
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${GITHUB_REF_NAME#v}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" | tee -a $GITHUB_OUTPUT
-          echo VERSION="${GITHUB_REF_NAME#v}" >>$GITHUB_OUTPUT
+          echo VERSION="${GITHUB_REF_NAME#v}" | tee -a $GITHUB_OUTPUT
+          echo vendorsha=$(make -f Makefile.sync print-base) | tee -a $GITHUB_OUTPUT
 
   darwin-build:
     runs-on: macos-14-xlarge
@@ -53,6 +55,9 @@ jobs:
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod
+          cache-dependency-path: |
+            go.sum
+            Makefile.sync
       - run: |
           ./scripts/build_darwin.sh
       - name: Log build results
@@ -185,7 +190,7 @@ jobs:
       - uses: actions/cache@v4
         with:
           path: ${{ github.workspace }}\.ccache
-          key: ccache-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}
+          key: ccache-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}-${{ needs.setup-environment.outputs.vendorsha }}
       - name: Build target "${{ matrix.preset }}"
         run: |
           Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
@@ -249,6 +254,9 @@ jobs:
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod
+          cache-dependency-path: |
+            go.sum
+            Makefile.sync
       - name: Verify gcc is actually clang
         run: |
           $ErrorActionPreference='Continue'
@@ -302,6 +310,9 @@ jobs:
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod
+          cache-dependency-path: |
+            go.sum
+            Makefile.sync
       - uses: actions/download-artifact@v4
         with:
           pattern: depends-windows*

.github/workflows/test.yaml

@@ -22,6 +22,7 @@ jobs:
     runs-on: ubuntu-latest
     outputs:
       changed: ${{ steps.changes.outputs.changed }}
+      vendorsha: ${{ steps.changes.outputs.vendorsha }}
     steps:
       - uses: actions/checkout@v4
         with:
@@ -37,6 +38,7 @@ jobs:
           }
 
           echo changed=$(changed 'llama/llama.cpp/**/*' 'ml/backend/ggml/ggml/**/*') | tee -a $GITHUB_OUTPUT
+          echo vendorsha=$(make -f Makefile.sync print-base) | tee -a $GITHUB_OUTPUT
 
   linux:
     needs: [changes]
@@ -83,7 +85,7 @@ jobs:
       - uses: actions/cache@v4
         with:
           path: /github/home/.cache/ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}
+          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.vendorsha }}
       - run: |
           cmake --preset ${{ matrix.preset }} ${{ matrix.flags }}
           cmake --build --preset ${{ matrix.preset }} --parallel
@@ -178,7 +180,7 @@ jobs:
       - uses: actions/cache@v4
         with:
           path: ${{ github.workspace }}\.ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}
+          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.vendorsha }}
       - run: |
           Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
           Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
@@ -206,6 +208,9 @@ jobs:
       - uses: actions/setup-go@v5
         with:
           go-version-file: 'go.mod'
+          cache-dependency-path: |
+            go.sum
+            Makefile.sync
       - uses: actions/setup-node@v4
         with:
           node-version: '20'

CMakeLists.txt

@@ -54,6 +54,13 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cp
 
 add_compile_definitions(NDEBUG GGML_VERSION=0x0 GGML_COMMIT=0x0)
 
+# Define GGML version variables for shared library SOVERSION
+# These are required by ggml/src/CMakeLists.txt for proper library versioning
+set(GGML_VERSION_MAJOR 0)
+set(GGML_VERSION_MINOR 0)
+set(GGML_VERSION_PATCH 0)
+set(GGML_VERSION "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
+
 set(GGML_CPU ON)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
 set_property(TARGET ggml PROPERTY EXCLUDE_FROM_ALL TRUE)

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggml-org/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=7f8ef50cce40e3e7e4526a3696cb45658190e69a
+FETCH_HEAD=ec98e2002
 
 .PHONY: help
 help:
@@ -57,7 +57,7 @@ checkout: $(WORKDIR)
 $(WORKDIR):
 	git clone $(UPSTREAM) $(WORKDIR)
 
-.PHONE: format-patches
+.PHONY: format-patches
 format-patches: llama/patches
 	git -C $(WORKDIR) format-patch \
 		--no-signature \
@@ -66,7 +66,11 @@ format-patches: llama/patches
 		-o $(realpath $<) \
 		$(FETCH_HEAD)
 
-.PHONE: clean
+.PHONY: clean
 clean: checkout
 	@git -C $(WORKDIR) am --abort || true
 	$(RM) llama/patches/.*.patched
+
+.PHONY: print-base
+print-base:
+	@echo $(FETCH_HEAD)

README.md

@@ -555,7 +555,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Parakeet](https://github.com/parakeet-nest/parakeet) is a GoLang library, made to simplify the development of small generative AI applications with Ollama.
 - [Haverscript](https://github.com/andygill/haverscript) with [examples](https://github.com/andygill/haverscript/tree/main/examples)
 - [Ollama for Swift](https://github.com/mattt/ollama-swift)
-- [Swollama for Swift](https://github.com/marcusziade/Swollama) with [DocC](https://marcusziade.github.io/Swollama/documentation/swollama/)
+- [Swollama for Swift](https://github.com/guitaripod/Swollama) with [DocC](https://guitaripod.github.io/Swollama/documentation/swollama)
 - [GoLamify](https://github.com/prasad89/golamify)
 - [Ollama for Haskell](https://github.com/tusharad/ollama-haskell)
 - [multi-llm-ts](https://github.com/nbonamy/multi-llm-ts) (A Typescript/JavaScript library allowing access to different LLM in a unified API)

api/client.go

@@ -347,7 +347,7 @@ type CreateProgressFunc func(ProgressResponse) error
 // Create creates a model from a [Modelfile]. fn is a progress function that
 // behaves similarly to other methods (see [Client.Pull]).
 //
-// [Modelfile]: https://github.com/ollama/ollama/blob/main/docs/modelfile.md
+// [Modelfile]: https://github.com/ollama/ollama/blob/main/docs/modelfile.mdx
 func (c *Client) Create(ctx context.Context, req *CreateRequest, fn CreateProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/create", req, func(bts []byte) error {
 		var resp ProgressResponse

api/types.go

@@ -288,6 +288,7 @@ type ToolProperty struct {
 	Items       any                     `json:"items,omitempty"`
 	Description string                  `json:"description,omitempty"`
 	Enum        []any                   `json:"enum,omitempty"`
+	Properties  map[string]ToolProperty `json:"properties,omitempty"`
 }
 
 // ToTypeScriptType converts a ToolProperty to a TypeScript type string

api/types_test.go

@@ -504,6 +504,107 @@ func TestThinking_UnmarshalJSON(t *testing.T) {
 	}
 }
 
+func TestToolPropertyNestedProperties(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    string
+		expected ToolProperty
+	}{
+		{
+			name: "nested object properties",
+			input: `{
+				"type": "object",
+				"description": "Location details",
+				"properties": {
+					"address": {
+						"type": "string",
+						"description": "Street address"
+					},
+					"city": {
+						"type": "string",
+						"description": "City name"
+					}
+				}
+			}`,
+			expected: ToolProperty{
+				Type:        PropertyType{"object"},
+				Description: "Location details",
+				Properties: map[string]ToolProperty{
+					"address": {
+						Type:        PropertyType{"string"},
+						Description: "Street address",
+					},
+					"city": {
+						Type:        PropertyType{"string"},
+						Description: "City name",
+					},
+				},
+			},
+		},
+		{
+			name: "deeply nested properties",
+			input: `{
+				"type": "object",
+				"description": "Event",
+				"properties": {
+					"location": {
+						"type": "object",
+						"description": "Location",
+						"properties": {
+							"coordinates": {
+								"type": "object",
+								"description": "GPS coordinates",
+								"properties": {
+									"lat": {"type": "number", "description": "Latitude"},
+									"lng": {"type": "number", "description": "Longitude"}
+								}
+							}
+						}
+					}
+				}
+			}`,
+			expected: ToolProperty{
+				Type:        PropertyType{"object"},
+				Description: "Event",
+				Properties: map[string]ToolProperty{
+					"location": {
+						Type:        PropertyType{"object"},
+						Description: "Location",
+						Properties: map[string]ToolProperty{
+							"coordinates": {
+								Type:        PropertyType{"object"},
+								Description: "GPS coordinates",
+								Properties: map[string]ToolProperty{
+									"lat": {Type: PropertyType{"number"}, Description: "Latitude"},
+									"lng": {Type: PropertyType{"number"}, Description: "Longitude"},
+								},
+							},
+						},
+					},
+				},
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var prop ToolProperty
+			err := json.Unmarshal([]byte(tt.input), &prop)
+			require.NoError(t, err)
+			assert.Equal(t, tt.expected, prop)
+
+			// Round-trip test: marshal and unmarshal again
+			data, err := json.Marshal(prop)
+			require.NoError(t, err)
+
+			var prop2 ToolProperty
+			err = json.Unmarshal(data, &prop2)
+			require.NoError(t, err)
+			assert.Equal(t, tt.expected, prop2)
+		})
+	}
+}
+
 func TestToolFunctionParameters_String(t *testing.T) {
 	tests := []struct {
 		name     string

app/cmd/app/app.go

@@ -273,10 +273,6 @@ func main() {
 		Handler: uiServer.Handler(),
 	}
 
-	if _, err := uiServer.UserData(ctx); err != nil {
-		slog.Warn("failed to load user data", "error", err)
-	}
-
 	// Start the UI server
 	slog.Info("starting ui server", "port", port)
 	go func() {
@@ -320,6 +316,17 @@ func main() {
 		slog.Debug("no URL scheme request to handle")
 	}
 
+	go func() {
+		slog.Debug("waiting for ollama server to be ready")
+		if err := ui.WaitForServer(ctx, 10*time.Second); err != nil {
+			slog.Warn("ollama server not ready, continuing anyway", "error", err)
+		}
+
+		if _, err := uiServer.UserData(ctx); err != nil {
+			slog.Warn("failed to load user data", "error", err)
+		}
+	}()
+
 	osRun(cancel, hasCompletedFirstRun, startHidden)
 
 	slog.Info("shutting down desktop server")
@@ -361,7 +368,7 @@ func checkUserLoggedIn(uiServerPort int) bool {
 		return false
 	}
 
-	resp, err := http.Get(fmt.Sprintf("http://127.0.0.1:%d/api/v1/me", uiServerPort))
+	resp, err := http.Post(fmt.Sprintf("http://127.0.0.1:%d/api/me", uiServerPort), "application/json", nil)
 	if err != nil {
 		slog.Debug("failed to call local auth endpoint", "error", err)
 		return false

app/cmd/app/app_darwin.go

@@ -191,13 +191,6 @@ func LaunchNewApp() {
 	C.launchApp(appName)
 }
 
-// Send a request to the main app thread to load a UI page
-func sendUIRequestMessage(path string) {
-	p := C.CString(path)
-	defer C.free(unsafe.Pointer(p))
-	C.uiRequest(p)
-}
-
 func registerLaunchAgent(hasCompletedFirstRun bool) {
 	// Remove any stale Login Item registrations
 	C.unregisterSelfFromLoginItem()

app/cmd/app/app_windows.go

@@ -263,11 +263,6 @@ func createLoginShortcut() error {
 	return nil
 }
 
-// Send a request to the main app thread to load a UI page
-func sendUIRequestMessage(path string) {
-	wintray.SendUIRequestMessage(path)
-}
-
 func LaunchNewApp() {
 }
 

app/dialog/cocoa/dlg.m

@@ -169,7 +169,11 @@ DlgResult fileDlg(FileDlgParams* params) {
 	}
 	
 	NSArray* urls = [panel URLs];
-	if(self->params->allowMultiple && [urls count] >= 1) {
+	if([urls count] == 0) {
+		return DLG_CANCEL;
+	}
+	
+	if(self->params->allowMultiple) {
 		// For multiple files, we need to return all paths separated by null bytes
 		char* bufPtr = self->params->buf;
 		int remainingBuf = self->params->nbuf;
@@ -200,6 +204,12 @@ DlgResult fileDlg(FileDlgParams* params) {
 			bufPtr += pathLen + 1;
 		}
 		*bufPtr = '\0'; // Final null terminator
+	} else {
+		// Single file/directory selection - write path to buffer
+		NSURL* url = [urls firstObject];
+		if(![url getFileSystemRepresentation:self->params->buf maxLength:self->params->nbuf]) {
+			return DLG_URLFAIL;
+		}
 	}
 	
 	return DLG_OK;

app/dialog/dlgs_windows.go

@@ -15,7 +15,7 @@ const multiFileBufferSize = w32.MAX_PATH * 10
 type WinDlgError int
 
 func (e WinDlgError) Error() string {
-	return fmt.Sprintf("CommDlgExtendedError: %#x", e)
+	return fmt.Sprintf("CommDlgExtendedError: %#x", int(e))
 }
 
 func err() error {

app/server/server.go

@@ -224,9 +224,7 @@ func (s *Server) cmd(ctx context.Context) (*exec.Cmd, error) {
 		if _, err := os.Stat(settings.Models); err == nil {
 			env["OLLAMA_MODELS"] = settings.Models
 		} else {
-			slog.Warn("models path not accessible, clearing models setting", "path", settings.Models, "err", err)
+			slog.Warn("models path not accessible, using default", "path", settings.Models, "err", err)
-			settings.Models = ""
-			s.store.SetSettings(settings)
 		}
 	}
 	if settings.ContextLength > 0 {

app/ui/app/codegen/gotypes.gen.ts

@@ -469,26 +469,24 @@ export class HealthResponse {
 }
 export class User {
     id: string;
-    name: string;
     email: string;
-    avatarURL: string;
+    name: string;
-    plan: string;
+    bio?: string;
-    bio: string;
+    avatarurl?: string;
-    firstName: string;
+    firstname?: string;
-    lastName: string;
+    lastname?: string;
-    overThreshold: boolean;
+    plan?: string;
 
     constructor(source: any = {}) {
         if ('string' === typeof source) source = JSON.parse(source);
         this.id = source["id"];
-        this.name = source["name"];
         this.email = source["email"];
-        this.avatarURL = source["avatarURL"];
+        this.name = source["name"];
-        this.plan = source["plan"];
         this.bio = source["bio"];
-        this.firstName = source["firstName"];
+        this.avatarurl = source["avatarurl"];
-        this.lastName = source["lastName"];
+        this.firstname = source["firstname"];
-        this.overThreshold = source["overThreshold"];
+        this.lastname = source["lastname"];
+        this.plan = source["plan"];
     }
 }
 export class Attachment {

app/ui/app/src/api.ts

@@ -15,7 +15,7 @@ import {
 import { parseJsonlFromResponse } from "./util/jsonl-parsing";
 import { ollamaClient as ollama } from "./lib/ollama-client";
 import type { ModelResponse } from "ollama/browser";
-import { API_BASE } from "./lib/config";
+import { API_BASE, OLLAMA_DOT_COM } from "./lib/config";
 
 // Extend Model class with utility methods
 declare module "@/gotypes" {
@@ -27,7 +27,6 @@ declare module "@/gotypes" {
 Model.prototype.isCloud = function (): boolean {
   return this.model.endsWith("cloud");
 };
-
 // Helper function to convert Uint8Array to base64
 function uint8ArrayToBase64(uint8Array: Uint8Array): string {
   const chunkSize = 0x8000; // 32KB chunks to avoid stack overflow
@@ -42,9 +41,8 @@ function uint8ArrayToBase64(uint8Array: Uint8Array): string {
 }
 
 export async function fetchUser(): Promise<User | null> {
-  try {
+  const response = await fetch(`${API_BASE}/api/me`, {
-    const response = await fetch(`${API_BASE}/api/v1/me`, {
+    method: "POST",
-      method: "GET",
     headers: {
       "Content-Type": "application/json",
     },
@@ -52,34 +50,41 @@ export async function fetchUser(): Promise<User | null> {
 
   if (response.ok) {
     const userData: User = await response.json();
+
+    if (userData.avatarurl && !userData.avatarurl.startsWith("http")) {
+      userData.avatarurl = `${OLLAMA_DOT_COM}${userData.avatarurl}`;
+    }
+
     return userData;
   }
 
-    return null;
+  if (response.status === 401 || response.status === 403) {
-  } catch (error) {
-    console.error("Error fetching user:", error);
     return null;
   }
+
+  throw new Error(`Failed to fetch user: ${response.status}`);
 }
 
 export async function fetchConnectUrl(): Promise<string> {
-  const response = await fetch(`${API_BASE}/api/v1/connect`, {
+  const response = await fetch(`${API_BASE}/api/me`, {
-    method: "GET",
+    method: "POST",
     headers: {
       "Content-Type": "application/json",
     },
   });
 
-  if (!response.ok) {
+  if (response.status === 401) {
+    const data = await response.json();
+    if (data.signin_url) {
+      return data.signin_url;
+    }
+  }
+
   throw new Error("Failed to fetch connect URL");
 }
 
-  const data = await response.json();
-  return data.connect_url;
-}
-
 export async function disconnectUser(): Promise<void> {
-  const response = await fetch(`${API_BASE}/api/v1/disconnect`, {
+  const response = await fetch(`${API_BASE}/api/signout`, {
     method: "POST",
     headers: {
       "Content-Type": "application/json",
@@ -389,7 +394,8 @@ export async function getInferenceCompute(): Promise<InferenceCompute[]> {
 
 export async function fetchHealth(): Promise<boolean> {
   try {
-    const response = await fetch(`${API_BASE}/api/v1/health`, {
+    // Use the /api/version endpoint as a health check
+    const response = await fetch(`${API_BASE}/api/version`, {
       method: "GET",
       headers: {
         "Content-Type": "application/json",
@@ -398,7 +404,8 @@ export async function fetchHealth(): Promise<boolean> {
 
     if (response.ok) {
       const data = await response.json();
-      return data.healthy || false;
+      // If we get a version back, the server is healthy
+      return !!data.version;
     }
 
     return false;

app/ui/app/src/components/Settings.tsx

@@ -299,9 +299,9 @@ export default function Settings() {
                         </Button>
                       </div>
                     </div>
-                    {user?.avatarURL && (
+                    {user?.avatarurl && (
                       <img
-                        src={user.avatarURL}
+                        src={user.avatarurl}
                         alt={user?.name}
                         className="h-10 w-10 rounded-full bg-neutral-200 dark:bg-neutral-700 flex-shrink-0"
                         onError={(e) => {

app/ui/app/src/components/Thinking.tsx

@@ -50,6 +50,9 @@ export default function Thinking({
   // Position content to show bottom when collapsed
   useEffect(() => {
     if (isCollapsed && contentRef.current && wrapperRef.current) {
+      requestAnimationFrame(() => {
+        if (!contentRef.current || !wrapperRef.current) return;
+
         const contentHeight = contentRef.current.scrollHeight;
         const wrapperHeight = wrapperRef.current.clientHeight;
         if (contentHeight > wrapperHeight) {
@@ -57,14 +60,23 @@ export default function Thinking({
           contentRef.current.style.transform = `translateY(${translateY}px)`;
           setHasOverflow(true);
         } else {
+          contentRef.current.style.transform = "translateY(0)";
           setHasOverflow(false);
         }
+      });
     } else if (contentRef.current) {
       contentRef.current.style.transform = "translateY(0)";
       setHasOverflow(false);
     }
   }, [thinking, isCollapsed]);
 
+  useEffect(() => {
+    if (activelyThinking && wrapperRef.current && !isCollapsed) {
+      // When expanded and actively thinking, scroll to bottom
+      wrapperRef.current.scrollTop = wrapperRef.current.scrollHeight;
+    }
+  }, [thinking, activelyThinking, isCollapsed]);
+
   const handleToggle = () => {
     setIsCollapsed(!isCollapsed);
     setHasUserInteracted(true);

app/ui/app/src/hooks/useChats.ts

@@ -7,6 +7,7 @@ import { createQueryBatcher } from "./useQueryBatcher";
 import { useRefetchModels } from "./useModels";
 import { useStreamingContext } from "@/contexts/StreamingContext";
 import { useSettings } from "./useSettings";
+import { getModelCapabilities } from "@/api";
 
 export const useChats = () => {
   return useQuery({
@@ -606,6 +607,24 @@ export const useSendMessage = (chatId: string) => {
               queryClient.setQueryData(["staleModels"], newStaleMap);
 
               queryClient.invalidateQueries({ queryKey: ["models"] });
+
+              // Fetch fresh capabilities for the downloaded model
+              getModelCapabilities(selectedModel.model)
+                .then((capabilities) => {
+                  queryClient.setQueryData(
+                    ["modelCapabilities", selectedModel.model],
+                    capabilities,
+                  );
+                })
+                .catch((error) => {
+                  console.error(
+                    "Failed to fetch capabilities after download:",
+                    error,
+                  );
+                  queryClient.invalidateQueries({
+                    queryKey: ["modelCapabilities", selectedModel.model],
+                  });
+                });
             }
             break;
           }

app/ui/app/src/hooks/useDownloadModel.ts

@@ -1,114 +0,0 @@
-import { useMutation, useQueryClient } from "@tanstack/react-query";
-import { useState } from "react";
-import { pullModel } from "@/api";
-import { useSelectedModel } from "./useSelectedModel";
-import { useSettings } from "./useSettings";
-
-interface DownloadProgress {
-  status: string;
-  digest?: string;
-  total?: number;
-  completed?: number;
-  done?: boolean;
-}
-
-export function useDownloadModel(chatId?: string) {
-  const queryClient = useQueryClient();
-  const { selectedModel } = useSelectedModel(chatId);
-  const { setSettings } = useSettings();
-  const [downloadProgress, setDownloadProgress] =
-    useState<DownloadProgress | null>(null);
-  const [abortController, setAbortController] =
-    useState<AbortController | null>(null);
-  const [downloadingChatIds, setDownloadingChatIds] = useState<Set<string>>(
-    new Set(),
-  );
-
-  const mutation = useMutation({
-    mutationFn: async (modelName: string) => {
-      const controller = new AbortController();
-      setAbortController(controller);
-      setDownloadProgress({ status: "Starting download..." });
-      if (chatId) {
-        setDownloadingChatIds((prev) => new Set(prev).add(chatId));
-      }
-
-      try {
-        for await (const progress of pullModel(modelName, controller.signal)) {
-          setDownloadProgress(progress);
-
-          if (progress.status === "success") {
-            // Update selected model to indicate it's now available locally
-            if (selectedModel && selectedModel.model === modelName) {
-              setSettings({ SelectedModel: modelName });
-            }
-            // Invalidate models query to refresh the list
-            await queryClient.invalidateQueries({ queryKey: ["models"] });
-            break;
-          }
-        }
-      } finally {
-        setAbortController(null);
-        if (chatId) {
-          setDownloadingChatIds((prev) => {
-            const newSet = new Set(prev);
-            newSet.delete(chatId);
-            return newSet;
-          });
-        }
-      }
-    },
-    onSuccess: () => {
-      setDownloadProgress(null);
-      if (chatId) {
-        setDownloadingChatIds((prev) => {
-          const newSet = new Set(prev);
-          newSet.delete(chatId);
-          return newSet;
-        });
-      }
-    },
-    onError: (error: Error) => {
-      const status =
-        error.name === "AbortError" ? "Download cancelled" : "Download failed";
-      setDownloadProgress({ status, done: true });
-
-      // Clear error message after delay
-      const delay = error.name === "AbortError" ? 1500 : 3000;
-      setTimeout(() => {
-        setDownloadProgress(null);
-        if (chatId) {
-          setDownloadingChatIds((prev) => {
-            const newSet = new Set(prev);
-            newSet.delete(chatId);
-            return newSet;
-          });
-        }
-      }, delay);
-    },
-  });
-
-  const cancelDownload = () => {
-    if (abortController) {
-      abortController.abort();
-      setAbortController(null);
-      if (chatId) {
-        setDownloadingChatIds((prev) => {
-          const newSet = new Set(prev);
-          newSet.delete(chatId);
-          return newSet;
-        });
-      }
-    }
-  };
-
-  return {
-    downloadModel: mutation.mutate,
-    isDownloading:
-      mutation.isPending && chatId ? downloadingChatIds.has(chatId) : false,
-    downloadProgress:
-      chatId && downloadingChatIds.has(chatId) ? downloadProgress : null,
-    error: mutation.error,
-    cancelDownload,
-  };
-}

app/ui/app/src/hooks/useUser.ts

@@ -1,29 +1,20 @@
 import { useQuery, useMutation, useQueryClient } from "@tanstack/react-query";
-import { useEffect, useState } from "react";
 import { fetchUser, fetchConnectUrl, disconnectUser } from "@/api";
 
 export function useUser() {
   const queryClient = useQueryClient();
-  const [initialDataLoaded, setInitialDataLoaded] = useState(false);
-
-  // Wait for initial data to be loaded
-  useEffect(() => {
-    const initialPromise = window.__initialUserDataPromise;
-    if (initialPromise) {
-      initialPromise.finally(() => {
-        setInitialDataLoaded(true);
-      });
-    } else {
-      setInitialDataLoaded(true);
-    }
-  }, []);
 
   const userQuery = useQuery({
     queryKey: ["user"],
-    queryFn: () => fetchUser(),
+    queryFn: async () => {
+      const result = await fetchUser();
+      return result;
+    },
     staleTime: 5 * 60 * 1000, // Consider data stale after 5 minutes
     gcTime: 10 * 60 * 1000, // Keep in cache for 10 minutes
-    initialData: null, // Start with null to prevent flashing
+    retry: 10,
+    retryDelay: (attemptIndex) => Math.min(500 * attemptIndex, 2000),
+    refetchOnMount: true, // Always fetch when component mounts
   });
 
   // Mutation to refresh user data
@@ -49,14 +40,15 @@ export function useUser() {
     },
   });
 
+  const isLoading = userQuery.isLoading || userQuery.isFetching;
+  const isAuthenticated = Boolean(userQuery.data?.name);
+
   return {
     user: userQuery.data,
-    isLoading:
+    isLoading,
-      !initialDataLoaded ||
-      (userQuery.isLoading && userQuery.data === undefined), // Show loading until initial data is loaded
     isError: userQuery.isError,
     error: userQuery.error,
-    isAuthenticated: Boolean(userQuery.data?.name),
+    isAuthenticated,
     refreshUser: refreshUser.mutate,
     isRefreshing: refreshUser.isPending,
     refetchUser: userQuery.refetch,

app/ui/app/src/lib/config.ts

@@ -8,3 +8,6 @@ export const API_BASE = import.meta.env.DEV ? DEV_API_URL : "";
 export const OLLAMA_HOST = import.meta.env.DEV
   ? DEV_API_URL
   : window.location.origin;
+
+export const OLLAMA_DOT_COM =
+  import.meta.env.VITE_OLLAMA_DOT_COM_URL || "https://ollama.com";

app/ui/app/src/main.tsx

@@ -5,13 +5,6 @@ import { QueryClient, QueryClientProvider } from "@tanstack/react-query";
 import { routeTree } from "./routeTree.gen";
 import { fetchUser } from "./api";
 import { StreamingProvider } from "./contexts/StreamingContext";
-import { User } from "@/gotypes";
-
-declare global {
-  interface Window {
-    __initialUserDataPromise?: Promise<User | null>;
-  }
-}
 
 const queryClient = new QueryClient({
   defaultOptions: {
@@ -24,27 +17,11 @@ const queryClient = new QueryClient({
   },
 });
 
-// Track initial user data fetch
+fetchUser().then((userData) => {
-let initialUserDataPromise: Promise<User | null> | null = null;
+  if (userData) {
-
-// Initialize user data on app startup
-const initializeUserData = async () => {
-  try {
-    const userData = await fetchUser();
     queryClient.setQueryData(["user"], userData);
-    return userData;
-  } catch (error) {
-    console.error("Error initializing user data:", error);
-    queryClient.setQueryData(["user"], null);
-    return null;
   }
-};
+});
-
-// Start initialization immediately and track the promise
-initialUserDataPromise = initializeUserData();
-
-// Export the promise so hooks can await it
-window.__initialUserDataPromise = initialUserDataPromise;
 
 const router = createRouter({
   routeTree,

app/ui/responses/types.go

@@ -102,14 +102,13 @@ type HealthResponse struct {
 
 type User struct {
 	ID        string `json:"id"`
-	Name          string `json:"name"`
 	Email     string `json:"email"`
-	AvatarURL     string `json:"avatarURL"`
+	Name      string `json:"name"`
-	Plan          string `json:"plan"`
+	Bio       string `json:"bio,omitempty"`
-	Bio           string `json:"bio"`
+	AvatarURL string `json:"avatarurl,omitempty"`
-	FirstName     string `json:"firstName"`
+	FirstName string `json:"firstname,omitempty"`
-	LastName      string `json:"lastName"`
+	LastName  string `json:"lastname,omitempty"`
-	OverThreshold bool   `json:"overThreshold"`
+	Plan      string `json:"plan,omitempty"`
 }
 
 type Attachment struct {

app/ui/ui.go

@@ -12,18 +12,17 @@ import (
 	"log/slog"
 	"net/http"
 	"net/http/httputil"
-	"net/url"
 	"os"
 	"runtime"
 	"runtime/debug"
 	"slices"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/google/uuid"
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/app/auth"
 	"github.com/ollama/ollama/app/server"
 	"github.com/ollama/ollama/app/store"
 	"github.com/ollama/ollama/app/tools"
@@ -118,40 +117,66 @@ func (s *Server) log() *slog.Logger {
 
 // ollamaProxy creates a reverse proxy handler to the Ollama server
 func (s *Server) ollamaProxy() http.Handler {
-	ollamaHost := os.Getenv("OLLAMA_HOST")
+	var (
-	if ollamaHost == "" {
+		proxy   http.Handler
-		ollamaHost = "http://127.0.0.1:11434"
+		proxyMu sync.Mutex
-	}
+	)
 
-	if !strings.HasPrefix(ollamaHost, "http://") && !strings.HasPrefix(ollamaHost, "https://") {
-		ollamaHost = "http://" + ollamaHost
-	}
-
-	target, err := url.Parse(ollamaHost)
-	if err != nil {
-		s.log().Error("failed to parse OLLAMA_HOST", "error", err, "host", ollamaHost)
 	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			http.Error(w, "failed to configure proxy", http.StatusInternalServerError)
+		proxyMu.Lock()
-		})
+		p := proxy
+		proxyMu.Unlock()
+
+		if p == nil {
+			proxyMu.Lock()
+			if proxy == nil {
+				var err error
+				for i := range 2 {
+					if i > 0 {
+						s.log().Warn("ollama server not ready, retrying", "attempt", i+1)
+						time.Sleep(1 * time.Second)
 					}
 
+					err = WaitForServer(context.Background(), 10*time.Second)
+					if err == nil {
+						break
+					}
+				}
+
+				if err != nil {
+					proxyMu.Unlock()
+					s.log().Error("ollama server not ready after retries", "error", err)
+					http.Error(w, "Ollama server is not ready", http.StatusServiceUnavailable)
+					return
+				}
+
+				target := envconfig.Host()
 				s.log().Info("configuring ollama proxy", "target", target.String())
 
-	proxy := httputil.NewSingleHostReverseProxy(target)
+				newProxy := httputil.NewSingleHostReverseProxy(target)
 
-	originalDirector := proxy.Director
+				originalDirector := newProxy.Director
-	proxy.Director = func(req *http.Request) {
+				newProxy.Director = func(req *http.Request) {
 					originalDirector(req)
 					req.Host = target.Host
 					s.log().Debug("proxying request", "method", req.Method, "path", req.URL.Path, "target", target.Host)
 				}
 
-	proxy.ErrorHandler = func(w http.ResponseWriter, r *http.Request, err error) {
+				newProxy.ErrorHandler = func(w http.ResponseWriter, r *http.Request, err error) {
 					s.log().Error("proxy error", "error", err, "path", r.URL.Path, "target", target.String())
 					http.Error(w, "proxy error: "+err.Error(), http.StatusBadGateway)
 				}
 
-	return proxy
+				proxy = newProxy
+				p = newProxy
+			} else {
+				p = proxy
+			}
+			proxyMu.Unlock()
+		}
+
+		p.ServeHTTP(w, r)
+	})
 }
 
 type errHandlerFunc func(http.ResponseWriter, *http.Request) error
@@ -264,11 +289,10 @@ func (s *Server) Handler() http.Handler {
 	ollamaProxy := s.ollamaProxy()
 	mux.Handle("GET /api/tags", ollamaProxy)
 	mux.Handle("POST /api/show", ollamaProxy)
-
+	mux.Handle("GET /api/version", ollamaProxy)
-	mux.Handle("GET /api/v1/me", handle(s.me))
+	mux.Handle("HEAD /api/version", ollamaProxy)
-	mux.Handle("POST /api/v1/disconnect", handle(s.disconnect))
+	mux.Handle("POST /api/me", ollamaProxy)
-	mux.Handle("GET /api/v1/connect", handle(s.connectURL))
+	mux.Handle("POST /api/signout", ollamaProxy)
-	mux.Handle("GET /api/v1/health", handle(s.health))
 
 	// React app - catch all non-API routes and serve the React app
 	mux.Handle("GET /", s.appHandler())
@@ -338,7 +362,7 @@ func (s *Server) doSelfSigned(ctx context.Context, method, path string) (*http.R
 }
 
 // UserData fetches user data from ollama.com API for the current ollama key
-func (s *Server) UserData(ctx context.Context) (*responses.User, error) {
+func (s *Server) UserData(ctx context.Context) (*api.UserResponse, error) {
 	resp, err := s.doSelfSigned(ctx, http.MethodPost, "/api/me")
 	if err != nil {
 		return nil, fmt.Errorf("failed to call ollama.com/api/me: %w", err)
@@ -349,7 +373,7 @@ func (s *Server) UserData(ctx context.Context) (*responses.User, error) {
 		return nil, fmt.Errorf("unexpected status code: %d", resp.StatusCode)
 	}
 
-	var user responses.User
+	var user api.UserResponse
 	if err := json.NewDecoder(resp.Body).Decode(&user); err != nil {
 		return nil, fmt.Errorf("failed to parse user response: %w", err)
 	}
@@ -368,29 +392,27 @@ func (s *Server) UserData(ctx context.Context) (*responses.User, error) {
 	return &user, nil
 }
 
-func waitForServer(ctx context.Context) error {
+// WaitForServer waits for the Ollama server to be ready
-	timeout := time.Now().Add(10 * time.Second)
+func WaitForServer(ctx context.Context, timeout time.Duration) error {
-	// TODO: this avoids an error on first load of the app
+	deadline := time.Now().Add(timeout)
-	// however we should either show a loading state or
+	for time.Now().Before(deadline) {
-	// wait for the Ollama server to be ready before redirecting
-	for {
 		c, err := api.ClientFromEnvironment()
 		if err != nil {
 			return err
 		}
 		if _, err := c.Version(ctx); err == nil {
-			break
+			slog.Debug("ollama server is ready")
-		}
+			return nil
-		if time.Now().After(timeout) {
-			return fmt.Errorf("timeout waiting for Ollama server to be ready")
 		}
 		time.Sleep(10 * time.Millisecond)
 	}
-	return nil
+	return errors.New("timeout waiting for Ollama server to be ready")
 }
 
 func (s *Server) createChat(w http.ResponseWriter, r *http.Request) error {
-	waitForServer(r.Context())
+	if err := WaitForServer(r.Context(), 10*time.Second); err != nil {
+		return err
+	}
 
 	id, err := uuid.NewV7()
 	if err != nil {
@@ -1438,129 +1460,6 @@ func (s *Server) settings(w http.ResponseWriter, r *http.Request) error {
 	})
 }
 
-func (s *Server) me(w http.ResponseWriter, r *http.Request) error {
-	if r.Method != http.MethodGet {
-		http.Error(w, "Method Not Allowed", http.StatusMethodNotAllowed)
-		return nil
-	}
-
-	user, err := s.UserData(r.Context())
-	if err != nil {
-		// If fetching from API fails, try to return cached user data if available
-		if cachedUser, cacheErr := s.Store.User(); cacheErr == nil && cachedUser != nil {
-			s.log().Info("API request failed, returning cached user data", "error", err)
-			responseUser := &responses.User{
-				Name:  cachedUser.Name,
-				Email: cachedUser.Email,
-				Plan:  cachedUser.Plan,
-			}
-			w.Header().Set("Content-Type", "application/json")
-			w.WriteHeader(http.StatusOK)
-			return json.NewEncoder(w).Encode(responseUser)
-		}
-
-		s.log().Error("failed to get user data", "error", err)
-		w.WriteHeader(http.StatusInternalServerError)
-		return json.NewEncoder(w).Encode(responses.Error{
-			Error: "failed to get user data",
-		})
-	}
-
-	w.Header().Set("Content-Type", "application/json")
-	w.WriteHeader(http.StatusOK)
-	return json.NewEncoder(w).Encode(user)
-}
-
-func (s *Server) disconnect(w http.ResponseWriter, r *http.Request) error {
-	if r.Method != http.MethodPost {
-		http.Error(w, "Method Not Allowed", http.StatusMethodNotAllowed)
-		return nil
-	}
-
-	if err := s.Store.ClearUser(); err != nil {
-		s.log().Warn("failed to clear cached user data", "error", err)
-	}
-
-	// Get the SSH public key to encode for the delete request
-	pubKey, err := ollamaAuth.GetPublicKey()
-	if err != nil {
-		s.log().Error("failed to get public key", "error", err)
-		w.WriteHeader(http.StatusInternalServerError)
-		return json.NewEncoder(w).Encode(responses.Error{
-			Error: "failed to get public key",
-		})
-	}
-
-	// Encode the key using base64 URL encoding
-	encodedKey := base64.RawURLEncoding.EncodeToString([]byte(pubKey))
-
-	// Call the /api/user/keys/{encodedKey} endpoint with DELETE
-	resp, err := s.doSelfSigned(r.Context(), http.MethodDelete, fmt.Sprintf("/api/user/keys/%s", encodedKey))
-	if err != nil {
-		s.log().Error("failed to call ollama.com/api/user/keys", "error", err)
-		w.WriteHeader(http.StatusInternalServerError)
-		return json.NewEncoder(w).Encode(responses.Error{
-			Error: "failed to disconnect from ollama.com",
-		})
-	}
-	defer resp.Body.Close()
-
-	if resp.StatusCode != http.StatusOK {
-		s.log().Error("disconnect request failed", "status", resp.StatusCode)
-		w.WriteHeader(http.StatusInternalServerError)
-		return json.NewEncoder(w).Encode(responses.Error{
-			Error: "failed to disconnect from ollama.com",
-		})
-	}
-
-	w.Header().Set("Content-Type", "application/json")
-	w.WriteHeader(http.StatusOK)
-	return json.NewEncoder(w).Encode(map[string]string{"status": "disconnected"})
-}
-
-func (s *Server) connectURL(w http.ResponseWriter, r *http.Request) error {
-	if r.Method != http.MethodGet {
-		http.Error(w, "Method Not Allowed", http.StatusMethodNotAllowed)
-		return nil
-	}
-
-	connectURL, err := auth.BuildConnectURL(OllamaDotCom)
-	if err != nil {
-		s.log().Error("failed to build connect URL", "error", err)
-		w.WriteHeader(http.StatusInternalServerError)
-		return json.NewEncoder(w).Encode(responses.Error{
-			Error: "failed to build connect URL",
-		})
-	}
-
-	w.Header().Set("Content-Type", "application/json")
-	w.WriteHeader(http.StatusOK)
-	return json.NewEncoder(w).Encode(map[string]string{
-		"connect_url": connectURL,
-	})
-}
-
-func (s *Server) health(w http.ResponseWriter, r *http.Request) error {
-	if r.Method != http.MethodGet {
-		http.Error(w, "Method Not Allowed", http.StatusMethodNotAllowed)
-		return nil
-	}
-
-	healthy := false
-	c, err := api.ClientFromEnvironment()
-	if err == nil {
-		if _, err := c.Version(r.Context()); err == nil {
-			healthy = true
-		}
-	}
-
-	w.Header().Set("Content-Type", "application/json")
-	w.WriteHeader(http.StatusOK)
-	return json.NewEncoder(w).Encode(responses.HealthResponse{
-		Healthy: healthy,
-	})
-}
-
 func (s *Server) getInferenceCompute(w http.ResponseWriter, r *http.Request) error {
 	ctx, cancel := context.WithTimeout(r.Context(), 500*time.Millisecond)
 	defer cancel()

app/wintray/eventloop.go

@@ -158,16 +158,16 @@ func (t *winTray) wndProc(hWnd windows.Handle, message uint32, wParam, lParam ui
 	case uint32(UI_REQUEST_MSG_ID):
 		// Requests for the UI must always come from the main event thread
 		l := int(wParam)
-		path := unsafe.String((*byte)(unsafe.Pointer(lParam)), l)
+		path := unsafe.String((*byte)(unsafe.Pointer(lParam)), l) //nolint:govet,gosec
 		t.app.UIRun(path)
 	case WM_COPYDATA:
 		// Handle URL scheme requests from other instances
 		if lParam != 0 {
-			cds := (*COPYDATASTRUCT)(unsafe.Pointer(lParam))
+			cds := (*COPYDATASTRUCT)(unsafe.Pointer(lParam)) //nolint:govet,gosec
 			if cds.DwData == 1 {                             // Our identifier for URL scheme messages
 				// Convert the data back to string
 				data := make([]byte, cds.CbData)
-				copy(data, (*[1 << 30]byte)(unsafe.Pointer(cds.LpData))[:cds.CbData:cds.CbData])
+				copy(data, (*[1 << 30]byte)(unsafe.Pointer(cds.LpData))[:cds.CbData:cds.CbData]) //nolint:govet,gosec
 				urlScheme := string(data)
 				handleURLSchemeRequest(urlScheme)
 				lResult = 1 // Return non-zero to indicate success

cmd/bench/README.md

@@ -15,7 +15,7 @@ A Go-based command-line tool for benchmarking Ollama models with configurable pa
 
 ```
 go build -o ollama-bench bench.go
-./bench -model gpt-oss:20b -epochs 6 -format csv
+./ollama-bench -model gpt-oss:20b -epochs 6 -format csv
 ```
 
 Using Go Run (without building)
@@ -29,31 +29,32 @@ go run bench.go -model gpt-oss:20b -epochs 3
 ### Basic Example
 
 ```
-./bench -model gemma3 -epochs 6
+./ollama-bench -model gemma3 -epochs 6
 ```
 
 ### Benchmark Multiple Models
 
 ```
-./bench -model gemma3,gemma3n -epochs 6 -max-tokens 100 -p "Write me a short story" | tee gemma.bench
+./ollama-bench -model gemma3,gemma3n -epochs 6 -max-tokens 100 -p "Write me a short story" | tee gemma.bench
 benchstat -col /name gemma.bench
 ```
 
 ### With Image Prompt
 
 ```
-./bench -model qwen3-vl -image photo.jpg -epochs 6 -max-tokens 100 -p "Describe this image"
+./ollama-bench -model qwen3-vl -image photo.jpg -epochs 6 -max-tokens 100 -p "Describe this image"
 ```
 
 ### Advanced Example
 
 ```
-./bench -model llama3 -epochs 10 -temperature 0.7 -max-tokens 500 -seed 42 -format csv -output results.csv
+./ollama-bench -model llama3 -epochs 10 -temperature 0.7 -max-tokens 500 -seed 42 -format csv -output results.csv
 ```
 
 ## Command Line Options
 
 | Option  	| Description | Default |
+|----------|-------------|---------|
 | -model	| Comma-separated list of models to benchmark	| (required)		|
 | -epochs	| Number of iterations per model		| 1			|
 | -max-tokens	| Maximum tokens for model response		| 0 (unlimited)		|

convert/convert.go

@@ -182,6 +182,8 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &llama4Model{}
 	case "Mistral3ForConditionalGeneration":
 		conv = &mistral3Model{}
+	case "Ministral3ForCausalLM":
+		conv = &mistral3CausalModel{}
 	case "MixtralForCausalLM":
 		conv = &mixtralModel{}
 	case "GemmaForCausalLM":
@@ -200,8 +202,12 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &qwen25VLModel{}
 	case "Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration":
 		conv = &qwen3VLModel{}
+	case "Olmo3ForCausalLM":
+		conv = &olmoModel{}
 	case "BertModel":
 		conv = &bertModel{}
+	case "NomicBertModel", "NomicBertMoEModel":
+		conv = &nomicbertModel{}
 	case "CohereForCausalLM":
 		conv = &commandrModel{}
 	case "GptOssForCausalLM":

convert/convert_gemma3.go

@@ -2,6 +2,7 @@ package convert
 
 import (
 	"cmp"
+	"slices"
 
 	"github.com/ollama/ollama/fs/ggml"
 )
@@ -33,9 +34,19 @@ type gemma3Model struct {
 	HeadDim                  uint32   `json:"head_dim"`
 	FinalLogitSoftcap        float32  `json:"final_logit_softcapping"`
 	RopeLocalTheta           float32  `json:"rope_local_base_freq"`
-	RopeGlobalTheta          float32 `json:"rope_global_base_freq"`
+	RopeTheta                float32  `json:"rope_theta"`
 	SlidingWindow            uint32   `json:"sliding_window"`
+	SlidingWindowPattern     *uint32  `json:"sliding_window_pattern"`
+	LayerTypes               []string `json:"layer_types"`
 	MultiModalTokensPerImage uint32   `json:"mm_tokens_per_image"`
+	RopeScaling              *struct {
+		Type                          string  `json:"rope_type"`
+		Factor                        float32 `json:"factor"`
+		OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
+		ExtrapolationFactor           float32 `json:"extrapolation_factor"`
+		BetaFast                      float32 `json:"beta_fast"`
+		BetaSlow                      float32 `json:"beta_slow"`
+	} `json:"rope_scaling"`
 }
 
 const (
@@ -81,9 +92,38 @@ func (p *gemma3Model) KV(t *Tokenizer) ggml.KV {
 		kv["gemma3.attention.key_length"] = p.HeadDim
 		kv["gemma3.attention.value_length"] = p.HeadDim
 		kv["gemma3.attention.sliding_window"] = p.SlidingWindow
-		kv["gemma3.final_logit_softcapping"] = cmp.Or(p.FinalLogitSoftcap, 30)
+
+		// The sliding window pattern is either provided as the sliding_window_pattern
+		// key (an int) or as the layer_types key (a list of strings).
+		if p.SlidingWindowPattern != nil || len(p.LayerTypes) > 0 {
+			kv["gemma3.attention.sliding_window_pattern"] = slices.Collect(func(yield func(bool) bool) {
+				for i := range numBlocks {
+					var isLocal bool
+					if len(p.LayerTypes) > 0 && int(i) < len(p.LayerTypes) {
+						isLocal = p.LayerTypes[i] == "sliding_attention"
+					} else if p.SlidingWindowPattern != nil && *p.SlidingWindowPattern > 0 {
+						isLocal = (i+1)%*p.SlidingWindowPattern != 0
+					}
+					if !yield(isLocal) {
+						break
+					}
+				}
+			})
+		}
+		if p.FinalLogitSoftcap > 0 {
+			kv["gemma3.final_logit_softcapping"] = p.FinalLogitSoftcap
+		}
 		kv["gemma3.rope.local.freq_base"] = cmp.Or(p.RopeLocalTheta, 10000.0)
-		kv["gemma3.rope.global.freq_base"] = cmp.Or(p.RopeGlobalTheta, 1000000.0)
+		kv["gemma3.rope.freq_base"] = cmp.Or(p.RopeTheta, 1000000.0)
+		if p.RopeScaling != nil && p.RopeScaling.Type == "yarn" && p.RopeScaling.Factor > 0 {
+			kv["gemma3.rope.scaling.type"] = "yarn"
+			kv["gemma3.rope.scaling.factor"] = p.RopeScaling.Factor
+			kv["gemma3.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeddings
+			kv["gemma3.rope.scaling.extrapolation_factor"] = cmp.Or(p.RopeScaling.ExtrapolationFactor, float32(1.0))
+			kv["gemma3.rope.scaling.beta_fast"] = cmp.Or(p.RopeScaling.BetaFast, float32(64.0))
+			kv["gemma3.rope.scaling.beta_slow"] = cmp.Or(p.RopeScaling.BetaSlow, float32(1.0))
+		}
+
 		kv["gemma3.embedding_length"] = p.HiddenSize
 		kv["gemma3.feed_forward_length"] = p.IntermediateSize
 	default:

convert/convert_mistral.go

@@ -33,10 +33,12 @@ type mistral3Model struct {
 			BetaFast                  float32  `json:"beta_fast"`
 			BetaSlow                  float32  `json:"beta_slow"`
 			Factor                    float32  `json:"factor"`
-			ScalingBeta               float32 `json:"llama_4_scaling_beta"`
+			Llama4ScalingBeta         *float32 `json:"llama_4_scaling_beta"`
 			OrigMaxPositionEmbeddings uint32   `json:"original_max_position_embeddings"`
 			RopeType                  string   `json:"rope_type"`
 			RopeTheta                 float32  `json:"rope_theta"`
+			Mscale                    *float32 `json:"mscale"`
+			MscaleAllDim              *float32 `json:"mscale_all_dim"`
 		} `json:"rope_parameters"`
 	} `json:"text_config"`
 	VisionModel struct {
@@ -50,6 +52,9 @@ type mistral3Model struct {
 		HeadDim           uint32  `json:"head_dim"`
 		HiddenAct         string  `json:"hidden_act"`
 		RopeTheta         float32 `json:"rope_theta"`
+		RopeParameters    struct {
+			RopeTheta float32 `json:"rope_theta"`
+		} `json:"rope_parameters"`
 	} `json:"vision_config"`
 	MultiModalProjectorBias bool   `json:"multimodal_projector_bias"`
 	ProjectorHiddenAct      string `json:"projector_hidden_act"`
@@ -72,10 +77,22 @@ func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
 	kv["mistral3.attention.value_length"] = p.TextModel.HeadDim
 	kv["mistral3.rope.dimension_count"] = cmp.Or(p.TextModel.HeadDim, p.TextModel.HiddenSize/p.TextModel.NumAttentionHeads)
 	kv["mistral3.rope.freq_base"] = cmp.Or(p.TextModel.RopeTheta, p.TextModel.RopeParameters.RopeTheta)
+	kv["mistral3.rope.scaling.factor"] = p.TextModel.RopeParameters.Factor
+	kv["mistral3.rope.scaling.type"] = p.TextModel.RopeParameters.RopeType
+	kv["mistral3.rope.scaling.beta_fast"] = p.TextModel.RopeParameters.BetaFast
+	kv["mistral3.rope.scaling.beta_slow"] = p.TextModel.RopeParameters.BetaSlow
 
+	if p.TextModel.RopeParameters.Mscale != nil {
+		kv["mistral3.rope.scaling.mscale"] = *p.TextModel.RopeParameters.Mscale
+	}
+	if p.TextModel.RopeParameters.MscaleAllDim != nil {
+		kv["mistral3.rope.scaling.mscale_all_dim"] = *p.TextModel.RopeParameters.MscaleAllDim
+	}
 	if p.TextModel.RopeParameters.OrigMaxPositionEmbeddings > 0 {
 		kv["mistral3.rope.scaling.original_context_length"] = p.TextModel.RopeParameters.OrigMaxPositionEmbeddings
-		kv["mistral3.rope.scaling_beta"] = p.TextModel.RopeParameters.ScalingBeta
+	}
+	if p.TextModel.RopeParameters.Llama4ScalingBeta != nil {
+		kv["mistral3.rope.scaling_beta"] = *p.TextModel.RopeParameters.Llama4ScalingBeta
 	}
 
 	// Vision configuration
@@ -88,7 +105,7 @@ func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
 	kv["mistral3.vision.patch_size"] = p.VisionModel.PatchSize
 	kv["mistral3.vision.num_channels"] = p.VisionModel.NumChannels
 	// kv["mistral3.vision.attention.layer_norm_epsilon"] = 1e-05 // Default value
-	kv["mistral3.vision.rope.freq_base"] = p.VisionModel.RopeTheta
+	kv["mistral3.vision.rope.freq_base"] = cmp.Or(p.VisionModel.RopeTheta, p.VisionModel.RopeParameters.RopeTheta)
 
 	// Multimodal configuration
 	kv["mistral3.image_token_index"] = p.ImageTokenIndex

convert/convert_mistral_causal.go

@@ -0,0 +1,181 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type mistral3CausalModel struct {
+	ModelParameters
+
+	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RopeTheta             float32 `json:"rope_theta"`
+	RMSNormEPS            float32 `json:"rms_norm_eps"`
+	HeadDim               uint32  `json:"head_dim"`
+	SlidingWindow         *uint32 `json:"sliding_window"`
+	HiddenAct             string  `json:"hidden_act"`
+	VocabSize             uint32  `json:"vocab_size"`
+	RopeParameters        struct {
+		BetaFast                  float32  `json:"beta_fast"`
+		BetaSlow                  float32  `json:"beta_slow"`
+		Factor                    float32  `json:"factor"`
+		Llama4ScalingBeta         *float32 `json:"llama_4_scaling_beta"`
+		OrigMaxPositionEmbeddings uint32   `json:"original_max_position_embeddings"`
+		RopeType                  string   `json:"rope_type"`
+		RopeTheta                 float32  `json:"rope_theta"`
+		Mscale                    *float32 `json:"mscale"`
+		MscaleAllDim              *float32 `json:"mscale_all_dim"`
+	} `json:"rope_parameters"`
+}
+
+func (p *mistral3CausalModel) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "mistral3"
+	kv["mistral3.vocab_size"] = p.VocabSize
+
+	// Text configuration
+	kv["mistral3.block_count"] = p.NumHiddenLayers
+	kv["mistral3.context_length"] = p.MaxPositionEmbeddings
+	kv["mistral3.embedding_length"] = p.HiddenSize
+	kv["mistral3.feed_forward_length"] = p.IntermediateSize
+	kv["mistral3.attention.head_count"] = p.NumAttentionHeads
+	kv["mistral3.attention.head_count_kv"] = p.NumKeyValueHeads
+	kv["mistral3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["mistral3.attention.key_length"] = p.HeadDim
+	kv["mistral3.attention.value_length"] = p.HeadDim
+	kv["mistral3.rope.dimension_count"] = cmp.Or(p.HeadDim, p.HiddenSize/p.NumAttentionHeads)
+	kv["mistral3.rope.freq_base"] = cmp.Or(p.RopeTheta, p.RopeParameters.RopeTheta)
+	kv["mistral3.rope.scaling.factor"] = p.RopeParameters.Factor
+	kv["mistral3.rope.scaling.type"] = p.RopeParameters.RopeType
+	kv["mistral3.rope.scaling.beta_fast"] = p.RopeParameters.BetaFast
+	kv["mistral3.rope.scaling.beta_slow"] = p.RopeParameters.BetaSlow
+
+	if p.RopeParameters.Mscale != nil {
+		kv["mistral3.rope.scaling.mscale"] = *p.RopeParameters.Mscale
+	}
+
+	if p.RopeParameters.MscaleAllDim != nil {
+		kv["mistral3.rope.scaling.mscale_all_dim"] = *p.RopeParameters.MscaleAllDim
+	}
+
+	if p.RopeParameters.OrigMaxPositionEmbeddings > 0 {
+		kv["mistral3.rope.scaling.original_context_length"] = p.RopeParameters.OrigMaxPositionEmbeddings
+		kv["mistral3.rope.scaling_beta"] = *p.RopeParameters.Llama4ScalingBeta
+	}
+
+	if p.RopeParameters.Llama4ScalingBeta != nil {
+		kv["mistral3.rope.scaling_beta"] = *p.RopeParameters.Llama4ScalingBeta
+	}
+
+	return kv
+}
+
+func (p *mistral3CausalModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+
+	for _, t := range ts {
+		if !strings.HasPrefix(t.Name(), "v.") {
+			if strings.HasSuffix(t.Name(), ".attn_q.weight") ||
+				strings.HasSuffix(t.Name(), ".attn_k.weight") {
+				t.SetRepacker(p.repack)
+			}
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *mistral3CausalModel) Replacements() []string {
+	return []string{
+		"model.norm", "output_norm",
+		"model.", "",
+		"layers", "blk",
+		"transformer.layers", "blk",
+		"vision_tower", "v",
+		"ln_pre", "encoder_norm",
+		"input_layernorm", "attn_norm",
+		"post_attention_layernorm", "ffn_norm",
+		"embed_tokens", "token_embd",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.down_proj", "ffn_down",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.up_proj", "ffn_up",
+		"attention.q_proj", "attn_q",
+		"attention.k_proj", "attn_k",
+		"attention.v_proj", "attn_v",
+		"attention.o_proj", "attn_output",
+		"attention_norm", "attn_norm",
+		"feed_forward.gate_proj", "ffn_gate",
+		"feed_forward.down_proj", "ffn_down",
+		"feed_forward.up_proj", "ffn_up",
+		"multi_modal_projector", "mm",
+		"ffn_norm", "ffn_norm",
+		"lm_head", "output",
+	}
+}
+
+func (p *mistral3CausalModel) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	var dims []int
+	for _, dim := range shape {
+		dims = append(dims, int(dim))
+	}
+
+	var heads uint32
+	if strings.HasSuffix(name, ".attn_q.weight") {
+		heads = p.NumAttentionHeads
+	} else if strings.HasSuffix(name, ".attn_k.weight") {
+		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+	} else {
+		return nil, fmt.Errorf("unknown tensor for repack: %s", name)
+	}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+	if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}

convert/convert_nomicbert.go

@@ -0,0 +1,213 @@
+package convert
+
+import (
+	"cmp"
+	"encoding/json"
+	"io/fs"
+	"path/filepath"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type nomicbertModel struct {
+	ModelParameters
+	NLayers               uint32  `json:"n_layers"`
+	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	LayerNormEPS          float32 `json:"layer_norm_eps"`
+	LayerNormEpsilon      float32 `json:"layer_norm_epsilon"`
+	RopeFreqBase          float32 `json:"rope_theta"`
+	normalizeEmbeddings   bool
+	PoolingType           uint32
+
+	// MoE parameters (only present in v2 models)
+	NumExperts      uint32 `json:"num_local_experts"`
+	NumExpertsUsed  uint32 `json:"num_experts_per_tok"`
+	MoEEveryNLayers uint32 `json:"moe_every_n_layers"`
+}
+
+var (
+	_ ModelConverter = (*nomicbertModel)(nil)
+	_ moreParser     = (*nomicbertModel)(nil)
+)
+
+func (p *nomicbertModel) parseMore(fsys fs.FS) error {
+	bts, err := fs.ReadFile(fsys, "modules.json")
+	if err != nil {
+		return err
+	}
+
+	var modules []struct {
+		Type string `json:"type"`
+		Path string `json:"path"`
+	}
+
+	if err := json.Unmarshal(bts, &modules); err != nil {
+		return err
+	}
+
+	var pooling string
+	for _, m := range modules {
+		switch m.Type {
+		case "sentence_transformers.models.Pooling":
+			pooling = m.Path
+		case "sentence_transformers.models.Normalize":
+			p.normalizeEmbeddings = true
+		}
+	}
+
+	if pooling != "" {
+		bts, err := fs.ReadFile(fsys, filepath.Join(pooling, "config.json"))
+		if err != nil {
+			return err
+		}
+
+		var pc struct {
+			PoolingModeCLSToken   bool `json:"pooling_mode_cls_token"`
+			PoolingModeMeanTokens bool `json:"pooling_mode_mean_tokens"`
+		}
+
+		if err := json.Unmarshal(bts, &pc); err != nil {
+			return err
+		}
+
+		if pc.PoolingModeMeanTokens {
+			p.PoolingType = 1
+		} else if pc.PoolingModeCLSToken {
+			p.PoolingType = 2
+		}
+	}
+
+	return nil
+}
+
+func (p *nomicbertModel) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+
+	// Determine architecture based on MoE parameters (following qwen3 pattern)
+	arch := "nomic-bert"
+	if p.MoEEveryNLayers > 0 {
+		arch += "-moe"
+	}
+
+	kv["general.architecture"] = arch
+	kv["attention.causal"] = false
+	kv["pooling_type"] = p.PoolingType
+	kv["normalize_embeddings"] = p.normalizeEmbeddings
+
+	kv["block_count"] = cmp.Or(p.NLayers, p.NumHiddenLayers)
+
+	if contextLength := p.MaxPositionEmbeddings; contextLength > 0 {
+		kv["context_length"] = contextLength
+	}
+
+	if embeddingLength := p.HiddenSize; embeddingLength > 0 {
+		kv["embedding_length"] = p.HiddenSize
+	}
+
+	if feedForwardLength := p.IntermediateSize; feedForwardLength > 0 {
+		kv["feed_forward_length"] = p.IntermediateSize
+	}
+
+	if headCount := p.NumAttentionHeads; headCount > 0 {
+		kv["attention.head_count"] = p.NumAttentionHeads
+	}
+
+	if kvHeadCount := p.NumKeyValueHeads; kvHeadCount > 0 {
+		kv["attention.head_count_kv"] = p.NumKeyValueHeads
+	}
+
+	if layerNormEpsilon := cmp.Or(p.LayerNormEPS, p.LayerNormEpsilon); layerNormEpsilon > 0 {
+		kv["attention.layer_norm_epsilon"] = layerNormEpsilon
+	}
+
+	if p.RopeFreqBase > 0 {
+		kv["rope.freq_base"] = p.RopeFreqBase
+	}
+
+	// MoE specific parameters (only if MoE is enabled)
+	if p.NumExperts > 0 {
+		kv["expert_count"] = p.NumExperts
+	}
+
+	if p.NumExpertsUsed > 0 {
+		kv["expert_used_count"] = p.NumExpertsUsed
+	}
+
+	if p.MoEEveryNLayers > 0 {
+		kv["moe_every_n_layers"] = p.MoEEveryNLayers
+	}
+
+	kv["tokenizer.ggml.model"] = "bert"
+	kv["tokenizer.ggml.token_type_count"] = uint32(2)
+
+	// convert to phantom space tokens
+	for i, e := range t.Tokens {
+		switch {
+		case strings.HasPrefix(e, "[") && strings.HasSuffix(e, "]"):
+			// noop - keep special tokens as-is
+		case strings.HasPrefix(e, "##"):
+			t.Tokens[i] = e[2:]
+		default:
+			t.Tokens[i] = "\u2581" + e
+		}
+	}
+
+	kv["tokenizer.ggml.tokens"] = t.Tokens
+
+	return kv
+}
+
+func (p *nomicbertModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	out := make([]*ggml.Tensor, 0, len(ts))
+	for _, t := range ts {
+		if slices.Contains([]string{
+			"embeddings.position_ids",
+			"pooler.dense.weight",
+			"pooler.dense.bias",
+		}, t.Name()) {
+			continue
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (nomicbertModel) Replacements() []string {
+	return []string{
+		"encoder.layer", "blk",
+		"encoder.layers", "blk",
+		"embeddings.word_embeddings", "token_embd",
+		"embeddings.token_type_embeddings", "token_types",
+		"embeddings.LayerNorm", "token_embd_norm",
+
+		"attention.self.qkv", "attn_qkv",
+
+		"attention.output.dense", "attn_output",
+		"attention.output.LayerNorm", "attn_output_norm",
+
+		"mlp.up", "ffn_up",
+		"mlp.down", "ffn_down",
+
+		"mlp.router", "ffn_gate_inp",
+		"mlp.experts.up", "ffn_up_exps",
+		"mlp.experts.down", "ffn_down_exps",
+
+		"intermediate.dense", "ffn_up",
+		"output.dense", "ffn_down",
+		"output.LayerNorm", "layer_output_norm",
+	}
+}

convert/convert_olmo.go

@@ -0,0 +1,117 @@
+package convert
+
+import (
+	"cmp"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type ropeScaling struct {
+	Factor                    float32 `json:"factor"`
+	OriginalMaxPositionEmbeds uint32  `json:"original_max_position_embeddings"`
+	AttentionFactor           float32 `json:"attention_factor"`
+	BetaFast                  float32 `json:"beta_fast"`
+	BetaSlow                  float32 `json:"beta_slow"`
+	RopeType                  string  `json:"rope_type"`
+	ExtrapolationFactor       float32 `json:"extrapolation_factor"`
+}
+
+type olmoModel struct {
+	ModelParameters
+
+	HiddenSize            uint32       `json:"hidden_size"`
+	NumHiddenLayers       uint32       `json:"num_hidden_layers"`
+	IntermediateSize      uint32       `json:"intermediate_size"`
+	NumAttentionHeads     uint32       `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32       `json:"num_key_value_heads"`
+	MaxPositionEmbeddings uint32       `json:"max_position_embeddings"`
+	RMSNormEPS            float32      `json:"rms_norm_eps"`
+	RopeTheta             float32      `json:"rope_theta"`
+	RopeScaling           *ropeScaling `json:"rope_scaling"`
+	SlidingWindow         uint32       `json:"sliding_window"`
+	LayerTypes            []string     `json:"layer_types"`
+}
+
+var _ ModelConverter = (*olmoModel)(nil)
+
+func (p *olmoModel) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "olmo3"
+	kv["olmo3.block_count"] = p.NumHiddenLayers
+	kv["olmo3.context_length"] = p.MaxPositionEmbeddings
+	kv["olmo3.embedding_length"] = p.HiddenSize
+	kv["olmo3.feed_forward_length"] = p.IntermediateSize
+	kv["olmo3.attention.head_count"] = p.NumAttentionHeads
+	kv["olmo3.attention.head_count_kv"] = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+
+	if p.RopeTheta > 0 {
+		kv["olmo3.rope.freq_base"] = p.RopeTheta
+	}
+
+	if p.RopeScaling != nil {
+		if p.RopeScaling.Factor > 0 {
+			kv["olmo3.rope.scaling.factor"] = p.RopeScaling.Factor
+		}
+		if p.RopeScaling.OriginalMaxPositionEmbeds > 0 {
+			kv["olmo3.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeds
+		}
+		if p.RopeScaling.AttentionFactor > 0 {
+			kv["olmo3.rope.scaling.attn_factor"] = p.RopeScaling.AttentionFactor
+		}
+		if p.RopeScaling.RopeType != "" {
+			kv["olmo3.rope.scaling.type"] = p.RopeScaling.RopeType
+		}
+	}
+
+	if p.RMSNormEPS > 0 {
+		kv["olmo3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	}
+
+	if p.SlidingWindow > 0 {
+		kv["olmo3.attention.sliding_window"] = p.SlidingWindow
+	}
+
+	if len(p.LayerTypes) > 0 {
+		slidingPattern := make([]bool, len(p.LayerTypes))
+		for i, layerType := range p.LayerTypes {
+			slidingPattern[i] = (layerType == "sliding_attention")
+		}
+		kv["olmo3.attention.sliding_window_pattern"] = slidingPattern
+	}
+
+	return kv
+}
+
+func (p *olmoModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	out := make([]*ggml.Tensor, 0, len(ts))
+	for _, t := range ts {
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *olmoModel) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.layers", "blk",
+		"model.norm", "output_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"self_attn.q_norm", "attn_q_norm",
+		"self_attn.k_norm", "attn_k_norm",
+		"post_attention_layernorm", "post_attention_norm",
+		"post_feedforward_layernorm", "post_ffw_norm",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+	}
+}

docs/api.md

@@ -50,7 +50,7 @@ Generate a response for a given prompt with a provided model. This is a streamin
 Advanced parameters (optional):
 
 - `format`: the format to return a response in. Format can be `json` or a JSON schema
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
 - `system`: system message to (overrides what is defined in the `Modelfile`)
 - `template`: the prompt template to use (overrides what is defined in the `Modelfile`)
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
@@ -507,7 +507,7 @@ The `message` object has the following fields:
 Advanced parameters (optional):
 
 - `format`: the format to return a response in. Format can be `json` or a JSON schema.
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
@@ -1189,7 +1189,7 @@ If you are creating a model from a safetensors directory or from a GGUF file, yo
 - `template`: (optional) the prompt template for the model
 - `license`: (optional) a string or list of strings containing the license or licenses for the model
 - `system`: (optional) a string containing the system prompt for the model
-- `parameters`: (optional) a dictionary of parameters for the model (see [Modelfile](./modelfile.md#valid-parameters-and-values) for a list of parameters)
+- `parameters`: (optional) a dictionary of parameters for the model (see [Modelfile](./modelfile.mdx#valid-parameters-and-values) for a list of parameters)
 - `messages`: (optional) a list of message objects used to create a conversation
 - `stream`: (optional) if `false` the response will be returned as a single response object, rather than a stream of objects
 - `quantize` (optional): quantize a non-quantized (e.g. float16) model
@@ -1698,7 +1698,7 @@ Generate embeddings from a model
 Advanced parameters:
 
 - `truncate`: truncates the end of each input to fit within context length. Returns error if `false` and context length is exceeded. Defaults to `true`
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 - `dimensions`: number of dimensions for the embedding
 
@@ -1817,7 +1817,7 @@ Generate embeddings from a model
 
 Advanced parameters:
 
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
 ### Examples

docs/tools/extract-examples/README.md

@@ -0,0 +1,46 @@
+# extract-examples
+
+Extracts code examples from MDX files to a temp directory so you can run them.
+
+## Usage
+
+```shell
+go run docs/tools/extract-examples/main.go <mdx-file>
+```
+
+## Example
+
+```shell
+go run docs/tools/extract-examples/main.go docs/api/openai-compatibility.mdx
+```
+
+Output:
+
+```
+Extracting code examples to: /var/folders/vq/wfm2g6k917d3ldzpjdxc8ph00000gn/T/mdx-examples-3271754368
+
+  - 01_basic.py
+  - 01_basic.js
+  - 01_basic.sh
+  - 02_responses.py
+  - 02_responses.js
+  - 02_responses.sh
+  - 03_vision.py
+  - 03_vision.js
+  - 03_vision.sh
+
+Extracted 9 file(s) to /var/folders/vq/wfm2g6k917d3ldzpjdxc8ph00000gn/T/mdx-examples-3271754368
+
+To run examples:
+
+  cd /var/folders/vq/wfm2g6k917d3ldzpjdxc8ph00000gn/T/mdx-examples-3271754368
+  npm install   # for JS examples
+
+then run individual files with `node file.js`, `python file.py`, `bash file.sh`
+```
+
+## How it works
+
+- Parses MDX files looking for fenced code blocks with filenames (e.g., ` ```python basic.py `)
+- Groups examples by their `<CodeGroup>` and prefixes filenames with `01_`, `02_`, etc.
+- Writes all extracted files to a temp directory

docs/tools/extract-examples/main.go

@@ -0,0 +1,137 @@
+package main
+
+import (
+	"bufio"
+	"fmt"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+)
+
+func main() {
+	if len(os.Args) < 2 {
+		fmt.Fprintln(os.Stderr, "Usage: go run extract-examples.go <mdx-file>")
+		os.Exit(1)
+	}
+
+	mdxFile := os.Args[1]
+
+	f, err := os.Open(mdxFile)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Error: %v\n", err)
+		os.Exit(1)
+	}
+	defer f.Close()
+
+	// Create temp directory
+	tempDir, err := os.MkdirTemp("", "mdx-examples-*")
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Error creating temp dir: %v\n", err)
+		os.Exit(1)
+	}
+
+	fmt.Printf("Extracting code examples to: %s\n\n", tempDir)
+
+	// Patterns
+	codeBlockStart := regexp.MustCompile("^```([a-zA-Z0-9_-]+)\\s+([^\\s]+)$")
+	codeGroupStart := regexp.MustCompile("^<CodeGroup")
+	codeGroupEnd := regexp.MustCompile("^</CodeGroup>")
+
+	scanner := bufio.NewScanner(f)
+	inCodeBlock := false
+	inCodeGroup := false
+	var currentFile string
+	var content strings.Builder
+	count := 0
+	codeGroupNum := 0
+
+	for scanner.Scan() {
+		line := scanner.Text()
+
+		// Track CodeGroup boundaries
+		if codeGroupStart.MatchString(line) {
+			inCodeGroup = true
+			codeGroupNum++
+			continue
+		}
+		if codeGroupEnd.MatchString(line) {
+			inCodeGroup = false
+			continue
+		}
+
+		if inCodeBlock {
+			if line == "```" {
+				// End of code block - write file
+				if currentFile != "" {
+					outPath := filepath.Join(tempDir, currentFile)
+					if err := os.WriteFile(outPath, []byte(content.String()), 0o644); err != nil {
+						fmt.Fprintf(os.Stderr, "Error writing %s: %v\n", currentFile, err)
+					} else {
+						fmt.Printf("  - %s\n", currentFile)
+						count++
+					}
+				}
+				inCodeBlock = false
+				currentFile = ""
+				content.Reset()
+			} else {
+				content.WriteString(line)
+				content.WriteString("\n")
+			}
+		} else {
+			if matches := codeBlockStart.FindStringSubmatch(line); matches != nil {
+				inCodeBlock = true
+				filename := matches[2]
+				// Prefix with CodeGroup number if inside a CodeGroup
+				if inCodeGroup {
+					currentFile = fmt.Sprintf("%02d_%s", codeGroupNum, filename)
+				} else {
+					currentFile = filename
+				}
+				content.Reset()
+			}
+		}
+	}
+
+	if err := scanner.Err(); err != nil {
+		fmt.Fprintf(os.Stderr, "Error reading file: %v\n", err)
+		os.Exit(1)
+	}
+
+	// Write package.json for JavaScript dependencies
+	packageJSON := `{
+  "name": "mdx-examples",
+  "type": "module",
+  "dependencies": {
+    "openai": "^4",
+    "ollama": "^0.5"
+  }
+}
+`
+	if err := os.WriteFile(filepath.Join(tempDir, "package.json"), []byte(packageJSON), 0o644); err != nil {
+		fmt.Fprintf(os.Stderr, "Error writing package.json: %v\n", err)
+	}
+
+	// Write pyproject.toml for Python dependencies
+	pyprojectTOML := `[project]
+name = "mdx-examples"
+version = "0.0.0"
+dependencies = [
+    "openai",
+    "ollama",
+]
+`
+	if err := os.WriteFile(filepath.Join(tempDir, "pyproject.toml"), []byte(pyprojectTOML), 0o644); err != nil {
+		fmt.Fprintf(os.Stderr, "Error writing pyproject.toml: %v\n", err)
+	}
+
+	fmt.Printf("\n")
+	fmt.Printf("Extracted %d file(s) to %s\n", count, tempDir)
+	fmt.Printf("\n")
+	fmt.Printf("To run examples:\n")
+	fmt.Printf("\n")
+	fmt.Printf("  cd %s\n  npm install   # for JS examples\n", tempDir)
+	fmt.Printf("\n")
+	fmt.Printf("then run individual files with `node file.js`, `python file.py`, `bash file.sh`\n")
+}

fs/ggml/ggml.go

@@ -13,6 +13,7 @@ import (
 
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/util/bufioutil"
+	"github.com/ollama/ollama/ml"
 )
 
 type GGML struct {
@@ -240,18 +241,20 @@ func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
 
 func (kv KV) OllamaEngineRequired() bool {
 	return slices.Contains([]string{
+		"bert",
+		"deepseek2",
+		"deepseekocr",
 		"gemma3",
 		"gemma3n",
 		"gptoss", "gpt-oss",
 		"llama4",
 		"mistral3",
 		"mllama",
+		"nomic-bert",
+		"olmo3",
 		"qwen25vl",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
-		"deepseekocr",
-		"deepseek2",
-		"nomic-bert",
 	}, kv.Architecture())
 }
 
@@ -550,7 +553,7 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, error) {
 	}, nil
 }
 
-func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string, useFlashAttention bool) (kv []uint64, partialOffload, fullOffload uint64) {
+func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string, useFlashAttention ml.FlashAttentionType) (kv []uint64, partialOffload, fullOffload uint64) {
 	context *= uint64(numParallel)
 
 	embedding := f.KV().EmbeddingLength()
@@ -791,7 +794,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 		}
 
 		partialOffload = 2 * f.KV().HeadCountMax() / cmp.Or(f.KV().HeadCountKVMin(), 1) * kvTotal / 6
-		if useFlashAttention {
+		if useFlashAttention == ml.FlashAttentionEnabled {
 			// rough estimate of graph size with flash attention on
 			partialOffload = (4*uint64(numParallel) + context>>10 + 110) * format.MebiByte
 		}
@@ -809,6 +812,14 @@ func (f GGML) SupportsKVCacheType(cacheType string) bool {
 	return slices.Contains([]string{"q8_0", "q4_0"}, cacheType)
 }
 
+// KVCacheTypeIsQuantized checks if the requested cache type is a quantized type
+func (f GGML) KVCacheTypeIsQuantized(cacheType string) bool {
+	if cacheType == "" || cacheType == "f16" || cacheType == "f32" || cacheType == "bf16" {
+		return false
+	}
+	return true
+}
+
 // SupportsFlashAttention checks if the model supports flash attention
 func (f GGML) SupportsFlashAttention() bool {
 	_, isEmbedding := f.KV()[fmt.Sprintf("%s.pooling_type", f.KV().Architecture())]
@@ -829,9 +840,11 @@ func (f GGML) SupportsFlashAttention() bool {
 // FlashAttention checks if the model should enable flash attention
 func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
+		"bert",
 		"gemma3",
 		"gptoss", "gpt-oss",
 		"mistral3",
+		"olmo3",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
 	}, f.KV().String("general.architecture"))

fs/ggml/gguf.go

@@ -597,6 +597,10 @@ func ggufWriteKV(ws io.WriteSeeker, arch, k string, v any) error {
 
 	var err error
 	switch v := v.(type) {
+	case int32:
+		err = writeGGUF(ws, ggufTypeInt32, v)
+	case int64:
+		err = writeGGUF(ws, ggufTypeInt64, v)
 	case uint32, FileType:
 		err = writeGGUF(ws, ggufTypeUint32, v)
 	case uint64:
@@ -611,6 +615,10 @@ func ggufWriteKV(ws io.WriteSeeker, arch, k string, v any) error {
 		err = writeGGUFArray(ws, ggufTypeInt32, v)
 	case *array[int32]:
 		err = writeGGUFArray(ws, ggufTypeInt32, v.values)
+	case []int64:
+		err = writeGGUFArray(ws, ggufTypeInt64, v)
+	case *array[int64]:
+		err = writeGGUFArray(ws, ggufTypeInt64, v.values)
 	case []uint32:
 		err = writeGGUFArray(ws, ggufTypeUint32, v)
 	case *array[uint32]:

fs/ggml/gguf_test.go

@@ -42,6 +42,10 @@ func TestWriteGGUF(t *testing.T) {
 				"general.architecture": "test",
 				"general.alignment":    uint32(16),
 				"test.key":             "value",
+				"test.int32_key":       int32(-42),
+				"test.int64_key":       int64(-9223372036854775808),
+				"test.int32_array":     []int32{-1, 0, 1, 2147483647, -2147483648},
+				"test.int64_array":     []int64{-1, 0, 1, 9223372036854775807, -9223372036854775808},
 				"attention.key":        "value2",
 				"tokenizer.key":        "value3",
 				"adapter.key":          "value4",
@@ -55,7 +59,7 @@ func TestWriteGGUF(t *testing.T) {
 			}
 			defer r.Close()
 
-			ff, err := Decode(r, 0)
+			ff, err := Decode(r, -1)
 			if err != nil {
 				t.Fatal(err)
 			}
@@ -65,15 +69,19 @@ func TestWriteGGUF(t *testing.T) {
 				"general.alignment":       uint32(16),
 				"general.parameter_count": uint64(54),
 				"test.key":                "value",
+				"test.int32_key":          int32(-42),
+				"test.int64_key":          int64(-9223372036854775808),
+				"test.int32_array":        &array[int32]{size: 5, values: []int32{-1, 0, 1, 2147483647, -2147483648}},
+				"test.int64_array":        &array[int64]{size: 5, values: []int64{-1, 0, 1, 9223372036854775807, -9223372036854775808}},
 				"test.attention.key":      "value2",
 				"tokenizer.key":           "value3",
 				"adapter.key":             "value4",
-			}, ff.KV()); diff != "" {
+			}, ff.KV(), cmp.AllowUnexported(array[int32]{}, array[int64]{})); diff != "" {
 				t.Errorf("Mismatch (-want +got):\n%s", diff)
 			}
 
 			if diff := cmp.Diff(Tensors{
-				Offset: 800,
+				Offset: 992,
 				items: []*Tensor{
 					{Name: "blk.0.attn_k.weight", Offset: 0, Shape: []uint64{2, 3}},
 					{Name: "blk.0.attn_norm.weight", Offset: 32, Shape: []uint64{2, 3}},

integration/embed_test.go

@@ -4,7 +4,9 @@ package integration
 
 import (
 	"context"
+	"errors"
 	"math"
+	"strings"
 	"testing"
 	"time"
 
@@ -204,8 +206,8 @@ func TestAllMiniLMEmbed(t *testing.T) {
 		t.Fatalf("expected %v, got %v (similarity: %f)", expected[0:5], res.Embeddings[0][0:5], sim)
 	}
 
-	if res.PromptEvalCount != 6 {
+	if res.PromptEvalCount != 8 {
-		t.Fatalf("expected 6 prompt tokens, got %d", res.PromptEvalCount)
+		t.Fatalf("expected 8 prompt tokens, got %d", res.PromptEvalCount)
 	}
 }
 
@@ -251,8 +253,8 @@ func TestAllMiniLMBatchEmbed(t *testing.T) {
 		t.Fatalf("expected %v, got %v (similarity: %f)", expected[1][0:5], res.Embeddings[1][0:5], sim)
 	}
 
-	if res.PromptEvalCount != 12 {
+	if res.PromptEvalCount != 16 {
-		t.Fatalf("expected 12 prompt tokens, got %d", res.PromptEvalCount)
+		t.Fatalf("expected 16 prompt tokens, got %d", res.PromptEvalCount)
 	}
 }
 
@@ -275,7 +277,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 	cases := []struct {
 		name    string
 		request api.EmbedRequest
-		check   func(*api.EmbedResponse, error)
+		check   func(*testing.T, *api.EmbedResponse, error)
 	}{
 		{
 			name: "target truncation",
@@ -283,7 +285,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Model: "all-minilm",
 				Input: "why",
 			},
-			check: func(got *api.EmbedResponse, err error) {
+			check: func(t *testing.T, got *api.EmbedResponse, err error) {
 				if err != nil {
 					t.Fatal(err)
 				}
@@ -300,10 +302,11 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Input:   "why is the sky blue?",
 				Options: map[string]any{"num_ctx": 3},
 			},
-			check: func(got *api.EmbedResponse, err error) {
+			check: func(t *testing.T, got *api.EmbedResponse, err error) {
 				if err != nil {
 					t.Fatal(err)
 				}
+				t.Logf("PromptEvalCount: want=%d got=%d", want.PromptEvalCount, got.PromptEvalCount)
 				if diff := cmp.Diff(want.Embeddings[0], got.Embeddings[0]); diff != "" {
 					t.Errorf("embedding mismatch (-want +got):\n%s", diff)
 				}
@@ -317,10 +320,11 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Truncate: &truncTrue,
 				Options:  map[string]any{"num_ctx": 3},
 			},
-			check: func(got *api.EmbedResponse, err error) {
+			check: func(t *testing.T, got *api.EmbedResponse, err error) {
 				if err != nil {
 					t.Fatal(err)
 				}
+				t.Logf("PromptEvalCount: want=%d got=%d", want.PromptEvalCount, got.PromptEvalCount)
 				if diff := cmp.Diff(want.Embeddings[0], got.Embeddings[0]); diff != "" {
 					t.Errorf("embedding mismatch (-want +got):\n%s", diff)
 				}
@@ -334,21 +338,21 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Truncate: &truncFalse,
 				Options:  map[string]any{"num_ctx": 3},
 			},
-			check: func(res *api.EmbedResponse, err error) {
+			check: func(t *testing.T, res *api.EmbedResponse, err error) {
-				if err.Error() != "input exceeds maximum context length" {
+				if err.Error() != "the input length exceeds the context length" {
 					t.Fatalf("expected truncation error, got: %v", err)
 				}
 			},
 		},
 		{
-			name: "input after truncate error",
+			name: "input after truncate error with context length of 1",
 			request: api.EmbedRequest{
 				Model:    "all-minilm",
 				Input:    "why is the sky blue?",
 				Truncate: &truncTrue,
 				Options:  map[string]any{"num_ctx": 1},
 			},
-			check: func(res *api.EmbedResponse, err error) {
+			check: func(t *testing.T, res *api.EmbedResponse, err error) {
 				if err.Error() != "input after truncation exceeds maximum context length" {
 					t.Fatalf("expected truncation error, got: %v", err)
 				}
@@ -362,7 +366,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Truncate: &truncTrue,
 				Options:  map[string]any{"num_ctx": 0},
 			},
-			check: func(res *api.EmbedResponse, err error) {
+			check: func(t *testing.T, res *api.EmbedResponse, err error) {
 				if err.Error() != "input after truncation exceeds maximum context length" {
 					t.Fatalf("expected truncation error, got: %v", err)
 				}
@@ -375,7 +379,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Input:   "why is the sky blue? Why is the sky blue? hi there my",
 				Options: map[string]any{"num_ctx": 16},
 			},
-			check: func(res *api.EmbedResponse, err error) {
+			check: func(t *testing.T, res *api.EmbedResponse, err error) {
 				if err != nil {
 					t.Fatal(err)
 				}
@@ -385,7 +389,8 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 
 	for _, req := range cases {
 		t.Run(req.name, func(t *testing.T) {
-			req.check(embedTestHelper(ctx, client, t, req.request))
+			resp, err := embedTestHelper(ctx, client, t, req.request)
+			req.check(t, resp, err)
 		})
 	}
 }
@@ -409,3 +414,230 @@ func embedTestHelper(ctx context.Context, client *api.Client, t *testing.T, req
 
 	return client.Embed(ctx, &req)
 }
+
+func TestEmbedTruncation(t *testing.T) {
+	// Use test deadline if set, otherwise default to 2 minutes
+	timeout := 2 * time.Minute
+	if deadline, ok := t.Deadline(); ok {
+		timeout = time.Until(deadline) - 10*time.Second // Reserve 10s buffer
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), timeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	for _, model := range libraryEmbedModels {
+		model := model
+		t.Run(model, func(t *testing.T) {
+			// Check if we're running out of time (reserve 20s for current model)
+			if deadline, ok := t.Deadline(); ok && time.Until(deadline) < 20*time.Second {
+				t.Skip("skipping remaining tests to avoid timeout")
+			}
+
+			// Give each model its own budget to account for first-time pulls/loads
+			mctx, mcancel := context.WithTimeout(ctx, 3*time.Minute)
+			defer mcancel()
+
+			t.Run("truncation batch", func(t *testing.T) {
+				truncTrue := true
+				req := api.EmbedRequest{
+					Model:    model,
+					Input:    []string{"short", strings.Repeat("long ", 100), "medium text"},
+					Truncate: &truncTrue,
+					Options:  map[string]any{"num_ctx": 30},
+				}
+
+				res, err := embedTestHelper(mctx, client, t, req)
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				if len(res.Embeddings) != 3 {
+					t.Fatalf("expected 3 embeddings, got %d", len(res.Embeddings))
+				}
+
+				if res.PromptEvalCount > 90 {
+					t.Fatalf("expected tokens <= 90 (3 × 30 max), got %d", res.PromptEvalCount)
+				}
+			})
+
+			t.Run("runner token count accuracy", func(t *testing.T) {
+				baseline := api.EmbedRequest{Model: model, Input: "test"}
+				baseRes, err := embedTestHelper(mctx, client, t, baseline)
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				batch := api.EmbedRequest{
+					Model: model,
+					Input: []string{"test", "test", "test"},
+				}
+				batchRes, err := embedTestHelper(mctx, client, t, batch)
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				expectedCount := baseRes.PromptEvalCount * 3
+				if batchRes.PromptEvalCount < expectedCount-2 || batchRes.PromptEvalCount > expectedCount+2 {
+					t.Fatalf("expected ~%d tokens (3 × %d), got %d",
+						expectedCount, baseRes.PromptEvalCount, batchRes.PromptEvalCount)
+				}
+			})
+		})
+	}
+}
+
+// TestEmbedLargeInput tests that embedding models can handle large inputs that would exceed typical batch sizes.
+func TestEmbedLargeInput(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	for _, model := range libraryEmbedModels {
+		model := model
+		t.Run(model, func(t *testing.T) {
+			mctx, mcancel := context.WithTimeout(ctx, 2*time.Minute)
+			defer mcancel()
+
+			// Test with progressively larger inputs
+			testCases := []struct {
+				name       string
+				inputWords int
+			}{
+				{"medium_input_256_words", 256},
+				{"large_input_512_words", 512},
+				{"very_large_input_800_words", 800},
+			}
+
+			for _, tc := range testCases {
+				t.Run(tc.name, func(t *testing.T) {
+					words := make([]string, tc.inputWords)
+					for i := range words {
+						words[i] = "word"
+					}
+					input := strings.Join(words, " ")
+
+					req := api.EmbedRequest{
+						Model:     model,
+						Input:     input,
+						KeepAlive: &api.Duration{Duration: 30 * time.Second},
+					}
+
+					res, err := embedTestHelper(mctx, client, t, req)
+					if err != nil {
+						t.Fatalf("embedding failed for %d words: %v", tc.inputWords, err)
+					}
+
+					if len(res.Embeddings) != 1 {
+						t.Fatalf("expected 1 embedding, got %d", len(res.Embeddings))
+					}
+
+					if len(res.Embeddings[0]) == 0 {
+						t.Fatal("expected non-empty embedding")
+					}
+
+					t.Logf("Successfully embedded %d words (%d tokens)", tc.inputWords, res.PromptEvalCount)
+				})
+			}
+		})
+	}
+}
+
+// TestEmbedStatusCode tests that errors from the embedding endpoint
+// properly preserve their HTTP status codes when returned to the client.
+// This test specifically checks the error handling path in EmbedHandler
+// where api.StatusError errors should maintain their original status code.
+func TestEmbedStatusCode(t *testing.T) {
+	// Use test deadline if set, otherwise default to 2 minutes
+	timeout := 2 * time.Minute
+	if deadline, ok := t.Deadline(); ok {
+		timeout = time.Until(deadline) - 10*time.Second // Reserve 10s buffer
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), timeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	for _, model := range libraryEmbedModels {
+		model := model
+		t.Run(model, func(t *testing.T) {
+			// Check if we're running out of time (reserve 20s for current model)
+			if deadline, ok := t.Deadline(); ok && time.Until(deadline) < 20*time.Second {
+				t.Skip("skipping remaining tests to avoid timeout")
+			}
+
+			mctx, mcancel := context.WithTimeout(ctx, 3*time.Minute)
+			defer mcancel()
+
+			// Pull the model if needed
+			if err := PullIfMissing(mctx, client, model); err != nil {
+				t.Fatal(err)
+			}
+
+			t.Run("truncation error status code", func(t *testing.T) {
+				truncFalse := false
+				longInput := strings.Repeat("word ", 100)
+
+				req := api.EmbedRequest{
+					Model:    model,
+					Input:    longInput,
+					Truncate: &truncFalse,
+					Options:  map[string]any{"num_ctx": 10},
+				}
+
+				_, err := embedTestHelper(mctx, client, t, req)
+				if err == nil {
+					t.Fatal("expected error when truncate=false with long input")
+				}
+
+				// Check that it's a StatusError with the correct status code
+				var statusErr api.StatusError
+				if !errors.As(err, &statusErr) {
+					t.Fatalf("expected api.StatusError, got %T: %v", err, err)
+				}
+
+				// The error should be a 4xx client error (likely 400 Bad Request)
+				// not a 500 Internal Server Error
+				if statusErr.StatusCode < 400 || statusErr.StatusCode >= 500 {
+					t.Errorf("expected 4xx status code, got %d", statusErr.StatusCode)
+				}
+
+				// Verify the error message is meaningful
+				if !strings.Contains(err.Error(), "context length") {
+					t.Errorf("expected error message to mention context length, got: %v", err)
+				}
+			})
+
+			t.Run("batch truncation error status code", func(t *testing.T) {
+				truncFalse := false
+				req := api.EmbedRequest{
+					Model: model,
+					Input: []string{
+						"short input",
+						strings.Repeat("very long input ", 100),
+						"another short input",
+					},
+					Truncate: &truncFalse,
+					Options:  map[string]any{"num_ctx": 10},
+				}
+
+				_, err := embedTestHelper(mctx, client, t, req)
+				if err == nil {
+					t.Fatal("expected error when one input exceeds context with truncate=false")
+				}
+
+				// Check that it's a StatusError with the correct status code
+				var statusErr api.StatusError
+				if !errors.As(err, &statusErr) {
+					t.Fatalf("expected api.StatusError, got %T: %v", err, err)
+				}
+
+				// The error should be a 4xx client error, not a 500 Internal Server Error
+				if statusErr.StatusCode < 400 || statusErr.StatusCode >= 500 {
+					t.Errorf("expected 4xx status code, got %d", statusErr.StatusCode)
+				}
+			})
+		})
+	}
+}

kvcache/causal.go

@@ -140,10 +140,6 @@ func (c *Causal) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity
 		c.config.CachePadding = 1
 	}
 
-	if c.config.MaskBatchPadding == 0 {
-		c.config.MaskBatchPadding = 1
-	}
-
 	if c.config.MaskDType == ml.DTypeOther {
 		c.config.MaskDType = ml.DTypeF32
 	}
@@ -364,15 +360,12 @@ func roundUp(length, pad int) int {
 // token in the history should apply. This is based on both the sequence and causality (the
 // position of the history is not ahead of the token in the batch).
 func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
-	// Align and pad the two dimensions as required by the backend
-	batchSize := roundUp(c.curBatchSize, c.config.MaskBatchPadding)
-
 	c.curCellRange.min = roundDown(c.curCellRange.min, c.config.CachePadding)
 	c.curCellRange.max = roundUp(c.curCellRange.max+1, c.config.CachePadding) - 1
 
 	length := c.curCellRange.max - c.curCellRange.min + 1
 
-	mask := make([]float32, batchSize*length)
+	mask := make([]float32, c.curBatchSize*length)
 
 	for i := range c.curBatchSize {
 		enabled := !slices.Contains(c.opts.Except, i)
@@ -386,13 +379,7 @@ func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
 		}
 	}
 
-	// Mask out any padding tokens we added. For padding that we added to the cache history, this
+	maskTensor := ctx.Input().FromFloats(mask, length, c.curBatchSize)
-	// has already been masked out because the sequence doesn't match.
-	for i := c.curBatchSize * length; i < len(mask); i++ {
-		mask[i] = float32(math.Inf(-1))
-	}
-
-	maskTensor := ctx.Input().FromFloats(mask, length, batchSize)
 
 	if c.config.MaskDType != ml.DTypeF32 {
 		maskTensor = maskTensor.Cast(ctx, c.config.MaskDType)

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "7f8ef50cce40e3e7e4526a3696cb45658190e69a";
+char const *LLAMA_COMMIT = "ec98e2002";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/.rsync-filter

@@ -17,6 +17,9 @@ include /tools/mtmd/clip.cpp
 include /tools/mtmd/mtmd.cpp
 include /tools/mtmd/mtmd-audio.cpp
 include /tools/mtmd/mtmd-helper.cpp
+include /tools/mtmd/models/
+include /tools/mtmd/models/*.h
+include /tools/mtmd/models/*.cpp
 include /src/
 include /src/llama.*
 include /src/llama-*.*

llama/llama.cpp/common/common.cpp

@@ -694,7 +694,7 @@ bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_over
 
 // Validate if a filename is safe to use
 // To validate a full path, split the path by the OS-specific path separator, and validate each part with this function
-bool fs_validate_filename(const std::string & filename) {
+bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
     if (!filename.length()) {
         // Empty filename invalid
         return false;
@@ -754,10 +754,14 @@ bool fs_validate_filename(const std::string & filename) {
             || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
             || c == 0xFFFD // Replacement Character (UTF-8)
             || c == 0xFEFF // Byte Order Mark (BOM)
-            || c == '/' || c == '\\' || c == ':' || c == '*' // Illegal characters
+            || c == ':' || c == '*' // Illegal characters
             || c == '?' || c == '"' || c == '<' || c == '>' || c == '|') {
             return false;
         }
+        if (!allow_subdirs && (c == '/' || c == '\\')) {
+            // Subdirectories not allowed, reject path separators
+            return false;
+        }
     }
 
     // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
@@ -782,11 +786,29 @@ bool fs_validate_filename(const std::string & filename) {
 #include <iostream>
 
 
+#ifdef _WIN32
+static std::wstring utf8_to_wstring(const std::string & str) {
+    if (str.empty()) {
+        return std::wstring();
+    }
+
+    int size = MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), NULL, 0);
+
+    if (size <= 0) {
+        return std::wstring();
+    }
+
+    std::wstring wstr(size, 0);
+    MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), &wstr[0], size);
+
+    return wstr;
+}
+#endif
+
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
-    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring wpath = utf8_to_wstring(path);
-    std::wstring wpath = converter.from_bytes(path);
 
     // if the path already exists, check whether it's a directory
     const DWORD attributes = GetFileAttributesW(wpath.c_str());
@@ -859,6 +881,11 @@ bool fs_create_directory_with_parents(const std::string & path) {
 #endif // _WIN32
 }
 
+bool fs_is_directory(const std::string & path) {
+    std::filesystem::path dir(path);
+    return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
+}
+
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
     auto ensure_trailing_slash = [](std::string p) {
@@ -893,6 +920,8 @@ std::string fs_get_cache_directory() {
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
         cache_directory = std::getenv("LOCALAPPDATA");
+#elif defined(__EMSCRIPTEN__)
+        GGML_ABORT("not implemented on this platform");
 #else
 #  error Unknown architecture
 #endif
@@ -912,7 +941,7 @@ std::string fs_get_cache_file(const std::string & filename) {
     return cache_directory + filename;
 }
 
-std::vector<common_file_info> fs_list_files(const std::string & path) {
+std::vector<common_file_info> fs_list(const std::string & path, bool include_directories) {
     std::vector<common_file_info> files;
     if (path.empty()) return files;
 
@@ -929,12 +958,20 @@ std::vector<common_file_info> fs_list_files(const std::string & path) {
                 common_file_info info;
                 info.path   = p.string();
                 info.name   = p.filename().string();
+                info.is_dir = false;
                 try {
                     info.size = static_cast<size_t>(std::filesystem::file_size(p));
                 } catch (const std::filesystem::filesystem_error &) {
                     info.size = 0;
                 }
                 files.push_back(std::move(info));
+            } else if (include_directories && std::filesystem::is_directory(p)) {
+                common_file_info info;
+                info.path   = p.string();
+                info.name   = p.filename().string();
+                info.size   = 0; // Directories have no size
+                info.is_dir = true;
+                files.push_back(std::move(info));
             }
         } catch (const std::filesystem::filesystem_error &) {
             // skip entries we cannot inspect
@@ -945,36 +982,71 @@ std::vector<common_file_info> fs_list_files(const std::string & path) {
     return files;
 }
 
+//
+// TTY utils
+//
+
+bool tty_can_use_colors() {
+    // 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;
+}
 
 //
 // Model utils
 //
 
-static inline void common_init_sampler_from_model(
+// TODO: move to common/sampling
+static void common_init_sampler_from_model(
     const llama_model * model,
     common_params_sampling & sparams) {
 
     const uint64_t config = sparams.user_sampling_config;
 
     auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) {
-        if (config & user_config) return;
+        if (config & user_config) {
+            return;
+        }
 
         char buf[64] = {0};
         if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
             char * end = nullptr;
             int32_t v = strtol(buf, &end, 10);
-            if (end && end != buf) dst = v;
+            if (end && end != buf) {
+                dst = v;
+            }
         }
     };
 
     auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
-        if (config & user_config) return;
+        if (config & user_config) {
+            return;
+        }
 
         char buf[128] = {0};
         if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
             char * end = nullptr;
             float v = strtof(buf, &end);
-            if (end && end != buf) dst = v;
+            if (end && end != buf) {
+                dst = v;
+            }
         }
     };
 
@@ -1002,31 +1074,125 @@ static inline void common_init_sampler_from_model(
     get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA),    sparams.mirostat_eta,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA);
 }
 
-struct common_init_result common_init_from_params(common_params & params) {
+struct common_init_result::impl {
-    common_init_result iparams;
+    impl() = default;
+    ~impl() = default;
+
+    llama_model_ptr   model;
+    llama_context_ptr context;
+
+    std::vector<llama_adapter_lora_ptr> lora;
+
+    std::vector<common_sampler_ptr> samplers;
+};
+
+common_init_result::common_init_result(common_params & params) :
+    pimpl(new impl{}) {
     auto mparams = common_model_params_to_llama(params);
+    auto cparams = common_context_params_to_llama(params);
+
+    if (params.fit_params) {
+        LOG_INF("%s: fitting params to device memory, to report bugs during this step use -fit off (or --verbose if you can't)\n", __func__);
+        llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
+            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target, params.fit_params_min_ctx,
+            params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
+    }
 
     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', try reducing --n-gpu-layers if you're running out of VRAM\n",
+        return;
-            __func__, params.model.path.c_str());
-        return iparams;
     }
 
-    common_init_sampler_from_model(model, params.sampling);
+    pimpl->model.reset(model);
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
-    auto cparams = common_context_params_to_llama(params);
+    // updates params.sampling
+    // TODO: fix naming
+    common_init_sampler_from_model(model, params.sampling);
+
+    if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
+        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
+        params.sampling.ignore_eos = false;
+    }
+
+    // initialize once
+    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
+        if (llama_vocab_is_eog(vocab, i)) {
+            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(vocab, i).c_str(), -INFINITY);
+            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
+        }
+    }
+
+    if (params.sampling.ignore_eos) {
+        // add EOG biases to the active set of logit biases
+        params.sampling.logit_bias.insert(
+                params.sampling.logit_bias.end(),
+                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
+    }
+
+    //if (params.sampling.penalty_last_n == -1) {
+    //    LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+    //    params.sampling.penalty_last_n = llama_n_ctx(lctx);
+    //}
+
+    //if (params.sampling.dry_penalty_last_n == -1) {
+    //    LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+    //    params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
+    //}
+
+    pimpl->samplers.resize(cparams.n_seq_max);
+
+    for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
+        pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
+    }
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
-            __func__, params.model.path.c_str());
+        return;
-        llama_model_free(model);
-        return iparams;
     }
 
+    pimpl->context.reset(lctx);
+}
+
+llama_model * common_init_result::model() {
+    return pimpl->model.get();
+}
+
+llama_context * common_init_result::context() {
+    return pimpl->context.get();
+}
+
+common_sampler * common_init_result::sampler(llama_seq_id seq_id) {
+    return pimpl->samplers[seq_id].get();
+}
+
+std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
+    return pimpl->lora;
+}
+
+void common_init_result::free_context() {
+    pimpl->context.reset();
+}
+
+common_init_result_ptr common_init_from_params(common_params & params) {
+    common_init_result_ptr res(new common_init_result(params));
+
+    llama_model * model = res->model();
+    if (model == NULL) {
+        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
+        return res;
+    }
+
+    llama_context * lctx = res->context();
+    if (lctx == NULL) {
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
+        return res;
+    }
+
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
     if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
         LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
@@ -1038,10 +1204,7 @@ struct common_init_result common_init_from_params(common_params & params) {
 
         const auto cvec = common_control_vector_load(params.control_vectors);
         if (cvec.n_embd == -1) {
-            llama_free(lctx);
+            return res;
-            llama_model_free(model);
-
-            return iparams;
         }
 
         int err = llama_apply_adapter_cvec(
@@ -1052,10 +1215,7 @@ struct common_init_result common_init_from_params(common_params & params) {
                 params.control_vector_layer_start,
                 params.control_vector_layer_end);
         if (err) {
-            llama_free(lctx);
+            return res;
-            llama_model_free(model);
-
-            return iparams;
         }
     }
 
@@ -1079,10 +1239,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
 
         if (!ok) {
-            llama_free(lctx);
+            return res;
-            llama_model_free(model);
-
-            return iparams;
         }
     }
 
@@ -1092,9 +1249,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
         if (lora == nullptr) {
             LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
-            llama_free(lctx);
+            return res;
-            llama_model_free(model);
-            return iparams;
         }
 
         char buf[1024];
@@ -1103,43 +1258,13 @@ struct common_init_result common_init_from_params(common_params & params) {
         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
+        res->lora().emplace_back(std::move(lora)); // copy to list of loaded adapters
     }
 
     if (!params.lora_init_without_apply) {
         common_set_adapter_lora(lctx, params.lora_adapters);
     }
 
-    if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
-        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
-        params.sampling.ignore_eos = false;
-    }
-
-    // initialize once
-    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
-        if (llama_vocab_is_eog(vocab, i)) {
-            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
-            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
-        }
-    }
-
-    if (params.sampling.ignore_eos) {
-        // add EOG biases to the active set of logit biases
-        params.sampling.logit_bias.insert(
-                params.sampling.logit_bias.end(),
-                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
-    }
-
-    if (params.sampling.penalty_last_n == -1) {
-        LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-        params.sampling.penalty_last_n = llama_n_ctx(lctx);
-    }
-
-    if (params.sampling.dry_penalty_last_n == -1) {
-        LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-        params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
-    }
-
     if (params.warmup) {
         LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
@@ -1178,12 +1303,11 @@ struct common_init_result common_init_from_params(common_params & params) {
         llama_set_warmup(lctx, false);
     }
 
-    iparams.model.reset(model);
+    return res;
-    iparams.context.reset(lctx);
-
-    return iparams;
 }
 
+common_init_result::~common_init_result() = default;
+
 std::string get_model_endpoint() {
     const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
     // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
@@ -1192,7 +1316,9 @@ std::string get_model_endpoint() {
     std::string model_endpoint = "https://huggingface.co/";
     if (endpoint_env) {
         model_endpoint = endpoint_env;
-        if (model_endpoint.back() != '/') model_endpoint += '/';
+        if (model_endpoint.back() != '/') {
+            model_endpoint += '/';
+        }
     }
     return model_endpoint;
 }

llama/llama.cpp/common/common.h

@@ -12,6 +12,10 @@
 #include <vector>
 #include <map>
 
+#if defined(_WIN32) && !defined(_WIN32_WINNT)
+#define _WIN32_WINNT 0x0A00
+#endif
+
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
 #else
@@ -26,8 +30,6 @@
     fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
-#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
-
 struct common_time_meas {
     common_time_meas(int64_t & t_acc, bool disable = false);
     ~common_time_meas();
@@ -80,7 +82,8 @@ int32_t cpu_get_num_math();
 enum llama_example {
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
-    LLAMA_EXAMPLE_MAIN,
+    LLAMA_EXAMPLE_COMPLETION,
+    LLAMA_EXAMPLE_CLI,
     LLAMA_EXAMPLE_EMBEDDING,
     LLAMA_EXAMPLE_PERPLEXITY,
     LLAMA_EXAMPLE_RETRIEVAL,
@@ -96,6 +99,7 @@ enum llama_example {
     LLAMA_EXAMPLE_TTS,
     LLAMA_EXAMPLE_DIFFUSION,
     LLAMA_EXAMPLE_FINETUNE,
+    LLAMA_EXAMPLE_FIT_PARAMS,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -192,7 +196,6 @@ struct common_params_sampling {
 
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 
-
     std::vector<enum common_sampler_type> samplers = {
         COMMON_SAMPLER_TYPE_PENALTIES,
         COMMON_SAMPLER_TYPE_DRY,
@@ -213,6 +216,10 @@ struct common_params_sampling {
     std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
     std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
+    bool has_logit_bias() const {
+        return !logit_bias.empty();
+    }
+
     // print the parameters into a string
     std::string print() const;
 };
@@ -223,6 +230,7 @@ struct common_params_model {
     std::string hf_repo     = ""; // HF repo                                                // NOLINT
     std::string hf_file     = ""; // HF file                                                // NOLINT
     std::string docker_repo = ""; // Docker repo                                            // NOLINT
+    std::string name        = ""; // in format <user>/<model>[:<tag>] (tag is optional)     // NOLINT
 };
 
 struct common_params_speculative {
@@ -299,8 +307,8 @@ struct lr_opt {
 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_predict             =    -1; // max. number of new tokens to predict, -1 == no limit
-    int32_t n_ctx                 =  4096; // context size
+    int32_t n_ctx                 =     0; // context size, 0 == context the model was trained with
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt
@@ -324,6 +332,9 @@ struct common_params {
     int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM (-1 - use default)
     int32_t main_gpu           = 0;                // the GPU that is used for scratch and small tensors
     float   tensor_split[128]  = {0};              // how split tensors should be distributed across GPUs
+    bool    fit_params         = true;             // whether to fit unset model/context parameters to free device memory
+    size_t  fit_params_target  = 1024 * 1024*1024; // margin per device in bytes for fitting parameters to free memory
+    int32_t fit_params_min_ctx = 4096;             // minimum context size to set when trying to reduce memory use
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 
@@ -369,7 +380,7 @@ struct common_params {
 
     std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
 
-    int32_t verbosity                  = 0;
+    int32_t verbosity                  = 3;  // LOG_LEVEL_INFO
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
     bool    offline                    = false;
@@ -403,6 +414,7 @@ struct common_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool no_perf           = false; // disable performance metrics
+    bool show_timings      = true;  // show timing information on CLI
     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
@@ -459,7 +471,7 @@ struct common_params {
     std::string public_path   = "";                                                                         // NOLINT
     std::string api_prefix    = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
-    bool use_jinja = false;                                                                                 // NOLINT
+    bool use_jinja = true;                                                                                  // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     int reasoning_budget = -1;
@@ -478,9 +490,16 @@ struct common_params {
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
 
+    // router server configs
+    std::string models_dir    = ""; // directory containing models for the router server
+    std::string models_preset = ""; // directory containing model presets for the router server
+    int models_max = 4;             // maximum number of models to load simultaneously
+    bool models_autoload = true;    // automatically load models when requested via the router server
+
     bool log_json = false;
 
     std::string slot_save_path;
+    std::string media_path; // path to directory for loading media files
 
     float slot_prompt_similarity = 0.1f;
 
@@ -631,8 +650,9 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 // Filesystem utils
 //
 
-bool fs_validate_filename(const std::string & filename);
+bool fs_validate_filename(const std::string & filename, bool allow_subdirs = false);
 bool fs_create_directory_with_parents(const std::string & path);
+bool fs_is_directory(const std::string & path);
 
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
@@ -641,22 +661,44 @@ struct common_file_info {
     std::string path;
     std::string name;
     size_t      size = 0; // in bytes
+    bool        is_dir = false;
 };
-std::vector<common_file_info> fs_list_files(const std::string & path);
+std::vector<common_file_info> fs_list(const std::string & path, bool include_directories);
+
+//
+// TTY utils
+//
+
+// Auto-detect if colors can be enabled based on terminal and environment
+bool tty_can_use_colors();
 
 //
 // Model utils
 //
 
-// note: defines object's lifetime
+struct common_sampler;
-struct common_init_result {
-    llama_model_ptr   model;
-    llama_context_ptr context;
 
-    std::vector<llama_adapter_lora_ptr> lora;
+// note: defines the model, context, samplers, ets. lifetimes
+struct common_init_result {
+    common_init_result(common_params & params);
+    ~common_init_result();
+
+    llama_model * model();
+    llama_context * context();
+    common_sampler * sampler(llama_seq_id seq_id);
+
+    std::vector<llama_adapter_lora_ptr> & lora();
+
+    void free_context();
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
 };
 
-struct common_init_result     common_init_from_params(common_params & params);
+using common_init_result_ptr = std::unique_ptr<common_init_result>;
+
+common_init_result_ptr common_init_from_params(common_params & params);
 
 struct llama_model_params     common_model_params_to_llama  (      common_params & params);
 struct llama_context_params   common_context_params_to_llama(const common_params & params);

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

@@ -305,8 +305,9 @@ static std::string format_literal(const std::string & literal) {
 
 std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }
 
-class SchemaConverter {
+class common_schema_converter {
 private:
+    friend class common_schema_info;
     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;
@@ -729,7 +730,7 @@ private:
     }
 
 public:
-    SchemaConverter(
+    common_schema_converter(
         const std::function<json(const std::string &)> & fetch_json,
         bool dotall)
           : _fetch_json(fetch_json), _dotall(dotall)
@@ -974,7 +975,7 @@ public:
 
     void check_errors() {
         if (!_errors.empty()) {
-            throw std::runtime_error("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
+            throw std::invalid_argument("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
         }
         if (!_warnings.empty()) {
             fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", string_join(_warnings, "; ").c_str());
@@ -990,6 +991,134 @@ public:
     }
 };
 
+// common_schema_info implementation (pimpl)
+
+common_schema_info::common_schema_info()
+    : impl_(std::make_unique<common_schema_converter>(
+        [](const std::string &) { return json(); },
+        false)) {}
+
+common_schema_info::~common_schema_info() = default;
+
+common_schema_info::common_schema_info(common_schema_info &&) noexcept = default;
+common_schema_info & common_schema_info::operator=(common_schema_info &&) noexcept = default;
+
+void common_schema_info::resolve_refs(nlohmann::ordered_json & schema) {
+    impl_->resolve_refs(schema, "");
+}
+
+// Determines if a JSON schema can resolve to a string type through any path.
+// Some models emit raw string values rather than JSON-encoded strings for string parameters.
+// If any branch of the schema (via oneOf, anyOf, $ref, etc.) permits a string, this returns
+// true, allowing callers to handle the value as a raw string for simplicity.
+bool common_schema_info::resolves_to_string(const nlohmann::ordered_json & schema) {
+    std::unordered_set<std::string> visited_refs;
+
+    std::function<bool(const json &)> check = [&](const json & s) -> bool {
+        if (!s.is_object()) {
+            return false;
+        }
+
+        // Handle $ref
+        if (s.contains("$ref")) {
+            const std::string & ref = s["$ref"];
+            if (visited_refs.find(ref) != visited_refs.end()) {
+                // Circular reference, assume not a string to be safe
+                return false;
+            }
+            visited_refs.insert(ref);
+            auto it = impl_->_refs.find(ref);
+            if (it != impl_->_refs.end()) {
+                return check(it->second);
+            }
+            return false;
+        }
+
+        // Check type field
+        if (s.contains("type")) {
+            const json & schema_type = s["type"];
+            if (schema_type.is_string()) {
+                if (schema_type == "string") {
+                    return true;
+                }
+            } else if (schema_type.is_array()) {
+                // Type can be an array like ["string", "null"]
+                for (const auto & t : schema_type) {
+                    if (t == "string") {
+                        return true;
+                    }
+                }
+            }
+        }
+
+        // Check oneOf/anyOf - if any alternative can be a string
+        if (s.contains("oneOf")) {
+            for (const auto & alt : s["oneOf"]) {
+                if (check(alt)) {
+                    return true;
+                }
+            }
+        }
+        if (s.contains("anyOf")) {
+            for (const auto & alt : s["anyOf"]) {
+                if (check(alt)) {
+                    return true;
+                }
+            }
+        }
+
+        // Check allOf - all components must be compatible with string type
+        if (s.contains("allOf")) {
+            bool all_string = true;
+            for (const auto & component : s["allOf"]) {
+                if (!check(component)) {
+                    all_string = false;
+                    break;
+                }
+            }
+            if (all_string) {
+                return true;
+            }
+        }
+
+        // Check const - if the constant value is a string
+        if (s.contains("const")) {
+            if (s["const"].is_string()) {
+                return true;
+            }
+        }
+
+        // Check enum - if any enum value is a string
+        if (s.contains("enum")) {
+            for (const auto & val : s["enum"]) {
+                if (val.is_string()) {
+                    return true;
+                }
+            }
+        }
+
+        // String-specific keywords imply string type
+        if (s.contains("pattern") || s.contains("minLength") || s.contains("maxLength")) {
+            return true;
+        }
+
+        // Check format - many formats imply string
+        if (s.contains("format")) {
+            const std::string & fmt = s["format"];
+            if (fmt == "date" || fmt == "time" || fmt == "date-time" ||
+                fmt == "uri" || fmt == "email" || fmt == "hostname" ||
+                fmt == "ipv4" || fmt == "ipv6" || fmt == "uuid" ||
+                fmt.find("uuid") == 0) {
+                return true;
+            }
+        }
+
+        return false;
+    };
+
+    return check(schema);
+}
+
 std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
 #ifdef LLAMA_USE_LLGUIDANCE
     if (!force_gbnf) {
@@ -1006,7 +1135,7 @@ std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
 }
 
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options) {
-    SchemaConverter converter([&](const std::string &) { return json(); }, options.dotall);
+    common_schema_converter converter([&](const std::string &) { return json(); }, options.dotall);
     common_grammar_builder builder {
         /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
             return converter._add_rule(name, rule);

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

@@ -3,11 +3,31 @@
 #include <nlohmann/json_fwd.hpp>
 
 #include <functional>
+#include <memory>
 #include <string>
 
 std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
                                    bool force_gbnf = false);
 
+class common_schema_converter;
+
+// Probes a JSON schema to extract information about its structure and type constraints.
+class common_schema_info {
+    std::unique_ptr<common_schema_converter> impl_;
+
+  public:
+    common_schema_info();
+    ~common_schema_info();
+
+    common_schema_info(const common_schema_info &) = delete;
+    common_schema_info & operator=(const common_schema_info &) = delete;
+    common_schema_info(common_schema_info &&) noexcept;
+    common_schema_info & operator=(common_schema_info &&) noexcept;
+
+    void resolve_refs(nlohmann::ordered_json & schema);
+    bool resolves_to_string(const nlohmann::ordered_json & schema);
+};
+
 struct common_grammar_builder {
     std::function<std::string(const std::string &, const std::string &)> add_rule;
     std::function<std::string(const std::string &, const nlohmann::ordered_json &)> add_schema;

llama/llama.cpp/common/log.cpp

@@ -1,3 +1,4 @@
+#include "common.h"
 #include "log.h"
 
 #include <chrono>
@@ -26,30 +27,6 @@ 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();
 }
@@ -391,7 +368,7 @@ struct common_log * common_log_main() {
     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());
+        log.set_colors(tty_can_use_colors());
     });
 
     return &log;
@@ -422,7 +399,7 @@ void common_log_set_file(struct common_log * log, const char * file) {
 
 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());
+        log->set_colors(tty_can_use_colors());
         return;
     }
 
@@ -443,8 +420,27 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
 
+void common_log_flush(struct common_log * log) {
+    log->pause();
+    log->resume();
+}
+
+static int common_get_verbosity(enum ggml_log_level level) {
+    switch (level) {
+        case GGML_LOG_LEVEL_DEBUG: return LOG_LEVEL_DEBUG;
+        case GGML_LOG_LEVEL_INFO:  return LOG_LEVEL_INFO;
+        case GGML_LOG_LEVEL_WARN:  return LOG_LEVEL_WARN;
+        case GGML_LOG_LEVEL_ERROR: return LOG_LEVEL_ERROR;
+        case GGML_LOG_LEVEL_CONT:  return LOG_LEVEL_INFO; // same as INFO
+        case GGML_LOG_LEVEL_NONE:
+        default:
+            return LOG_LEVEL_OUTPUT;
+    }
+}
+
 void common_log_default_callback(enum ggml_log_level level, const char * text, void * /*user_data*/) {
-    if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
+    auto verbosity = common_get_verbosity(level);
+    if (verbosity <= common_log_verbosity_thold) {
         common_log_add(common_log_main(), level, "%s", text);
     }
 }

llama/llama.cpp/common/log.h

@@ -21,8 +21,14 @@
 #    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
 #endif
 
-#define LOG_DEFAULT_DEBUG 1
+#define LOG_LEVEL_DEBUG  4
-#define LOG_DEFAULT_LLAMA 0
+#define LOG_LEVEL_INFO   3
+#define LOG_LEVEL_WARN   2
+#define LOG_LEVEL_ERROR  1
+#define LOG_LEVEL_OUTPUT 0 // output data from tools
+
+#define LOG_DEFAULT_DEBUG LOG_LEVEL_DEBUG
+#define LOG_DEFAULT_LLAMA LOG_LEVEL_INFO
 
 enum log_colors {
     LOG_COLORS_AUTO     = -1,
@@ -67,16 +73,18 @@ void common_log_add(struct common_log * log, enum ggml_log_level level, const ch
 //   0.00.090.578 I llm_load_tensors: offloading 32 repeating layers to GPU
 //   0.00.090.579 I llm_load_tensors: offloading non-repeating layers to GPU
 //
-// I - info    (stdout, V = 0)
-// W - warning (stderr, V = 0)
-// E - error   (stderr, V = 0)
 // D - debug   (stderr, V = LOG_DEFAULT_DEBUG)
+// I - info    (stdout, V = LOG_DEFAULT_INFO)
+// W - warning (stderr, V = LOG_DEFAULT_WARN)
+// E - error   (stderr, V = LOG_DEFAULT_ERROR)
+// O - output  (stdout, V = LOG_DEFAULT_OUTPUT)
 //
 
 void common_log_set_file      (struct common_log * log, const char * file); // 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
+void common_log_flush         (struct common_log * log);                    // flush all pending log messages
 
 // helper macros for logging
 // use these to avoid computing log arguments if the verbosity of the log is higher than the threshold
@@ -95,14 +103,14 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps);   // w
         } \
     } while (0)
 
-#define LOG(...)             LOG_TMPL(GGML_LOG_LEVEL_NONE, 0,         __VA_ARGS__)
+#define LOG(...)             LOG_TMPL(GGML_LOG_LEVEL_NONE, LOG_LEVEL_OUTPUT, __VA_ARGS__)
 #define LOGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_NONE, verbosity,        __VA_ARGS__)
 
-#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  0,                 __VA_ARGS__)
+#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_LEVEL_DEBUG,  __VA_ARGS__)
-#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  0,                 __VA_ARGS__)
+#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  LOG_LEVEL_INFO,   __VA_ARGS__)
-#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0,                 __VA_ARGS__)
+#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  LOG_LEVEL_WARN,   __VA_ARGS__)
-#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
+#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, LOG_LEVEL_ERROR,  __VA_ARGS__)
-#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  0,                 __VA_ARGS__)
+#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  LOG_LEVEL_INFO,   __VA_ARGS__) // same as INFO
 
 #define LOG_INFV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  verbosity, __VA_ARGS__)
 #define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  verbosity, __VA_ARGS__)

llama/llama.cpp/common/sampling.cpp

@@ -104,9 +104,10 @@ struct ring_buffer {
 struct common_sampler {
     common_params_sampling params;
 
-    struct llama_sampler * grmr;
     struct llama_sampler * chain;
 
+    bool grammar;
+
     ring_buffer<llama_token> prev;
 
     std::vector<llama_token_data> cur;
@@ -116,7 +117,6 @@ struct common_sampler {
     void reset() {
         prev.clear();
 
-        llama_sampler_reset(grmr);
         llama_sampler_reset(chain);
     }
 
@@ -167,10 +167,15 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
     lparams.no_perf = params.no_perf;
 
-    struct llama_sampler * grmr;
+    llama_sampler * chain = llama_sampler_chain_init(lparams);
+
+    bool grammar = false;
+    std::vector<llama_sampler *> samplers;
+
     if (params.grammar.compare(0, 11, "%llguidance") == 0) {
 #ifdef LLAMA_USE_LLGUIDANCE
-        grmr = llama_sampler_init_llg(vocab, "lark", params.grammar.c_str());
+        samplers.push_back(llama_sampler_init_llg(vocab, "lark", params.grammar.c_str()));
+        grammar = true;
 #else
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
@@ -217,30 +222,23 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
             trigger_patterns_c.push_back(regex.c_str());
         }
 
-        grmr = params.grammar_lazy
+        if (!params.grammar.empty()) {
-             ? llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
+             if (params.grammar_lazy) {
+                 samplers.push_back(
+                         llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
                              trigger_patterns_c.data(), trigger_patterns_c.size(),
-                                                        trigger_tokens.data(), trigger_tokens.size())
+                             trigger_tokens.data(),     trigger_tokens.size()));
-             :      llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
+             } else {
-        if (!grmr) {
+                 samplers.push_back(llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root"));
-            return nullptr;
+             }
+
+             grammar = true;
         }
     }
 
-    auto * result = new common_sampler {
+    if (params.has_logit_bias()) {
-        /* .params = */ params,
+        samplers.push_back(llama_sampler_init_logit_bias(llama_vocab_n_tokens(vocab), params.logit_bias.size(), params.logit_bias.data()));
-        /* .grmr   = */ grmr,
+    }
-        /* .chain  = */ llama_sampler_chain_init(lparams),
-        /* .prev   = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
-        /* .cur    = */ {},
-        /* .cur_p  = */ {},
-    };
-
-    llama_sampler_chain_add(result->chain,
-            llama_sampler_init_logit_bias(
-                llama_vocab_n_tokens(vocab),
-                params.logit_bias.size(),
-                params.logit_bias.data()));
 
     if (params.mirostat == 0) {
         for (const auto & cnstr : params.samplers) {
@@ -253,58 +251,70 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                             c_breakers.push_back(str.c_str());
                         }
 
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                        samplers.push_back(llama_sampler_init_dry    (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
                     }
                     break;
                 case COMMON_SAMPLER_TYPE_TOP_K:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k       (params.top_k));
+                    samplers.push_back(llama_sampler_init_top_k      (params.top_k));
                     break;
                 case COMMON_SAMPLER_TYPE_TOP_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p       (params.top_p, params.min_keep));
+                    samplers.push_back(llama_sampler_init_top_p      (params.top_p, params.min_keep));
                     break;
                 case COMMON_SAMPLER_TYPE_TOP_N_SIGMA:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma));
+                    samplers.push_back(llama_sampler_init_top_n_sigma(params.top_n_sigma));
                     break;
                 case COMMON_SAMPLER_TYPE_MIN_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p       (params.min_p, params.min_keep));
+                    samplers.push_back(llama_sampler_init_min_p      (params.min_p, params.min_keep));
                     break;
                 case COMMON_SAMPLER_TYPE_XTC:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc         (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                    samplers.push_back(llama_sampler_init_xtc        (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
                     break;
                 case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical     (params.typ_p, params.min_keep));
+                    samplers.push_back(llama_sampler_init_typical    (params.typ_p, params.min_keep));
                     break;
                 case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext    (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                    samplers.push_back(llama_sampler_init_temp_ext   (params.temp, params.dynatemp_range, params.dynatemp_exponent));
                     break;
                 case COMMON_SAMPLER_TYPE_INFILL:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill      (vocab));
+                    samplers.push_back(llama_sampler_init_infill     (vocab));
                     break;
                 case COMMON_SAMPLER_TYPE_PENALTIES:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_penalties   (params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    samplers.push_back(llama_sampler_init_penalties  (params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
                     break;
                 default:
                     GGML_ASSERT(false && "unknown sampler type");
             }
         }
-        llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
+
+        samplers.push_back(llama_sampler_init_dist(params.seed));
     } else if (params.mirostat == 1) {
-        llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
+        samplers.push_back(llama_sampler_init_temp(params.temp));
-        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_vocab_n_tokens(vocab), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
+        samplers.push_back(llama_sampler_init_mirostat(llama_vocab_n_tokens(vocab), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
     } else if (params.mirostat == 2) {
-        llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
+        samplers.push_back(llama_sampler_init_temp(params.temp));
-        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
+        samplers.push_back(llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
     } else {
         GGML_ASSERT(false && "unknown mirostat version");
     }
 
+    for (auto * smpl : samplers) {
+        llama_sampler_chain_add(chain, smpl);
+    }
+
+    auto * result = new common_sampler {
+        /* .params  = */ params,
+        /* .chain   = */ chain,
+        /* .grammar = */ grammar,
+        /* .prev    = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
+        /* .cur     = */ {},
+        /* .cur_p   = */ {},
+    };
+
     return result;
 }
 
 void common_sampler_free(struct common_sampler * gsmpl) {
     if (gsmpl) {
-        llama_sampler_free(gsmpl->grmr);
-
         llama_sampler_free(gsmpl->chain);
 
         delete gsmpl;
@@ -314,11 +324,24 @@ void common_sampler_free(struct common_sampler * gsmpl) {
 void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
     const auto tm = gsmpl->tm();
 
-    if (accept_grammar) {
+    if (gsmpl->grammar) {
-        llama_sampler_accept(gsmpl->grmr, token);
+        const int n_smpl = llama_sampler_chain_n(gsmpl->chain);
-    }
 
+        for (int i = 0; i < n_smpl; i++) {
+            auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
+
+            // the grammar sampler is always the first one
+            if (i == 0) {
+                if (accept_grammar) {
+                    llama_sampler_accept(smpl, token);
+                }
+            } else {
+                llama_sampler_accept(smpl, token);
+            }
+        }
+    } else {
         llama_sampler_accept(gsmpl->chain, token);
+    }
 
     gsmpl->prev.push_back(token);
 }
@@ -330,8 +353,8 @@ void common_sampler_reset(struct common_sampler * gsmpl) {
 struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
     return new common_sampler {
         /* .params  = */ gsmpl->params,
-        /* .grmr   = */ llama_sampler_clone(gsmpl->grmr),
         /* .chain   = */ llama_sampler_clone(gsmpl->chain),
+        /* .grammar = */ gsmpl->grammar,
         /* .prev    = */ gsmpl->prev,
         /* .cur     = */ gsmpl->cur,
         /* .cur_p   = */ gsmpl->cur_p,
@@ -383,58 +406,33 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
     }
 }
 
-llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
+struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl) {
+    return gsmpl->chain;
+}
+
+llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx) {
     llama_synchronize(ctx);
 
     // start measuring sampling time after the llama_context synchronization in order to not measure any ongoing async operations
     const auto tm = gsmpl->tm();
 
-    gsmpl->set_logits(ctx, idx);
+    llama_token id = LLAMA_TOKEN_NULL;
 
-    auto & grmr  = gsmpl->grmr;
     auto & chain = gsmpl->chain;
     auto & cur_p = gsmpl->cur_p; // initialized by set_logits
 
-    if (grammar_first) {
+    gsmpl->set_logits(ctx, idx);
-        llama_sampler_apply(grmr, &cur_p);
-    }
 
     llama_sampler_apply(chain, &cur_p);
 
     GGML_ASSERT(cur_p.selected != -1 && "no selected token during sampling - check your sampling configuration");
 
-    const llama_token id = cur_p.data[cur_p.selected].id;
+    id = cur_p.data[cur_p.selected].id;
 
-    if (grammar_first) {
     return id;
 }
 
-    // check if it the sampled token fits the grammar
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft) {
-    {
-        llama_token_data       single_token_data       = { id, 1.0f, 0.0f };
-        llama_token_data_array single_token_data_array = { &single_token_data, 1, -1, false };
-
-        llama_sampler_apply(grmr, &single_token_data_array);
-
-        const bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
-        if (is_valid) {
-            return id;
-        }
-    }
-
-    // resampling:
-    // if the token is not valid, sample again, but first apply the grammar sampler and then the sampling chain
-    gsmpl->set_logits(ctx, idx);
-
-    llama_sampler_apply(grmr,  &cur_p);
-    llama_sampler_apply(chain, &cur_p);
-
-    GGML_ASSERT(cur_p.selected != -1 && "no selected token during re-sampling - check your sampling configuration");
-
-    return cur_p.data[cur_p.selected].id;
-}
-
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first) {
     GGML_ASSERT(idxs.size() == draft.size() + 1 && "idxs.size() must be draft.size() + 1");
 
     std::vector<llama_token> result;
@@ -442,7 +440,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
 
     size_t i = 0;
     for (; i < draft.size(); i++) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
 
         common_sampler_accept(gsmpl, id, true);
 
@@ -454,7 +452,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
     }
 
     if (i == draft.size()) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
 
         common_sampler_accept(gsmpl, id, true);
 
@@ -464,13 +462,13 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
     return result;
 }
 
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first) {
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft) {
     std::vector<int> idxs(draft.size() + 1);
     for (size_t i = 0; i < idxs.size(); ++i) {
         idxs[i] = i;
     }
 
-    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft, grammar_first);
+    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft);
 }
 
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
@@ -515,7 +513,8 @@ std::string common_sampler_print(const struct common_sampler * gsmpl) {
 
     for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) {
         const auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
-        result += std::string("-> ") + llama_sampler_name(smpl) + " ";
+        result += std::string("-> ");
+        result += std::string(llama_sampler_name(smpl)) + " ";
     }
 
     return result;

llama/llama.cpp/common/sampling.h

@@ -48,6 +48,8 @@ struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
 // arguments can be nullptr to skip printing
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl);
 
+struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl);
+
 // extended sampling implementation:
 //
 // - set logits
@@ -55,10 +57,7 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
 // - check if the token fits the grammar (if any)
 // - if not: resample by first applying the grammar constraints and then sampling again (slower path)
 //
-// if grammar_first is true, the grammar is applied before the samplers (slower)
+llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx);
-// useful in cases where all the resulting candidates (not just the sampled one) must fit the grammar
-//
-llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
 
 // generalized version of common_sampler_sample
 //
@@ -76,10 +75,10 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
 //
 // returns at least 1 token, up to idxs.size()
 //
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first = false);
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft);
 
 // assume idxs == [ 0, 1, 2, ..., draft.size() ]
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first = false);
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft);
 
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 
@@ -107,3 +106,9 @@ std::vector<enum common_sampler_type> common_sampler_types_from_chars(const std:
 
 llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab,
                 const char * grammar_kind, const char * grammar_data);
+
+struct common_sampler_deleter {
+    void operator()(common_sampler * s) { common_sampler_free(s); }
+};
+
+typedef std::unique_ptr<common_sampler, common_sampler_deleter> common_sampler_ptr;

llama/llama.cpp/include/llama.h

@@ -313,6 +313,7 @@ extern "C" {
         bool check_tensors;   // validate model tensor data
         bool use_extra_bufts; // use extra buffer types (used for weight repacking)
         bool no_host;         // bypass host buffer allowing extra buffers to be used
+        bool no_alloc;        // only load metadata and simulate memory allocations
     };
 
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
@@ -466,10 +467,24 @@ extern "C" {
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 
+    // fits mparams and cparams to free device memory (assumes system memory is unlimited)
+    // returns true if the parameters could be successfully modified to fit device memory
+    // this function is NOT thread safe because it modifies the global llama logger state
+    LLAMA_API bool llama_params_fit(
+                                   const char   * path_model,
+                    struct llama_model_params   * mparams,
+                    struct llama_context_params * cparams,
+                                          float * tensor_split,          // writable buffer for tensor split, needs at least llama_max_devices elements
+        struct llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
+                                         size_t   margin,                // margin of memory to leave per device in bytes
+                                       uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
+                            enum ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
+
     LLAMA_API int64_t llama_time_us(void);
 
     LLAMA_API size_t llama_max_devices(void);
     LLAMA_API size_t llama_max_parallel_sequences(void);
+    LLAMA_API size_t llama_max_tensor_buft_overrides(void);
 
     LLAMA_API bool llama_supports_mmap       (void);
     LLAMA_API bool llama_supports_mlock      (void);
@@ -1354,6 +1369,8 @@ extern "C" {
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
+    // The logger state is global so these functions are NOT thread safe.
+    LLAMA_API void llama_log_get(ggml_log_callback * log_callback, void ** user_data);
     LLAMA_API void llama_log_set(ggml_log_callback   log_callback, void *  user_data);
 
     //

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

@@ -3,6 +3,7 @@
 #include "ggml.h" // ggml_op
 
 #include <string>
+#include <set>
 
 //
 // gguf constants (sync with gguf.py)
@@ -79,6 +80,7 @@ enum llm_arch {
     LLM_ARCH_JAIS,
     LLM_ARCH_NEMOTRON,
     LLM_ARCH_NEMOTRON_H,
+    LLM_ARCH_NEMOTRON_H_MOE,
     LLM_ARCH_EXAONE,
     LLM_ARCH_EXAONE4,
     LLM_ARCH_RWKV6,
@@ -116,6 +118,7 @@ enum llm_arch {
     LLM_ARCH_COGVLM,
     LLM_ARCH_RND1,
     LLM_ARCH_PANGU_EMBED,
+    LLM_ARCH_MISTRAL3,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -209,6 +212,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_OUTPUT_SCALE,
     LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
+    LLM_KV_ATTENTION_TEMPERATURE_SCALE,
     LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
@@ -315,6 +319,7 @@ enum llm_tensor {
     LLM_TENSOR_DENSE_3_OUT,
     LLM_TENSOR_OUTPUT,
     LLM_TENSOR_OUTPUT_NORM,
+    LLM_TENSOR_OUTPUT_NORM_LFM2, // fix for wrong tensor name
     LLM_TENSOR_ROPE_FREQS,
     LLM_TENSOR_ROPE_FACTORS_LONG,
     LLM_TENSOR_ROPE_FACTORS_SHORT,
@@ -379,6 +384,7 @@ enum llm_tensor {
     LLM_TENSOR_SSM_DT,
     LLM_TENSOR_SSM_DT_NORM,
     LLM_TENSOR_SSM_A,
+    LLM_TENSOR_SSM_A_NOSCAN,        // qwen3next special case with MUL instead of SSM_SCAN
     LLM_TENSOR_SSM_B_NORM,
     LLM_TENSOR_SSM_C_NORM,
     LLM_TENSOR_SSM_D,
@@ -525,6 +531,10 @@ struct LLM_TN_IMPL {
     const int bid;
     const int xid;
 
+    const std::set<llm_tensor> model_tensors;
+
+    LLM_TN_IMPL(llm_arch arch, llm_tensor tensor, const char * suffix, int bid, int xid);
+
     std::string str() const;
 
     operator std::string() const {
@@ -546,11 +556,11 @@ struct LLM_TN {
     llm_arch arch;
 
     LLM_TN_IMPL operator()(llm_tensor tensor, const char * suffix, int bid = -1, int xid = -1) const {
-        return { arch, tensor, suffix, bid, xid };
+        return LLM_TN_IMPL(arch, tensor, suffix, bid, xid);
     }
 
     LLM_TN_IMPL operator()(llm_tensor tensor, int bid = -1, int xid = -1) const {
-        return { arch, tensor, nullptr, bid, xid };
+        return LLM_TN_IMPL(arch, tensor, nullptr, bid, xid);
     }
 };
 

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

@@ -695,6 +695,8 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
     udata->seq_idx   .resize(LLAMA_MAX_SEQ, -1);
     udata->output    .resize(n_tokens);
 
+    udata->seq_id_data.reserve(n_tokens);
+
     seq_set_t seq_set_unq;
 
     for (size_t i = 0; i < idxs.size(); ++i) {
@@ -716,11 +718,13 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
         }
 
         udata->n_seq_id[i] = batch.n_seq_id[idxs[i]];
-        udata->seq_id[i]   = batch.seq_id[idxs[i]];
         udata->output[i]   = batch.logits[idxs[i]];
 
         for (int s = 0; s < udata->n_seq_id[i]; ++s) {
-            seq_set_unq.set(udata->seq_id[i][s]);
+            const llama_seq_id seq_id = batch.seq_id[idxs[i]][s];
+
+            udata->seq_id_data.push_back(seq_id);
+            seq_set_unq.set(seq_id);
         }
 
         if (udata->output[i]) {
@@ -728,6 +732,12 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
         }
     }
 
+    llama_seq_id * seq_id_ptr = udata->seq_id_data.data();
+    for (size_t i = 0; i < idxs.size(); ++i) {
+        udata->seq_id[i] = seq_id_ptr;
+        seq_id_ptr += udata->n_seq_id[i];
+    }
+
     for (uint32_t s = 0; s < n_seq_max; ++s) {
         if (seq_set_unq.test(s)) {
             udata->seq_idx[s] = udata->seq_id_unq.size();

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

@@ -56,13 +56,15 @@ struct llama_ubatch {
         std::vector<float>          embd;
         std::vector<llama_pos>      pos;
         std::vector<int32_t>        n_seq_id;
-        std::vector<llama_seq_id *> seq_id;
+        std::vector<llama_seq_id *> seq_id;      // these point into the seq_id_data below
         std::vector<llama_seq_id>   seq_id_unq;
         std::vector<int32_t>        seq_idx;
         std::vector<int8_t>         output;
+
+        std::vector<llama_seq_id> seq_id_data;
     };
 
-    // the llama_ubatch pointers above point to this data if set. otherwise - points to non-owning data
+    // the llama_ubatch pointers above point to this data if set. otherwise - point to external non-owning data
     std::shared_ptr<data_t> data;
 };
 

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

@@ -9,6 +9,7 @@
 #include "llama-model.h"
 
 #include <cinttypes>
+#include <cmath>
 #include <cstring>
 #include <limits>
 #include <stdexcept>
@@ -72,6 +73,43 @@ llama_context::llama_context(
         cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f;
     }
 
+    if (cparams.yarn_ext_factor != 0) {
+        static auto get_mscale = [](float scale, float mscale) {
+            return scale <= 1.0f ? 1.0f : (0.1f * mscale * logf(scale) + 1.0f);
+        };
+
+        const float factor = 1.0f / cparams.rope_freq_scale;
+
+        // ref: https://github.com/huggingface/transformers/blob/6d00f6b0a5679c36510f203e4226e36f517c3032/src/transformers/modeling_rope_utils.py#L336-L348
+        if (hparams.rope_yarn_log_mul != 0.0f) {
+            // note: here we assume `mscale == 1.0f`
+            // TODO: start reading the actual value of mscale and handle the case where it is not 1.0f
+                  float mscale          = 1.0f;
+            const float mscale_all_dims = hparams.rope_yarn_log_mul;
+
+            // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+            // special-case DEEPSEEK v2:
+            // https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/blob/main/config.json#L42-L43
+            if (model.arch == LLM_ARCH_DEEPSEEK2 && mscale_all_dims != 1.0f) {
+                mscale = mscale_all_dims;
+            }
+
+            cparams.yarn_attn_factor = get_mscale(factor, mscale) / get_mscale(factor, mscale_all_dims);
+
+            LLAMA_LOG_WARN("%s: setting new yarn_attn_factor = %.4f (mscale == %.1f, mscale_all_dim = %.1f)\n",
+                    __func__, cparams.yarn_attn_factor, mscale, mscale_all_dims);
+        } else {
+            cparams.yarn_attn_factor = get_mscale(factor, 1.0f);
+        }
+
+        // when YARN is applied with yarn_ext_factor != 0.0f, we need to cancel this factor:
+        // https://github.com/ggml-org/llama.cpp/blob/a81a569577cc38b32558958b048228150be63eae/ggml/src/ggml-cpu/ops.cpp#L5541-L5544
+        //
+        // ref: https://github.com/ggml-org/llama.cpp/discussions/7416
+        //      https://github.com/ggml-org/llama.cpp/pull/17945
+        cparams.yarn_attn_factor *= 1.0f / (1.0f + 0.1f * logf(factor));
+    }
+
     cparams.yarn_attn_factor *= hparams.rope_attn_factor;
 
     if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
@@ -93,14 +131,6 @@ llama_context::llama_context(
     // 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_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;
-    }
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
 
     cparams.op_offload = params.op_offload;
@@ -228,6 +258,7 @@ llama_context::llama_context(
 
         backend_buft.clear();
         backend_ptrs.clear();
+        backend_buf_exp_size.clear();
 
         for (auto & backend : backends) {
             auto * buft = ggml_backend_get_default_buffer_type(backend.get());
@@ -244,11 +275,15 @@ llama_context::llama_context(
 
             backend_buft.push_back(buft);
             backend_ptrs.push_back(backend.get());
+            backend_buf_exp_size.push_back(0);
         }
 
         LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size());
 
-        const size_t max_nodes = this->graph_max_nodes();
+        const uint32_t n_seqs = cparams.n_seq_max;
+        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
+
+        const size_t max_nodes = this->graph_max_nodes(n_tokens);
 
         LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes);
 
@@ -300,9 +335,6 @@ llama_context::llama_context(
 
         cross.v_embd.clear();
 
-        const uint32_t n_seqs = 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;
 
@@ -359,7 +391,8 @@ llama_context::llama_context(
 
         // 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());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get(),
+                model.hparams.no_alloc, model.hparams.no_alloc ? backend_buf_exp_size.data() : nullptr);
             if (!gf) {
                 if (pipeline_parallel) {
                     LLAMA_LOG_WARN("%s: compute buffer allocation failed, retrying without pipeline parallelism\n", __func__);
@@ -377,7 +410,7 @@ llama_context::llama_context(
 
         // reserve with tg (token generation) graph to get the number of splits and nodes
         {
-            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get());
+            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get(), model.hparams.no_alloc);
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute tg buffers");
             }
@@ -392,7 +425,7 @@ llama_context::llama_context(
             //
             // auto * gf = graph_reserve(n_tokens, 1, n_tokens, mctx.get());
             //
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get(), model.hparams.no_alloc);
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -401,11 +434,13 @@ llama_context::llama_context(
         for (size_t i = 0; i < backend_ptrs.size(); ++i) {
             ggml_backend_t             backend = backend_ptrs[i];
             ggml_backend_buffer_type_t buft    = backend_buft[i];
-            size_t size = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+            if (!model.hparams.no_alloc) {
-            if (size > 1) {
+                backend_buf_exp_size[i] = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+            }
+            if (backend_buf_exp_size[i] > 1) {
                 LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__,
                         ggml_backend_buft_name(buft),
-                        size / 1024.0 / 1024.0);
+                        backend_buf_exp_size[i] / 1024.0 / 1024.0);
             }
         }
 
@@ -424,6 +459,23 @@ llama_context::llama_context(
 }
 
 llama_context::~llama_context() {
+    // FIXME this currently results in a use-after-free bug if the model is freed before the context
+    // if (!model.hparams.no_alloc) {
+    //     for (size_t i = 0; i < backend_ptrs.size(); ++i) {
+    //         ggml_backend_t             backend = backend_ptrs[i];
+    //         ggml_backend_buffer_type_t buft    = backend_buft[i];
+
+    //         const size_t size_exp = backend_buf_exp_size[i];
+    //         const size_t size_act = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+    //         if (size_exp == size_act) {
+    //             LLAMA_LOG_DEBUG("%s: %10s compute buffer size is %8.4f MiB, matches expectation of %8.4f MiB\n",
+    //                 __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+    //         } else {
+    //             LLAMA_LOG_WARN("%s: %10s compute buffer size of %8.4f MiB, does not match expectation of %8.4f MiB\n",
+    //                 __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+    //         }
+    //     }
+    // }
     ggml_opt_free(opt_ctx);
 }
 
@@ -1325,6 +1377,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
             // This doesn't happen often, but may be annoying in some cases (like the HellaSwag benchmark)
             LLAMA_LOG_INFO("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
+            synchronize();
             buf_output = nullptr;
             logits = nullptr;
             embd = nullptr;
@@ -1385,9 +1438,9 @@ void llama_context::output_reorder() {
 // graph
 //
 
-uint32_t llama_context::graph_max_nodes() const {
+uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
     if (model.arch == LLM_ARCH_QWEN3NEXT) {
-        return std::max<uint32_t>(8192u, 32u*model.n_tensors());
+        return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
     }
     return std::max<uint32_t>(1024u, 8u*model.n_tensors());
 }
@@ -1396,7 +1449,8 @@ 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, bool split_only) {
+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, size_t * sizes) {
     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);
 
@@ -1433,8 +1487,13 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
 
     // initialize scheduler with the specified graph
     if (split_only) {
+        if (sizes) {
+            ggml_backend_sched_reserve_size(sched.get(), gf, sizes);
+        } else {
             ggml_backend_sched_split_graph(sched.get(), gf);
+        }
     } else if (!ggml_backend_sched_reserve(sched.get(), gf)) {
+        GGML_ASSERT(!sizes);
         LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
         return nullptr;
     }
@@ -2056,15 +2115,26 @@ void llama_context::perf_reset() {
 
 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()) {
+    for (const auto & [buft, size] : model.memory_breakdown()) {
-        ret[buft_size.first].model += buft_size.second;
+        ret[buft].model += size;
     }
-    for (const auto & buft_size : memory->memory_breakdown()) {
+    if (memory) {
-        ret[buft_size.first].context += buft_size.second;
+        for (const auto & [buft, size] : memory->memory_breakdown()) {
+            ret[buft].context += size;
         }
+    }
+    if (model.hparams.no_alloc) {
+        for (size_t i = 0; i < backends.size(); ++i) {
+            ggml_backend_t             backend = backends[i].get();
+            ggml_backend_buffer_type_t buft    = ggml_backend_sched_get_buffer_type(sched.get(), backend);
+            ret[buft].compute += backend_buf_exp_size[i];
+        }
+    } else {
         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);
+            ggml_backend_buffer_type_t buft    = ggml_backend_sched_get_buffer_type(sched.get(), backend);
+            ret[buft].compute += ggml_backend_sched_get_buffer_size(sched.get(), backend);
+        }
     }
     return ret;
 }

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

@@ -26,6 +26,10 @@ 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
+
+    size_t total() const {
+        return model + context + compute;
+    }
 };
 
 struct llama_context {
@@ -197,7 +201,7 @@ private:
     //
 
 public:
-    uint32_t graph_max_nodes() const;
+    uint32_t graph_max_nodes(uint32_t n_tokens) const;
 
     // can reuse the llm_graph_result instance of the context (for example to update a memory module)
     llm_graph_result * get_gf_res_reserve() const;
@@ -206,7 +210,8 @@ 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, bool split_only = false);
+    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, size_t * sizes = nullptr);
 
 private:
     llm_graph_params graph_params(
@@ -281,9 +286,10 @@ private:
 
     std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
 
-    // buffer types used for the compute buffer of each backend
+    // pointers and buffer types used for the compute buffer of each backend
     std::vector<ggml_backend_t>             backend_ptrs;
     std::vector<ggml_backend_buffer_type_t> backend_buft;
+    std::vector<size_t>                     backend_buf_exp_size; // expected buffer sizes
 
     llm_graph_result_ptr gf_res_prev;
     llm_graph_result_ptr gf_res_reserve;

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

@@ -181,6 +181,52 @@ static std::pair<uint32_t, const char *> parse_char(const char * src) {
     throw std::runtime_error("unexpected end of input");
 }
 
+static std::pair<uint32_t, const char *> parse_token(const llama_vocab * vocab, const char * src) {
+    const char * pos = src;
+    if (*pos != '<') {
+        throw std::runtime_error(std::string("expecting '<' at ") + pos);
+    }
+    pos++;
+
+    // Parse <[id]>
+    if (*pos == '[') {
+        pos++;
+        const char * int_end = parse_int(pos);
+        uint32_t token_id = std::stoul(std::string(pos, int_end - pos));
+        pos = int_end;
+        if (*pos != ']') {
+            throw std::runtime_error(std::string("expecting ']' at ") + pos);
+        }
+        pos++;
+        if (*pos != '>') {
+            throw std::runtime_error(std::string("expecting '>' at ") + pos);
+        }
+        pos++;
+        return std::make_pair(token_id, pos);
+    }
+
+    if (vocab == nullptr) {
+        throw std::runtime_error(std::string("no vocab to parse token at ") + src);
+    }
+
+    // Parse <token> and tokenize to obtain the token id
+    while (*pos != 0 && *pos != '>') {
+        pos++;
+    }
+    if (*pos != '>') {
+        throw std::runtime_error(std::string("expecting '>' at ") + pos);
+    }
+    pos++;
+
+    llama_token tokens[2];
+    int32_t n_tokens = vocab->tokenize(src, static_cast<int32_t>(pos - src), tokens, 2, false, true);
+    if (n_tokens != 1) {
+        // must tokenize to exactly 1 token
+        throw std::runtime_error("invalid token '" + std::string(src, pos - src) + "'");
+    }
+    return std::make_pair(tokens[0], pos);
+}
+
 static void print_grammar_char(FILE * file, uint32_t c) {
     if (0x20 <= c && c <= 0x7f) {
         fprintf(file, "%c", static_cast<char>(c));
@@ -212,6 +258,8 @@ static void print_rule_binary(FILE * file, const llama_grammar_rule & rule) {
             case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
             case LLAMA_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
             case LLAMA_GRETYPE_CHAR_ANY:       fprintf(file, "CHAR_ANY");       break;
+            case LLAMA_GRETYPE_TOKEN:          fprintf(file, "TOKEN");          break;
+            case LLAMA_GRETYPE_TOKEN_NOT:      fprintf(file, "TOKEN_NOT");      break;
         }
         switch (elem.type) {
             case LLAMA_GRETYPE_END:
@@ -228,6 +276,17 @@ static void print_rule_binary(FILE * file, const llama_grammar_rule & rule) {
                 print_grammar_char(file, elem.value);
                 fprintf(file, "\") ");
                 break;
+            case LLAMA_GRETYPE_TOKEN:
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
+            case LLAMA_GRETYPE_TOKEN_NOT:
+                fprintf(file, "!");
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
         }
     }
     fprintf(file, "\n");
@@ -284,6 +343,17 @@ static void print_rule(
             case LLAMA_GRETYPE_CHAR_ANY:
                 fprintf(file, ".");
                 break;
+            case LLAMA_GRETYPE_TOKEN:
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
+            case LLAMA_GRETYPE_TOKEN_NOT:
+                fprintf(file, "!");
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
         }
         if (is_char_element(elem)) {
             switch (rule[i + 1].type) {
@@ -444,6 +514,17 @@ const char * llama_grammar_parser::parse_sequence(
                 }
             }
             pos = parse_space(pos + 1, is_nested);
+        } else if (*pos == '<' || *pos == '!') { // token
+            auto type = LLAMA_GRETYPE_TOKEN;
+            if (*pos == '!') { // token inverse
+                type = LLAMA_GRETYPE_TOKEN_NOT;
+                pos++;
+            }
+            auto token_pair = parse_token(vocab, pos);
+            const char * token_end  = token_pair.second;
+            last_sym_start = rule.size();
+            rule.push_back({type, token_pair.first});
+            pos = parse_space(token_end, is_nested);
         } else if (is_word_char(*pos)) { // rule reference
             const char * name_end    = parse_name(pos);
             uint32_t ref_rule_id = get_symbol_id(pos, name_end - pos);
@@ -691,6 +772,21 @@ static bool llama_grammar_match_partial_char(
     return !is_positive_char;
 }
 
+// returns true iff token matches the rule at pos (regular or inverse)
+// asserts that pos is pointing to a token element
+static bool llama_grammar_match_token(
+    const llama_grammar_element * pos,
+    const llama_token             token) {
+    GGML_ASSERT(pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT);
+    if (pos->type == LLAMA_GRETYPE_TOKEN) {
+        return pos->value == static_cast<uint32_t>(token);
+    }
+    if (pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+        return pos->value != static_cast<uint32_t>(token);
+    }
+    return false;
+}
+
 // transforms a grammar pushdown stack into N possible stacks, all ending
 // at a character range (terminal element)
 static void llama_grammar_advance_stack(
@@ -738,6 +834,8 @@ static void llama_grammar_advance_stack(
         case LLAMA_GRETYPE_CHAR:
         case LLAMA_GRETYPE_CHAR_NOT:
         case LLAMA_GRETYPE_CHAR_ANY:
+        case LLAMA_GRETYPE_TOKEN:
+        case LLAMA_GRETYPE_TOKEN_NOT:
             if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) {
                 // only add the stack if it's not a duplicate of one we already have
                 new_stacks.emplace_back(stack);
@@ -831,26 +929,38 @@ llama_grammar_stacks & llama_grammar_get_stacks(struct llama_grammar * grammar)
     return grammar->stacks;
 }
 
+static void llama_grammar_accept_chr(
+        struct llama_grammar       & grammar,
+        const llama_grammar_stack  & stack,
+              uint32_t               chr,
+              llama_grammar_stacks & new_stacks) {
+    if (stack.empty()) {
+        return;
+    }
+
+    const llama_grammar_element * pos = stack.back();
+
+    // ignore if this turns into a token
+    if (pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+        return;
+    }
+
+    auto match = llama_grammar_match_char(pos, chr);
+    if (match.first) {
+        llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
+        if (!llama_grammar_is_end_of_sequence(match.second)) {
+            new_stack.push_back(match.second);
+        }
+        llama_grammar_advance_stack(grammar.rules, new_stack, new_stacks);
+    }
+}
+
 void llama_grammar_accept(struct llama_grammar * grammar, uint32_t chr) {
     llama_grammar_stacks stacks_new;
     stacks_new.reserve(grammar->stacks.size());
 
     for (const auto & stack : grammar->stacks) {
-        if (stack.empty()) {
+        llama_grammar_accept_chr(*grammar, stack, chr, stacks_new);
-            continue;
-        }
-
-        auto match = llama_grammar_match_char(stack.back(), chr);
-        if (match.first) {
-            const llama_grammar_element * pos = match.second;
-
-            // update top of stack to next element, if any
-            llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
-            if (!llama_grammar_is_end_of_sequence(pos)) {
-                new_stack.push_back(pos);
-            }
-            llama_grammar_advance_stack(grammar->rules, new_stack, stacks_new);
-        }
     }
 
     grammar->stacks = std::move(stacks_new);
@@ -875,6 +985,22 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
 
     const llama_grammar_element * stack_pos = stack.back();
 
+    // if the top of the stack is a token rule, then we only need to check the token id
+    if (stack_pos->type == LLAMA_GRETYPE_TOKEN || stack_pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+        for (const auto & tok : candidates) {
+            if (*tok.code_points == 0) {
+                // reached the end of a token consumed by char rules, reject iff it ended
+                // in a partial response
+                if (tok.partial_utf8.n_remain != 0) {
+                    rejects.push_back(tok);
+                }
+            } else if (!llama_grammar_match_token(stack_pos, tok.id)) {
+                rejects.push_back(tok);
+            }
+        }
+        return rejects;
+    }
+
     llama_grammar_candidates next_candidates;
     next_candidates.reserve(candidates.size());
 
@@ -887,7 +1013,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
                 rejects.push_back(tok);
             }
         } else if (llama_grammar_match_char(stack_pos, *tok.code_points).first) {
-            next_candidates.push_back({ tok.index, tok.code_points + 1, tok.partial_utf8 });
+            next_candidates.push_back({ tok.index, tok.code_points + 1, tok.partial_utf8, tok.id });
         } else {
             rejects.push_back(tok);
         }
@@ -905,7 +1031,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
 
     auto next_rejects = llama_grammar_reject_candidates(rules, next_stacks, next_candidates);
     for (const auto & tok : next_rejects) {
-        rejects.push_back({ tok.index, tok.code_points - 1, tok.partial_utf8 });
+        rejects.push_back({ tok.index, tok.code_points - 1, tok.partial_utf8, tok.id });
     }
 
     return rejects;
@@ -978,6 +1104,7 @@ struct llama_grammar * llama_grammar_init_impl(
         /* .lazy = */                     false,
         /* .awaiting_trigger = */         false,
         /* .trigger_buffer = */           "",
+        /* .trigger_buffer_positions = */ {},
         /* .trigger_tokens = */           {},
         /* .trigger_patterns = */         {},
     };
@@ -993,7 +1120,7 @@ struct llama_grammar * llama_grammar_init_impl(
                             size_t num_trigger_patterns,
                const llama_token * trigger_tokens,
                             size_t num_trigger_tokens) {
-    llama_grammar_parser parser;
+    llama_grammar_parser parser(vocab);
 
     // if there is a grammar, parse it
     // rules will be empty (default) if there are parse errors
@@ -1085,6 +1212,7 @@ struct llama_grammar * llama_grammar_init_impl(
         /* .lazy = */                     lazy,
         /* .awaiting_trigger = */         lazy,
         /* .trigger_buffer = */           "",
+        /* .trigger_buffer_positions = */ {},
         std::move(vec_trigger_tokens),
         std::move(vec_trigger_patterns),
     };
@@ -1108,6 +1236,7 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
         grammar.lazy,
         grammar.awaiting_trigger,
         grammar.trigger_buffer,
+        grammar.trigger_buffer_positions,
         grammar.trigger_tokens,
         grammar.trigger_patterns,
     };
@@ -1164,7 +1293,7 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
             cur_p->data[i].logit = -INFINITY;
         } else {
             candidates_decoded.push_back(decode_utf8(piece, grammar.partial_utf8));
-            candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second });
+            candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second, id });
         }
     }
 
@@ -1184,10 +1313,12 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
         if (std::find(grammar.trigger_tokens.begin(), grammar.trigger_tokens.end(), token) != grammar.trigger_tokens.end()) {
             grammar.awaiting_trigger = false;
             grammar.trigger_buffer.clear();
-            llama_grammar_accept_str(grammar, piece);
+            llama_grammar_accept_token(grammar, token, piece);
             LLAMA_LOG_DEBUG("Grammar triggered on token %u (`%s`)", token, piece.c_str());
             return;
         } else {
+            auto position = std::make_pair(grammar.trigger_buffer.size(), grammar.trigger_buffer.size() + piece.size());
+            grammar.trigger_buffer_positions.push_back(std::make_pair(token, position));
             grammar.trigger_buffer += piece;
 
             std::smatch match;
@@ -1205,10 +1336,23 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
                     if (start == std::string::npos) {
                         start = match.position(0);
                     }
+
+                    // replay tokens that overlap with [start, end)
+                    for (const auto & [tok, tok_pos] : grammar.trigger_buffer_positions) {
+                        auto [tok_start, tok_end] = tok_pos;
+                        if (tok_end <= start) {
+                            continue;
+                        }
+
+                        size_t piece_start = (tok_start < start) ? start : tok_start; // allow for partial token pieces
+                        size_t piece_len = tok_end - piece_start;
+                        auto tok_piece = grammar.trigger_buffer.substr(piece_start, piece_len);
+                        llama_grammar_accept_token(grammar, tok, tok_piece);
+                    }
+
                     auto constrained_str = grammar.trigger_buffer.substr(start);
-                    // std::string constrained_str(match[1].first, grammar.trigger_buffer.end());
                     grammar.trigger_buffer.clear();
-                    llama_grammar_accept_str(grammar, constrained_str);
+                    grammar.trigger_buffer_positions.clear();
                     LLAMA_LOG_DEBUG("Grammar triggered on regex: '%s'\n", constrained_str.c_str());
                     return;
                 }
@@ -1228,7 +1372,7 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
         GGML_ABORT("grammar error: end of grammar token received but grammar stack is not empty");
     }
 
-    llama_grammar_accept_str(grammar, piece);
+    llama_grammar_accept_token(grammar, token, piece);
 }
 
 void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string & piece) {
@@ -1246,6 +1390,61 @@ void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string
     }
 }
 
+void llama_grammar_accept_token(struct llama_grammar & grammar, llama_token token, const std::string & piece) {
+    // Note terminating 0 in decoded string
+    const auto   decoded     = decode_utf8(piece, grammar.partial_utf8);
+    const auto & code_points = decoded.first;
+
+    llama_grammar_stacks stacks_new;
+    stacks_new.reserve(grammar.stacks.size());
+
+    for (const auto & stack : grammar.stacks) {
+        if (stack.empty()) {
+            continue;
+        }
+
+        const llama_grammar_element * pos = stack.back();
+
+        if (pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+            if (llama_grammar_match_token(pos, token)) {
+                llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
+                if (!llama_grammar_is_end_of_sequence(pos + 1)) {
+                    new_stack.push_back(pos + 1);
+                }
+                llama_grammar_advance_stack(grammar.rules, new_stack, stacks_new);
+            }
+        } else {
+            llama_grammar_stacks current_stacks = {stack};
+
+            for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
+                llama_grammar_stacks next_stacks;
+
+                for (const auto & cur_stack : current_stacks) {
+                    llama_grammar_accept_chr(grammar, cur_stack, *it, next_stacks);
+                }
+
+                current_stacks = std::move(next_stacks);
+                if (current_stacks.empty()) {
+                    break;
+                }
+            }
+
+            for (auto & surviving_stack : current_stacks) {
+                if (std::find(stacks_new.begin(), stacks_new.end(), surviving_stack) == stacks_new.end()) {
+                    stacks_new.emplace_back(surviving_stack);
+                }
+            }
+        }
+    }
+
+    grammar.stacks = std::move(stacks_new);
+    grammar.partial_utf8 = decoded.second;
+
+    if (grammar.stacks.empty()) {
+        throw std::runtime_error("Unexpected empty grammar stack after accepting piece: " + piece + " (" + std::to_string(token) + ")");
+    }
+}
+
 
 const std::string & ollama_vocab::token_to_piece(const uint32_t token) const {
     try {

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

@@ -47,11 +47,17 @@ enum llama_gretype {
 
     // any character (.)
     LLAMA_GRETYPE_CHAR_ANY       = 7,
+
+    // terminal element: token (<[token-id]>)
+    LLAMA_GRETYPE_TOKEN          = 8,
+
+    // inverse token (!<[token-id]>)
+    LLAMA_GRETYPE_TOKEN_NOT      = 9,
 };
 
 typedef struct llama_grammar_element {
     enum llama_gretype type;
-    uint32_t           value; // Unicode code point or rule ID
+    uint32_t           value; // Unicode code point, rule ID, or token ID
 } llama_grammar_element;
 
 struct llama_partial_utf8 {
@@ -63,6 +69,7 @@ struct llama_grammar_candidate {
     size_t               index;
     const uint32_t     * code_points;
     llama_partial_utf8   partial_utf8;
+    llama_token          id;
 };
 
 using llama_grammar_rule  = std::vector<      llama_grammar_element>;
@@ -88,10 +95,13 @@ std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_stack(
         const llama_grammar_candidates & candidates);
 
 struct llama_grammar_parser {
+    const llama_vocab * vocab;
     std::map<std::string, uint32_t> symbol_ids;
 
     llama_grammar_rules rules;
 
+    llama_grammar_parser(const struct llama_vocab * vocab = nullptr) : vocab(vocab) {}
+
     llama_grammar_stack c_rules() const;
 
     uint32_t get_symbol_id(const char * src, size_t len);
@@ -123,6 +133,9 @@ struct llama_grammar_trigger_pattern {
 };
 
 struct llama_grammar {
+    // maintain a list of llama_tokens and their positions in the trigger_buffer
+    using token_pos = std::pair<llama_token, std::pair<size_t, size_t>>;
+
     // note: allow null vocab for testing (not great)
     const llama_vocab * vocab;
     const ollama_vocab * o_vocab;
@@ -139,6 +152,7 @@ struct llama_grammar {
     bool                     lazy             = false;
     bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only
     std::string              trigger_buffer;           // Output buffered by lazy grammar. Will be cleared once trigger is found.
+    std::vector<token_pos>   trigger_buffer_positions; // Tokens buffered by lazy grammar. Used to replay when a trigger is found.
     std::vector<llama_token> trigger_tokens;           // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
     std::vector<llama_grammar_trigger_pattern>
                              trigger_patterns;         // Regular expressions that trigger a lazy grammar. Must be a full match of the entire generated
@@ -185,3 +199,8 @@ void llama_grammar_accept_impl(
 void llama_grammar_accept_str(
               struct llama_grammar & grammar,
                  const std::string & piece);
+
+void llama_grammar_accept_token(
+              struct llama_grammar & grammar,
+                       llama_token   token,
+                 const std::string & piece);

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

@@ -71,11 +71,14 @@ void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && attn_scale) {
         const int64_t n_tokens = ubatch->n_tokens;
 
+        GGML_ASSERT(f_attn_temp_scale != 0.0f);
+        GGML_ASSERT(n_attn_temp_floor_scale != 0);
+
         std::vector<float> attn_scale_data(n_tokens, 0.0f);
         for (int i = 0; i < n_tokens; ++i) {
             const float pos = ubatch->pos[i];
             attn_scale_data[i] = std::log(
-                std::floor((pos + 1.0f) / n_attn_temp_floor_scale) + 1.0
+                std::floor((pos + f_attn_temp_offset) / n_attn_temp_floor_scale) + 1.0
             ) * f_attn_temp_scale + 1.0;
         }
 
@@ -251,6 +254,24 @@ void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+bool llm_graph_input_rs::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_memory_recurrent_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= s_copy->ne[0] == mctx->get_n_rs();
+
+    res &= s_copy_main->ne[0]  == params.ubatch.n_seqs;
+    res &= s_copy_extra->ne[0] == mctx->get_n_rs() - params.ubatch.n_seqs;
+
+    res &= head == mctx->get_head();
+    res &= rs_z == mctx->get_rs_z();
+
+    return res;
+}
+
 void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
@@ -382,7 +403,7 @@ bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) {
   //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
 
     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 &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
 
     return res;
 }
@@ -413,10 +434,10 @@ bool llm_graph_input_attn_kv_iswa::can_reuse(const llm_graph_params & params) {
   //res &= self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
 
     res &= self_kq_mask->ne[0] == mctx->get_base()->get_n_kv();
-    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+    res &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
 
     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 &= self_kq_mask_swa->ne[1] == params.ubatch.n_tokens;
 
     return res;
 }
@@ -449,7 +470,7 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
             }
         }
 
-        for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
+        for (int i = n_tokens; i < n_tokens; ++i) {
             for (int j = 0; j < n_enc; ++j) {
                 data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY;
             }
@@ -458,8 +479,46 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
-    inp_attn->set_input(ubatch);
+    mctx->get_attn()->set_input_k_idxs(inp_attn->self_k_idxs, ubatch);
-    inp_rs->set_input(ubatch);
+    mctx->get_attn()->set_input_v_idxs(inp_attn->self_v_idxs, ubatch);
+
+    mctx->get_attn()->set_input_kq_mask(inp_attn->self_kq_mask, ubatch, cparams.causal_attn);
+
+    const int64_t n_rs = mctx->get_recr()->get_n_rs();
+
+    if (inp_rs->s_copy) {
+        GGML_ASSERT(ggml_backend_buffer_is_host(inp_rs->s_copy->buffer));
+        int32_t * data = (int32_t *) inp_rs->s_copy->data;
+
+        // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
+        for (uint32_t i = 0; i < n_rs; ++i) {
+            data[i] = mctx->get_recr()->s_copy(i);
+        }
+    }
+}
+
+bool llm_graph_input_mem_hybrid::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_memory_hybrid_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= inp_attn->self_k_idxs->ne[0] == params.ubatch.n_tokens;
+  //res &= inp_attn->self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+    res &= inp_attn->self_kq_mask->ne[0] == mctx->get_attn()->get_n_kv();
+    res &= inp_attn->self_kq_mask->ne[1] == params.ubatch.n_tokens;
+
+    res &= inp_rs->s_copy->ne[0] == mctx->get_recr()->get_n_rs();
+
+    res &= inp_rs->s_copy_main->ne[0]  == params.ubatch.n_seqs;
+    res &= inp_rs->s_copy_extra->ne[0] == mctx->get_recr()->get_n_rs() - params.ubatch.n_seqs;
+
+    res &= inp_rs->head == mctx->get_recr()->get_head();
+    res &= inp_rs->rs_z == mctx->get_recr()->get_rs_z();
+
+    return res;
 }
 
 //
@@ -810,9 +869,6 @@ ggml_tensor * llm_graph_context::build_ffn(
             GGML_ABORT("fatal error");
     }
 
-    //expand here so that we can fuse ffn gate
-    ggml_build_forward_expand(gf, cur);
-
     if (gate && type_gate == LLM_FFN_PAR) {
         cur = ggml_mul(ctx0, cur, tmp);
         cb(cur, "ffn_gate_par", il);
@@ -973,7 +1029,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
         // mask out the other groups
         selection_probs = ggml_get_rows(ctx0, selection_groups, expert_groups); // [n_exp_per_group, n_group_used, n_tokens]
-        selection_probs = ggml_set_rows(ctx0, ggml_scale_bias(ctx0, selection_groups, 0.0f, -INFINITY), selection_probs, expert_groups); // [n_exp_per_group, n_expert_groups, n_tokens]
+        selection_probs = ggml_set_rows(ctx0, ggml_fill(ctx0, selection_groups, -INFINITY), selection_probs, expert_groups); // [n_exp_per_group, n_expert_groups, n_tokens]
         selection_probs = ggml_reshape_2d(ctx0, selection_probs, n_expert, n_tokens); // [n_expert, n_tokens]
         cb(selection_probs, "ffn_moe_probs_masked", il);
     }
@@ -1089,13 +1145,19 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 cur = ggml_relu(ctx0, cur);
                 cb(cur, "ffn_moe_relu", il);
             } break;
+        case LLM_FFN_RELU_SQR:
+            if (gate_exps) {
+                // TODO: add support for gated squared relu
+                GGML_ABORT("fatal error: gated squared relu not implemented");
+            } else {
+                cur = ggml_relu(ctx0, cur);
+                cur = ggml_sqr(ctx0, cur);
+                cb(cur, "ffn_moe_relu_sqr", il);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
 
-    //expand here so that we can fuse ffn gate
-    ggml_build_forward_expand(gf, cur);
-
     experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 
@@ -1206,7 +1268,7 @@ ggml_tensor * llm_graph_context::build_inp_pos() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
-    auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
+    auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale, hparams.f_attn_temp_offset);
 
     auto & cur = inp->attn_scale;
 
@@ -1473,13 +1535,13 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
     auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);
 
     // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
-    inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+    inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens, 1, 1);
     ggml_set_input(inp->self_kq_mask);
 
     inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
 
     if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
-        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens, 1, 1);
         ggml_set_input(inp->self_kq_mask_swa);
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
@@ -1561,7 +1623,7 @@ static std::unique_ptr<llm_graph_input_attn_kv> build_attn_inp_kv_impl(
         inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1704,7 +1766,7 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
 
     const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
 
-    inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+    inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, n_tokens, 1, 1);
     ggml_set_input(inp->cross_kq_mask);
 
     inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->cross_kq_mask, GGML_TYPE_F16) : inp->cross_kq_mask;
@@ -1770,7 +1832,7 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
         inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1784,7 +1846,7 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
         inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp->self_kq_mask_swa);
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
@@ -1844,6 +1906,9 @@ static std::unique_ptr<llm_graph_input_rs> build_rs_inp_impl(
     inp->s_copy_main  = ggml_view_1d(ctx0, inp->s_copy, n_seqs, 0);
     inp->s_copy_extra = ggml_view_1d(ctx0, inp->s_copy, n_rs - n_seqs, n_seqs * inp->s_copy->nb[0]);
 
+    inp->head = mctx_cur->get_head();
+    inp->rs_z = mctx_cur->get_rs_z();
+
     return inp;
 }
 
@@ -1915,7 +1980,7 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
     auto inp_rs   = build_rs_inp_impl     (ctx0, ubatch, mctx_cur->get_recr());
     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);
+    auto inp = std::make_unique<llm_graph_input_mem_hybrid>(cparams, std::move(inp_attn), std::move(inp_rs), mctx_cur);
 
     return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
 }

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

@@ -132,8 +132,8 @@ public:
 // temperature tuning, used by llama4
 class llm_graph_input_attn_temp : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
+    llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale, float f_attn_temp_offset)
-        : n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
+        : n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale), f_attn_temp_offset(f_attn_temp_offset) {}
     virtual ~llm_graph_input_attn_temp() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -142,6 +142,7 @@ public:
 
     const uint32_t n_attn_temp_floor_scale;
     const float    f_attn_temp_scale;
+    const float    f_attn_temp_offset;
 };
 
 class llm_graph_input_pos_bucket : public llm_graph_input_i {
@@ -224,6 +225,8 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * s_copy;  // I32 [n_rs]
 
     // views of s_copy, computed once per graph
@@ -232,6 +235,10 @@ public:
     ggml_tensor * s_copy_extra;  // I32 [n_rs - n_seqs]
 
     const llama_memory_recurrent_context * mctx;
+
+    // used in view offsets, need to match for valid graph reuse
+    uint32_t head;
+    int32_t rs_z;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -364,22 +371,28 @@ public:
 class llm_graph_input_mem_hybrid : public llm_graph_input_i {
 public:
     llm_graph_input_mem_hybrid(
+            const llama_cparams & cparams,
             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)),
         inp_rs(std::move(inp_rs)),
+        cparams(cparams),
         mctx(mctx) { }
     virtual ~llm_graph_input_mem_hybrid() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     std::unique_ptr<llm_graph_input_attn_kv> inp_attn;
     std::unique_ptr<llm_graph_input_rs>      inp_rs;
 
     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_cparams cparams;
+
     const llama_memory_hybrid_context * mctx;
 };
 

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

@@ -1,6 +1,8 @@
 #include "llama-hparams.h"
 
 #include "ggml.h"
+
+#include <algorithm>
 #include <cassert>
 
 void llama_hparams::set_swa_pattern(uint32_t n_pattern, bool dense_first) {
@@ -237,3 +239,7 @@ bool llama_hparams::is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama
 
     return false;
 }
+
+bool llama_hparams::use_mrope() const {
+    return rope_sections[0] > 0 && rope_sections[1] > 0;
+}

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

@@ -34,6 +34,7 @@ struct llama_hparams_convnext {
 
 struct llama_hparams {
     bool vocab_only;
+    bool no_alloc;
     bool rope_finetuned;
     bool use_par_res;
     bool swin_norm;
@@ -109,6 +110,7 @@ struct llama_hparams {
     float    rope_freq_base_train_swa;
     float    rope_freq_scale_train;
     float    rope_freq_scale_train_swa;
+
     uint32_t n_ctx_orig_yarn;
     float    rope_yarn_log_mul = 0.0f;
 
@@ -164,8 +166,9 @@ struct llama_hparams {
     // llama4 smallthinker
     uint32_t n_moe_layer_step        = 0;
     uint32_t n_no_rope_layer_step    = 4;
-    uint32_t n_attn_temp_floor_scale = 8192;
+    uint32_t n_attn_temp_floor_scale = 0;
-    float    f_attn_temp_scale       = 0.1;
+    float    f_attn_temp_scale       = 0.0f;
+    float    f_attn_temp_offset      = 0.0f; // offset position index
 
     // gemma3n altup
     uint32_t n_altup      = 4; // altup_num_inputs
@@ -272,7 +275,8 @@ struct llama_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);
+
+    bool use_mrope() const;
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
-

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

@@ -25,6 +25,10 @@ time_meas::~time_meas() {
     }
 }
 
+void llama_log_get(ggml_log_callback * log_callback, void ** user_data) {
+    ggml_log_get(log_callback, user_data);
+}
+
 void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     ggml_log_set(log_callback, user_data);
     g_logger_state.log_callback = log_callback ? log_callback : llama_log_callback_default;

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

@@ -37,7 +37,7 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void *
 template <typename T>
 struct no_init {
     T value;
-    no_init() { /* do nothing */ }
+    no_init() = default;
 };
 
 struct time_meas {

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

@@ -175,7 +175,15 @@ llama_kv_cache::llama_kv_cache(
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
     for (auto & [buft, ctx] : ctx_map) {
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
+        ggml_backend_buffer_t buf;
+        if (model.hparams.no_alloc) {
+            buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
+            for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != nullptr; t = ggml_get_next_tensor(ctx.get(), t)) {
+                t->buffer = buf; // set dummy buffer for KV cache so that the backend scheduler won't try to allocate it
+            }
+        } else {
+            buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft); // real buffer
+        }
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for kv cache");
         }
@@ -482,9 +490,18 @@ llama_pos llama_kv_cache::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, buf] : ctxs_bufs) {
+    for (const auto & [ctx, buf] : ctxs_bufs) {
-        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+        ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf.get());
+
+        if (hparams.no_alloc) {
+            GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) == nullptr);
+            ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
+        } else {
+            // GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) != nullptr); // multi_buffer does not have a defined base
+            ret[buft] += ggml_backend_buffer_get_size(buf.get());
         }
+    }
+
     return ret;
 }
 
@@ -1233,7 +1250,6 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
 
     // n_tps == n_tokens_per_stream
     const int64_t n_tps = n_tokens/n_stream;
-    const int64_t n_tps_pad = GGML_PAD(n_tps, GGML_KQ_MASK_PAD);
 
     std::fill(data, data + ggml_nelements(dst), -INFINITY);
 
@@ -1266,7 +1282,7 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
                 const llama_pos p1_x = is_2d ? ubatch->pos[i + ubatch->n_tokens*2] : 0;
                 const llama_pos p1_y = is_2d ? ubatch->pos[i + ubatch->n_tokens]   : 0;
 
-                const uint64_t idst = n_kv*(h*n_stream*n_tps_pad + s*n_tps_pad + ii);
+                const uint64_t idst = n_kv*(h*n_stream*n_tps + s*n_tps + ii);
 
                 for (uint32_t j = 0; j < n_kv; ++j) {
                     if (cells.is_empty(j)) {
@@ -1373,6 +1389,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
     const auto & yarn_ext_factor  = cparams.yarn_ext_factor;
     const auto & yarn_beta_fast   = cparams.yarn_beta_fast;
     const auto & yarn_beta_slow   = cparams.yarn_beta_slow;
+    const auto & yarn_attn_factor = cparams.yarn_attn_factor;
 
     const auto & n_rot     = hparams.n_rot;
     const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
@@ -1383,12 +1400,6 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
                                 ? LLAMA_ROPE_TYPE_NEOX
                                 : hparams.rope_type;
 
-    // See llm_build_deepseek2() for why attn_factor has to be scaled for YaRN RoPE to work correctly.
-    // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
-    const float yarn_attn_factor = model.arch == LLM_ARCH_DEEPSEEK2
-                                    ? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale))
-                                    : cparams.yarn_attn_factor;
-
     ggml_tensor * tmp;
 
     if (ggml_is_quantized(cur->type)) {
@@ -1550,9 +1561,11 @@ void llama_kv_cache::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama
 
         const uint32_t strm = seq_id == -1 ? s : seq_to_stream[seq_id];
 
+        slot_info sinfo;
+
         bool res = true;
-        res = res && state_read_meta(io, strm, cell_count, seq_id);
+        res = res && state_read_meta(io, strm, cell_count, sinfo, seq_id);
-        res = res && state_read_data(io, strm, cell_count);
+        res = res && state_read_data(io, strm, cell_count, sinfo);
 
         if (!res) {
             if (seq_id == -1) {
@@ -1691,7 +1704,7 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
     }
 }
 
-bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, llama_seq_id dest_seq_id) {
+bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, slot_info & sinfo, llama_seq_id dest_seq_id) {
     auto & cells = v_cells[strm];
     auto & head  = v_heads[strm];
 
@@ -1728,7 +1741,7 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             ubatch.seq_id[i]   = &dest_seq_id;
         }
 
-        const auto sinfo = find_slot(ubatch, true);
+        sinfo = find_slot(ubatch, false);
         if (sinfo.empty()) {
             LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
             return false;
@@ -1738,20 +1751,16 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
         //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
         apply_ubatch(sinfo, ubatch);
 
-        const auto head_cur = sinfo.head();
+        LLAMA_LOG_DEBUG("%s: cell_count = %d, dest_seq_id = %d\n", __func__, cell_count, dest_seq_id);
 
-        // keep the head at the old position because we will read the KV data into it in state_read_data()
+        // DEBUG CHECK: verify that all cells were allocated and have correct seq_id and pos values
-        head = head_cur;
+        GGML_ASSERT(sinfo.n_stream() == 1);
-
+        GGML_ASSERT(sinfo.idxs[0].size() == cell_count);
-        LLAMA_LOG_DEBUG("%s: head_cur = %d, head = %d, cell_count = %d, dest_seq_id = %d\n", __func__, head_cur, head, cell_count, dest_seq_id);
+        for (uint32_t i = 0; i < cell_count; ++i) {
-
+            const uint32_t idx = sinfo.idxs[0][i];
-        // DEBUG CHECK: head_cur should be our first cell, head_cur + cell_count - 1 should be our last cell (verify seq_id and pos values)
+            GGML_ASSERT(cells.pos_get(idx) == ubatch.pos[i]);
-        // Assume that this is one contiguous block of cells
+            GGML_ASSERT(cells.seq_has(idx, dest_seq_id));
-        GGML_ASSERT(head_cur + cell_count <= cells.size());
+        }
-        GGML_ASSERT(cells.pos_get(head_cur)                  == ubatch.pos[0]);
-        GGML_ASSERT(cells.pos_get(head_cur + cell_count - 1) == ubatch.pos[cell_count - 1]);
-        GGML_ASSERT(cells.seq_has(head_cur,                  dest_seq_id));
-        GGML_ASSERT(cells.seq_has(head_cur + cell_count - 1, dest_seq_id));
     } else {
         // whole KV cache restore
 
@@ -1784,15 +1793,24 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             }
         }
 
+        // Create contiguous slot_info for whole cache restore
+        sinfo.s0 = strm;
+        sinfo.s1 = strm;
+        sinfo.resize(1);
+        sinfo.strm[0] = strm;
+        sinfo.idxs[0].resize(cell_count);
+        for (uint32_t i = 0; i < cell_count; ++i) {
+            sinfo.idxs[0][i] = i;
+        }
+
         head = 0;
     }
 
     return true;
 }
 
-bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count) {
+bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, const slot_info & sinfo) {
     auto & cells = v_cells[strm];
-    auto & head  = v_heads[strm];
 
     uint32_t v_trans;
     uint32_t n_layer;
@@ -1842,8 +1860,17 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
         }
 
         if (cell_count) {
-            // Read and set the keys for the whole cell range
+            if (sinfo.is_contiguous()) {
-            ggml_backend_tensor_set(k, io.read(cell_count * k_size_row), head * k_size_row, cell_count * k_size_row);
+                // Fast path: contiguous cells, single memcpy
+                ggml_backend_tensor_set(k, io.read(cell_count * k_size_row), sinfo.head() * k_size_row, cell_count * k_size_row);
+            } else {
+                // Slow path: scatter to non-contiguous positions
+                const void * src = io.read(cell_count * k_size_row);
+                for (uint32_t i = 0; i < cell_count; ++i) {
+                    const size_t dst_offset = sinfo.idxs[0][i] * k_size_row;
+                    ggml_backend_tensor_set(k, (const char*)src + i * k_size_row, dst_offset, k_size_row);
+                }
+            }
         }
     }
 
@@ -1874,8 +1901,17 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             }
 
             if (cell_count) {
-                // Read and set the values for the whole cell range
+                if (sinfo.is_contiguous()) {
-                ggml_backend_tensor_set(v, io.read(cell_count * v_size_row), head * v_size_row, cell_count * v_size_row);
+                    // Fast path: contiguous cells, single memcpy
+                    ggml_backend_tensor_set(v, io.read(cell_count * v_size_row), sinfo.head() * v_size_row, cell_count * v_size_row);
+                } else {
+                    // Slow path: scatter to non-contiguous positions
+                    const void * src = io.read(cell_count * v_size_row);
+                    for (uint32_t i = 0; i < cell_count; ++i) {
+                        const size_t dst_offset = sinfo.idxs[0][i] * v_size_row;
+                        ggml_backend_tensor_set(v, (const char*)src + i * v_size_row, dst_offset, v_size_row);
+                    }
+                }
             }
         }
     } else {
@@ -1914,11 +1950,23 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             }
 
             if (cell_count) {
-                // For each row in the transposed matrix, read the values for the whole cell range
+                if (sinfo.is_contiguous()) {
+                    // Fast path: contiguous cells
+                    const uint32_t h = sinfo.head();
                     for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                    const size_t dst_offset = (head + j * cells.size()) * v_size_el;
+                        const size_t dst_offset = (h + j * cells.size()) * v_size_el;
                         ggml_backend_tensor_set(v, io.read(cell_count * v_size_el), dst_offset, cell_count * v_size_el);
                     }
+                } else {
+                    // Slow path: scatter to non-contiguous positions
+                    for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                        const void * src = io.read(cell_count * v_size_el);
+                        for (uint32_t i = 0; i < cell_count; ++i) {
+                            const size_t dst_offset = (sinfo.idxs[0][i] + j * cells.size()) * v_size_el;
+                            ggml_backend_tensor_set(v, (const char*)src + i * v_size_el, dst_offset, v_size_el);
+                        }
+                    }
+                }
             }
         }
     }

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

@@ -72,6 +72,23 @@ public:
         void clear() {
             idxs.clear();
         }
+
+        // check if indices are contiguous starting from head()
+        bool is_contiguous() const {
+            if (idxs.empty() || idxs[0].empty()) {
+                return true;
+            }
+            if (idxs.size() > 1) {
+                return false;
+            }
+            const uint32_t h = idxs[0][0];
+            for (size_t i = 0; i < idxs[0].size(); ++i) {
+                if (idxs[0][i] != h + i) {
+                    return false;
+                }
+            }
+            return true;
+        }
     };
 
     using slot_info_vec_t = std::vector<slot_info>;
@@ -264,8 +281,8 @@ private:
     void state_write_meta(llama_io_write_i & io, const cell_ranges_t & cr, llama_seq_id seq_id = -1) const;
     void state_write_data(llama_io_write_i & io, const cell_ranges_t & cr) const;
 
-    bool state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
+    bool state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count,       slot_info & sinfo, llama_seq_id dest_seq_id = -1);
-    bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count);
+    bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, const slot_info & sinfo);
 };
 
 class llama_kv_cache_context : public llama_memory_context_i {

llama/llama.cpp/src/llama-mmap.cpp

@@ -485,7 +485,7 @@ struct llama_mlock::impl {
         if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) {
             suggest = false;
         }
-        if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size)) {
+        if (suggest && ((uint64_t)lock_limit.rlim_max > (uint64_t)lock_limit.rlim_cur + size)) {
             suggest = false;
         }
 #endif

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

@@ -473,6 +473,7 @@ llama_model_loader::llama_model_loader(
         std::vector<std::string> & splits,
         bool use_mmap,
         bool check_tensors,
+        bool no_alloc,
         const llama_model_kv_override * param_overrides_p,
         const llama_model_tensor_buft_override * param_tensor_buft_overrides_p) {
     int trace = 0;
@@ -716,6 +717,7 @@ llama_model_loader::llama_model_loader(
 
     this->use_mmap = use_mmap;
     this->check_tensors = check_tensors;
+    this->no_alloc = no_alloc;
 }
 
 std::string llama_model_loader::get_arch_name() const {

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

@@ -71,6 +71,7 @@ struct llama_model_loader {
 
     bool use_mmap = false;
     bool check_tensors;
+    bool no_alloc;
 
     llama_files files;
     llama_ftype ftype;
@@ -97,6 +98,7 @@ struct llama_model_loader {
         std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
         bool use_mmap,
         bool check_tensors,
+        bool no_alloc,
         const llama_model_kv_override * param_overrides_p,
         const llama_model_tensor_buft_override * param_tensor_buft_overrides_p);
 

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

@@ -120,6 +120,8 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_16B_A1B:       return "16B.A1B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
+        case LLM_TYPE_31B_A3_5B:     return "31B.A3.5B";
+        case LLM_TYPE_80B_A3B:       return "80B.A3B";
         case LLM_TYPE_100B_A6B:      return "100B.A6B";
         case LLM_TYPE_106B_A12B:     return "106B.A12B";
         case LLM_TYPE_230B_A10B:     return "230B.A10B";
@@ -423,8 +425,8 @@ static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, llama_split_mode s
 }
 
 struct llama_model::impl {
-    impl() {}
+    impl() = default;
-    ~impl() {}
+    ~impl() = default;
 
     uint64_t n_elements = 0;
 
@@ -461,7 +463,7 @@ llama_model::llama_model(const llama_model_params & params) : params(params), pi
     pimpl->has_tensor_overrides = params.tensor_buft_overrides && params.tensor_buft_overrides[0].pattern;
 }
 
-llama_model::~llama_model() {}
+llama_model::~llama_model() = default;
 
 void llama_model::load_stats(llama_model_loader & ml) {
     pimpl->n_elements = ml.n_elements;
@@ -665,6 +667,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 } else {
                     hparams.swa_type                = LLAMA_SWA_TYPE_CHUNKED;
                     hparams.n_swa                   = 8192;
+                    hparams.n_attn_temp_floor_scale = 8192;
+                    hparams.f_attn_temp_scale       = 0.1f;
+                    hparams.f_attn_temp_offset      = 1.0f;
                     hparams.set_swa_pattern(4);   // pattern: 3 chunked - 1 full
                 }
 
@@ -1262,18 +1267,25 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_GEMMA3:
             {
+                const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+                if (found_swa && hparams.n_swa > 0) {
                     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                     hparams.set_swa_pattern(6);
 
                     hparams.rope_freq_base_train_swa  = 10000.0f;
                     hparams.rope_freq_scale_train_swa = 1.0f;
+                } else {
+                    hparams.swa_type = LLAMA_SWA_TYPE_NONE;
+                }
 
-                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
+                hparams.f_final_logit_softcapping = 0.0f;
+                ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
                 switch (hparams.n_layer) {
                     case 18: type = LLM_TYPE_270M; break;
                     case 26: type = LLM_TYPE_1B; break;
+                    case 32: type = LLM_TYPE_8B; break; // Rnj-1
                     case 34: type = LLM_TYPE_4B; break;
                     case 48: type = LLM_TYPE_12B; break;
                     case 62: type = LLM_TYPE_27B; break;
@@ -1597,8 +1609,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
                 ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale);
 
-                switch (hparams.n_layer) {
+                switch (hparams.n_ff_exp) {
-                    case 28: type = LLM_TYPE_20B; break;
+                    case 1408: type = LLM_TYPE_16B; break;
+                    case 1792: type = LLM_TYPE_20B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -1624,7 +1637,18 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     // that have no expert_gating_func model parameter set
                     hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;
                 }
-                ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, false);
+
+                if (ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, 0.0f)) {
+                    // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+                    // cancel the factor from the convert script
+                    hparams.rope_yarn_log_mul /= 0.1f;
+                }
+
+                // (optional) temperature tuning - used by mistral-large
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE,  hparams.f_attn_temp_scale,       false);
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.n_attn_temp_floor_scale, false);
+
+                hparams.f_attn_temp_offset = 0.0f;
 
                 switch (hparams.n_layer) {
                     case 27: type = LLM_TYPE_16B; break;
@@ -1666,6 +1690,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
         case LLM_ARCH_GLM4:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
                 switch (hparams.n_layer) {
                     case 40: type = LLM_TYPE_9B; break;
                     case 61: type = LLM_TYPE_32B; break;
@@ -1676,6 +1701,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             {
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,     hparams.n_ff_exp);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
 
                 // MoE parameters
                 ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert);
@@ -1774,6 +1800,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
             } break;
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             {
                 ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
                 ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
@@ -1789,7 +1816,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp,        false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp,      false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
+
                 switch (hparams.n_layer) {
+                    case 52: type = LLM_TYPE_31B_A3_5B; break; // Nemotron-H_MOE 31B
                     case 56: type = LLM_TYPE_9B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
@@ -2258,7 +2292,33 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
 
                 switch (hparams.n_layer) {
-                    case 80: type = LLM_TYPE_80B_A3B; break;
+                    case 48: type = LLM_TYPE_80B_A3B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_MISTRAL3:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE, hparams.f_attn_temp_scale, false);
+
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_FAST, hparams.yarn_beta_fast,    false);
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, hparams.yarn_beta_slow,    false);
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,   hparams.rope_yarn_log_mul, 0.0f);
+
+                hparams.f_attn_temp_offset = 0.0f;
+
+                // TODO: maybe add n_attn_temp_floor_scale as a separate KV?
+                if (hparams.f_attn_temp_scale != 0.0f) {
+                    hparams.n_attn_temp_floor_scale = hparams.n_ctx_orig_yarn;
+                    if (hparams.n_attn_temp_floor_scale == 0) {
+                        throw std::runtime_error("invalid n_ctx_orig_yarn for attention temperature scaling");
+                    }
+                }
+
+                switch (hparams.n_layer) {
+                    case 26: type = LLM_TYPE_3B; break;
+                    case 34: type = LLM_TYPE_8B; break;
+                    case 40: type = LLM_TYPE_14B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -2575,6 +2635,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             case LLM_ARCH_MINICPM:
             case LLM_ARCH_GRANITE:
             case LLM_ARCH_GRANITE_MOE:
+            case LLM_ARCH_MISTRAL3:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -3353,9 +3414,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
 
                         // optional bias tensors
-                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, 0);
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
-                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
-                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
@@ -5124,6 +5185,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
                 } break;
             case LLM_ARCH_NEMOTRON_H:
+            case LLM_ARCH_NEMOTRON_H_MOE:
                 {
                     // mamba2 Mixer SSM params
                     // NOTE: int64_t for tensor dimensions
@@ -5134,6 +5196,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_group    = hparams.ssm_n_group;
                     const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
 
+                    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+                    const int64_t n_ff_shexp = hparams.n_ff_shexp;
+
                     // embeddings
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -5183,6 +5248,19 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias",   i), {n_embd_k_gqa_i}, TENSOR_NOT_REQUIRED);
                             layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias",   i), {n_embd_v_gqa_i}, TENSOR_NOT_REQUIRED);
                             layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd},         TENSOR_NOT_REQUIRED);
+                        }  else {
+                            if (n_expert != 0) {
+                                layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, 0);
+                                layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, 0);
+
+                                // MoE branch
+                                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+                                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+                                // Shared expert branch
+                                layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
+                                layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
+
                             } else {
                                 // mlp layers
                                 layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, 0);
@@ -5191,6 +5269,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                                 layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);
                             }
                         }
+                    }
                 } break;
             case LLM_ARCH_EXAONE:
                 {
@@ -6530,7 +6609,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.ssm_in         = create_tensor(tn(LLM_TENSOR_SSM_IN,         "weight", i), { n_embd, qkvz_dim }, 0);
                             layer.ssm_conv1d     = create_tensor(tn(LLM_TENSOR_SSM_CONV1D,     "weight", i), { hparams.ssm_d_conv, conv_dim }, 0);
                             layer.ssm_dt         = create_tensor(tn(LLM_TENSOR_SSM_DT,         "bias",   i), { hparams.ssm_dt_rank }, 0);
-                            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A,                    i), { hparams.ssm_dt_rank }, 0);
+                            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN,             i), { hparams.ssm_dt_rank }, 0);
                             layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", i), { n_embd, ba_dim }, 0);
                             layer.ssm_norm       = create_tensor(tn(LLM_TENSOR_SSM_NORM,       "weight", i), { head_v_dim }, 0);
                             layer.ssm_out        = create_tensor(tn(LLM_TENSOR_SSM_OUT,        "weight", i), { value_dim, n_embd }, 0);
@@ -6599,9 +6678,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
         std::vector<ggml_backend_buffer_ptr> bufs;
         if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported && is_default_buft) {
+            GGML_ASSERT(!ml.no_alloc);
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                 // only the mmap region containing the tensors in the model is mapped to the backend buffer
-                // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
+                // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer,
+                //     then we could just use metal for all layers
                 // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
                 void * addr = nullptr;
                 size_t first, last; // NOLINT
@@ -6617,9 +6698,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 bufs.emplace_back(buf);
                 buf_map.emplace(idx, buf);
             }
+        } else {
+            ggml_backend_buffer_t buf;
+            if (ml.no_alloc) {
+                buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
+                for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
+                    t->buffer = buf; // set dummy buffer for weights so that the backend scheduler won't try to allocate them
+                }
+            } else {
+                buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); // real buffer
             }
-        else {
-            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
             if (buf == nullptr) {
                 throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
             }
@@ -6674,6 +6762,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         }
     }
 
+    if (ml.no_alloc) {
+        return true;
+    }
+
     // load tensor data
     for (auto & [ctx, buf_map] : ctx_buf_maps) {
         if (!ml.load_all_data(ctx, buf_map, use_mlock ? &pimpl->mlock_mmaps : NULL, params.progress_callback, params.progress_callback_user_data)) {
@@ -6716,11 +6808,20 @@ size_t llama_model::n_devices() const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_model::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, bufs] : pimpl->ctxs_bufs) {
+    for (const auto & [ctx, bufs] : pimpl->ctxs_bufs) {
+        if (hparams.no_alloc) {
+            GGML_ASSERT(bufs.size() == 1);
+            ggml_backend_buffer_t buf = bufs[0].get();
+            GGML_ASSERT(ggml_backend_buffer_get_base(buf) == nullptr);
+            ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf);
+            ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
+        } else {
             for (const auto & buf : bufs) {
+                // GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) != nullptr); // multi_buffer does not have a defined base
                 ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
             }
         }
+    }
     return ret;
 }
 
@@ -6763,6 +6864,7 @@ void llama_model::print_info() const {
     // hparams
     LLAMA_LOG_INFO("%s: arch             = %s\n",     __func__, arch_name().c_str());
     LLAMA_LOG_INFO("%s: vocab_only       = %d\n",     __func__, hparams.vocab_only);
+    LLAMA_LOG_INFO("%s: no_alloc         = %d\n",     __func__, hparams.no_alloc);
 
     if (!hparams.vocab_only) {
         LLAMA_LOG_INFO("%s: n_ctx_train      = %u\n",     __func__, hparams.n_ctx_train);
@@ -6797,6 +6899,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
         LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
+        LLAMA_LOG_INFO("%s: rope_yarn_log_mul= %.4f\n",   __func__, hparams.rope_yarn_log_mul);
         LLAMA_LOG_INFO("%s: rope_finetuned   = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
         // MRoPE (Multi-axis Rotary Position Embedding) sections
         if (const auto & s = hparams.rope_sections; s[0] || s[1] || s[2] || s[3]) {
@@ -6819,7 +6922,8 @@ void llama_model::print_info() const {
         arch == LLM_ARCH_PLAMO2 ||
         arch == LLM_ARCH_GRANITE_HYBRID ||
         arch == LLM_ARCH_QWEN3NEXT ||
-        arch == LLM_ARCH_NEMOTRON_H) {
+        arch == LLM_ARCH_NEMOTRON_H ||
+        arch == LLM_ARCH_NEMOTRON_H_MOE) {
         LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
         LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
         LLAMA_LOG_INFO("%s: ssm_d_state      = %u\n",     __func__, hparams.ssm_d_state);
@@ -6860,7 +6964,6 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
         LLAMA_LOG_INFO("%s: expert_weights_norm  = %d\n",     __func__, hparams.expert_weights_norm);
         LLAMA_LOG_INFO("%s: expert_gating_func   = %s\n",     __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
-        LLAMA_LOG_INFO("%s: rope_yarn_log_mul    = %.4f\n",   __func__, hparams.rope_yarn_log_mul);
     }
 
     if (arch == LLM_ARCH_QWEN2MOE) {
@@ -6875,7 +6978,8 @@ void llama_model::print_info() const {
     if (arch == LLM_ARCH_MINICPM ||
         arch == LLM_ARCH_GRANITE ||
         arch == LLM_ARCH_GRANITE_MOE ||
-        arch == LLM_ARCH_GRANITE_HYBRID) {
+        arch == LLM_ARCH_GRANITE_HYBRID ||
+        arch == LLM_ARCH_NEMOTRON_H_MOE) {
         LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
         LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
         LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
@@ -7056,7 +7160,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                     if (arch == LLM_ARCH_FALCON_H1) {
                         filter_attn = [&](int32_t) { return true; };
                         filter_recr = [&](int32_t) { return true; };
-                    } else if (arch == LLM_ARCH_NEMOTRON_H) {
+                    } else if (arch == LLM_ARCH_NEMOTRON_H || arch == LLM_ARCH_NEMOTRON_H_MOE) {
                         filter_attn = [&](int32_t il) {
                             return !hparams.is_recurrent(il) && hparams.n_ff(il) == 0;
                         };
@@ -7304,7 +7408,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             } break;
         case LLM_ARCH_GEMMA3:
             {
-                llm = std::make_unique<llm_build_gemma3_iswa>(*this, params);
+                if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
+                    llm = std::make_unique<llm_build_gemma3<true>>(*this, params);
+                } else {
+                    llm = std::make_unique<llm_build_gemma3<false>>(*this, params);
+                }
             } break;
         case LLM_ARCH_GEMMA3N:
             {
@@ -7423,6 +7531,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
                 llm = std::make_unique<llm_build_nemotron>(*this, params);
             } break;
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             {
                 llm = std::make_unique<llm_build_nemotron_h>(*this, params);
             } break;
@@ -7569,6 +7678,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_qwen3next>(*this, params);
             } break;
+        case LLM_ARCH_MISTRAL3:
+            {
+                llm = std::make_unique<llm_build_mistral3>(*this, params);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -7607,6 +7720,7 @@ llama_model_params llama_model_default_params() {
         /*.check_tensors               =*/ false,
         /*.use_extra_bufts             =*/ true,
         /*.no_host                     =*/ false,
+        /*.no_alloc                    =*/ false,
     };
 
     return result;
@@ -7706,6 +7820,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ARWKV7:
         case LLM_ARCH_WAVTOKENIZER_DEC:
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             return LLAMA_ROPE_TYPE_NONE;
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values
@@ -7726,7 +7841,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_PLM:
         case LLM_ARCH_CHATGLM:
-        case LLM_ARCH_GLM4:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
         case LLM_ARCH_GRANITE_HYBRID:
@@ -7738,6 +7852,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ARCEE:
         case LLM_ARCH_ERNIE4_5:
         case LLM_ARCH_ERNIE4_5_MOE:
+        case LLM_ARCH_MISTRAL3:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2
@@ -7788,7 +7903,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_SMALLTHINKER:
-        case LLM_ARCH_GLM4_MOE:
         case LLM_ARCH_SEED_OSS:
         case LLM_ARCH_GROVEMOE:
         case LLM_ARCH_APERTUS:
@@ -7805,6 +7919,11 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_QWEN3VLMOE:
             return LLAMA_ROPE_TYPE_IMROPE;
 
+        case LLM_ARCH_GLM4:
+            return model->hparams.use_mrope() ? LLAMA_ROPE_TYPE_MROPE : LLAMA_ROPE_TYPE_NORM;
+        case LLM_ARCH_GLM4_MOE:
+            return model->hparams.use_mrope() ? LLAMA_ROPE_TYPE_MROPE : LLAMA_ROPE_TYPE_NEOX;
+
         // all model arches should be listed explicitly here
         case LLM_ARCH_UNKNOWN:
             GGML_ABORT("unknown architecture");

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

@@ -114,6 +114,7 @@ enum llm_type {
     LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
+    LLM_TYPE_31B_A3_5B,
     LLM_TYPE_80B_A3B, // Qwen3 Next
     LLM_TYPE_100B_A6B,
     LLM_TYPE_106B_A12B, // GLM-4.5-Air

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

@@ -596,7 +596,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     }
 
     std::vector<std::string> splits = {};
-    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides, nullptr);
+    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr);
     ml.init_mappings(false); // no prefetching
 
     llama_model model(llama_model_default_params());
@@ -666,7 +666,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
     std::map<int, std::string> mapped;
     int blk_id = 0;
-    int pruned_attention_w = 0;
 
     // make a list of weights
     std::vector<const llama_model_loader::llama_tensor_weight *> tensors;
@@ -674,11 +673,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     for (const auto & it : ml.weights_map) {
         const std::string remapped_name(remap_layer(it.first, prune_list, mapped, blk_id));
         if (remapped_name.empty()) {
-            if (it.first.find("attn_v.weight") != std::string::npos ||
-                it.first.find("attn_qkv.weight") != std::string::npos ||
-                it.first.find("attn_kv_b.weight") != std::string::npos) {
-                    pruned_attention_w++;
-            }
             LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
             continue;
         }
@@ -703,7 +697,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         });
     }
 
-    bool is_clip_model = false;
     for (const auto * it : tensors) {
         const struct ggml_tensor * tensor = it->tensor;
 
@@ -717,32 +710,10 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         } else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
             qs.has_output = true;
         }
-
-        is_clip_model |= name.rfind("mm.", 0) == 0; // check the "mm." prefix
     }
 
     qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
 
-    // sanity checks for models that have attention layers
-    if (qs.n_attention_wv != 0 && !is_clip_model)
-    {
-        const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
-        // attention layers have a non-zero number of kv heads
-        int32_t n_layer_attn = 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)) {
-            // now n_layer_attn is the number of attention layers in the encoder
-            // for each decoder block, there are 2 attention layers
-            n_layer_attn += 2 * model.hparams.dec_n_layer;
-        }
-
-        // note: for linear-attention models (such as Qwen3 Next) this is the number of linear layers
-        const int32_t n_layer_recr = std::count(model.hparams.recurrent_layer_arr.begin(), model.hparams.recurrent_layer_arr.end(), true);
-
-        LLAMA_LOG_INFO("%s: n_layer_attn = %d, n_layer_recr = %d, pruned_attention_w = %d\n", __func__, n_layer_attn, n_layer_recr, pruned_attention_w);
-
-        GGML_ASSERT((qs.n_attention_wv == n_layer_attn - pruned_attention_w - n_layer_recr) && "n_attention_wv is unexpected");
-    }
-
     size_t total_size_org = 0;
     size_t total_size_new = 0;
 

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

@@ -1884,7 +1884,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 clean_spaces = false;
             } else if (
                     tokenizer_pre == "qwen2" ||
-                    tokenizer_pre == "deepseek-r1-qwen") {
+                    tokenizer_pre == "deepseek-r1-qwen" ||
+                    tokenizer_pre == "kormo") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
                 clean_spaces = false;
             } else if (
@@ -3243,8 +3244,7 @@ void llama_vocab::impl::print_info() const {
 llama_vocab::llama_vocab() : pimpl(new impl(*this)) {
 }
 
-llama_vocab::~llama_vocab() {
+llama_vocab::~llama_vocab() = default;
-}
 
 void llama_vocab::load(llama_model_loader & ml, const LLM_KV & kv) {
     pimpl->load(ml, kv);

llama/llama.cpp/src/llama.cpp

@@ -1,6 +1,9 @@
+#include "llama.h"
+
 #include "llama-impl.h"
 
 #include "llama-chat.h"
+#include "llama-context.h"
 #include "llama-mmap.h"
 #include "llama-vocab.h"
 #include "llama-model-loader.h"
@@ -11,11 +14,14 @@
 #include "ggml-backend.h"
 
 #include <algorithm>
+#include <cassert>
+#include <cinttypes>
 #include <cstddef>
 #include <cstdint>
 #include <cstdio>
 #include <cstring>
 #include <ctime>
+#include <stdexcept>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -37,6 +43,646 @@ const char * llama_flash_attn_type_name(enum llama_flash_attn_type flash_attn_ty
     GGML_ABORT("fatal error");
 }
 
+struct llama_device_memory_data {
+    int64_t total;
+    int64_t free;
+    llama_memory_breakdown_data mb;
+};
+
+static std::vector<llama_device_memory_data> llama_get_device_memory_data(
+        const char * path_model, const llama_model_params * mparams, const llama_context_params * cparams,
+        std::vector<ggml_backend_dev_t> & devs, uint32_t & hp_ngl, uint32_t & hp_n_ctx_train, uint32_t & hp_n_expert,
+        const ggml_log_level log_level) {
+    struct user_data_t {
+        struct {
+            ggml_log_callback callback;
+            void * user_data;
+        } original_logger;
+        ggml_log_level min_level; // prints below this log level go to debug log
+    };
+    user_data_t ud;
+    llama_log_get(&ud.original_logger.callback, &ud.original_logger.user_data);
+    ud.min_level = log_level;
+
+    llama_log_set([](ggml_log_level level, const char * text, void * user_data) {
+        const user_data_t * ud = (const user_data_t *) user_data;
+        const ggml_log_level level_eff = level >= ud->min_level ? level : GGML_LOG_LEVEL_DEBUG;
+        ud->original_logger.callback(level_eff, text, ud->original_logger.user_data);
+    }, &ud);
+
+    llama_model_params mparams_copy = *mparams;
+    mparams_copy.no_alloc = true;
+    mparams_copy.use_mmap = false;
+
+    llama_model * model = llama_model_load_from_file(path_model, mparams_copy);
+    if (model == nullptr) {
+        llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+        throw std::runtime_error("failed to load model");
+    }
+
+    llama_context * ctx = llama_init_from_model(model, *cparams);
+    if (ctx == nullptr) {
+        llama_model_free(model);
+        llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+        throw std::runtime_error("failed to create llama_context from model");
+    }
+
+    std::vector<llama_device_memory_data> ret(model->devices.size());
+
+    std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> memory_breakdown = ctx->memory_breakdown();
+
+    for (const auto & [buft, mb] : memory_breakdown) {
+        if (ggml_backend_buft_is_host(buft)) {
+            continue;
+        }
+
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        if (!dev) {
+            continue;
+        }
+        for (size_t i = 0; i < ret.size(); i++) {
+            if (model->devices[i] == dev) {
+                ret[i].mb.model   += mb.model;
+                ret[i].mb.context += mb.context;
+                ret[i].mb.compute += mb.compute;
+                break;
+            }
+        }
+    }
+    for (size_t i = 0; i < ret.size(); i++) {
+        size_t free, total;
+        ggml_backend_dev_memory(model->devices[i], &free, &total);
+        ret[i].free  = free;
+        ret[i].total = total;
+    }
+
+    devs           = model->devices;
+    hp_ngl         = model->hparams.n_layer;
+    hp_n_ctx_train = model->hparams.n_ctx_train;
+    hp_n_expert    = model->hparams.n_expert;
+
+    llama_memory_breakdown_print(ctx); // goes to debug log
+
+    llama_free(ctx);
+    llama_model_free(model);
+    llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+    return ret;
+}
+
+// enum to identify part of a layer for distributing its tensors:
+enum layer_fraction_t {
+    LAYER_FRACTION_NONE = 0, // nothing
+    LAYER_FRACTION_ATTN = 1, // attention
+    LAYER_FRACTION_UP   = 2, // attention + up
+    LAYER_FRACTION_GATE = 3, // attention + up + gate
+    LAYER_FRACTION_MOE  = 4, // everything but sparse MoE weights
+};
+// this enum is only used in llama_params_fit_impl but needs to be defined outside of it to fix a Windows compilation issue
+
+static void llama_params_fit_impl(
+        const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
+        float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
+        size_t margin_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
+    constexpr int64_t MiB = 1024*1024;
+    const int64_t margin = margin_s; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
+    typedef std::vector<llama_device_memory_data> dmds_t;
+    const llama_model_params default_mparams = llama_model_default_params();
+
+    std::vector<ggml_backend_dev_t> devs;
+    uint32_t hp_ngl = 0; // hparams.n_gpu_layers
+    uint32_t hp_nct = 0; // hparams.n_ctx_train
+    uint32_t hp_nex = 0; // hparams.n_expert
+
+    // step 1: get data for default parameters and check whether any changes are necessary in the first place
+
+    LLAMA_LOG_DEBUG("%s: getting device memory data for initial parameters:\n", __func__);
+    const dmds_t dmds_full = llama_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+    const size_t nd = devs.size(); // number of devices
+    if (nd == 0) {
+        LLAMA_LOG_INFO("%s: no devices with dedicated memory found\n", __func__);
+        return;
+    }
+
+    std::vector<std::string> dev_names;
+    {
+        dev_names.reserve(nd);
+        size_t max_length = 0;
+        for (ggml_backend_dev_t dev : devs) {
+            std::string name = ggml_backend_dev_name(dev);
+            name += " (";
+            name += ggml_backend_dev_description(dev);
+            name += ")";
+            dev_names.push_back(name);
+            max_length = std::max(max_length, name.length());
+        }
+        for (std::string & dn : dev_names) {
+            dn.insert(dn.end(), max_length - dn.length(), ' ');
+        }
+    }
+
+    int64_t sum_total          = 0;
+    int64_t sum_projected_free = 0;
+    int64_t min_projected_free = INT64_MAX;
+    int64_t sum_projected_used = 0;
+    int64_t sum_projected_ctx  = 0;
+
+    if (nd > 1) {
+        LLAMA_LOG_INFO("%s: projected memory use with initial parameters [MiB]:\n", __func__);
+    }
+    for (size_t id = 0; id < nd; id++) {
+        const llama_device_memory_data & dmd = dmds_full[id];
+
+        const int64_t projected_used = dmd.mb.total();
+        const int64_t projected_free = dmd.free - projected_used;
+
+        sum_total          += dmd.total;
+        sum_projected_used += projected_used;
+        sum_projected_free += projected_free;
+        min_projected_free  = std::min(min_projected_free, projected_free);
+        sum_projected_ctx  += dmd.mb.context;
+
+        if (nd > 1) {
+            LLAMA_LOG_INFO("%s:   - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " %s\n",
+                __func__, dev_names[id].c_str(), dmd.total/MiB, projected_used/MiB, std::abs(projected_free)/MiB,
+                projected_free >= 0 ? "surplus" : "deficit");
+        }
+    }
+    assert(sum_total >= 0 && sum_projected_used >= 0 && sum_projected_ctx >= 0);
+    assert(sum_projected_used >= sum_projected_ctx);
+    LLAMA_LOG_INFO("%s: projected to use %" PRId64 " MiB of device memory vs. %" PRId64 " MiB of free device memory\n",
+        __func__, sum_projected_used/MiB, sum_total/MiB);
+    if (min_projected_free >= margin) {
+        if (nd == 1) {
+            LLAMA_LOG_INFO("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n",
+                __func__, min_projected_free/MiB, margin/MiB);
+            return;
+        }
+        LLAMA_LOG_INFO("%s: will leave at least %" PRId64 " >= %" PRId64 " MiB of free memory on all devices, no changes needed\n",
+            __func__, min_projected_free/MiB, margin/MiB);
+        return;
+    }
+
+    // step 2: try reducing memory use by reducing the context size
+
+    {
+        int64_t global_surplus = sum_projected_free - int64_t(nd)*margin;
+        if (global_surplus < 0) {
+            LLAMA_LOG_INFO(nd == 1 ?
+                "%s: cannot fulfill margin of %" PRId64 " MiB, need to reduce device memory by %" PRId64 " MiB\n" :
+                "%s: cannot fulfill margin of %" PRId64 " MiB on all devices, need to use %" PRId64 " MiB less in total\n",
+                __func__, margin/MiB, -global_surplus/MiB);
+            if (cparams->n_ctx == 0) {
+                if (hp_nct > n_ctx_min) {
+                    const int64_t bytes_per_ctx = sum_projected_ctx / hp_nct;
+                    const uint32_t ctx_reduction = std::min(
+                        uint32_t((-global_surplus + bytes_per_ctx - 1) / bytes_per_ctx), hp_nct - n_ctx_min);
+                    cparams->n_ctx = hp_nct - ctx_reduction;
+                    const int64_t memory_reduction = ctx_reduction * bytes_per_ctx;
+                    global_surplus += memory_reduction;
+                    LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
+                        __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
+                    if (global_surplus >= 0) {
+                        if (nd == 1) {
+                            LLAMA_LOG_INFO("%s: entire model can be fit by reducing context\n", __func__);
+                            return;
+                        }
+                        LLAMA_LOG_INFO("%s: entire model should be fit across devices by reducing context\n", __func__);
+                    }
+                } else {
+                    LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
+                        __func__, hp_nct, n_ctx_min);
+                }
+            } else {
+                LLAMA_LOG_INFO("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx);
+            }
+        }
+    }
+
+    if (mparams->n_gpu_layers != default_mparams.n_gpu_layers) {
+        throw std::runtime_error("n_gpu_layers already set by user to " + std::to_string(mparams->n_gpu_layers) + ", abort");
+    }
+    if (nd > 1) {
+        if (!tensor_split) {
+            throw std::runtime_error("did not provide a buffer to write the tensor_split to, abort");
+        }
+        if (mparams->tensor_split) {
+            for (size_t id = 0; id < nd; id++) {
+                if (mparams->tensor_split[id] != 0.0f) {
+                    throw std::runtime_error("model_params::tensor_split already set by user, abort");
+                }
+            }
+        }
+        if (mparams->split_mode == LLAMA_SPLIT_MODE_ROW) {
+            throw std::runtime_error("changing weight allocation for LLAMA_SPLIT_MODE_ROW not implemented, abort");
+        }
+        if (hp_ngl < 2*nd) {
+            throw std::runtime_error("model has only " + std::to_string(hp_ngl) + " layers but need at least "
+                + std::to_string(2*nd) + " to fit memory for " + std::to_string(nd) + " devices, abort");
+        }
+    }
+    if (!tensor_buft_overrides) {
+        throw std::runtime_error("did not provide buffer to set tensor_buft_overrides, abort");
+    }
+    if (mparams->tensor_buft_overrides && (mparams->tensor_buft_overrides->pattern || mparams->tensor_buft_overrides->buft)) {
+        throw std::runtime_error("model_params::tensor_buft_overrides already set by user, abort");
+    }
+
+    // step 3: iteratively fill the back to front with "dense" layers
+    //   - for a dense model simply fill full layers, giving each device a contiguous slice of the model
+    //   - for a MoE model, same as dense model but with all MoE tensors in system memory
+
+    // utility function that returns a static C string matching the tensors for a specific layer index and layer fraction:
+    auto get_overflow_pattern = [&](const size_t il, const layer_fraction_t lf) -> const char * {
+        constexpr size_t n_strings = 1000;
+        if (il >= n_strings) {
+            throw std::runtime_error("at most " + std::to_string(n_strings) + " model layers are supported");
+        }
+        switch (lf) {
+            case LAYER_FRACTION_ATTN: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(up|gate|down).*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_UP: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(gate|down).*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_GATE: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_down.*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_MOE: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(up|down|gate)_(ch|)exps";
+                }
+                return patterns[il].c_str();
+            }
+            default:
+                GGML_ABORT("fatal error");
+        }
+    };
+
+    struct ngl_t {
+        uint32_t n_layer = 0; // number of total layers
+        uint32_t n_part  = 0; // number of partial layers, <= n_layer
+
+        // for the first partial layer varying parts can overflow, all further layers use LAYER_FRACTION_MOE:
+        layer_fraction_t overflow_type = LAYER_FRACTION_MOE;
+    };
+
+    const size_t ntbo = llama_max_tensor_buft_overrides();
+
+    // utility function to set n_gpu_layers and tensor_split
+    auto set_ngl_tensor_split_tbo = [&](
+            const std::vector<ngl_t> & ngl_per_device,
+            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts,
+            llama_model_params & mparams,
+            const bool add_nonrepeating) {
+        mparams.n_gpu_layers = 0;
+        for (size_t id = 0; id < nd; id++) {
+            mparams.n_gpu_layers += ngl_per_device[id].n_layer;
+            if (nd > 1) {
+                tensor_split[id] = ngl_per_device[id].n_layer;
+            }
+        }
+        assert(uint32_t(mparams.n_gpu_layers) <= hp_ngl);
+        uint32_t il0 = hp_ngl - mparams.n_gpu_layers; // start index for tensor buft overrides
+
+        if (add_nonrepeating) {
+            mparams.n_gpu_layers += 1;
+            tensor_split[nd - 1] += 1;
+        }
+        mparams.tensor_split = tensor_split;
+
+        size_t itbo = 0;
+        for (size_t id = 0; id < nd; id++) {
+            il0 += ngl_per_device[id].n_layer - ngl_per_device[id].n_part;
+            for (uint32_t il = il0; il < il0 + ngl_per_device[id].n_part; il++) {
+                if (itbo + 1 >= ntbo) {
+                    tensor_buft_overrides[itbo].pattern = nullptr;
+                    tensor_buft_overrides[itbo].buft    = nullptr;
+                    itbo++;
+                    mparams.tensor_buft_overrides = tensor_buft_overrides;
+                    throw std::runtime_error("llama_params_fit_n_tensor_buft_overrides() == "
+                        + std::to_string(ntbo) + " is insufficient for model\n");
+                }
+                tensor_buft_overrides[itbo].pattern = get_overflow_pattern(il, il == il0 ? ngl_per_device[id].overflow_type : LAYER_FRACTION_MOE);
+                tensor_buft_overrides[itbo].buft = overflow_bufts[id];
+                itbo++;
+            }
+            il0 += ngl_per_device[id].n_part;
+        }
+        tensor_buft_overrides[itbo].pattern = nullptr;
+        tensor_buft_overrides[itbo].buft    = nullptr;
+        itbo++;
+        mparams.tensor_buft_overrides = tensor_buft_overrides;
+    };
+
+    // utility function that returns the memory use per device for given numbers of layers per device
+    auto get_memory_for_layers = [&](
+            const char * func_name,
+            const std::vector<ngl_t> & ngl_per_device,
+            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts,
+            const bool add_nonrepeating) -> std::vector<int64_t> {
+        llama_model_params mparams_copy = *mparams;
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy, add_nonrepeating);
+
+        const dmds_t dmd_nl = llama_get_device_memory_data(
+            path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+
+        LLAMA_LOG_DEBUG("%s: memory for test allocation by device:\n", func_name);
+        for (size_t id = 0; id < nd; id++) {
+            const ngl_t & n = ngl_per_device[id];
+            LLAMA_LOG_DEBUG(
+                "%s: id=%zu, n_layer=%2" PRIu32 ", n_part=%2" PRIu32 ", overflow_type=%d, mem=%6" PRId64 " MiB\n",
+                func_name, id, n.n_layer, n.n_part, int(n.overflow_type), dmd_nl[id].mb.total()/MiB);
+        }
+
+        std::vector<int64_t> ret;
+        ret.reserve(nd);
+        for (const llama_device_memory_data & dmd : dmd_nl) {
+            ret.push_back(dmd.mb.total());
+        }
+        return ret;
+    };
+
+    int64_t global_surplus_cpu_moe = 0;
+    if (hp_nex > 0) {
+        const static std::string pattern_moe_all = "blk\\.\\d+\\.ffn_(up|down|gate)_(ch|)exps"; // matches all MoE tensors
+        ggml_backend_buffer_type_t cpu_buft = ggml_backend_cpu_buffer_type();
+        tensor_buft_overrides[0] = {pattern_moe_all.c_str(), cpu_buft};
+        tensor_buft_overrides[1] = {nullptr, nullptr};
+        mparams->tensor_buft_overrides = tensor_buft_overrides;
+
+        LLAMA_LOG_DEBUG("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
+        const dmds_t dmds_cpu_moe = llama_get_device_memory_data(
+            path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+
+        for (const llama_device_memory_data & dmd : dmds_cpu_moe) {
+            global_surplus_cpu_moe += dmd.free;
+            global_surplus_cpu_moe -= int64_t(dmd.mb.total()) + margin;
+        }
+
+        if (global_surplus_cpu_moe > 0) {
+            LLAMA_LOG_INFO("%s: with only dense weights in device memory there is a total surplus of %" PRId64 " MiB\n",
+                __func__, global_surplus_cpu_moe/MiB);
+        } else {
+            LLAMA_LOG_INFO("%s: with only dense weights in device memory there is still a total deficit of %" PRId64 " MiB\n",
+                __func__, -global_surplus_cpu_moe/MiB);
+        }
+
+        // reset
+        tensor_buft_overrides[0] = {nullptr, nullptr};
+        mparams->tensor_buft_overrides = tensor_buft_overrides;
+    }
+
+    std::vector<int64_t> targets; // maximum acceptable memory use per device
+    targets.reserve(nd);
+    for (size_t id = 0; id < nd; id++) {
+        targets.push_back(dmds_full[id].free - margin);
+        LLAMA_LOG_DEBUG("%s: id=%zu, target=%" PRId64 " MiB\n", __func__, id, targets[id]/MiB);
+    }
+
+    // whether for the optimal memory use we expect to load at least some MoE tensors:
+    const bool partial_moe = hp_nex > 0 && global_surplus_cpu_moe > 0;
+
+    std::vector<ggml_backend_buffer_type_t> overflow_bufts; // which bufts the partial layers of a device overflow to:
+    overflow_bufts.reserve(nd);
+    for (size_t id = 0; id < nd - 1; ++id) {
+        overflow_bufts.push_back(ggml_backend_dev_buffer_type(devs[id + 1]));
+    }
+    overflow_bufts.push_back(ggml_backend_cpu_buffer_type());
+
+    std::vector<ngl_t> ngl_per_device(nd);
+    std::vector<int64_t> mem = get_memory_for_layers(__func__, ngl_per_device, overflow_bufts, partial_moe);
+    if (hp_nex > 0) {
+        for (size_t id = 0; id < nd; id++) {
+            ngl_per_device[id].overflow_type = LAYER_FRACTION_MOE;
+        }
+    }
+
+    // optimize the number of layers per device using the method of false position:
+    //   - ngl_per_device has 0 layers for each device, lower bound
+    //   - try a "high" configuration where a device is given all unassigned layers
+    //   - interpolate the memory use / layer between low and high linearly to get a guess where it meets our target
+    //   - check memory use of our guess, replace either the low or high bound
+    //   - once we only have a difference of a single layer, stop and return the lower bound that just barely still fits
+    if (hp_nex == 0) {
+        LLAMA_LOG_INFO("%s: filling dense layers back-to-front:\n", __func__);
+    } else {
+        LLAMA_LOG_INFO("%s: filling dense-only layers back-to-front:\n", __func__);
+    }
+    uint32_t n_unassigned = hp_ngl;
+    for (int id = nd - 1; id >= 0; id--) {
+        std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
+        ngl_per_device_high[id].n_layer = n_unassigned;
+        if (hp_nex > 0) {
+            ngl_per_device_high[id].n_part = ngl_per_device_high[id].n_layer;
+        }
+        if (ngl_per_device_high[id].n_layer > 0) {
+            std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts, partial_moe);
+            if (mem_high[id] > targets[id]) {
+                uint32_t delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                while (delta > 1) {
+                    uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
+                    step_size = std::max(step_size, uint32_t(1));
+                    step_size = std::min(step_size, delta - 1);
+
+                    std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+                    ngl_per_device_test[id].n_layer += step_size;
+                    if (hp_nex) {
+                        ngl_per_device_test[id].n_part += step_size;
+                    }
+                    const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+
+                    if (mem_test[id] <= targets[id]) {
+                        ngl_per_device  = ngl_per_device_test;
+                        mem             = mem_test;
+                        n_unassigned   -= ngl_per_device[id].n_layer;
+                        LLAMA_LOG_DEBUG("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+                    } else {
+                        ngl_per_device_high = ngl_per_device_test;
+                        mem_high            = mem_test;
+                        LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+                    }
+                    delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                }
+            } else {
+                ngl_per_device  = ngl_per_device_high;
+                n_unassigned   -= ngl_per_device[id].n_layer;
+                LLAMA_LOG_DEBUG("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+            }
+        }
+
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LLAMA_LOG_INFO(
+            "%s:   - %s: %2" PRIu32 " layers, %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, mem[id]/MiB, projected_margin/MiB);
+    }
+    if (hp_nex == 0 || global_surplus_cpu_moe <= 0) {
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams, partial_moe);
+        return;
+    }
+
+    // step 4: for a MoE model where all dense tensors fit,
+    //     convert the dense-only layers in the back to full layers in the front until all devices are full
+    // essentially the same procedure as for the dense-only layers except front-to-back
+    // also, try fitting at least part of one more layer to reduce waste for "small" GPUs with e.g. 24 GiB VRAM
+
+    size_t id_dense_start = nd;
+    for (int id = nd - 1; id >= 0; id--) {
+        if (ngl_per_device[id].n_layer > 0) {
+            id_dense_start = id;
+            continue;
+        }
+        break;
+    }
+    assert(id_dense_start < nd);
+
+    LLAMA_LOG_INFO("%s: converting dense-only layers to full layers and filling them front-to-back with overflow to next device/system memory:\n", __func__);
+    for (size_t id = 0; id <= id_dense_start; id++) {
+        std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
+        for (size_t jd = id_dense_start; jd < nd; jd++) {
+            const uint32_t n_layer_move = ngl_per_device_high[jd].n_layer;
+            ngl_per_device_high[id].n_layer += n_layer_move;
+            ngl_per_device_high[jd].n_layer -= n_layer_move;
+            ngl_per_device_high[jd].n_part = 0;
+        }
+        size_t id_dense_start_high = nd - 1;
+        std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts, partial_moe);
+
+        if (mem_high[id] > targets[id]) {
+            assert(ngl_per_device_high[id].n_layer >= ngl_per_device_high[id].n_part);
+            assert(ngl_per_device[id].n_layer >= ngl_per_device[id].n_part);
+            assert((ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
+                   >= ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+            uint32_t delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
+                - (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+            while (delta > 1) {
+                uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
+                step_size = std::max(step_size, uint32_t(1));
+                step_size = std::min(step_size, delta - 1);
+
+                std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+                size_t id_dense_start_test = id_dense_start;
+                uint32_t n_converted_test = 0;
+                for (;id_dense_start_test < nd; id_dense_start_test++) {
+                    const uint32_t n_convert_jd = std::min(step_size - n_converted_test, ngl_per_device_test[id_dense_start_test].n_part);
+                    ngl_per_device_test[id_dense_start_test].n_layer -= n_convert_jd;
+                    ngl_per_device_test[id_dense_start_test].n_part -= n_convert_jd;
+                    ngl_per_device_test[id].n_layer += n_convert_jd;
+                    n_converted_test += n_convert_jd;
+
+                    if (ngl_per_device_test[id_dense_start_test].n_layer > 0) {
+                        break;
+                    }
+                }
+                const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+
+                if (mem_test[id] <= targets[id]) {
+                    ngl_per_device = ngl_per_device_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                } else {
+                    ngl_per_device_high = ngl_per_device_test;
+                    mem_high            = mem_test;
+                    id_dense_start_high = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start_high=%zu\n",
+                        __func__, id, ngl_per_device_high[id].n_layer, ngl_per_device_high[id].n_part, id_dense_start_high);
+                }
+                delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
+                    - (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+            }
+        } else {
+            ngl_per_device = ngl_per_device_high;
+            id_dense_start = id_dense_start_high;
+            LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start=%zu\n",
+                __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+        }
+
+        // try to fit at least part of one more layer
+        if (ngl_per_device[id_dense_start].n_layer > 0) {
+            std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+            size_t id_dense_start_test = id_dense_start;
+            ngl_per_device_test[id_dense_start_test].n_layer--;
+            ngl_per_device_test[id_dense_start_test].n_part--;
+            ngl_per_device_test[id].n_layer++;
+            ngl_per_device_test[id].n_part++;
+            if (ngl_per_device_test[id_dense_start_test].n_layer == 0) {
+                id_dense_start_test++;
+            }
+            ngl_per_device_test[id].overflow_type = LAYER_FRACTION_UP;
+            LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_UP\n", __func__);
+            std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+            if (mem_test[id] < targets[id]) {
+                ngl_per_device = ngl_per_device_test;
+                mem            = mem_test;
+                id_dense_start = id_dense_start_test;
+                LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", UP), id_dense_start=%zu\n",
+                    __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+
+                ngl_per_device_test[id].overflow_type = LAYER_FRACTION_GATE;
+                LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_GATE\n", __func__);
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+                if (mem_test[id] < targets[id]) {
+                    ngl_per_device = ngl_per_device_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", GATE), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                }
+            } else {
+                ngl_per_device_test[id].overflow_type = LAYER_FRACTION_ATTN;
+                LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_ATTN\n", __func__);
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+                if (mem_test[id] < targets[id]) {
+                    ngl_per_device = ngl_per_device_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", ATTN), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                }
+            }
+        }
+
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LLAMA_LOG_INFO(
+            "%s:   - %s: %2" PRIu32 " layers (%2" PRIu32 " overflowing), %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
+    }
+
+    set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams, partial_moe);
+}
+
+bool llama_params_fit(
+        const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
+        float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
+        size_t margin_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
+    const int64_t t0_us = llama_time_us();
+    bool ok = true;
+    try {
+        llama_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margin_s, n_ctx_min, log_level);
+        LLAMA_LOG_INFO("%s: successfully fit params to free device memory\n", __func__);
+    } catch (const std::runtime_error & e) {
+        LLAMA_LOG_WARN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());
+        ok = false;
+    }
+    const int64_t t1_us = llama_time_us();
+    LLAMA_LOG_INFO("%s: fitting params to free memory took %.2f seconds\n", __func__, (t1_us - t0_us) * 1e-6);
+    return ok;
+}
+
 struct llama_sampler_chain_params llama_sampler_chain_default_params() {
     struct llama_sampler_chain_params result = {
         /*.no_perf                     =*/ true,
@@ -49,6 +695,10 @@ size_t llama_max_devices(void) {
     return 16;
 }
 
+size_t llama_max_tensor_buft_overrides() {
+    return 4096;
+}
+
 bool llama_supports_mmap(void) {
     return llama_mmap::SUPPORTED;
 }
@@ -108,11 +758,12 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
     model.t_start_us = tm.t_start_us;
 
     try {
-        llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides, params.tensor_buft_overrides);
+        llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
 
         ml.print_info();
 
         model.hparams.vocab_only = params.vocab_only;
+        model.hparams.no_alloc   = params.no_alloc;
 
         try {
             model.load_arch(ml);

llama/llama.cpp/src/models/deepseek2.cpp

@@ -1,7 +1,5 @@
 #include "models.h"
 
-
-
 llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
@@ -20,9 +18,15 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
 
     // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
     // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
-    const float mscale      = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+    // And also: https://github.com/ggml-org/llama.cpp/pull/17945 [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+
+    // first cancel the adjustment from llama_hparams::yarn_attn_factor_adjust to get the original attn_factor
+    GGML_ASSERT(ext_factor >= 0.0f);
+    const float attn_factor_org = attn_factor * (1.0f + 0.1f * logf(1.0f / freq_scale));
+
+    // use the original attn_factor to pre-scale the kq_scale
+    const float mscale   = attn_factor_org * (1.0f + 0.1f * hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
     const float kq_scale = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k));
-    const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
 
     ggml_tensor * cur;
     ggml_tensor * inpL;
@@ -30,6 +34,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
     // {n_embd, n_tokens}
     inpL = build_inp_embd(model.tok_embd);
 
+    // (optional) temperature tuning - used by mistral-large
+    ggml_tensor * inp_attn_scale = nullptr;
+    if (hparams.f_attn_temp_scale != 0.0f) {
+        inp_attn_scale = build_inp_attn_scale();
+    }
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -128,6 +138,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 ggml_tensor * Vcur = kv_cmpr;
                 cb(Vcur, "Vcur", il);
 
+                if (inp_attn_scale) {
+                    // apply llama 4 temperature scaling
+                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                    cb(Qcur, "Qcur_attn_temp_scaled", il);
+                }
+
                 // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
@@ -160,6 +176,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
                 cb(Kcur, "Kcur", il);
 
+                if (inp_attn_scale) {
+                    // apply llama 4 temperature scaling
+                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                    cb(Qcur, "Qcur_attn_temp_scaled", il);
+                }
+
                 // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
                 cur = build_attn(inp_attn,
                             model.layers[il].wo, NULL,

llama/llama.cpp/src/models/gemma3.cpp

@@ -1,6 +1,7 @@
 #include "models.h"
 
-llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+template <bool iswa>
+llm_build_gemma3<iswa>::llm_build_gemma3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_k;
 
     ggml_tensor * cur;
@@ -17,13 +18,28 @@ llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const ll
     ggml_tensor * inp_pos = build_inp_pos();
 
     // TODO: is causal == true correct? might need some changes
-    auto * inp_attn = build_attn_inp_kv_iswa();
+    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
 
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     for (int il = 0; il < n_layer; ++il) {
-        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+        float freq_base_l  = 0.0f;
-        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+        float freq_scale_l = 0.0f;
+
+        if constexpr (iswa) {
+            freq_base_l  = model.get_rope_freq_base (cparams, il);
+            freq_scale_l = model.get_rope_freq_scale(cparams, il);
+        } else {
+            freq_base_l  = freq_base;
+            freq_scale_l = freq_scale;
+        }
 
         // norm
         cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
@@ -102,7 +118,7 @@ llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const ll
         cur = build_norm(cur,
                 model.layers[il].ffn_post_norm, NULL,
                 LLM_NORM_RMS, -1);
-        cb(cur, "ffn_post_norm", -1);
+        cb(cur, "ffn_post_norm", il);
 
         cur = ggml_add(ctx0, cur, sa_out);
 
@@ -124,8 +140,17 @@ llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const ll
     // lm_head
     cur = build_lora_mm(model.output, cur);
 
+    if (hparams.f_final_logit_softcapping) {
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+    }
+
     cb(cur, "result_output", -1);
     res->t_logits = cur;
 
     ggml_build_forward_expand(gf, cur);
 }
+
+template struct llm_build_gemma3<false>;
+template struct llm_build_gemma3<true>;

llama/llama.cpp/src/models/glm4-moe.cpp

@@ -5,11 +5,20 @@ llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_grap
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 
+    int sections[4];
+    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
     ggml_tensor * cur;
     ggml_tensor * inpL;
 
     inpL = build_inp_embd(model.tok_embd);
 
+    bool use_mrope = hparams.use_mrope();
+    if (ubatch.embd && !use_mrope) {
+        // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results
+        GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");
+    }
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -60,17 +69,25 @@ llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_grap
                 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
                 cb(Kcur, "Kcur_normed", il);
             }
-            Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
 
-            Kcur = ggml_rope_ext(
+            if (use_mrope) {
-                    ctx0, Kcur, inp_pos, nullptr,
+                Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
+                            ext_factor, attn_factor, beta_fast, beta_slow);
-                    );
+
+                Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,
+                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow);
+            } else {
+                // Normal RoPE
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
+                                    rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
+                                    rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+            }
 
             cb(Qcur, "Qcur", il);
             cb(Kcur, "Kcur", il);

llama/llama.cpp/src/models/glm4.cpp

@@ -8,11 +8,20 @@ llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 
+    int sections[4];
+    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
     ggml_tensor * cur;
     ggml_tensor * inpL;
 
     inpL = build_inp_embd(model.tok_embd);
 
+    bool use_mrope = hparams.use_mrope();
+    if (ubatch.embd && !use_mrope) {
+        // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results
+        GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");
+    }
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -63,11 +72,25 @@ llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params
                 Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
                                     cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
             }
-            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+
+            if (use_mrope) {
+                Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,
+                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
                             ext_factor, attn_factor, beta_fast, beta_slow);
 
-            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,
+                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
                             ext_factor, attn_factor, beta_fast, beta_slow);
+            } else {
+                // Normal RoPE
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
+                                    rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
+                                    rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+            }
 
             cb(Qcur, "Qcur", il);
             cb(Kcur, "Kcur", il);

llama/llama.cpp/src/models/mistral3.cpp

@@ -0,0 +1,160 @@
+#include "models.h"
+
+llm_build_mistral3::llm_build_mistral3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // (optional) temperature tuning
+    ggml_tensor * inp_attn_scale = nullptr;
+    if (hparams.f_attn_temp_scale != 0.0f) {
+        inp_attn_scale = build_inp_attn_scale();
+    }
+
+    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;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            if (inp_attn_scale) {
+                // apply llama 4 temperature scaling
+                Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                cb(Qcur, "Qcur_attn_temp_scaled", il);
+            }
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network (non-MoE)
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+            cb(cur, "ffn_moe_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/models.h

@@ -179,8 +179,9 @@ struct llm_build_gemma2_iswa : public llm_graph_context {
     llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params);
 };
 
-struct llm_build_gemma3_iswa : public llm_graph_context {
+template <bool iswa>
-    llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params);
+struct llm_build_gemma3 : public llm_graph_context {
+    llm_build_gemma3(const llama_model & model, const llm_graph_params & params);
 };
 
 struct llm_build_gemma3n_iswa : public llm_graph_context {
@@ -322,6 +323,10 @@ struct llm_build_minimax_m2 : public llm_graph_context {
     llm_build_minimax_m2(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_mistral3 : public llm_graph_context {
+    llm_build_mistral3(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_mpt : public llm_graph_context {
     llm_build_mpt(const llama_model & model, const llm_graph_params & params);
 };
@@ -436,23 +441,13 @@ private:
                 ggml_tensor * cur,
                 ggml_tensor * causal_mask,
                 ggml_tensor * identity,
+                ggml_tensor * diag_mask,
                         int   il);
 
     ggml_tensor * build_layer_ffn(
                 ggml_tensor * cur,
                         int   il);
 
-    ggml_tensor * build_delta_net_recurrent(
-                ggml_tensor * q,
-                ggml_tensor * k,
-                ggml_tensor * v,
-                ggml_tensor * g,
-                ggml_tensor * beta,
-                ggml_tensor * state,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                        int   il);
-
     ggml_tensor * build_delta_net_chunking(
                 ggml_tensor * q,
                 ggml_tensor * k,
@@ -462,6 +457,16 @@ private:
                 ggml_tensor * state,
                 ggml_tensor * causal_mask,
                 ggml_tensor * identity,
+                ggml_tensor * diag_mask,
+                        int   il);
+
+    ggml_tensor * build_delta_net_autoregressive(
+                ggml_tensor * q,
+                ggml_tensor * k,
+                ggml_tensor * v,
+                ggml_tensor * g,
+                ggml_tensor * beta,
+                ggml_tensor * state,
                 int           il);
 
     ggml_tensor * build_norm_gated(

llama/llama.cpp/src/models/nemotron-h.cpp

@@ -107,12 +107,41 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
 }
 
 ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il) {
+    if (model.layers[il].ffn_gate_inp == nullptr) {
         cur = build_ffn(cur,
                 model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
                 NULL,                      NULL,                        NULL,
                 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-            NULL, LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
         cb(cur, "ffn_out", il);
+    } else {
+        ggml_tensor * ffn_inp = cur;
+        ggml_tensor * moe_out =
+            build_moe_ffn(ffn_inp,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    nullptr, // no gate
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_RELU_SQR, hparams.expert_weights_norm,
+                    true, hparams.expert_weights_scale,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        ggml_tensor * ffn_shexp = build_ffn(ffn_inp,
+                    model.layers[il].ffn_up_shexp,  NULL, NULL,
+                    NULL /* no gate */           ,  NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
+        cb(ffn_shexp, "ffn_shexp", il);
+
+        cur = ggml_add(ctx0, moe_out, ffn_shexp);
+        cb(cur, "ffn_out", il);
+    }
 
     cur = build_cvec(cur, il);
     cb(cur, "l_out", il);

llama/llama.cpp/src/models/qwen2.cpp

@@ -31,16 +31,25 @@ llm_build_qwen2::llm_build_qwen2(const llama_model & model, const llm_graph_para
         {
             // compute Q and K and RoPE them
             ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
                 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
                 cb(Qcur, "Qcur", il);
+            }
 
             ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
                 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
                 cb(Kcur, "Kcur", il);
+            }
 
             ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
                 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
                 cb(Vcur, "Vcur", il);
+            }
 
             Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
             Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);

llama/llama.cpp/src/models/qwen3next.cpp

@@ -17,13 +17,15 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     ggml_tensor * causal_mask =
-        ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, ubatch.n_seq_tokens, ubatch.n_seq_tokens), 1.0f),
+        ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
                     GGML_TRI_TYPE_LOWER);
 
-    ggml_tensor * identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, ubatch.n_seq_tokens), 1.0f));
+    ggml_tensor * identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
+    ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
 
     ggml_build_forward_expand(gf, causal_mask);
     ggml_build_forward_expand(gf, identity);
+    ggml_build_forward_expand(gf, diag_mask);
 
     for (int il = 0; il < n_layer; ++il) {
         ggml_tensor * inpSA = inpL;
@@ -34,7 +36,7 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
-            cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, il);
+            cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
         } else {
             // Full attention layer
             cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);
@@ -93,14 +95,8 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
         ggml_tensor * state,
         ggml_tensor * causal_mask,
         ggml_tensor * identity,
+        ggml_tensor * diag_mask,
         int           il) {
-    GGML_ASSERT(ggml_is_contiguous(q));
-    GGML_ASSERT(ggml_is_contiguous(k));
-    GGML_ASSERT(ggml_is_contiguous(v));
-    GGML_ASSERT(ggml_is_contiguous(g));
-    GGML_ASSERT(ggml_is_contiguous(beta));
-    GGML_ASSERT(ggml_is_contiguous(state));
-
     const int64_t S_k      = q->ne[0];
     const int64_t H_k      = q->ne[1];
     const int64_t n_tokens = q->ne[2];
@@ -120,15 +116,10 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
 
     GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
 
-    // TODO: can this ever be false?
-    const bool use_qk_l2norm = true;
-
-    if (use_qk_l2norm) {
     const float eps_norm = hparams.f_norm_rms_eps;
 
     q = ggml_l2_norm(ctx0, q, eps_norm);
     k = ggml_l2_norm(ctx0, k, eps_norm);
-    }
 
     const float scale = 1.0f / sqrtf(S_v);
 
@@ -136,8 +127,6 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
 
     beta = ggml_sigmoid(ctx0, beta);
 
-    ggml_tensor * causal_diag_mask = ggml_add(ctx0, causal_mask, identity);
-
     cb(q, "q_in", il);
     cb(k, "k_in", il);
     cb(v, "v_in", il);
@@ -188,36 +177,21 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
     cb(v_beta, "v_beta", il);
     cb(k_beta, "k_beta", il);
 
-    ggml_tensor * chunked_mask =
+    q      = ggml_reshape_4d(ctx0, q,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-        ggml_view_4d(ctx0, causal_mask, chunk_size,
+    k      = ggml_reshape_4d(ctx0, k,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-                chunk_size,         causal_mask->ne[2], causal_mask->ne[3],
+    k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
-                causal_mask->nb[1], causal_mask->nb[2], causal_mask->nb[3], 0);
+    v      = ggml_reshape_4d(ctx0, v,      S_v, chunk_size, n_chunks, H_v * n_seqs);
+    v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
 
-    ggml_tensor * chunked_diag_mask =
+    g    = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
-        ggml_view_4d(ctx0, causal_diag_mask, chunk_size,
+    beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
-                chunk_size,              causal_diag_mask->ne[2], causal_diag_mask->ne[3],
-                causal_diag_mask->nb[1], causal_diag_mask->nb[2], causal_diag_mask->nb[3], 0);
-
-    ggml_tensor * chunked_identity =
-        ggml_view_4d(ctx0, identity, chunk_size,
-            chunk_size,      identity->ne[2], identity->ne[3],
-            identity->nb[1], identity->nb[2], identity->nb[3], 0);
-
-    q      = ggml_cont_4d(ctx0, q,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k      = ggml_cont_4d(ctx0, k,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k_beta = ggml_cont_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
-    v      = ggml_cont_4d(ctx0, v,      S_v, chunk_size, n_chunks, H_v * n_seqs);
-    v_beta = ggml_cont_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
-
-    g    = ggml_cont_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
-    beta = ggml_cont_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
 
     ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
 
     cb(g_cumsum, "g_cumsum", il);
 
-    ggml_tensor * gcs_i = ggml_cont_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
+    ggml_tensor * gcs_i = ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
-    ggml_tensor * gcs_j = ggml_cont_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
+    ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
 
     ggml_tensor * gcs_j_broadcast =
         ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
@@ -226,23 +200,23 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
 
     cb(decay_mask, "decay_mask", il);
 
-    decay_mask = ggml_mul(ctx0, decay_mask, chunked_diag_mask);
+    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
     decay_mask = ggml_exp(ctx0, decay_mask);
-    decay_mask = ggml_mul(ctx0, decay_mask, chunked_diag_mask);
+    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
 
     ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
 
     ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
-    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, chunked_mask));
+    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
 
     cb(attn, "attn_pre_solve", il);
 
-    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, chunked_mask);
+    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
-    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, chunked_identity, attn_lower), attn_lower);
+    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
 
     ggml_tensor * lin_solve  = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
-    attn                     = ggml_mul(ctx0, lin_solve, chunked_mask);
+    attn                     = ggml_mul(ctx0, lin_solve, causal_mask);
-    attn                     = ggml_add(ctx0, attn, chunked_identity);
+    attn                     = ggml_add(ctx0, attn, identity);
 
     cb(attn, "attn_solved", il);
 
@@ -291,7 +265,7 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
         // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
         attn = ggml_mul_mat(ctx0, k_chunk, q_chunk);
         attn = ggml_mul(ctx0, attn, decay_mask_chunk);
-        attn = ggml_mul(ctx0, attn, ggml_add(ctx0, chunked_identity, chunked_mask));
+        attn = ggml_mul(ctx0, attn, diag_mask);
 
         ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
 
@@ -361,23 +335,14 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
     return ggml_concat(ctx0, flat_output, flat_state, 0);
 }
 
-ggml_tensor * llm_build_qwen3next::build_delta_net_recurrent(
+ggml_tensor * llm_build_qwen3next::build_delta_net_autoregressive(
         ggml_tensor * q,
         ggml_tensor * k,
         ggml_tensor * v,
         ggml_tensor * g,
         ggml_tensor * beta,
         ggml_tensor * state,
-        ggml_tensor * causal_mask,
-        ggml_tensor * identity,
         int           il) {
-    GGML_ASSERT(ggml_is_contiguous(q));
-    GGML_ASSERT(ggml_is_contiguous(k));
-    GGML_ASSERT(ggml_is_contiguous(v));
-    GGML_ASSERT(ggml_is_contiguous(g));
-    GGML_ASSERT(ggml_is_contiguous(beta));
-    GGML_ASSERT(ggml_is_contiguous(state));
-
     const int64_t S_k      = q->ne[0];
     const int64_t H_k      = q->ne[1];
     const int64_t n_tokens = q->ne[2];
@@ -386,6 +351,7 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_recurrent(
     const int64_t S_v = v->ne[0];
     const int64_t H_v = v->ne[1];
 
+    GGML_ASSERT(n_tokens == 1);  // This function is optimized for single token processing
     GGML_ASSERT(v->ne[2] == n_tokens);
     GGML_ASSERT(k->ne[2] == n_tokens);
     GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
@@ -397,215 +363,65 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_recurrent(
 
     GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
 
-    // TODO: can this ever be false?
-    const bool use_qk_l2norm = true;
-
-    if (use_qk_l2norm) {
     const float eps_norm = hparams.f_norm_rms_eps;
 
     q = ggml_l2_norm(ctx0, q, eps_norm);
     k = ggml_l2_norm(ctx0, k, eps_norm);
-    }
 
     const float scale = 1.0f / sqrtf(S_v);
 
     q    = ggml_scale(ctx0, q, scale);
-
     beta = ggml_sigmoid(ctx0, beta);
 
-    ggml_tensor * causal_diag_mask = ggml_add(ctx0, causal_mask, identity);
-
     cb(q, "q_in", il);
     cb(k, "k_in", il);
     cb(v, "v_in", il);
     cb(beta, "beta_in", il);
     cb(g, "g_in", il);
 
-    q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
-
-    beta  = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
     state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
 
-    cb(q, "q_perm", il);
+    ggml_tensor * g_t    = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
-    cb(k, "k_perm", il);
+    ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
-    cb(v, "v_perm", il);
-    cb(beta, "beta_perm", il);
-    cb(g, "g_perm", il);
-    cb(state, "state_in", il);
 
-    GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
+    // Apply exponential to g_t
-    GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
+    g_t = ggml_exp(ctx0, g_t);
-    GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
-    GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
 
-    ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
+    // Apply the gated delta rule for the single timestep
-    ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
+    // last_recurrent_state = last_recurrent_state * g_t
+    state = ggml_mul(ctx0, state, g_t);
 
-    ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
+    // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
+    ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
+    ggml_tensor * kv_mem         = ggml_mul(ctx0, state, k_t_unsqueezed);
+    // we need to sum over dim=-2, so we transpose, sum, then transpose again
+    kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
 
-    cb(k_beta, "k_beta", il);
+    // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
-    cb(v_beta, "v_beta", il);
+    ggml_tensor * v_t    = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
-    cb(g_cumsum, "g_cumsum", il);
+    // delta = (v_t - kv_mem) * beta_t
+    ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem);  // both should be [S_v, 1, H_v, n_seqs]
+    ggml_tensor * delta  = ggml_mul(ctx0, v_diff, beta_t);
 
-    ggml_tensor * gcs_i = ggml_cont_4d(ctx0, g_cumsum, n_tokens, 1, H_v, n_seqs);  // [chunk_size, 1, n_tokens, n_seqs]
+    // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
-    ggml_tensor * gcs_j = ggml_cont_4d(ctx0, g_cumsum, 1, n_tokens, H_v, n_seqs);  // [1, chunk_size, n_tokens, n_seqs]
+    ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
+    state                   = ggml_add(ctx0, state, k_t_delta);
 
-    // Broadcast both tensors to [chunk_size, chunk_size, H_v, n_seqs]
+    // Compute the attention output
-    // ggml_tensor * gcs_i_broadcast =
+    // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
-    //     ggml_repeat_4d(ctx0, gcs_i, GGML_DELTA_NET_CHUNK, GGML_DELTA_NET_CHUNK, num_chunks * H_v,
+    ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs);  // unsqueeze q_t
-    //                     n_seqs);  // [chunk_size, 1, H_v, n_seqs] -> [chunk_size, chunk_size, H_v, n_seqs]
+    ggml_tensor * state_q        = ggml_mul(ctx0, state, q_t_unsqueezed);
-    // Don't need this, this one will get auto-broadcast
+    // again, since it's over dim = -2, transpose, sum, transpose back
-    ggml_tensor * gcs_j_broadcast =
+    ggml_tensor * core_attn_out =
-        ggml_repeat_4d(ctx0, gcs_j, n_tokens, n_tokens, H_v, n_seqs);  // [1, chunk_size, H_v, n_seqs] -> [chunk_size, chunk_size, H_v, n_seqs]
+        ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
-
-    ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
-
-    // Apply lower triangular mask to ensure attention is causal (only past tokens influence current)
-    decay_mask = ggml_mul(ctx0, decay_mask, causal_diag_mask);
-    // Apply exponential to get the decay mask values
-    decay_mask = ggml_exp(ctx0, decay_mask);
-    // Apply lower triangular mask again to ensure only lower triangular values remain
-    decay_mask = ggml_mul(ctx0, decay_mask, causal_diag_mask);
-
-    cb(decay_mask, "decay_mask", il);
-
-    // attn = -((k_beta @ key.transpose(-1, -2)) * decay_mask).masked_fill(mask, 0)
-    ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
-
-    cb(kmulkbeta, "kmulkbeta", il);
-
-    ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
-    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
-
-    cb(attn, "attn_pre_rec", il);
-
-    // for i in range(1, chunk_size):
-    //          row = attn[..., i, :i].clone()
-    //          sub = attn[..., :i, :i].clone()
-    //          attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2)
-    // attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device)
-    //
-    // We reduce this to a linear triangular solve: AX = B, where B = attn, A = I - tril(A)
-    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
-    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
-
-    ggml_tensor * lin_solve  = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
-    attn                     = ggml_mul(ctx0, lin_solve, causal_mask);
-    attn                     = ggml_add(ctx0, attn, identity);
-
-    // value = attn @ v_beta
-    v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
-
-    cb(v, "value_beta", il);
-
-    // k_cumdecay = attn @ (k_beta * g.exp().unsqueeze(-1))
-    ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
-    ggml_tensor * gexp       = ggml_exp(ctx0, g_cumsum_t);
-
-    cb(gexp, "g_cum_exp", il);
-
-    ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
-
-    cb(kbeta_gexp, "kbeta_gexp", il);
-
-    ggml_tensor * k_cumdecay =
-        ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
-
-    cb(k_cumdecay, "k_cumdecay", il);
-
-    // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
-    attn = ggml_mul_mat(ctx0, k, q);
-    attn = ggml_mul(ctx0, attn, decay_mask);
-    attn = ggml_mul(ctx0, attn, ggml_add(ctx0, identity, causal_mask));
-
-    cb(attn, "attn_decay_key", il);
-
-    ggml_tensor * state_t = ggml_cont(ctx0, ggml_transpose(ctx0, state));
-
-    // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
-    ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay);
-
-    cb(v_prime, "v_prime", il);
-
-    // v_new = v_i - v_prime
-    ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v, v_prime), v_prime);
-
-    ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
-
-    cb(v_new, "v_new", il);
-
-    // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
-    ggml_tensor * q_g_exp    = ggml_mul(ctx0, q, gexp);
-    ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
-
-    cb(attn_inter, "attn_inter", il);
-
-    // core_attn_out[:, :, i] = attn_inter + attn @ v_new
-    ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn);
-
-    cb(v_attn, "v_attn", il);
-
-    ggml_tensor * core_attn_out = ggml_add(ctx0, attn_inter, v_attn);
-
-    cb(core_attn_out, "core_attn_out", il);
-
-    // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
-    // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
-    // key_gdiff = key * g_diff.unsqueeze(-1)
-    // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
-    // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
-
-    ggml_tensor * g_cum_last =
-        ggml_cont(ctx0, ggml_view_4d(ctx0, g_cumsum_t, g_cumsum_t->ne[0], 1, g_cumsum_t->ne[2], g_cumsum_t->ne[3],
-                                    g_cumsum_t->nb[1], g_cumsum_t->nb[2], g_cumsum_t->nb[3],
-                                    g_cumsum_t->nb[0] * (g_cumsum_t->ne[1] - 1)));
-
-    cb(g_cum_last, "g_cum_last", il);
-
-    ggml_tensor * gexp_last =
-        ggml_reshape_4d(ctx0, ggml_exp(ctx0, g_cum_last), 1, 1, g_cum_last->ne[0] * g_cum_last->ne[2], g_cum_last->ne[3]);
-
-    cb(gexp_last, "gexp_last", il);
-
-    ggml_tensor * g_cum_last_3d =
-        ggml_reshape_3d(ctx0, g_cum_last, g_cum_last->ne[0], g_cum_last->ne[2], g_cum_last->ne[3]);
-
-    cb(g_cum_last_3d, "g_cum_last_3d", il);
-
-    ggml_tensor * g_cumsum_3d = ggml_reshape_3d(ctx0, g_cumsum, g_cumsum->ne[0], g_cumsum->ne[2], g_cumsum->ne[3]);
-
-    cb(g_cumsum_3d, "g_cumsum_3d", il);
-
-    ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum_3d, g_cum_last_3d));
-
-    cb(g_diff, "g_diff", il);
-
-    ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
-
-    cb(g_diff_exp, "g_diff_exp", il);
-
-    ggml_tensor * key_gdiff = ggml_mul(ctx0, k,
-                                    ggml_reshape_4d(ctx0, g_diff_exp, 1, g_diff_exp->ne[0], g_diff_exp->ne[1],
-                                                    g_diff_exp->ne[2] * g_diff_exp->ne[3]));
-
-    cb(key_gdiff, "key_gdiff", il);
-
-    ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)));
-
-    cb(kgdmulvnew, "kgdmulvnew", il);
-
-    state = ggml_add(ctx0, ggml_mul(ctx0, state, gexp_last), kgdmulvnew);
 
+    // core_attn_out should be [S_v, 1, H_v, n_seqs] after this
+    cb(core_attn_out, "output_tokens", il);
     cb(state, "new_state", il);
 
-    // flatten output
+    // flatten output, no need to permute since n_tokens is 1 so [S_v, 1, H_v, n_seqs] and [S_v, H_v, 1, n_seqs] are equivalent memory-layout wise
-    ggml_tensor * flat_output =
+    ggml_tensor * flat_output = ggml_reshape_1d(ctx0, core_attn_out, S_v * H_v * n_tokens * n_seqs);
-        ggml_cont_1d(ctx0, ggml_permute(ctx0, core_attn_out, 0, 2, 1, 3), S_v * H_v * n_tokens * n_seqs);
+    ggml_tensor * flat_state  = ggml_reshape_1d(ctx0, state, S_v * S_v * H_v * n_seqs);
-
-    ggml_tensor * flat_state = ggml_cont_1d(ctx0, state, S_v * S_v * H_v * n_seqs);
 
     return ggml_concat(ctx0, flat_output, flat_state, 0);
 }
@@ -712,6 +528,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
         ggml_tensor *        cur,
         ggml_tensor *        causal_mask,
         ggml_tensor *        identity,
+        ggml_tensor *        diag_mask,
         int                  il) {
     const auto * mctx_cur = inp->mctx;
 
@@ -737,11 +554,11 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     cb(mixed_ba, "linear_attn_mixed_ba", il);
 
     int64_t       qkvz_new_dim        = 2 * head_k_dim + 2 * head_v_dim * (num_v_heads / num_k_heads);
-    ggml_tensor * mixed_qkvz_reshaped = ggml_cont_4d(ctx0, mixed_qkvz, qkvz_new_dim, num_k_heads, n_seq_tokens, n_seqs);
+    ggml_tensor * mixed_qkvz_reshaped = ggml_reshape_4d(ctx0, mixed_qkvz, qkvz_new_dim, num_k_heads, n_seq_tokens, n_seqs);
 
     // Reshape mixed_ba: [batch, seq_len, hidden_size] -> [batch, seq_len, num_k_heads, 2*num_v_heads/num_k_heads]
     int64_t       ba_new_dim        = 2 * num_v_heads / num_k_heads;
-    ggml_tensor * mixed_ba_reshaped = ggml_cont_4d(ctx0, mixed_ba, ba_new_dim, num_k_heads, n_seq_tokens, n_seqs);
+    ggml_tensor * mixed_ba_reshaped = ggml_reshape_4d(ctx0, mixed_ba, ba_new_dim, num_k_heads, n_seq_tokens, n_seqs);
 
     // Split mixed_ba into b and a (beta and alpha parameters)
     int64_t split_sizes_ba[2] = {
@@ -762,8 +579,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     ggml_tensor * beta  = ggml_cont_3d(ctx0, b, num_v_heads, n_seq_tokens, n_seqs);
     ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);
 
-    GGML_ASSERT(ggml_nelements(beta) + ggml_nelements(alpha) == ggml_nelements(mixed_ba));
-
     ggml_tensor * alpha_biased   = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
     ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
     cb(alpha_softplus, "a_softplus", il);
@@ -799,9 +614,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
                                    (split_sizes_qkvz[0] + split_sizes_qkvz[1] + split_sizes_qkvz[2]) * sizeof(float));
     cb(z, "z", il);
 
-    GGML_ASSERT(ggml_nelements(query) + ggml_nelements(key) + ggml_nelements(value) + ggml_nelements(z) ==
-                ggml_nelements(mixed_qkvz));
-
     // After creating query, key, and value_reshaped, reshape each to flatten the head dimensions
     // query: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]
     ggml_tensor * query_flat = ggml_cont_3d(ctx0, query, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);
@@ -925,10 +737,13 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     cb(k_conv, "k_conv_predelta", il);
     cb(v_conv, "v_conv_predelta", il);
 
-    // Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens
+    // Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens
-    ggml_tensor * attn_out = n_seq_tokens > CHUNK_SIZE ?
+    ggml_tensor * attn_out;
-        build_delta_net_chunking (q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, il) :
+    if (n_seq_tokens == 1) {
-        build_delta_net_recurrent(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, il);
+        attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
+    } else {
+        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
+    }
     cb(attn_out, "attn_out", il);
 
     // The tensors were concatenated 1d, so we need to extract them 1d as well

llama/llama.cpp/src/unicode.cpp

@@ -520,7 +520,7 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
 
 // use std::wregex to split the text
 static std::vector<size_t> unicode_regex_split_stl(const std::wstring & wtext, const std::wstring & regex_expr, const std::vector<size_t> & offsets) {
-    std::wregex expr(regex_expr);
+    std::wregex expr(regex_expr, std::regex_constants::optimize | std::regex_constants::nosubs);
     std::vector<size_t> bpe_offsets; // store the offset of each word
     bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
     size_t start = 0;
@@ -550,7 +550,7 @@ static std::vector<size_t> unicode_regex_split_stl(const std::wstring & wtext, c
 
 // use std::regex to split the text
 static std::vector<size_t> unicode_regex_split_stl(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
-    std::regex expr(regex_expr);
+    std::regex expr(regex_expr, std::regex_constants::optimize | std::regex_constants::nosubs);
     std::vector<size_t> bpe_offsets; // store the offset of each word
     bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
     size_t start = 0;

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

@@ -0,0 +1,121 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-cpp.h"
+#include "clip.h"
+#include "clip-impl.h"
+#include "clip-model.h"
+
+#include <vector>
+#include <functional>
+
+#define DEFAULT_INTERPOLATION_MODE (GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ANTIALIAS)
+
+struct clip_graph {
+    const clip_model & model;
+    const clip_hparams & hparams;
+    projector_type proj_type;
+
+    // we only support single image per batch
+    const clip_image_f32 & img;
+
+    const int patch_size;
+    const int n_patches_x;
+    const int n_patches_y;
+    const int n_patches;
+    const int n_embd;
+    const int n_head;
+    const int d_head;
+    const int n_layer;
+    const int n_mmproj_embd;
+    const float eps;
+    const float kq_scale;
+    const clip_flash_attn_type flash_attn_type;
+
+    // for debugging
+    const bool debug_graph;
+    std::vector<ggml_tensor *> & debug_print_tensors;
+
+    ggml_context_ptr ctx0_ptr;
+    ggml_context * ctx0;
+    ggml_cgraph * gf;
+
+    clip_graph(clip_ctx * ctx, const clip_image_f32 & img);
+
+    virtual ~clip_graph() = default;
+    virtual ggml_cgraph * build() = 0;
+
+    //
+    // utility functions
+    //
+    void cb(ggml_tensor * cur0, const char * name, int il) const;
+
+    // siglip2 naflex
+    ggml_tensor * resize_position_embeddings(uint32_t interpolation_mode = DEFAULT_INTERPOLATION_MODE);
+
+    // 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
+    ggml_tensor * build_vit(
+                ggml_tensor * inp,
+                int64_t n_pos,
+                norm_type norm_t,
+                ffn_op_type ffn_t,
+                ggml_tensor * learned_pos_embd,
+                std::function<ggml_tensor *(ggml_tensor *, const clip_layer &)> add_pos);
+
+    // build the input after conv2d (inp_raw --> patches)
+    // returns tensor with shape [n_embd, n_patches]
+    ggml_tensor * build_inp();
+
+    ggml_tensor * build_inp_raw(int channels = 3);
+
+    ggml_tensor * build_norm(
+            ggml_tensor * cur,
+            ggml_tensor * mw,
+            ggml_tensor * mb,
+            norm_type type,
+            float norm_eps,
+            int il) const;
+
+    ggml_tensor * build_ffn(
+            ggml_tensor * cur,
+            ggml_tensor * up,
+            ggml_tensor * up_b,
+            ggml_tensor * gate,
+            ggml_tensor * gate_b,
+            ggml_tensor * down,
+            ggml_tensor * down_b,
+            ffn_op_type type_op,
+            int il) const;
+
+    ggml_tensor * build_attn(
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur,
+            ggml_tensor * k_cur,
+            ggml_tensor * v_cur,
+            ggml_tensor * kq_mask,
+            float kq_scale,
+            int il) const;
+
+    // implementation of the 2D RoPE without adding a new op in ggml
+    // this is not efficient (use double the memory), but works on all backends
+    // TODO: there was a more efficient which relies on ggml_view and ggml_rope_ext_inplace, but the rope inplace does not work well with non-contiguous tensors ; we should fix that and revert back to the original implementation in https://github.com/ggml-org/llama.cpp/pull/13065
+    ggml_tensor * build_rope_2d(
+        ggml_context * ctx0,
+        ggml_tensor * cur,
+        ggml_tensor * pos_a, // first half
+        ggml_tensor * pos_b, // second half
+        const float freq_base,
+        const bool interleave_freq
+    );
+
+    // aka pixel_shuffle / pixel_unshuffle / patch_merger (Kimi-VL)
+    // support dynamic resolution
+    ggml_tensor * build_patch_merge_permute(ggml_tensor * cur, int scale_factor);
+
+    // Generic function to stack frames for audio processing
+    // Abstracts out the StackAudioFrames logic used by ultravox
+    ggml_tensor * build_stack(ggml_tensor * cur, int32_t stack_factor, int32_t n_embed);
+};

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

@@ -1,3 +1,5 @@
+#pragma once
+
 #include "ggml.h"
 #include "gguf.h"
 #include "clip.h"
@@ -13,6 +15,8 @@
 
 // Internal header for clip.cpp
 
+#define MTMD_INTERNAL_HEADER
+
 #define KEY_FTYPE               "general.file_type"
 #define KEY_NAME                "general.name"
 #define KEY_DESCRIPTION         "general.description"
@@ -64,6 +68,7 @@
 #define TN_PATCH_EMBD      "v.patch_embd.weight"  // not rename tensor with ".0" postfix for backwrad compat
 #define TN_PATCH_EMBD_1    "v.patch_embd.weight.1"
 #define TN_PATCH_BIAS      "v.patch_embd.bias"
+#define TN_NORM_EMBD       "v.norm_embd.%s"
 #define TN_ATTN_QKV        "%s.blk.%d.attn_qkv.%s"
 #define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
 #define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
@@ -82,6 +87,10 @@
 #define TN_LN_PRE          "%s.pre_ln.%s"
 #define TN_LN_POST         "%s.post_ln.%s"
 #define TN_LLAVA_PROJ      "mm.%d.%s"
+#define TN_MM_UP           "mm.up.%s"
+#define TN_MM_GATE         "mm.gate.%s"
+#define TN_MM_DOWN         "mm.down.%s"
+#define TN_MM_POST_NORM    "mm.post_norm.%s"
 #define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
@@ -91,7 +100,7 @@
 #define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
 #define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
 #define TN_MM_PROJECTOR    "mm.model.fc.weight"         // idefics3
-#define TN_MM_PATCH_MERGER "mm.patch_merger.weight"     // mistral small 3.1
+#define TN_MM_PATCH_MERGER "mm.patch_merger.%s"         // mistral small 3.1, glm4v
 #define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
 #define TN_TOK_GLM_BOI     "adapter.boi"                // glm-edge (these embeddings are not in text model)
 #define TN_TOK_GLM_EOI     "adapter.eoi"                // glm-edge (these embeddings are not in text model)
@@ -132,6 +141,10 @@
 // align x to upper multiple of n
 #define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
 
+// forward declaration
+// TODO: improve this later
+struct clip_ctx;
+
 enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
@@ -149,6 +162,7 @@ enum projector_type {
     PROJECTOR_TYPE_INTERNVL,
     PROJECTOR_TYPE_LLAMA4,
     PROJECTOR_TYPE_QWEN2A,
+    PROJECTOR_TYPE_GLMA,
     PROJECTOR_TYPE_QWEN25O, // will be replaced by QWEN2A or QWEN25VL depending on clip_ctx
     PROJECTOR_TYPE_VOXTRAL,
     PROJECTOR_TYPE_LFM2,
@@ -156,6 +170,7 @@ enum projector_type {
     PROJECTOR_TYPE_LIGHTONOCR,
     PROJECTOR_TYPE_COGVLM,
     PROJECTOR_TYPE_JANUS_PRO,
+    PROJECTOR_TYPE_GLM4V,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -175,6 +190,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_INTERNVL,  "internvl"},
     { PROJECTOR_TYPE_LLAMA4,    "llama4"},
     { PROJECTOR_TYPE_QWEN2A,    "qwen2a"},
+    { PROJECTOR_TYPE_GLMA,      "glma"},
     { PROJECTOR_TYPE_QWEN25O,   "qwen2.5o"},
     { PROJECTOR_TYPE_VOXTRAL,   "voxtral"},
     { PROJECTOR_TYPE_LFM2,      "lfm2"},
@@ -182,6 +198,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_LIGHTONOCR,"lightonocr"},
     { PROJECTOR_TYPE_COGVLM,    "cogvlm"},
     { PROJECTOR_TYPE_JANUS_PRO, "janus_pro"},
+    { PROJECTOR_TYPE_GLM4V,     "glm4v"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {
@@ -485,6 +502,8 @@ static void print_tensor_data(ggml_tensor * t, uint8_t * data, int64_t n) {
     }
 }
 
+void clip_debug_encode(clip_ctx * ctx, int h, int w, float fill_value);
+
 //
 // API used internally with mtmd
 //

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

@@ -0,0 +1,300 @@
+#pragma once
+
+#include "ggml.h"
+#include "clip.h"
+#include "clip-impl.h"
+
+#include <vector>
+#include <unordered_set>
+#include <cstdint>
+#include <cmath>
+
+enum ffn_op_type {
+    FFN_GELU,
+    FFN_GELU_ERF,
+    FFN_SILU,
+    FFN_GELU_QUICK,
+};
+
+enum norm_type {
+    NORM_TYPE_NORMAL,
+    NORM_TYPE_RMS,
+};
+
+enum patch_merge_type {
+    PATCH_MERGE_FLAT,
+    PATCH_MERGE_SPATIAL_UNPAD,
+};
+
+struct clip_hparams {
+    int32_t image_size = 0;
+    int32_t patch_size = 0;
+    int32_t n_embd = 0;
+    int32_t n_ff = 0;
+    int32_t projection_dim = 0;
+    int32_t n_head = 0;
+    int32_t n_layer = 0;
+    // idefics3
+    int32_t image_longest_edge = 0;
+    int32_t image_min_pixels = -1;
+    int32_t image_max_pixels = -1;
+    int32_t n_merge = 0; // number of patch merges **per-side**
+
+    float image_mean[3];
+    float image_std[3];
+
+    // for models using dynamic image size, we need to have a smaller image size to warmup
+    // otherwise, user will get OOM everytime they load the model
+    int32_t warmup_image_size = 0;
+    int32_t warmup_audio_size = 3000;
+
+    ffn_op_type ffn_op = FFN_GELU;
+
+    patch_merge_type mm_patch_merge_type = PATCH_MERGE_FLAT;
+
+    float eps = 1e-6;
+    float rope_theta = 0.0;
+
+    std::vector<clip_image_size> image_res_candidates; // for llava-uhd style models
+    int32_t image_crop_resolution;
+    std::unordered_set<int32_t> vision_feature_layer;
+    int32_t attn_window_size = 0;
+    int32_t n_wa_pattern = 0;
+
+    // audio
+    int32_t n_mel_bins = 0; // whisper preprocessor
+    int32_t proj_stack_factor = 0; // ultravox
+
+    // audio-to-mel preprocessor params
+    int32_t audio_chunk_len   = -1; // in seconds
+    int32_t audio_sample_rate = -1;
+    int32_t audio_n_fft       = -1;
+    int32_t audio_window_len  = -1;
+    int32_t audio_hop_len     = -1;
+
+    // legacy
+    bool has_llava_projector = false;
+    int minicpmv_version = 0;
+    int32_t minicpmv_query_num = 0;         // MiniCPM-V query number
+
+    // custom value provided by user, can be undefined if not set
+    int32_t custom_image_min_tokens = -1;
+    int32_t custom_image_max_tokens = -1;
+
+    void set_limit_image_tokens(int n_tokens_min, int n_tokens_max) {
+        const int cur_merge = n_merge == 0 ? 1 : n_merge;
+        const int patch_area = patch_size * patch_size * cur_merge * cur_merge;
+        image_min_pixels = (custom_image_min_tokens > 0 ? custom_image_min_tokens : n_tokens_min) * patch_area;
+        image_max_pixels = (custom_image_max_tokens > 0 ? custom_image_max_tokens : n_tokens_max) * patch_area;
+        warmup_image_size = static_cast<int>(std::sqrt(image_max_pixels));
+    }
+
+    void set_warmup_n_tokens(int n_tokens) {
+        int n_tok_per_side = static_cast<int>(std::sqrt(n_tokens));
+        GGML_ASSERT(n_tok_per_side * n_tok_per_side == n_tokens && "n_tokens must be n*n");
+        const int cur_merge = n_merge == 0 ? 1 : n_merge;
+        warmup_image_size = n_tok_per_side * patch_size * cur_merge;
+        // TODO: support warmup size for custom token numbers
+    }
+};
+
+struct clip_layer {
+    // attention
+    ggml_tensor * k_w = nullptr;
+    ggml_tensor * k_b = nullptr;
+    ggml_tensor * q_w = nullptr;
+    ggml_tensor * q_b = nullptr;
+    ggml_tensor * v_w = nullptr;
+    ggml_tensor * v_b = nullptr;
+    ggml_tensor * qkv_w = nullptr;
+    ggml_tensor * qkv_b = nullptr;
+
+    ggml_tensor * o_w = nullptr;
+    ggml_tensor * o_b = nullptr;
+
+    ggml_tensor * k_norm = nullptr;
+    ggml_tensor * q_norm = nullptr;
+
+    // layernorm 1
+    ggml_tensor * ln_1_w = nullptr;
+    ggml_tensor * ln_1_b = nullptr;
+
+    ggml_tensor * ff_up_w = nullptr;
+    ggml_tensor * ff_up_b = nullptr;
+    ggml_tensor * ff_gate_w = nullptr;
+    ggml_tensor * ff_gate_b = nullptr;
+    ggml_tensor * ff_down_w = nullptr;
+    ggml_tensor * ff_down_b = nullptr;
+
+    // layernorm 2
+    ggml_tensor * ln_2_w = nullptr;
+    ggml_tensor * ln_2_b = nullptr;
+
+    // layer scale (no bias)
+    ggml_tensor * ls_1_w = nullptr;
+    ggml_tensor * ls_2_w = nullptr;
+
+    // qwen3vl deepstack merger
+    ggml_tensor * deepstack_norm_w = nullptr;
+    ggml_tensor * deepstack_norm_b = nullptr;
+    ggml_tensor * deepstack_fc1_w = nullptr;
+    ggml_tensor * deepstack_fc1_b = nullptr;
+    ggml_tensor * deepstack_fc2_w = nullptr;
+    ggml_tensor * deepstack_fc2_b = nullptr;
+
+    bool has_deepstack() const {
+        return deepstack_fc1_w != nullptr;
+    }
+};
+
+struct clip_model {
+    clip_modality modality = CLIP_MODALITY_VISION;
+    projector_type proj_type = PROJECTOR_TYPE_MLP;
+    clip_hparams hparams;
+
+    // embeddings
+    ggml_tensor * class_embedding = nullptr;
+    ggml_tensor * patch_embeddings_0 = nullptr;
+    ggml_tensor * patch_embeddings_1 = nullptr;  // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL)
+    ggml_tensor * patch_bias = nullptr;
+    ggml_tensor * position_embeddings = nullptr;
+    ggml_tensor * norm_embd_w = nullptr;
+    ggml_tensor * norm_embd_b = nullptr;
+
+    ggml_tensor * pre_ln_w = nullptr;
+    ggml_tensor * pre_ln_b = nullptr;
+
+    std::vector<clip_layer> layers;
+
+    int32_t n_deepstack_layers = 0; // used by Qwen3-VL, calculated from clip_layer
+
+    ggml_tensor * post_ln_w;
+    ggml_tensor * post_ln_b;
+
+    ggml_tensor * projection; // TODO: rename it to fc (fully connected layer)
+    ggml_tensor * mm_fc_w;
+    ggml_tensor * mm_fc_b;
+    ggml_tensor * mm_ffn_up_w = nullptr;
+    ggml_tensor * mm_ffn_up_b = nullptr;
+    ggml_tensor * mm_ffn_gate_w = nullptr;
+    ggml_tensor * mm_ffn_gate_b = nullptr;
+    ggml_tensor * mm_ffn_down_w = nullptr;
+    ggml_tensor * mm_ffn_down_b = nullptr;
+    ggml_tensor * mm_post_norm_w = nullptr;
+    ggml_tensor * mm_post_norm_b = nullptr;
+
+    // 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;
+    ggml_tensor * mm_2_b = nullptr;
+
+    ggml_tensor * image_newline = nullptr;
+
+    // Yi type models with mlp+normalization projection
+    ggml_tensor * mm_1_w = nullptr; // Yi type models have 0, 1, 3, 4
+    ggml_tensor * mm_1_b = nullptr;
+    ggml_tensor * mm_3_w = nullptr;
+    ggml_tensor * mm_3_b = nullptr;
+    ggml_tensor * mm_4_w = nullptr;
+    ggml_tensor * mm_4_b = nullptr;
+
+    // GLMV-Edge projection
+    ggml_tensor * mm_model_adapter_conv_w = nullptr;
+    ggml_tensor * mm_model_adapter_conv_b = nullptr;
+
+    // MobileVLM projection
+    ggml_tensor * mm_model_mlp_1_w = nullptr;
+    ggml_tensor * mm_model_mlp_1_b = nullptr;
+    ggml_tensor * mm_model_mlp_3_w = nullptr;
+    ggml_tensor * mm_model_mlp_3_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_0_0_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_0_1_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_0_1_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc1_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc1_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc2_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc2_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_2_0_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_2_1_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_2_1_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_0_0_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_0_1_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_0_1_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc1_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc1_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc2_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc2_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_2_0_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_2_1_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_2_1_b = nullptr;
+
+    // MobileVLM_V2 projection
+    ggml_tensor * mm_model_mlp_0_w = nullptr;
+    ggml_tensor * mm_model_mlp_0_b = nullptr;
+    ggml_tensor * mm_model_mlp_2_w = nullptr;
+    ggml_tensor * mm_model_mlp_2_b = nullptr;
+    ggml_tensor * mm_model_peg_0_w = nullptr;
+    ggml_tensor * mm_model_peg_0_b = nullptr;
+
+    // MINICPMV projection
+    ggml_tensor * mm_model_pos_embed_k = nullptr;
+    ggml_tensor * mm_model_query = nullptr;
+    ggml_tensor * mm_model_proj = nullptr;
+    ggml_tensor * mm_model_kv_proj = nullptr;
+    ggml_tensor * mm_model_attn_q_w = nullptr;
+    ggml_tensor * mm_model_attn_q_b = nullptr;
+    ggml_tensor * mm_model_attn_k_w = nullptr;
+    ggml_tensor * mm_model_attn_k_b = nullptr;
+    ggml_tensor * mm_model_attn_v_w = nullptr;
+    ggml_tensor * mm_model_attn_v_b = nullptr;
+    ggml_tensor * mm_model_attn_o_w = nullptr;
+    ggml_tensor * mm_model_attn_o_b = nullptr;
+    ggml_tensor * mm_model_ln_q_w = nullptr;
+    ggml_tensor * mm_model_ln_q_b = nullptr;
+    ggml_tensor * mm_model_ln_kv_w = nullptr;
+    ggml_tensor * mm_model_ln_kv_b = nullptr;
+    ggml_tensor * mm_model_ln_post_w = nullptr;
+    ggml_tensor * mm_model_ln_post_b = nullptr;
+
+    // gemma3
+    ggml_tensor * mm_input_proj_w = nullptr;
+    ggml_tensor * mm_soft_emb_norm_w = nullptr;
+
+    // pixtral, glm4v
+    ggml_tensor * token_embd_img_break = nullptr;
+    ggml_tensor * mm_patch_merger_w = nullptr;
+    ggml_tensor * mm_patch_merger_b = nullptr;
+
+    // ultravox / whisper encoder
+    ggml_tensor * conv1d_1_w = nullptr;
+    ggml_tensor * conv1d_1_b = nullptr;
+    ggml_tensor * conv1d_2_w = nullptr;
+    ggml_tensor * conv1d_2_b = nullptr;
+    ggml_tensor * mm_norm_pre_w = nullptr;
+    ggml_tensor * mm_norm_pre_b = nullptr;
+    ggml_tensor * mm_norm_mid_w = nullptr;
+
+    // cogvlm
+    ggml_tensor * mm_post_fc_norm_w = nullptr;
+    ggml_tensor * mm_post_fc_norm_b = nullptr;
+    ggml_tensor * mm_h_to_4h_w = nullptr;
+    ggml_tensor * mm_gate_w = nullptr;
+    ggml_tensor * mm_4h_to_h_w = nullptr;
+    ggml_tensor * mm_boi = nullptr;
+    ggml_tensor * mm_eoi = nullptr;
+
+    bool audio_has_avgpool() const {
+        return proj_type == PROJECTOR_TYPE_QWEN2A
+            || proj_type == PROJECTOR_TYPE_VOXTRAL;
+    }
+
+    bool audio_has_stack_frames() const {
+        return proj_type == PROJECTOR_TYPE_ULTRAVOX
+            || proj_type == PROJECTOR_TYPE_VOXTRAL;
+    }
+};
+
+const clip_hparams * clip_get_hparams(const struct clip_ctx * ctx);

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

@@ -7,6 +7,8 @@
 
 // !!! Internal header, to be used by mtmd only !!!
 
+#define MTMD_INTERNAL_HEADER
+
 struct clip_ctx;
 
 struct clip_image_size {
@@ -34,6 +36,7 @@ struct clip_context_params {
     enum clip_flash_attn_type flash_attn_type;
     int image_min_tokens;
     int image_max_tokens;
+    bool warmup;
 };
 
 struct clip_init_result {
@@ -101,7 +104,7 @@ bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct
 
 int clip_is_minicpmv(const struct clip_ctx * ctx);
 bool clip_is_glm(const struct clip_ctx * ctx);
-bool clip_is_qwen2vl(const struct clip_ctx * ctx);
+bool clip_is_mrope(const struct clip_ctx * ctx);
 bool clip_is_llava(const struct clip_ctx * ctx);
 bool clip_is_gemma3(const struct clip_ctx * ctx);
 

llama/llama.cpp/tools/mtmd/models/cogvlm.cpp

@@ -0,0 +1,98 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_cogvlm::build() {
+    GGML_ASSERT(model.class_embedding != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+
+    const int n_pos = n_patches + 1; // +1 for [CLS]
+
+    // build input and concatenate class embedding
+    ggml_tensor * inp = build_inp();
+    inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+
+    inp = ggml_add(ctx0, inp, model.position_embeddings);
+    cb(inp, "inp_pos", -1);
+
+    ggml_tensor * inpL = inp;
+
+    for (int il = 0; il < n_layer; il++) {
+        auto & layer = model.layers[il];
+        ggml_tensor * cur = inpL;
+
+        cur = ggml_mul_mat(ctx0, layer.qkv_w, cur);
+
+        cur = ggml_add(ctx0, cur, layer.qkv_b);
+
+        ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
+            cur->nb[1], 0);
+        ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
+            cur->nb[1], n_embd * sizeof(float));
+        ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
+            cur->nb[1], 2 * n_embd * sizeof(float));
+
+        cb(Qcur, "Qcur", il);
+        cb(Kcur, "Kcur", il);
+        cb(Vcur, "Vcur", il);
+
+        cur = build_attn(layer.o_w, layer.o_b,
+            Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+        cb(cur, "attn_out", il);
+
+        cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, eps, il);
+        cb(cur, "attn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, inpL);
+        inpL = cur;
+
+        cur = build_ffn(cur,
+            layer.ff_up_w, layer.ff_up_b,
+            layer.ff_gate_w, layer.ff_gate_b,
+            layer.ff_down_w, layer.ff_down_b,
+            hparams.ffn_op, il);
+
+        cb(cur, "ffn_out", il);
+
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, eps, il);
+        cb(cur, "ffn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, inpL);
+        cb(cur, "layer_out", il);
+        inpL = cur;
+
+    }
+
+    // remove CLS token (like build_llama4 does)
+    ggml_tensor * cur = ggml_view_2d(ctx0, inpL,
+        n_embd, n_patches,
+        ggml_row_size(inpL->type, n_embd), 0);
+
+    // Multiply with mm_model_proj
+    cur = ggml_mul_mat(ctx0, model.mm_model_proj, cur);
+
+    // Apply layernorm, weight, bias
+    cur = build_norm(cur, model.mm_post_fc_norm_w, model.mm_post_fc_norm_b, NORM_TYPE_NORMAL, 1e-5, -1);
+
+    // Apply GELU
+    cur = ggml_gelu_inplace(ctx0, cur);
+
+    // Branch 1: multiply with mm_h_to_4h_w
+    ggml_tensor * h_to_4h = ggml_mul_mat(ctx0, model.mm_h_to_4h_w, cur);
+
+    // Branch 2: multiply with mm_gate_w
+    ggml_tensor * gate = ggml_mul_mat(ctx0, model.mm_gate_w, cur);
+
+    // Apply silu
+    gate = ggml_swiglu_split(ctx0, gate, h_to_4h);
+
+    // Apply mm_4h_to_h_w
+    cur = ggml_mul_mat(ctx0, model.mm_4h_to_h_w, gate);
+
+    // Concatenate with boi and eoi
+    cur = ggml_concat(ctx0, model.mm_boi, cur, 1);
+    cur = ggml_concat(ctx0, cur, model.mm_eoi, 1);
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

llama/llama.cpp/tools/mtmd/models/glm4v.cpp

@@ -0,0 +1,120 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_glm4v::build() {
+    GGML_ASSERT(model.patch_bias != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+    GGML_ASSERT(model.class_embedding == nullptr);
+
+    const int batch_size = 1;
+
+    norm_type norm_t = NORM_TYPE_RMS;
+
+    ggml_tensor * inp_raw = build_inp_raw();
+    ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
+    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches * 4);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    GGML_ASSERT(img.nx % (patch_size * 2) == 0);
+    GGML_ASSERT(img.ny % (patch_size * 2) == 0);
+
+    // second conv dimension
+    {
+        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_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_permute(ctx0, inp, 0, 2, 1, 3);
+        inp = ggml_cont_3d(
+            ctx0, inp,
+            n_embd, n_patches_x * n_patches_y, batch_size);
+    }
+
+    // add patch bias
+    inp = ggml_add(ctx0, inp, model.patch_bias);
+    cb(inp, "patch_bias", -1);
+
+    // pos-conv norm
+    inp = build_norm(inp, model.norm_embd_w, model.norm_embd_b, norm_t, eps, -1);
+
+    // calculate absolute position embedding and apply
+    ggml_tensor * learned_pos_embd = resize_position_embeddings(GGML_SCALE_MODE_BICUBIC);
+    learned_pos_embd = ggml_cont_4d(
+        ctx0, learned_pos_embd,
+        n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+    learned_pos_embd = ggml_reshape_4d(
+        ctx0, learned_pos_embd,
+        n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+    learned_pos_embd = ggml_permute(ctx0, learned_pos_embd, 0, 2, 1, 3);
+    learned_pos_embd = ggml_cont_3d(
+        ctx0, learned_pos_embd,
+        n_embd, n_patches_x * n_patches_y, batch_size);
+    cb(learned_pos_embd, "learned_pos_embd", -1);
+
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        return ggml_rope_multi(
+                    ctx0, cur, positions, nullptr,
+                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION,
+                    32768, hparams.rope_theta, 1, 0, 1, 32, 1);
+    };
+
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            norm_t,
+                            hparams.ffn_op,
+                            learned_pos_embd,
+                            add_pos);
+
+    cb(cur, "vit_out", -1);
+    // cb(ggml_sum(ctx0, cur), "vit_out_sum", -1);
+
+    // GLM4V projector
+    // ref: https://github.com/huggingface/transformers/blob/40dc11cd3eb4126652aa41ef8272525affd4a636/src/transformers/models/glm4v/modeling_glm4v.py#L116-L130
+
+    // patch merger (downsample)
+    {
+        int n_merge = hparams.n_merge;
+        GGML_ASSERT(n_merge > 0);
+
+        int n_token_out = n_patches / n_merge / n_merge;
+        cur = ggml_reshape_4d(ctx0, cur, n_embd, n_merge, n_merge, n_token_out);
+        cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 2, 0, 1, 3)); // [n_merge, n_merge, n_embd, n_token_out]
+        cur = ggml_conv_2d(ctx0, model.mm_patch_merger_w, cur, n_merge, n_merge, 0, 0, 1, 1);
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[2], n_token_out); // [n_embd_out, n_token_out]
+
+        cur = ggml_add(ctx0, cur, model.mm_patch_merger_b);
+    }
+
+    // FC projector
+    {
+        cur = ggml_mul_mat(ctx0, model.projection, cur);
+        // default LayerNorm (post_projection_norm)
+        cur = build_norm(cur, model.mm_post_norm_w, model.mm_post_norm_b, NORM_TYPE_NORMAL, 1e-5, -1);
+        cur = ggml_gelu_erf(ctx0, cur);
+        cb(cur, "after_fc_proj", -1);
+    }
+
+    // FFN projector
+    {
+        cur = build_ffn(cur,
+            model.mm_ffn_up_w, model.mm_ffn_up_b,
+            model.mm_ffn_gate_w, model.mm_ffn_gate_b,
+            model.mm_ffn_down_w, model.mm_ffn_down_b,
+            hparams.ffn_op, -1);
+        cb(cur, "after_ffn_proj", -1);
+        // cb(ggml_sum(ctx0, cur), "merged_sum", -1);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

llama/llama.cpp/tools/mtmd/models/internvl.cpp

@@ -0,0 +1,69 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_internvl::build() {
+    GGML_ASSERT(model.class_embedding != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+
+    const int n_pos = n_patches + 1;
+    ggml_tensor * inp = build_inp();
+
+    // add CLS token
+    inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+
+    // The larger models use a different ViT, which uses RMS norm instead of layer norm
+    // ref: https://github.com/ggml-org/llama.cpp/pull/13443#issuecomment-2869786188
+    norm_type norm_t = (hparams.n_embd == 3200 && hparams.n_layer == 45)
+        ? NORM_TYPE_RMS // 6B ViT (Used by InternVL 2.5/3 - 26B, 38B, 78B)
+        : NORM_TYPE_NORMAL; // 300M ViT (Used by all smaller InternVL models)
+
+    ggml_tensor * cur = build_vit(
+                            inp, n_pos,
+                            norm_t,
+                            hparams.ffn_op,
+                            model.position_embeddings,
+                            nullptr);
+
+    // remove CLS token
+    cur = ggml_view_2d(ctx0, cur,
+        n_embd, n_patches,
+        ggml_row_size(cur->type, n_embd), 0);
+
+    // pixel shuffle
+    {
+        const int scale_factor = model.hparams.n_merge;
+        const int bsz    = 1; // batch size, always 1 for now since we don't support batching
+        const int height = n_patches_y;
+        const int width  = n_patches_x;
+        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_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_cont_2d(ctx0, cur,
+            n_embd * scale_factor * scale_factor,
+            cur->ne[1] * cur->ne[2]);
+    }
+
+    // projector (always using GELU activation)
+    {
+        // projector LayerNorm uses pytorch's default eps = 1e-5
+        // ref: https://huggingface.co/OpenGVLab/InternVL3-8B-Instruct/blob/a34d3e4e129a5856abfd6aa6de79776484caa14e/modeling_internvl_chat.py#L79
+        cur = build_norm(cur, model.mm_0_w, model.mm_0_b, NORM_TYPE_NORMAL, 1e-5, -1);
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_3_w, model.mm_3_b,
+            FFN_GELU,
+            -1);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

llama/llama.cpp/tools/mtmd/models/kimivl.cpp

@@ -0,0 +1,63 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_kimivl::build() {
+    // 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.n_merge;
+        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 = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU,
+            -1);
+        cb(cur, "proj_out", -1);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

llama/llama.cpp/tools/mtmd/models/llama4.cpp

@@ -0,0 +1,96 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_llama4::build() {
+    GGML_ASSERT(model.class_embedding != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+
+    const int n_pos = n_patches + 1; // +1 for [CLS]
+
+    // 2D input positions
+    ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+    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_pos);
+    ggml_set_name(pos_w, "pos_w");
+    ggml_set_input(pos_w);
+
+    ggml_tensor * inp = build_inp_raw();
+
+    // Llama4UnfoldConvolution
+    {
+        ggml_tensor * kernel = ggml_reshape_4d(ctx0, model.patch_embeddings_0,
+                                                patch_size, patch_size, 3, n_embd);
+        inp = ggml_im2col(ctx0, kernel, inp, patch_size, patch_size, 0, 0, 1, 1, true, inp->type);
+        inp = ggml_mul_mat(ctx0, model.patch_embeddings_0, inp);
+        inp = ggml_reshape_2d(ctx0, inp, n_embd, n_patches);
+        cb(inp, "patch_conv", -1);
+    }
+
+    // add CLS token
+    inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+
+    // 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
+        // ref: https://github.com/huggingface/transformers/blob/40a493c7ed4f19f08eadb0639cf26d49bfa5e180/src/transformers/models/llama4/modeling_llama4.py#L1312
+        // ref: https://github.com/Blaizzy/mlx-vlm/blob/a57156aa87b33cca6e5ee6cfc14dd4ef8f611be6/mlx_vlm/models/llama4/vision.py#L441
+        return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);
+    };
+    ggml_tensor * cur = build_vit(
+                            inp, n_pos,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            model.position_embeddings,
+                            add_pos);
+
+    // remove CLS token
+    cur = ggml_view_2d(ctx0, cur,
+        n_embd, n_patches,
+        ggml_row_size(cur->type, n_embd), 0);
+
+    // pixel shuffle
+    // based on Llama4VisionPixelShuffleMLP
+    // https://github.com/huggingface/transformers/blob/2932f318a20d9e54cc7aea052e040164d85de7d6/src/transformers/models/llama4/modeling_llama4.py#L1151
+    {
+        const int scale_factor = model.hparams.n_merge;
+        const int bsz = 1; // batch size, always 1 for now since we don't support batching
+        GGML_ASSERT(scale_factor > 0);
+        GGML_ASSERT(n_patches_x == n_patches_y); // llama4 only supports square images
+        cur = ggml_reshape_4d(ctx0, cur,
+            n_embd * scale_factor,
+            n_patches_x / scale_factor,
+            n_patches_y,
+            bsz);
+        cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+        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);
+        // flatten to 2D
+        cur = ggml_cont_2d(ctx0, cur,
+            n_embd * scale_factor * scale_factor,
+            n_patches / scale_factor / scale_factor);
+        cb(cur, "pixel_shuffle", -1);
+    }
+
+    // based on Llama4VisionMLP2 (always uses GELU activation, no bias)
+    {
+        cur = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, cur);
+        cur = ggml_gelu(ctx0, cur);
+        cur = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, cur);
+        cur = ggml_gelu(ctx0, cur);
+        cb(cur, "adapter_mlp", -1);
+    }
+
+    // Llama4MultiModalProjector
+    cur = ggml_mul_mat(ctx0, model.mm_model_proj, cur);
+    cb(cur, "projected", -1);
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

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

@@ -0,0 +1,374 @@
+#include "models.h"
+
+// this graph is used by llava, granite and glm
+// due to having embedding_stack (used by granite), we cannot reuse build_vit
+ggml_cgraph * clip_graph_llava::build() {
+    const int batch_size = 1;
+    const int n_pos = n_patches + (model.class_embedding ? 1 : 0);
+
+    GGML_ASSERT(n_patches_x == n_patches_y && "only square images supported");
+
+    // Calculate the deepest feature layer based on hparams and projector type
+    int max_feature_layer = n_layer;
+    {
+        // Get the index of the second to last layer; this is the default for models that have a llava projector
+        int il_last = hparams.n_layer - 1;
+        int deepest_feature_layer = -1;
+
+        if (proj_type == PROJECTOR_TYPE_MINICPMV || proj_type == PROJECTOR_TYPE_GLM_EDGE) {
+            il_last += 1;
+        }
+
+        // If we set explicit vision feature layers, only go up to the deepest one
+        // NOTE: only used by granite-vision models for now
+        for (const auto & feature_layer : hparams.vision_feature_layer) {
+            if (feature_layer > deepest_feature_layer) {
+                deepest_feature_layer = feature_layer;
+            }
+        }
+        max_feature_layer = deepest_feature_layer < 0 ? il_last : deepest_feature_layer;
+    }
+
+    ggml_tensor * inp = build_inp();
+
+    // concat class_embeddings and patch_embeddings
+    if (model.class_embedding) {
+        inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+    }
+
+    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    inp = ggml_add(ctx0, inp, ggml_get_rows(ctx0, model.position_embeddings, positions));
+
+    ggml_tensor * inpL = inp;
+
+    // pre-layernorm
+    if (model.pre_ln_w) {
+        inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, NORM_TYPE_NORMAL, eps, -1);
+        cb(inpL, "pre_ln", -1);
+    }
+
+    std::vector<ggml_tensor *> embedding_stack;
+    const auto & vision_feature_layer = hparams.vision_feature_layer;
+
+    // loop over layers
+    for (int il = 0; il < max_feature_layer; il++) {
+        auto & layer = model.layers[il];
+        ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states
+
+        // If this is an embedding feature layer, save the output.
+        // NOTE: 0 index here refers to the input to the encoder.
+        if (vision_feature_layer.find(il) != vision_feature_layer.end()) {
+            embedding_stack.push_back(cur);
+        }
+
+        // layernorm1
+        cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, eps, il);
+        cb(cur, "layer_inp_normed", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = ggml_mul_mat(ctx0, layer.q_w, cur);
+            if (layer.q_b) {
+                Qcur = ggml_add(ctx0, Qcur, layer.q_b);
+            }
+
+            ggml_tensor * Kcur = ggml_mul_mat(ctx0, layer.k_w, cur);
+            if (layer.k_b) {
+                Kcur = ggml_add(ctx0, Kcur, layer.k_b);
+            }
+
+            ggml_tensor * Vcur = ggml_mul_mat(ctx0, layer.v_w, cur);
+            if (layer.v_b) {
+                Vcur = ggml_add(ctx0, Vcur, layer.v_b);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_pos);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(layer.o_w, layer.o_b,
+                Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, inpL);
+
+        inpL = cur; // inpL = residual, cur = hidden_states
+
+        cb(cur, "ffn_inp", il);
+
+        // layernorm2
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, eps, il);
+        cb(cur, "ffn_inp_normed", il);
+
+        // ffn
+        cur = build_ffn(cur,
+            layer.ff_up_w, layer.ff_up_b,
+            layer.ff_gate_w, layer.ff_gate_b,
+            layer.ff_down_w, layer.ff_down_b,
+            hparams.ffn_op, il);
+
+        cb(cur, "ffn_out", il);
+
+        // residual 2
+        cur = ggml_add(ctx0, inpL, cur);
+        cb(cur, "layer_out", il);
+
+        inpL = cur;
+    }
+
+    // post-layernorm
+    if (model.post_ln_w) {
+        inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, NORM_TYPE_NORMAL, eps, -1);
+    }
+
+    ggml_tensor * embeddings = inpL;
+
+    // process vision feature layers (used by granite)
+    {
+        // final layer is a vision feature layer
+        if (vision_feature_layer.find(max_feature_layer) != vision_feature_layer.end()) {
+            embedding_stack.push_back(inpL);
+        }
+
+        // If feature layers are explicitly set, stack them (if we have multiple)
+        if (!embedding_stack.empty()) {
+            embeddings = embedding_stack[0];
+            for (size_t i = 1; i < embedding_stack.size(); i++) {
+                embeddings = ggml_concat(ctx0, embeddings, embedding_stack[i], 0);
+            }
+        }
+    }
+
+    // llava projector (also used by granite)
+    if (hparams.has_llava_projector) {
+        embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
+
+        ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+        ggml_set_name(patches, "patches");
+        ggml_set_input(patches);
+
+        // shape [1, 576, 1024]
+        // ne is whcn, ne = [1024, 576, 1, 1]
+        embeddings = ggml_get_rows(ctx0, embeddings, patches);
+
+        // print_tensor_info(embeddings, "embeddings");
+
+        // llava projector
+        if (proj_type == PROJECTOR_TYPE_MLP) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+
+            embeddings = ggml_gelu(ctx0, embeddings);
+            if (model.mm_2_w) {
+                embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
+                embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+            }
+        }
+        else if (proj_type == PROJECTOR_TYPE_MLP_NORM) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+            // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
+            // First LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w),
+                                model.mm_1_b);
+
+            // GELU activation
+            embeddings = ggml_gelu(ctx0, embeddings);
+
+            // Second linear layer
+            embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_3_b);
+
+            // Second LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w),
+                                model.mm_4_b);
+        }
+        else if (proj_type == PROJECTOR_TYPE_LDP) {
+            // MobileVLM projector
+            int n_patch = 24;
+            ggml_tensor * mlp_1 = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, embeddings);
+            mlp_1 = ggml_add(ctx0, mlp_1, model.mm_model_mlp_1_b);
+            mlp_1 = ggml_gelu(ctx0, mlp_1);
+            ggml_tensor * mlp_3 = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, mlp_1);
+            mlp_3 = ggml_add(ctx0, mlp_3, model.mm_model_mlp_3_b);
+            // mlp_3 shape = [1, 576, 2048], ne = [2048, 576, 1, 1]
+
+            // block 1
+            ggml_tensor * block_1 = nullptr;
+            {
+                // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24]
+                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);
+
+                // layer norm
+                // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_0_1_w), model.mm_model_block_1_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+
+                // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // hardswish
+                ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+                // block_1_hw shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1], block_1 shape = [1, 2048], ne = [2048, 1, 1, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_2_1_w), model.mm_model_block_1_block_2_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // residual
+                block_1 = ggml_add(ctx0, mlp_3, block_1);
+            }
+
+            // block_2
+            {
+                // stride = 2
+                block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1);
+
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // layer norm
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_0_1_w), model.mm_model_block_2_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // hardswish
+                ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                // not sure the parameters is right for globalAvgPooling
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+
+                // block_1_hw shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1], block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_2_1_w), model.mm_model_block_2_block_2_1_b);
+                block_1 = ggml_reshape_3d(ctx0, block_1, block_1->ne[0], block_1->ne[1] * block_1->ne[2], block_1->ne[3]);
+                // block_1 shape = [1, 144, 2048], ne = [2048, 144, 1]
+            }
+            embeddings = block_1;
+        }
+        else if (proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            int n_patch = 24;
+            ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+            mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b);
+            mlp_0 = ggml_gelu(ctx0, mlp_0);
+            ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0);
+            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_permute(ctx0, mlp_2, 1, 0, 2, 3);
+            // mlp_2 ne = [576, 2048, 1, 1]
+            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]
+            ggml_tensor * peg_0 = ggml_conv_2d_dw(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
+            peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b);
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, mlp_2);
+            peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]);
+            embeddings = peg_0;
+        }
+        else {
+            GGML_ABORT("fatal error");
+        }
+    }
+
+    // glm projector
+    else if (proj_type == PROJECTOR_TYPE_GLM_EDGE) {
+        size_t gridsz = (size_t)sqrt(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));
+        embeddings = ggml_add(ctx0, embeddings, model.mm_model_adapter_conv_b);
+        // GLU
+        {
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
+            embeddings = ggml_gelu_inplace(ctx0, embeddings);
+            ggml_tensor * x = embeddings;
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, embeddings);
+            x = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w,x);
+            embeddings = ggml_swiglu_split(ctx0, embeddings, x);
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, embeddings);
+        }
+        // arrangement of BOI/EOI token embeddings
+        // note: these embeddings are not present in text model, hence we cannot process them as text tokens
+        // see: https://huggingface.co/THUDM/glm-edge-v-2b/blob/main/siglip.py#L53
+        {
+            embeddings = ggml_concat(ctx0, model.mm_boi, embeddings, 1); // BOI
+            embeddings = ggml_concat(ctx0, embeddings, model.mm_eoi, 1); // EOI
+        }
+    }
+
+    else {
+        GGML_ABORT("llava: unknown projector type");
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    return gf;
+}

llama/llama.cpp/tools/mtmd/models/minicpmv.cpp

@@ -0,0 +1,114 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_minicpmv::build() {
+    GGML_ASSERT(model.class_embedding == nullptr);
+    const int n_pos       = n_patches;
+    const int n_embd_proj = n_mmproj_embd;
+
+    // position embeddings for the projector (not for ViT)
+    // see: https://huggingface.co/openbmb/MiniCPM-o-2_6/blob/main/resampler.py#L70
+    // base frequency omega
+    ggml_tensor * omega = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_embd_proj / 4);
+    ggml_set_name(omega, "omega");
+    ggml_set_input(omega);
+
+    // 2D input positions (using float for sinusoidal embeddings)
+    ggml_tensor * pos_h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_pos);
+    ggml_set_name(pos_h, "pos_h");
+    ggml_set_input(pos_h);
+    ggml_tensor * pos_w = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_pos);
+    ggml_set_name(pos_w, "pos_w");
+    ggml_set_input(pos_w);
+
+    // for selecting learned pos embd, used by ViT
+    struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    ggml_tensor * learned_pos_embd = ggml_get_rows(ctx0, model.position_embeddings, positions);
+
+    ggml_tensor * inp = build_inp();
+    ggml_tensor * embeddings = build_vit(
+                            inp, n_pos,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            learned_pos_embd,
+                            nullptr);
+
+    // resampler projector (it is just another transformer)
+
+    ggml_tensor * q = model.mm_model_query;
+    ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings);
+
+    // norm
+    q = build_norm(q, model.mm_model_ln_q_w,  model.mm_model_ln_q_b,  NORM_TYPE_NORMAL, eps, -1);
+    v = build_norm(v, model.mm_model_ln_kv_w, model.mm_model_ln_kv_b, NORM_TYPE_NORMAL, eps, -1);
+
+    // calculate sinusoidal pos embd
+    ggml_tensor * pos_embed = nullptr;
+    {
+        // outer product
+        ggml_tensor * omega_b = ggml_repeat_4d(ctx0, omega, omega->ne[0], n_pos, 1, 1); // n_pos rows
+        ggml_tensor * theta_x = ggml_mul(ctx0, omega_b, pos_w);
+        ggml_tensor * theta_y = ggml_mul(ctx0, omega_b, pos_h);
+        // sin and cos
+        ggml_tensor * pos_embd_x = ggml_concat(
+            ctx0,
+            ggml_sin(ctx0, theta_x),
+            ggml_cos(ctx0, theta_x),
+            0 // concat on first dim
+        );
+        ggml_tensor * pos_embd_y = ggml_concat(
+            ctx0,
+            ggml_sin(ctx0, theta_y),
+            ggml_cos(ctx0, theta_y),
+            0 // concat on first dim
+        );
+        pos_embed = ggml_concat(ctx0, pos_embd_x, pos_embd_y, 0);
+    }
+
+    // k = v + pos_embed
+    ggml_tensor * k = ggml_add(ctx0, v, pos_embed);
+
+    // attention
+    {
+        const int d_head = 128;
+        int n_head = n_embd_proj/d_head;
+        // Use actual config value if available, otherwise fall back to hardcoded values
+        int num_query = 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);
+        ggml_tensor * K = ggml_add(ctx0,
+            ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k),
+            model.mm_model_attn_k_b);
+        ggml_tensor * V = ggml_add(ctx0,
+            ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v),
+            model.mm_model_attn_v_b);
+
+        Q = ggml_reshape_3d(ctx0, Q, d_head, n_head, num_query);
+        K = ggml_reshape_3d(ctx0, K, d_head, n_head, n_pos);
+        V = ggml_reshape_3d(ctx0, V, d_head, n_head, n_pos);
+
+        cb(Q, "resampler_Q", -1);
+        cb(K, "resampler_K", -1);
+        cb(V, "resampler_V", -1);
+
+        float resampler_kq_scale = 1.0f/ sqrtf(float(d_head));
+        embeddings = build_attn(
+            model.mm_model_attn_o_w,
+            model.mm_model_attn_o_b,
+            Q, K, V, nullptr, resampler_kq_scale, -1);
+        cb(embeddings, "resampler_attn_out", -1);
+    }
+    // layernorm
+    embeddings = build_norm(embeddings, model.mm_model_ln_post_w, model.mm_model_ln_post_b, NORM_TYPE_NORMAL, eps, -1);
+
+    // projection
+    embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings);
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    return gf;
+}

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

@@ -0,0 +1,6 @@
+package models
+
+// #cgo CXXFLAGS: -std=c++17
+// #cgo CPPFLAGS: -I${SRCDIR}/../../../include -I${SRCDIR}/../../../common -I${SRCDIR}/../../../vendor
+// #cgo CPPFLAGS: -I${SRCDIR}/../../../../../ml/backend/ggml/ggml/include
+import "C"