refactor: bpe and spm tokenizer merges

- merge candidates and pairs which are essentially the same other than
  the type for rank/score
- use binaryheap in sentencepiece instead of implement custom structure
- update merging algorithm so it uses about 15% less allocations

.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\"'" >>$GITHUB_OUTPUT
-          echo VERSION="${GITHUB_REF_NAME#v}" >>$GITHUB_OUTPUT
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${GITHUB_REF_NAME#v}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" | tee -a $GITHUB_OUTPUT
+          echo VERSION="${GITHUB_REF_NAME#v}" | 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'

.golangci.yaml

@@ -1,77 +1,51 @@
 version: "2"
 linters:
-  default: none
   enable:
     - asasalint
     - bidichk
     - bodyclose
     - containedctx
-    - copyloopvar
-    - errcheck
-    - errorlint
-    - exptostd
     - gocheckcompilerdirectives
-    - gocritic
-    - govet
-    - ineffassign
     - intrange
     - makezero
     - misspell
-    - modernize
     - nilerr
-    - nilnil
     - nolintlint
     - nosprintfhostport
-    - perfsprint
-    - prealloc
-    - sloglint
-    - staticcheck
     - unconvert
-    - unused
-    - usestdlibvars
     - usetesting
     - wastedassign
     - whitespace
+  disable:
+    - errcheck
+    - usestdlibvars
   settings:
-    errcheck:
-      exclude-functions:
-        - fmt.Fprintf
-    perfsprint:
-      strconcat: false
-      concat-loop: false
+    govet:
+      disable:
+        - unusedresult
     staticcheck:
       checks:
         - all
-        # Using a deprecated function, variable, constant or field.
-        # https://staticcheck.dev/docs/checks/#SA1019
+        - -QF* # disable quick fix suggestions
         - -SA1019
-        # Incorrect or missing package comment.
-        # https://staticcheck.dev/docs/checks/#ST1000
-        - -ST1000
-        # Poorly chosen identifier.
-        # https://staticcheck.dev/docs/checks/#ST1003
-        - -ST1003
-        # The documentation of an exported function should start with the function's name.
-        # https://staticcheck.dev/docs/checks/#ST1020
-        - -ST1020
-        # The documentation of an exported type should start with type's name.
-        # https://staticcheck.dev/docs/checks/#ST1021
-        - -ST1021
-        # The documentation of an exported variable or constant should start with variable's name.
-        # https://staticcheck.dev/docs/checks/#ST1022
-        - -ST1022
-    usestdlibvars:
-      http-method: false
-      http-status-code: false
-
+        - -ST1000 # package comment format
+        - -ST1003 # underscores in package names
+        - -ST1005 # error strings should not be capitalized
+        - -ST1012 # error var naming (ErrFoo)
+        - -ST1016 # receiver name consistency
+        - -ST1020 # comment on exported function format
+        - -ST1021 # comment on exported type format
+        - -ST1022 # comment on exported var format
+        - -ST1023 # omit type from declaration
+severity:
+  default: error
+  rules:
+    - linters:
+        - gofmt
+        - goimports
+        - intrange
+      severity: info
 formatters:
   enable:
-    - gci
     - gofmt
     - gofumpt
-  settings:
-    gci:
-      sections:
-        - standard
-        - default
-        - localmodule

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=3cfa9c3f125763305b4226bc032f1954f08990dc
+FETCH_HEAD=17f7f4baad8b3a716ee139da7bb56ae984e8c0fa
 
 .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/examples/chat/main.go

@@ -15,19 +15,19 @@ func main() {
 	}
 
 	messages := []api.Message{
-		api.Message{
+		{
 			Role:    "system",
 			Content: "Provide very brief, concise responses",
 		},
-		api.Message{
+		{
 			Role:    "user",
 			Content: "Name some unusual animals",
 		},
-		api.Message{
+		{
 			Role:    "assistant",
 			Content: "Monotreme, platypus, echidna",
 		},
-		api.Message{
+		{
 			Role:    "user",
 			Content: "which of these is the most dangerous?",
 		},

api/types.go

@@ -283,11 +283,12 @@ func (pt PropertyType) String() string {
 }
 
 type ToolProperty struct {
-	AnyOf       []ToolProperty `json:"anyOf,omitempty"`
-	Type        PropertyType   `json:"type,omitempty"`
-	Items       any            `json:"items,omitempty"`
-	Description string         `json:"description,omitempty"`
-	Enum        []any          `json:"enum,omitempty"`
+	AnyOf       []ToolProperty          `json:"anyOf,omitempty"`
+	Type        PropertyType            `json:"type,omitempty"`
+	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

@@ -169,37 +169,47 @@ 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;
 		
-  // Calculate total required buffer size first
-  int totalSize = 0;
-  for(NSURL* url in urls) {
-      char tempBuf[PATH_MAX];
-      if(![url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX]) {
-          return DLG_URLFAIL;
-      }
-      totalSize += strlen(tempBuf) + 1; // +1 for null terminator
-  }
-  totalSize += 1; // Final null terminator
+		// Calculate total required buffer size first
+		int totalSize = 0;
+		for(NSURL* url in urls) {
+			char tempBuf[PATH_MAX];
+			if(![url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX]) {
+				return DLG_URLFAIL;
+			}
+			totalSize += strlen(tempBuf) + 1; // +1 for null terminator
+		}
+		totalSize += 1; // Final null terminator
 
-  if(totalSize > self->params->nbuf) {
-      // Not enough buffer space
-      return DLG_URLFAIL;
-  }
+		if(totalSize > self->params->nbuf) {
+			// Not enough buffer space
+			return DLG_URLFAIL;
+		}
 
-  // Now actually copy the paths (we know we have space)
-  bufPtr = self->params->buf;
-  for(NSURL* url in urls) {
-      char tempBuf[PATH_MAX];
-      [url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX];
-      int pathLen = strlen(tempBuf);
-      strcpy(bufPtr, tempBuf);
-      bufPtr += pathLen + 1;
-  }
-  *bufPtr = '\0'; // Final null terminator
+		// Now actually copy the paths (we know we have space)
+		bufPtr = self->params->buf;
+		for(NSURL* url in urls) {
+			char tempBuf[PATH_MAX];
+			[url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX];
+			int pathLen = strlen(tempBuf);
+			strcpy(bufPtr, tempBuf);
+			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)
-			settings.Models = ""
-			s.store.SetSettings(settings)
+			slog.Warn("models path not accessible, using default", "path", settings.Models, "err", err)
 		}
 	}
 	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;
-    plan: string;
-    bio: string;
-    firstName: string;
-    lastName: string;
-    overThreshold: boolean;
+    name: string;
+    bio?: string;
+    avatarurl?: string;
+    firstname?: string;
+    lastname?: string;
+    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.plan = source["plan"];
+        this.name = source["name"];
         this.bio = source["bio"];
-        this.firstName = source["firstName"];
-        this.lastName = source["lastName"];
-        this.overThreshold = source["overThreshold"];
+        this.avatarurl = source["avatarurl"];
+        this.firstname = source["firstname"];
+        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
@@ -42,44 +41,50 @@ function uint8ArrayToBase64(uint8Array: Uint8Array): string {
 }
 
 export async function fetchUser(): Promise<User | null> {
-  try {
-    const response = await fetch(`${API_BASE}/api/v1/me`, {
-      method: "GET",
-      headers: {
-        "Content-Type": "application/json",
-      },
-    });
-
-    if (response.ok) {
-      const userData: User = await response.json();
-      return userData;
-    }
-
-    return null;
-  } catch (error) {
-    console.error("Error fetching user:", error);
-    return null;
-  }
-}
-
-export async function fetchConnectUrl(): Promise<string> {
-  const response = await fetch(`${API_BASE}/api/v1/connect`, {
-    method: "GET",
+  const response = await fetch(`${API_BASE}/api/me`, {
+    method: "POST",
     headers: {
       "Content-Type": "application/json",
     },
   });
 
-  if (!response.ok) {
-    throw new Error("Failed to fetch connect URL");
+  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;
   }
 
-  const data = await response.json();
-  return data.connect_url;
+  if (response.status === 401 || response.status === 403) {
+    return null;
+  }
+
+  throw new Error(`Failed to fetch user: ${response.status}`);
+}
+
+export async function fetchConnectUrl(): Promise<string> {
+  const response = await fetch(`${API_BASE}/api/me`, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+    },
+  });
+
+  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");
 }
 
 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",
@@ -204,12 +209,10 @@ export async function* sendMessage(
     data: uint8ArrayToBase64(att.data),
   }));
 
-  // Only send think parameter when actually requesting thinking
-  // Don't send false as it causes issues with some providers
+  // Send think parameter when it's explicitly set (true, false, or a non-empty string).
   const shouldSendThink =
     think !== undefined &&
-    ((typeof think === "boolean" && think) ||
-      (typeof think === "string" && think !== ""));
+    (typeof think === "boolean" || (typeof think === "string" && think !== ""));
 
   const response = await fetch(`${API_BASE}/api/v1/chat/${chatId}`, {
     method: "POST",
@@ -391,7 +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",
@@ -400,7 +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

@@ -50,21 +50,33 @@ export default function Thinking({
   // Position content to show bottom when collapsed
   useEffect(() => {
     if (isCollapsed && contentRef.current && wrapperRef.current) {
-      const contentHeight = contentRef.current.scrollHeight;
-      const wrapperHeight = wrapperRef.current.clientHeight;
-      if (contentHeight > wrapperHeight) {
-        const translateY = -(contentHeight - wrapperHeight);
-        contentRef.current.style.transform = `translateY(${translateY}px)`;
-        setHasOverflow(true);
-      } else {
-        setHasOverflow(false);
-      }
+      requestAnimationFrame(() => {
+        if (!contentRef.current || !wrapperRef.current) return;
+
+        const contentHeight = contentRef.current.scrollHeight;
+        const wrapperHeight = wrapperRef.current.clientHeight;
+        if (contentHeight > wrapperHeight) {
+          const translateY = -(contentHeight - wrapperHeight);
+          contentRef.current.style.transform = `translateY(${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:
-      !initialDataLoaded ||
-      (userQuery.isLoading && userQuery.data === undefined), // Show loading until initial data is loaded
+    isLoading,
     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
-let initialUserDataPromise: Promise<User | null> | null = null;
-
-// Initialize user data on app startup
-const initializeUserData = async () => {
-  try {
-    const userData = await fetchUser();
+fetchUser().then((userData) => {
+  if (userData) {
     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

@@ -101,15 +101,14 @@ type HealthResponse struct {
 }
 
 type User struct {
-	ID            string `json:"id"`
-	Name          string `json:"name"`
-	Email         string `json:"email"`
-	AvatarURL     string `json:"avatarURL"`
-	Plan          string `json:"plan"`
-	Bio           string `json:"bio"`
-	FirstName     string `json:"firstName"`
-	LastName      string `json:"lastName"`
-	OverThreshold bool   `json:"overThreshold"`
+	ID        string `json:"id"`
+	Email     string `json:"email"`
+	Name      string `json:"name"`
+	Bio       string `json:"bio,omitempty"`
+	AvatarURL string `json:"avatarurl,omitempty"`
+	FirstName string `json:"firstname,omitempty"`
+	LastName  string `json:"lastname,omitempty"`
+	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")
-	if ollamaHost == "" {
-		ollamaHost = "http://127.0.0.1:11434"
-	}
+	var (
+		proxy   http.Handler
+		proxyMu sync.Mutex
+	)
 
-	if !strings.HasPrefix(ollamaHost, "http://") && !strings.HasPrefix(ollamaHost, "https://") {
-		ollamaHost = "http://" + ollamaHost
-	}
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		proxyMu.Lock()
+		p := proxy
+		proxyMu.Unlock()
 
-	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)
-		})
-	}
+		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)
+					}
 
-	s.log().Info("configuring ollama proxy", "target", target.String())
+					err = WaitForServer(context.Background(), 10*time.Second)
+					if err == nil {
+						break
+					}
+				}
 
-	proxy := httputil.NewSingleHostReverseProxy(target)
+				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
+				}
 
-	originalDirector := proxy.Director
-	proxy.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)
-	}
+				target := envconfig.Host()
+				s.log().Info("configuring ollama proxy", "target", target.String())
 
-	proxy.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)
-	}
+				newProxy := httputil.NewSingleHostReverseProxy(target)
 
-	return proxy
+				originalDirector := newProxy.Director
+				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)
+				}
+
+				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)
+				}
+
+				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/v1/me", handle(s.me))
-	mux.Handle("POST /api/v1/disconnect", handle(s.disconnect))
-	mux.Handle("GET /api/v1/connect", handle(s.connectURL))
-	mux.Handle("GET /api/v1/health", handle(s.health))
+	mux.Handle("GET /api/version", ollamaProxy)
+	mux.Handle("HEAD /api/version", ollamaProxy)
+	mux.Handle("POST /api/me", ollamaProxy)
+	mux.Handle("POST /api/signout", ollamaProxy)
 
 	// 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 {
-	timeout := time.Now().Add(10 * time.Second)
-	// TODO: this avoids an error on first load of the app
-	// however we should either show a loading state or
-	// wait for the Ollama server to be ready before redirecting
-	for {
+// WaitForServer waits for the Ollama server to be ready
+func WaitForServer(ctx context.Context, timeout time.Duration) error {
+	deadline := time.Now().Add(timeout)
+	for time.Now().Before(deadline) {
 		c, err := api.ClientFromEnvironment()
 		if err != nil {
 			return err
 		}
 		if _, err := c.Version(ctx); err == nil {
-			break
-		}
-		if time.Now().After(timeout) {
-			return fmt.Errorf("timeout waiting for Ollama server to be ready")
+			slog.Debug("ollama server is ready")
+			return nil
 		}
 		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))
-			if cds.DwData == 1 { // Our identifier for URL scheme messages
+			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)		|

cmd/bench/bench.go

@@ -48,8 +48,8 @@ func OutputMetrics(w io.Writer, format string, metrics []Metrics, verbose bool)
 	case "benchstat":
 		if verbose {
 			printHeader := func() {
-				fmt.Printf("sysname: %s\n", runtime.GOOS)
-				fmt.Printf("machine: %s\n", runtime.GOARCH)
+				fmt.Fprintf(w, "sysname: %s\n", runtime.GOOS)
+				fmt.Fprintf(w, "machine: %s\n", runtime.GOARCH)
 			}
 			once.Do(printHeader)
 		}
@@ -147,6 +147,17 @@ func BenchmarkChat(fOpt flagOptions) error {
 		return err
 	}
 
+	var out io.Writer = os.Stdout
+	if fOpt.outputFile != nil && *fOpt.outputFile != "" {
+		f, err := os.OpenFile(*fOpt.outputFile, os.O_CREATE|os.O_WRONLY, 0o644)
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "ERROR: cannot open output file %s: %v\n", *fOpt.outputFile, err)
+			return err
+		}
+		defer f.Close()
+		out = f
+	}
+
 	for _, model := range models {
 		for range *fOpt.epochs {
 			options := make(map[string]interface{})
@@ -241,13 +252,14 @@ func BenchmarkChat(fOpt flagOptions) error {
 				},
 			}
 
-			OutputMetrics(os.Stdout, *fOpt.format, metrics, *fOpt.verbose)
+			OutputMetrics(out, *fOpt.format, metrics, *fOpt.verbose)
 
 			if *fOpt.keepAlive > 0 {
 				time.Sleep(time.Duration(*fOpt.keepAlive*float64(time.Second)) + 200*time.Millisecond)
 			}
 		}
 	}
+
 	return nil
 }
 

cmd/cmd.go

@@ -1430,7 +1430,7 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		latest.Summary()
 	}
 
-	return &api.Message{Role: role, Content: fullResponse.String()}, nil
+	return &api.Message{Role: role, Thinking: thinkingContent.String(), Content: fullResponse.String()}, nil
 }
 
 func generate(cmd *cobra.Command, opts runOptions) error {

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":
@@ -200,14 +202,20 @@ 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":
 		conv = &gptossModel{}
 	case "DeepseekOCRForCausalLM":
 		conv = &deepseekocr{}
+	case "DeepseekV3ForCausalLM":
+		conv = &deepseek2Model{}
 	default:
 		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}

convert/convert_deepseek2.go

@@ -0,0 +1,173 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"log/slog"
+	"regexp"
+	"strconv"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type deepseek2Model struct {
+	ModelParameters               // architectures, vocab_size
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	HiddenLayers          uint32  `json:"num_hidden_layers"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RMSNormEPS            float32 `json:"rms_norm_eps"`
+
+	RopeTheta     float32 `json:"rope_theta"`
+	QKNopeHeadDim uint32  `json:"qk_nope_head_dim"`
+	QKRopeHeadDim uint32  `json:"qk_rope_head_dim"`
+	KVLoraRank    uint32  `json:"kv_lora_rank"`
+	QLoraRank     uint32  `json:"q_lora_rank"`
+	VHeadDim      uint32  `json:"v_head_dim"`
+
+	ExpertCount            uint32  `json:"n_routed_experts"`
+	ExpertSharedCount      uint32  `json:"n_shared_experts"`
+	ExpertIntermediateSize uint32  `json:"moe_intermediate_size"`
+	ExpertUsedCount        uint32  `json:"num_experts_per_tok"`
+	ExpertWeightsNorm      bool    `json:"norm_topk_prob"`
+	ExpertWeightsScale     float32 `json:"routed_scaling_factor"`
+
+	ScoringFunc            string `json:"scoring_func"`
+	LeadingDenseBlockCount uint32 `json:"first_k_dense_replace"`
+
+	RopeScaling struct {
+		Factor                        float32 `json:"factor"`
+		OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
+		Type                          string  `json:"type"`
+		MScaleAllDim                  float32 `json:"mscale_all_dim"`
+	} `json:"rope_scaling"`
+
+	Architecture string
+}
+
+func (p *deepseek2Model) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "deepseek2"
+	kv["general.type"] = "model"
+	kv["deepseek2.block_count"] = p.HiddenLayers
+
+	numHeads := p.NumAttentionHeads
+	numKVHeads := p.NumKeyValueHeads
+
+	kv["deepseek2.attention.head_count"] = numHeads
+	kv["deepseek2.attention.head_count_kv"] = numKVHeads
+	kv["deepseek2.attention.key_length"] = p.QKNopeHeadDim + p.QKRopeHeadDim
+	kv["deepseek2.attention.kv_lora_rank"] = p.KVLoraRank
+	kv["deepseek2.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["deepseek2.attention.q_lora_rank"] = p.QLoraRank
+	kv["deepseek2.attention.value_length"] = p.VHeadDim
+	kv["deepseek2.context_length"] = p.MaxPositionEmbeddings
+	kv["deepseek2.embedding_length"] = p.HiddenSize
+	kv["deepseek2.expert_count"] = p.ExpertCount
+	kv["deepseek2.expert_feed_forward_length"] = p.ExpertIntermediateSize
+	kv["deepseek2.expert_shared_count"] = p.ExpertSharedCount
+
+	var scoringFunc uint32
+	switch p.ScoringFunc {
+	case "softmax":
+		// not currently supported in the model, but needed for Deepseek-OCR
+		scoringFunc = 1
+	case "sigmoid":
+		scoringFunc = 2
+	}
+	kv["deepseek2.expert_gating_func"] = scoringFunc
+	kv["deepseek2.expert_used_count"] = p.ExpertUsedCount
+	kv["deepseek2.expert_weights_norm"] = p.ExpertWeightsNorm
+	kv["deepseek2.expert_weights_scale"] = p.ExpertWeightsScale
+	kv["deepseek2.feed_forward_length"] = p.IntermediateSize
+	kv["deepseek2.leading_dense_block_count"] = p.LeadingDenseBlockCount
+
+	kv["deepseek2.rope.dimension_count"] = p.QKRopeHeadDim
+	kv["deepseek2.rope.freq_base"] = cmp.Or(p.RopeTheta, 10000.0)
+	kv["deepseek2.rope.scaling.factor"] = p.RopeScaling.Factor
+	kv["deepseek2.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeddings
+	kv["deepseek2.rope.scaling.type"] = p.RopeScaling.Type
+	kv["deepseek2.rope.scaling.yarn_log_multiplier"] = 0.1 * p.RopeScaling.MScaleAllDim
+
+	kv["tokenizer.ggml.pre"] = "deepseek-v3"
+
+	return kv
+}
+
+func (p *deepseek2Model) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"language_model.", "",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.kv_a_proj_with_mqa", "attn_kv_a_mqa",
+		"self_attn.kv_a_layernorm", "attn_kv_a_norm",
+		"self_attn.kv_b_proj", "attn_kv_b",
+		"self_attn.q_a_proj", "attn_q_a",
+		"self_attn.q_a_layernorm", "attn_q_a_norm",
+		"self_attn.q_b_proj", "attn_q_b",
+		"self_attn.o_proj", "attn_output",
+		"post_attention_layernorm", "ffn_norm",
+		"mlp.shared_experts.down_proj", "ffn_down_shexp",
+		"mlp.shared_experts.gate_proj", "ffn_gate_shexp",
+		"mlp.shared_experts.up_proj", "ffn_up_shexp",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"mlp.gate.e_score_correction_bias", "exp_probs_b.bias",
+		"mlp.gate", "ffn_gate_inp",
+	}
+}
+
+func (p *deepseek2Model) Tensors(s []Tensor) (out []*ggml.Tensor) {
+	merges := make([]merge, p.HiddenLayers*3)
+	for i := range p.HiddenLayers {
+		merges[i*3+0] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.gate_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
+		}
+		merges[i*3+1] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.up_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
+		}
+		merges[i*3+2] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.down_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
+		}
+	}
+
+	skipLayer := func(n string, minValue uint32) bool {
+		re := regexp.MustCompile(`^blk\.(\d+)`)
+		matches := re.FindStringSubmatch(n)
+		if matches == nil {
+			return false
+		}
+
+		blkNum, err := strconv.Atoi(matches[1])
+		if err != nil {
+			return false
+		}
+
+		return uint32(blkNum) >= minValue
+	}
+
+	out, s = mergeTensors(s, merges...)
+	for _, t := range s {
+		// skip any additional layers (such as the Multi-Token Prediction layer)
+		if skipLayer(t.Name(), p.HiddenLayers) {
+			slog.Debug("skipping layer", "name", t.Name())
+			continue
+		}
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+	return out
+}

convert/convert_gemma3.go

@@ -2,6 +2,7 @@ package convert
 
 import (
 	"cmp"
+	"slices"
 
 	"github.com/ollama/ollama/fs/ggml"
 )
@@ -26,16 +27,26 @@ type gemma3Model struct {
 		NumChannels       uint32  `json:"num_channels"`        // num_channels 3
 		PatchSize         uint32  `json:"patch_size"`          // patch_size 14
 	} `json:"vision_config"`
-	MaxPositionEmbeddings    uint32  `json:"max_position_embeddings"`
-	NumAttentionHeads        uint32  `json:"num_attention_heads"`
-	NumKeyValueHeads         uint32  `json:"num_key_value_heads"`
-	RMSNormEPS               float32 `json:"rms_norm_eps"`
-	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"`
-	SlidingWindow            uint32  `json:"sliding_window"`
-	MultiModalTokensPerImage uint32  `json:"mm_tokens_per_image"`
+	MaxPositionEmbeddings    uint32   `json:"max_position_embeddings"`
+	NumAttentionHeads        uint32   `json:"num_attention_heads"`
+	NumKeyValueHeads         uint32   `json:"num_key_value_heads"`
+	RMSNormEPS               float32  `json:"rms_norm_eps"`
+	HeadDim                  uint32   `json:"head_dim"`
+	FinalLogitSoftcap        float32  `json:"final_logit_softcapping"`
+	RopeLocalTheta           float32  `json:"rope_local_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

@@ -29,6 +29,17 @@ type mistral3Model struct {
 		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"`
 	} `json:"text_config"`
 	VisionModel struct {
 		NumAttentionHeads uint32  `json:"num_attention_heads"`
@@ -41,6 +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"`
@@ -61,8 +75,25 @@ func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
 	kv["mistral3.attention.layer_norm_rms_epsilon"] = p.TextModel.RMSNormEPS
 	kv["mistral3.attention.key_length"] = p.TextModel.HeadDim
 	kv["mistral3.attention.value_length"] = p.TextModel.HeadDim
-	kv["mistral3.rope.dimension_count"] = p.TextModel.HiddenSize / p.TextModel.NumHiddenLayers
-	kv["mistral3.rope.freq_base"] = p.TextModel.RopeTheta
+	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
+	}
+	if p.TextModel.RopeParameters.Llama4ScalingBeta != nil {
+		kv["mistral3.rope.scaling_beta"] = *p.TextModel.RopeParameters.Llama4ScalingBeta
+	}
 
 	// Vision configuration
 	kv["mistral3.vision.block_count"] = p.VisionModel.NumHiddenLayers
@@ -74,7 +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",
+	}
+}

discover/runner.go

@@ -147,7 +147,7 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 			wg.Add(1)
 			go func(i int) {
 				defer wg.Done()
-				extraEnvs := ml.GetVisibleDevicesEnv(devices[i : i+1])
+				extraEnvs := ml.GetVisibleDevicesEnv(devices[i:i+1], true)
 				devices[i].AddInitValidation(extraEnvs)
 				if len(bootstrapDevices(ctx2ndPass, devices[i].LibraryPath, extraEnvs)) == 0 {
 					slog.Debug("filtering device which didn't fully initialize",
@@ -333,7 +333,8 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 			defer cancel()
 
 			// Apply any dev filters to avoid re-discovering unsupported devices, and get IDs correct
-			devFilter := ml.GetVisibleDevicesEnv(devices)
+			// We avoid CUDA filters here to keep ROCm from failing to discover GPUs in a mixed environment
+			devFilter := ml.GetVisibleDevicesEnv(devices, false)
 
 			for dir := range libDirs {
 				updatedDevices := bootstrapDevices(ctx, []string{ml.LibOllamaPath, dir}, devFilter)

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/capabilities/tool-calling.mdx

@@ -15,7 +15,7 @@ Also known as "single-shot" tool calling.
     ```shell
     curl -s http://localhost:11434/api/chat -H "Content-Type: application/json" -d '{
       "model": "qwen3",
-      "messages": [{"role": "user", "content": "What's the temperature in New York?"}],
+      "messages": [{"role": "user", "content": "What is the temperature in New York?"}],
       "stream": false,
       "tools": [
         {
@@ -41,7 +41,7 @@ Also known as "single-shot" tool calling.
     curl -s http://localhost:11434/api/chat -H "Content-Type: application/json" -d '{
       "model": "qwen3",
       "messages": [
-        {"role": "user", "content": "What's the temperature in New York?"},
+        {"role": "user", "content": "What is the temperature in New York?"},
         {
           "role": "assistant",
           "tool_calls": [
@@ -90,7 +90,7 @@ Also known as "single-shot" tool calling.
       }
       return temperatures.get(city, "Unknown")
 
-    messages = [{"role": "user", "content": "What's the temperature in New York?"}]
+    messages = [{"role": "user", "content": "What is the temperature in New York?"}]
 
     # pass functions directly as tools in the tools list or as a JSON schema
     response = chat(model="qwen3", messages=messages, tools=[get_temperature], think=True)
@@ -146,7 +146,7 @@ Also known as "single-shot" tool calling.
       },
     ]
 
-    const messages = [{ role: 'user', content: "What's the temperature in New York?" }]
+    const messages = [{ role: 'user', content: "What is the temperature in New York?" }]
 
     const response = await ollama.chat({
       model: 'qwen3',
@@ -609,7 +609,7 @@ def get_temperature(city: str) -> str:
   return temperatures.get(city, 'Unknown')
 
 
-messages = [{'role': 'user', 'content': "What's the temperature in New York?"}]
+messages = [{'role': 'user', 'content': "What is the temperature in New York?"}]
 
 while True:
   stream = chat(
@@ -684,7 +684,7 @@ const getTemperatureTool = {
 }
 
 async function agentLoop() {
-  const messages = [{ role: 'user', content: "What's the temperature in New York?" }]
+  const messages = [{ role: 'user', content: "What is the temperature in New York?" }]
 
   while (true) {
     const stream = await ollama.chat({

docs/development.md

@@ -49,6 +49,8 @@ Install prerequisites:
     - [Ninja](https://github.com/ninja-build/ninja/releases)
 - (Optional) NVIDIA GPU support
     - [CUDA SDK](https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_network)
+- (Optional) VULKAN GPU support
+    - [VULKAN SDK](https://vulkan.lunarg.com/sdk/home) - useful for AMD/Intel GPUs
 
 Then, configure and build the project:
 
@@ -57,6 +59,17 @@ cmake -B build
 cmake --build build --config Release
 ```
 
+> Building for Vulkan requires VULKAN_SDK environment variable:
+> 
+> PowerShell
+> ```powershell
+> $env:VULKAN_SDK="C:\VulkanSDK\<version>"
+> ```
+> CMD
+> ```cmd
+> set VULKAN_SDK=C:\VulkanSDK\<version>
+> ```
+
 > [!IMPORTANT]
 > Building for ROCm requires additional flags:
 > ```
@@ -65,6 +78,7 @@ cmake --build build --config Release
 > ```
 
 
+
 Lastly, run Ollama:
 
 ```shell
@@ -84,7 +98,9 @@ Install prerequisites:
     - [ROCm](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/install/quick-start.html)
 - (Optional) NVIDIA GPU support
     - [CUDA SDK](https://developer.nvidia.com/cuda-downloads)
-
+- (Optional) VULKAN GPU support
+    - [VULKAN SDK](https://vulkan.lunarg.com/sdk/home) - useful for AMD/Intel GPUs
+    - Or install via package manager: `sudo apt install vulkan-sdk` (Ubuntu/Debian) or `sudo dnf install vulkan-sdk` (Fedora/CentOS)
 > [!IMPORTANT]
 > Ensure prerequisites are in `PATH` before running CMake.
 

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())]
@@ -829,8 +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 {
-		t.Fatalf("expected 6 prompt tokens, got %d", res.PromptEvalCount)
+	if res.PromptEvalCount != 8 {
+		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 {
-		t.Fatalf("expected 12 prompt tokens, got %d", res.PromptEvalCount)
+	if res.PromptEvalCount != 16 {
+		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) {
-				if err.Error() != "input exceeds maximum context length" {
+			check: func(t *testing.T, res *api.EmbedResponse, err error) {
+				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)
+				}
+			})
+		})
+	}
+}

integration/llm_image_test.go

@@ -33,6 +33,9 @@ func TestVisionModels(t *testing.T) {
 			// Qwen 3 VL mixture of experts
 			model: "qwen3-vl:30b",
 		},
+		{
+			model: "ministral-3",
+		},
 	}
 
 	for _, v := range testCases {

integration/tools_test.go

@@ -30,6 +30,7 @@ func TestAPIToolCalling(t *testing.T) {
 		"mistral":       6,
 		"qwen2.5":       6,
 		"qwen2":         6,
+		"ministral-3":   20,
 		"mistral-nemo":  9,
 		"mistral-small": 16,
 		"mixtral:8x22b": 80,

integration/utils_test.go

@@ -38,6 +38,7 @@ var (
 
 	// Note: add newer models at the top of the list to test them first
 	ollamaEngineChatModels = []string{
+		"ministral-3",
 		"qwen3-coder:30b",
 		"gpt-oss:20b",
 		"gemma3n:e2b",
@@ -167,6 +168,7 @@ var (
 		"medllama2",
 		"megadolphin",
 		"minicpm-v",
+		"ministral-3",
 		"mistral-large",
 		"mistral-nemo",
 		"mistral-openorca",
@@ -270,6 +272,7 @@ var (
 		"mistral",
 		"qwen2.5",
 		"qwen2",
+		"ministral-3",
 		"mistral-nemo",
 		"mistral-small",
 		"mixtral:8x22b",

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "3cfa9c3f125763305b4226bc032f1954f08990dc";
+char const *LLAMA_COMMIT = "17f7f4baad8b3a716ee139da7bb56ae984e8c0fa";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/.rsync-filter

@@ -22,6 +22,9 @@ include /src/llama.*
 include /src/llama-*.*
 include /src/unicode-data.*
 include /src/unicode.*
+include /src/models/
+include /src/models/*.h
+include /src/models/*.cpp
 include /vendor/
 include /vendor/miniaudio/
 include /vendor/miniaudio/*.h

llama/llama.cpp/common/common.cpp

@@ -8,6 +8,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
+#include "sampling.h"
 
 #include <algorithm>
 #include <cinttypes>
@@ -26,7 +27,6 @@
 #include <sstream>
 #include <string>
 #include <thread>
-#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 
@@ -60,6 +60,14 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+common_time_meas::common_time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
+
+common_time_meas::~common_time_meas() {
+    if (t_start_us >= 0) {
+        t_acc += ggml_time_us() - t_start_us;
+    }
+}
+
 //
 // CPU utils
 //
@@ -355,11 +363,7 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD
 }
 
 void common_init() {
-    llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
-        if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
-            common_log_add(common_log_main(), level, "%s", text);
-        }
-    }, NULL);
+    llama_log_set(common_log_default_callback, NULL);
 
 #ifdef NDEBUG
     const char * build_type = "";
@@ -690,7 +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;
@@ -750,10 +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
@@ -778,11 +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 = converter.from_bytes(path);
+    std::wstring wpath = utf8_to_wstring(path);
 
     // if the path already exists, check whether it's a directory
     const DWORD attributes = GetFileAttributesW(wpath.c_str());
@@ -855,6 +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) {
@@ -889,6 +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
@@ -908,11 +941,130 @@ std::string fs_get_cache_file(const std::string & filename) {
     return cache_directory + filename;
 }
 
+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;
+
+    std::filesystem::path dir(path);
+    if (!std::filesystem::exists(dir) || !std::filesystem::is_directory(dir)) {
+        return files;
+    }
+
+    for (const auto & entry : std::filesystem::directory_iterator(dir)) {
+        try {
+            // Only include regular files (skip directories)
+            const auto & p = entry.path();
+            if (std::filesystem::is_regular_file(p)) {
+                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
+            continue;
+        }
+    }
+
+    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(
+    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;
+
+        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;
+        }
+    };
+
+    auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
+        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;
+        }
+    };
+
+    // Sampling sequence
+    if (!(config & common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS)) {
+        char buf[512] = {0};
+        if (llama_model_meta_val_str(model, llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE), buf, sizeof(buf)) > 0) {
+            const std::vector<std::string> sampler_names = string_split<std::string>(std::string(buf), ';');
+            if (!sampler_names.empty()) {
+                sparams.samplers = common_sampler_types_from_names(sampler_names, true);
+            }
+        }
+    }
+
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_K),           sparams.top_k,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_P),           sparams.top_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIN_P),           sparams.min_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY), sparams.xtc_probability, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD),   sparams.xtc_threshold,   common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TEMP),            sparams.temp,            common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP);
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N),  sparams.penalty_last_n,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT),  sparams.penalty_repeat,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT);
+    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT),        sparams.mirostat,        common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT);
+    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU),    sparams.mirostat_tau,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU);
+    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) {
     common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
@@ -924,6 +1076,8 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
+    common_init_sampler_from_model(model, params.sampling);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     auto cparams = common_context_params_to_llama(params);

llama/llama.cpp/common/common.h

@@ -2,17 +2,19 @@
 
 #pragma once
 
+#include "ggml-opt.h"
+#include "llama-cpp.h"
+
 #include <set>
 #include <sstream>
 #include <string>
 #include <string_view>
 #include <vector>
 #include <map>
-#include <sstream>
-#include <cmath>
 
-#include "ggml-opt.h"
-#include "llama-cpp.h"
+#if defined(_WIN32) && !defined(_WIN32_WINNT)
+#define _WIN32_WINNT 0x0A00
+#endif
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -28,7 +30,14 @@
     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();
+
+    const int64_t t_start_us;
+
+    int64_t & t_acc;
+};
 
 struct common_adapter_lora_info {
     std::string path;
@@ -133,6 +142,22 @@ struct common_grammar_trigger {
     llama_token token = LLAMA_TOKEN_NULL;
 };
 
+enum common_params_sampling_config : uint64_t {
+    COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS        = 1 << 0,
+    COMMON_PARAMS_SAMPLING_CONFIG_TOP_K           = 1 << 1,
+    COMMON_PARAMS_SAMPLING_CONFIG_TOP_P           = 1 << 2,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIN_P           = 1 << 3,
+    COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY = 1 << 4,
+    COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD   = 1 << 5,
+    COMMON_PARAMS_SAMPLING_CONFIG_TEMP            = 1 << 6,
+    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N  = 1 << 7,
+    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT  = 1 << 8,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT        = 1 << 9,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU    = 1 << 10,
+    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA    = 1 << 11,
+};
+
+
 // sampling parameters
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@@ -165,6 +190,8 @@ struct common_params_sampling {
     bool    no_perf            = false; // disable performance metrics
     bool    timing_per_token   = false;
 
+    uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers
+
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 
 
@@ -198,6 +225,7 @@ 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 {
@@ -344,7 +372,7 @@ 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;
@@ -406,6 +434,8 @@ struct common_params {
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
+    int image_min_tokens = -1;
+    int image_max_tokens = -1;
 
     // finetune
     struct lr_opt lr;
@@ -451,14 +481,21 @@ 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
+    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;
 
     // batched-bench params
-    bool is_pp_shared = false;
+    bool is_pp_shared   = false;
+    bool is_tg_separate = false;
 
     std::vector<int32_t> n_pp;
     std::vector<int32_t> n_tg;
@@ -505,6 +542,10 @@ struct common_params {
     // return false from callback to abort model loading or true to continue
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
+
+    bool has_speculative() const {
+        return !speculative.model.path.empty() || !speculative.model.hf_repo.empty();
+    }
 };
 
 // call once at the start of a program if it uses libcommon
@@ -599,12 +640,28 @@ 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);
 
+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(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
 //

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

@@ -268,10 +268,10 @@ static bool is_reserved_name(const std::string & name) {
 }
 
 std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
-std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"]");
+std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"\\\\]");
 std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
 std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
-    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}
+    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };
 
 std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
@@ -303,6 +303,8 @@ static std::string format_literal(const std::string & literal) {
     return "\"" + escaped + "\"";
 }
 
+std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }
+
 class SchemaConverter {
 private:
     friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
@@ -601,7 +603,10 @@ private:
     }
 
     std::string _resolve_ref(const std::string & ref) {
-        std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
+        auto it = ref.find('#');
+        std::string ref_fragment = it != std::string::npos ? ref.substr(it + 1) : ref;
+        static const std::regex nonalphanumeric_regex(R"([^a-zA-Z0-9-]+)");
+        std::string ref_name = "ref" + std::regex_replace(ref_fragment, nonalphanumeric_regex, "-");
         if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
             _refs_being_resolved.insert(ref);
             json resolved = _refs[ref];
@@ -774,11 +779,24 @@ public:
                         std::vector<std::string> tokens = string_split(pointer, "/");
                         for (size_t i = 1; i < tokens.size(); ++i) {
                             std::string sel = tokens[i];
-                            if (target.is_null() || !target.contains(sel)) {
+                            if (target.is_object() && target.contains(sel)) {
+                                target = target[sel];
+                            } else if (target.is_array()) {
+                                size_t sel_index;
+                                try {
+                                    sel_index = std::stoul(sel);
+                                } catch (const std::invalid_argument & e) {
+                                    sel_index = target.size();
+                                }
+                                if (sel_index >= target.size()) {
+                                    _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
+                                    return;
+                                }
+                                target = target[sel_index];
+                            } else {
                                 _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
                                 return;
                             }
-                            target = target[sel];
                         }
                         _refs[ref] = target;
                     }
@@ -956,7 +974,7 @@ 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());

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

@@ -18,4 +18,6 @@ struct common_grammar_options {
     bool dotall = false;
 };
 
+std::string gbnf_format_literal(const std::string & literal);
+
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options = {});

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;
     }
 
@@ -442,3 +419,23 @@ void common_log_set_prefix(struct common_log * log, bool prefix) {
 void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
+
+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*/) {
+    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_DEFAULT_LLAMA 0
+#define LOG_LEVEL_DEBUG  4
+#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,
@@ -36,6 +42,8 @@ extern int common_log_verbosity_thold;
 
 void common_log_set_verbosity_thold(int verbosity); // not thread-safe
 
+void common_log_default_callback(enum ggml_log_level level, const char * text, void * user_data);
+
 // the common_log uses an internal worker thread to print/write log messages
 // when the worker thread is paused, incoming log messages are discarded
 struct common_log;
@@ -65,10 +73,11 @@ 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
@@ -93,14 +102,14 @@ 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 LOGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_NONE, verbosity, __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_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  0,                 __VA_ARGS__)
-#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0,                 __VA_ARGS__)
-#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
-#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  0,                 __VA_ARGS__)
+#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_LEVEL_DEBUG,  __VA_ARGS__)
+#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  LOG_LEVEL_INFO,   __VA_ARGS__)
+#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  LOG_LEVEL_WARN,   __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,  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

@@ -3,9 +3,10 @@
 #include "common.h"
 #include "log.h"
 
-#include <cmath>
-#include <unordered_map>
 #include <algorithm>
+#include <cmath>
+#include <cstring>
+#include <unordered_map>
 
 // the ring buffer works similarly to std::deque, but with a fixed capacity
 // TODO: deduplicate with llama-impl.h
@@ -112,6 +113,13 @@ struct common_sampler {
 
     llama_token_data_array cur_p;
 
+    void reset() {
+        prev.clear();
+
+        llama_sampler_reset(grmr);
+        llama_sampler_reset(chain);
+    }
+
     void set_logits(struct llama_context * ctx, int idx) {
         const auto * logits = llama_get_logits_ith(ctx, idx);
 
@@ -128,6 +136,12 @@ struct common_sampler {
 
         cur_p = { cur.data(), cur.size(), -1, false };
     }
+
+    common_time_meas tm() {
+        return common_time_meas(t_total_us, params.no_perf);
+    }
+
+    mutable int64_t t_total_us = 0;
 };
 
 std::string common_params_sampling::print() const {
@@ -298,6 +312,8 @@ 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) {
         llama_sampler_accept(gsmpl->grmr, token);
     }
@@ -308,9 +324,7 @@ void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, boo
 }
 
 void common_sampler_reset(struct common_sampler * gsmpl) {
-    llama_sampler_reset(gsmpl->grmr);
-
-    llama_sampler_reset(gsmpl->chain);
+    gsmpl->reset();
 }
 
 struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
@@ -327,16 +341,54 @@ struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl) {
     // TODO: measure grammar performance
 
+    const double t_sampling_ms = gsmpl ? 1e-3*gsmpl->t_total_us : 0;
+
+    llama_perf_sampler_data data_smpl;
+    llama_perf_context_data data_ctx;
+
+    memset(&data_smpl, 0, sizeof(data_smpl));
+    memset(&data_ctx,  0, sizeof(data_ctx));
+
     if (gsmpl) {
-        llama_perf_sampler_print(gsmpl->chain);
+        auto & data = data_smpl;
+
+        data = llama_perf_sampler(gsmpl->chain);
+
+        // note: the sampling time includes the samplers time + extra time spent in common/sampling
+        LOG_INF("%s:    sampling time = %10.2f ms\n", __func__, t_sampling_ms);
+        LOG_INF("%s:    samplers time = %10.2f ms / %5d tokens\n", __func__, data.t_sample_ms, data.n_sample);
     }
+
     if (ctx) {
-        llama_perf_context_print(ctx);
+        auto & data = data_ctx;
+
+        data = llama_perf_context(ctx);
+
+        const double t_end_ms = 1e-3 * ggml_time_us();
+
+        const double t_total_ms = t_end_ms - data.t_start_ms;
+        const double t_unacc_ms = t_total_ms - (t_sampling_ms + data.t_p_eval_ms + data.t_eval_ms);
+        const double t_unacc_pc = 100.0 * t_unacc_ms /  t_total_ms;
+
+        LOG_INF("%s:        load time = %10.2f ms\n", __func__, data.t_load_ms);
+        LOG_INF("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
+                __func__, data.t_p_eval_ms, data.n_p_eval, data.t_p_eval_ms / data.n_p_eval, 1e3 / data.t_p_eval_ms * data.n_p_eval);
+        LOG_INF("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+                __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval);
+        LOG_INF("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval));
+        LOG_INF("%s: unaccounted time = %10.2f ms / %5.1f %%      (total - sampling - prompt eval - eval) / (total)\n", __func__, t_unacc_ms, t_unacc_pc);
+        LOG_INF("%s:    graphs reused = %10d\n", __func__, data.n_reused);
+
         llama_memory_breakdown_print(ctx);
     }
 }
 
 llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
+    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);
 
     auto & grmr  = gsmpl->grmr;
@@ -428,6 +480,8 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
 // helpers
 
 llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort) {
+    const auto tm = gsmpl->tm();
+
     auto * res = &gsmpl->cur_p;
 
     if (do_sort && !res->sorted) {

llama/llama.cpp/include/llama.h

@@ -83,6 +83,7 @@ extern "C" {
         LLAMA_ROPE_TYPE_NORM   = 0,
         LLAMA_ROPE_TYPE_NEOX   = GGML_ROPE_TYPE_NEOX,
         LLAMA_ROPE_TYPE_MROPE  = GGML_ROPE_TYPE_MROPE,
+        LLAMA_ROPE_TYPE_IMROPE = GGML_ROPE_TYPE_IMROPE,
         LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION,
     };
 
@@ -245,6 +246,21 @@ extern "C" {
         LLAMA_KV_OVERRIDE_TYPE_STR,
     };
 
+    enum llama_model_meta_key {
+        LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE,
+        LLAMA_MODEL_META_KEY_SAMPLING_TOP_K,
+        LLAMA_MODEL_META_KEY_SAMPLING_TOP_P,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIN_P,
+        LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY,
+        LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD,
+        LLAMA_MODEL_META_KEY_SAMPLING_TEMP,
+        LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N,
+        LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU,
+        LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA,
+    };
+
     struct llama_model_kv_override {
         enum llama_model_kv_override_type tag;
 
@@ -460,7 +476,11 @@ extern "C" {
     LLAMA_API bool llama_supports_gpu_offload(void);
     LLAMA_API bool llama_supports_rpc        (void);
 
+    // NOTE: After creating a llama_context, it is recommended to query the actual values using these functions
+    //       In some cases the requested values via llama_context_params may differ from the actual values used by the context
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_ctx_seq  (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);
@@ -481,6 +501,7 @@ extern "C" {
 
     LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_embd     (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_embd_inp (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_layer    (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head_kv  (const struct llama_model * model);
@@ -512,6 +533,9 @@ extern "C" {
     // Get the number of metadata key/value pairs
     LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model);
 
+    // Get sampling metadata key name. Returns nullptr if the key is invalid
+    LLAMA_API const char * llama_model_meta_key_str(enum llama_model_meta_key key);
+
     // Get metadata key name by index
     LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
 
@@ -584,7 +608,7 @@ extern "C" {
     LLAMA_API int32_t llama_adapter_meta_val_str_by_index(const struct llama_adapter_lora * adapter, int32_t i, char * buf, size_t buf_size);
 
     // Manually free a LoRA adapter
-    // Note: loaded adapters will be free when the associated model is deleted
+    // NOTE: loaded adapters will be free when the associated model is deleted
     LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter);
 
     // Get the invocation tokens if the current lora is an alora
@@ -1110,8 +1134,6 @@ extern "C" {
     //        // sample from the logits of the last token in the batch
     //        const llama_token id = llama_sampler_sample(smpl, ctx, -1);
     //
-    //        // accepting the token updates the internal state of certain samplers (e.g. grammar, repetition, etc.)
-    //        llama_sampler_accept(smpl, id);
     //        ...
     //    }
     //

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

@@ -32,6 +32,9 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_QWEN2VL,          "qwen2vl"          },
     { LLM_ARCH_QWEN3,            "qwen3"            },
     { LLM_ARCH_QWEN3MOE,         "qwen3moe"         },
+    { LLM_ARCH_QWEN3NEXT,        "qwen3next"        },
+    { LLM_ARCH_QWEN3VL,          "qwen3vl"          },
+    { LLM_ARCH_QWEN3VLMOE,       "qwen3vlmoe"       },
     { LLM_ARCH_PHI2,             "phi2"             },
     { LLM_ARCH_PHI3,             "phi3"             },
     { LLM_ARCH_PHIMOE,           "phimoe"           },
@@ -72,6 +75,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_JAIS,             "jais"             },
     { LLM_ARCH_NEMOTRON,         "nemotron"         },
     { LLM_ARCH_NEMOTRON_H,       "nemotron_h"       },
+    { LLM_ARCH_NEMOTRON_H_MOE,   "nemotron_h_moe"   },
     { LLM_ARCH_EXAONE,           "exaone"           },
     { LLM_ARCH_EXAONE4,          "exaone4"          },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
@@ -89,6 +93,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_BAILINGMOE2,      "bailingmoe2"      },
     { LLM_ARCH_DOTS1,            "dots1"            },
     { LLM_ARCH_ARCEE,            "arcee"            },
+    { LLM_ARCH_AFMOE,            "afmoe"            },
     { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
     { LLM_ARCH_ERNIE4_5_MOE,     "ernie4_5-moe"     },
     { LLM_ARCH_HUNYUAN_MOE,      "hunyuan-moe"      },
@@ -104,23 +109,40 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_SEED_OSS,         "seed_oss"         },
     { LLM_ARCH_GROVEMOE,         "grovemoe"         },
     { LLM_ARCH_APERTUS,          "apertus"          },
+    { LLM_ARCH_MINIMAX_M2,       "minimax-m2"       },
+    { LLM_ARCH_COGVLM,           "cogvlm"           },
+    { LLM_ARCH_RND1,             "rnd1"             },
+    { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
+    { LLM_ARCH_MISTRAL3,         "mistral3"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
 static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
-    { LLM_KV_GENERAL_TYPE,                 "general.type"                          },
-    { LLM_KV_GENERAL_ARCHITECTURE,         "general.architecture"                  },
-    { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version"          },
-    { LLM_KV_GENERAL_ALIGNMENT,            "general.alignment"                     },
-    { LLM_KV_GENERAL_FILE_TYPE,            "general.file_type"                     },
-    { LLM_KV_GENERAL_NAME,                 "general.name"                          },
-    { LLM_KV_GENERAL_AUTHOR,               "general.author"                        },
-    { LLM_KV_GENERAL_VERSION,              "general.version"                       },
-    { LLM_KV_GENERAL_URL,                  "general.url"                           },
-    { LLM_KV_GENERAL_DESCRIPTION,          "general.description"                   },
-    { LLM_KV_GENERAL_LICENSE,              "general.license"                       },
-    { LLM_KV_GENERAL_SOURCE_URL,           "general.source.url"                    },
-    { LLM_KV_GENERAL_SOURCE_HF_REPO,       "general.source.huggingface.repository" },
+    { LLM_KV_GENERAL_TYPE,                     "general.type"                          },
+    { LLM_KV_GENERAL_ARCHITECTURE,             "general.architecture"                  },
+    { LLM_KV_GENERAL_QUANTIZATION_VERSION,     "general.quantization_version"          },
+    { LLM_KV_GENERAL_ALIGNMENT,                "general.alignment"                     },
+    { LLM_KV_GENERAL_FILE_TYPE,                "general.file_type"                     },
+    { LLM_KV_GENERAL_SAMPLING_SEQUENCE,        "general.sampling.sequence"             },
+    { LLM_KV_GENERAL_SAMPLING_TOP_K,           "general.sampling.top_k"                },
+    { LLM_KV_GENERAL_SAMPLING_TOP_P,           "general.sampling.top_p"                },
+    { LLM_KV_GENERAL_SAMPLING_MIN_P,           "general.sampling.min_p"                },
+    { LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY, "general.sampling.xtc_probability"      },
+    { LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD,   "general.sampling.xtc_threshold"        },
+    { LLM_KV_GENERAL_SAMPLING_TEMP,            "general.sampling.temp"                 },
+    { LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N,  "general.sampling.penalty_last_n"       },
+    { LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT,  "general.sampling.penalty_repeat"       },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT,        "general.sampling.mirostat"             },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU,    "general.sampling.mirostat_tau"         },
+    { LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA,    "general.sampling.mirostat_eta"         },
+    { LLM_KV_GENERAL_NAME,                     "general.name"                          },
+    { LLM_KV_GENERAL_AUTHOR,                   "general.author"                        },
+    { LLM_KV_GENERAL_VERSION,                  "general.version"                       },
+    { LLM_KV_GENERAL_URL,                      "general.url"                           },
+    { LLM_KV_GENERAL_DESCRIPTION,              "general.description"                   },
+    { LLM_KV_GENERAL_LICENSE,                  "general.license"                       },
+    { LLM_KV_GENERAL_SOURCE_URL,               "general.source.url"                    },
+    { LLM_KV_GENERAL_SOURCE_HF_REPO,           "general.source.huggingface.repository" },
 
     { LLM_KV_VOCAB_SIZE,                        "%s.vocab_size"                        },
     { LLM_KV_CONTEXT_LENGTH,                    "%s.context_length"                    },
@@ -146,6 +168,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERTS_PER_GROUP,                 "%s.experts_per_group"                 },
     { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
     { LLM_KV_NEXTN_PREDICT_LAYERS,              "%s.nextn_predict_layers"              },
+    { LLM_KV_NUM_DEEPSTACK_LAYERS,              "%s.n_deepstack_layers"                },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
@@ -184,6 +207,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
     { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
+    { LLM_KV_ATTENTION_TEMPERATURE_SCALE,            "%s.attention.temperature_scale"            },
     { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
@@ -329,6 +353,36 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_AFMOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_GATE,       "blk.%d.attn_gate" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,  "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
+        },
+    },
     {
         LLM_ARCH_LLAMA4,
         {
@@ -781,6 +835,77 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_QWEN3NEXT,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_POST_NORM,     "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_SSM_A_NOSCAN,       "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_CONV1D,         "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,             "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_BETA_ALPHA,     "blk.%d.ssm_ba" },
+            { LLM_TENSOR_SSM_IN,             "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_NORM,           "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,            "blk.%d.ssm_out" },
+        },
+    },
+    {
+        LLM_ARCH_QWEN3VL,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_QWEN3VLMOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_PHI2,
         {
@@ -1641,6 +1766,39 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_NEMOTRON_H_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,     "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,    "output_norm" },
+            { LLM_TENSOR_OUTPUT,         "output" },
+            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm" },
+            // mamba(2) ssm layers
+            { LLM_TENSOR_SSM_IN,         "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,     "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,         "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,          "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,          "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,       "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,        "blk.%d.ssm_out" },
+            // attention layers
+            { LLM_TENSOR_ATTN_Q,         "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,         "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,         "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,       "blk.%d.attn_output" },
+            // dense FFN
+            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
+            // MoE FFN (for MoE layers)
+            { LLM_TENSOR_FFN_GATE_INP,   "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_UP_EXPS,    "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,  "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,"blk.%d.exp_probs_b" },
+            // MoE shared expert layer
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
+        },
+    },
     {
         LLM_ARCH_EXAONE,
         {
@@ -2168,7 +2326,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SHORTCONV_INPROJ,  "blk.%d.shortconv.in_proj" },
             { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
             { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
-            { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,       "token_embd_norm" }, // note: wrong tensor name
             { LLM_TENSOR_OUTPUT,            "output" },
         }
     },
@@ -2190,7 +2348,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SHORTCONV_INPROJ,  "blk.%d.shortconv.in_proj" },
             { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
             { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
-            { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,       "token_embd_norm" }, // note: wrong tensor name
             { LLM_TENSOR_FFN_GATE_INP,      "blk.%d.ffn_gate_inp" },
             { LLM_TENSOR_FFN_GATE_EXPS,     "blk.%d.ffn_gate_exps" },
             { LLM_TENSOR_FFN_DOWN_EXPS,     "blk.%d.ffn_down_exps" },
@@ -2332,6 +2490,110 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_CHEXPS,      "blk.%d.ffn_up_chexps" },
         },
     },
+    {
+        LLM_ARCH_MINIMAX_M2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+        },
+    },
+    {
+        LLM_ARCH_PANGU_EMBED,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_COGVLM,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_VISEXP_ATTN_QKV, "blk.%d.vis_attn_qkv" },
+            { LLM_TENSOR_VISEXP_ATTN_OUT, "blk.%d.vis_attn_output" },
+            { LLM_TENSOR_VISEXP_FFN_GATE, "blk.%d.vis_gate" },
+            { LLM_TENSOR_VISEXP_FFN_DOWN, "blk.%d.vis_down" },
+            { LLM_TENSOR_VISEXP_FFN_UP,   "blk.%d.vis_up" },
+        },
+    },
+    {
+        LLM_ARCH_RND1,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
+    {
+        LLM_ARCH_MISTRAL3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
+            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
+            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2340,11 +2602,21 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
     },
 };
 
+// declare information about the model weight tensors:
+// - the layer in which the tensor is going to be used. this is needed in order to assign the correct buffer type for the weight
+// - the operator which is going to use the weight. this is needed to determine if the respective backend supports the operator
+//
+// for example, input layers are usually assigned to CPU/host buffer types
+//
+// a mismatch between the declared information and the actual layer/op in which the tensor is used can lead to sub-optimal
+//   assignment of the buffer types and extra overhead during computation
+// example: https://github.com/ggml-org/llama.cpp/pull/17548
+//
 static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_TOKEN_EMBD,                 {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_POS_EMBD,                   {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
-    {LLM_TENSOR_TOKEN_EMBD_NORM,            {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_TOKEN_TYPES,                {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_TOKEN_EMBD_NORM,            {LLM_TENSOR_LAYER_INPUT, GGML_OP_MUL}},
     {LLM_TENSOR_OUTPUT,                     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS,                        {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS_OUT,                    {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
@@ -2361,6 +2633,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_ATTN_V,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_QKV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_OUT,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_GATE,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_FFN_GATE,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_FFN_DOWN,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_FFN_UP,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -2398,6 +2671,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SSM_X,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SSM_DT,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SSM_OUT,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_SSM_BETA_ALPHA,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_A1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -2419,6 +2693,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_ACT,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_DIV}},
     {LLM_TENSOR_SSM_CONV1D,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SSM_A,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_SCAN}},
+    {LLM_TENSOR_SSM_A_NOSCAN,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, // a version of SSM_A used for MUL instead of SSM_SCAN
     {LLM_TENSOR_SSM_DT_NORM,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_SSM_B_NORM,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_SSM_C_NORM,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@@ -2509,6 +2784,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SHORTCONV_CONV,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SHORTCONV_INPROJ,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SHORTCONV_OUTPROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_ATTN_QKV,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_ATTN_OUT,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_GATE,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_DOWN,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_UP,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     // NextN/MTP tensors are currently ignored (reserved for future MTP support)
     // These tensors only exist in the last layer(s) and are treated as output tensors
     {LLM_TENSOR_NEXTN_EH_PROJ,              {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
@@ -2592,6 +2872,8 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
+        case LLM_ARCH_QWEN3NEXT:
             return true;
         default:
             return false;
@@ -2603,6 +2885,7 @@ bool llm_arch_is_diffusion(const llm_arch & arch) {
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
         case LLM_ARCH_LLADA_MOE:
+        case LLM_ARCH_RND1:
             return true;
         default:
             return false;

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

@@ -36,6 +36,9 @@ enum llm_arch {
     LLM_ARCH_QWEN2VL,
     LLM_ARCH_QWEN3,
     LLM_ARCH_QWEN3MOE,
+    LLM_ARCH_QWEN3NEXT,
+    LLM_ARCH_QWEN3VL,
+    LLM_ARCH_QWEN3VLMOE,
     LLM_ARCH_PHI2,
     LLM_ARCH_PHI3,
     LLM_ARCH_PHIMOE,
@@ -76,6 +79,7 @@ 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,
@@ -93,6 +97,7 @@ enum llm_arch {
     LLM_ARCH_BAILINGMOE2,
     LLM_ARCH_DOTS1,
     LLM_ARCH_ARCEE,
+    LLM_ARCH_AFMOE,
     LLM_ARCH_ERNIE4_5,
     LLM_ARCH_ERNIE4_5_MOE,
     LLM_ARCH_HUNYUAN_MOE,
@@ -108,6 +113,11 @@ enum llm_arch {
     LLM_ARCH_SEED_OSS,
     LLM_ARCH_GROVEMOE,
     LLM_ARCH_APERTUS,
+    LLM_ARCH_MINIMAX_M2,
+    LLM_ARCH_COGVLM,
+    LLM_ARCH_RND1,
+    LLM_ARCH_PANGU_EMBED,
+    LLM_ARCH_MISTRAL3,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -117,6 +127,18 @@ enum llm_kv {
     LLM_KV_GENERAL_QUANTIZATION_VERSION,
     LLM_KV_GENERAL_ALIGNMENT,
     LLM_KV_GENERAL_FILE_TYPE,
+    LLM_KV_GENERAL_SAMPLING_SEQUENCE,
+    LLM_KV_GENERAL_SAMPLING_TOP_K,
+    LLM_KV_GENERAL_SAMPLING_TOP_P,
+    LLM_KV_GENERAL_SAMPLING_MIN_P,
+    LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY,
+    LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD,
+    LLM_KV_GENERAL_SAMPLING_TEMP,
+    LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N,
+    LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU,
+    LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA,
     LLM_KV_GENERAL_NAME,
     LLM_KV_GENERAL_AUTHOR,
     LLM_KV_GENERAL_VERSION,
@@ -150,6 +172,7 @@ enum llm_kv {
     LLM_KV_EXPERTS_PER_GROUP,
     LLM_KV_MOE_EVERY_N_LAYERS,
     LLM_KV_NEXTN_PREDICT_LAYERS,
+    LLM_KV_NUM_DEEPSTACK_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
@@ -188,6 +211,7 @@ 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,
@@ -308,6 +332,7 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_POST_NORM,
     LLM_TENSOR_ATTN_ROT_EMBD,
     LLM_TENSOR_ATTN_SINKS,
+    LLM_TENSOR_ATTN_GATE,
     LLM_TENSOR_FFN_GATE_INP,
     LLM_TENSOR_FFN_GATE_INP_SHEXP,
     LLM_TENSOR_FFN_NORM,
@@ -357,11 +382,13 @@ 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,
     LLM_TENSOR_SSM_NORM,
     LLM_TENSOR_SSM_OUT,
+    LLM_TENSOR_SSM_BETA_ALPHA,      // qwen3next
     LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
     LLM_TENSOR_TIME_MIX_W2,
@@ -458,6 +485,11 @@ enum llm_tensor {
     LLM_TENSOR_SHORTCONV_CONV,
     LLM_TENSOR_SHORTCONV_INPROJ,
     LLM_TENSOR_SHORTCONV_OUTPROJ,
+    LLM_TENSOR_VISEXP_ATTN_QKV,
+    LLM_TENSOR_VISEXP_ATTN_OUT,
+    LLM_TENSOR_VISEXP_FFN_GATE,
+    LLM_TENSOR_VISEXP_FFN_DOWN,
+    LLM_TENSOR_VISEXP_FFN_UP,
     LLM_TENSOR_NEXTN_EH_PROJ,
     LLM_TENSOR_NEXTN_EMBED_TOKENS,
     LLM_TENSOR_NEXTN_ENORM,

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

@@ -215,6 +215,7 @@ bool llama_batch_allocr::init(
             /*.n_seq_tokens =*/ (uint32_t) 1,
             /*.n_seqs       =*/ (uint32_t) batch.n_tokens,
             /*.n_seqs_unq   =*/ (uint32_t) this->seq_id_unq.size(),
+            /*.n_pos        =*/ n_pos_per_embd,
             /*.token        =*/ batch.token,
             /*.embd         =*/ batch.embd,
             /*.pos          =*/ batch.pos,
@@ -251,46 +252,72 @@ bool llama_batch_allocr::init(
     // consistency checks
     //
 
-    for (uint32_t s = 0; s < n_seq_max; ++s) {
-        if (seq_pos[s].empty()) {
-            continue;
-        }
+    if (n_pos_per_embd > 1) {
+        // M-RoPE case: allow position to "jump" forward only (non-continuous positions are allowed)
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
+            if (seq_pos[s].empty()) {
+                continue;
+            }
 
-        const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
-
-        if (p0 >= 0) {
-            bool ok = true;
+            const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
 
             if (batch.token) {
+                if (p0 >= 0 && p0 >= seq_pos_min(s)) {
+                    LLAMA_LOG_ERROR(
+                            "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                            " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                            " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                            " for M-RoPE, it is required that the position satisfies: X < Y\n",
+                            __func__, s, s, p0, s, seq_pos_min(s));
+
+                    return false;
+                }
+            } else {
+                // embedding inputs can have overlapping positions
+                if (p0 >= 0 && p0 > seq_pos_min(s)) {
+                    LLAMA_LOG_ERROR(
+                            "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                            " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                            " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                            " for M-RoPE, it is required that the position satisfies: X <= Y\n",
+                            __func__, s, s, p0, s, seq_pos_min(s));
+
+                    return false;
+                }
+            }
+        }
+    } else {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
+            if (seq_pos[s].empty()) {
+                continue;
+            }
+
+            const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
+
+            if (p0 >= 0) {
+                bool ok = true;
+
                 if (seq_pos_min(s) != p0 + 1) {
                     ok = false;
                 }
-            } else {
-                assert(batch.embd);
 
-                // for embeddings (typically used as vision input), we allow them to have repeating positions
-                // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762
-                if (seq_pos_min(s) != p0 && seq_pos_min(s) != p0 + 1) {
-                    ok = false;
+                if (!ok) {
+                    LLAMA_LOG_ERROR(
+                            "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                            " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                            " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                            " it is required that the sequence positions remain consecutive: Y = X + 1\n",
+                            __func__, s, s, p0, s, seq_pos_min(s));
+
+                    return false;
                 }
             }
 
-            if (!ok) {
-                LLAMA_LOG_ERROR(
-                        "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
-                        " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
-                        " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
-                        " it is required that the sequence positions remain consecutive: Y = X + 1\n",
-                        __func__, s, s, p0, s, seq_pos_min(s));
-
+            if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) {
+                LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s);
                 return false;
             }
         }
-
-        if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) {
-            LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s);
-            return false;
-        }
     }
 
     if (memory) {
@@ -389,6 +416,7 @@ llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t
         /*.n_seq_tokens =*/ n_seq_tokens,
         /*.n_seqs       =*/ n_seqs,
         /*.n_seqs_unq   =*/ n_seqs,
+        /*.n_pos        =*/ n_pos_per_embd,
 
         /*.token        =*/ udata->token.data(),
         /*.embd         =*/ nullptr,
@@ -655,10 +683,8 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
 
     auto udata = std::make_shared<llama_ubatch::data_t>();
 
-    const int32_t n_pos_cur = batch.embd ? n_pos_per_embd : 1;
-
     const int64_t n_embd_all = batch.embd ? (int64_t) n_tokens*n_embd : 0;
-    const int64_t n_pos_all  =              (int64_t) n_tokens*n_pos_cur;
+    const int64_t n_pos_all  =              (int64_t) n_tokens*n_pos_per_embd;
 
     udata->token     .resize(n_tokens);
     udata->embd      .resize(n_embd_all);
@@ -680,8 +706,13 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
             memcpy(udata->embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float));
         }
 
-        for (int j = 0; j < n_pos_cur; ++j) {
-            udata->pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]];
+        for (size_t j = 0; j < (size_t)n_pos_per_embd; ++j) {
+            // if we are using M-RoPE
+            //     if the current batch is text, we need to broadcast the same position across all RoPE sections
+            //     otherwise, the input batch is image embeddings, we copy the positions as-is
+            // if we are not using M-RoPE, there is only one position per token (this loop runs only once)
+            size_t src_off = batch.token ? 0 : j*batch.n_tokens;
+            udata->pos[j*n_tokens + i] = batch.pos[src_off + idxs[i]];
         }
 
         udata->n_seq_id[i] = batch.n_seq_id[idxs[i]];
@@ -710,6 +741,7 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
         /*.n_seq_tokens =*/ n_tokens/n_seqs,
         /*.n_seqs       =*/ n_seqs,
         /*.n_seqs_unq   =*/ (uint32_t) udata->seq_id_unq.size(),
+        /*.n_pos        =*/ n_pos_per_embd,
 
         /*.token        =*/ batch.token ? udata->token.data() : nullptr,
         /*.embd         =*/ batch.embd ? udata->embd.data() : nullptr,

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

@@ -17,6 +17,16 @@ struct llama_ubatch {
         return b_equal_seqs != 0;
     }
 
+    // typical for M-RoPE cases:
+    //   0 - sequantial position of the tokens/embeddings in the sequence
+    //   1 - y position in the image
+    //   2 - x position in the image
+    //   3 - other
+    bool is_pos_2d() const {
+        // TODO @ngxson : we may need to check for model arch when more models use >1 positions
+        return n_pos >= 3;
+    }
+
     uint32_t b_equal_seqs; // note: this is a boolean, but we use an int32_t for alignment
                            //       otherwise address sanitizer complains
     // TODO: whole_seqs for embeddings?
@@ -25,6 +35,7 @@ struct llama_ubatch {
     uint32_t n_seq_tokens; // tokens per sequence set
     uint32_t n_seqs;       // sequence sets in the ubatch
     uint32_t n_seqs_unq;   // unique sequence ids in the ubatch
+    uint32_t n_pos;        // number of position inputs for each token/embedding
 
     // seq_id_unq: unique sequence ids in the ubatch
     // seq_idx:    indices of the unique sequence ids in the ubatch in [0, n_seqs_unq)
@@ -33,7 +44,7 @@ struct llama_ubatch {
     //                          // size               | idx | val
     llama_token  *  token;      // [n_tokens]         | i   | id, token
     float        *  embd;       // [n_embd, n_tokens] | i   | embd
-    llama_pos    *  pos;        // [n_tokens]         | i   | pos
+    llama_pos    *  pos;        // [n_tokens*n_pos]   | i   | pos
     int32_t      *  n_seq_id;   // [n_tokens]         | i   | -
     llama_seq_id ** seq_id;     // [n_tokens]         | s   | s0, s1, seq_id
     llama_seq_id *  seq_id_unq; // [n_seqs_unq]       | s   | seq_id

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

@@ -73,6 +73,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
     { "seed_oss",          LLM_CHAT_TEMPLATE_SEED_OSS          },
     { "grok-2",            LLM_CHAT_TEMPLATE_GROK_2            },
+    { "pangu-embedded",    LLM_CHAT_TEMPLATE_PANGU_EMBED       },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -213,6 +214,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_SEED_OSS;
     } else if (tmpl_contains("'Assistant: '  + message['content'] + '<|separator|>")) {
         return LLM_CHAT_TEMPLATE_GROK_2;
+    } else if (tmpl_contains(LU8("[unused9]系统：[unused10]"))) {
+        return LLM_CHAT_TEMPLATE_PANGU_EMBED;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -813,6 +816,35 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "Assistant:";
         }
+    }else if (tmpl == LLM_CHAT_TEMPLATE_PANGU_EMBED) {
+        // [unused9]系统：xxx[unused10]
+        // [unused9]用户：xxx[unused10]
+        // [unused9]助手：xxx[unused10]
+        // ...
+        for (size_t i = 0; i < chat.size(); ++i) {
+            const auto & msg = chat[i];
+            const std::string & role = msg->role;
+            const std::string & content = msg->content;
+
+            if (i == 0 && role != "system") {
+                ss << "[unused9]系统：[unused10]";
+            }
+
+            if (role == "system") {
+                ss << "[unused9]系统：" << content << "[unused10]";
+            } else if (role == "user") {
+                ss << "[unused9]用户：" << content << "[unused10]";
+            } else if (role == "assistant") {
+                ss << "[unused9]助手：" << content << "[unused10]";
+            } else if (role == "tool") {
+                ss << "[unused9]工具：" << content << "[unused10]";
+            } else if (role == "function") {
+                ss << "[unused9]方法：" << content << "[unused10]";
+            }
+        }
+        if (add_ass) {
+            ss << "[unused9]助手：";
+        }
     } else {
         // template not supported
         return -1;

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

@@ -53,6 +53,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_KIMI_K2,
     LLM_CHAT_TEMPLATE_SEED_OSS,
     LLM_CHAT_TEMPLATE_GROK_2,
+    LLM_CHAT_TEMPLATE_PANGU_EMBED,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

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

@@ -1,5 +1,6 @@
 #include "llama-context.h"
 
+#include "llama-arch.h"
 #include "llama-impl.h"
 #include "llama-batch.h"
 #include "llama-io.h"
@@ -21,6 +22,8 @@ llama_context::llama_context(
               llama_context_params params) :
     model(model),
     balloc(std::make_unique<llama_batch_allocr>(model.hparams.n_pos_per_embd())) {
+    // TODO warning when creating llama_context with awkward ctx size that is not a power of 2,
+    //     may need to be backend-dependent
     LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);
 
     t_start_us = model.t_start_us;
@@ -112,11 +115,28 @@ llama_context::llama_context(
         }
     }
 
-    const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
+    // ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732
+    cparams.n_ctx = GGML_PAD(cparams.n_ctx, 256);
+
+    if (cparams.kv_unified) {
+        cparams.n_ctx_seq = cparams.n_ctx;
+    } else {
+        cparams.n_ctx_seq = cparams.n_ctx / cparams.n_seq_max;
+        cparams.n_ctx_seq = GGML_PAD(cparams.n_ctx_seq, 256);
+
+        if (cparams.n_ctx_seq == 0) {
+            throw std::runtime_error("n_ctx_seq == 0");
+        }
+
+        if (cparams.n_ctx != cparams.n_ctx_seq * cparams.n_seq_max) {
+            cparams.n_ctx =  cparams.n_ctx_seq * cparams.n_seq_max;
+            LLAMA_LOG_WARN("%s: n_ctx is not divisible by n_seq_max - rounding down to %u\n", __func__, cparams.n_ctx);
+        }
+    }
 
     LLAMA_LOG_INFO("%s: n_seq_max     = %u\n",   __func__, cparams.n_seq_max);
     LLAMA_LOG_INFO("%s: n_ctx         = %u\n",   __func__, cparams.n_ctx);
-    LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n",   __func__, n_ctx_per_seq);
+    LLAMA_LOG_INFO("%s: n_ctx_seq     = %u\n",   __func__, cparams.n_ctx_seq);
     LLAMA_LOG_INFO("%s: n_batch       = %u\n",   __func__, cparams.n_batch);
     LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
     LLAMA_LOG_INFO("%s: causal_attn   = %d\n",   __func__, cparams.causal_attn);
@@ -125,14 +145,14 @@ llama_context::llama_context(
     LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
     LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);
 
-    if (n_ctx_per_seq < hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    if (cparams.n_ctx_seq < hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
+                __func__, cparams.n_ctx_seq, hparams.n_ctx_train);
     }
 
-    if (n_ctx_per_seq > hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    if (cparams.n_ctx_seq > hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
+                __func__, cparams.n_ctx_seq, hparams.n_ctx_train);
     }
 
     if (!hparams.vocab_only) {
@@ -228,7 +248,10 @@ llama_context::llama_context(
 
         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);
 
@@ -268,9 +291,7 @@ llama_context::llama_context(
         if (pipeline_parallel) {
             LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get()));
         }
-    }
 
-    if (!hparams.vocab_only) {
         llama_memory_context_ptr mctx;
         if (memory) {
             LLAMA_LOG_DEBUG("%s: reserving full memory module\n", __func__);
@@ -282,9 +303,6 @@ llama_context::llama_context(
 
         cross.v_embd.clear();
 
-        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
-        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
-
         // avoid reserving graphs with zero outputs - assume one output per sequence
         n_outputs = n_seqs;
 
@@ -343,7 +361,14 @@ llama_context::llama_context(
         {
             auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
             if (!gf) {
-                throw std::runtime_error("failed to allocate compute pp buffers");
+                if (pipeline_parallel) {
+                    LLAMA_LOG_WARN("%s: compute buffer allocation failed, retrying without pipeline parallelism\n", __func__);
+                    sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, false, cparams.op_offload));
+                    gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+                }
+                if (!gf) {
+                    throw std::runtime_error("failed to allocate compute pp buffers");
+                }
             }
 
             n_splits_pp = ggml_backend_sched_get_n_splits(sched.get());
@@ -448,8 +473,8 @@ uint32_t llama_context::n_ctx() const {
     return cparams.n_ctx;
 }
 
-uint32_t llama_context::n_ctx_per_seq() const {
-    return cparams.n_ctx / cparams.n_seq_max;
+uint32_t llama_context::n_ctx_seq() const {
+    return cparams.n_ctx_seq;
 }
 
 uint32_t llama_context::n_batch() const {
@@ -518,7 +543,7 @@ bool llama_context::memory_update(bool optimize) {
             throw std::runtime_error("failed to initialize memory context");
         }
 
-        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
@@ -803,7 +828,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     const auto & hparams = model.hparams;
 
-    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd_inp();
     const int64_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
@@ -972,7 +997,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     const auto & hparams = model.hparams;
 
     const int64_t n_vocab = vocab.n_tokens();
-    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd_inp();
 
     const bool output_all = false;
 
@@ -1223,7 +1248,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
         // make the outputs have the same order they had in the user-provided batch
         // note: this is mostly relevant for recurrent models atm
-        if (!sorted_output) {
+        if (!sorted_output && n_outputs > 1) {
             GGML_ASSERT((size_t) n_outputs == out_ids.size());
 
             // TODO: is there something more efficient which also minimizes swaps?
@@ -1360,7 +1385,10 @@ 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>(n_tokens * 40, 32u * model.n_tensors());
+    }
     return std::max<uint32_t>(1024u, 8u*model.n_tensors());
 }
 
@@ -2129,7 +2157,7 @@ void llama_context::opt_epoch_iter(
             batch.logits  [pos_batch]    = true;
         }
 
-        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
+        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd_inp(), cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
             LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
             return;
         }
@@ -2377,6 +2405,10 @@ uint32_t llama_n_ctx(const llama_context * ctx) {
     return ctx->n_ctx();
 }
 
+uint32_t llama_n_ctx_seq(const llama_context * ctx) {
+    return ctx->n_ctx_seq();
+}
+
 uint32_t llama_n_batch(const llama_context * ctx) {
     return ctx->n_batch();
 }

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

@@ -43,11 +43,11 @@ struct llama_context {
 
     ggml_backend_sched_t get_sched() const;
 
-    uint32_t n_ctx()         const;
-    uint32_t n_ctx_per_seq() const;
-    uint32_t n_batch()       const;
-    uint32_t n_ubatch()      const;
-    uint32_t n_seq_max()     const;
+    uint32_t n_ctx()     const;
+    uint32_t n_ctx_seq() const;
+    uint32_t n_batch()   const;
+    uint32_t n_ubatch()  const;
+    uint32_t n_seq_max() const;
 
     uint32_t n_threads()       const;
     uint32_t n_threads_batch() const;
@@ -197,7 +197,7 @@ 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;

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

@@ -8,6 +8,7 @@
 
 struct llama_cparams {
     uint32_t n_ctx;           // context size used during inference
+    uint32_t n_ctx_seq;       // context for a single sequence
     uint32_t n_batch;
     uint32_t n_ubatch;
     uint32_t n_seq_max;

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

@@ -6,8 +6,10 @@
 
 #include <cmath>
 #include <algorithm>
+#include <cstdint>
 #include <stdexcept>
 
+#define MAX_REPETITION_THRESHOLD 2000
 //
 // helpers
 //
@@ -179,6 +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));
@@ -210,6 +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:
@@ -226,6 +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");
@@ -282,6 +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) {
@@ -345,8 +417,10 @@ const char * llama_grammar_parser::parse_sequence(
     size_t last_sym_start = rule.size();
     const char * pos = src;
 
-    auto handle_repetitions = [&](int min_times, int max_times) {
-
+    // use UINT64_MAX as the empty value because we aligned to the proper uint64_t type so -1 can't be used
+    // (though it's technically the same as -1 now)
+    auto handle_repetitions = [&](uint64_t min_times, uint64_t max_times) {
+        bool no_max = max_times == UINT64_MAX;
         if (last_sym_start == rule.size()) {
             throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
         }
@@ -373,20 +447,20 @@ const char * llama_grammar_parser::parse_sequence(
             rule.resize(last_sym_start);
         } else {
             // Repeat the previous elements (min_times - 1) times
-            for (int i = 1; i < min_times; i++) {
+            for (uint64_t i = 1; i < min_times; i++) {
                 rule.insert(rule.end(), prev_rule.begin(), prev_rule.end());
             }
         }
 
         uint32_t last_rec_rule_id = 0;
-        auto n_opt = max_times < 0 ? 1 : max_times - min_times;
+        auto n_opt = no_max ? 1 : max_times - min_times;
 
         llama_grammar_rule rec_rule(prev_rule);
-        for (int i = 0; i < n_opt; i++) {
+        for (uint64_t i = 0; i < n_opt; i++) {
             rec_rule.resize(prev_rule.size());
             uint32_t rec_rule_id = generate_symbol_id( rule_name);
-            if (i > 0 || max_times < 0) {
-                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
+            if (i > 0 || no_max) {
+                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, no_max ? rec_rule_id : last_rec_rule_id});
             }
             rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
             rec_rule.push_back({LLAMA_GRETYPE_END, 0});
@@ -440,6 +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);
@@ -478,10 +563,10 @@ const char * llama_grammar_parser::parse_sequence(
                 throw std::runtime_error(std::string("expecting an int at ") + pos);
             }
             const char * int_end = parse_int(pos);
-            int min_times = std::stoul(std::string(pos, int_end - pos));
+            uint64_t min_times = std::stoul(std::string(pos, int_end - pos));
             pos = parse_space(int_end, is_nested);
 
-            int max_times = -1;
+            uint64_t max_times = UINT64_MAX; // default: no max limit
 
             if (*pos == '}') {
                 max_times = min_times;
@@ -502,6 +587,10 @@ const char * llama_grammar_parser::parse_sequence(
             } else {
                 throw std::runtime_error(std::string("expecting ',' at ") + pos);
             }
+            bool has_max = max_times != UINT64_MAX;
+            if (min_times > MAX_REPETITION_THRESHOLD || (has_max && max_times > MAX_REPETITION_THRESHOLD)) {
+                throw std::runtime_error(std::string("number of repetitions exceeds sane defaults, please reduce the number of repetitions"));
+            }
             handle_repetitions(min_times, max_times);
         } else {
             break;
@@ -683,6 +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(
@@ -730,6 +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);
@@ -823,26 +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()) {
-            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);
-        }
+        llama_grammar_accept_chr(*grammar, stack, chr, stacks_new);
     }
 
     grammar->stacks = std::move(stacks_new);
@@ -867,6 +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());
 
@@ -879,7 +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);
         }
@@ -897,7 +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;
@@ -966,12 +1100,13 @@ struct llama_grammar * llama_grammar_init_impl(
         ollama_vocab,
         std::move(vec_rules),
         std::move(stacks),
-        /* .partial_utf8 = */     {},
-        /* .lazy =*/              false,
-        /* .awaiting_trigger = */ false,
-        /* .trigger_buffer = */   "",
-        /* .trigger_tokens   = */ {},
-        /* .trigger_patterns    = */ {},
+        /* .partial_utf8 = */             {},
+        /* .lazy = */                     false,
+        /* .awaiting_trigger = */         false,
+        /* .trigger_buffer = */           "",
+        /* .trigger_buffer_positions = */ {},
+        /* .trigger_tokens = */           {},
+        /* .trigger_patterns = */         {},
     };
 }
 
@@ -985,7 +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
@@ -1073,10 +1208,11 @@ struct llama_grammar * llama_grammar_init_impl(
         ollama_vocab,
         std::move(vec_rules),
         std::move(stacks),
-        /* .partial_utf8 = */     {},
-        /* .lazy = */             lazy,
-        /* .awaiting_trigger = */ lazy,
-        /* .trigger_buffer = */   "",
+        /* .partial_utf8 = */             {},
+        /* .lazy = */                     lazy,
+        /* .awaiting_trigger = */         lazy,
+        /* .trigger_buffer = */           "",
+        /* .trigger_buffer_positions = */ {},
         std::move(vec_trigger_tokens),
         std::move(vec_trigger_patterns),
     };
@@ -1100,6 +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,
     };
@@ -1156,7 +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 });
         }
     }
 
@@ -1176,10 +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;
@@ -1197,10 +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;
                 }
@@ -1220,7 +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) {
@@ -1238,6 +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

@@ -71,6 +71,9 @@ 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];
@@ -958,25 +961,25 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         // organize experts into n_expert_groups
         ggml_tensor * selection_groups = ggml_reshape_3d(ctx0, selection_probs, n_exp_per_group, hparams.n_expert_groups, n_tokens); // [n_exp_per_group, n_expert_groups, n_tokens]
 
-        ggml_tensor * group_scores = ggml_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
+        ggml_tensor * group_scores = ggml_argsort_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
         group_scores = ggml_get_rows(ctx0, ggml_reshape_4d(ctx0, selection_groups, 1, selection_groups->ne[0], selection_groups->ne[1], selection_groups->ne[2]), group_scores); // [1, 2, n_expert_groups, n_tokens]
 
         // get top n_group_used expert groups
         group_scores = ggml_sum_rows(ctx0, ggml_reshape_3d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2], group_scores->ne[3])); // [1, n_expert_groups, n_tokens]
         group_scores = ggml_reshape_2d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2]); // [n_expert_groups, n_tokens]
 
-        ggml_tensor * expert_groups = ggml_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
+        ggml_tensor * expert_groups = ggml_argsort_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
         cb(expert_groups, "ffn_moe_group_topk", il);
 
         // 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);
     }
 
     // select experts
-    ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
+    ggml_tensor * selected_experts = ggml_argsort_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
     cb(selected_experts->src[0], "ffn_moe_argsort", il);
     cb(selected_experts, "ffn_moe_topk", il);
 
@@ -1006,10 +1009,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens]
         cb(weights_sum, "ffn_moe_weights_sum", il);
 
-        if (arch == LLM_ARCH_BAILINGMOE2) {
-            weights_sum = ggml_scale_bias(ctx0, weights_sum, 1.0, 1e-20);
-            cb(weights_sum, "ffn_moe_weights_sum_biased", il);
-        }
+        // Avoid division by zero, clamp to smallest number representable by F16
+        weights_sum = ggml_clamp(ctx0, weights_sum, 6.103515625e-5, INFINITY);
+        cb(weights_sum, "ffn_moe_weights_sum_clamped", il);
 
         weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens]
         cb(weights, "ffn_moe_weights_norm", il);
@@ -1087,6 +1089,16 @@ 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");
     }
@@ -1137,7 +1149,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
 // input embeddings with optional lora
 ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
-    const int64_t n_embd = hparams.n_embd;
+    const int64_t n_embd = hparams.n_embd_inp();
 
     auto inp = std::make_unique<llm_graph_input_embd>();
 
@@ -1274,7 +1286,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
     //    return cur;
     //}
 
-    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd;
+    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp();
     const auto n_enc  = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
 
     cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);
@@ -1587,9 +1599,10 @@ ggml_tensor * llm_graph_context::build_attn(
             int       il) const {
     // these nodes are added to the graph together so that they are not reordered
     // by doing so, the number of splits in the graph is reduced
+    // expand k later to enable rope fusion which directly writes into k-v cache
     ggml_build_forward_expand(gf, q_cur);
-    ggml_build_forward_expand(gf, k_cur);
     ggml_build_forward_expand(gf, v_cur);
+    ggml_build_forward_expand(gf, k_cur);
 
     const auto * mctx_cur = inp->mctx;
 
@@ -2030,7 +2043,7 @@ int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buck
 
     if (bidirectional) {
         relative_bucket += (relative_position > 0) * n_buckets;
-        relative_position = abs(relative_position);
+        relative_position = std::abs(relative_position);
     } else {
         relative_position = -std::min<int32_t>(relative_position, 0);
     }

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

@@ -60,6 +60,16 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const {
     return n_head/n_head_kv;
 }
 
+uint32_t llama_hparams::n_embd_inp() const {
+    uint32_t n_embd_inp = n_embd;
+
+    if (n_deepstack_layers > 0) {
+        n_embd_inp += n_embd * n_deepstack_layers;
+    }
+
+    return n_embd_inp;
+}
+
 uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 
@@ -148,7 +158,7 @@ bool llama_hparams::is_recurrent(uint32_t il) const {
 }
 
 uint32_t llama_hparams::n_pos_per_embd() const {
-    return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
+    return rope_type == LLAMA_ROPE_TYPE_MROPE || rope_type == LLAMA_ROPE_TYPE_IMROPE ? 4 : 1;
 }
 
 bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {

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

@@ -6,7 +6,7 @@
 
 // bump if necessary
 #define LLAMA_MAX_LAYERS  512
-#define LLAMA_MAX_EXPERTS 384  // Kimi-K2
+#define LLAMA_MAX_EXPERTS 512 // Qwen3 Next
 
 enum llama_expert_gating_func_type {
     LLAMA_EXPERT_GATING_FUNC_TYPE_NONE           = 0,
@@ -164,8 +164,8 @@ 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;
-    float    f_attn_temp_scale       = 0.1;
+    uint32_t n_attn_temp_floor_scale = 0;
+    float    f_attn_temp_scale       = 0.0f;
 
     // gemma3n altup
     uint32_t n_altup      = 4; // altup_num_inputs
@@ -185,6 +185,9 @@ struct llama_hparams {
     std::array<float, LLAMA_MAX_LAYERS> xielu_beta;
     std::array<float, LLAMA_MAX_LAYERS> xielu_eps;
 
+    // qwen3vl deepstack
+    uint32_t n_deepstack_layers = 0;
+
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
@@ -226,6 +229,9 @@ struct llama_hparams {
 
     uint32_t n_gqa(uint32_t il = 0) const;
 
+    // dimension of main + auxiliary input embeddings
+    uint32_t n_embd_inp() const;
+
     // dimension of key embeddings across all k-v heads
     uint32_t n_embd_k_gqa(uint32_t il = 0) const;
 

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

@@ -20,10 +20,10 @@ static llama_logger_state g_logger_state;
 time_meas::time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
 
 time_meas::~time_meas() {
-        if (t_start_us >= 0) {
-            t_acc += ggml_time_us() - t_start_us;
-        }
+    if (t_start_us >= 0) {
+        t_acc += ggml_time_us() - t_start_us;
     }
+}
 
 void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     ggml_log_set(log_callback, user_data);

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-iswa.cpp

@@ -45,7 +45,9 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
 
     const uint32_t size_base = kv_size;
 
-    uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch, n_pad));
+    // note: the SWA cache is always padded to 256 for performance
+    //       https://github.com/ggml-org/llama.cpp/issues/17037
+    uint32_t size_swa = GGML_PAD(std::min(size_base, hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch), 256);
 
     // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size
     if (swa_full) {

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

@@ -8,6 +8,7 @@
 #include <algorithm>
 #include <cassert>
 #include <cmath>
+#include <cstring>
 #include <limits>
 #include <map>
 #include <stdexcept>
@@ -37,8 +38,15 @@ llama_kv_cache::llama_kv_cache(
 
     const uint32_t n_layer_kv = hparams.n_layer_kv();
 
+    // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
+    struct ggml_backend_buft_comparator {
+        bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
+            return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
+        }
+    };
+    std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
+
     // create a context for each buffer type
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
@@ -53,13 +61,12 @@ llama_kv_cache::llama_kv_cache(
                 return nullptr;
             }
 
-            ctx_map[buft] = ctx;
-            ctxs.emplace_back(ctx);
+            ctx_map.emplace(buft, ctx);
 
             return ctx;
         }
 
-        return it->second;
+        return it->second.get();
     };
 
     GGML_ASSERT(n_stream == 1 || n_stream == n_seq_max);
@@ -167,11 +174,8 @@ llama_kv_cache::llama_kv_cache(
     }
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
-    for (auto it : ctx_map) {
-        auto * buft = it.first;
-        auto * ctx  = it.second;
-
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+    for (auto & [buft, ctx] : ctx_map) {
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for kv cache");
         }
@@ -179,7 +183,7 @@ llama_kv_cache::llama_kv_cache(
         LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
 
         ggml_backend_buffer_clear(buf, 0);
-        bufs.emplace_back(buf);
+        ctxs_bufs.emplace_back(std::move(ctx), buf);
     }
 
     {
@@ -203,7 +207,7 @@ void llama_kv_cache::clear(bool data) {
     }
 
     if (data) {
-        for (auto & buf : bufs) {
+        for (auto & [_, buf] : ctxs_bufs) {
             ggml_backend_buffer_clear(buf.get(), 0);
         }
     }
@@ -334,6 +338,8 @@ void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, ll
             llama_pos pos   = v_cells[s0].pos_get(i);
             llama_pos shift = v_cells[s0].get_shift(i);
 
+            llama_kv_cell_ext ext = v_cells[s0].ext_get(i);
+
             if (shift != 0) {
                 pos -= shift;
                 assert(pos >= 0);
@@ -345,6 +351,8 @@ void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, ll
             if (shift != 0) {
                 v_cells[s1].pos_add(i, shift);
             }
+
+            v_cells[s1].ext_set(i, ext);
         }
     }
 
@@ -379,6 +387,7 @@ void llama_kv_cache::seq_keep(llama_seq_id seq_id) {
 
 void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
+    GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_add() is only supported for n_pos_per_embd() == 1");
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
     auto & head  = v_heads[seq_to_stream[seq_id]];
@@ -423,6 +432,7 @@ void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, ll
 
 void llama_kv_cache::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
+    GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_div() is only supported for n_pos_per_embd() == 1");
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
 
@@ -472,8 +482,8 @@ 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 ggml_backend_buffer_ptr & buf_ptr : bufs) {
-        ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get());
+    for (const auto & [_, buf] : ctxs_bufs) {
+        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
     }
     return ret;
 }
@@ -896,6 +906,14 @@ void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch &
 
             cells.pos_set(idx, ubatch.pos[i]);
 
+            if (ubatch.is_pos_2d()) {
+                llama_kv_cell_ext ext {
+                    /*.x =*/ ubatch.pos[i + ubatch.n_tokens*2],
+                    /*.y =*/ ubatch.pos[i + ubatch.n_tokens],
+                };
+                cells.ext_set(idx, ext);
+            }
+
             for (int32_t s = 0; s < ubatch.n_seq_id[i]; s++) {
                 cells.seq_add(idx, ubatch.seq_id[i][s]);
             }
@@ -957,10 +975,14 @@ bool llama_kv_cache::get_has_shift() const {
 uint32_t llama_kv_cache::get_n_kv(const slot_info & sinfo) const {
     uint32_t result = 0;
 
+    // pad the n_kv value so that the graph remains constant across batches and can be reused
+    // note: this also helps some backends with performance (f.ex https://github.com/ggml-org/llama.cpp/pull/16812#issuecomment-3455112220)
+    const uint32_t n_pad_cur = std::max(n_pad, 256u);
+
     for (uint32_t s = 0; s < sinfo.n_stream(); ++s) {
         const auto & cells = v_cells[sinfo.strm[s]];
 
-        result = std::max(std::min(cells.size(), std::max(n_pad, GGML_PAD(cells.used_max_p1(), n_pad))), result);
+        result = std::max(std::min(cells.size(), std::max(n_pad_cur, GGML_PAD(cells.used_max_p1(), n_pad_cur))), result);
     }
 
     return result;
@@ -1239,6 +1261,11 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
 
                 const llama_pos p1 = ubatch->pos[i];
 
+                // for M-RoPE
+                const bool is_2d = ubatch->is_pos_2d();
+                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);
 
                 for (uint32_t j = 0; j < n_kv; ++j) {
@@ -1258,6 +1285,14 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
                         continue;
                     }
 
+                    // M-RoPE causal mask
+                    if (causal_attn && is_2d && p0 == p1) {
+                        const auto & p0_ext = cells.ext_get(j);
+                        if (p0_ext.is_2d_gt(p1_x, p1_y)) {
+                            continue;
+                        }
+                    }
+
                     // apply SWA if any
                     if (is_masked_swa(p0, p1)) {
                         continue;
@@ -1298,7 +1333,7 @@ void llama_kv_cache::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch
 size_t llama_kv_cache::total_size() const {
     size_t size = 0;
 
-    for (const auto & buf : bufs) {
+    for (const auto & [_, buf] : ctxs_bufs) {
         size += ggml_backend_buffer_get_size(buf.get());
     }
 
@@ -1340,7 +1375,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
     const auto & yarn_beta_slow  = cparams.yarn_beta_slow;
 
     const auto & n_rot     = hparams.n_rot;
-    const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE
+    const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
                                 // @ngxson : this is a workaround
                                 // for M-RoPE, we want to rotate the whole vector when doing KV shift
                                 // a normal RoPE should work, we just need to use the correct ordering
@@ -1551,6 +1586,9 @@ void llama_kv_cache::state_write_meta(llama_io_write_i & io, const cell_ranges_t
             io.write(&pos,      sizeof(pos));
             io.write(&n_seq_id, sizeof(n_seq_id));
 
+            // TODO: we also need to save llama_kv_cell_ext when apply_ubatch() support loading it
+            //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
+
             for (const auto & seq_id : seq_ids) {
                 io.write(&seq_id, sizeof(seq_id));
             }
@@ -1696,6 +1734,8 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             return false;
         }
 
+        // TODO: we cannot yet restore llama_kv_cell_ext as the apply_ubatch() does not support it yet
+        //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
         apply_ubatch(sinfo, ubatch);
 
         const auto head_cur = sinfo.head();
@@ -2010,8 +2050,3 @@ void llama_kv_cache_context::set_input_kq_mask(ggml_tensor * dst, const llama_ub
 void llama_kv_cache_context::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
     kv->set_input_pos_bucket(dst, ubatch);
 }
-
-uint32_t llama_kv_cache::get_padding(const llama_cparams & cparams) {
-    // the FA kernels require padding to avoid extra runtime boundary checks
-    return cparams.flash_attn ? 256u : 32u;
-}

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

@@ -19,8 +19,6 @@ struct llama_context;
 
 class llama_kv_cache : public llama_memory_i {
 public:
-    static uint32_t get_padding(const llama_cparams & cparams);
-
     struct stream_copy_info {
         bool empty() const {
             assert(ssrc.size() == sdst.size());
@@ -217,8 +215,8 @@ private:
     // this is the SWA type of the cache - not to be confused with the model SWA type
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
-    std::vector<ggml_context_ptr>        ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
+    // ggml contexts for the KV cache along with the allocated backend buffers:
+    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
 
     // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
     // note: this is not part of the KV state and it's only used to speed-up the find_slot() method

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

@@ -5,9 +5,27 @@
 
 #include <bitset>
 #include <cassert>
-#include <vector>
-#include <set>
+#include <cstring>
 #include <map>
+#include <set>
+#include <vector>
+
+struct llama_kv_cell_ext {
+    // 2D spatial positions, typically used for M-RoPE
+    llama_pos x = 0;
+    llama_pos y = 0;
+
+    // return true if the current 2D spatial position is greater than other
+    bool is_2d_gt(llama_pos ox, llama_pos oy) const {
+        return (y > oy) || (y == oy && x > ox);
+    }
+
+    void reset() {
+        static_assert(std::is_trivially_copyable_v<llama_kv_cell_ext>);
+
+        memset(this, 0, sizeof(*this));
+    }
+};
 
 // meta information about KV cells that can be part of multiple sequences at the same time
 // TODO: add unit tests
@@ -16,6 +34,7 @@ public:
     void reset() {
         for (uint32_t i = 0; i < pos.size(); ++i) {
             pos[i]   = -1;
+            ext[i].reset();
             shift[i] =  0;
             seq[i].reset();
         }
@@ -43,6 +62,7 @@ public:
 
     void resize(uint32_t n) {
         pos.resize(n);
+        ext.resize(n);
         shift.resize(n);
         seq.resize(n);
 
@@ -108,6 +128,7 @@ public:
             const auto idx = i + j;
 
             res.pos[j] = pos[idx];
+            res.ext[j] = ext[idx];
             res.seq[j] = seq[idx];
 
             assert(shift[idx] == 0);
@@ -126,6 +147,7 @@ public:
             const auto idx = idxs[j];
 
             res.pos[j] = pos[idx];
+            res.ext[j] = ext[idx];
             res.seq[j] = seq[idx];
 
             assert(shift[idx] == 0);
@@ -154,6 +176,7 @@ public:
             }
 
             pos[idx] = other.pos[j];
+            ext[idx] = other.ext[j];
             seq[idx] = other.seq[j];
 
             if (pos[idx] != -1) {
@@ -184,6 +207,7 @@ public:
             }
 
             pos[idx] = other.pos[j];
+            ext[idx] = other.ext[j];
             seq[idx] = other.seq[j];
 
             if (pos[idx] != -1) {
@@ -203,6 +227,7 @@ public:
         seq[i].reset();
 
         pos[i] = -1;
+        ext[i].reset();
         shift[i] = 0;
 
         used.erase(i);
@@ -221,6 +246,7 @@ public:
 
         if (seq[i].none()) {
             pos[i] = -1;
+            ext[i].reset();
             shift[i] = 0;
 
             used.erase(i);
@@ -250,6 +276,7 @@ public:
             seq[i].reset();
 
             pos[i] = -1;
+            ext[i].reset();
             shift[i] = 0;
 
             used.erase(i);
@@ -340,6 +367,13 @@ public:
         return pos[i];
     }
 
+    const llama_kv_cell_ext & ext_get(uint32_t i) const {
+        assert(i < pos.size());
+        assert(pos[i] != -1);
+
+        return ext[i];
+    }
+
     // note: call only if the cell is not empty
     llama_pos get_shift(uint32_t i) const {
         assert(i < pos.size());
@@ -368,6 +402,11 @@ public:
         used.insert(i);
     }
 
+    void ext_set(uint32_t i, llama_kv_cell_ext p) {
+        assert(i < ext.size());
+        ext[i] = p;
+    }
+
     // pos[i] = pos[i] + d
     // sets "has_shift" to true
     // note: call only if the cell is not empty
@@ -424,6 +463,9 @@ private:
 
     std::vector<llama_pos> pos;
 
+    // stores extra info per cell
+    std::vector<llama_kv_cell_ext> ext;
+
     // this array accumulates any applied shifts to the pos array since the last reset_shift() call
     // this is used to queue multiple updates to the pos array, which in the end can be applied in one go:
     //

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

@@ -7,6 +7,7 @@
 
 #include <algorithm>
 #include <cassert>
+#include <cstring>
 #include <limits>
 #include <map>
 #include <stdexcept>
@@ -32,8 +33,15 @@ llama_memory_recurrent::llama_memory_recurrent(
     cells.clear();
     cells.resize(mem_size);
 
+    // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
+    struct ggml_backend_buft_comparator {
+        bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
+            return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
+        }
+    };
+    std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
+
     // create a context for each buffer type
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
@@ -48,13 +56,12 @@ llama_memory_recurrent::llama_memory_recurrent(
                 return nullptr;
             }
 
-            ctx_map[buft] = ctx;
-            ctxs.emplace_back(ctx);
+            ctx_map.emplace(buft, ctx);
 
             return ctx;
         }
 
-        return it->second;
+        return it->second.get();
     };
 
     r_l.resize(n_layer);
@@ -93,17 +100,14 @@ llama_memory_recurrent::llama_memory_recurrent(
     }
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
-    for (auto it : ctx_map) {
-        auto * buft = it.first;
-        auto * ctx  = it.second;
-
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+    for (auto & [buft, ctx] : ctx_map) {
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for rs cache");
         }
         ggml_backend_buffer_clear(buf, 0);
         LLAMA_LOG_INFO("%s: %10s RS buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
-        bufs.emplace_back(buf);
+        ctxs_bufs.emplace_back(std::move(ctx), buf);
     }
 
     {
@@ -129,7 +133,7 @@ void llama_memory_recurrent::clear(bool data) {
     used = 0;
 
     if (data) {
-        for (auto & buf : bufs) {
+        for (auto & [_, buf] : ctxs_bufs) {
             ggml_backend_buffer_clear(buf.get(), 0);
         }
     }
@@ -147,7 +151,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         p1 = std::numeric_limits<llama_pos>::max();
     }
 
-    // models like Mamba or RWKV can't have a state partially erased
+    // models like Mamba or RWKV can't have a state partially erased at the end
+    // of the sequence because their state isn't preserved for previous tokens
     if (seq_id >= (int64_t) size) {
         // could be fatal
         return false;
@@ -156,8 +161,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         int32_t & tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
             const auto & cell = cells[tail_id];
-            // partial intersection is invalid
-            if ((0 < p0 && p0 < cell.pos) || (0 < p1 && p1 <= cell.pos)) {
+            // partial intersection is invalid if it includes the final pos
+            if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
                 //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
                 return false;
             }
@@ -364,8 +369,8 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_memory_recurrent::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const ggml_backend_buffer_ptr & buf_ptr : bufs) {
-        ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get());
+    for (const auto & [_, buf] : ctxs_bufs) {
+        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
     }
     return ret;
 }
@@ -662,7 +667,7 @@ bool llama_memory_recurrent::get_can_shift() const {
 
 size_t llama_memory_recurrent::total_size() const {
     size_t size = 0;
-    for (const auto & buf : bufs) {
+    for (const auto & [_, buf] : ctxs_bufs) {
         size += ggml_backend_buffer_get_size(buf.get());
     }
 

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

@@ -109,8 +109,8 @@ private:
 
     const uint32_t n_seq_max = 1;
 
-    std::vector<ggml_context_ptr>        ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
+    // ggml contexts for the KV cache along with the allocated backend buffers:
+    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
 
     size_t total_size() const;
 

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.h

@@ -77,6 +77,7 @@ enum llm_type {
     LLM_TYPE_16B,
     LLM_TYPE_20B,
     LLM_TYPE_22B,
+    LLM_TYPE_26B,
     LLM_TYPE_27B,
     LLM_TYPE_30B,
     LLM_TYPE_32B,
@@ -113,8 +114,11 @@ 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
+    LLM_TYPE_230B_A10B, // Minimax M2
     LLM_TYPE_235B_A22B,
     LLM_TYPE_300B_A47B, // Ernie MoE big
     LLM_TYPE_355B_A32B, // GLM-4.5
@@ -234,6 +238,7 @@ struct llama_layer {
     struct ggml_tensor * wk_enc    = nullptr;
     struct ggml_tensor * wv_enc    = nullptr;
     struct ggml_tensor * wo_enc    = nullptr;
+    struct ggml_tensor * wqkv_gate = nullptr;
 
     // attention bias
     struct ggml_tensor * bq   = nullptr;
@@ -307,6 +312,9 @@ struct llama_layer {
     struct ggml_tensor * ssm_conv1d_b = nullptr;
     struct ggml_tensor * ssm_dt_b     = nullptr;
 
+    // qwen3next
+    struct ggml_tensor * ssm_beta_alpha = nullptr;
+
     // rwkv
     struct ggml_tensor * time_mix_w1         = nullptr;
     struct ggml_tensor * time_mix_w2         = nullptr;
@@ -385,6 +393,13 @@ struct llama_layer {
     // openai-moe
     struct ggml_tensor * attn_sinks = nullptr;
 
+    // cogvlm
+    struct ggml_tensor * visexp_attn_wqkv = nullptr;
+    struct ggml_tensor * visexp_attn_wo   = nullptr;
+    struct ggml_tensor * visexp_ffn_gate  = nullptr;
+    struct ggml_tensor * visexp_ffn_down  = nullptr;
+    struct ggml_tensor * visexp_ffn_up    = nullptr;
+
     // xIELU activation parameters for Apertus
     struct ggml_tensor * ffn_act_alpha_n = nullptr;
     struct ggml_tensor * ffn_act_alpha_p = nullptr;
@@ -503,9 +518,8 @@ struct llama_model {
 
     ggml_tensor * get_rope_factors(const llama_cparams & cparams, int il) const;
 
-    // note: can mutate `cparams`
     // TODO: move this to new llm_arch_model_i interface
-    llama_memory_i * create_memory(const llama_memory_params & params, llama_cparams & cparams) const;
+    llama_memory_i * create_memory(const llama_memory_params & params, const llama_cparams & cparams) const;
 
     // TODO: move this to new llm_arch_model_i interface
     ggml_cgraph * build_graph(const llm_graph_params & params) const;

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

@@ -653,7 +653,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                 gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
                 // Setting type to UINT32. See https://github.com/ggml-org/llama.cpp/pull/14182 for context
-                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)abs(o.val_i64));
+                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)std::abs(o.val_i64));
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
                 gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {
@@ -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;
@@ -674,14 +673,11 @@ 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;
-        } else if (remapped_name != it.first) {
+        }
+
+        if (remapped_name != it.first) {
             ggml_set_name(it.second.tensor, remapped_name.c_str());
             LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor));
         }
@@ -701,7 +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;
 
@@ -715,26 +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_attn_layer = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
-        if (llama_model_has_encoder(&model)) {
-            // now n_attn_layer is the number of attention layers in the encoder
-            // for each decoder block, there are 2 attention layers
-            n_attn_layer += 2 * model.hparams.dec_n_layer;
-        }
-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
-    }
-
     size_t total_size_org = 0;
     size_t total_size_new = 0;
 

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

@@ -4,6 +4,7 @@
 #include "llama-vocab.h"
 #include "llama-grammar.h"
 
+#include <array>
 #include <algorithm>
 #include <cassert>
 #include <cfloat>
@@ -471,9 +472,6 @@ static void llama_sampler_chain_reset(struct llama_sampler * smpl) {
     for (auto * smpl : chain->samplers) {
         llama_sampler_reset(smpl);
     }
-
-    chain->t_sample_us = 0;
-    chain->n_sample    = 0;
 }
 
 static struct llama_sampler * llama_sampler_chain_clone(const struct llama_sampler * smpl) {
@@ -1625,10 +1623,12 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
     auto * ctx = new llama_sampler_grammar;
 
     if (grammar_str != nullptr && grammar_str[0] != '\0') {
+        std::string trigger_pattern;
+        llama_grammar * grammar = nullptr;
         // TODO: remove trigger_words support.
         if (trigger_words != nullptr && num_trigger_words > 0) {
             GGML_ASSERT(trigger_patterns == nullptr && num_trigger_patterns == 0);
-            std::string trigger_pattern("[\\s\\S]*?(");
+            trigger_pattern = "[\\s\\S]*?(";
             for (size_t i = 0; i < num_trigger_words; ++i) {
                 static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
                 if (i > 0) {
@@ -1637,15 +1637,17 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
                 trigger_pattern += std::regex_replace(trigger_words[i], special_chars, "\\$0");
             }
             trigger_pattern += ")[\\s\\S]*";
-            const auto * trigger_pattern_c = trigger_pattern.c_str();
-            trigger_patterns = &trigger_pattern_c;
-            num_trigger_patterns = 1;
+
+            std::array<const char *, 1> tmp_trigger_patterns = { trigger_pattern.c_str() };
+            grammar = llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, tmp_trigger_patterns.data(), tmp_trigger_patterns.size(), trigger_tokens, num_trigger_tokens);
+        } else {
+            grammar = llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens);
         }
         *ctx = {
             /* .vocab        = */ vocab,
             /* .grammar_str  = */ grammar_str,
             /* .grammar_root = */ grammar_root,
-            /* .grammar      = */ llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens),
+            /* .grammar      = */ grammar,
         };
         if (!ctx->grammar) {
             delete ctx;
@@ -2665,8 +2667,7 @@ struct llama_perf_sampler_data llama_perf_sampler(const struct llama_sampler * c
 void llama_perf_sampler_print(const struct llama_sampler * chain) {
     const auto data = llama_perf_sampler(chain);
 
-    LLAMA_LOG_INFO("%s:    sampling time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, data.t_sample_ms, data.n_sample, data.t_sample_ms / data.n_sample, 1e3 / data.t_sample_ms * data.n_sample);
+    LLAMA_LOG_INFO("%s:    samplers time = %10.2f ms / %5d runs\n", __func__, data.t_sample_ms, data.n_sample);
 }
 
 void llama_perf_sampler_reset(struct llama_sampler * chain) {
@@ -2676,5 +2677,6 @@ void llama_perf_sampler_reset(struct llama_sampler * chain) {
 
     auto * ctx = (struct llama_sampler_chain *) chain->ctx;
 
-    ctx->t_sample_us = ctx->n_sample = 0;
+    ctx->t_sample_us = 0;
+    ctx->n_sample    = 0;
 }

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

@@ -401,6 +401,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                 };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_GPT4O:
+            case LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2:
                 regex_exprs = {
                     // original regex from tokenizer.json
                     // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
@@ -442,6 +443,17 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_AFMOE:
+                regex_exprs = {
+                    // Digit handling - uses custom implementation in unicode.cpp
+                    // Groups digits with leading 1-2 based on total length modulo 3
+                    "\\p{AFMoE_digits}",
+                    // CJK and Asian scripts (using direct Unicode literals)
+                    "[一-鿿㐀-䶿豈-﫿぀-ゟ゠-ヿ･-ﾟ⼀-⿟เ-๿຀-໿ក-៿က-႟ꩠ-ꩿꧠ-꧿가-힯ᄀ-ᇿ]+",
+                    // Main BPE pattern
+                    "[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\\r\\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -1012,7 +1024,7 @@ private:
         }
     private:
         uint32_t get_node(size_t index) {
-            if (index > xcda_array_size) {
+            if (index >= xcda_array_size) {
                 throw std::runtime_error("Index out of array bounds in XCDA array!");
             }
             return xcda_array[index];
@@ -1269,6 +1281,7 @@ struct llm_tokenizer_plamo2 : llm_tokenizer {
 
         // Build suffix list in lexicographical order of reversed strings
         std::vector<std::string> suffixes;
+        suffixes.reserve(suffix_to_score.size() + 1);
         for (const auto & pair : suffix_to_score) {
             suffixes.push_back(pair.first);
         }
@@ -1981,6 +1994,14 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "grok-2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GROK_2;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "afmoe") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_AFMOE;
+                clean_spaces = false;
+            } else if (
+                tokenizer_pre == "minimax-m2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2;
+                clean_spaces = false;
             } else {
                 LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
                 pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
@@ -3222,8 +3243,7 @@ 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-vocab.h

@@ -49,6 +49,8 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE   = 38,
     LLAMA_VOCAB_PRE_TYPE_GROK_2          = 39,
     LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING = 40,
+    LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2      = 41,
+    LLAMA_VOCAB_PRE_TYPE_AFMOE           = 42,
 };
 
 struct LLM_KV;

llama/llama.cpp/src/llama.go

@@ -5,4 +5,8 @@ package llama
 // #cgo CPPFLAGS: -I${SRCDIR}/../../../ml/backend/ggml/ggml/include
 // #cgo windows CPPFLAGS: -D_WIN32_WINNT=0x0602
 import "C"
-import _ "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
+
+import (
+	_ "github.com/ollama/ollama/llama/llama.cpp/src/models"
+	_ "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
+)

llama/llama.cpp/src/models/afmoe.cpp

@@ -0,0 +1,187 @@
+#include "models.h"
+
+llm_build_afmoe::llm_build_afmoe(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_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // MuP scaling: embeddings * sqrt(hidden_size)
+    // mup_enabled = true, hidden_size = 1024, scale = 32.0
+    inpL = ggml_scale(ctx0, inpL, sqrtf(float(n_embd)));
+    cb(inpL, "inp_embd_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+    auto * inp_attn = build_attn_inp_kv_iswa();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const float kq_scale = 1.0f/sqrtf(float(n_embd_head));
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // dual attention normalization (pre)
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * attn_inp = cur;  // save input for gate computation
+
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            // compute gate from input
+            ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
+            cb(gate, "attn_gate_proj", 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);
+
+            // Q/K normalization
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+
+            // RoPE only for sliding_attention layers
+            const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
+                                ((il + 1) % hparams.n_no_rope_layer_step) != 0;
+            if (use_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);
+                cb(Qcur, "Qcur_rope", il);
+
+                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(Kcur, "Kcur_rope", il);
+            }
+
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            cur = build_attn(inp_attn,
+                    NULL, NULL,  // wo will be applied after gating
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+
+            // attention gating: attn_out * sigmoid(gate) BEFORE o_proj
+            gate = ggml_sigmoid(ctx0, gate);
+            cb(gate, "attn_gate_sig", il);
+            cur = ggml_mul(ctx0, cur, gate);
+            cb(cur, "attn_gated", il);
+
+            // now apply output projection
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            cb(cur, "attn_o_proj", il);
+        }
+
+        // dual attention normalization (post)
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", 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);
+
+        // dual ffn normalization (pre)
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // MoE or dense FFN
+        if ((uint32_t)il >= hparams.n_layer_dense_lead) {
+            // MoE layer with sigmoid routing, normalization, and scaling
+            ggml_tensor * moe_out = 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,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU,
+                    hparams.expert_weights_norm,           // norm_w (route_norm=True)
+                    hparams.expert_weights_scale,          // scale_w
+                    hparams.expert_weights_scale,          // w_scale (route_scale=2.826)
+                    (llama_expert_gating_func_type) hparams.expert_gating_func,
+                    il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // shared expert
+            if (hparams.n_expert_shared > 0) {
+                ggml_tensor * ffn_shexp = build_ffn(cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            } else {
+                cur = moe_out;
+            }
+        } else {
+            // dense layer
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // dual ffn normalization (post)
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        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/apertus.cpp

@@ -0,0 +1,125 @@
+#include "models.h"
+
+
+
+llm_build_apertus::llm_build_apertus(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);
+
+    ggml_tensor * inp_pos  = build_inp_pos();
+    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;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            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);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            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_pos", il);
+            cb(Kcur, "Kcur_pos", il);
+            cb(Vcur, "Vcur_pos", 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 with xIELU activation
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // Up projection
+            ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur);
+            cb(up, "ffn_up", il);
+
+            float alpha_n_val = hparams.xielu_alpha_n[il];
+            float alpha_p_val = hparams.xielu_alpha_p[il];
+            float beta_val    = hparams.xielu_beta[il];
+            float eps_val     = hparams.xielu_eps[il];
+
+            // Apply xIELU activation
+            ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val);
+            cb(activated, "ffn_xielu", il);
+
+            // Down projection
+            cur = build_lora_mm(model.layers[il].ffn_down, activated);
+            cb(cur, "ffn_down", 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, nullptr, 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/arcee.cpp

@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+llm_build_arcee::llm_build_arcee(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();
+
+    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);
+
+            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
+        // ARCEE uses relu^2 instead of silu
+        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,   NULL, NULL,
+                NULL,                      NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/arctic.cpp

@@ -0,0 +1,138 @@
+#include "models.h"
+
+
+llm_build_arctic::llm_build_arctic(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();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    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
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            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, 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);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), 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
+        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,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
+        cb(ffn_out, "ffn_out", il);
+
+        // MoE
+        cur = build_norm(inpSA,
+                model.layers[il].ffn_norm_exps, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm_exps", 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_out);
+        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/arwkv7.cpp

@@ -0,0 +1,86 @@
+#include "models.h"
+
+
+llm_build_arwkv7::llm_build_arwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) {
+    GGML_ASSERT(n_embd == hparams.n_embd_r());
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * v_first = nullptr;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * rs_inp = build_rs_inp();
+
+    const auto n_embd = hparams.n_embd;
+    const auto n_seq_tokens = ubatch.n_seq_tokens;
+    const auto n_seqs = ubatch.n_seqs;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const llama_layer * layer = &model.layers[il];
+        inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
+
+        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
+
+        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
+        cb(att_norm, "attn_norm", il);
+
+        ggml_tensor * x_prev = ggml_concat(
+                ctx0,
+                token_shift,
+                ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
+                1
+                );
+
+        cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il);
+
+        token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
+        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
+        ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        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, model.output_norm_b, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    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/baichuan.cpp

@@ -0,0 +1,122 @@
+#include "models.h"
+
+
+llm_build_baichuan::llm_build_baichuan(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 = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            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);
+
+            switch (model.type) {
+                case LLM_TYPE_7B:
+                    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
+                            );
+                    break;
+                case LLM_TYPE_13B:
+                    break;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), 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
+        {
+            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,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        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/bailingmoe.cpp

@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    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();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    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_rot, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), 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);
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            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, hparams.expert_weights_norm,
+                    false, hparams.expert_weights_scale,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // FFN shared expert
+        {
+            ggml_tensor * ffn_shexp = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        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/bailingmoe2.cpp

@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+
+llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    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();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+    for (int il = 0; il < n_transformer_layers; ++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
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 0 * sizeof(float) * (n_embd));
+            ggml_tensor * Kcur = 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));
+            ggml_tensor * 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 = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_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 = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            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);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_transformer_layers - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpSA);
+        cb(sa_out, "sa_out", il);
+
+        // MoE branch
+        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * moe_out = 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,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        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);
+}