llama: improve README sync documentation

Add numbered steps and clarify the conflict resolution workflow.
Add note about checking for upstreamed patches before updating.

.github/ISSUE_TEMPLATE/10_bug_report.yml

@@ -13,7 +13,7 @@ body:
     id: logs
     attributes:
       label: Relevant log output
-      description: Please copy and paste any relevant log output. See [Troubleshooting Guide](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.mdx#how-to-troubleshoot-issues) for details.
+      description: Please copy and paste any relevant log output. See [Troubleshooting Guide](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) for details.
       render: shell
     validations:
       required: false

.github/workflows/release.yaml

@@ -372,17 +372,13 @@ jobs:
           outputs: type=local,dest=dist/${{ matrix.os }}-${{ matrix.arch }}
           cache-from: type=registry,ref=${{ vars.DOCKER_REPO }}:latest
           cache-to: type=inline
-      - name: Deduplicate CUDA libraries
-        run: |
-          ./scripts/deduplicate_cuda_libs.sh dist/${{ matrix.os }}-${{ matrix.arch }}
       - run: |
           for COMPONENT in bin/* lib/ollama/*; do
             case "$COMPONENT" in
-              bin/ollama*)               echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              bin/ollama)                echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/cuda_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/vulkan*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/mlx*)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/cuda_jetpack5)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
               lib/ollama/cuda_jetpack6)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
               lib/ollama/rocm)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;

CMakeLists.txt

@@ -48,10 +48,9 @@ if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
     set(GGML_CPU_ALL_VARIANTS ON)
 endif()
 
-if(APPLE)
+if (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
     set(CMAKE_BUILD_RPATH "@loader_path")
     set(CMAKE_INSTALL_RPATH "@loader_path")
-    set(CMAKE_BUILD_WITH_INSTALL_RPATH ON)
 endif()
 
 set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
@@ -190,21 +189,13 @@ if(MLX_ENGINE)
     install(TARGETS mlx mlxc
         RUNTIME_DEPENDENCIES
             DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
-            PRE_INCLUDE_REGEXES cublas cublasLt cudart nvrtc nvrtc-builtins cudnn nccl openblas gfortran
+            PRE_INCLUDE_REGEXES cublas cublasLt cudart nvrtc cudnn nccl
             PRE_EXCLUDE_REGEXES ".*"
         RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
         LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
         FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
     )
 
-    # Install the Metal library for macOS arm64 (must be colocated with the binary)
-    # Metal backend is only built for arm64, not x86_64
-    if(APPLE AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
-        install(FILES ${CMAKE_BINARY_DIR}/_deps/mlx-build/mlx/backend/metal/kernels/mlx.metallib
-            DESTINATION ${OLLAMA_INSTALL_DIR}
-            COMPONENT MLX)
-    endif()
-
     # Manually install cudart and cublas since they might not be picked up as direct dependencies
     if(CUDAToolkit_FOUND)
         file(GLOB CUDART_LIBS

Dockerfile

@@ -161,9 +161,6 @@ ARG GOFLAGS="'-ldflags=-w -s'"
 ENV CGO_ENABLED=1
 ARG CGO_CFLAGS
 ARG CGO_CXXFLAGS
-RUN mkdir -p dist/bin
-RUN --mount=type=cache,target=/root/.cache/go-build \
-    go build -tags mlx -trimpath -buildmode=pie -o dist/bin/ollama-mlx .
 
 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama
@@ -185,7 +182,6 @@ COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
 COPY --from=cuda-13 dist/lib/ollama /lib/ollama/
 COPY --from=vulkan  dist/lib/ollama  /lib/ollama/
 COPY --from=mlx     /go/src/github.com/ollama/ollama/dist/lib/ollama /lib/ollama/
-COPY --from=mlx     /go/src/github.com/ollama/ollama/dist/bin/ /bin/
 
 FROM --platform=linux/arm64 scratch AS arm64
 # COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggml-org/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=ec98e2002
+FETCH_HEAD=b1377188784f9aea26b8abde56d4aee8c733eec7
 
 .PHONY: help
 help:

README.md

@@ -48,7 +48,7 @@ ollama run gemma3
 
 ## Model library
 
-Ollama supports a list of models available on [ollama.com/library](https://ollama.com/library "ollama model library")
+Ollama supports a list of models available on [ollama.com/library](https://ollama.com/library 'ollama model library')
 
 Here are some example models that can be downloaded:
 
@@ -79,7 +79,7 @@ Here are some example models that can be downloaded:
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`           |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored`   |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`               |
-| Granite-3.3        | 8B         | 4.9GB | `ollama run granite3.3`          |
+| Granite-3.3         | 8B         | 4.9GB | `ollama run granite3.3`          |
 
 > [!NOTE]
 > You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
@@ -260,38 +260,6 @@ Finally, in a separate shell, run a model:
 ./ollama run llama3.2
 ```
 
-## Building with MLX (experimental)
-
-First build the MLX libraries:
-
-```shell
-cmake --preset MLX
-cmake --build --preset MLX --parallel
-cmake --install build --component MLX
-```
-
-Next, build the `ollama-mlx` binary, which is a separate build of the Ollama runtime with MLX support enabled (needs to be in the same directory as `ollama`):
-
-```shell
-go build -tags mlx -o ollama-mlx .
-```
-
-Finally, start the server:
-
-```
-./ollama serve
-```
-
-### Building MLX with CUDA
-
-When building with CUDA, use the preset "MLX CUDA 13" or "MLX CUDA 12" to enable CUDA with default architectures:
-
-```shell
-cmake --preset 'MLX CUDA 13'
-cmake --build --preset 'MLX CUDA 13' --parallel
-cmake --install build --component MLX
-```
-
 ## REST API
 
 Ollama has a REST API for running and managing models.
@@ -322,7 +290,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 ### Web & Desktop
 
-- [Onyx](https://github.com/onyx-dot-app/onyx)
 - [Open WebUI](https://github.com/open-webui/open-webui)
 - [SwiftChat (macOS with ReactNative)](https://github.com/aws-samples/swift-chat)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
@@ -454,7 +421,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [AppFlowy](https://github.com/AppFlowy-IO/AppFlowy) (AI collaborative workspace with Ollama, cross-platform and self-hostable)
 - [Lumina](https://github.com/cushydigit/lumina.git) (A lightweight, minimal React.js frontend for interacting with Ollama servers)
 - [Tiny Notepad](https://pypi.org/project/tiny-notepad) (A lightweight, notepad-like interface to chat with ollama available on PyPI)
-- [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.)
+- [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.) 
 - [GPTranslate](https://github.com/philberndt/GPTranslate) (A fast and lightweight, AI powered desktop translation application written with Rust and Tauri. Features real-time translation with OpenAI/Azure/Ollama.)
 - [ollama launcher](https://github.com/NGC13009/ollama-launcher) (A launcher for Ollama, aiming to provide users with convenient functions such as ollama server launching, management, or configuration.)
 - [ai-hub](https://github.com/Aj-Seven/ai-hub) (AI Hub supports multiple models via API keys and Chat support via Ollama API.)
@@ -526,7 +493,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Database
 
 - [pgai](https://github.com/timescale/pgai) - PostgreSQL as a vector database (Create and search embeddings from Ollama models using pgvector)
-  - [Get started guide](https://github.com/timescale/pgai/blob/main/docs/vectorizer-quick-start.md)
+   - [Get started guide](https://github.com/timescale/pgai/blob/main/docs/vectorizer-quick-start.md)
 - [MindsDB](https://github.com/mindsdb/mindsdb/blob/staging/mindsdb/integrations/handlers/ollama_handler/README.md) (Connects Ollama models with nearly 200 data platforms and apps)
 - [chromem-go](https://github.com/philippgille/chromem-go/blob/v0.5.0/embed_ollama.go) with [example](https://github.com/philippgille/chromem-go/tree/v0.5.0/examples/rag-wikipedia-ollama)
 - [Kangaroo](https://github.com/dbkangaroo/kangaroo) (AI-powered SQL client and admin tool for popular databases)
@@ -669,7 +636,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [llama.cpp](https://github.com/ggml-org/llama.cpp) project founded by Georgi Gerganov.
 
 ### Observability
-
 - [Opik](https://www.comet.com/docs/opik/cookbook/ollama) is an open-source platform to debug, evaluate, and monitor your LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards. Opik supports native integration to Ollama.
 - [Lunary](https://lunary.ai/docs/integrations/ollama) is the leading open-source LLM observability platform. It provides a variety of enterprise-grade features such as real-time analytics, prompt templates management, PII masking, and comprehensive agent tracing.
 - [OpenLIT](https://github.com/openlit/openlit) is an OpenTelemetry-native tool for monitoring Ollama Applications & GPUs using traces and metrics.
@@ -678,5 +644,4 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [MLflow Tracing](https://mlflow.org/docs/latest/llms/tracing/index.html#automatic-tracing) is an open source LLM observability tool with a convenient API to log and visualize traces, making it easy to debug and evaluate GenAI applications.
 
 ### Security
-
 - [Ollama Fortress](https://github.com/ParisNeo/ollama_proxy_server)

api/client.go

@@ -165,7 +165,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	return nil
 }
 
-const maxBufferSize = 8 * format.MegaByte
+const maxBufferSize = 512 * format.KiloByte
 
 func (c *Client) stream(ctx context.Context, method, path string, data any, fn func([]byte) error) error {
 	var buf io.Reader
@@ -281,20 +281,6 @@ func (c *Client) Generate(ctx context.Context, req *GenerateRequest, fn Generate
 	})
 }
 
-// XGenerate generates images using the experimental /x/generate endpoint.
-// This endpoint is specifically designed for image generation models and
-// supports parameters like width, height, and steps.
-func (c *Client) XGenerate(ctx context.Context, req *GenerateRequest, fn GenerateResponseFunc) error {
-	return c.stream(ctx, http.MethodPost, "/x/generate", req, func(bts []byte) error {
-		var resp GenerateResponse
-		if err := json.Unmarshal(bts, &resp); err != nil {
-			return err
-		}
-
-		return fn(resp)
-	})
-}
-
 // ChatResponseFunc is a function that [Client.Chat] invokes every time
 // a response is received from the service. If this function returns an error,
 // [Client.Chat] will stop generating and return this error.

api/types.go

@@ -97,15 +97,6 @@ type GenerateRequest struct {
 	// request, for multimodal models.
 	Images []ImageData `json:"images,omitempty"`
 
-	// Width is the width of the generated image (for image generation models).
-	Width int32 `json:"width,omitempty"`
-
-	// Height is the height of the generated image (for image generation models).
-	Height int32 `json:"height,omitempty"`
-
-	// Steps is the number of diffusion steps (for image generation models).
-	Steps int32 `json:"steps,omitempty"`
-
 	// Options lists model-specific options. For example, temperature can be
 	// set through this field, if the model supports it.
 	Options map[string]any `json:"options"`
@@ -869,18 +860,6 @@ type GenerateResponse struct {
 	// Logprobs contains log probability information for the generated tokens,
 	// if requested via the Logprobs parameter.
 	Logprobs []Logprob `json:"logprobs,omitempty"`
-
-	// Status describes the current phase of generation (e.g., "generating image").
-	Status string `json:"status,omitempty"`
-
-	// Total is the total count for the current phase (e.g., total steps).
-	Total int64 `json:"total,omitempty"`
-
-	// Completed is the completed count for the current phase.
-	Completed int64 `json:"completed,omitempty"`
-
-	// Images contains base64-encoded generated images for image generation models.
-	Images []string `json:"images,omitempty"`
 }
 
 // ModelDetails provides details about a model.

app/cmd/app/app_darwin.m

@@ -14,7 +14,6 @@ extern NSString *SystemWidePath;
 @interface AppDelegate () <NSWindowDelegate, WKNavigationDelegate, WKUIDelegate>
 @property(strong, nonatomic) NSStatusItem *statusItem;
 @property(assign, nonatomic) BOOL updateAvailable;
-@property(assign, nonatomic) BOOL systemShutdownInProgress;
 @end
 
 @implementation AppDelegate
@@ -41,13 +40,6 @@ bool firstTimeRun,startHidden; // Set in run before initialization
 }
 
 - (void)applicationDidFinishLaunching:(NSNotification *)aNotification {
-    // Register for system shutdown/restart notification so we can allow termination
-    [[[NSWorkspace sharedWorkspace] notificationCenter]
-        addObserver:self
-           selector:@selector(systemWillPowerOff:)
-               name:NSWorkspaceWillPowerOffNotification
-             object:nil];
-
     // if we're in development mode, set the app icon
     NSString *bundlePath = [[NSBundle mainBundle] bundlePath];
     if (![bundlePath hasSuffix:@".app"]) {
@@ -286,18 +278,7 @@ bool firstTimeRun,startHidden; // Set in run before initialization
     [NSApp activateIgnoringOtherApps:YES];
 }
 
-- (void)systemWillPowerOff:(NSNotification *)notification {
-    // Set flag so applicationShouldTerminate: knows to allow termination.
-    // The system will call applicationShouldTerminate: after posting this notification.
-    self.systemShutdownInProgress = YES;
-}
-
 - (NSApplicationTerminateReply)applicationShouldTerminate:(NSApplication *)sender {
-    // Allow termination if the system is shutting down or restarting
-    if (self.systemShutdownInProgress) {
-        return NSTerminateNow;
-    }
-    // Otherwise just hide the app (for Cmd+Q, close button, etc.)
     [NSApp hide:nil];
     [NSApp setActivationPolicy:NSApplicationActivationPolicyAccessory];
     return NSTerminateCancel;

cmd/cmd.go

@@ -46,9 +46,8 @@ import (
 	"github.com/ollama/ollama/types/syncmap"
 	"github.com/ollama/ollama/version"
 	xcmd "github.com/ollama/ollama/x/cmd"
-	"github.com/ollama/ollama/x/create"
-	xcreateclient "github.com/ollama/ollama/x/create/client"
 	"github.com/ollama/ollama/x/imagegen"
+	imagegenclient "github.com/ollama/ollama/x/imagegen/client"
 )
 
 const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
@@ -94,87 +93,14 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
-	// Validate model name early to fail fast
-	modelName := args[0]
-	name := model.ParseName(modelName)
-	if !name.IsValid() {
-		return fmt.Errorf("invalid model name: %s", modelName)
-	}
-
-	// Check for --experimental flag for safetensors model creation
-	experimental, _ := cmd.Flags().GetBool("experimental")
-	if experimental {
-		// Get Modelfile content - either from -f flag or default to "FROM ."
-		var reader io.Reader
-		filename, err := getModelfileName(cmd)
-		if os.IsNotExist(err) || filename == "" {
-			// No Modelfile specified or found - use default
-			reader = strings.NewReader("FROM .\n")
-		} else if err != nil {
-			return err
-		} else {
-			f, err := os.Open(filename)
-			if err != nil {
-				return err
-			}
-			defer f.Close()
-			reader = f
-		}
-
-		// Parse the Modelfile
-		modelfile, err := parser.ParseFile(reader)
-		if err != nil {
-			return fmt.Errorf("failed to parse Modelfile: %w", err)
-		}
-
-		// Extract FROM path and configuration
-		var modelDir string
-		mfConfig := &xcreateclient.ModelfileConfig{}
-
-		for _, cmd := range modelfile.Commands {
-			switch cmd.Name {
-			case "model":
-				modelDir = cmd.Args
-			case "template":
-				mfConfig.Template = cmd.Args
-			case "system":
-				mfConfig.System = cmd.Args
-			case "license":
-				mfConfig.License = cmd.Args
-			}
-		}
-
-		if modelDir == "" {
-			modelDir = "."
-		}
-
-		// Resolve relative paths based on Modelfile location
-		if !filepath.IsAbs(modelDir) && filename != "" {
-			modelDir = filepath.Join(filepath.Dir(filename), modelDir)
-		}
-
-		quantize, _ := cmd.Flags().GetString("quantize")
-		return xcreateclient.CreateModel(xcreateclient.CreateOptions{
-			ModelName: modelName,
-			ModelDir:  modelDir,
-			Quantize:  quantize,
-			Modelfile: mfConfig,
-		}, p)
-	}
-
 	var reader io.Reader
 
 	filename, err := getModelfileName(cmd)
 	if os.IsNotExist(err) {
 		if filename == "" {
 			// No Modelfile found - check if current directory is an image gen model
-			if create.IsTensorModelDir(".") {
-				quantize, _ := cmd.Flags().GetString("quantize")
-				return xcreateclient.CreateModel(xcreateclient.CreateOptions{
-					ModelName: modelName,
-					ModelDir:  ".",
-					Quantize:  quantize,
-				}, p)
+			if imagegen.IsTensorModelDir(".") {
+				return imagegenclient.CreateModel(args[0], ".", p)
 			}
 			reader = strings.NewReader("FROM .\n")
 		} else {
@@ -207,7 +133,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	}
 	spinner.Stop()
 
-	req.Model = modelName
+	req.Model = args[0]
 	quantize, _ := cmd.Flags().GetString("quantize")
 	if quantize != "" {
 		req.Quantize = quantize
@@ -538,6 +464,14 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 
 	name := args[0]
 
+	// Check if this is a known image generation model (skip Show/Pull)
+	if imagegen.HasTensorLayers(name) {
+		if opts.Prompt == "" && !interactive {
+			return errors.New("image generation models require a prompt. Usage: ollama run " + name + " \"your prompt here\"")
+		}
+		return imagegen.RunCLI(cmd, name, opts.Prompt, interactive, opts.KeepAlive)
+	}
+
 	info, err := func() (*api.ShowResponse, error) {
 		showReq := &api.ShowRequest{Name: name}
 		info, err := client.Show(cmd.Context(), showReq)
@@ -599,18 +533,9 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return generateEmbedding(cmd, name, opts.Prompt, opts.KeepAlive, truncate, dimensions)
 	}
 
-	// Check if this is an image generation model
-	if slices.Contains(info.Capabilities, model.CapabilityImageGeneration) {
-		if opts.Prompt == "" && !interactive {
-			return errors.New("image generation models require a prompt. Usage: ollama run " + name + " \"your prompt here\"")
-		}
-		return imagegen.RunCLI(cmd, name, opts.Prompt, interactive, opts.KeepAlive)
-	}
-
 	// Check for experimental flag
 	isExperimental, _ := cmd.Flags().GetBool("experimental")
 	yoloMode, _ := cmd.Flags().GetBool("experimental-yolo")
-	enableWebsearch, _ := cmd.Flags().GetBool("experimental-websearch")
 
 	if interactive {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
@@ -640,7 +565,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 
 		// Use experimental agent loop with tools
 		if isExperimental {
-			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive, yoloMode, enableWebsearch)
+			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive, yoloMode)
 		}
 
 		return generateInteractive(cmd, opts)
@@ -746,11 +671,7 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 
 			bar, ok := bars[resp.Digest]
 			if !ok {
-				msg := resp.Status
-				if msg == "" {
-					msg = fmt.Sprintf("pushing %s...", resp.Digest[7:19])
-				}
-				bar = progress.NewBar(msg, resp.Total, resp.Completed)
+				bar = progress.NewBar(fmt.Sprintf("pushing %s...", resp.Digest[7:19]), resp.Total, resp.Completed)
 				bars[resp.Digest] = bar
 				p.Add(resp.Digest, bar)
 			}
@@ -916,6 +837,11 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
 }
 
 func ShowHandler(cmd *cobra.Command, args []string) error {
+	// Check if this is an image generation model
+	if imagegen.HasTensorLayers(args[0]) {
+		return imagegen.Show(args[0], os.Stdout)
+	}
+
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
@@ -1815,22 +1741,15 @@ func NewCLI() *cobra.Command {
 	rootCmd.Flags().BoolP("version", "v", false, "Show version information")
 
 	createCmd := &cobra.Command{
-		Use:   "create MODEL",
-		Short: "Create a model",
-		Args:  cobra.ExactArgs(1),
-		PreRunE: func(cmd *cobra.Command, args []string) error {
-			// Skip server check for experimental mode (writes directly to disk)
-			if experimental, _ := cmd.Flags().GetBool("experimental"); experimental {
-				return nil
-			}
-			return checkServerHeartbeat(cmd, args)
-		},
-		RunE: CreateHandler,
+		Use:     "create MODEL",
+		Short:   "Create a model",
+		Args:    cobra.ExactArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE:    CreateHandler,
 	}
 
 	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\")")
 	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_K_M)")
-	createCmd.Flags().Bool("experimental", false, "Enable experimental safetensors model creation")
 
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
@@ -1867,7 +1786,6 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Int("dimensions", 0, "Truncate output embeddings to specified dimension (embedding models only)")
 	runCmd.Flags().Bool("experimental", false, "Enable experimental agent loop with tools")
 	runCmd.Flags().Bool("experimental-yolo", false, "Skip all tool approval prompts (use with caution)")
-	runCmd.Flags().Bool("experimental-websearch", false, "Enable web search tool in experimental mode")
 
 	// Image generation flags (width, height, steps, seed, etc.)
 	imagegen.RegisterFlags(runCmd)

cmd/cmd_test.go

@@ -1547,79 +1547,6 @@ func TestRunOptions_Copy_ThinkValueVariants(t *testing.T) {
 	}
 }
 
-func TestShowInfoImageGen(t *testing.T) {
-	var b bytes.Buffer
-	err := showInfo(&api.ShowResponse{
-		Details: api.ModelDetails{
-			Family:            "ZImagePipeline",
-			ParameterSize:     "10.3B",
-			QuantizationLevel: "FP8",
-		},
-		Capabilities: []model.Capability{model.CapabilityImageGeneration},
-		Requires:     "0.14.0",
-	}, false, &b)
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	expect := "  Model\n" +
-		"    architecture    ZImagePipeline    \n" +
-		"    parameters      10.3B             \n" +
-		"    quantization    FP8               \n" +
-		"    requires        0.14.0            \n" +
-		"\n" +
-		"  Capabilities\n" +
-		"    image    \n" +
-		"\n"
-	if diff := cmp.Diff(expect, b.String()); diff != "" {
-		t.Errorf("unexpected output (-want +got):\n%s", diff)
-	}
-}
-
-func TestPushProgressMessage(t *testing.T) {
-	tests := []struct {
-		name    string
-		status  string
-		digest  string
-		wantMsg string
-	}{
-		{
-			name:    "uses status when provided",
-			status:  "uploading model",
-			digest:  "sha256:abc123456789def",
-			wantMsg: "uploading model",
-		},
-		{
-			name:    "falls back to digest when status empty",
-			status:  "",
-			digest:  "sha256:abc123456789def",
-			wantMsg: "pushing abc123456789...",
-		},
-		{
-			name:    "handles short digest gracefully",
-			status:  "",
-			digest:  "sha256:abc",
-			wantMsg: "pushing sha256:abc...",
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			msg := tt.status
-			if msg == "" {
-				if len(tt.digest) >= 19 {
-					msg = fmt.Sprintf("pushing %s...", tt.digest[7:19])
-				} else {
-					msg = fmt.Sprintf("pushing %s...", tt.digest)
-				}
-			}
-			if msg != tt.wantMsg {
-				t.Errorf("got %q, want %q", msg, tt.wantMsg)
-			}
-		})
-	}
-}
-
 func TestRunOptions_Copy_Independence(t *testing.T) {
 	// Test that modifications to original don't affect copy
 	originalThink := &api.ThinkValue{Value: "original"}

cmd/interactive.go

@@ -116,7 +116,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		Prompt:         ">>> ",
 		AltPrompt:      "... ",
 		Placeholder:    "Send a message (/? for help)",
-		AltPlaceholder: "Press Enter to send",
+		AltPlaceholder: `Use """ to end multi-line input`,
 	})
 	if err != nil {
 		return err

docs/api.md

@@ -16,7 +16,6 @@
 - [Generate Embeddings](#generate-embeddings)
 - [List Running Models](#list-running-models)
 - [Version](#version)
-- [Experimental: Generate an image](#generate-an-image-experimental)
 
 ## Conventions
 
@@ -1868,85 +1867,3 @@ curl http://localhost:11434/api/version
   "version": "0.5.1"
 }
 ```
-
-## Experimental Endpoints
-
-### Generate an image (Experimental)
-
-```
-POST /x/generate
-```
-
-> [!WARNING]
-> This endpoint is experimental and may change in future versions.
-
-Generate an image using an image generation model. This endpoint is specifically designed for diffusion-based image generation models.
-
-#### Parameters
-
-- `model`: (required) the [model name](#model-names) of an image generation model
-- `prompt`: the text prompt describing the image to generate
-
-Image generation parameters (optional):
-
-- `width`: width of the generated image in pixels (default: model-specific, typically 1024)
-- `height`: height of the generated image in pixels (default: model-specific, typically 1024)
-- `steps`: number of diffusion steps (default: model-specific)
-
-Other parameters:
-
-- `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
-
-#### Response
-
-The response is streamed as JSON objects showing generation progress:
-
-- `status`: describes the current phase (e.g., "generating image")
-- `total`: total number of steps
-- `completed`: number of completed steps
-- `done`: whether generation is complete
-
-The final response includes:
-
-- `images`: array of base64-encoded generated images
-- `total_duration`: time spent generating the image
-- `load_duration`: time spent loading the model
-
-#### Examples
-
-##### Request
-
-```shell
-curl http://localhost:11434/x/generate -d '{
-  "model": "flux",
-  "prompt": "a sunset over mountains",
-  "width": 1024,
-  "height": 768
-}'
-```
-
-##### Response (streaming)
-
-```json
-{
-  "model": "flux",
-  "created_at": "2024-01-15T10:30:00.000000Z",
-  "status": "generating image",
-  "completed": 5,
-  "total": 20,
-  "done": false
-}
-```
-
-##### Final Response
-
-```json
-{
-  "model": "flux",
-  "created_at": "2024-01-15T10:30:15.000000Z",
-  "images": ["iVBORw0KGgoAAAANSUhEUg..."],
-  "done": true,
-  "total_duration": 15000000000,
-  "load_duration": 2000000000
-}
-```

docs/api/anthropic-compatibility.mdx

@@ -21,7 +21,6 @@ ollama pull glm-4.7:cloud
 To use Ollama with tools that expect the Anthropic API (like Claude Code), set these environment variables:
 
 ```shell
-export ANTHROPIC_AUTH_TOKEN=ollama  # required but ignored
 export ANTHROPIC_BASE_URL=http://localhost:11434
 export ANTHROPIC_API_KEY=ollama  # required but ignored
 ```
@@ -248,13 +247,12 @@ curl -X POST http://localhost:11434/v1/messages \
 [Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend:
 
 ```shell
-ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
 ```
 
 Or set the environment variables in your shell profile:
 
 ```shell
-export ANTHROPIC_AUTH_TOKEN=ollama
 export ANTHROPIC_BASE_URL=http://localhost:11434
 export ANTHROPIC_API_KEY=ollama
 ```

docs/capabilities/web-search.mdx

@@ -110,7 +110,7 @@ More Ollama [Python example](https://github.com/ollama/ollama-python/blob/main/e
 import { Ollama } from "ollama";
 
 const client = new Ollama();
-const results = await client.webSearch("what is ollama?");
+const results = await client.webSearch({ query: "what is ollama?" });
 console.log(JSON.stringify(results, null, 2));
 ```
 
@@ -213,7 +213,7 @@ models](https://ollama.com/models)\n\nAvailable for macOS, Windows, and Linux',
 import { Ollama } from "ollama";
 
 const client = new Ollama();
-const fetchResult = await client.webFetch("https://ollama.com");
+const fetchResult = await client.webFetch({ url: "https://ollama.com" });
 console.log(JSON.stringify(fetchResult, null, 2));
 ```
 

docs/docs.json

@@ -111,9 +111,7 @@
               "/integrations/zed",
               "/integrations/roo-code",
               "/integrations/n8n",
-              "/integrations/xcode",
-              "/integrations/onyx",
-              "/integrations/marimo"
+              "/integrations/xcode"
             ]
           },
           {

docs/faq.mdx

@@ -22,7 +22,7 @@ Please refer to the [GPU docs](./gpu).
 
 ## How can I specify the context window size?
 
-By default, Ollama uses a context window size of 4096 tokens.
+By default, Ollama uses a context window size of 2048 tokens.
 
 This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use:
 

docs/integrations/claude-code.mdx

@@ -25,7 +25,6 @@ Claude Code connects to Ollama using the Anthropic-compatible API.
 1. Set the environment variables:
 
 ```shell
-export ANTHROPIC_AUTH_TOKEN=ollama
 export ANTHROPIC_BASE_URL=http://localhost:11434
 export ANTHROPIC_API_KEY=ollama
 ```
@@ -39,7 +38,7 @@ claude --model qwen3-coder
 Or run with environment variables inline:
 
 ```shell
-ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
 ```
 
 ## Connecting to ollama.com

docs/integrations/marimo.mdx

@@ -1,73 +0,0 @@
----
-title: marimo
----
-
-## Install
-
-Install [marimo](https://marimo.io). You can use `pip` or `uv` for this. You 
-can also use `uv` to create a sandboxed environment for marimo by running:
-
-```
-uvx marimo edit --sandbox notebook.py
-```
-
-## Usage with Ollama
-
-1. In marimo, go to the user settings and go to the AI tab. From here
-you can find and configure Ollama as an AI provider. For local use you
-would typically point the base url to `http://localhost:11434/v1`.
-
-<div style={{ display: 'flex', justifyContent: 'center' }}>
-  <img 
-    src="/images/marimo-settings.png" 
-    alt="Ollama settings in marimo"
-    width="50%"
-  />
-</div>
-
-2. Once the AI provider is set up, you can turn on/off specific AI models you'd like to access. 
-
-<div style={{ display: 'flex', justifyContent: 'center' }}>
-  <img 
-    src="/images/marimo-models.png" 
-    alt="Selecting an Ollama model"
-    width="50%"
-  />
-</div>
-
-3. You can also add a model to the list of available models by scrolling to the bottom and using the UI there. 
-
-<div style={{ display: 'flex', justifyContent: 'center' }}>
-  <img 
-    src="/images/marimo-add-model.png" 
-    alt="Adding a new Ollama model"
-    width="50%"
-  />
-</div>
-
-4. Once configured, you can now use Ollama for AI chats in marimo.
-
-<div style={{ display: 'flex', justifyContent: 'center' }}>
-  <img 
-    src="/images/marimo-chat.png" 
-    alt="Configure code completion"
-    width="50%"
-  />
-</div>
-
-4. Alternatively, you can now use Ollama for **inline code completion** in marimo. This can be configured in the "AI Features" tab. 
-
-<div style={{ display: 'flex', justifyContent: 'center' }}>
-  <img 
-    src="/images/marimo-code-completion.png" 
-    alt="Configure code completion"
-    width="50%"
-  />
-</div>
-
-
-## Connecting to ollama.com
-
-1. Sign in to ollama cloud via `ollama signin` 
-2. In the ollama model settings add a model that ollama hosts, like `gpt-oss:120b`.
-3. You can now refer to this model in marimo!

docs/integrations/onyx.mdx

@@ -1,63 +0,0 @@
----
-title: Onyx
----
-
-## Overview
-[Onyx](http://onyx.app/) is a self-hostable Chat UI that integrates with all Ollama models. Features include:
-- Creating custom Agents
-- Web search
-- Deep Research
-- RAG over uploaded documents and connected apps
-- Connectors to applications like Google Drive, Email, Slack, etc.
-- MCP and OpenAPI Actions support
-- Image generation
-- User/Groups management, RBAC, SSO, etc.
-
-Onyx can be deployed for single users or large organizations.
-
-## Install Onyx
-
-Deploy Onyx with the [quickstart guide](https://docs.onyx.app/deployment/getting_started/quickstart).
-
-<Info>
-Resourcing/scaling docs [here](https://docs.onyx.app/deployment/getting_started/resourcing).
-</Info>
-
-## Usage with Ollama 
-
-1. Login to your Onyx deployment (create an account first).
-<div style={{ display: 'flex', justifyContent: 'center' }}>
-  <img 
-    src="/images/onyx-login.png" 
-    alt="Onyx Login Page"
-    width="75%"
-  />
-</div>
-2. In the set-up process select `Ollama` as the LLM provider.
-<div style={{ display: 'flex', justifyContent: 'center' }}>
-  <img 
-    src="/images/onyx-ollama-llm.png" 
-    alt="Onyx Set Up Form"
-    width="75%"
-  />
-</div>
-3. Provide your **Ollama API URL** and select your models.
-<Note>If you're running Onyx in Docker, to access your computer's local network use `http://host.docker.internal` instead of `http://127.0.0.1`.</Note>
-<div style={{ display: 'flex', justifyContent: 'center' }}>
-  <img 
-    src="/images/onyx-ollama-form.png" 
-    alt="Selecting Ollama Models"
-    width="75%"
-  />
-</div>
-
-You can also easily connect up Onyx Cloud with the `Ollama Cloud` tab of the setup.
-
-## Send your first query
-<div style={{ display: 'flex', justifyContent: 'center' }}>
-  <img 
-    src="/images/onyx-query.png" 
-    alt="Onyx Query Example"
-    width="75%"
-  />
-</div>

docs/linux.mdx

@@ -1,5 +1,5 @@
 ---
-title: Linux
+title: "Linux"
 ---
 
 ## Install
@@ -13,15 +13,14 @@ curl -fsSL https://ollama.com/install.sh | sh
 ## Manual install
 
 <Note>
-  If you are upgrading from a prior version, you should remove the old libraries
-  with `sudo rm -rf /usr/lib/ollama` first.
+  If you are upgrading from a prior version, you should remove the old libraries with `sudo rm -rf /usr/lib/ollama` first.
 </Note>
 
 Download and extract the package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst \
-    | sudo tar x -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz \
+    | sudo tar zx -C /usr
 ```
 
 Start Ollama:
@@ -41,8 +40,8 @@ ollama -v
 If you have an AMD GPU, also download and extract the additional ROCm package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tar.zst \
-    | sudo tar x -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tgz \
+    | sudo tar zx -C /usr
 ```
 
 ### ARM64 install
@@ -50,8 +49,8 @@ curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tar.zst \
 Download and extract the ARM64-specific package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-arm64.tar.zst \
-    | sudo tar x -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-arm64.tgz \
+    | sudo tar zx -C /usr
 ```
 
 ### Adding Ollama as a startup service (recommended)
@@ -113,11 +112,7 @@ sudo systemctl status ollama
 ```
 
 <Note>
-  While AMD has contributed the `amdgpu` driver upstream to the official linux
-  kernel source, the version is older and may not support all ROCm features. We
-  recommend you install the latest driver from
-  https://www.amd.com/en/support/linux-drivers for best support of your Radeon
-  GPU.
+  While AMD has contributed the `amdgpu` driver upstream to the official linux kernel source, the version is older and may not support all ROCm features. We recommend you install the latest driver from https://www.amd.com/en/support/linux-drivers for best support of your Radeon GPU.
 </Note>
 
 ## Customizing
@@ -146,8 +141,8 @@ curl -fsSL https://ollama.com/install.sh | sh
 Or by re-downloading Ollama:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst \
-    | sudo tar x -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz \
+    | sudo tar zx -C /usr
 ```
 
 ## Installing specific versions
@@ -196,4 +191,4 @@ Remove the downloaded models and Ollama service user and group:
 sudo userdel ollama
 sudo groupdel ollama
 sudo rm -r /usr/share/ollama
-```
+```

docs/troubleshooting.md

@@ -0,0 +1,3 @@
+# Troubleshooting
+
+For troubleshooting, see [https://docs.ollama.com/troubleshooting](https://docs.ollama.com/troubleshooting)

integration/tools_test.go

@@ -131,7 +131,7 @@ func TestAPIToolCalling(t *testing.T) {
 					t.Errorf("unexpected tool called: got %q want %q", lastToolCall.Function.Name, "get_weather")
 				}
 
-				if _, ok := lastToolCall.Function.Arguments.Get("location"); !ok {
+				if _, ok := lastToolCall.Function.Arguments["location"]; !ok {
 					t.Errorf("expected tool arguments to include 'location', got: %s", lastToolCall.Function.Arguments.String())
 				}
 			case <-ctx.Done():

llama/README.md

@@ -14,25 +14,28 @@ make -f Makefile.sync apply-patches
 
 ### Updating Base Commit
 
-**Pin to new base commit**
+To update to a new base commit:
 
-To change the base commit, update `FETCH_HEAD` in Makefile.sync.
+1. **Update FETCH_HEAD** in `Makefile.sync` to the new commit hash.
 
-When updating to a newer base commit, the existing patches may not apply cleanly and require manual merge resolution.
+2. **Check for upstreamed patches**: Before applying, review if any patches have been merged upstream. Remove those patches from `./patches/` to avoid conflicts.
 
-Start by applying the patches. If any of the patches have conflicts, the `git am` will stop at the first failure.
+3. **Apply patches**:
+   ```shell
+   make -f Makefile.sync apply-patches
+   ```
 
-```shell
-make -f Makefile.sync apply-patches
-```
+4. **Resolve conflicts** (if any): When `git am` fails on a patch:
+   - Fix conflicts in `./vendor/`
+   - Stage the resolved files: `git -C llama/vendor add <file>`
+   - Continue: `git -C llama/vendor am --continue`
+   - Re-run: `make -f Makefile.sync apply-patches`
+   - Repeat until all patches are applied.
 
-If there are conflicts, you will see an error message. Resolve the conflicts in `./vendor/`, and continue the patch series with `git am --continue` and rerun `make -f Makefile.sync apply-patches`. Repeat until all patches are successfully applied.
-
-Once all patches are applied, commit the changes to the tracking repository.
-
-```shell
-make -f Makefile.sync format-patches sync
-```
+5. **Regenerate patches and sync**:
+   ```shell
+   make -f Makefile.sync format-patches sync
+   ```
 
 ### Generating Patches
 

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "ec98e2002";
+char const *LLAMA_COMMIT = "b1377188784f9aea26b8abde56d4aee8c733eec7";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/common/common.cpp

@@ -251,7 +251,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
         case GGML_SCHED_PRIO_REALTIME: p = -20; break;
     }
 
-    if (!setpriority(PRIO_PROCESS, 0, p)) {
+    if (setpriority(PRIO_PROCESS, 0, p) != 0) {
         LOG_WRN("failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno);
         return false;
     }
@@ -1078,12 +1078,15 @@ struct common_init_result::impl {
     impl() = default;
     ~impl() = default;
 
+    // note: the order in which model, context, etc. are declared matters because their destructors will be called bottom-to-top
+
     llama_model_ptr   model;
     llama_context_ptr context;
 
     std::vector<llama_adapter_lora_ptr> lora;
 
     std::vector<common_sampler_ptr> samplers;
+    std::vector<llama_sampler_seq_config> samplers_seq_config;
 };
 
 common_init_result::common_init_result(common_params & params) :
@@ -1092,9 +1095,9 @@ common_init_result::common_init_result(common_params & params) :
     auto cparams = common_context_params_to_llama(params);
 
     if (params.fit_params) {
-        LOG_INF("%s: fitting params to device memory, to report bugs during this step use -fit off (or --verbose if you can't)\n", __func__);
+        LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__);
         llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
-            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target, params.fit_params_min_ctx,
+            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target.data(), params.fit_params_min_ctx,
             params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 
@@ -1107,6 +1110,25 @@ common_init_result::common_init_result(common_params & params) :
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
+    // load and optionally apply lora adapters (must be loaded before context creation)
+    for (auto & la : params.lora_adapters) {
+        llama_adapter_lora_ptr lora;
+        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
+        if (lora == nullptr) {
+            LOG_ERR("%s: failed to load lora adapter '%s'\n", __func__, la.path.c_str());
+            pimpl->model.reset(model);
+            return;
+        }
+
+        char buf[1024];
+        la.ptr = lora.get();
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
+        la.task_name = buf;
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
+        la.prompt_prefix = buf;
+        pimpl->lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
+    }
+
     // updates params.sampling
     // TODO: fix naming
     common_init_sampler_from_model(model, params.sampling);
@@ -1141,10 +1163,19 @@ common_init_result::common_init_result(common_params & params) :
     //    params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
     //}
 
+    // init the backend samplers as part of the context creation
     pimpl->samplers.resize(cparams.n_seq_max);
+    pimpl->samplers_seq_config.resize(cparams.n_seq_max);
 
     for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
         pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
+        pimpl->samplers_seq_config[i] = { i, common_sampler_get(pimpl->samplers[i].get()) };
+    }
+
+    // TODO: temporarily gated behind a flag
+    if (params.sampling.backend_sampling) {
+        cparams.samplers   = pimpl->samplers_seq_config.data();
+        cparams.n_samplers = pimpl->samplers_seq_config.size();
     }
 
     llama_context * lctx = llama_init_from_model(model, cparams);
@@ -1168,6 +1199,12 @@ common_sampler * common_init_result::sampler(llama_seq_id seq_id) {
     return pimpl->samplers[seq_id].get();
 }
 
+void common_init_result::reset_samplers() {
+    for (int i = 0; i < (int) pimpl->samplers.size(); ++i) {
+        llama_sampler_reset(common_sampler_get(pimpl->samplers[i].get()));
+    }
+}
+
 std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
     return pimpl->lora;
 }
@@ -1243,24 +1280,6 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         }
     }
 
-    // load and optionally apply lora adapters
-    for (auto & la : params.lora_adapters) {
-        llama_adapter_lora_ptr lora;
-        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
-        if (lora == nullptr) {
-            LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
-            return res;
-        }
-
-        char buf[1024];
-        la.ptr = lora.get();
-        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
-        la.task_name = buf;
-        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
-        la.prompt_prefix = buf;
-        res->lora().emplace_back(std::move(lora)); // copy to list of loaded adapters
-    }
-
     if (!params.lora_init_without_apply) {
         common_set_adapter_lora(lctx, params.lora_adapters);
     }
@@ -1301,6 +1320,9 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
         llama_set_warmup(lctx, false);
+
+        // reset samplers to reset RNG state after warmup to the seeded state
+        res->reset_samplers();
     }
 
     return res;
@@ -1339,14 +1361,12 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.devices = params.devices.data();
     }
 
-    if (params.n_gpu_layers != -1) {
-        mparams.n_gpu_layers = params.n_gpu_layers;
-    }
-
+    mparams.n_gpu_layers    = params.n_gpu_layers;
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
+    mparams.use_direct_io   = params.use_direct_io;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
     mparams.use_extra_bufts = !params.no_extra_bufts;

llama/llama.cpp/common/common.h

@@ -80,6 +80,7 @@ int32_t cpu_get_num_math();
 //
 
 enum llama_example {
+    LLAMA_EXAMPLE_DEBUG,
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
     LLAMA_EXAMPLE_COMPLETION,
@@ -216,6 +217,8 @@ struct common_params_sampling {
     std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
     std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
+    bool backend_sampling = false;
+
     bool has_logit_bias() const {
         return !logit_bias.empty();
     }
@@ -329,12 +332,14 @@ struct common_params {
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM (-1 - use default)
-    int32_t main_gpu           = 0;                // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]  = {0};              // how split tensors should be distributed across GPUs
-    bool    fit_params         = true;             // whether to fit unset model/context parameters to free device memory
-    size_t  fit_params_target  = 1024 * 1024*1024; // margin per device in bytes for fitting parameters to free memory
-    int32_t fit_params_min_ctx = 4096;             // minimum context size to set when trying to reduce memory use
+    int32_t n_gpu_layers       = -1;   // number of layers to store in VRAM, -1 is auto, <= -2 is all
+    int32_t main_gpu           = 0;    // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]  = {0};  // how split tensors should be distributed across GPUs
+    bool    fit_params         = true; // whether to fit unset model/context parameters to free device memory
+    int32_t fit_params_min_ctx = 4096; // minimum context size to set when trying to reduce memory use
+
+    // margin per device in bytes for fitting parameters to free memory:
+    std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 
@@ -370,6 +375,11 @@ struct common_params {
     std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
     std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
 
+    // llama-debug specific options
+    std::string logits_output_dir = "data"; // directory for saving logits output files                     // NOLINT
+    bool        save_logits       = false;  // whether to save logits to files                              // NOLINT
+    std::vector<std::string> tensor_filter; // filter tensor names for debug output (regex)                 // NOLINT
+
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
@@ -420,7 +430,8 @@ struct common_params {
     bool kv_unified        = false; // enable unified KV cache
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
-    bool use_mmap          = true;  // use mmap for faster loads
+    bool use_mmap          = true;  // enable mmap to use filesystem cache
+    bool use_direct_io     = true;  // read from disk without buffering for faster model loading
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
@@ -475,7 +486,8 @@ struct common_params {
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     int reasoning_budget = -1;
-    bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
+    bool prefill_assistant = true; // if true, any trailing assistant message will be prefilled into the response
+    int sleep_idle_seconds = -1;   // if >0, server will sleep after this many seconds of idle time
 
     std::vector<std::string> api_keys;
 
@@ -484,8 +496,11 @@ struct common_params {
 
     std::map<std::string, std::string> default_template_kwargs;
 
+    // webui configs
+    bool webui = true;
+    std::string webui_config_json;
+
     // "advanced" endpoints are disabled by default for better security
-    bool webui            = true;
     bool endpoint_slots   = true;
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
@@ -685,7 +700,9 @@ struct common_init_result {
 
     llama_model * model();
     llama_context * context();
+
     common_sampler * sampler(llama_seq_id seq_id);
+    void reset_samplers();
 
     std::vector<llama_adapter_lora_ptr> & lora();
 

llama/llama.cpp/common/sampling.cpp

@@ -104,10 +104,9 @@ struct ring_buffer {
 struct common_sampler {
     common_params_sampling params;
 
+    struct llama_sampler * grmr;
     struct llama_sampler * chain;
 
-    bool grammar;
-
     ring_buffer<llama_token> prev;
 
     std::vector<llama_token_data> cur;
@@ -121,17 +120,34 @@ struct common_sampler {
     }
 
     void set_logits(struct llama_context * ctx, int idx) {
-        const auto * logits = llama_get_logits_ith(ctx, idx);
+        const float *       sampled_probs  = llama_get_sampled_probs_ith     (ctx, idx);
+        const float *       sampled_logits = llama_get_sampled_logits_ith    (ctx, idx);
+        const llama_token * sampled_ids    = llama_get_sampled_candidates_ith(ctx, idx);
 
         const llama_model * model = llama_get_model(ctx);
         const llama_vocab * vocab = llama_model_get_vocab(model);
 
         const int n_vocab = llama_vocab_n_tokens(vocab);
 
-        cur.resize(n_vocab);
-
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
+        if (sampled_probs) {
+            const uint32_t sampled_probs_count = llama_get_sampled_probs_count_ith(ctx, idx);
+            cur.resize(sampled_probs_count);
+            for (uint32_t i = 0; i < sampled_probs_count; ++i) {
+                cur[i] = llama_token_data{sampled_ids[i], sampled_logits[i], sampled_probs[i]};
+            }
+        } else if (sampled_logits) {
+            const uint32_t sampled_logits_count = llama_get_sampled_logits_count_ith(ctx, idx);
+            cur.resize(sampled_logits_count);
+            for (uint32_t i = 0; i < sampled_logits_count; i++) {
+                cur[i] = llama_token_data{sampled_ids[i], sampled_logits[i], 0.0f};
+            }
+        } else {
+            const auto * logits = llama_get_logits_ith(ctx, idx);
+            GGML_ASSERT(logits != nullptr);
+            cur.resize(n_vocab);
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
+            }
         }
 
         cur_p = { cur.data(), cur.size(), -1, false };
@@ -160,45 +176,50 @@ std::string common_params_sampling::print() const {
     return std::string(result);
 }
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
+struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params) {
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
 
     lparams.no_perf = params.no_perf;
 
+    llama_sampler * grmr = nullptr;
     llama_sampler * chain = llama_sampler_chain_init(lparams);
 
-    bool grammar = false;
     std::vector<llama_sampler *> samplers;
 
     if (params.grammar.compare(0, 11, "%llguidance") == 0) {
 #ifdef LLAMA_USE_LLGUIDANCE
-        samplers.push_back(llama_sampler_init_llg(vocab, "lark", params.grammar.c_str()));
-        grammar = true;
+        grmr = llama_sampler_init_llg(vocab, "lark", params.grammar.c_str());
 #else
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
     } else {
         std::vector<std::string> trigger_patterns;
-        std::vector<std::string> patterns_anywhere;
         std::vector<llama_token> trigger_tokens;
         for (const auto & trigger : params.grammar_triggers) {
             switch (trigger.type) {
                 case COMMON_GRAMMAR_TRIGGER_TYPE_WORD:
                 {
                     const auto & word = trigger.value;
-                    patterns_anywhere.push_back(regex_escape(word));
+                    trigger_patterns.push_back(regex_escape(word));
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN:
                 {
-                    patterns_anywhere.push_back(trigger.value);
+                    trigger_patterns.push_back(trigger.value);
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL:
                 {
-                    trigger_patterns.push_back(trigger.value);
+                    const auto & pattern = trigger.value;
+                    std::string anchored = "^$";
+                    if (!pattern.empty()) {
+                        anchored = (pattern.front() != '^' ? "^" : "")
+                            + pattern
+                            + (pattern.back() != '$' ? "$" : "");
+                    }
+                    trigger_patterns.push_back(anchored);
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN:
@@ -212,10 +233,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
             }
         }
 
-        if (!patterns_anywhere.empty()) {
-            trigger_patterns.push_back("^[\\s\\S]*?(" + string_join(patterns_anywhere, "|") + ")[\\s\\S]*");
-        }
-
         std::vector<const char *> trigger_patterns_c;
         trigger_patterns_c.reserve(trigger_patterns.size());
         for (const auto & regex : trigger_patterns) {
@@ -224,15 +241,12 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
         if (!params.grammar.empty()) {
              if (params.grammar_lazy) {
-                 samplers.push_back(
-                         llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
-                             trigger_patterns_c.data(), trigger_patterns_c.size(),
-                             trigger_tokens.data(),     trigger_tokens.size()));
+                 grmr = llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
+                         trigger_patterns_c.data(), trigger_patterns_c.size(),
+                         trigger_tokens.data(), trigger_tokens.size());
              } else {
-                 samplers.push_back(llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root"));
+                 grmr = llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
              }
-
-             grammar = true;
         }
     }
 
@@ -301,10 +315,16 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
         llama_sampler_chain_add(chain, smpl);
     }
 
+    if (grmr && params.backend_sampling) {
+        LOG_WRN("%s: backend sampling is not compatible with grammar, disabling\n", __func__);
+
+        params.backend_sampling = false;
+    }
+
     auto * result = new common_sampler {
         /* .params  = */ params,
+        /* .grmr    = */ grmr,
         /* .chain   = */ chain,
-        /* .grammar = */ grammar,
         /* .prev    = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
         /* .cur     = */ {},
         /* .cur_p   = */ {},
@@ -315,6 +335,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
 void common_sampler_free(struct common_sampler * gsmpl) {
     if (gsmpl) {
+        llama_sampler_free(gsmpl->grmr);
         llama_sampler_free(gsmpl->chain);
 
         delete gsmpl;
@@ -324,25 +345,12 @@ 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 (gsmpl->grammar) {
-        const int n_smpl = llama_sampler_chain_n(gsmpl->chain);
-
-        for (int i = 0; i < n_smpl; i++) {
-            auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
-
-            // the grammar sampler is always the first one
-            if (i == 0) {
-                if (accept_grammar) {
-                    llama_sampler_accept(smpl, token);
-                }
-            } else {
-                llama_sampler_accept(smpl, token);
-            }
-        }
-    } else {
-        llama_sampler_accept(gsmpl->chain, token);
+    if (gsmpl->grmr && accept_grammar) {
+        llama_sampler_accept(gsmpl->grmr, token);
     }
 
+    llama_sampler_accept(gsmpl->chain, token);
+
     gsmpl->prev.push_back(token);
 }
 
@@ -353,8 +361,8 @@ void common_sampler_reset(struct common_sampler * gsmpl) {
 struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
     return new common_sampler {
         /* .params  = */ gsmpl->params,
+        /* .grmr    = */ llama_sampler_clone(gsmpl->grmr),
         /* .chain   = */ llama_sampler_clone(gsmpl->chain),
-        /* .grammar = */ gsmpl->grammar,
         /* .prev    = */ gsmpl->prev,
         /* .cur     = */ gsmpl->cur,
         /* .cur_p   = */ gsmpl->cur_p,
@@ -410,7 +418,7 @@ struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl) {
     return gsmpl->chain;
 }
 
-llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx) {
+llama_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
@@ -418,11 +426,61 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
 
     llama_token id = LLAMA_TOKEN_NULL;
 
+    auto & grmr  = gsmpl->grmr;
     auto & chain = gsmpl->chain;
     auto & cur_p = gsmpl->cur_p; // initialized by set_logits
 
+    // Check if a backend sampler has already sampled a token in which case we
+    // return that token id directly.
+    {
+        id = llama_get_sampled_token_ith(ctx, idx);
+
+        if (id != LLAMA_TOKEN_NULL) {
+            LOG_DBG("%s: Backend sampler selected token: '%d'. Will not run any CPU samplers\n", __func__, id);
+
+            GGML_ASSERT(!gsmpl->grmr && "using grammar in combination with backend sampling is not supported");
+
+            // TODO: simplify
+            gsmpl->cur.resize(1);
+            gsmpl->cur[0] = { id, 0.0f, 1.0f };
+            cur_p = { gsmpl->cur.data(), gsmpl->cur.size(), 0, true };
+
+            return id;
+        }
+    }
+
     gsmpl->set_logits(ctx, idx);
 
+    if (grammar_first) {
+        llama_sampler_apply(grmr, &cur_p);
+    }
+
+    llama_sampler_apply(chain, &cur_p);
+
+    id = cur_p.data[cur_p.selected].id;
+
+    if (grammar_first) {
+        return id;
+    }
+
+    // check if it the sampled token fits the grammar (grammar-based rejection sampling)
+    {
+        llama_token_data       single_token_data       = { id, 1.0f, 0.0f };
+        llama_token_data_array single_token_data_array = { &single_token_data, 1, -1, false };
+
+        llama_sampler_apply(grmr, &single_token_data_array);
+
+        const bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
+        if (is_valid) {
+            return id;
+        }
+    }
+
+    // resampling:
+    // if the token is not valid, sample again, but first apply the grammar sampler and then the sampling chain
+    gsmpl->set_logits(ctx, idx);
+
+    llama_sampler_apply(grmr,  &cur_p);
     llama_sampler_apply(chain, &cur_p);
 
     GGML_ASSERT(cur_p.selected != -1 && "no selected token during sampling - check your sampling configuration");
@@ -432,7 +490,7 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
     return id;
 }
 
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft) {
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first) {
     GGML_ASSERT(idxs.size() == draft.size() + 1 && "idxs.size() must be draft.size() + 1");
 
     std::vector<llama_token> result;
@@ -440,7 +498,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
 
     size_t i = 0;
     for (; i < draft.size(); i++) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
 
         common_sampler_accept(gsmpl, id, true);
 
@@ -452,7 +510,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
     }
 
     if (i == draft.size()) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
 
         common_sampler_accept(gsmpl, id, true);
 
@@ -462,13 +520,13 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
     return result;
 }
 
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft) {
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first) {
     std::vector<int> idxs(draft.size() + 1);
     for (size_t i = 0; i < idxs.size(); ++i) {
         idxs[i] = i;
     }
 
-    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft);
+    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft, grammar_first);
 }
 
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {

llama/llama.cpp/common/sampling.h

@@ -36,7 +36,8 @@ struct common_sampler;
 
 // llama_sampler API overloads
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params);
+// note: can mutate params in some cases
+struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params);
 
 void common_sampler_free(struct common_sampler * gsmpl);
 
@@ -48,6 +49,7 @@ struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
 // arguments can be nullptr to skip printing
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl);
 
+// get the underlying llama_sampler_chain
 struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl);
 
 // extended sampling implementation:
@@ -57,7 +59,10 @@ struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl);
 // - check if the token fits the grammar (if any)
 // - if not: resample by first applying the grammar constraints and then sampling again (slower path)
 //
-llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx);
+// if grammar_first is true, the grammar is applied before the samplers (slower)
+// useful in cases where all the resulting candidates (not just the sampled one) must fit the grammar
+//
+llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
 
 // generalized version of common_sampler_sample
 //
@@ -75,10 +80,10 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
 //
 // returns at least 1 token, up to idxs.size()
 //
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft);
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first = false);
 
 // assume idxs == [ 0, 1, 2, ..., draft.size() ]
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft);
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first = false);
 
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 

llama/llama.cpp/include/llama.h

@@ -286,7 +286,7 @@ extern "C" {
         // NULL-terminated list of buffer types to use for tensors that match a pattern
         const struct llama_model_tensor_buft_override * tensor_buft_overrides;
 
-        int32_t n_gpu_layers; // number of layers to store in VRAM
+        int32_t n_gpu_layers; // number of layers to store in VRAM, a negative value means all layers
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
         // the GPU that is used for the entire model when split_mode is LLAMA_SPLIT_MODE_NONE
@@ -309,6 +309,7 @@ extern "C" {
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool vocab_only;      // only load the vocabulary, no weights
         bool use_mmap;        // use mmap if possible
+        bool use_direct_io;   // use direct io, takes precedence over use_mmap
         bool use_mlock;       // force system to keep model in RAM
         bool check_tensors;   // validate model tensor data
         bool use_extra_bufts; // use extra buffer types (used for weight repacking)
@@ -316,6 +317,11 @@ extern "C" {
         bool no_alloc;        // only load metadata and simulate memory allocations
     };
 
+    struct llama_sampler_seq_config {
+        llama_seq_id           seq_id;
+        struct llama_sampler * sampler;
+    };
+
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
     //       https://github.com/ggml-org/llama.cpp/pull/7544
     struct llama_context_params {
@@ -364,6 +370,12 @@ extern "C" {
         bool kv_unified;  // use a unified buffer across the input sequences when computing the attention
                           // try to disable when n_seq_max > 1 for improved performance when the sequences do not share a large prefix
                           // ref: https://github.com/ggml-org/llama.cpp/pull/14363
+
+        // [EXPERIMENTAL]
+        // backend sampler chain configuration (make sure the caller keeps the sampler chains alive)
+        // note: the samplers must be sampler chains (i.e. use llama_sampler_chain_init)
+        struct llama_sampler_seq_config * samplers;
+        size_t                            n_samplers;
     };
 
     // model quantization parameters
@@ -467,16 +479,23 @@ extern "C" {
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 
+    enum llama_params_fit_status {
+        LLAMA_PARAMS_FIT_STATUS_SUCCESS = 0, // found allocations that are projected to fit
+        LLAMA_PARAMS_FIT_STATUS_FAILURE = 1, // could not find allocations that are projected to fit
+        LLAMA_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occured, e.g. because no model could be found at the specified path
+    };
+
     // fits mparams and cparams to free device memory (assumes system memory is unlimited)
-    // returns true if the parameters could be successfully modified to fit device memory
-    // this function is NOT thread safe because it modifies the global llama logger state
-    LLAMA_API bool llama_params_fit(
+    //   - returns true if the parameters could be successfully modified to fit device memory
+    //   - this function is NOT thread safe because it modifies the global llama logger state
+    //   - only parameters that have the same value as in llama_default_model_params are modified
+    LLAMA_API enum llama_params_fit_status llama_params_fit(
                                    const char   * path_model,
                     struct llama_model_params   * mparams,
                     struct llama_context_params * cparams,
                                           float * tensor_split,          // writable buffer for tensor split, needs at least llama_max_devices elements
         struct llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
-                                         size_t   margin,                // margin of memory to leave per device in bytes
+                                         size_t * margins,               // margins of memory to leave per device in bytes
                                        uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
                             enum ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
 
@@ -517,6 +536,7 @@ 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_embd_out (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);
@@ -600,6 +620,8 @@ extern "C" {
     //
 
     // Load a LoRA adapter from file
+    // The adapter is valid as long as the associated model is not freed
+    // All adapters must be loaded before context creation
     LLAMA_API struct llama_adapter_lora * llama_adapter_lora_init(
             struct llama_model * model,
             const char * path_lora);
@@ -983,6 +1005,32 @@ extern "C" {
     // otherwise: float[n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);
 
+    //
+    // backend sampling API [EXPERIMENTAL]
+    // note: use only if the llama_context was created with at least one llama_sampler_seq_config
+    //
+
+    // Get the backend sampled token for the ith token.
+    // Returns LLAMA_TOKEN_NULL if no token was sampled.
+    LLAMA_API llama_token llama_get_sampled_token_ith(struct llama_context * ctx, int32_t i);
+
+    // Get the backend sampled probabilites for the ith token
+    // The index matches llama_get_sampled_token_ith().
+    // Returns NULL if no probabilites were generated.
+    LLAMA_API float *  llama_get_sampled_probs_ith      (struct llama_context * ctx, int32_t i);
+    LLAMA_API uint32_t llama_get_sampled_probs_count_ith(struct llama_context * ctx, int32_t i);
+
+    // Get the backend sampled logits for the ith token
+    // Returns NULL if no logits were sampled.
+    LLAMA_API float *  llama_get_sampled_logits_ith      (struct llama_context * ctx, int32_t i);
+    LLAMA_API uint32_t llama_get_sampled_logits_count_ith(struct llama_context * ctx, int32_t i);
+
+    // Get the backend sampled candidates (token ids) for the ith token
+    // These are needed to map probability/logit indices to vocab token ids.
+    // Returns NULL if no candidates were sampled.
+    LLAMA_API llama_token * llama_get_sampled_candidates_ith      (struct llama_context * ctx, int32_t i);
+    LLAMA_API uint32_t      llama_get_sampled_candidates_count_ith(struct llama_context * ctx, int32_t i);
+
     //
     // Vocab
     //
@@ -1154,11 +1202,16 @@ extern "C" {
     //
     //    llama_sampler_free(smpl);
     //
-    // TODO: In the future, llama_sampler will be utilized to offload the sampling to the backends (e.g. GPU).
-    //
 
     typedef void * llama_sampler_context_t;
 
+    struct llama_sampler_data {
+        struct ggml_tensor * logits;
+        struct ggml_tensor * probs;
+        struct ggml_tensor * sampled;
+        struct ggml_tensor * candidates;
+    };
+
     // user code can implement the interface below in order to create custom llama_sampler
     struct llama_sampler_i {
         const char *           (*name)  (const struct llama_sampler * smpl);                                 // can be NULL
@@ -1168,17 +1221,45 @@ extern "C" {
         struct llama_sampler * (*clone) (const struct llama_sampler * smpl);                                 // can be NULL if ctx is NULL
         void                   (*free)  (      struct llama_sampler * smpl);                                 // can be NULL if ctx is NULL
 
-        // TODO: API for internal libllama usage for appending the sampling to an existing ggml_cgraph
-        //void (*apply_ggml) (struct llama_sampler * smpl, ...);
+        // [EXPERIMENTAL]
+        // backend sampling interface:
+
+        // return true if the backend supports all ops needed by the sampler
+        // note: call once per sampler
+        bool (*backend_init)(struct llama_sampler * smpl, ggml_backend_buffer_type_t buft);
+
+        // call after .backend_apply()
+        void (*backend_accept)(
+                struct llama_sampler * smpl,
+                struct ggml_context  * ctx,
+                struct ggml_cgraph   * gf,
+                struct ggml_tensor   * selected_token);
+
+        // call after .backend_init()
+        void (*backend_apply)(
+                struct llama_sampler      * smpl,
+                struct ggml_context       * ctx,
+                struct ggml_cgraph        * gf,
+                struct llama_sampler_data * data);
+
+        // called before graph execution to set inputs for the current ubatch
+        void (*backend_set_input)(struct llama_sampler * smpl);
     };
 
     struct llama_sampler {
-        const struct llama_sampler_i * iface;
-        llama_sampler_context_t        ctx;
+        struct llama_sampler_i * iface;
+
+        llama_sampler_context_t ctx;
     };
 
+    // [EXPERIMENTAL]
+    // attach a sampler to the context
+    // note: prefer initializing the context with llama_context_params.samplers when possible
+    // note: changing the samplers of a context can cause graph reallocations and degraded performance
+    LLAMA_API bool llama_set_sampler(struct llama_context * ctx, llama_seq_id seq_id, struct llama_sampler * smpl);
+
     // mirror of llama_sampler_i:
-    LLAMA_API struct llama_sampler * llama_sampler_init  (const struct llama_sampler_i * iface, llama_sampler_context_t ctx);
+    LLAMA_API struct llama_sampler * llama_sampler_init  (      struct llama_sampler_i * iface, llama_sampler_context_t ctx);
     LLAMA_API const char *           llama_sampler_name  (const struct llama_sampler * smpl);
     LLAMA_API void                   llama_sampler_accept(      struct llama_sampler * smpl, llama_token token);
     LLAMA_API void                   llama_sampler_apply (      struct llama_sampler * smpl, llama_token_data_array * cur_p);
@@ -1194,7 +1275,15 @@ extern "C" {
 
     // important: takes ownership of the sampler object and will free it when llama_sampler_free is called
     LLAMA_API void                   llama_sampler_chain_add(      struct llama_sampler * chain, struct llama_sampler * smpl);
-    LLAMA_API struct llama_sampler * llama_sampler_chain_get(const struct llama_sampler * chain, int32_t i);
+
+    // return NULL if:
+    //   - the sampler is NULL
+    //   - the sampler is not a llama_sampler_chain
+    //   - the index is out of bounds, unless i == -1
+    //   - if i == -1, returns the chain itself (can be used to check if the sampler is a chain)
+    LLAMA_API struct llama_sampler * llama_sampler_chain_get(      struct llama_sampler * chain, int32_t i);
+
+    // the total number of samplers in the chain
     LLAMA_API int                    llama_sampler_chain_n  (const struct llama_sampler * chain);
 
     // after removing a sampler, the chain will no longer own it, and it will not be freed when the chain is freed
@@ -1203,7 +1292,9 @@ extern "C" {
     // available samplers:
 
     LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
-    LLAMA_API struct llama_sampler * llama_sampler_init_dist  (uint32_t seed);
+
+    /// seed == LLAMA_DEFAULT_SEED to use a random seed.
+    LLAMA_API struct llama_sampler * llama_sampler_init_dist(uint32_t seed);
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     /// Setting k <= 0 makes this a noop

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

@@ -146,9 +146,11 @@ llama_adapter_lora_weight * llama_adapter_lora::get_weight(ggml_tensor * w) {
     return nullptr;
 }
 
-static void llama_adapter_lora_init_impl(llama_model & model, const char * path_lora, llama_adapter_lora & adapter) {
+static void llama_adapter_lora_init_impl(const char * path_lora, llama_adapter_lora & adapter) {
     LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
 
+    llama_model & model = adapter.model;
+
     ggml_context * ctx_init;
     gguf_init_params meta_gguf_params = {
         /* .no_alloc = */ true,
@@ -411,14 +413,17 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
         }
     }
 
+    // update number of nodes used
+    model.n_lora_nodes += adapter.get_n_nodes();
+
     LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
 }
 
 llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * path_lora) {
-    llama_adapter_lora * adapter = new llama_adapter_lora();
+    llama_adapter_lora * adapter = new llama_adapter_lora(*model);
 
     try {
-        llama_adapter_lora_init_impl(*model, path_lora, *adapter);
+        llama_adapter_lora_init_impl(path_lora, *adapter);
         return adapter;
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
@@ -469,6 +474,10 @@ int32_t llama_adapter_meta_val_str_by_index(const llama_adapter_lora * adapter,
 }
 
 void llama_adapter_lora_free(llama_adapter_lora * adapter) {
+    // update number of nodes used
+    GGML_ASSERT(adapter->model.n_lora_nodes >= adapter->get_n_nodes());
+    adapter->model.n_lora_nodes -= adapter->get_n_nodes();
+
     delete adapter;
 }
 

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

@@ -59,6 +59,8 @@ struct llama_adapter_lora_weight {
 };
 
 struct llama_adapter_lora {
+    llama_model & model;
+
     // map tensor name to lora_a_b
     std::unordered_map<std::string, llama_adapter_lora_weight> ab_map;
 
@@ -73,10 +75,14 @@ struct llama_adapter_lora {
     // activated lora (aLoRA)
     std::vector<llama_token> alora_invocation_tokens;
 
-    llama_adapter_lora() = default;
+    llama_adapter_lora(llama_model & model) : model(model) {}
     ~llama_adapter_lora() = default;
 
     llama_adapter_lora_weight * get_weight(ggml_tensor * w);
+
+    uint32_t get_n_nodes() const {
+        return ab_map.size() * 6u; // a, b, scale, add, 2 x mul_mat
+    }
 };
 
 using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;

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

@@ -20,6 +20,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_STARCODER,        "starcoder"        },
     { LLM_ARCH_REFACT,           "refact"           },
     { LLM_ARCH_BERT,             "bert"             },
+    { LLM_ARCH_MODERN_BERT,      "modern-bert"      },
     { LLM_ARCH_NOMIC_BERT,       "nomic-bert"       },
     { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
     { LLM_ARCH_NEO_BERT,         "neo-bert"         },
@@ -41,6 +42,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_PHIMOE,           "phimoe"           },
     { LLM_ARCH_PLAMO,            "plamo"            },
     { LLM_ARCH_PLAMO2,           "plamo2"           },
+    { LLM_ARCH_PLAMO3,           "plamo3"           },
     { LLM_ARCH_CODESHELL,        "codeshell"        },
     { LLM_ARCH_ORION,            "orion"            },
     { LLM_ARCH_INTERNLM2,        "internlm2"        },
@@ -115,6 +117,9 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_RND1,             "rnd1"             },
     { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
     { LLM_ARCH_MISTRAL3,         "mistral3"         },
+    { LLM_ARCH_MIMO2,            "mimo2"           },
+    { LLM_ARCH_LLAMA_EMBED,      "llama-embed"      },
+    { LLM_ARCH_MAINCODER,        "maincoder"        },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -148,6 +153,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_VOCAB_SIZE,                        "%s.vocab_size"                        },
     { LLM_KV_CONTEXT_LENGTH,                    "%s.context_length"                    },
     { LLM_KV_EMBEDDING_LENGTH,                  "%s.embedding_length"                  },
+    { LLM_KV_EMBEDDING_LENGTH_OUT,              "%s.embedding_length_out"              },
     { LLM_KV_FEATURES_LENGTH,                   "%s.features_length"                   },
     { LLM_KV_BLOCK_COUNT,                       "%s.block_count"                       },
     { LLM_KV_LEADING_DENSE_BLOCK_COUNT,         "%s.leading_dense_block_count"         },
@@ -205,6 +211,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_GATE_LORA_RANK,               "%s.attention.gate_lora_rank"               },
     { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
     { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
+    { LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,       "%s.attention.sliding_window_pattern"       },
     { 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"           },
@@ -216,6 +223,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,       "%s.rope.dimension_sections"              },
     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
+    { LLM_KV_ROPE_FREQ_BASE_SWA,            "%s.rope.freq_base_swa"                   },
     { LLM_KV_ROPE_SCALE_LINEAR,             "%s.rope.scale_linear"                    },
     { LLM_KV_ROPE_SCALING_TYPE,             "%s.rope.scaling.type"                    },
     { LLM_KV_ROPE_SCALING_FACTOR,           "%s.rope.scaling.factor"                  },
@@ -500,6 +508,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
         case LLM_ARCH_LLAMA:
         case LLM_ARCH_DECI:
         case LLM_ARCH_MISTRAL3:
+        case LLM_ARCH_LLAMA_EMBED:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
                 LLM_TENSOR_OUTPUT_NORM,
@@ -781,6 +790,20 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_CLS,
                 LLM_TENSOR_CLS_OUT,
             };
+        case LLM_ARCH_MODERN_BERT:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_TOKEN_EMBD_NORM,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_ATTN_QKV,
+                LLM_TENSOR_FFN_DOWN,
+                LLM_TENSOR_FFN_UP,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_CLS,
+                LLM_TENSOR_CLS_OUT,
+            };
         case LLM_ARCH_JINA_BERT_V2:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
@@ -1060,6 +1083,22 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_ATTN_POST_NORM,
                 LLM_TENSOR_FFN_POST_NORM,
             };
+        case LLM_ARCH_PLAMO3:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_QKV,
+                LLM_TENSOR_ATTN_Q_NORM,
+                LLM_TENSOR_ATTN_K_NORM,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_ATTN_POST_NORM,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_POST_NORM,
+                LLM_TENSOR_FFN_DOWN,
+                LLM_TENSOR_FFN_UP,
+            };
         case LLM_ARCH_CODESHELL:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
@@ -2040,6 +2079,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_TOKEN_EMBD,
                 LLM_TENSOR_OUTPUT_NORM_LFM2,
                 LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_DENSE_2_OUT,
             };
         case LLM_ARCH_LFM2MOE:
             return {
@@ -2058,7 +2098,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_SHORTCONV_INPROJ,
                 LLM_TENSOR_SHORTCONV_OUTPROJ,
                 LLM_TENSOR_TOKEN_EMBD,
-                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT_NORM_LFM2,
                 LLM_TENSOR_FFN_GATE_INP,
                 LLM_TENSOR_FFN_GATE_EXPS,
                 LLM_TENSOR_FFN_DOWN_EXPS,
@@ -2174,11 +2214,49 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_VISEXP_FFN_DOWN,
                 LLM_TENSOR_VISEXP_FFN_UP,
             };
+        case LLM_ARCH_MIMO2:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_Q,
+                LLM_TENSOR_ATTN_K,
+                LLM_TENSOR_ATTN_V,
+                LLM_TENSOR_ATTN_SINKS,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_GATE,
+                LLM_TENSOR_FFN_DOWN,
+                LLM_TENSOR_FFN_UP,
+                LLM_TENSOR_FFN_GATE_INP,
+                LLM_TENSOR_FFN_GATE_EXPS,
+                LLM_TENSOR_FFN_DOWN_EXPS,
+                LLM_TENSOR_FFN_UP_EXPS,
+                LLM_TENSOR_FFN_EXP_PROBS_B,
+            };
         case LLM_ARCH_GPTJ:
         case LLM_ARCH_UNKNOWN:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
             };
+        case LLM_ARCH_MAINCODER:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_Q,
+                LLM_TENSOR_ATTN_Q_NORM,
+                LLM_TENSOR_ATTN_K,
+                LLM_TENSOR_ATTN_K_NORM,
+                LLM_TENSOR_ATTN_V,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_GATE,
+                LLM_TENSOR_FFN_DOWN,
+                LLM_TENSOR_FFN_UP,
+            };
         case LLM_ARCH_SOLAR:
             return {
                 LLM_TENSOR_TOKEN_EMBD,

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

@@ -24,6 +24,7 @@ enum llm_arch {
     LLM_ARCH_STARCODER,
     LLM_ARCH_REFACT,
     LLM_ARCH_BERT,
+    LLM_ARCH_MODERN_BERT,
     LLM_ARCH_NOMIC_BERT,
     LLM_ARCH_NOMIC_BERT_MOE,
     LLM_ARCH_NEO_BERT,
@@ -45,6 +46,7 @@ enum llm_arch {
     LLM_ARCH_PHIMOE,
     LLM_ARCH_PLAMO,
     LLM_ARCH_PLAMO2,
+    LLM_ARCH_PLAMO3,
     LLM_ARCH_CODESHELL,
     LLM_ARCH_ORION,
     LLM_ARCH_INTERNLM2,
@@ -119,6 +121,9 @@ enum llm_arch {
     LLM_ARCH_RND1,
     LLM_ARCH_PANGU_EMBED,
     LLM_ARCH_MISTRAL3,
+    LLM_ARCH_MIMO2,
+    LLM_ARCH_LLAMA_EMBED,
+    LLM_ARCH_MAINCODER,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -152,6 +157,7 @@ enum llm_kv {
     LLM_KV_VOCAB_SIZE,
     LLM_KV_CONTEXT_LENGTH,
     LLM_KV_EMBEDDING_LENGTH,
+    LLM_KV_EMBEDDING_LENGTH_OUT,
     LLM_KV_FEATURES_LENGTH,
     LLM_KV_BLOCK_COUNT,
     LLM_KV_LEADING_DENSE_BLOCK_COUNT,
@@ -209,6 +215,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_GATE_LORA_RANK,
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
+    LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_OUTPUT_SCALE,
     LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
@@ -220,6 +227,7 @@ enum llm_kv {
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
     LLM_KV_ROPE_FREQ_BASE,
+    LLM_KV_ROPE_FREQ_BASE_SWA,
     LLM_KV_ROPE_SCALE_LINEAR,
     LLM_KV_ROPE_SCALING_TYPE,
     LLM_KV_ROPE_SCALING_FACTOR,

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

@@ -74,6 +74,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "seed_oss",          LLM_CHAT_TEMPLATE_SEED_OSS          },
     { "grok-2",            LLM_CHAT_TEMPLATE_GROK_2            },
     { "pangu-embedded",    LLM_CHAT_TEMPLATE_PANGU_EMBED       },
+    { "solar-open",        LLM_CHAT_TEMPLATE_SOLAR_OPEN        },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -216,6 +217,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_GROK_2;
     } else if (tmpl_contains(LU8("[unused9]系统：[unused10]"))) {
         return LLM_CHAT_TEMPLATE_PANGU_EMBED;
+    } else if (tmpl_contains("<|begin|>") && tmpl_contains("<|end|>") && tmpl_contains("<|content|>")) {
+        return LLM_CHAT_TEMPLATE_SOLAR_OPEN;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -845,6 +848,14 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "[unused9]助手：";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_SOLAR_OPEN) {
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|begin|>" << role << "<|content|>" << message->content << "<|end|>";
+        }
+        if (add_ass) {
+            ss << "<|begin|>assistant";
+        }
     } else {
         // template not supported
         return -1;

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

@@ -54,6 +54,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_SEED_OSS,
     LLM_CHAT_TEMPLATE_GROK_2,
     LLM_CHAT_TEMPLATE_PANGU_EMBED,
+    LLM_CHAT_TEMPLATE_SOLAR_OPEN,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

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

@@ -60,6 +60,25 @@ llama_context::llama_context(
     cparams.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
 
+    // Initialize backend samplers here so they are part of the sampling graph
+    // before the reserve passes run later in this function. This avoids a later
+    // re-reserve when graph nodes change.
+    if (params.samplers != nullptr && params.n_samplers > 0) {
+        for (size_t i = 0; i < params.n_samplers; ++i) {
+            const auto & config = params.samplers[i];
+
+            if (llama_sampler_chain_get(config.sampler, -1) == nullptr) {
+                throw std::runtime_error("the backend samplers must be of type llama_sampler_chain");
+            }
+
+            if (set_sampler(config.seq_id, config.sampler)) {
+                const int n_samplers = llama_sampler_chain_n(config.sampler);
+
+                LLAMA_LOG_INFO("%s: setting backend sampler for seq_id %d (n = %d)\n", __func__, config.seq_id, n_samplers);
+            }
+        }
+    }
+
     auto rope_scaling_type = params.rope_scaling_type;
     if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED) {
         rope_scaling_type = hparams.rope_scaling_type_train;
@@ -231,7 +250,10 @@ llama_context::llama_context(
         // graph outputs buffer
         {
             // resized during inference when a batch uses more outputs
-            if (output_reserve(params.n_seq_max) < params.n_seq_max) {
+            // Create a dummy batch for initialization.
+            llama_batch dummy_batch = {};
+            dummy_batch.n_tokens = 0;
+            if (output_reserve(params.n_seq_max, dummy_batch) < params.n_seq_max) {
                 throw std::runtime_error("failed to reserve initial output buffer");
             }
 
@@ -294,8 +316,8 @@ llama_context::llama_context(
         // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
         bool pipeline_parallel =
             model.n_devices() > 1 &&
-            model.params.n_gpu_layers > (int) model.hparams.n_layer &&
-            model.params.split_mode == LLAMA_SPLIT_MODE_LAYER &&
+            model.n_gpu_layers() > model.hparams.n_layer &&
+            model.split_mode() == LLAMA_SPLIT_MODE_LAYER &&
             cparams.offload_kqv &&
             !model.has_tensor_overrides();
 
@@ -456,26 +478,35 @@ llama_context::llama_context(
             LLAMA_LOG_INFO("%s: graph splits = %d (with bs=%d), %d (with bs=1)\n", __func__, n_splits_pp, n_tokens, n_splits_tg);
         }
     }
+
+    // Initialize the full vocabulary token ids for backend samplers.
+    {
+        const int n_vocab = model.vocab.n_tokens();
+
+        sampling.token_ids_full_vocab.resize(n_vocab);
+        for (int i = 0; i < n_vocab; ++i) {
+            sampling.token_ids_full_vocab[i] = i;
+        }
+    }
 }
 
 llama_context::~llama_context() {
-    // FIXME this currently results in a use-after-free bug if the model is freed before the context
-    // if (!model.hparams.no_alloc) {
-    //     for (size_t i = 0; i < backend_ptrs.size(); ++i) {
-    //         ggml_backend_t             backend = backend_ptrs[i];
-    //         ggml_backend_buffer_type_t buft    = backend_buft[i];
+    if (!model.hparams.no_alloc) {
+        for (size_t i = 0; i < backend_ptrs.size(); ++i) {
+            ggml_backend_t             backend = backend_ptrs[i];
+            ggml_backend_buffer_type_t buft    = backend_buft[i];
 
-    //         const size_t size_exp = backend_buf_exp_size[i];
-    //         const size_t size_act = ggml_backend_sched_get_buffer_size(sched.get(), backend);
-    //         if (size_exp == size_act) {
-    //             LLAMA_LOG_DEBUG("%s: %10s compute buffer size is %8.4f MiB, matches expectation of %8.4f MiB\n",
-    //                 __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
-    //         } else {
-    //             LLAMA_LOG_WARN("%s: %10s compute buffer size of %8.4f MiB, does not match expectation of %8.4f MiB\n",
-    //                 __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
-    //         }
-    //     }
-    // }
+            const size_t size_exp = backend_buf_exp_size[i];
+            const size_t size_act = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+            if (size_exp == size_act) {
+                LLAMA_LOG_DEBUG("%s: %10s compute buffer size is %8.4f MiB, matches expectation of %8.4f MiB\n",
+                    __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+            } else {
+                LLAMA_LOG_WARN("%s: %10s compute buffer size of %8.4f MiB, does not match expectation of %8.4f MiB\n",
+                    __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+            }
+        }
+    }
     ggml_opt_free(opt_ctx);
 }
 
@@ -617,6 +648,35 @@ float * llama_context::get_logits() {
     return logits;
 }
 
+int64_t llama_context::output_resolve_row(int32_t i) const {
+    int64_t j = -1;
+
+    // support negative indices (last output row)
+    if (i < 0) {
+        j = n_outputs + i;
+        if (j < 0) {
+            throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs));
+        }
+    } else if ((size_t) i >= output_ids.size()) {
+        throw std::runtime_error(format("out of range [0, %zu)", output_ids.size()));
+    } else {
+        // use output_ids to translate the batch token index into a row number
+        // that holds this token's data.
+        j = output_ids[i];
+    }
+
+    if (j < 0) {
+        // the batch token was not configured to output anything
+        throw std::runtime_error(format("batch.logits[%d] != true", i));
+    }
+
+    if (j >= n_outputs) {
+        throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
+    }
+
+    return j;
+}
+
 float * llama_context::get_logits_ith(int32_t i) {
     int64_t j = -1;
 
@@ -627,6 +687,7 @@ float * llama_context::get_logits_ith(int32_t i) {
             throw std::runtime_error("no logits");
         }
 
+        // TODO: use output_resolve_row()
         if (i < 0) {
             j = n_outputs + i;
             if (j < 0) {
@@ -663,6 +724,10 @@ float * llama_context::get_embeddings() {
     return embd;
 }
 
+llama_token * llama_context::get_sampled_tokens()  const{
+    return sampling.sampled;
+}
+
 float * llama_context::get_embeddings_ith(int32_t i) {
     int64_t j = -1;
 
@@ -673,6 +738,7 @@ float * llama_context::get_embeddings_ith(int32_t i) {
             throw std::runtime_error("no embeddings");
         }
 
+        // TODO: use output_resolve_row()
         if (i < 0) {
             j = n_outputs + i;
             if (j < 0) {
@@ -692,7 +758,8 @@ float * llama_context::get_embeddings_ith(int32_t i) {
             throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
-        return embd + j*model.hparams.n_embd;
+        const uint32_t n_embd_out = model.hparams.get_n_embd_out();
+        return embd + j*n_embd_out;
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what());
 #ifndef NDEBUG
@@ -712,6 +779,136 @@ float * llama_context::get_embeddings_seq(llama_seq_id seq_id) {
     return it->second.data();
 }
 
+llama_token llama_context::get_sampled_token_ith(int32_t idx) {
+    output_reorder();
+
+    if (sampling.sampled == nullptr) {
+        return LLAMA_TOKEN_NULL;
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        GGML_ASSERT(row < (int64_t) sampling.sampled_size);
+        return sampling.sampled[row];
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled token id %d, reason: %s\n", __func__, idx, err.what());
+        return LLAMA_TOKEN_NULL;
+    }
+}
+
+float * llama_context::get_sampled_probs_ith(int32_t idx) {
+    output_reorder();
+
+    if (sampling.probs == nullptr) {
+        return nullptr;
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if ((size_t) row >= sampling.probs_count.size() || sampling.probs_count[row] == 0) {
+            return nullptr;
+        }
+        return sampling.probs + row*model.vocab.n_tokens();
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled probs id %d, reason: %s\n", __func__, idx, err.what());
+        return nullptr;
+    }
+}
+
+float * llama_context::get_sampled_logits_ith(int32_t idx) {
+    output_reorder();
+
+    if (sampling.logits == nullptr) {
+        return nullptr;
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if ((size_t) row >= sampling.logits_count.size() || sampling.logits_count[row] == 0) {
+            return nullptr;
+        }
+        return sampling.logits + row*model.vocab.n_tokens();
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled logits id %d, reason: %s\n", __func__, idx, err.what());
+        return nullptr;
+    }
+}
+
+const llama_token * llama_context::get_sampled_candidates_ith(int32_t idx) {
+    output_reorder();
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if (sampling.candidates != nullptr &&
+            (size_t) row < sampling.candidates_count.size() &&
+            sampling.candidates_count[row] > 0) {
+            return sampling.candidates + row*model.vocab.n_tokens();
+        }
+    } catch (const std::exception & err) {
+        // fallback to full vocab list
+    }
+
+    return sampling.token_ids_full_vocab.data();
+}
+
+size_t llama_context::get_sampled_candidates_count(int32_t idx) {
+    output_reorder();
+
+    if (sampling.candidates == nullptr) {
+        return 0;
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if ((size_t) row >= sampling.candidates_count.size()) {
+            return 0;
+        }
+        return sampling.candidates_count[row];
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled candidates count id %d, reason: %s\n", __func__, idx, err.what());
+        return 0;
+    }
+}
+
+size_t llama_context::get_sampled_logits_count(int32_t idx) {
+    output_reorder();
+
+    if (sampling.logits == nullptr) {
+        return model.vocab.n_tokens();
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if ((size_t) row >= sampling.logits_count.size()) {
+            return 0;
+        }
+        return sampling.logits_count[row];
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled logits count id %d, reason: %s\n", __func__, idx, err.what());
+        return 0;
+    }
+}
+
+size_t llama_context::get_sampled_probs_count(int32_t idx) {
+    output_reorder();
+
+    if (sampling.probs == nullptr) {
+        return 0;
+    }
+
+    try {
+        const int64_t row = output_resolve_row(idx);
+        if ((size_t) row >= sampling.probs_count.size()) {
+            return 0;
+        }
+        return sampling.probs_count[row];
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid backend sampled probs count id %d, reason: %s\n", __func__, idx, err.what());
+        return 0;
+    }
+}
+
+
 void llama_context::attach_threadpool(
            ggml_threadpool_t threadpool,
            ggml_threadpool_t threadpool_batch) {
@@ -768,6 +965,42 @@ void llama_context::set_warmup(bool value) {
     cparams.warmup = value;
 }
 
+bool llama_context::set_sampler(llama_seq_id seq_id, llama_sampler * sampler) {
+    LLAMA_LOG_DEBUG("%s: seq_id = %d, sampler = %p\n", __func__, (int) seq_id, (void *) sampler);
+
+    const bool can_offload =
+        sampler &&
+        sampler->iface->backend_init &&
+        sampler->iface->backend_apply &&
+        llama_sampler_chain_n(sampler) > 0;
+
+    if (sampler && can_offload) {
+        ggml_backend_buffer_type_t buft = ggml_backend_dev_buffer_type(model.dev_output());
+        auto * host_buft = ggml_backend_dev_host_buffer_type(model.dev_output());
+        if (host_buft) {
+            buft = host_buft;
+        }
+
+        sampler->iface->backend_init(sampler, buft);
+
+        sampling.samplers[seq_id] = sampler;
+
+        return true;
+    }
+
+    if (sampler && !can_offload) {
+        LLAMA_LOG_WARN("%s: sampler '%s' for seq_id = %d, cannot be offloaded to the backend\n", __func__, llama_sampler_name(sampler), seq_id);
+
+        sampling.samplers.erase(seq_id);
+
+        return false;
+    }
+
+    sampling.samplers.erase(seq_id);
+
+    return true;
+}
+
 void llama_context::set_adapter_lora(
             llama_adapter_lora * adapter,
             float scale) {
@@ -908,7 +1141,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
     n_queued_tokens += n_tokens;
 
     // reserve output buffer
-    if (output_reserve(n_tokens) < n_tokens) {
+    if (output_reserve(n_tokens, batch_inp) < n_tokens) {
         LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens);
         return -2;
     };
@@ -962,9 +1195,10 @@ int llama_context::encode(const llama_batch & batch_inp) {
                 {
                     // extract token embeddings
                     GGML_ASSERT(embd != nullptr);
+                    const uint32_t n_embd_out = hparams.get_n_embd_out();
 
-                    GGML_ASSERT(n_tokens*n_embd <= (int64_t) embd_size);
-                    ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd*sizeof(float));
+                    GGML_ASSERT(n_tokens*n_embd_out <= (int64_t) embd_size);
+                    ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd_out*sizeof(float));
                 } break;
             case LLAMA_POOLING_TYPE_MEAN:
             case LLAMA_POOLING_TYPE_CLS:
@@ -1032,6 +1266,112 @@ int llama_context::encode(const llama_batch & batch_inp) {
     return 0;
 }
 
+static std::map<llama_seq_id, uint32_t> build_seq_to_output_row(const llama_ubatch & ubatch, uint32_t row_offset) {
+    std::map<llama_seq_id, uint32_t> seq_to_row;
+    // how many output tokens we have seen so far for this ubatch.
+    uint32_t local = 0;
+    for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
+        // skip tokens that are not output.
+        if (!ubatch.output[i]) {
+            continue;
+        }
+
+        const llama_seq_id seq_id = ubatch.seq_id[i][0];
+        // row_offset is the number of output tokens before this ubatch.
+        seq_to_row[seq_id] = row_offset + local;
+        ++local;
+    }
+    return seq_to_row;
+}
+
+static void copy_tensor_async_ints(
+    const std::map<llama_seq_id, ggml_tensor*> & tensor_map,
+    llama_token * sampled,
+    size_t sampled_size,
+    const std::map<llama_seq_id, uint32_t> & seq_to_row,
+    ggml_backend_sched_t sched) {
+    if (sampled == nullptr) {
+        return;
+    }
+
+    for (const auto & [seq_id, tensor] : tensor_map) {
+        auto it = seq_to_row.find(seq_id);
+        if (it == seq_to_row.end()) {
+            continue;
+        }
+
+        const uint32_t row = it->second;
+        GGML_ASSERT(row < sampled_size);
+
+        GGML_ASSERT(ggml_is_contiguous(tensor) && "sampled tokens tensor must be contiguous for async copy");
+
+        ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(sched, tensor);
+        ggml_backend_tensor_get_async(backend, tensor, sampled + row, 0, sizeof(sampled[row]));
+    }
+}
+
+static void copy_tensor_async_floats(
+    const std::map<llama_seq_id, ggml_tensor*> & tensor_map,
+    float * dst,
+    size_t stride,
+    std::vector<uint32_t> & counts,
+    const std::map<llama_seq_id, uint32_t> & seq_to_row,
+    ggml_backend_sched_t sched) {
+    if (dst == nullptr) {
+        return;
+    }
+
+    for (const auto & [seq_id, tensor] : tensor_map) {
+        auto it = seq_to_row.find(seq_id);
+        if (it == seq_to_row.end()) {
+            continue;
+        }
+
+        const uint32_t row = it->second;
+        GGML_ASSERT(row < counts.size());
+
+        GGML_ASSERT(ggml_is_contiguous(tensor) && "logits/probs tensor must be contiguous for async copy");
+
+        ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(sched, tensor);
+        float * row_ptr = dst + (size_t) row * stride;
+        ggml_backend_tensor_get_async(backend, tensor, row_ptr, 0, ggml_nbytes(tensor));
+
+        // Update the actual number of logits/probabilities that were written for this row.
+        counts[row] = ggml_nelements(tensor);
+    }
+}
+
+static void copy_tensor_async_candidates(
+    const std::map<llama_seq_id, ggml_tensor*> & tensor_map,
+    llama_token * dst,
+    size_t stride,
+    std::vector<uint32_t> & counts,
+    const std::map<llama_seq_id, uint32_t> & seq_to_row,
+    ggml_backend_sched_t sched) {
+    if (dst == nullptr) {
+        return;
+    }
+
+    for (const auto & [seq_id, tensor] : tensor_map) {
+        auto it = seq_to_row.find(seq_id);
+        if (it == seq_to_row.end()) {
+            continue;
+        }
+
+        const uint32_t row = it->second;
+        GGML_ASSERT(row < counts.size());
+
+        GGML_ASSERT(ggml_is_contiguous(tensor) && "candidates tensor must be contiguous for async copy");
+
+        ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(sched, tensor);
+        llama_token * row_ptr = dst + (size_t) row * stride;
+        ggml_backend_tensor_get_async(backend, tensor, row_ptr, 0, ggml_nbytes(tensor));
+
+        // Update the actual number of candidates that were written.
+        counts[row] = ggml_nelements(tensor);
+    }
+}
+
 int llama_context::decode(const llama_batch & batch_inp) {
     GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT
 
@@ -1052,8 +1392,35 @@ int llama_context::decode(const llama_batch & batch_inp) {
     const int64_t n_embd  = hparams.n_embd_inp();
 
     const bool output_all = false;
+    const bool has_samplers = !sampling.samplers.empty();
 
-    if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, output_all)) {
+    const uint32_t n_seq_max = cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max;
+
+    // TODO: avoid this workaround in the future
+    if (has_samplers && batch_inp.logits) {
+        std::vector<int32_t> seq_output_count(n_seq_max, 0);
+
+        for (int32_t i = 0; i < batch_inp.n_tokens; ++i) {
+            if (batch_inp.logits[i] == 0) {
+                continue;
+            }
+
+            const int ns = batch_inp.n_seq_id ? batch_inp.n_seq_id[i] : 1;
+
+            for (int32_t s = 0; s < ns; ++s) {
+                const llama_seq_id seq_id = batch_inp.seq_id ? batch_inp.seq_id[i][s] : 0;
+
+                seq_output_count[seq_id]++;
+                if (seq_output_count[seq_id] > 1) {
+                    LLAMA_LOG_ERROR("%s: backend sampling requires at most one output token per sequence (seq_id %d had %d)\n",
+                            __func__, seq_id, seq_output_count[seq_id]);
+                    return -1;
+                }
+            }
+        }
+    }
+
+    if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, n_seq_max, output_all)) {
         LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
         return -1;
     }
@@ -1134,7 +1501,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     }
 
     // reserve output buffer
-    if (output_reserve(n_outputs_all) < n_outputs_all) {
+    if (output_reserve(n_outputs_all, balloc->get_batch()) < n_outputs_all) {
         LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
         return -2;
     };
@@ -1207,7 +1574,10 @@ int llama_context::decode(const llama_batch & batch_inp) {
         }
 
         // extract logits
-        if (t_logits && n_outputs > 0) {
+        // For multi-sequence batches that mix backend samplers and CPU sampler
+        // this is currently inefficient as we copy all logits even for the
+        // backend sampled tokens.
+        if (logits && t_logits && n_outputs > 0) {
             ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits);
             GGML_ASSERT(backend_res != nullptr);
             GGML_ASSERT(logits != nullptr);
@@ -1222,7 +1592,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
         }
 
         // extract embeddings
-        if (t_embd && n_outputs > 0) {
+        if (embd && t_embd && n_outputs > 0) {
             ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd);
             GGML_ASSERT(backend_embd != nullptr);
 
@@ -1231,12 +1601,13 @@ int llama_context::decode(const llama_batch & batch_inp) {
                     {
                         // extract token embeddings
                         GGML_ASSERT(embd != nullptr);
-                        float * embd_out = embd + n_outputs_prev*n_embd;
+                        const uint32_t n_embd_out = hparams.get_n_embd_out();
+                        float * embd_out = embd + n_outputs_prev*n_embd_out;
 
                         if (n_outputs) {
                             GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all);
-                            GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd <= (int64_t) embd_size);
-                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_out, 0, n_outputs*n_embd*sizeof(float));
+                            GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd_out <= (int64_t) embd_size);
+                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_out, 0, n_outputs*n_embd_out*sizeof(float));
                         }
                     } break;
                 case LLAMA_POOLING_TYPE_MEAN:
@@ -1276,6 +1647,22 @@ int llama_context::decode(const llama_batch & batch_inp) {
             }
         }
 
+        // This flag indicates whether a backend sampler has actually sampled a specific
+        // token, or if it has produced probabilites. If true, we can skip the normal copying of logits and embeddings.
+        const bool has_sampled = !res->t_sampled.empty() || !res->t_sampled_probs.empty() || !res->t_sampled_logits.empty();
+
+        if (has_samplers && has_sampled) {
+            const auto seq_to_output_row = build_seq_to_output_row(ubatch, n_outputs_prev);
+            const auto stride = n_vocab;
+
+            // async copy the sampling data from the backend to the host
+            copy_tensor_async_ints(res->t_sampled, sampling.sampled, sampling.sampled_size, seq_to_output_row, sched.get());
+
+            copy_tensor_async_floats    (res->t_sampled_logits, sampling.logits,     stride, sampling.logits_count,     seq_to_output_row, sched.get());
+            copy_tensor_async_floats    (res->t_sampled_probs,  sampling.probs,      stride, sampling.probs_count,      seq_to_output_row, sched.get());
+            copy_tensor_async_candidates(res->t_candidates,     sampling.candidates, stride, sampling.candidates_count, seq_to_output_row, sched.get());
+        }
+
         n_outputs_prev += n_outputs;
     } while (mctx->next());
 
@@ -1339,15 +1726,15 @@ int llama_context::decode(const llama_batch & batch_inp) {
 // output
 //
 
-uint32_t llama_context::output_reserve(int32_t n_outputs) {
+uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & batch) {
     const auto & hparams = model.hparams;
     const auto & vocab   = model.vocab;
 
     const int64_t n_outputs_max = std::max<int64_t>(n_outputs, n_seq_max());
 
-    const auto n_batch = cparams.n_batch;
-    const auto n_vocab = vocab.n_tokens();
-    const auto n_embd  = hparams.n_embd;
+    const auto n_batch    = cparams.n_batch;
+    const auto n_vocab    = vocab.n_tokens();
+    const auto n_embd_out = hparams.get_n_embd_out();
 
     bool has_logits = true;
     bool has_embd   = cparams.embeddings;
@@ -1358,8 +1745,53 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
         has_embd   = true;
     }
 
-    logits_size = has_logits ? n_vocab*n_outputs_max : 0;
-    embd_size   = has_embd   ?  n_embd*n_outputs_max : 0;
+    // Check which sampling modes are needed for the current batch.
+    // TODO: avoid this branching by working with the worst-case
+    bool has_sampling = false;
+    bool cpu_logits   = false;
+
+    if (batch.logits) {
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            if (!batch.logits[i]) {
+                continue;
+            }
+            for (int32_t j = 0; j < batch.n_seq_id[i]; j++) {
+                llama_seq_id seq_id = batch.seq_id[i][j];
+                if (sampling.samplers.find(seq_id) != sampling.samplers.end()) {
+                    has_sampling = true;
+                } else {
+                    cpu_logits = true;
+                }
+            }
+        }
+    } else {
+        // When batch.logits is nullptr (when loading state with a dummy batch),
+        // allocate CPU logits.
+        cpu_logits = true;
+    }
+
+    size_t backend_float_count = 0;
+    size_t backend_token_count = 0;
+
+    // Allocate CPU logits buffer only if needed by sequences in this batch
+    logits_size = (has_logits && cpu_logits) ? n_vocab*n_outputs_max : 0;
+    embd_size   = has_embd ? n_embd_out*n_outputs_max : 0;
+
+    // TODO: avoid this branching by working with the worst-case
+    if (!has_sampling) {
+        sampling.logits_size     = 0;
+        sampling.probs_size      = 0;
+        sampling.sampled_size    = 0;
+        sampling.candidates_size = 0;
+    } else {
+        sampling.logits_size     = n_vocab*n_outputs_max;
+        sampling.probs_size      = n_vocab*n_outputs_max;
+        sampling.sampled_size    =         n_outputs_max;
+        sampling.candidates_size = n_vocab*n_outputs_max;
+
+        backend_float_count = sampling.logits_size  + sampling.probs_size;
+        backend_token_count = sampling.sampled_size + sampling.candidates_size;
+    }
 
     if (output_ids.empty()) {
         // init, never resized afterwards
@@ -1367,7 +1799,9 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
     }
 
     const size_t prev_size = buf_output ? ggml_backend_buffer_get_size(buf_output.get()) : 0;
-    const size_t new_size  = (logits_size + embd_size) * sizeof(float);
+    const size_t new_size  =
+        (logits_size + embd_size + backend_float_count) * sizeof(float) +
+        (                          backend_token_count) * sizeof(llama_token);
 
     // alloc only when more than the current capacity is required
     // TODO: also consider shrinking the buffer
@@ -1375,9 +1809,11 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
         if (buf_output) {
 #ifndef NDEBUG
             // This doesn't happen often, but may be annoying in some cases (like the HellaSwag benchmark)
-            LLAMA_LOG_INFO("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+            LLAMA_LOG_DEBUG("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
             synchronize();
+
+            // TODO: not needed?
             buf_output = nullptr;
             logits = nullptr;
             embd = nullptr;
@@ -1399,8 +1835,49 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
 
     float * output_base = (float *) ggml_backend_buffer_get_base(buf_output.get());
 
-    logits = has_logits ? output_base               : nullptr;
-    embd   = has_embd   ? output_base + logits_size : nullptr;
+    logits = nullptr;
+    embd   = nullptr;
+
+    size_t offset = 0;
+    uint8_t * base = (uint8_t *) output_base;
+
+    logits = (has_logits && cpu_logits) ? output_base : nullptr;
+    offset += logits_size * sizeof(float);
+
+    embd = has_embd ? (float *) (base + offset) : nullptr;
+    offset += embd_size * sizeof(float);
+
+    sampling.logits     = nullptr;
+    sampling.probs      = nullptr;
+    sampling.sampled    = nullptr;
+    sampling.candidates = nullptr;
+
+    if (has_sampling) {
+        sampling.logits = (float *) (base + offset);
+        offset += sampling.logits_size * sizeof(float);
+
+        sampling.probs = (float *) (base + offset);
+        offset += sampling.probs_size * sizeof(float);
+
+        sampling.sampled = (llama_token *) (base + offset);
+        offset += sampling.sampled_size * sizeof(llama_token);
+
+        sampling.candidates = (llama_token *) (base + offset);
+        offset += sampling.candidates_size * sizeof(llama_token);
+
+        // The count vectors keep track of the actual number of logits/probs/candidates
+        // copied from the backend for each output row.
+
+        sampling.logits_count.resize(n_outputs_max);
+        sampling.probs_count.resize(n_outputs_max);
+        sampling.candidates_count.resize(n_outputs_max);
+
+        std::fill(sampling.logits_count.begin(),     sampling.logits_count.end(),     0);
+        std::fill(sampling.probs_count.begin(),      sampling.probs_count.end(),      0);
+        std::fill(sampling.candidates_count.begin(), sampling.candidates_count.end(), 0);
+
+        std::fill_n(sampling.sampled, sampling.sampled_size, LLAMA_TOKEN_NULL);
+    }
 
     // set all ids as invalid (negative)
     std::fill(output_ids.begin(), output_ids.end(), -1);
@@ -1429,6 +1906,40 @@ void llama_context::output_reorder() {
                 std::swap(embd[i0*n_embd + k], embd[i1*n_embd + k]);
             }
         }
+
+        if (sampling.logits && sampling.logits_size > 0) {
+            for (uint64_t k = 0; k < n_vocab; ++k) {
+                std::swap(sampling.logits[i0*n_vocab + k], sampling.logits[i1*n_vocab + k]);
+            }
+        }
+
+        if (sampling.probs && sampling.probs_size > 0) {
+            for (uint64_t k = 0; k < n_vocab; ++k) {
+                std::swap(sampling.probs[i0*n_vocab + k], sampling.probs[i1*n_vocab + k]);
+            }
+        }
+
+        if (sampling.candidates && sampling.candidates_size > 0) {
+            for (uint64_t k = 0; k < n_vocab; ++k) {
+                std::swap(sampling.candidates[i0*n_vocab + k], sampling.candidates[i1*n_vocab + k]);
+            }
+        }
+
+        if (sampling.sampled && sampling.sampled_size > 0) {
+            std::swap(sampling.sampled[i0], sampling.sampled[i1]);
+        }
+
+        if (!sampling.logits_count.empty()) {
+            std::swap(sampling.logits_count[i0], sampling.logits_count[i1]);
+        }
+
+        if (!sampling.probs_count.empty()) {
+            std::swap(sampling.probs_count[i0], sampling.probs_count[i1]);
+        }
+
+        if (!sampling.candidates_count.empty()) {
+            std::swap(sampling.candidates_count[i0], sampling.candidates_count[i1]);
+        }
     }
 
     output_swaps.clear();
@@ -1442,7 +1953,9 @@ 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());
+    uint32_t res = std::max<uint32_t>(1024u, 8u*model.n_tensors());
+    res += model.n_lora_nodes;
+    return res;
 }
 
 llm_graph_result * llama_context::get_gf_res_reserve() const {
@@ -1456,7 +1969,7 @@ ggml_cgraph * llama_context::graph_reserve(
 
     if (n_tokens % n_seqs != 0) {
         n_tokens = ((n_tokens + (n_seqs - 1)) / n_seqs) * n_seqs; // round to next multiple of n_seqs
-        n_outputs = std::min(n_outputs, n_tokens);
+        n_outputs = std::max(n_outputs, n_tokens);
 
         LLAMA_LOG_DEBUG("%s: making n_tokens a multiple of n_seqs - n_tokens = %u, n_seqs = %u, n_outputs = %u\n", __func__, n_tokens, n_seqs, n_outputs);
     }
@@ -1475,6 +1988,15 @@ ggml_cgraph * llama_context::graph_reserve(
     llama_batch_allocr balloc(model.hparams.n_pos_per_embd());
     llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs);
 
+    // set one output token per sequence in order to activate all backend samplers
+    std::vector<llama_seq_id> seq_ids(n_seqs);
+    for (uint32_t i = 0; i < n_seqs; ++i) {
+        seq_ids[i] = i;
+        ubatch.n_seq_id[i] = 1;
+        ubatch.seq_id[i] = &seq_ids[i];
+        ubatch.output[i] = true;
+    }
+
     auto * res = gf_res_reserve.get();
 
     const auto gparams = graph_params(res, ubatch, mctx, LLM_GRAPH_TYPE_DEFAULT);
@@ -1505,7 +2027,7 @@ llm_graph_params llama_context::graph_params(
                         llm_graph_result * res,
                       const llama_ubatch & ubatch,
             const llama_memory_context_i * mctx,
-            llm_graph_type   gtype) const {
+                          llm_graph_type   gtype) const {
     return {
         /*.arch        =*/ model.arch,
         /*.hparams     =*/ model.hparams,
@@ -1518,6 +2040,7 @@ llm_graph_params llama_context::graph_params(
         /*.loras       =*/ &loras,
         /*.mctx        =*/ mctx,
         /*.cross       =*/ &cross,
+        /*.samplers    =*/ sampling.samplers,
         /*.n_outputs   =*/ n_outputs,
         /*.cb          =*/ graph_get_cb(),
         /*.res         =*/ res,
@@ -1570,7 +2093,7 @@ llm_graph_cb llama_context::graph_get_cb() const {
 
         // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
         // FIXME: fix in ggml_backend_sched
-        const bool full_offload = model.params.n_gpu_layers > (int) model.hparams.n_layer;
+        const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer;
         if (ubatch.n_tokens < 32 || full_offload) {
             if (il != -1 && strcmp(name, "norm") == 0) {
                 const auto & dev_layer = model.dev_layer(il);
@@ -1973,6 +2496,9 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
         }
     }
 
+    // TODO: handle sampling buffers and samplers state ?
+    //       https://github.com/ggml-org/llama.cpp/pull/17004
+
     if (memory != nullptr) {
         LLAMA_LOG_DEBUG("%s: - writing memory module\n", __func__);
         memory->state_write(io);
@@ -2005,7 +2531,10 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
         auto n_outputs = this->n_outputs;
         io.read_to(&n_outputs, sizeof(n_outputs));
 
-        if (n_outputs > output_reserve(n_outputs)) {
+        // Create a dummy batch for state loading.
+        llama_batch dummy_batch = {};
+        dummy_batch.n_tokens = 0;
+        if (n_outputs > output_reserve(n_outputs, dummy_batch)) {
             throw std::runtime_error("could not reserve outputs");
         }
 
@@ -2059,6 +2588,9 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
         }
     }
 
+    // TODO: handle sampling buffers and samplers state ?
+    //       https://github.com/ggml-org/llama.cpp/pull/17004
+
     if (memory) {
         LLAMA_LOG_DEBUG("%s: - reading memory module\n", __func__);
 
@@ -2247,7 +2779,7 @@ void llama_context::opt_epoch_iter(
         }
 
         // reserve output buffer
-        if (output_reserve(n_outputs_all) < n_outputs_all) {
+        if (output_reserve(n_outputs_all, balloc->get_batch()) < n_outputs_all) {
             LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
             GGML_ABORT("TODO: handle this error");
         };
@@ -2392,6 +2924,8 @@ llama_context_params llama_context_default_params() {
         /*.op_offload                  =*/ true,
         /*.swa_full                    =*/ true,
         /*.kv_unified                  =*/ false,
+        /*.sampler                     =*/ nullptr,
+        /*.n_sampler                   =*/ 0,
     };
 
     return result;
@@ -2551,7 +3085,15 @@ float * llama_get_logits(llama_context * ctx) {
 float * llama_get_logits_ith(llama_context * ctx, int32_t i) {
     ctx->synchronize();
 
-    return ctx->get_logits_ith(i);
+    float * res = nullptr;
+
+    res = ctx->get_sampled_logits_ith(i);
+
+    if (!res) {
+        res = ctx->get_logits_ith(i);
+    }
+
+    return res;
 }
 
 float * llama_get_embeddings(llama_context * ctx) {
@@ -2572,6 +3114,52 @@ float * llama_get_embeddings_seq(llama_context * ctx, llama_seq_id seq_id) {
     return ctx->get_embeddings_seq(seq_id);
 }
 
+bool llama_set_sampler(llama_context * ctx, llama_seq_id seq_id, llama_sampler * smpl) {
+    return ctx->set_sampler(seq_id, smpl);
+}
+
+llama_token llama_get_sampled_token_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return ctx->get_sampled_token_ith(i);
+}
+
+float * llama_get_sampled_probs_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return ctx->get_sampled_probs_ith(i);
+}
+
+float * llama_get_sampled_logits_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return ctx->get_sampled_logits_ith(i);
+}
+
+llama_token * llama_get_sampled_candidates_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return const_cast<llama_token *>(ctx->get_sampled_candidates_ith(i));
+}
+
+uint32_t llama_get_sampled_candidates_count_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return static_cast<uint32_t>(ctx->get_sampled_candidates_count(i));
+}
+
+uint32_t llama_get_sampled_logits_count_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return static_cast<uint32_t>(ctx->get_sampled_logits_count(i));
+}
+
+uint32_t llama_get_sampled_probs_count_ith(llama_context * ctx, int32_t i) {
+    ctx->synchronize();
+
+    return static_cast<uint32_t>(ctx->get_sampled_probs_count(i));
+}
+
 // llama adapter API
 
 int32_t llama_set_adapter_lora(

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

@@ -70,6 +70,18 @@ struct llama_context {
     float * get_embeddings_ith(int32_t i);
     float * get_embeddings_seq(llama_seq_id seq_id);
 
+    llama_token * get_sampled_tokens() const;
+    llama_token   get_sampled_token_ith(int32_t idx);
+
+    float * get_sampled_logits_ith(int32_t idx);
+    size_t  get_sampled_logits_count(int32_t idx);
+
+    float * get_sampled_probs_ith(int32_t idx);
+    size_t  get_sampled_probs_count(int32_t idx);
+
+    const llama_token * get_sampled_candidates_ith(int32_t idx);
+    size_t get_sampled_candidates_count(int32_t idx);
+
     void attach_threadpool(
             ggml_threadpool_t threadpool,
             ggml_threadpool_t threadpool_batch);
@@ -192,10 +204,13 @@ private:
 
     // Make sure enough space is available for outputs.
     // Returns max number of outputs for which space was reserved.
-    uint32_t output_reserve(int32_t n_outputs);
+    uint32_t output_reserve(int32_t n_outputs, const llama_batch & batch);
 
     void output_reorder();
 
+    // map the output row index `i` to batch index
+    int64_t output_resolve_row(int32_t i) const;
+
     //
     // graph
     //
@@ -213,6 +228,8 @@ public:
     ggml_cgraph * graph_reserve(
         uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false, size_t * sizes = nullptr);
 
+    bool set_sampler(llama_seq_id seq_id, llama_sampler * sampler);
+
 private:
     llm_graph_params graph_params(
                         llm_graph_result * res,
@@ -252,6 +269,31 @@ private:
     size_t  embd_size = 0; // capacity (of floats) for embeddings
     float * embd      = nullptr;
 
+    // TODO: simplify
+    struct sampling_info {
+        std::map<llama_seq_id, llama_sampler *> samplers;
+
+        float       * logits      = nullptr;
+        size_t        logits_size = 0;
+
+        llama_token * sampled      = nullptr;
+        size_t        sampled_size = 0;
+
+        float       * probs        = nullptr;
+        size_t        probs_size   = 0;
+
+        llama_token * candidates   = nullptr;
+        size_t        candidates_size = 0;
+
+        std::vector<uint32_t> logits_count;
+        std::vector<uint32_t> probs_count;
+        std::vector<uint32_t> candidates_count;
+
+        std::vector<llama_token> token_ids_full_vocab;
+    };
+
+    sampling_info sampling;
+
     // sequence embeddings output (map of [n_embd] vectors)
     // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
     std::map<llama_seq_id, std::vector<float>> embd_seq;

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

@@ -369,6 +369,44 @@ static void print_rule(
     fprintf(file, "\n");
 }
 
+//
+// Regex utilities
+//
+
+size_t llama_grammar_trigger_pattern::find(const std::string & input) const {
+    auto find_start_pos = [](const std::smatch & match) {
+        // get from the first matched capturing group to the end of the string
+        size_t start = std::string::npos;
+        for (auto i = 1u; i < match.size(); i++) {
+            if (match.length(i) > 0) {
+                start = match.position(i);
+                break;
+            }
+        }
+        if (start == std::string::npos) {
+            start = match.position(0);
+        }
+        return start;
+    };
+
+    if (!pattern.empty() && pattern.front() == '^' && pattern.back() == '$') {
+        // match against the entire input
+        std::smatch match;
+        if (std::regex_match(input, match, regex)) {
+            return find_start_pos(match);
+        }
+    }
+
+    // search anywhere
+    std::smatch match;
+    if (std::regex_search(input, match, regex)) {
+        return find_start_pos(match);
+    }
+
+    return std::string::npos;
+}
+
+
 //
 // implementation
 //
@@ -1321,21 +1359,10 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
             grammar.trigger_buffer_positions.push_back(std::make_pair(token, position));
             grammar.trigger_buffer += piece;
 
-            std::smatch match;
             for (const auto & trigger_pattern : grammar.trigger_patterns) {
-                if (std::regex_match(grammar.trigger_buffer, match, trigger_pattern.regex)) {
+                auto start = trigger_pattern.find(grammar.trigger_buffer);
+                if (start != std::string::npos) {
                     grammar.awaiting_trigger = false;
-                    // get from the first matched capturing group to the end of the string
-                    size_t start = std::string::npos;
-                    for (auto i = 1u; i < match.size(); i++) {
-                        if (match.length(i) > 0) {
-                            start = match.position(i);
-                            break;
-                        }
-                    }
-                    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) {

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

@@ -130,6 +130,8 @@ struct llama_grammar_parser {
 struct llama_grammar_trigger_pattern {
     std::string pattern;
     std::regex  regex;
+
+    size_t find(const std::string & input) const;
 };
 
 struct llama_grammar {

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

@@ -12,6 +12,7 @@
 #include <cassert>
 #include <cmath>
 #include <cstring>
+#include <unordered_set>
 
 void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
     if (ubatch->token) {
@@ -32,7 +33,7 @@ bool llm_graph_input_embd::can_reuse(const llm_graph_params & params) {
     bool res = true;
 
     res &= (!tokens && !params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens);
-    res &= (!embd   && !params.ubatch.embd)  || (embd   &&   embd->ne[0] == params.ubatch.n_tokens);
+    res &= (!embd   && !params.ubatch.embd)  || (embd   &&   embd->ne[1] == params.ubatch.n_tokens);
 
     return res;
 }
@@ -62,7 +63,7 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
 bool llm_graph_input_pos::can_reuse(const llm_graph_params & params) {
     bool res = true;
 
-    res &= pos->ne[0] == params.ubatch.n_tokens;
+    res &= pos->ne[0] == params.ubatch.n_tokens*n_pos_per_embd;
 
     return res;
 }
@@ -521,6 +522,43 @@ bool llm_graph_input_mem_hybrid::can_reuse(const llm_graph_params & params) {
     return res;
 }
 
+void llm_graph_input_sampling::set_input(const llama_ubatch * ubatch) {
+    // set the inputs only for the active samplers in the current ubatch
+    std::unordered_set<llama_seq_id> active_samplers;
+    for (uint32_t i = 0; i < ubatch->n_tokens; i++) {
+        if (ubatch->output[i]) {
+            llama_seq_id seq_id = ubatch->seq_id[i][0];
+            active_samplers.insert(seq_id);
+        }
+    }
+
+    for (auto seq_id : active_samplers) {
+        if (samplers.find(seq_id) == samplers.end()) {
+            continue;
+        }
+
+        auto & sampler = samplers[seq_id];
+
+        if (sampler->iface->backend_set_input) {
+            sampler->iface->backend_set_input(sampler);
+        }
+    }
+}
+
+bool llm_graph_input_sampling::can_reuse(const llm_graph_params & params) {
+    if (samplers.size() != params.samplers.size()) {
+        return false;
+    }
+
+    for (const auto & [seq_id, sampler] : params.samplers) {
+        if (samplers[seq_id] != sampler) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
 //
 // llm_graph_result
 //
@@ -541,6 +579,10 @@ void llm_graph_result::reset() {
     t_logits      = nullptr;
     t_embd        = nullptr;
     t_embd_pooled = nullptr;
+    t_sampled.clear();
+    t_sampled_probs.clear();
+    t_sampled_logits.clear();
+    t_candidates.clear();
 
     params = {};
 
@@ -565,6 +607,38 @@ void llm_graph_result::set_inputs(const llama_ubatch * ubatch) {
     }
 }
 
+void llm_graph_result::set_outputs() {
+    if (t_logits != nullptr) {
+        ggml_set_output(t_logits);
+    }
+    if (t_embd != nullptr) {
+        ggml_set_output(t_embd);
+    }
+    if (t_embd_pooled != nullptr) {
+        ggml_set_output(t_embd_pooled);
+    }
+    for (auto & [seq_id, t] : t_sampled) {
+        if (t != nullptr) {
+            ggml_set_output(t);
+        }
+    }
+    for (auto & [seq_id, t] : t_sampled_probs) {
+        if (t != nullptr) {
+            ggml_set_output(t);
+        }
+    }
+    for (auto & [seq_id, t] : t_sampled_logits) {
+        if (t != nullptr) {
+            ggml_set_output(t);
+        }
+    }
+    for (auto & [seq_id, t] : t_candidates) {
+        if (t != nullptr) {
+            ggml_set_output(t);
+        }
+    }
+}
+
 bool llm_graph_result::can_reuse(const llm_graph_params & params) {
     if (!this->params.allow_reuse(params)) {
         if (debug > 1) {
@@ -646,6 +720,7 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     loras            (params.loras),
     mctx             (params.mctx),
     cross            (params.cross),
+    samplers         (params.samplers),
     cb_func          (params.cb),
     res              (params.res),
     ctx0             (res->get_ctx()),
@@ -1251,6 +1326,10 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
 
     res->add_input(std::move(inp));
 
+    // make sure the produced embeddings are immediately materialized in the ggml graph
+    // ref: https://github.com/ggml-org/llama.cpp/pull/18599
+    ggml_build_forward_expand(gf, cur);
+
     return cur;
 }
 
@@ -1834,8 +1913,10 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
 
         inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
+        ggml_set_name(inp->self_kq_mask, "self_kq_mask");
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
+        ggml_set_name(inp->self_kq_mask_cnv, "self_kq_mask_cnv");
     }
 
     {
@@ -1848,8 +1929,10 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
 
         inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp->self_kq_mask_swa);
+        ggml_set_name(inp->self_kq_mask_swa, "self_kq_mask_swa");
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
+        ggml_set_name(inp->self_kq_mask_swa_cnv, "self_kq_mask_swa_cnv");
     }
 
     return (llm_graph_input_attn_kv_iswa *) res->add_input(std::move(inp));
@@ -1988,14 +2071,18 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
 void llm_graph_context::build_dense_out(
     ggml_tensor * dense_2,
     ggml_tensor * dense_3) const {
-    if (!cparams.embeddings || dense_2 == nullptr || dense_3 == nullptr) {
+    if (!cparams.embeddings || !(dense_2 || dense_3)) {
         return;
     }
     ggml_tensor * cur = res->t_embd_pooled != nullptr ? res->t_embd_pooled : res->t_embd;
     GGML_ASSERT(cur != nullptr && "missing t_embd_pooled/t_embd");
 
-    cur = ggml_mul_mat(ctx0, dense_2, cur);
-    cur = ggml_mul_mat(ctx0, dense_3, cur);
+    if (dense_2) {
+        cur = ggml_mul_mat(ctx0, dense_2, cur);
+    }
+    if (dense_3) {
+        cur = ggml_mul_mat(ctx0, dense_3, cur);
+    }
     cb(cur, "result_embd_pooled", -1);
     res->t_embd_pooled = cur;
     ggml_build_forward_expand(gf, cur);
@@ -2086,6 +2173,87 @@ void llm_graph_context::build_pooling(
     ggml_build_forward_expand(gf, cur);
 }
 
+void llm_graph_context::build_sampling() const {
+    if (samplers.empty() || !res->t_logits) {
+        return;
+    }
+
+    auto inp_sampling = std::make_unique<llm_graph_input_sampling>(samplers);
+    res->add_input(std::move(inp_sampling));
+
+    std::map<llama_seq_id, int32_t> seq_to_logit_row;
+    int32_t logit_row_idx = 0;
+
+    for (uint32_t i = 0; i < ubatch.n_tokens; i++) {
+        if (ubatch.output[i]) {
+            llama_seq_id seq_id = ubatch.seq_id[i][0];
+            seq_to_logit_row[seq_id] = logit_row_idx;
+            logit_row_idx++;
+        }
+    }
+
+    // res->t_logits will contain logits for all tokens that want the logits calculated (logits=1 or output=1)
+    GGML_ASSERT(res->t_logits != nullptr && "missing t_logits tensor");
+
+    // add a dummy row of logits
+    // this trick makes the graph static, regardless of which samplers are activated
+    // this is important in order to minimize graph reallocations
+    // TODO: use `ggml_build_forward_select()` when available (https://github.com/ggml-org/llama.cpp/pull/18550)
+    ggml_tensor * logits_t = ggml_pad(ctx0, res->t_logits, 0, 1, 0, 0);
+
+    for (const auto & [seq_id, sampler] : samplers) {
+        const auto it = seq_to_logit_row.find(seq_id);
+
+        // inactive samplers always work on the first row
+        const auto row_idx = seq_to_logit_row.find(seq_id) != seq_to_logit_row.end() ? it->second : 0;
+
+        ggml_tensor * logits_seq = ggml_view_1d(ctx0, logits_t, logits_t->ne[0], row_idx * logits_t->nb[1]);
+        ggml_format_name(logits_seq, "logits_seq_%d", seq_id);
+
+        struct llama_sampler_data data = {
+            /*.logits      =*/ logits_seq,
+            /*.probs       =*/ nullptr,
+            /*.sampled     =*/ nullptr,
+            /*.candidates  =*/ nullptr,
+        };
+
+        assert(sampler->iface->backend_apply);
+        sampler->iface->backend_apply(sampler, ctx0, gf, &data);
+
+        if (data.sampled != nullptr) {
+            res->t_sampled[seq_id] = data.sampled;
+            ggml_build_forward_expand(gf, data.sampled);
+        }
+
+        if (data.probs != nullptr) {
+            res->t_sampled_probs[seq_id] = data.probs;
+            ggml_build_forward_expand(gf, data.probs);
+        }
+
+        if (data.logits != nullptr) {
+            res->t_sampled_logits[seq_id] = data.logits;
+            ggml_build_forward_expand(gf, data.logits);
+        }
+
+        if (data.candidates != nullptr) {
+            res->t_candidates[seq_id] = data.candidates;
+            ggml_build_forward_expand(gf, data.candidates);
+        }
+    }
+
+    // TODO: Call llama_sampler_accept_ggml after all samplers have been applied.
+    /*
+    for (const auto & [seq_id, sampler] : samplers) {
+        if (auto it = res->t_sampled.find(seq_id); it != res->t_sampled.end()) {
+            ggml_tensor * selected_token = it->second;
+            if (selected_token != nullptr) {
+                llama_sampler_accept_ggml(sampler, ctx0, gf, selected_token);
+            }
+        }
+    }
+    */
+}
+
 int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) {
     // TODO move to hparams if a T5 variant appears that uses a different value
     const int64_t max_distance = 128;

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

@@ -10,6 +10,7 @@
 #include <memory>
 #include <set>
 #include <functional>
+#include <map>
 
 struct ggml_cgraph;
 struct ggml_context;
@@ -396,6 +397,18 @@ public:
     const llama_memory_hybrid_context * mctx;
 };
 
+class llm_graph_input_sampling : public llm_graph_input_i {
+public:
+    llm_graph_input_sampling(std::map<llama_seq_id, llama_sampler *> samplers) :
+        samplers(std::move(samplers)) { }
+    virtual ~llm_graph_input_sampling() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+    bool can_reuse(const llm_graph_params & params) override;
+
+    std::map<llama_seq_id, llama_sampler *> samplers;
+};
+
 //
 // llm_graph_result
 //
@@ -429,6 +442,23 @@ struct llm_graph_params {
     const llama_memory_context_i * mctx;
     const llama_cross            * cross;
 
+    std::map<llama_seq_id, llama_sampler *> samplers;
+
+    static bool samplers_equal(
+          const std::map<llama_seq_id, llama_sampler *> & lhs,
+          const std::map<llama_seq_id, llama_sampler *> & rhs) {
+        if (lhs.size() != rhs.size()) {
+            return false;
+        }
+        for (const auto & [seq_id, sampler] : lhs) {
+            auto it = rhs.find(seq_id);
+            if (it == rhs.end() || it->second != sampler) {
+                return false;
+            }
+        }
+        return true;
+    }
+
     uint32_t n_outputs;
 
     llm_graph_cb cb;
@@ -468,15 +498,36 @@ struct llm_graph_params {
             return false;
         }
 
+        if (n_outputs != other.n_outputs) {
+            return false;
+        }
+
+        if (!samplers_equal(samplers, other.samplers)) {
+            return false;
+        }
+
+        if (samplers.size() > 0) {
+            if (!ubatch.data || !other.ubatch.data) {
+                return false;
+            }
+
+            // check that the outputs are the same for all samplers
+            for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
+                if (ubatch.output[i]    != other.ubatch.output[i] ||
+                    ubatch.seq_id[i][0] != other.ubatch.seq_id[i][0]) {
+                    return false;
+                }
+            }
+        }
+
         return
             cparams.embeddings  == other.cparams.embeddings  &&
             cparams.causal_attn == other.cparams.causal_attn &&
-            arch      == other.arch  &&
-            gtype     == other.gtype &&
-            cvec      == other.cvec  &&
-            loras     == other.loras &&
-            cross     == other.cross &&
-            n_outputs == other.n_outputs;
+            arch  == other.arch  &&
+            gtype == other.gtype &&
+            cvec  == other.cvec  &&
+            loras == other.loras &&
+            cross == other.cross;
     }
 };
 
@@ -499,6 +550,7 @@ public:
     void reset();
 
     void set_inputs(const llama_ubatch * ubatch);
+    void set_outputs();
 
     // try to update the existing graph result using the new graph parameters in order to reuse it
     // this can only be done if we determine that the resulting graph using the new graph parameters
@@ -517,6 +569,11 @@ public:
     ggml_tensor * t_embd        = nullptr;
     ggml_tensor * t_embd_pooled = nullptr;
 
+    std::map<llama_seq_id, ggml_tensor*> t_sampled_logits;
+    std::map<llama_seq_id, ggml_tensor*> t_candidates;
+    std::map<llama_seq_id, ggml_tensor*> t_sampled;
+    std::map<llama_seq_id, ggml_tensor*> t_sampled_probs;
+
     std::vector<llm_graph_input_ptr> inputs;
 
     ggml_context_ptr ctx_compute;
@@ -592,6 +649,8 @@ struct llm_graph_context {
     const llama_memory_context_i * mctx;
     const llama_cross            * cross;
 
+    std::map<llama_seq_id, llama_sampler *> samplers;
+
     const llm_graph_cb & cb_func;
 
     llm_graph_result * res;
@@ -832,6 +891,12 @@ struct llm_graph_context {
             ggml_tensor * cls_out,
             ggml_tensor * cls_out_b) const;
 
+    //
+    // sampling (backend sampling)
+    //
+
+    void build_sampling() const;
+
     //
     // dense (out)
     //

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

@@ -72,6 +72,10 @@ uint32_t llama_hparams::n_embd_inp() const {
     return n_embd_inp;
 }
 
+uint32_t llama_hparams::get_n_embd_out() const {
+    return n_embd_out > 0 ? n_embd_out : n_embd;
+}
+
 uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 

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

@@ -107,9 +107,9 @@ struct llama_hparams {
 
     float    rope_attn_factor = 1.0f;
     float    rope_freq_base_train;
-    float    rope_freq_base_train_swa;
+    float    rope_freq_base_train_swa  = 10000.0f;
     float    rope_freq_scale_train;
-    float    rope_freq_scale_train_swa;
+    float    rope_freq_scale_train_swa = 1.0f;
 
     uint32_t n_ctx_orig_yarn;
     float    rope_yarn_log_mul = 0.0f;
@@ -125,10 +125,11 @@ struct llama_hparams {
     llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
     // the size of the sliding window (0 - no SWA)
     uint32_t n_swa = 0;
-    // if swa_layers[il] == true, then layer il is SWA
-    // if swa_layers[il] == false, then layer il is dense (i.e. non-SWA)
+    // if swa_layers[il] == 1, then layer il is SWA
+    // if swa_layers[il] == 0, then layer il is dense (i.e. non-SWA)
     // by default, all layers are dense
-    std::array<bool, LLAMA_MAX_LAYERS> swa_layers;
+    // note: using uint32_t type for compatibility reason
+    std::array<uint32_t, LLAMA_MAX_LAYERS> swa_layers;
 
     // for State Space Models
     uint32_t ssm_d_conv  = 0;
@@ -163,6 +164,9 @@ struct llama_hparams {
     // for Classifiers
     uint32_t n_cls_out = 1;
 
+    // output embedding dimension (0 = use n_embd)
+    uint32_t n_embd_out = 0;
+
     // llama4 smallthinker
     uint32_t n_moe_layer_step        = 0;
     uint32_t n_no_rope_layer_step    = 4;
@@ -235,6 +239,9 @@ struct llama_hparams {
     // dimension of main + auxiliary input embeddings
     uint32_t n_embd_inp() const;
 
+    // dimension of output embeddings
+    uint32_t get_n_embd_out() 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-kv-cache.h

@@ -305,7 +305,7 @@ public:
             bool do_shift,
             stream_copy_info sc_info);
 
-    // used to create a batch procesing context from a batch
+    // used to create a batch processing context from a batch
     llama_kv_cache_context(
             llama_kv_cache * kv,
             slot_info_vec_t sinfos,

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

@@ -13,9 +13,10 @@
 #ifdef __has_include
     #if __has_include(<unistd.h>)
         #include <unistd.h>
+        #include <fcntl.h>
+        #include <sys/stat.h>
         #if defined(_POSIX_MAPPED_FILES)
             #include <sys/mman.h>
-            #include <fcntl.h>
         #endif
         #if defined(_POSIX_MEMLOCK_RANGE)
             #include <sys/resource.h>
@@ -74,7 +75,7 @@ struct llama_file::impl {
         return ret;
     }
 
-    impl(const char * fname, const char * mode) {
+    impl(const char * fname, const char * mode, [[maybe_unused]] const bool use_direct_io = false) {
         fp = ggml_fopen(fname, mode);
         if (fp == NULL) {
             throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
@@ -109,7 +110,7 @@ struct llama_file::impl {
         }
     }
 
-    void read_raw(void * ptr, size_t len) const {
+    void read_raw(void * ptr, size_t len) {
         size_t bytes_read = 0;
         while (bytes_read < len) {
             size_t chunk_size = std::min<size_t>(len - bytes_read, 64*1024*1024);
@@ -126,7 +127,7 @@ struct llama_file::impl {
         }
     }
 
-    uint32_t read_u32() const {
+    uint32_t read_u32() {
         uint32_t val;
         read_raw(&val, sizeof(val));
         return val;
@@ -153,16 +154,55 @@ struct llama_file::impl {
         write_raw(&val, sizeof(val));
     }
 
+    bool has_direct_io() const {
+        return true;
+    }
+
     ~impl() {
         if (fp) {
             std::fclose(fp);
         }
     }
 #else
-    impl(const char * fname, const char * mode) {
-        fp = ggml_fopen(fname, mode);
+    impl(const char * fname, const char * mode, [[maybe_unused]] const bool use_direct_io = false) : fname(fname) {
+#ifdef __linux__
+        // Try unbuffered I/O for read only
+        if (use_direct_io && std::strcmp(mode, "rb") == 0) {
+            if (init_fd()) {
+                return;
+            }
+            LLAMA_LOG_WARN("Failed to open file '%s' with error: %s. Falling back to buffered I/O",
+                           fname, strerror(errno));
+        }
+#endif
+        init_fp(mode);
+    }
+
+#ifdef __linux__
+    bool init_fd() {
+        fd = open(fname.c_str(), O_RDONLY | O_DIRECT);
+
+        if (fd != -1) {
+            struct stat file_stats{};
+            fstat(fd, &file_stats);
+
+            size = file_stats.st_size;
+            alignment = file_stats.st_blksize;
+
+            off_t ret = lseek(fd, 0, SEEK_SET);
+            if (ret == -1) {
+                throw std::runtime_error(format("seek error: %s", strerror(errno)));
+            }
+            return true;
+        }
+        return false;
+    }
+#endif
+
+    void init_fp(const char * mode) {
+        fp = ggml_fopen(fname.c_str(), mode);
         if (fp == NULL) {
-            throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
+            throw std::runtime_error(format("failed to open %s: %s", fname.c_str(), strerror(errno)));
         }
         seek(0, SEEK_END);
         size = tell();
@@ -170,46 +210,118 @@ struct llama_file::impl {
     }
 
     size_t tell() const {
-// TODO: this ifdef is never true?
-#ifdef _WIN32
-        __int64 ret = _ftelli64(fp);
-#else
-        long ret = std::ftell(fp);
-#endif
-        if (ret == -1) {
-            throw std::runtime_error(format("ftell error: %s", strerror(errno)));
+        if (fd == -1) {
+            long ret = std::ftell(fp);
+            if (ret == -1) {
+                throw std::runtime_error(format("ftell error: %s", strerror(errno)));
+            }
+
+            return (size_t) ret;
         }
 
-        return (size_t) ret;
+        off_t pos = lseek(fd, 0, SEEK_CUR);
+        if (pos == -1) {
+            throw std::runtime_error(format("lseek error: %s", strerror(errno)));
+        }
+        return (size_t) pos;
     }
 
     void seek(size_t offset, int whence) const {
-// TODO: this ifdef is never true?
-#ifdef _WIN32
-        int ret = _fseeki64(fp, (__int64) offset, whence);
-#else
-        int ret = std::fseek(fp, (long) offset, whence);
-#endif
-        if (ret != 0) {
+        off_t ret = 0;
+        if (fd == -1) {
+            ret = std::fseek(fp, (long) offset, whence);
+        } else {
+            ret = lseek(fd, offset, whence);
+        }
+        if (ret == -1) {
             throw std::runtime_error(format("seek error: %s", strerror(errno)));
         }
     }
 
-    void read_raw(void * ptr, size_t len) const {
+    void read_raw_unsafe(void * ptr, size_t len) {
         if (len == 0) {
             return;
         }
         errno = 0;
-        std::size_t ret = std::fread(ptr, len, 1, fp);
-        if (ferror(fp)) {
-            throw std::runtime_error(format("read error: %s", strerror(errno)));
-        }
-        if (ret != 1) {
-            throw std::runtime_error("unexpectedly reached end of file");
+        if (fd == -1) {
+            std::size_t ret = std::fread(ptr, len, 1, fp);
+            if (ferror(fp)) {
+                throw std::runtime_error(format("read error: %s", strerror(errno)));
+            }
+            if (ret != 1) {
+                throw std::runtime_error("unexpectedly reached end of file");
+            }
+        } else {
+            size_t bytes_read = 0;
+            while (bytes_read < len) {
+                const size_t to_read = len - bytes_read;
+                ssize_t ret = ::read(fd, reinterpret_cast<char *>(ptr) + bytes_read, to_read);
+
+                if (ret == -1) {
+                    if (errno == EINTR) {
+                        continue;  // Interrupted by signal, retry
+                    }
+                    // Fallback to std::fread in case the DMA controller cannot access the buffer
+                    if (errno == EFAULT) {
+                        auto curr_off = tell();
+                        close(fd);
+                        fd = -1;
+                        alignment = 1;
+                        init_fp("rb");
+                        seek(curr_off, SEEK_SET);
+                        read_raw_unsafe(ptr, len);
+                        return;
+                    }
+                    throw std::runtime_error(format("read error: %s", strerror(errno)));
+                }
+                if (ret == 0) {
+                    // EOF: allow if this read was only pulling alignment padding past file end
+                    off_t pos = lseek(fd, 0, SEEK_CUR);
+                    if (pos != -1 && (size_t) pos == size) {
+                        std::memset(reinterpret_cast<char *>(ptr) + bytes_read, 0, len - bytes_read);
+                        return;
+                    }
+                    throw std::runtime_error("unexpectedly reached end of file");
+                }
+
+                bytes_read += (size_t) ret;
+            }
         }
     }
 
-    uint32_t read_u32() const {
+    void read_aligned_chunk(void * dest, size_t size) {
+        size_t offset = tell();
+        off_t aligned_offset = offset & ~(alignment - 1);
+        off_t offset_from_alignment = offset - aligned_offset;
+        size_t bytes_to_read = (offset_from_alignment + size + alignment - 1) & ~(alignment - 1);
+
+        void * raw_buffer = nullptr;
+        int ret = posix_memalign(&raw_buffer, alignment, bytes_to_read);
+        if (ret != 0) {
+            throw std::runtime_error(format("posix_memalign failed with error %d", ret));
+        }
+
+        struct aligned_buffer_deleter {
+            void operator()(void * p) const { free(p); }
+        };
+        std::unique_ptr<void, aligned_buffer_deleter> buffer(raw_buffer);
+
+        seek(aligned_offset, SEEK_SET);
+        read_raw_unsafe(buffer.get(), bytes_to_read);
+
+        uintptr_t actual_data = reinterpret_cast<uintptr_t>(buffer.get()) + offset_from_alignment;
+        memcpy(dest, reinterpret_cast<void *>(actual_data), size);
+    }
+
+    void read_raw(void * ptr, size_t len) {
+        if (has_direct_io()) {
+            read_aligned_chunk(ptr, len);
+        } else {
+            read_raw_unsafe(ptr, len);
+        }
+    }
+
+    uint32_t read_u32() {
         uint32_t ret;
         read_raw(&ret, sizeof(ret));
         return ret;
@@ -230,23 +342,41 @@ struct llama_file::impl {
         write_raw(&val, sizeof(val));
     }
 
+    bool has_direct_io() const {
+        return fd != -1 && alignment > 1;
+    }
+
     ~impl() {
-        if (fp) {
+        if (fd != -1) {
+            close(fd);
+        } else {
             std::fclose(fp);
         }
     }
+    int fd = -1;
+    std::string fname;
 #endif
 
-    FILE * fp;
-    size_t size;
+    size_t read_alignment() const {
+        return alignment;
+    }
+
+    size_t alignment = 1;
+
+    FILE * fp{};
+    size_t size{};
 };
 
-llama_file::llama_file(const char * fname, const char * mode) : pimpl(std::make_unique<impl>(fname, mode)) {}
+llama_file::llama_file(const char * fname, const char * mode, const bool use_direct_io) :
+    pimpl(std::make_unique<impl>(fname, mode, use_direct_io)) {}
 llama_file::~llama_file() = default;
 
 size_t llama_file::tell() const { return pimpl->tell(); }
 size_t llama_file::size() const { return pimpl->size; }
 
+size_t llama_file::read_alignment() const { return pimpl->read_alignment(); }
+bool llama_file::has_direct_io() const { return pimpl->has_direct_io(); }
+
 int llama_file::file_id() const {
 #ifdef _WIN32
     return _fileno(pimpl->fp);
@@ -260,9 +390,14 @@ int llama_file::file_id() const {
 }
 
 void llama_file::seek(size_t offset, int whence) const { pimpl->seek(offset, whence); }
-void llama_file::read_raw(void * ptr, size_t len) const { pimpl->read_raw(ptr, len); }
+void llama_file::read_raw(void * ptr, size_t len) { pimpl->read_raw(ptr, len); }
+#ifdef _WIN32
+void llama_file::read_raw_unsafe(void * ptr, size_t len) { pimpl->read_raw(ptr, len); }
+#else
+void llama_file::read_raw_unsafe(void * ptr, size_t len) { pimpl->read_raw_unsafe(ptr, len); }
+#endif
 
-uint32_t llama_file::read_u32() const { return pimpl->read_u32(); }
+uint32_t llama_file::read_u32() { return pimpl->read_u32(); }
 
 void llama_file::write_raw(const void * ptr, size_t len) const { pimpl->write_raw(ptr, len); }
 void llama_file::write_u32(uint32_t val) const { pimpl->write_u32(val); }

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

@@ -3,6 +3,7 @@
 #include <cstdint>
 #include <memory>
 #include <vector>
+#include <cstdio>
 
 struct llama_file;
 struct llama_mmap;
@@ -13,7 +14,7 @@ using llama_mmaps  = std::vector<std::unique_ptr<llama_mmap>>;
 using llama_mlocks = std::vector<std::unique_ptr<llama_mlock>>;
 
 struct llama_file {
-    llama_file(const char * fname, const char * mode);
+    llama_file(const char * fname, const char * mode, bool use_direct_io = false);
     ~llama_file();
 
     size_t tell() const;
@@ -23,12 +24,16 @@ struct llama_file {
 
     void seek(size_t offset, int whence) const;
 
-    void read_raw(void * ptr, size_t len) const;
-    uint32_t read_u32() const;
+    void read_raw(void * ptr, size_t len);
+    void read_raw_unsafe(void * ptr, size_t len);
+    void read_aligned_chunk(void * dest, size_t size);
+    uint32_t read_u32();
 
     void write_raw(const void * ptr, size_t len) const;
     void write_u32(uint32_t val) const;
 
+    size_t read_alignment() const;
+    bool has_direct_io() const;
 private:
     struct impl;
     std::unique_ptr<impl> pimpl;

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

@@ -462,6 +462,29 @@ namespace GGUFMeta {
         return get_key_or_arr(llm_kv(kid), result, n, required);
     }
 
+    bool llama_model_loader::get_key_or_arr(enum llm_kv kid, uint32_t & result, bool required) {
+        const std::string key = llm_kv(kid);
+
+        const int id = gguf_find_key(meta.get(), key.c_str());
+
+        if (id < 0) {
+            if (required) {
+                throw std::runtime_error(format("key not found in model: %s", key.c_str()));
+            }
+            return false;
+        }
+
+        // throw and error if type is an array
+        if (gguf_get_kv_type(meta.get(), id) == GGUF_TYPE_ARRAY) {
+            if (required) {
+                throw std::runtime_error(format("expected scalar, found array for key: %s", key.c_str()));
+            }
+            return false;
+        }
+
+        return get_key(key, result, required);
+    }
+
     // TODO: this is not very clever - figure out something better
     template bool llama_model_loader::get_key_or_arr<std::array<int, 4>>(enum llm_kv kid, std::array<int, 4> & result, uint32_t n, bool required);
     template bool llama_model_loader::get_key_or_arr<std::array<uint32_t, 512>>(enum llm_kv kid, std::array<uint32_t, 512> & result, uint32_t n, bool required);
@@ -472,6 +495,7 @@ llama_model_loader::llama_model_loader(
         const std::string & fname,
         std::vector<std::string> & splits,
         bool use_mmap,
+        bool use_direct_io,
         bool check_tensors,
         bool no_alloc,
         const llama_model_kv_override * param_overrides_p,
@@ -504,9 +528,17 @@ llama_model_loader::llama_model_loader(
     get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
     llm_kv = LLM_KV(llm_arch_from_string(arch_name));
 
-    files.emplace_back(new llama_file(fname.c_str(), "rb"));
+    files.emplace_back(new llama_file(fname.c_str(), "rb", use_direct_io));
     contexts.emplace_back(ctx);
 
+    use_direct_io = use_direct_io && files.back()->has_direct_io();
+
+    // Disable mmap in case Direct I/O is enabled and available
+    if (use_direct_io && use_mmap) {
+        use_mmap = false;
+        LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
+    }
+
     // Save tensors data offset of the main file.
     // For subsidiary files, `meta` tensor data offset must not be used,
     // so we build a unified tensors index for weights.
@@ -572,7 +604,7 @@ llama_model_loader::llama_model_loader(
                 }
             }
 
-            files.emplace_back(new llama_file(fname_split, "rb"));
+            files.emplace_back(new llama_file(fname_split, "rb", use_direct_io));
             contexts.emplace_back(ctx);
 
             // Save tensors data offset info of the shard.
@@ -716,6 +748,7 @@ llama_model_loader::llama_model_loader(
     }
 
     this->use_mmap = use_mmap;
+    this->use_direct_io = use_direct_io;
     this->check_tensors = check_tensors;
     this->no_alloc = no_alloc;
 }
@@ -935,7 +968,15 @@ bool llama_model_loader::load_all_data(
     // 4 staging buffers for async uploads, each sized 1MB seems to be a good default for single NVMe drives.
     // NVMe raid configurations might require more / larger buffers.
     constexpr size_t n_buffers = 4;
-    constexpr size_t buffer_size = 1 * 1024 * 1024; // 1MB
+
+    size_t alignment = 1;
+    for (const auto & file : files) {
+        alignment = std::max(file->read_alignment(), alignment);
+    }
+
+    // Buffer size: balance between memory usage and I/O efficiency
+    // 64MB works well for NVMe drives
+    const size_t buffer_size = alignment != 1 ? 64 * 1024 * 1024 + 2 * alignment : 1 * 1024 * 1024;
 
     std::vector<ggml_backend_buffer_t> host_buffers;
     std::vector<ggml_backend_event_t> events;
@@ -985,6 +1026,7 @@ bool llama_model_loader::load_all_data(
         // If the backend is supported, create pinned memory buffers and events for synchronisation.
         for (size_t idx = 0; idx < n_buffers; ++idx) {
             auto * buf = ggml_backend_buft_alloc_buffer(host_buft, buffer_size);
+
             if (!buf) {
                 LLAMA_LOG_DEBUG("%s: failed to allocate host buffer for async uploads for device %s\n", func,
                     ggml_backend_dev_name(dev));
@@ -1066,6 +1108,7 @@ bool llama_model_loader::load_all_data(
             }
         } else {
             const auto & file = files.at(weight->idx);
+
             if (ggml_backend_buffer_is_host(cur->buffer)) {
                 file->seek(weight->offs, SEEK_SET);
                 file->read_raw(cur->data, n_size);
@@ -1077,19 +1120,54 @@ bool llama_model_loader::load_all_data(
             } else {
                 // If upload_backend is valid load the tensor in chunks to pinned memory and upload the buffers asynchronously to the GPU.
                 if (upload_backend) {
-                    file->seek(weight->offs, SEEK_SET);
+                    size_t offset = weight->offs;
+                    alignment = file->read_alignment();
+                    size_t aligned_offset = offset & ~(alignment - 1);
+                    size_t offset_from_alignment = offset - aligned_offset;
+                    file->seek(aligned_offset, SEEK_SET);
+
+                    // Calculate aligned read boundaries
+                    size_t read_start = aligned_offset;
+                    size_t read_end = (offset + n_size + alignment - 1) & ~(alignment - 1);
 
                     size_t bytes_read = 0;
+                    size_t data_read = 0;  // Actual tensor data copied (excluding padding)
 
-                    while (bytes_read < n_size) {
-                        size_t read_iteration = std::min<size_t>(buffer_size, n_size - bytes_read);
+                    while (bytes_read < read_end - read_start) {
+                        size_t read_size = std::min<size_t>(buffer_size, read_end - read_start - bytes_read);
 
+                        // Align the destination pointer within the pinned buffer
+                        uintptr_t ptr_dest_aligned = (reinterpret_cast<uintptr_t>(host_ptrs[buffer_idx]) + alignment - 1) & ~(alignment - 1);
+
+                        // Wait for previous upload to complete before reusing buffer
                         ggml_backend_event_synchronize(events[buffer_idx]);
-                        file->read_raw(host_ptrs[buffer_idx], read_iteration);
-                        ggml_backend_tensor_set_async(upload_backend, cur, host_ptrs[buffer_idx], bytes_read, read_iteration);
+
+                        // Read aligned chunk from file
+                        file->read_raw_unsafe(reinterpret_cast<void *>(ptr_dest_aligned), read_size);
+
+                        // Calculate actual data portion (excluding alignment padding)
+                        uintptr_t ptr_data = ptr_dest_aligned;
+                        size_t data_to_copy = read_size;
+
+                        // Skip alignment padding at start of first chunk
+                        if (bytes_read == 0) {
+                            ptr_data += offset_from_alignment;
+                            data_to_copy -= offset_from_alignment;
+                        }
+
+                        // Trim alignment padding at end of last chunk
+                        if (aligned_offset + bytes_read + read_size > offset + n_size) {
+                            data_to_copy -= (read_end - (offset + n_size));
+                        }
+
+                        // Async upload actual data to GPU
+                        ggml_backend_tensor_set_async(upload_backend, cur,
+                                                      reinterpret_cast<void *>(ptr_data), data_read, data_to_copy);
                         ggml_backend_event_record(events[buffer_idx], upload_backend);
 
-                        bytes_read += read_iteration;
+                        data_read += data_to_copy;
+                        bytes_read += read_size;
+
                         ++buffer_idx;
                         buffer_idx %= n_buffers;
                     }

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

@@ -70,6 +70,7 @@ struct llama_model_loader {
     size_t   n_bytes    = 0;
 
     bool use_mmap = false;
+    bool use_direct_io = false;
     bool check_tensors;
     bool no_alloc;
 
@@ -97,6 +98,7 @@ struct llama_model_loader {
         const std::string & fname,
         std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
         bool use_mmap,
+        bool use_direct_io,
         bool check_tensors,
         bool no_alloc,
         const llama_model_kv_override * param_overrides_p,
@@ -131,6 +133,8 @@ struct llama_model_loader {
     template<typename T>
     bool get_key_or_arr(enum llm_kv kid, T & result, uint32_t n, bool required = true);
 
+    bool get_key_or_arr(enum llm_kv kid, uint32_t & result, bool required = true);
+
     std::string get_arch_name() const;
 
     enum llm_arch get_arch() const;

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

@@ -146,6 +146,9 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_VOCAB_SIZE,                        vocab.n_tokens());
     add_kv(LLM_KV_CONTEXT_LENGTH,                    hparams.n_ctx_train);
     add_kv(LLM_KV_EMBEDDING_LENGTH,                  hparams.n_embd);
+    if (hparams.n_embd_out > 0) {
+        add_kv(LLM_KV_EMBEDDING_LENGTH_OUT,          hparams.n_embd_out);
+    }
     add_kv(LLM_KV_BLOCK_COUNT,                       hparams.n_layer);
     add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead);
     add_kv(LLM_KV_FEED_FORWARD_LENGTH,               hparams.n_ff_arr, true);

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

@@ -31,12 +31,14 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_17M:           return "17M";
         case LLM_TYPE_22M:           return "22M";
         case LLM_TYPE_33M:           return "33M";
+        case LLM_TYPE_47M:           return "47M";
         case LLM_TYPE_60M:           return "60M";
         case LLM_TYPE_70M:           return "70M";
         case LLM_TYPE_80M:           return "80M";
         case LLM_TYPE_109M:          return "109M";
         case LLM_TYPE_137M:          return "137M";
         case LLM_TYPE_140M:          return "140M";
+        case LLM_TYPE_149M:          return "149M";
         case LLM_TYPE_160M:          return "160M";
         case LLM_TYPE_190M:          return "190M";
         case LLM_TYPE_220M:          return "220M";
@@ -46,6 +48,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_335M:          return "335M";
         case LLM_TYPE_350M:          return "350M";
         case LLM_TYPE_360M:          return "360M";
+        case LLM_TYPE_395M:          return "395M";
         case LLM_TYPE_410M:          return "410M";
         case LLM_TYPE_450M:          return "450M";
         case LLM_TYPE_475M:          return "475M";
@@ -123,10 +126,12 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_31B_A3_5B:     return "31B.A3.5B";
         case LLM_TYPE_80B_A3B:       return "80B.A3B";
         case LLM_TYPE_100B_A6B:      return "100B.A6B";
+        case LLM_TYPE_102B_A12B:     return "102B.A12B";
         case LLM_TYPE_106B_A12B:     return "106B.A12B";
         case LLM_TYPE_230B_A10B:     return "230B.A10B";
         case LLM_TYPE_235B_A22B:     return "235B.A22B";
         case LLM_TYPE_300B_A47B:     return "300B.A47B";
+        case LLM_TYPE_310B_A15B:     return "310B.A15B";
         case LLM_TYPE_355B_A32B:     return "355B.A32B";
         case LLM_TYPE_E2B:           return "E2B";
         case LLM_TYPE_E4B:           return "E4B";
@@ -502,6 +507,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
     ml.get_key(LLM_KV_CONTEXT_LENGTH,          hparams.n_ctx_train);
     ml.get_key(LLM_KV_EMBEDDING_LENGTH,        hparams.n_embd);
+    ml.get_key(LLM_KV_EMBEDDING_LENGTH_OUT,    hparams.n_embd_out, false);
     ml.get_key(LLM_KV_BLOCK_COUNT,             hparams.n_layer);
     ml.get_key(LLM_KV_EXPERT_COUNT,            hparams.n_expert,        false);
     ml.get_key(LLM_KV_EXPERT_USED_COUNT,       hparams.n_expert_used,   false);
@@ -573,6 +579,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     hparams.rope_scaling_type_train = llama_rope_scaling_type_from_string(rope_scaling);
     GGML_ASSERT(hparams.rope_scaling_type_train != LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED);
 
+    // TODO: Handle SWA metadata similarly when models start implementing it
     // rope_freq_scale (inverse of the kv) is optional
     float ropescale = 0.0f;
     if (!ml.get_key(LLM_KV_ROPE_SCALING_FACTOR, ropescale, false)) {
@@ -581,10 +588,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     }
     hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale;
 
-    // by default assume that the sliding-window layers use the same scaling type as the non-sliding-window layers
-    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
-    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
-
     ml.get_key(LLM_KV_ROPE_SCALING_ATTN_FACTOR, hparams.rope_attn_factor, false);
 
     // non-transformer models do not have attention heads
@@ -603,7 +606,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
         ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
 
-        if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON) {
+        if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON || arch == LLM_ARCH_LLAMA_EMBED) {
             if (hparams.n_rot != hparams.n_embd_head_k) {
                 throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
             }
@@ -627,6 +630,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     // arch-specific KVs
     switch (arch) {
         case LLM_ARCH_LLAMA:
+        case LLM_ARCH_LLAMA_EMBED:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
@@ -671,6 +675,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.f_attn_temp_scale       = 0.1f;
                     hparams.f_attn_temp_offset      = 1.0f;
                     hparams.set_swa_pattern(4);   // pattern: 3 chunked - 1 full
+
+                    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 }
 
                 switch (hparams.n_expert) {
@@ -716,6 +724,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 if (hparams.n_swa > 0) {
                     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                     hparams.set_swa_pattern(4);
+
+                    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 } else {
                     hparams.swa_type = LLAMA_SWA_TYPE_NONE;
                 }
@@ -875,6 +887,34 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_MODERN_BERT:
+            {
+                const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+                if (found_swa && hparams.n_swa > 0) {
+                    uint32_t swa_period = 3;
+                    hparams.swa_type = LLAMA_SWA_TYPE_SYMMETRIC;
+
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa);
+                    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+                    hparams.set_swa_pattern(swa_period);
+                } else {
+                    hparams.swa_type = LLAMA_SWA_TYPE_NONE;
+                }
+
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+                ml.get_key(LLM_KV_ATTENTION_CAUSAL,        hparams.causal_attn);
+                ml.get_key(LLM_KV_POOLING_TYPE,            hparams.pooling_type, false);
+
+                switch (hparams.n_layer) {
+                    case 12:
+                        type = LLM_TYPE_47M; break; // granite-embedding-small
+                    case 22:
+                        type = LLM_TYPE_149M; break; // modern-bert-base
+                    case 28:
+                        type = LLM_TYPE_395M; break; // modern-bert-large
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_JINA_BERT_V2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
@@ -1076,6 +1116,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_MAINCODER:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 32: type = LLM_TYPE_1B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_QWEN3VL:
             {
                 ml.get_key(LLM_KV_NUM_DEEPSTACK_LAYERS, hparams.n_deepstack_layers, false);
@@ -1194,6 +1242,25 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH,   hparams.n_embd_head_k, false);
                 ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false);
             } break;
+        case LLM_ARCH_PLAMO3:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+                if (found_swa && hparams.n_swa > 0) {
+                    uint32_t swa_period = 8;
+                    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa);
+                    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+                    hparams.set_swa_pattern(swa_period);
+                } else {
+                    hparams.swa_type = LLAMA_SWA_TYPE_NONE;
+                }
+
+                switch (hparams.n_layer) {
+                    case 24: type = LLM_TYPE_2B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_GPT2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -1247,7 +1314,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.n_swa = 4096; // default value of gemma 2
                 hparams.set_swa_pattern(2);
                 hparams.attn_soft_cap = true;
+                hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
 
+                ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,          hparams.rope_freq_base_train_swa, false);
                 ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING,      hparams.f_attn_logit_softcapping, false);
@@ -1272,8 +1342,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                     hparams.set_swa_pattern(6);
 
-                    hparams.rope_freq_base_train_swa  = 10000.0f;
-                    hparams.rope_freq_scale_train_swa = 1.0f;
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 } else {
                     hparams.swa_type = LLAMA_SWA_TYPE_NONE;
                 }
@@ -1303,10 +1372,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.set_swa_pattern(5);
 
                 hparams.n_layer_kv_from_start     = 20;
-                hparams.rope_freq_base_train_swa  = 10000.0f;
-                hparams.rope_freq_scale_train_swa = 1.0f;
                 hparams.f_attention_scale         = 1.0f;
 
+                ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,          hparams.rope_freq_base_train_swa, false);
                 ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
@@ -1322,9 +1390,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.set_swa_pattern(6);
 
                 hparams.causal_attn = false; // embeddings do not use causal attention
-                hparams.rope_freq_base_train_swa = 10000.0f;
-                hparams.rope_freq_scale_train_swa = 1.0f;
 
+                ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type);
@@ -1463,7 +1530,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             {
                 hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                 hparams.set_swa_pattern(4);
+                hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
 
+                ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,       hparams.rope_freq_base_train_swa, false);
                 ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
                 ml.get_key(LLM_KV_LOGIT_SCALE,              hparams.f_logit_scale);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,  hparams.f_norm_eps);
@@ -1502,6 +1572,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 if (found_swa && hparams.n_swa > 0) {
                     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                     hparams.set_swa_pattern(4);
+
+                    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                    hparams.rope_freq_scale_train_swa = 1.0; // See olmo2.cpp
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 } else {
                     hparams.swa_type = LLAMA_SWA_TYPE_NONE;
                 }
@@ -1629,7 +1703,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla, false);
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
                 ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
-                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);
                 ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,        hparams.expert_weights_norm, false);
                 ml.get_key(LLM_KV_EXPERT_GATING_FUNC,         hparams.expert_gating_func, false);
                 if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
@@ -1725,6 +1799,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 switch (hparams.n_layer) {
                     case 47: type = LLM_TYPE_106B_A12B; break; // GLM-4.5-Air (46 layers + 1 NextN layer)
+                    case 48: type = LLM_TYPE_102B_A12B; break; // Solar Open
                     case 93: type = LLM_TYPE_355B_A32B; break; // GLM-4.5 (92 layers + 1 NextN layer)
                     default: type = LLM_TYPE_UNKNOWN;
                 }
@@ -1843,6 +1918,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                     hparams.n_swa = 4096;
                     hparams.set_swa_pattern(4);
+
+                    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 }
 
                 ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
@@ -2160,6 +2239,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                 hparams.set_swa_pattern(2);
 
+                hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+                ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+
                 switch (hparams.n_layer) {
                     case 24: type = LLM_TYPE_20B; break;
                     case 36: type = LLM_TYPE_120B; break;
@@ -2204,6 +2287,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.swa_type      = LLAMA_SWA_TYPE_STANDARD;
                     hparams.n_swa         = 4096;
                     hparams.set_swa_pattern(4, true);
+
+                    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 } else {
                     hparams.swa_type             = LLAMA_SWA_TYPE_NONE;
                     hparams.n_no_rope_layer_step = hparams.n_layer;
@@ -2322,6 +2409,22 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_MIMO2:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,   hparams.n_swa);
+                ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,         hparams.rope_freq_base_train_swa);
+                ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer);
+
+                switch (hparams.n_layer) {
+                    case 48: type = LLM_TYPE_310B_A15B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         default: throw std::runtime_error("unsupported model architecture");
     }
 
@@ -2344,15 +2447,16 @@ void llama_model::load_vocab(llama_model_loader & ml) {
 
 bool llama_model::load_tensors(llama_model_loader & ml) {
     const auto & split_mode   = params.split_mode;
-    const auto & n_gpu_layers = params.n_gpu_layers;
     const auto & use_mlock    = params.use_mlock;
     const auto & tensor_split = params.tensor_split;
 
-    const int n_layer = hparams.n_layer;
+    const int n_layer      = hparams.n_layer;
+    const int n_gpu_layers = this->n_gpu_layers();
 
     const bool use_mmap_buffer = true;
 
-    LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (mmap = %s)\n", __func__, ml.use_mmap ? "true" : "false");
+    LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (mmap = %s, direct_io = %s)\n",
+        __func__, ml.use_mmap ? "true" : "false", ml.use_direct_io ? "true" : "false");
 
     // build a list of buffer types for the CPU and GPU devices
     pimpl->cpu_buft_list = make_cpu_buft_list(devices, params.use_extra_bufts, params.no_host);
@@ -2363,6 +2467,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         pimpl->gpu_buft_list.emplace(dev, std::move(buft_list));
     }
 
+    ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    if (cpu_dev == nullptr) {
+        throw std::runtime_error(format("%s: no CPU backend found", __func__));
+    }
+
     // calculate the split points
     bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + n_devices(), [](float x) { return x == 0.0f; });
     std::vector<float> splits(n_devices());
@@ -2373,6 +2482,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             size_t total;
             size_t free;
             ggml_backend_dev_memory(dev, &free, &total);
+
+            // devices can return 0 bytes for free and total memory if they do not
+            // have any to report. in this case, we will use the host memory as a fallback
+            // fixes: https://github.com/ggml-org/llama.cpp/issues/18577
+            if (free == 0 && total == 0) {
+                ggml_backend_dev_memory(cpu_dev, &free, &total);
+            }
             splits[i] = free;
         }
     } else {
@@ -2389,14 +2505,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         splits[i] /= split_sum;
     }
 
-    ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    if (cpu_dev == nullptr) {
-        throw std::runtime_error(format("%s: no CPU backend found", __func__));
-    }
-    const int i_gpu_start = std::max((int) hparams.n_layer - n_gpu_layers, (int) 0);
-    const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, (int)n_layer + 1);
+    const int i_gpu_start = std::max(int(hparams.n_layer) + 1 - n_gpu_layers, 0);
+    const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, int(n_layer) + 1);
     auto get_layer_buft_list = [&](int il) -> llama_model::impl::layer_dev {
-        const bool is_swa = il < (int) hparams.n_layer && hparams.is_swa(il);
+        const bool is_swa = il < int(hparams.n_layer) && hparams.is_swa(il);
         if (il < i_gpu_start || (il - i_gpu_start) >= act_gpu_layers) {
             LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s, is_swa = %d\n", il, ggml_backend_dev_name(cpu_dev), is_swa);
             return {cpu_dev, &pimpl->cpu_buft_list};
@@ -2636,6 +2748,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             case LLM_ARCH_GRANITE:
             case LLM_ARCH_GRANITE_MOE:
             case LLM_ARCH_MISTRAL3:
+            case LLM_ARCH_LLAMA_EMBED:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -3170,6 +3283,37 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_MODERN_BERT:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+                    tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
+
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+                    for(int i = 0; i < n_layer; ++i) {
+                        auto& layer = layers[i];
+
+                        if ( i != 0 ) {
+                            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        } else{
+                            // layer 0 uses identity
+                            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                        }
+
+
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, 3 * n_embd }, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT,   "weight", i), {n_embd, n_embd}, 0);
+
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, 2 * n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                    }
+
+                    cls       = create_tensor(tn(LLM_TENSOR_CLS,     "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
+                    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
+                    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
+
+                } break;
             case LLM_ARCH_NEO_BERT:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
@@ -3234,7 +3378,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.attn_norm_2_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
 
                         layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
-                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, layer.ffn_gate ? n_ff : n_ff * 2}, 0);
+
+                        const auto tn_ffn_up_weight = tn(LLM_TENSOR_FFN_UP, "weight", i);
+                        ggml_tensor * t_ffn_up = ml.get_tensor_meta(tn_ffn_up_weight.str().c_str());
+                        const int64_t n_ffn_up = t_ffn_up ? t_ffn_up->ne[1] : n_ff;
+
+                        GGML_ASSERT(n_ffn_up == n_ff || n_ffn_up == n_ff * 2);
+                        layer.ffn_up   = create_tensor(tn_ffn_up_weight, {n_embd, n_ffn_up}, 0);
+                        layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ffn_up}, TENSOR_NOT_REQUIRED);
 
                         layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
                         layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, 0);
@@ -3762,6 +3913,44 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, i), {n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_PLAMO3:
+                {
+                    const int64_t head_dim_q = hparams.n_embd_head_k;
+                    const int64_t head_dim_v = hparams.n_embd_head_v;
+
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        const int64_t num_attention_heads = hparams.n_head(i);
+                        const int64_t num_key_value_heads = hparams.n_head_kv(i);
+                        const int64_t q_proj_dim = num_attention_heads * head_dim_q;
+                        const int64_t k_proj_dim = num_key_value_heads * head_dim_q;
+                        const int64_t v_proj_dim = num_key_value_heads * head_dim_v;
+                        const int64_t n_ff_cur   = hparams.n_ff(i);
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i),
+                                {n_embd,q_proj_dim + k_proj_dim + v_proj_dim}, 0);
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {head_dim_q}, 0);
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {head_dim_q}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {num_attention_heads * head_dim_v, n_embd}, 0);
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, i), {n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, i), {n_embd}, 0);
+
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff_cur * 2}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff_cur, n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_GPT2:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -4652,7 +4841,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                     // output
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+                    // try to load output.weight, if not found, use token_embd (tied embeddings)
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    if (!output) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -4715,7 +4908,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                     // output
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+                    // try to load output.weight, if not found, use token_embd (tied embeddings)
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    if (!output) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -5082,9 +5279,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, flags);
                         layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, flags);
                         layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, flags);
-                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), { n_embd_head_k * n_head }, flags);
-                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), { n_embd_k_gqa }, flags);
-                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), { n_embd_v_gqa }, flags);
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), { n_embd_head_k * n_head }, TENSOR_NOT_REQUIRED | flags);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), { n_embd_k_gqa }, TENSOR_NOT_REQUIRED | flags);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), { n_embd_v_gqa }, TENSOR_NOT_REQUIRED | flags);
 
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, flags);
 
@@ -5196,9 +5393,6 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_group    = hparams.ssm_n_group;
                     const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
 
-                    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
-                    const int64_t n_ff_shexp = hparams.n_ff_shexp;
-
                     // embeddings
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -5250,6 +5444,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd},         TENSOR_NOT_REQUIRED);
                         }  else {
                             if (n_expert != 0) {
+                                const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+                                const int64_t n_ff_shexp = hparams.n_ff_shexp;
+
                                 layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, 0);
                                 layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, 0);
 
@@ -6279,8 +6476,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
-                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM_LFM2, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,           "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
 
                     if (output == NULL) {
                         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
@@ -6325,6 +6522,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.shortconv.out_proj = create_tensor(tn(LLM_TENSOR_SHORTCONV_OUTPROJ, "weight", i), {n_embd, n_embd}, 0);
                         }
                     }
+
+                    // for LFM2-ColBert-350M
+                    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.get_n_embd_out()}, TENSOR_NOT_REQUIRED);
                 } break;
             case LLM_ARCH_SMALLTHINKER:
                 {
@@ -6627,6 +6827,75 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_down_shexp     = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP,     "weight", i), { hparams.n_ff_shexp, n_embd }, 0);
                     }
                 } break;
+            case LLM_ARCH_MIMO2:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+                        uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i);
+                        uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i);
+                        uint32_t n_head = hparams.n_head(i);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_v * n_head, n_embd }, 0);
+
+                        layer.attn_norm  = create_tensor(tn(LLM_TENSOR_ATTN_NORM,  "weight", i), {n_embd}, 0);
+                        layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, TENSOR_NOT_REQUIRED);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        // non-MoE branch
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, TENSOR_NOT_REQUIRED);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED);
+
+                        // MoE branch
+                        int64_t n_ff_exp = hparams.n_ff_exp;
+                        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);
+                        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp,   n_expert}, TENSOR_NOT_REQUIRED);
+                        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, TENSOR_NOT_REQUIRED);
+                        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff_exp,   n_expert}, TENSOR_NOT_REQUIRED);
+                        layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+                    }
+                } break;
+            case LLM_ARCH_MAINCODER:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -6736,10 +7005,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     if (llama_supports_gpu_offload()) {
         const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
 
-        LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
-        if (n_gpu_layers > (int) hparams.n_layer) {
+        int n_repeating = n_gpu;
+        if (n_repeating > 0) {
             LLAMA_LOG_INFO("%s: offloading output layer to GPU\n", __func__);
+            n_repeating--;
         }
+        LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_repeating);
 
         const int max_backend_supported_layers = hparams.n_layer + 1;
         const int max_offloadable_layers       = hparams.n_layer + 1;
@@ -6806,6 +7077,14 @@ size_t llama_model::n_devices() const {
     return devices.size();
 }
 
+uint32_t llama_model::n_gpu_layers() const {
+    return params.n_gpu_layers >= 0 ? params.n_gpu_layers : hparams.n_layer + 1;
+}
+
+llama_split_mode llama_model::split_mode() const {
+    return params.split_mode;
+}
+
 std::map<ggml_backend_buffer_type_t, size_t> llama_model::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
     for (const auto & [ctx, bufs] : pimpl->ctxs_bufs) {
@@ -6898,6 +7177,10 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type.c_str());
         LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
         LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
+        if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+            LLAMA_LOG_INFO("%s: freq_base_swa    = %.1f\n",   __func__, hparams.rope_freq_base_train_swa);
+            LLAMA_LOG_INFO("%s: freq_scale_swa   = %g\n",     __func__, hparams.rope_freq_scale_train_swa);
+        }
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
         LLAMA_LOG_INFO("%s: rope_yarn_log_mul= %.4f\n",   __func__, hparams.rope_yarn_log_mul);
         LLAMA_LOG_INFO("%s: rope_finetuned   = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
@@ -7130,6 +7413,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
         case LLM_ARCH_NOMIC_BERT_MOE:
         case LLM_ARCH_NEO_BERT:
         case LLM_ARCH_WAVTOKENIZER_DEC:
+        case LLM_ARCH_MODERN_BERT:
         case LLM_ARCH_GEMMA_EMBEDDING:
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
@@ -7247,16 +7531,24 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
     switch (arch) {
         case LLM_ARCH_LLAMA:
             {
-                llm = std::make_unique<llm_build_llama>(*this, params);
+                llm = std::make_unique<llm_build_llama<false>>(*this, params);
             } break;
         case LLM_ARCH_LLAMA4:
             {
                 if (hparams.swa_type == LLAMA_SWA_TYPE_NONE) {
-                    llm = std::make_unique<llm_build_llama>(*this, params);
+                    llm = std::make_unique<llm_build_llama<false>>(*this, params);
                 } else {
                     llm = std::make_unique<llm_build_llama_iswa>(*this, params);
                 }
             } break;
+        case LLM_ARCH_LLAMA_EMBED:
+            {
+                llm = std::make_unique<llm_build_llama<true>>(*this, params);
+            } break;
+        case LLM_ARCH_MAINCODER:
+            {
+                llm = std::make_unique<llm_build_maincoder>(*this, params);
+            } break;
         case LLM_ARCH_DECI:
             {
                 llm = std::make_unique<llm_build_deci>(*this, params);
@@ -7289,6 +7581,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_bert>(*this, params);
             } break;
+        case LLM_ARCH_MODERN_BERT:
+            {
+                llm = std::make_unique<llm_build_modern_bert>(*this, params);
+            } break;
         case LLM_ARCH_NEO_BERT:
             {
                 llm = std::make_unique<llm_build_neo_bert>(*this, params);
@@ -7378,6 +7674,14 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_plamo2>(*this, params);
             } break;
+        case LLM_ARCH_PLAMO3:
+            {
+                if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+                    llm = std::make_unique<llm_build_plamo3<true>> (*this, params);
+                } else {
+                    llm = std::make_unique<llm_build_plamo3<false>>(*this, params);
+                }
+            } break;
         case LLM_ARCH_GPT2:
             {
                 llm = std::make_unique<llm_build_gpt2>(*this, params);
@@ -7682,6 +7986,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_mistral3>(*this, params);
             } break;
+        case LLM_ARCH_MIMO2:
+            {
+                llm = std::make_unique<llm_build_mimo2_iswa>(*this, params);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -7689,12 +7997,17 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
     // add on pooling layer
     llm->build_pooling(cls, cls_b, cls_out, cls_out_b);
 
+    // add backend sampling layers (if any)
+    llm->build_sampling();
+
     // if the gguf model was converted with --sentence-transformers-dense-modules
     // there will be two additional dense projection layers
     // dense linear projections are applied after pooling
     // TODO: move reranking logic here and generalize
     llm->build_dense_out(dense_2_out_layers, dense_3_out_layers);
 
+    llm->res->set_outputs();
+
     return llm->res->get_gf();
 }
 
@@ -7707,7 +8020,7 @@ llama_model_params llama_model_default_params() {
     llama_model_params result = {
         /*.devices                     =*/ nullptr,
         /*.tensor_buft_overrides       =*/ nullptr,
-        /*.n_gpu_layers                =*/ 999,
+        /*.n_gpu_layers                =*/ -1,
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ nullptr,
@@ -7716,6 +8029,7 @@ llama_model_params llama_model_default_params() {
         /*.kv_overrides                =*/ nullptr,
         /*.vocab_only                  =*/ false,
         /*.use_mmap                    =*/ true,
+        /*.use_direct_io               =*/ true,
         /*.use_mlock                   =*/ false,
         /*.check_tensors               =*/ false,
         /*.use_extra_bufts             =*/ true,
@@ -7750,6 +8064,10 @@ int32_t llama_model_n_embd_inp(const llama_model * model) {
     return model->hparams.n_embd_inp();
 }
 
+int32_t llama_model_n_embd_out(const llama_model * model) {
+    return model->hparams.get_n_embd_out();
+}
+
 int32_t llama_model_n_layer(const llama_model * model) {
     return model->hparams.n_layer;
 }
@@ -7853,6 +8171,8 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ERNIE4_5:
         case LLM_ARCH_ERNIE4_5_MOE:
         case LLM_ARCH_MISTRAL3:
+        case LLM_ARCH_LLAMA_EMBED:
+        case LLM_ARCH_MAINCODER:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2
@@ -7862,6 +8182,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_DBRX:
         case LLM_ARCH_BERT:
         case LLM_ARCH_JINA_BERT_V3:
+        case LLM_ARCH_MODERN_BERT:
         case LLM_ARCH_NOMIC_BERT:
         case LLM_ARCH_NOMIC_BERT_MOE:
         case LLM_ARCH_STABLELM:
@@ -7881,6 +8202,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_PHIMOE:
         case LLM_ARCH_PLAMO:
         case LLM_ARCH_PLAMO2:
+        case LLM_ARCH_PLAMO3:
         case LLM_ARCH_GEMMA:
         case LLM_ARCH_GEMMA2:
         case LLM_ARCH_GEMMA3:
@@ -7911,6 +8233,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_PANGU_EMBED:
         case LLM_ARCH_AFMOE:
         case LLM_ARCH_QWEN3NEXT:
+        case LLM_ARCH_MIMO2:
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:

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

@@ -24,12 +24,14 @@ enum llm_type {
     LLM_TYPE_17M,
     LLM_TYPE_22M,
     LLM_TYPE_33M,
+    LLM_TYPE_47M,
     LLM_TYPE_60M,
     LLM_TYPE_70M,
     LLM_TYPE_80M,
     LLM_TYPE_109M,
     LLM_TYPE_137M,
     LLM_TYPE_140M,
+    LLM_TYPE_149M,
     LLM_TYPE_160M,
     LLM_TYPE_190M,
     LLM_TYPE_220M,
@@ -39,6 +41,7 @@ enum llm_type {
     LLM_TYPE_335M,
     LLM_TYPE_350M,
     LLM_TYPE_360M,
+    LLM_TYPE_395M,
     LLM_TYPE_410M,
     LLM_TYPE_450M,
     LLM_TYPE_475M,
@@ -117,10 +120,12 @@ enum llm_type {
     LLM_TYPE_31B_A3_5B,
     LLM_TYPE_80B_A3B, // Qwen3 Next
     LLM_TYPE_100B_A6B,
+    LLM_TYPE_102B_A12B, // Solar-Open
     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_310B_A15B, // /MiMo-V2-Flash
     LLM_TYPE_355B_A32B, // GLM-4.5
     LLM_TYPE_E2B,
     LLM_TYPE_E4B,
@@ -465,8 +470,6 @@ struct llama_model {
     struct ggml_tensor * dense_2_out_layers = nullptr;
     struct ggml_tensor * dense_3_out_layers = nullptr;
 
-    llama_model_params params;
-
     // gguf metadata
     std::unordered_map<std::string, std::string> gguf_kv;
 
@@ -476,6 +479,9 @@ struct llama_model {
     // for quantize-stats only
     std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
 
+    // for keeping track of extra nodes used by lora adapters
+    uint32_t n_lora_nodes = 0;
+
     int64_t t_load_us  = 0;
     int64_t t_start_us = 0;
 
@@ -497,6 +503,9 @@ struct llama_model {
     size_t n_tensors() const;
     size_t n_devices() const;
 
+    uint32_t n_gpu_layers() const;
+    llama_split_mode split_mode() const;
+
     std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const;
 
     // total number of parameters in the model
@@ -525,6 +534,8 @@ struct llama_model {
     ggml_cgraph * build_graph(const llm_graph_params & params) const;
 
 private:
+    llama_model_params params;
+
     struct impl;
     std::unique_ptr<impl> pimpl;
 };

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

@@ -596,7 +596,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     }
 
     std::vector<std::string> splits = {};
-    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr);
+    llama_model_loader ml(fname_inp, splits, use_mmap, /*use_direct_io*/ true, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr);
     ml.init_mappings(false); // no prefetching
 
     llama_model model(llama_model_default_params());

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

@@ -14,7 +14,19 @@ struct llama_grammar;
 struct llama_sampler_chain {
     llama_sampler_chain_params params;
 
-    std::vector<struct llama_sampler *> samplers;
+    // has .backend_init() been called?
+    bool is_init = false;
+
+    struct info {
+        bool is_backend;
+
+        llama_sampler * ptr;
+    };
+
+    std::vector<info> samplers;
+
+    // pre-allocated buffer for llama_sampler_sample to avoid repeated allocations
+    std::vector<llama_token_data> cur;
 
     // timing
 
@@ -24,9 +36,9 @@ struct llama_sampler_chain {
 };
 
 struct llama_sampler * llama_sampler_init_dry_testing(
-                         int32_t   context_size,
-                           float   dry_multiplier,
-                           float   dry_base,
-                         int32_t   dry_allowed_length,
-                         int32_t   dry_penalty_last_n,
-  const std::vector<std::vector<llama_token>>& seq_breakers);
+        int32_t context_size,
+        float   dry_multiplier,
+        float   dry_base,
+        int32_t dry_allowed_length,
+        int32_t dry_penalty_last_n,
+        const std::vector<std::vector<llama_token>> & seq_breakers);

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

@@ -314,6 +314,12 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][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;
+            case LLAMA_VOCAB_PRE_TYPE_YOUTU:
+                regex_exprs = {
+                    "[가-힣ㄱ-ㆎ]+|[！…“”‘’—：；，、-〿︰-﹏]+|[ㄅ-ㄯ]+|[一-龥぀-ゟ゠-ヿ]+",
+                    "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER:
                 regex_exprs = {
                     "[\r\n]",
@@ -355,6 +361,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
             case LLAMA_VOCAB_PRE_TYPE_STABLELM2:
             case LLAMA_VOCAB_PRE_TYPE_QWEN2:
             case LLAMA_VOCAB_PRE_TYPE_HUNYUAN:
+            case LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN:
                 regex_exprs = {
                     // original regex from tokenizer.json
                     // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
@@ -1849,6 +1856,11 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "deepseek-v3") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM;
                 clean_spaces = false;
+            } else if (
+                    tokenizer_pre == "youtu") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_YOUTU;
+                clean_spaces = false;
+                ignore_merges = true;
             } else if (
                     tokenizer_pre == "falcon") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_FALCON;
@@ -1867,7 +1879,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "jina-v2-es" ||
                     tokenizer_pre == "jina-v2-de" ||
                     tokenizer_pre == "a.x-4.0" ||
-                    tokenizer_pre == "mellum") {
+                    tokenizer_pre == "mellum"  ||
+                    tokenizer_pre == "modern-bert" ) {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                     tokenizer_pre == "jina-v1-en" ||
@@ -2003,6 +2016,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "minimax-m2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "solar-open") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN;
+                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;
@@ -2176,6 +2193,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
         //       for now, we apply this workaround to find the tokens based on their text
 
         for (const auto & t : token_to_id) {
+            auto & attr = id_to_token[t.second].attr;
+
             // find EOT token: "<|eot_id|>", "<|im_end|>", "<end_of_turn>", etc.
             if (special_eot_id == LLAMA_TOKEN_NULL) {
                 if (false
@@ -2191,10 +2210,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<end_of_utterance>" // smoldocling
                    ) {
                     special_eot_id = t.second;
-                    if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                    if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
                         LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
                                 __func__, t.second, t.first.c_str());
-                        id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                        attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_CONTROL);
                     }
                 }
             }
@@ -2205,10 +2224,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<|eom_id|>"
                         ) {
                     special_eom_id = t.second;
-                    if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                    if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
                         LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
                                 __func__, t.second, t.first.c_str());
-                        id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                        attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_CONTROL);
                     }
                 }
             }
@@ -2225,10 +2244,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<|code_prefix|>" // GLM-4.5
                         ) {
                     special_fim_pre_id = t.second;
-                    if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                    if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
                         LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
                                 __func__, t.second, t.first.c_str());
-                        id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                        attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_CONTROL);
                     }
                 }
             }
@@ -2245,10 +2264,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<|code_suffix|>" // GLM-4.5
                         ) {
                     special_fim_suf_id = t.second;
-                    if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                    if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
                         LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
                                 __func__, t.second, t.first.c_str());
-                        id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                        attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_CONTROL);
                     }
                 }
             }
@@ -2265,10 +2284,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<|code_middle|>" // GLM-4.5
                         ) {
                     special_fim_mid_id = t.second;
-                    if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                    if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
                         LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
                                 __func__, t.second, t.first.c_str());
-                        id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                        attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_CONTROL);
                     }
                 }
             }
@@ -2282,10 +2301,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<PAD>"
                         ) {
                     special_fim_pad_id = t.second;
-                    if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                    if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
                         LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
                                 __func__, t.second, t.first.c_str());
-                        id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                        attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_CONTROL);
                     }
                 }
             }
@@ -2300,10 +2319,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<reponame>"    // Granite
                         ) {
                     special_fim_rep_id = t.second;
-                    if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                    if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
                         LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
                                 __func__, t.second, t.first.c_str());
-                        id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                        attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_CONTROL);
                     }
                 }
             }
@@ -2314,15 +2333,41 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<|file_sep|>" // Qwen
                         ) {
                     special_fim_sep_id = t.second;
-                    if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                    if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
                         LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
                                 __func__, t.second, t.first.c_str());
-                        id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                        attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_CONTROL);
                     }
                 }
             }
         }
 
+        // auto-detect unused tokens: e.g. control tokens with the word "unused"
+        // ideally, these tokens should be marked as unused during conversion
+        {
+            uint32_t n_unused = 0;
+
+            for (const auto & t : token_to_id) {
+                auto & attr = id_to_token[t.second].attr;
+
+                if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                    continue;
+                }
+
+                if ((attr & LLAMA_TOKEN_ATTR_UNUSED) == 0) {
+                    if (strstr(t.first.c_str(), "unused") != NULL) {
+                        attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_UNUSED);
+                    }
+                }
+
+                if (attr & LLAMA_TOKEN_ATTR_UNUSED) {
+                    n_unused++;
+                }
+            }
+
+            LLAMA_LOG_INFO("%s: %u unused tokens\n", __func__, n_unused);
+        }
+
         // maintain a list of tokens that cause end-of-generation
         // this is currently determined based on the token text, which is obviously not ideal
         // ref: https://github.com/ggerganov/llama.cpp/issues/9606
@@ -2341,12 +2386,16 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
         }
 
         for (const auto & t : token_to_id) {
+            auto & attr = id_to_token[t.second].attr;
+
             if (false
                     || t.first == "<|eot_id|>"
                     || t.first == "<|im_end|>"
                     || t.first == "<|end|>"
                     || t.first == "<|return|>" // o200k_harmony
                     || t.first == "<|call|>"   // o200k_harmony
+                    || t.first == "<|flush|>"  // solar-open
+                    || t.first == "<|calls|>"  // solar-open
                     || t.first == "<end_of_turn>"
                     || t.first == "<|endoftext|>"
                     || t.first == "<|eom_id|>"
@@ -2356,24 +2405,28 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "<end_of_utterance>" // smoldocling
                ) {
                 special_eog_ids.insert(t.second);
-                if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
+                if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
                     LLAMA_LOG_WARN("%s: control-looking token: %6d '%s' was not control-type; this is probably a bug in the model. its type will be overridden\n",
                             __func__, t.second, t.first.c_str());
-                    id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_CONTROL;
+                    attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_CONTROL);
                 }
             } else {
-                // token is control, but not marked as EOG -> print a debug log
-                if (id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL && special_eog_ids.count(t.second) == 0) {
-                    LLAMA_LOG_DEBUG("%s: control token: %6d '%s' is not marked as EOG\n",
-                            __func__, t.second, t.first.c_str());
+                if (attr & LLAMA_TOKEN_ATTR_CONTROL && !(attr & LLAMA_TOKEN_ATTR_UNUSED)) {
+                    // token is control, but not marked as EOG -> print a debug log
+                    if (special_eog_ids.count(t.second) == 0) {
+                        LLAMA_LOG_DEBUG("%s: control token: %6d '%s' is not marked as EOG\n",
+                                __func__, t.second, t.first.c_str());
+                    }
                 }
             }
         }
 
         // @ngxson : quick hack for gpt-oss, always render these tokens
         for (const auto & t : token_to_id) {
+            auto & attr = id_to_token[t.second].attr;
+
             if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>" || t.first == "<|constrain|>") {
-                id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
+                attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_USER_DEFINED);
             }
         }
 
@@ -2393,34 +2446,42 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             LLAMA_LOG_WARN("%s: special_eom_id is not in special_eog_ids - the tokenizer config may be incorrect\n", __func__);
         }
 
-        // TODO: workaround for o200k_harmony tokenizer: the "<|end|>" token should not be EOG
-        //       we don't have a good way to detect this, so for now, if we have "<|return|>" and "<|call|>" tokens,
+        // TODO: workaround for o200k_harmony and solar-open tokenizer: the "<|end|>" token should not be EOG
+        //       we don't have a good way to detect this, so for now, if we have "<|return|>" and "<|call|>" tokens ("<|calls|>" and "<|flush|>" for solar-open),
         //       we remove the "<|end|>" token from the EOG list
         {
             bool has_return = false;
             bool has_call   = false;
             bool has_end    = false;
+            bool has_flush  = false;
 
             llama_token end_id = LLAMA_TOKEN_NULL;
 
             LLAMA_LOG_INFO("%s: printing all EOG tokens:\n", __func__);
             for (auto tid : special_eog_ids) {
-                LLAMA_LOG_INFO("%s:   - %d ('%s')\n", __func__, tid, id_to_token[tid].text.c_str());
+                auto & text = id_to_token[tid].text;
 
-                if (id_to_token[tid].text == "<|return|>") {
+                LLAMA_LOG_INFO("%s:   - %d ('%s')\n", __func__, tid, text.c_str());
+
+                if (text == "<|return|>") {
                     has_return = true;
-                } else if (id_to_token[tid].text == "<|call|>") {
+                } else if (text == "<|call|>" || text == "<|calls|>") {
                     has_call = true;
-                } else if (id_to_token[tid].text == "<|end|>") {
+                } else if (text == "<|flush|>") {
+                    has_flush = true;
+                } else if (text == "<|end|>") {
                     has_end = true;
                     end_id = tid;
                 }
             }
 
-            if (has_return && has_call && has_end) {
+            if ((has_return && has_call && has_end) || (has_call && has_flush && has_end)) {
                 special_eog_ids.erase(end_id);
-                id_to_token[end_id].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
-                LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
+
+                auto & attr = id_to_token[end_id].attr;
+                attr = (llama_token_attr) (attr | LLAMA_TOKEN_ATTR_USER_DEFINED);
+
+                LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>', or '<|calls|>' and '<|flush|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
             }
         }
     }
@@ -2518,6 +2579,13 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             for (const auto * token : {"<unk>", "<s>", "<|endoftext|>"}) {
                 _set_token_attr(token, LLAMA_TOKEN_ATTR_RSTRIP, false);
             }
+        } else if (_contains_any(model_name, {"modern-bert"})) {
+            if (token_to_id.count("[MASK]") == 0 ) {
+                LLAMA_LOG_WARN("%s: Mask token missing in vocab!\n", __func__);
+            }
+            else {
+                _set_token_attr("[MASK]", LLAMA_TOKEN_ATTR_LSTRIP, true);
+            }
         }
     }
 }

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

@@ -51,6 +51,8 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING = 40,
     LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2      = 41,
     LLAMA_VOCAB_PRE_TYPE_AFMOE           = 42,
+    LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN      = 43,
+    LLAMA_VOCAB_PRE_TYPE_YOUTU           = 44,
 };
 
 struct LLM_KV;

llama/llama.cpp/src/llama.cpp

@@ -71,8 +71,9 @@ static std::vector<llama_device_memory_data> llama_get_device_memory_data(
     }, &ud);
 
     llama_model_params mparams_copy = *mparams;
-    mparams_copy.no_alloc = true;
-    mparams_copy.use_mmap = false;
+    mparams_copy.no_alloc  = true;
+    mparams_copy.use_mmap  = false;
+    mparams_copy.use_mlock = false;
 
     llama_model * model = llama_model_load_from_file(path_model, mparams_copy);
     if (model == nullptr) {
@@ -110,8 +111,20 @@ static std::vector<llama_device_memory_data> llama_get_device_memory_data(
         }
     }
     for (size_t i = 0; i < ret.size(); i++) {
-        size_t free, total;
+        size_t free;
+        size_t total;
         ggml_backend_dev_memory(model->devices[i], &free, &total);
+
+        // devices can return 0 bytes for free and total memory if they do not
+        // have any to report. in this case, we will use the host memory as a fallback
+        // fixes: https://github.com/ggml-org/llama.cpp/issues/18577
+        if (free == 0 && total == 0) {
+            ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+            if (cpu_dev == nullptr) {
+                throw std::runtime_error(format("%s: no CPU backend found", __func__));
+            }
+            ggml_backend_dev_memory(cpu_dev, &free, &total);
+        }
         ret[i].free  = free;
         ret[i].total = total;
     }
@@ -139,12 +152,15 @@ enum layer_fraction_t {
 };
 // this enum is only used in llama_params_fit_impl but needs to be defined outside of it to fix a Windows compilation issue
 
+class llama_params_fit_exception : public std::runtime_error {
+    using std::runtime_error::runtime_error;
+};
+
 static void llama_params_fit_impl(
         const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
         float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
-        size_t margin_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
+        size_t * margins_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
     constexpr int64_t MiB = 1024*1024;
-    const int64_t margin = margin_s; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
     typedef std::vector<llama_device_memory_data> dmds_t;
     const llama_model_params default_mparams = llama_model_default_params();
 
@@ -163,6 +179,12 @@ static void llama_params_fit_impl(
         return;
     }
 
+    std::vector<int64_t> margins; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
+    margins.reserve(nd);
+    for (size_t id = 0; id < nd; id++) {
+        margins.push_back(margins_s[id]);
+    }
+
     std::vector<std::string> dev_names;
     {
         dev_names.reserve(nd);
@@ -180,11 +202,12 @@ static void llama_params_fit_impl(
         }
     }
 
-    int64_t sum_total          = 0;
-    int64_t sum_projected_free = 0;
-    int64_t min_projected_free = INT64_MAX;
-    int64_t sum_projected_used = 0;
-    int64_t sum_projected_ctx  = 0;
+    int64_t sum_free            = 0;
+    int64_t sum_projected_free  = 0;
+    int64_t sum_projected_used  = 0;
+    int64_t sum_projected_model = 0;
+    std::vector<int64_t> projected_free_per_device;
+    projected_free_per_device.reserve(nd);
 
     if (nd > 1) {
         LLAMA_LOG_INFO("%s: projected memory use with initial parameters [MiB]:\n", __func__);
@@ -194,59 +217,98 @@ static void llama_params_fit_impl(
 
         const int64_t projected_used = dmd.mb.total();
         const int64_t projected_free = dmd.free - projected_used;
+        projected_free_per_device.push_back(projected_free);
 
-        sum_total          += dmd.total;
-        sum_projected_used += projected_used;
-        sum_projected_free += projected_free;
-        min_projected_free  = std::min(min_projected_free, projected_free);
-        sum_projected_ctx  += dmd.mb.context;
+        sum_free            += dmd.free;
+        sum_projected_used  += projected_used;
+        sum_projected_free  += projected_free;
+        sum_projected_model += dmd.mb.model;
 
         if (nd > 1) {
-            LLAMA_LOG_INFO("%s:   - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " %s\n",
-                __func__, dev_names[id].c_str(), dmd.total/MiB, projected_used/MiB, std::abs(projected_free)/MiB,
-                projected_free >= 0 ? "surplus" : "deficit");
+            LLAMA_LOG_INFO("%s:   - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " free vs. target of %6" PRId64 "\n",
+                __func__, dev_names[id].c_str(), dmd.total/MiB, projected_used/MiB, projected_free/MiB, margins[id]/MiB);
         }
     }
-    assert(sum_total >= 0 && sum_projected_used >= 0 && sum_projected_ctx >= 0);
-    assert(sum_projected_used >= sum_projected_ctx);
+    assert(sum_free >= 0 && sum_projected_used >= 0);
     LLAMA_LOG_INFO("%s: projected to use %" PRId64 " MiB of device memory vs. %" PRId64 " MiB of free device memory\n",
-        __func__, sum_projected_used/MiB, sum_total/MiB);
-    if (min_projected_free >= margin) {
-        if (nd == 1) {
+        __func__, sum_projected_used/MiB, sum_free/MiB);
+    if (nd == 1) {
+        if (projected_free_per_device[0] >= margins[0]) {
             LLAMA_LOG_INFO("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n",
-                __func__, min_projected_free/MiB, margin/MiB);
+                __func__, projected_free_per_device[0]/MiB, margins[0]/MiB);
+            return;
+        }
+    } else {
+        bool changes_needed = false;
+        for (size_t id = 0; id < nd; id++) {
+            if (projected_free_per_device[id] < margins[id]) {
+                changes_needed = true;
+                break;
+            }
+        }
+        if (!changes_needed) {
+            LLAMA_LOG_INFO("%s: targets for free memory can be met on all devices, no changes needed\n", __func__);
             return;
         }
-        LLAMA_LOG_INFO("%s: will leave at least %" PRId64 " >= %" PRId64 " MiB of free memory on all devices, no changes needed\n",
-            __func__, min_projected_free/MiB, margin/MiB);
-        return;
     }
 
     // step 2: try reducing memory use by reducing the context size
 
     {
-        int64_t global_surplus = sum_projected_free - int64_t(nd)*margin;
+        int64_t global_surplus = sum_projected_free;
+        for (size_t id = 0; id < nd; id++) {
+            global_surplus -= margins[id];
+        }
         if (global_surplus < 0) {
-            LLAMA_LOG_INFO(nd == 1 ?
-                "%s: cannot fulfill margin of %" PRId64 " MiB, need to reduce device memory by %" PRId64 " MiB\n" :
-                "%s: cannot fulfill margin of %" PRId64 " MiB on all devices, need to use %" PRId64 " MiB less in total\n",
-                __func__, margin/MiB, -global_surplus/MiB);
+            if (nd == 1) {
+                LLAMA_LOG_INFO("%s: cannot meet free memory target of %" PRId64 " MiB, need to reduce device memory by %" PRId64 " MiB\n",
+                    __func__, margins[0]/MiB, -global_surplus/MiB);
+            } else {
+                LLAMA_LOG_INFO(
+                    "%s: cannot meet free memory targets on all devices, need to use %" PRId64 " MiB less in total\n",
+                    __func__, -global_surplus/MiB);
+            }
             if (cparams->n_ctx == 0) {
                 if (hp_nct > n_ctx_min) {
-                    const int64_t bytes_per_ctx = sum_projected_ctx / hp_nct;
-                    const uint32_t ctx_reduction = std::min(
-                        uint32_t((-global_surplus + bytes_per_ctx - 1) / bytes_per_ctx), hp_nct - n_ctx_min);
-                    cparams->n_ctx = hp_nct - ctx_reduction;
-                    const int64_t memory_reduction = ctx_reduction * bytes_per_ctx;
-                    global_surplus += memory_reduction;
-                    LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                        __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
-                    if (global_surplus >= 0) {
+                    int64_t sum_used_target = sum_free;
+                    for (size_t id = 0; id < nd; id++) {
+                        sum_used_target -= margins[id];
+                    }
+                    if (nd > 1) {
+                        // for multiple devices we need to be more conservative in terms of how much context we think can fit:
+                        //   - for dense models only whole layers can be assigned to devices
+                        //   - for MoE models only whole tensors can be assigned to devices, which we estimate to be <= 1/3 of a layer
+                        //   - on average we expect a waste of 0.5 layers/tensors per device
+                        //   - use slightly more than the expected average for nd devices to be safe
+                        const int64_t model_per_layer = sum_projected_model / std::min(uint32_t(mparams->n_gpu_layers), hp_ngl);
+                        sum_used_target -= (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6);
+                    }
+
+                    int64_t sum_projected_used_min_ctx = 0;
+                    cparams->n_ctx = n_ctx_min;
+                    const dmds_t dmds_min_ctx = llama_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+                    for (const auto & dmd : dmds_min_ctx) {
+                        sum_projected_used_min_ctx += dmd.mb.total();
+                    }
+                    if (sum_used_target > sum_projected_used_min_ctx) {
+                        // linear interpolation between minimum and maximum context size:
+                        cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx)
+                            / (sum_projected_used - sum_projected_used_min_ctx);
+                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
+
+                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (hp_nct - n_ctx_min);
+                        const int64_t memory_reduction = (hp_nct - cparams->n_ctx) * bytes_per_ctx;
+                        LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                         if (nd == 1) {
                             LLAMA_LOG_INFO("%s: entire model can be fit by reducing context\n", __func__);
                             return;
                         }
                         LLAMA_LOG_INFO("%s: entire model should be fit across devices by reducing context\n", __func__);
+                    } else {
+                        const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx;
+                        LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                     }
                 } else {
                     LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
@@ -259,32 +321,28 @@ static void llama_params_fit_impl(
     }
 
     if (mparams->n_gpu_layers != default_mparams.n_gpu_layers) {
-        throw std::runtime_error("n_gpu_layers already set by user to " + std::to_string(mparams->n_gpu_layers) + ", abort");
+        throw llama_params_fit_exception("n_gpu_layers already set by user to " + std::to_string(mparams->n_gpu_layers) + ", abort");
     }
     if (nd > 1) {
         if (!tensor_split) {
-            throw std::runtime_error("did not provide a buffer to write the tensor_split to, abort");
+            throw llama_params_fit_exception("did not provide a buffer to write the tensor_split to, abort");
         }
         if (mparams->tensor_split) {
             for (size_t id = 0; id < nd; id++) {
                 if (mparams->tensor_split[id] != 0.0f) {
-                    throw std::runtime_error("model_params::tensor_split already set by user, abort");
+                    throw llama_params_fit_exception("model_params::tensor_split already set by user, abort");
                 }
             }
         }
         if (mparams->split_mode == LLAMA_SPLIT_MODE_ROW) {
-            throw std::runtime_error("changing weight allocation for LLAMA_SPLIT_MODE_ROW not implemented, abort");
-        }
-        if (hp_ngl < 2*nd) {
-            throw std::runtime_error("model has only " + std::to_string(hp_ngl) + " layers but need at least "
-                + std::to_string(2*nd) + " to fit memory for " + std::to_string(nd) + " devices, abort");
+            throw llama_params_fit_exception("changing weight allocation for LLAMA_SPLIT_MODE_ROW not implemented, abort");
         }
     }
     if (!tensor_buft_overrides) {
-        throw std::runtime_error("did not provide buffer to set tensor_buft_overrides, abort");
+        throw llama_params_fit_exception("did not provide buffer to set tensor_buft_overrides, abort");
     }
     if (mparams->tensor_buft_overrides && (mparams->tensor_buft_overrides->pattern || mparams->tensor_buft_overrides->buft)) {
-        throw std::runtime_error("model_params::tensor_buft_overrides already set by user, abort");
+        throw llama_params_fit_exception("model_params::tensor_buft_overrides already set by user, abort");
     }
 
     // step 3: iteratively fill the back to front with "dense" layers
@@ -337,6 +395,11 @@ static void llama_params_fit_impl(
 
         // for the first partial layer varying parts can overflow, all further layers use LAYER_FRACTION_MOE:
         layer_fraction_t overflow_type = LAYER_FRACTION_MOE;
+
+        uint32_t n_full() const {
+            assert(n_layer >= n_part);
+            return n_layer - n_part;
+        }
     };
 
     const size_t ntbo = llama_max_tensor_buft_overrides();
@@ -345,8 +408,7 @@ static void llama_params_fit_impl(
     auto set_ngl_tensor_split_tbo = [&](
             const std::vector<ngl_t> & ngl_per_device,
             const std::vector<ggml_backend_buffer_type_t> & overflow_bufts,
-            llama_model_params & mparams,
-            const bool add_nonrepeating) {
+            llama_model_params & mparams) {
         mparams.n_gpu_layers = 0;
         for (size_t id = 0; id < nd; id++) {
             mparams.n_gpu_layers += ngl_per_device[id].n_layer;
@@ -354,29 +416,25 @@ static void llama_params_fit_impl(
                 tensor_split[id] = ngl_per_device[id].n_layer;
             }
         }
-        assert(uint32_t(mparams.n_gpu_layers) <= hp_ngl);
-        uint32_t il0 = hp_ngl - mparams.n_gpu_layers; // start index for tensor buft overrides
+        assert(uint32_t(mparams.n_gpu_layers) <= hp_ngl + 1);
+        uint32_t il0 = hp_ngl + 1 - mparams.n_gpu_layers; // start index for tensor buft overrides
 
-        if (add_nonrepeating) {
-            mparams.n_gpu_layers += 1;
-            tensor_split[nd - 1] += 1;
-        }
         mparams.tensor_split = tensor_split;
 
         size_t itbo = 0;
         for (size_t id = 0; id < nd; id++) {
-            il0 += ngl_per_device[id].n_layer - ngl_per_device[id].n_part;
+            il0 += ngl_per_device[id].n_full();
             for (uint32_t il = il0; il < il0 + ngl_per_device[id].n_part; il++) {
                 if (itbo + 1 >= ntbo) {
                     tensor_buft_overrides[itbo].pattern = nullptr;
                     tensor_buft_overrides[itbo].buft    = nullptr;
                     itbo++;
                     mparams.tensor_buft_overrides = tensor_buft_overrides;
-                    throw std::runtime_error("llama_params_fit_n_tensor_buft_overrides() == "
-                        + std::to_string(ntbo) + " is insufficient for model\n");
+                    throw llama_params_fit_exception("llama_max_tensor_buft_overrides() == "
+                        + std::to_string(ntbo) + " is insufficient for model");
                 }
                 tensor_buft_overrides[itbo].pattern = get_overflow_pattern(il, il == il0 ? ngl_per_device[id].overflow_type : LAYER_FRACTION_MOE);
-                tensor_buft_overrides[itbo].buft = overflow_bufts[id];
+                tensor_buft_overrides[itbo].buft = il == il0 ? overflow_bufts[id] : ggml_backend_cpu_buffer_type();
                 itbo++;
             }
             il0 += ngl_per_device[id].n_part;
@@ -391,10 +449,9 @@ static void llama_params_fit_impl(
     auto get_memory_for_layers = [&](
             const char * func_name,
             const std::vector<ngl_t> & ngl_per_device,
-            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts,
-            const bool add_nonrepeating) -> std::vector<int64_t> {
+            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts) -> std::vector<int64_t> {
         llama_model_params mparams_copy = *mparams;
-        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy, add_nonrepeating);
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy);
 
         const dmds_t dmd_nl = llama_get_device_memory_data(
             path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
@@ -427,9 +484,9 @@ static void llama_params_fit_impl(
         const dmds_t dmds_cpu_moe = llama_get_device_memory_data(
             path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
 
-        for (const llama_device_memory_data & dmd : dmds_cpu_moe) {
-            global_surplus_cpu_moe += dmd.free;
-            global_surplus_cpu_moe -= int64_t(dmd.mb.total()) + margin;
+        for (size_t id = 0; id < nd; id++) {
+            global_surplus_cpu_moe += dmds_cpu_moe[id].free;
+            global_surplus_cpu_moe -= int64_t(dmds_cpu_moe[id].mb.total()) + margins[id];
         }
 
         if (global_surplus_cpu_moe > 0) {
@@ -448,27 +505,18 @@ static void llama_params_fit_impl(
     std::vector<int64_t> targets; // maximum acceptable memory use per device
     targets.reserve(nd);
     for (size_t id = 0; id < nd; id++) {
-        targets.push_back(dmds_full[id].free - margin);
+        targets.push_back(dmds_full[id].free - margins[id]);
         LLAMA_LOG_DEBUG("%s: id=%zu, target=%" PRId64 " MiB\n", __func__, id, targets[id]/MiB);
     }
 
-    // whether for the optimal memory use we expect to load at least some MoE tensors:
-    const bool partial_moe = hp_nex > 0 && global_surplus_cpu_moe > 0;
-
-    std::vector<ggml_backend_buffer_type_t> overflow_bufts; // which bufts the partial layers of a device overflow to:
+    std::vector<ggml_backend_buffer_type_t> overflow_bufts; // which bufts the first partial layer of a device overflows to:
     overflow_bufts.reserve(nd);
-    for (size_t id = 0; id < nd - 1; ++id) {
-        overflow_bufts.push_back(ggml_backend_dev_buffer_type(devs[id + 1]));
+    for (size_t id = 0; id < nd; id++) {
+        overflow_bufts.push_back(ggml_backend_cpu_buffer_type());
     }
-    overflow_bufts.push_back(ggml_backend_cpu_buffer_type());
 
     std::vector<ngl_t> ngl_per_device(nd);
-    std::vector<int64_t> mem = get_memory_for_layers(__func__, ngl_per_device, overflow_bufts, partial_moe);
-    if (hp_nex > 0) {
-        for (size_t id = 0; id < nd; id++) {
-            ngl_per_device[id].overflow_type = LAYER_FRACTION_MOE;
-        }
-    }
+    std::vector<int64_t> mem = get_memory_for_layers(__func__, ngl_per_device, overflow_bufts);
 
     // optimize the number of layers per device using the method of false position:
     //   - ngl_per_device has 0 layers for each device, lower bound
@@ -476,22 +524,30 @@ static void llama_params_fit_impl(
     //   - interpolate the memory use / layer between low and high linearly to get a guess where it meets our target
     //   - check memory use of our guess, replace either the low or high bound
     //   - once we only have a difference of a single layer, stop and return the lower bound that just barely still fits
+    //   - the last device has the output layer, which cannot be a partial layer
     if (hp_nex == 0) {
         LLAMA_LOG_INFO("%s: filling dense layers back-to-front:\n", __func__);
     } else {
         LLAMA_LOG_INFO("%s: filling dense-only layers back-to-front:\n", __func__);
     }
-    uint32_t n_unassigned = hp_ngl;
     for (int id = nd - 1; id >= 0; id--) {
+        uint32_t n_unassigned = hp_ngl + 1;
+        for (size_t jd = id + 1; jd < nd; ++jd) {
+            assert(n_unassigned >= ngl_per_device[jd].n_layer);
+            n_unassigned -= ngl_per_device[jd].n_layer;
+        }
+
         std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
         ngl_per_device_high[id].n_layer = n_unassigned;
         if (hp_nex > 0) {
-            ngl_per_device_high[id].n_part = ngl_per_device_high[id].n_layer;
+            ngl_per_device_high[id].n_part = size_t(id) < nd - 1 ? ngl_per_device_high[id].n_layer : ngl_per_device_high[id].n_layer - 1;
         }
         if (ngl_per_device_high[id].n_layer > 0) {
-            std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts, partial_moe);
+            std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts);
             if (mem_high[id] > targets[id]) {
+                assert(ngl_per_device_high[id].n_layer > ngl_per_device[id].n_layer);
                 uint32_t delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                LLAMA_LOG_DEBUG("%s: start filling device %" PRIu32 ", delta=%" PRIu32 "\n", __func__, id, delta);
                 while (delta > 1) {
                     uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
                     step_size = std::max(step_size, uint32_t(1));
@@ -500,25 +556,26 @@ static void llama_params_fit_impl(
                     std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
                     ngl_per_device_test[id].n_layer += step_size;
                     if (hp_nex) {
-                        ngl_per_device_test[id].n_part += step_size;
+                        ngl_per_device_test[id].n_part += size_t(id) == nd - 1 && ngl_per_device_test[id].n_part == 0 ?
+                            step_size - 1 : step_size; // the first layer is the output layer which must always be full
                     }
-                    const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+                    const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
 
                     if (mem_test[id] <= targets[id]) {
-                        ngl_per_device  = ngl_per_device_test;
-                        mem             = mem_test;
-                        n_unassigned   -= ngl_per_device[id].n_layer;
+                        ngl_per_device = ngl_per_device_test;
+                        mem            = mem_test;
                         LLAMA_LOG_DEBUG("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
                     } else {
                         ngl_per_device_high = ngl_per_device_test;
                         mem_high            = mem_test;
-                        LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+                        LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device_high[id].n_layer);
                     }
                     delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
                 }
             } else {
-                ngl_per_device  = ngl_per_device_high;
-                n_unassigned   -= ngl_per_device[id].n_layer;
+                assert(ngl_per_device_high[id].n_layer == n_unassigned);
+                ngl_per_device = ngl_per_device_high;
+                mem            = mem_high;
                 LLAMA_LOG_DEBUG("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
             }
         }
@@ -529,7 +586,7 @@ static void llama_params_fit_impl(
             __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, mem[id]/MiB, projected_margin/MiB);
     }
     if (hp_nex == 0 || global_surplus_cpu_moe <= 0) {
-        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams, partial_moe);
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams);
         return;
     }
 
@@ -549,24 +606,20 @@ static void llama_params_fit_impl(
     assert(id_dense_start < nd);
 
     LLAMA_LOG_INFO("%s: converting dense-only layers to full layers and filling them front-to-back with overflow to next device/system memory:\n", __func__);
-    for (size_t id = 0; id <= id_dense_start; id++) {
+    for (size_t id = 0; id <= id_dense_start && id_dense_start < nd; id++) {
         std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
         for (size_t jd = id_dense_start; jd < nd; jd++) {
-            const uint32_t n_layer_move = ngl_per_device_high[jd].n_layer;
+            const uint32_t n_layer_move = jd < nd - 1 ? ngl_per_device_high[jd].n_layer : ngl_per_device_high[jd].n_layer - 1;
             ngl_per_device_high[id].n_layer += n_layer_move;
             ngl_per_device_high[jd].n_layer -= n_layer_move;
             ngl_per_device_high[jd].n_part = 0;
         }
         size_t id_dense_start_high = nd - 1;
-        std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts, partial_moe);
+        std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts);
 
         if (mem_high[id] > targets[id]) {
-            assert(ngl_per_device_high[id].n_layer >= ngl_per_device_high[id].n_part);
-            assert(ngl_per_device[id].n_layer >= ngl_per_device[id].n_part);
-            assert((ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
-                   >= ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
-            uint32_t delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
-                - (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+            assert(ngl_per_device_high[id].n_full() >= ngl_per_device[id].n_full());
+            uint32_t delta = ngl_per_device_high[id].n_full() - ngl_per_device[id].n_full();
             while (delta > 1) {
                 uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
                 step_size = std::max(step_size, uint32_t(1));
@@ -582,11 +635,11 @@ static void llama_params_fit_impl(
                     ngl_per_device_test[id].n_layer += n_convert_jd;
                     n_converted_test += n_convert_jd;
 
-                    if (ngl_per_device_test[id_dense_start_test].n_layer > 0) {
+                    if (ngl_per_device_test[id_dense_start_test].n_part > 0) {
                         break;
                     }
                 }
-                const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+                const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
 
                 if (mem_test[id] <= targets[id]) {
                     ngl_per_device = ngl_per_device_test;
@@ -601,32 +654,38 @@ static void llama_params_fit_impl(
                     LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start_high=%zu\n",
                         __func__, id, ngl_per_device_high[id].n_layer, ngl_per_device_high[id].n_part, id_dense_start_high);
                 }
-                delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
-                    - (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+                assert(ngl_per_device_high[id].n_full() >= ngl_per_device[id].n_full());
+                delta = ngl_per_device_high[id].n_full() - ngl_per_device[id].n_full();
             }
         } else {
             ngl_per_device = ngl_per_device_high;
+            mem            = mem_high;
             id_dense_start = id_dense_start_high;
             LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start=%zu\n",
                 __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
         }
 
         // try to fit at least part of one more layer
-        if (ngl_per_device[id_dense_start].n_layer > 0) {
+        if (ngl_per_device[id_dense_start].n_layer > (id < nd - 1 ? 0 : 1)) {
             std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
             size_t id_dense_start_test = id_dense_start;
             ngl_per_device_test[id_dense_start_test].n_layer--;
             ngl_per_device_test[id_dense_start_test].n_part--;
             ngl_per_device_test[id].n_layer++;
             ngl_per_device_test[id].n_part++;
-            if (ngl_per_device_test[id_dense_start_test].n_layer == 0) {
+            if (ngl_per_device_test[id_dense_start_test].n_part == 0) {
                 id_dense_start_test++;
             }
             ngl_per_device_test[id].overflow_type = LAYER_FRACTION_UP;
+            std::vector<ggml_backend_buffer_type_t> overflow_bufts_test = overflow_bufts;
+            if (id < nd - 1) {
+                overflow_bufts_test[id] = ggml_backend_dev_buffer_type(devs[id + 1]);
+            }
             LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_UP\n", __func__);
-            std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
-            if (mem_test[id] < targets[id]) {
+            std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+            if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
                 ngl_per_device = ngl_per_device_test;
+                overflow_bufts = overflow_bufts_test;
                 mem            = mem_test;
                 id_dense_start = id_dense_start_test;
                 LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", UP), id_dense_start=%zu\n",
@@ -634,9 +693,10 @@ static void llama_params_fit_impl(
 
                 ngl_per_device_test[id].overflow_type = LAYER_FRACTION_GATE;
                 LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_GATE\n", __func__);
-                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
-                if (mem_test[id] < targets[id]) {
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+                if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
                     ngl_per_device = ngl_per_device_test;
+                    overflow_bufts = overflow_bufts_test;
                     mem            = mem_test;
                     id_dense_start = id_dense_start_test;
                     LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", GATE), id_dense_start=%zu\n",
@@ -645,9 +705,10 @@ static void llama_params_fit_impl(
             } else {
                 ngl_per_device_test[id].overflow_type = LAYER_FRACTION_ATTN;
                 LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_ATTN\n", __func__);
-                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
-                if (mem_test[id] < targets[id]) {
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+                if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
                     ngl_per_device = ngl_per_device_test;
+                    overflow_bufts = overflow_bufts_test;
                     mem            = mem_test;
                     id_dense_start = id_dense_start_test;
                     LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", ATTN), id_dense_start=%zu\n",
@@ -662,30 +723,41 @@ static void llama_params_fit_impl(
             __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
     }
 
-    set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams, partial_moe);
+    // print info for devices that were not changed during the conversion from dense only to full layers:
+    for (size_t id = id_dense_start + 1; id < nd; id++) {
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LLAMA_LOG_INFO(
+            "%s:   - %s: %2" PRIu32 " layers (%2" PRIu32 " overflowing), %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
+    }
+
+    set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams);
 }
 
-bool llama_params_fit(
+enum llama_params_fit_status llama_params_fit(
         const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
         float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
-        size_t margin_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
+        size_t * margins, uint32_t n_ctx_min, enum ggml_log_level log_level) {
     const int64_t t0_us = llama_time_us();
-    bool ok = true;
+    llama_params_fit_status status = LLAMA_PARAMS_FIT_STATUS_SUCCESS;
     try {
-        llama_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margin_s, n_ctx_min, log_level);
+        llama_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, log_level);
         LLAMA_LOG_INFO("%s: successfully fit params to free device memory\n", __func__);
-    } catch (const std::runtime_error & e) {
+    } catch (const llama_params_fit_exception & e) {
         LLAMA_LOG_WARN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());
-        ok = false;
+        status = LLAMA_PARAMS_FIT_STATUS_FAILURE;
+    } catch (const std::runtime_error & e) {
+        LLAMA_LOG_ERROR("%s: encountered an error while trying to fit params to free device memory: %s\n", __func__, e.what());
+        status = LLAMA_PARAMS_FIT_STATUS_ERROR;
     }
     const int64_t t1_us = llama_time_us();
     LLAMA_LOG_INFO("%s: fitting params to free memory took %.2f seconds\n", __func__, (t1_us - t0_us) * 1e-6);
-    return ok;
+    return status;
 }
 
 struct llama_sampler_chain_params llama_sampler_chain_default_params() {
     struct llama_sampler_chain_params result = {
-        /*.no_perf                     =*/ true,
+        /*.no_perf =*/ true,
     };
 
     return result;
@@ -758,7 +830,7 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
     model.t_start_us = tm.t_start_us;
 
     try {
-        llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
+        llama_model_loader ml(fname, splits, params.use_mmap, params.use_direct_io, params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
 
         ml.print_info();
 

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

@@ -22,8 +22,15 @@ llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_para
     const float kq_scale = 1.0f/sqrtf(float(n_embd_head));
 
     for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
         ggml_tensor * inpSA = inpL;
 
+        // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
+        const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
+                              (il + 1) % hparams.n_no_rope_layer_step != 0;
+
         // dual attention normalization (pre)
         cur = build_norm(inpL,
                 model.layers[il].attn_norm, NULL,
@@ -56,19 +63,16 @@ llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_para
             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,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                         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,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Kcur, "Kcur_rope", il);
             }

llama/llama.cpp/src/models/bert.cpp

@@ -142,11 +142,13 @@ llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params
                     LLM_FFN_GELU, LLM_FFN_SEQ, il);
             cb(cur, "ffn_out", il);
         } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
+            const bool up_contains_gate = !model.layers[il].ffn_gate && model.layers[il].ffn_up->ne[1] != hparams.n_ff();
+            auto type_op = up_contains_gate ? LLM_FFN_GEGLU : LLM_FFN_GELU;
             cur = build_ffn(cur,
-                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
                     model.layers[il].ffn_gate, NULL, NULL,
                     model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
-                    model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il);
+                    type_op, LLM_FFN_PAR, il);
             cb(cur, "ffn_out", il);
         } else {
             cur = build_ffn(cur,

llama/llama.cpp/src/models/cogvlm.cpp

@@ -3,12 +3,14 @@
 llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
-    float         kq_scale    = 1.0f / sqrtf(float(n_embd_head));
+    const float   kq_scale    = 1.0f / sqrtf(float(n_embd_head));
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
     GGML_ASSERT(n_embd_head == hparams.n_rot);
 
-    ggml_tensor *inpL, *cur;
+    ggml_tensor * inpL;
+    ggml_tensor * cur;
+
     inpL = build_inp_embd(model.tok_embd);
 
     ggml_tensor * inp_pos = build_inp_pos();
@@ -44,7 +46,7 @@ llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_pa
         }
 
         ggml_tensor * inpSA = inpL;
-        cur                 = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cur = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
 
         // build self attention
         {

llama/llama.cpp/src/models/cohere2-iswa.cpp

@@ -21,6 +21,9 @@ llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const
 
     for (int il = 0; il < n_layer; ++il) {
         const bool is_swa = hparams.is_swa(il);
+        // UNUSED:
+        // const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+        // const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
 
         // norm
         cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);

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

@@ -215,7 +215,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 model.layers[il].ffn_exp_probs_b,
                 n_expert, n_expert_used,
                 LLM_FFN_SILU, hparams.expert_weights_norm,
-                true, hparams.expert_weights_scale,
+                hparams.expert_weights_scale, hparams.expert_weights_scale,
                 (llama_expert_gating_func_type) hparams.expert_gating_func,
                 il);
             cb(moe_out, "ffn_moe_out", il);

llama/llama.cpp/src/models/gemma-embedding.cpp

@@ -1,7 +1,5 @@
 #include "models.h"
 
-
-
 llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_k;
@@ -12,10 +10,8 @@ llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model,
     inpL = build_inp_embd(model.tok_embd);
 
     // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
-    if (ubatch.token) {
-        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-        cb(inpL, "inp_scaled", -1);
-    }
+    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
+    cb(inpL, "inp_scaled", -1);
 
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();

llama/llama.cpp/src/models/gemma2-iswa.cpp

@@ -19,6 +19,9 @@ llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const ll
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
         // norm
         cur = build_norm(inpL,
                 model.layers[il].attn_norm, NULL,
@@ -43,12 +46,12 @@ llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const ll
 
             Qcur = ggml_rope_ext(
                     ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                     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,
+                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                     ext_factor, attn_factor, beta_fast, beta_slow);
 
             cb(Qcur, "Qcur", il);

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

@@ -10,10 +10,9 @@ llm_build_gemma3<iswa>::llm_build_gemma3(const llama_model & model, const llm_gr
     inpL = build_inp_embd(model.tok_embd);
 
     // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
-    if (ubatch.token) {
-        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-        cb(inpL, "inp_scaled", -1);
-    }
+    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
+    cb(inpL, "inp_scaled", -1);
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 

llama/llama.cpp/src/models/gemma3n-iswa.cpp

@@ -1,7 +1,5 @@
 #include "models.h"
 
-
-
 llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params),
     model(model),
@@ -15,10 +13,9 @@ llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const
     inpL = build_inp_embd(model.tok_embd);
 
     // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
-    if (ubatch.token) {
-        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-        cb(inpL, "inp_scaled", -1);
-    }
+    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
+    cb(inpL, "inp_scaled", -1);
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -248,7 +245,7 @@ ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) {
 // equivalent to get_per_layer_inputs() in python code
 // output shape: [n_embd_altup, n_layer, n_tokens]
 ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
-    auto          inp = std::make_unique<llm_graph_input_embd>();
+    auto inp = std::make_unique<llm_graph_input_embd>();
     ggml_tensor * inp_per_layer;
     if (ubatch.token) {
         inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
@@ -258,10 +255,20 @@ ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
         inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
         inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_altup));
         cb(inp_per_layer, "inp_per_layer_selected", -1);
+        res->add_input(std::move(inp));
     } else {
-        GGML_ABORT("TODO: support embd input");
+        // Vision embedding path: use padding token (ID=0) embedding
+        const int64_t embd_size = model.tok_embd_per_layer->ne[0];  // n_embd_altup * n_layer
+
+        // Extract and dequantize padding token embedding (column 0)
+        ggml_tensor * padding_q = ggml_view_1d(ctx0, model.tok_embd_per_layer, embd_size, 0);
+        ggml_tensor * padding_f32 = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, embd_size);
+        inp_per_layer = ggml_cpy(ctx0, padding_q, padding_f32);
+
+        // Reshape to [n_embd_altup, n_layer, 1]
+        inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, 1);
+        cb(inp_per_layer, "inp_per_layer_vision", -1);
     }
-    res->add_input(std::move(inp));
     return inp_per_layer;
 }
 
@@ -279,7 +286,7 @@ ggml_tensor * llm_build_gemma3n_iswa::project_per_layer_inputs(ggml_tensor * inp
                                               -1);  // [n_embd_altup, n_layer, n_tokens]
     cb(per_layer_proj, "per_layer_proj", -1);
 
-    inp_per_layer = ggml_add(ctx0, inp_per_layer, per_layer_proj);
+    inp_per_layer = ggml_add(ctx0, per_layer_proj, inp_per_layer);
     inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);
     cb(inp_per_layer, "inp_per_layer", -1);
 

llama/llama.cpp/src/models/llama-iswa.cpp

@@ -25,8 +25,12 @@ llm_build_llama_iswa::llm_build_llama_iswa(const llama_model & model, const llm_
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
         ggml_tensor * inpSA = inpL;
 
+        // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
         const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
                               (il + 1) % hparams.n_no_rope_layer_step != 0;
 
@@ -67,13 +71,13 @@ llm_build_llama_iswa::llm_build_llama_iswa(const llama_model & model, const llm_
             if (use_rope) {
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                         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,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                         ext_factor, attn_factor, beta_fast, beta_slow
                         );
             } else if (inp_attn_scale) {

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

@@ -1,6 +1,7 @@
 #include "models.h"
 
-llm_build_llama::llm_build_llama(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+template <bool embed>
+llm_build_llama<embed>::llm_build_llama(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);
@@ -14,7 +15,14 @@ llm_build_llama::llm_build_llama(const llama_model & model, const llm_graph_para
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
-    auto * inp_attn = build_attn_inp_kv();
+    using inp_attn_type = std::conditional_t<embed, llm_graph_input_attn_no_cache, llm_graph_input_attn_kv>;
+
+    inp_attn_type * inp_attn = nullptr;
+    if constexpr (embed) {
+        inp_attn = build_attn_inp_no_cache();
+    } else {
+        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;
 
@@ -145,11 +153,16 @@ llm_build_llama::llm_build_llama(const llama_model & model, const llm_graph_para
     cb(cur, "result_norm", -1);
     res->t_embd = cur;
 
-    // lm_head
-    cur = build_lora_mm(model.output, cur);
+    if constexpr (!embed) {
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
 
-    cb(cur, "result_output", -1);
-    res->t_logits = cur;
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+    }
 
     ggml_build_forward_expand(gf, cur);
 }
+
+template struct llm_build_llama<false>;
+template struct llm_build_llama<true>;

llama/llama.cpp/src/models/maincoder.cpp

@@ -0,0 +1,117 @@
+#include "models.h"
+
+llm_build_maincoder::llm_build_maincoder(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
+                    );
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            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_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/mimo2-iswa.cpp

@@ -0,0 +1,123 @@
+
+#include "models.h"
+
+llm_build_mimo2_iswa::llm_build_mimo2_iswa(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);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+    auto * inp_attn = build_attn_inp_kv_iswa();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        uint32_t n_head_l    = hparams.n_head(il);
+        uint32_t n_head_kv_l = hparams.n_head_kv(il);
+        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        cur = inpL;
+
+        // self_attention
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", 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_k, n_head_l,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                ctx0, Qcur, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                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_l, freq_scale_l,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            ggml_tensor * sinks = model.layers[il].attn_sinks;
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, sinks, nullptr, 1.0f/sqrtf(float(n_embd_head_k)), 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);
+
+        // feed-forward network
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+            // dense branch
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_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, true, false,
+                                0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID, il);
+            cb(cur, "ffn_moe_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/models.h

@@ -303,6 +303,7 @@ struct llm_build_llada_moe : public llm_graph_context {
     llm_build_llada_moe(const llama_model & model, const llm_graph_params & params);
 };
 
+template <bool embed>
 struct llm_build_llama : public llm_graph_context {
     llm_build_llama(const llama_model & model, const llm_graph_params & params);
 };
@@ -311,10 +312,18 @@ struct llm_build_llama_iswa : public llm_graph_context {
     llm_build_llama_iswa(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_maincoder : public llm_graph_context {
+    llm_build_maincoder(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_mamba : public llm_graph_context_mamba {
     llm_build_mamba(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_mimo2_iswa : public llm_graph_context {
+    llm_build_mimo2_iswa(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_minicpm3 : public llm_graph_context {
     llm_build_minicpm3(const llama_model & model, const llm_graph_params & params);
 };
@@ -327,6 +336,10 @@ struct llm_build_mistral3 : public llm_graph_context {
     llm_build_mistral3(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_modern_bert : public llm_graph_context {
+    llm_build_modern_bert(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_mpt : public llm_graph_context {
     llm_build_mpt(const llama_model & model, const llm_graph_params & params);
 };
@@ -396,6 +409,11 @@ struct llm_build_plamo : public llm_graph_context {
     llm_build_plamo(const llama_model & model, const llm_graph_params & params);
 };
 
+template <bool iswa>
+struct llm_build_plamo3 : public llm_graph_context {
+    llm_build_plamo3(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_plm : public llm_graph_context {
     llm_build_plm(const llama_model & model, const llm_graph_params & params);
 };

llama/llama.cpp/src/models/modern-bert.cpp

@@ -0,0 +1,116 @@
+#include "models.h"
+
+llm_build_modern_bert::llm_build_modern_bert(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;
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // construct input embeddings (token, type, position)
+    inpL = build_inp_embd(model.tok_embd);
+    cb(inpL, "inp_embd", -1);
+
+    // embed layer norm
+    inpL = build_norm(inpL, model.tok_norm, nullptr, LLM_NORM, -1);
+    cb(inpL, "inp_norm", -1);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        cur = inpL;
+
+        // attention layer norm
+        if (model.layers[il].attn_norm) {
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM, il);
+            cb(cur, "attn_norm", il);
+        }
+
+        // self attention
+        cur = build_lora_mm(model.layers[il].wqkv, cur);
+        cb(cur, "wqkv", il);
+
+        const size_t type_size = ggml_type_size(cur->type);
+
+        ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*type_size, cur->nb[1], 0*type_size*(n_embd));
+        ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*type_size, cur->nb[1], 1*type_size*(n_embd));
+        ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*type_size, cur->nb[1], 1*type_size*(n_embd + n_embd_gqa));
+
+        // RoPE
+        Qcur = ggml_rope_ext(
+                ctx0, Qcur, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                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_l, freq_scale_l,
+                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, nullptr,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        cb(cur, "kqv_out", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // re-add the layer input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // attention layer norm
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM, 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_GEGLU, LLM_FFN_SEQ, il);
+
+        // attentions bypass the intermediate layer
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM, -1);
+    cb(cur, "final_norm_out", -1);
+
+    if (hparams.pooling_type == LLAMA_POOLING_TYPE_CLS) {
+        // extracting cls token
+        cur = ggml_view_1d(ctx0, cur, hparams.n_embd, 0);
+        cb(cur, "cls_pooled_embd", -1);
+    }
+
+    cb(cur, "res_embd", -1);
+    res->t_embd = cur;
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/openai-moe-iswa.cpp

@@ -14,6 +14,9 @@ llm_build_openai_moe_iswa::llm_build_openai_moe_iswa(const llama_model & model,
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
         ggml_tensor * inpSA = inpL;
 
         // norm
@@ -49,13 +52,13 @@ llm_build_openai_moe_iswa::llm_build_openai_moe_iswa(const llama_model & model,
 
             Qcur = ggml_rope_ext(
                     ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                     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,
+                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                     ext_factor, attn_factor, beta_fast, beta_slow
                     );
 

llama/llama.cpp/src/models/plamo3.cpp

@@ -0,0 +1,128 @@
+#include "models.h"
+
+template <bool iswa>
+llm_build_plamo3<iswa>::llm_build_plamo3(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t head_dim_q = hparams.n_embd_head_k;
+    const int64_t head_dim_v = hparams.n_embd_head_v;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL = build_inp_embd(model.tok_embd);
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * residual = inpL;
+
+        float freq_base_l  = 0.0f;
+        float freq_scale_l = 0.0f;
+        if constexpr (iswa) {
+            freq_base_l  = model.get_rope_freq_base (cparams, il);
+            freq_scale_l = model.get_rope_freq_scale(cparams, il);
+        } else {
+            freq_base_l  = freq_base;
+            freq_scale_l = freq_scale;
+        }
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
+        cb(cur, "wqkv", il);
+
+        const int32_t n_head    = hparams.n_head(il);
+        const int32_t n_head_kv = hparams.n_head_kv(il);
+
+        const int64_t q_offset = 0;
+        const int64_t k_offset = head_dim_q * n_head;
+        const int64_t v_offset = k_offset + head_dim_q * n_head_kv;
+
+        ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, head_dim_q, n_head, n_tokens,
+                head_dim_q * sizeof(float), qkv->nb[1], q_offset * ggml_element_size(qkv));
+        ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, head_dim_q, n_head_kv, n_tokens,
+                head_dim_q * sizeof(float), qkv->nb[1], k_offset * ggml_element_size(qkv));
+        ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, head_dim_v, n_head_kv, n_tokens,
+                head_dim_v * sizeof(float), qkv->nb[1], v_offset * ggml_element_size(qkv));
+
+        cb(Qcur, "Qcur", il);
+        cb(Kcur, "Kcur", il);
+        cb(Vcur, "Vcur", il);
+
+        Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+        cb(Qcur, "attn_q_norm", il);
+        Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+        cb(Kcur, "attn_k_norm", il);
+
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                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_l, freq_scale_l,
+                ext_factor, attn_factor, beta_fast, beta_slow);
+
+        const float attn_scale = 1.0f / sqrtf(float(head_dim_q));
+
+        cur = build_attn(inp_attn,
+                model.layers[il].wo, NULL,
+                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, attn_scale, il);
+        cb(cur, "attn_out", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);
+            residual = ggml_get_rows(ctx0, residual, inp_out_ids);
+        }
+
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, residual);
+        cb(cur, "attn_residual", il);
+
+        residual = cur;
+
+        cur = build_norm(cur, 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_SWIGLU, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        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, residual);
+        cb(cur, "ffn_residual", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+// Explicit template instantiations
+template struct llm_build_plamo3<false>;
+template struct llm_build_plamo3<true>;

llama/llama.cpp/src/models/smallthinker.cpp

@@ -26,10 +26,16 @@ llm_build_smallthinker<iswa>::llm_build_smallthinker(const llama_model & model,
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     for (int il = 0; il < n_layer; ++il) {
-        ggml_tensor * inpSA  = inpL;
-        ggml_tensor * probs  = nullptr;
+        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
 
-        probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL);  // [n_expert, n_tokens]
+        ggml_tensor * inpSA  = inpL;
+
+        // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
+        const bool use_rope = hparams.n_no_rope_layer_step == n_layer ||
+                              il % hparams.n_no_rope_layer_step != 0;
+
+        ggml_tensor * probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL);  // [n_expert, n_tokens]
         cb(probs, "ffn_moe_logits", il);
 
         // norm
@@ -52,11 +58,11 @@ llm_build_smallthinker<iswa>::llm_build_smallthinker(const llama_model & model,
             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);
 
-            if (hparams.n_no_rope_layer_step == n_layer || il % hparams.n_no_rope_layer_step != 0) {
-                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            if (use_rope) {
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                                     ext_factor, attn_factor, beta_fast, beta_slow);
 
-                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                                     ext_factor, attn_factor, beta_fast, beta_slow);
             }
             cb(Qcur, "Qcur", il);

llama/llama.cpp/src/unicode.cpp

@@ -985,6 +985,11 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
         { "\\p{P}", unicode_cpt_flags::PUNCTUATION },
         { "\\p{M}", unicode_cpt_flags::ACCENT_MARK },
         { "\\p{S}", unicode_cpt_flags::SYMBOL },
+        { "\\p{Lu}", unicode_cpt_flags::LETTER }, // Uppercase letter
+        { "\\p{Ll}", unicode_cpt_flags::LETTER }, // Lowercase letter
+        { "\\p{Lt}", unicode_cpt_flags::LETTER }, // Titlecase letter
+        { "\\p{Lm}", unicode_cpt_flags::LETTER }, // Modifier letter
+        { "\\p{Lo}", unicode_cpt_flags::LETTER }, // Other letter
     };
 
     static const std::map<int, int> k_ucat_cpt = {
@@ -1095,22 +1100,26 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                         continue;
                     }
 
-                    if (regex_expr[i + 0] == '\\' && i + 4 < regex_expr.size() &&
+                    // Match \p{...} Unicode properties of varying lengths
+                    if (regex_expr[i + 0] == '\\' && i + 3 < regex_expr.size() &&
                         regex_expr[i + 1] == 'p' &&
-                        regex_expr[i + 2] == '{' &&
-                        regex_expr[i + 4] == '}') {
-                        const std::string pat = regex_expr.substr(i, 5);
-                        if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {
-                            if (!inside) {
-                                regex_expr_collapsed += '[';
+                        regex_expr[i + 2] == '{') {
+                        // Find the closing brace
+                        size_t closing_brace = regex_expr.find('}', i + 3);
+                        if (closing_brace != std::string::npos && closing_brace <= i + 10) { // reasonable limit
+                            const std::string pat = regex_expr.substr(i, closing_brace - i + 1);
+                            if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {
+                                if (!inside) {
+                                    regex_expr_collapsed += '[';
+                                }
+                                regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));
+                                regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));
+                                if (!inside) {
+                                    regex_expr_collapsed += ']';
+                                }
+                                i = closing_brace;
+                                continue;
                             }
-                            regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));
-                            regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));
-                            if (!inside) {
-                                regex_expr_collapsed += ']';
-                            }
-                            i += 4;
-                            continue;
                         }
                     }
 

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

@@ -45,13 +45,14 @@
 #define KEY_SPATIAL_MERGE_SIZE  "clip.vision.spatial_merge_size"
 #define KEY_IS_DEEPSTACK_LAYERS "clip.vision.is_deepstack_layers"
 
-#define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
-#define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
-#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
-#define KEY_WIN_ATTN_PATTERN      "clip.vision.n_wa_pattern"
-#define KEY_ATTN_WINDOW_SIZE      "clip.vision.window_size"
-#define KEY_MINICPMV_VERSION      "clip.minicpmv_version"
-#define KEY_MINICPMV_QUERY_NUM    "clip.minicpmv_query_num"
+#define KEY_MM_PATCH_MERGE_TYPE    "clip.vision.mm_patch_merge_type"
+#define KEY_IMAGE_GRID_PINPOINTS   "clip.vision.image_grid_pinpoints"
+#define KEY_IMAGE_CROP_RESOLUTION  "clip.vision.image_crop_resolution"
+#define KEY_WIN_ATTN_PATTERN       "clip.vision.n_wa_pattern"
+#define KEY_WIN_ATTN_LAYER_INDEXES "clip.vision.wa_layer_indexes"
+#define KEY_ATTN_WINDOW_SIZE       "clip.vision.window_size"
+#define KEY_MINICPMV_VERSION       "clip.minicpmv_version"
+#define KEY_MINICPMV_QUERY_NUM     "clip.minicpmv_query_num"
 
 // audio-specific
 #define KEY_AUDIO_PROJ_TYPE     "clip.audio.projector_type" // for models with mixed modalities
@@ -138,6 +139,62 @@
 #define TN_TOK_BOI         "v.boi"
 #define TN_TOK_EOI         "v.eoi"
 
+// (conformer) lfm2
+#define TN_PRE_ENCODE_OUT  "a.pre_encode.out.%s"
+#define TN_FFN_NORM        "%s.blk.%d.ffn_norm.%s"
+#define TN_FFN_NORM_1      "%s.blk.%d.ffn_norm_1.%s"
+#define TN_FFN_UP_1        "%s.blk.%d.ffn_up_1.%s"
+#define TN_FFN_DOWN_1      "%s.blk.%d.ffn_down_1.%s"
+#define TN_POS_BIAS_U      "%s.blk.%d.pos_bias_u"
+#define TN_POS_BIAS_V      "%s.blk.%d.pos_bias_v"
+#define TN_NORM_CONV       "%s.blk.%d.norm_conv.%s"
+#define TN_LINEAR_POS      "%s.blk.%d.linear_pos.%s"
+#define TN_CONV_DW         "%s.blk.%d.conv_dw.%s"
+#define TN_CONV_NORM       "%s.blk.%d.conv_norm.%s"
+#define TN_CONV_PW1        "%s.blk.%d.conv_pw1.%s"
+#define TN_CONV_PW2        "%s.blk.%d.conv_pw2.%s"
+
+// mobilenetv5 (gemma3n) definitions
+#define TN_MNV5_STEM_CONV        "v.conv_stem.conv.weight"
+#define TN_MNV5_STEM_BIAS        "v.conv_stem.conv.bias"
+#define TN_MNV5_STEM_BN          "v.conv_stem.bn.weight"
+
+// Stage 0 Block (Edge Residual)
+#define TN_MNV5_BLK_S0_EXP_W     "v.blk.%d.%d.conv_exp.weight"
+#define TN_MNV5_BLK_S0_BN1_W     "v.blk.%d.%d.bn1.weight"
+#define TN_MNV5_BLK_S0_PWL_W     "v.blk.%d.%d.conv_pwl.weight"
+#define TN_MNV5_BLK_S0_BN2_W     "v.blk.%d.%d.bn2.weight"
+
+// Stage 1+ Block (Universal Inverted Residual)
+#define TN_MNV5_BLK_DW_START_W   "v.blk.%d.%d.dw_start.conv.weight"
+#define TN_MNV5_BLK_DW_START_BN  "v.blk.%d.%d.dw_start.bn.weight"
+#define TN_MNV5_BLK_DW_MID_W     "v.blk.%d.%d.dw_mid.conv.weight"
+#define TN_MNV5_BLK_DW_MID_BN    "v.blk.%d.%d.dw_mid.bn.weight"
+#define TN_MNV5_BLK_PW_EXP_W     "v.blk.%d.%d.pw_exp.conv.weight"
+#define TN_MNV5_BLK_PW_EXP_BN    "v.blk.%d.%d.pw_exp.bn.weight"
+#define TN_MNV5_BLK_PW_PROJ_W    "v.blk.%d.%d.pw_proj.conv.weight"
+#define TN_MNV5_BLK_PW_PROJ_BN   "v.blk.%d.%d.pw_proj.bn.weight"
+#define TN_MNV5_BLK_LAYER_SCALE  "v.blk.%d.%d.layer_scale.gamma"
+
+// Attention Components
+#define TN_MNV5_ATTN_Q_W         "v.blk.%d.%d.attn.query.proj.weight"
+#define TN_MNV5_ATTN_K_W         "v.blk.%d.%d.attn.key.proj.weight"
+#define TN_MNV5_ATTN_V_W         "v.blk.%d.%d.attn.value.proj.weight"
+#define TN_MNV5_ATTN_O_W         "v.blk.%d.%d.attn.output.proj.weight"
+#define TN_MNV5_ATTN_K_DW        "v.blk.%d.%d.attn.key.down_conv.weight"
+#define TN_MNV5_ATTN_K_NORM      "v.blk.%d.%d.attn.key.norm.weight"
+#define TN_MNV5_ATTN_V_DW        "v.blk.%d.%d.attn.value.down_conv.weight"
+#define TN_MNV5_ATTN_V_NORM      "v.blk.%d.%d.attn.value.norm.weight"
+#define TN_MNV5_ATTN_NORM        "v.blk.%d.%d.norm.weight" // Block norm used in attn blocks
+
+// MSFA
+#define TN_MNV5_MSFA_FFN_EXP_W   "v.msfa.ffn.pw_exp.conv.weight"
+#define TN_MNV5_MSFA_FFN_EXP_BN  "v.msfa.ffn.pw_exp.bn.weight"
+#define TN_MNV5_MSFA_FFN_PROJ_W  "v.msfa.ffn.pw_proj.conv.weight"
+#define TN_MNV5_MSFA_FFN_PROJ_BN "v.msfa.ffn.pw_proj.bn.weight"
+#define TN_MNV5_MSFA_NORM        "v.msfa.norm.weight"
+
+
 // align x to upper multiple of n
 #define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
 
@@ -155,6 +212,8 @@ enum projector_type {
     PROJECTOR_TYPE_QWEN2VL,
     PROJECTOR_TYPE_QWEN3VL,
     PROJECTOR_TYPE_GEMMA3,
+    PROJECTOR_TYPE_GEMMA3NV,
+    PROJECTOR_TYPE_GEMMA3NA,
     PROJECTOR_TYPE_IDEFICS3,
     PROJECTOR_TYPE_PIXTRAL,
     PROJECTOR_TYPE_QWEN25VL,
@@ -165,12 +224,15 @@ enum projector_type {
     PROJECTOR_TYPE_GLMA,
     PROJECTOR_TYPE_QWEN25O, // will be replaced by QWEN2A or QWEN25VL depending on clip_ctx
     PROJECTOR_TYPE_VOXTRAL,
+    PROJECTOR_TYPE_MUSIC_FLAMINGO,
     PROJECTOR_TYPE_LFM2,
     PROJECTOR_TYPE_KIMIVL,
     PROJECTOR_TYPE_LIGHTONOCR,
     PROJECTOR_TYPE_COGVLM,
     PROJECTOR_TYPE_JANUS_PRO,
+    PROJECTOR_TYPE_LFM2A,
     PROJECTOR_TYPE_GLM4V,
+    PROJECTOR_TYPE_YOUTUVL,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -184,6 +246,8 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_QWEN25VL,  "qwen2.5vl_merger"},
     { PROJECTOR_TYPE_QWEN3VL,   "qwen3vl_merger"},
     { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
+    { PROJECTOR_TYPE_GEMMA3NV,  "gemma3nv"},
+    { PROJECTOR_TYPE_GEMMA3NA,  "gemma3na"},
     { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
     { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
     { PROJECTOR_TYPE_ULTRAVOX,  "ultravox"},
@@ -193,12 +257,15 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_GLMA,      "glma"},
     { PROJECTOR_TYPE_QWEN25O,   "qwen2.5o"},
     { PROJECTOR_TYPE_VOXTRAL,   "voxtral"},
+    { PROJECTOR_TYPE_MUSIC_FLAMINGO, "musicflamingo"},
     { PROJECTOR_TYPE_LFM2,      "lfm2"},
     { PROJECTOR_TYPE_KIMIVL,    "kimivl"},
     { PROJECTOR_TYPE_LIGHTONOCR,"lightonocr"},
     { PROJECTOR_TYPE_COGVLM,    "cogvlm"},
     { PROJECTOR_TYPE_JANUS_PRO, "janus_pro"},
+    { PROJECTOR_TYPE_LFM2A,     "lfm2a"},
     { PROJECTOR_TYPE_GLM4V,     "glm4v"},
+    { PROJECTOR_TYPE_YOUTUVL,   "youtuvl"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {

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

@@ -4,6 +4,7 @@
 #include "clip.h"
 #include "clip-impl.h"
 
+#include <array>
 #include <vector>
 #include <unordered_set>
 #include <cstdint>
@@ -60,6 +61,7 @@ struct clip_hparams {
     std::unordered_set<int32_t> vision_feature_layer;
     int32_t attn_window_size = 0;
     int32_t n_wa_pattern = 0;
+    std::unordered_set<int32_t> wa_layer_indexes; // explicit layer indexes that use full attention (for irregular patterns like YoutuVL)
 
     // audio
     int32_t n_mel_bins = 0; // whisper preprocessor
@@ -142,11 +144,74 @@ struct clip_layer {
     ggml_tensor * deepstack_fc2_w = nullptr;
     ggml_tensor * deepstack_fc2_b = nullptr;
 
+    // lfm2
+    ggml_tensor * ff_norm_w     = nullptr;
+    ggml_tensor * ff_norm_b     = nullptr;
+    ggml_tensor * ff_norm_1_w   = nullptr;
+    ggml_tensor * ff_norm_1_b   = nullptr;
+    ggml_tensor * ff_up_1_w     = nullptr;
+    ggml_tensor * ff_up_1_b     = nullptr;
+    ggml_tensor * ff_down_1_w   = nullptr;
+    ggml_tensor * ff_down_1_b   = nullptr;
+    ggml_tensor * pos_bias_u    = nullptr;
+    ggml_tensor * pos_bias_v    = nullptr;
+    ggml_tensor * norm_conv_w   = nullptr;
+    ggml_tensor * norm_conv_b   = nullptr;
+    ggml_tensor * linear_pos_w  = nullptr;
+
+    ggml_tensor * conv_norm_w   = nullptr;
+    ggml_tensor * conv_norm_b   = nullptr;
+    ggml_tensor * conv_dw_w     = nullptr;
+    ggml_tensor * conv_dw_b     = nullptr;
+    ggml_tensor * conv_pw1_w    = nullptr;
+    ggml_tensor * conv_pw1_b    = nullptr;
+    ggml_tensor * conv_pw2_w    = nullptr;
+    ggml_tensor * conv_pw2_b    = nullptr;
+
     bool has_deepstack() const {
         return deepstack_fc1_w != nullptr;
     }
 };
 
+// Expanded MobileNetV5 block structure for Gemma3n vision encoder
+struct mobilenetv5_block {
+    // Stage 0 (Edge Residual)
+    ggml_tensor * s0_conv_exp_w = nullptr;
+    ggml_tensor * s0_bn1_w      = nullptr;
+    ggml_tensor * s0_conv_pwl_w = nullptr;
+    ggml_tensor * s0_bn2_w      = nullptr;
+
+    // Stage 1+ (Universal Inverted Residual)
+    ggml_tensor * dw_start_w    = nullptr;
+    ggml_tensor * dw_start_bn_w = nullptr;
+
+    ggml_tensor * pw_exp_w      = nullptr;
+    ggml_tensor * pw_exp_bn_w   = nullptr;
+
+    ggml_tensor * dw_mid_w      = nullptr;
+    ggml_tensor * dw_mid_bn_w   = nullptr;
+
+    ggml_tensor * pw_proj_w     = nullptr;
+    ggml_tensor * pw_proj_bn_w  = nullptr;
+
+    ggml_tensor * layer_scale_w = nullptr;
+
+    // Attention (MQA) components
+    ggml_tensor * attn_q_w = nullptr;
+    ggml_tensor * attn_k_w = nullptr;
+    ggml_tensor * attn_v_w = nullptr;
+    ggml_tensor * attn_o_w = nullptr;
+
+    // Optional downsampling/norm in attention
+    ggml_tensor * attn_k_dw_w   = nullptr;
+    ggml_tensor * attn_k_norm_w = nullptr;
+    ggml_tensor * attn_v_dw_w   = nullptr;
+    ggml_tensor * attn_v_norm_w = nullptr;
+
+    // Block norm (often present in attention blocks)
+    ggml_tensor * attn_norm_w   = nullptr;
+};
+
 struct clip_model {
     clip_modality modality = CLIP_MODALITY_VISION;
     projector_type proj_type = PROJECTOR_TYPE_MLP;
@@ -263,6 +328,23 @@ struct clip_model {
     ggml_tensor * mm_input_proj_w = nullptr;
     ggml_tensor * mm_soft_emb_norm_w = nullptr;
 
+    // mobilenetv5 for gemma3n
+    std::vector<mobilenetv5_block> mobilenet_blocks;
+    std::vector<int> mobilenet_stage_ends;
+    ggml_tensor * mobilenet_stem_conv_w = nullptr;
+    ggml_tensor * mobilenet_stem_conv_b = nullptr;
+    ggml_tensor * mobilenet_stem_norm_w = nullptr;
+    ggml_tensor * mm_post_proj_norm_w = nullptr;
+
+    // Multi-Scale Fusion Adapter (MSFA) components
+    ggml_tensor * msfa_concat_conv_w = nullptr;
+    ggml_tensor * msfa_concat_norm_w = nullptr;
+    ggml_tensor * msfa_ffn_expand_w = nullptr;
+    ggml_tensor * msfa_ffn_project_w = nullptr;
+    ggml_tensor * msfa_ffn_expand_bn = nullptr;
+    ggml_tensor * msfa_ffn_project_bn = nullptr;
+
+
     // pixtral, glm4v
     ggml_tensor * token_embd_img_break = nullptr;
     ggml_tensor * mm_patch_merger_w = nullptr;
@@ -286,9 +368,16 @@ struct clip_model {
     ggml_tensor * mm_boi = nullptr;
     ggml_tensor * mm_eoi = nullptr;
 
+    // lfm2 audio
+    std::array<ggml_tensor *, 7> pre_encode_conv_X_w = {nullptr};
+    std::array<ggml_tensor *, 7> pre_encode_conv_X_b = {nullptr};
+    ggml_tensor * pre_encode_out_w = nullptr;
+    ggml_tensor * pre_encode_out_b = nullptr;
+
     bool audio_has_avgpool() const {
         return proj_type == PROJECTOR_TYPE_QWEN2A
-            || proj_type == PROJECTOR_TYPE_VOXTRAL;
+            || proj_type == PROJECTOR_TYPE_VOXTRAL
+            || proj_type == PROJECTOR_TYPE_MUSIC_FLAMINGO;
     }
 
     bool audio_has_stack_frames() const {

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

@@ -801,6 +801,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             {
                 builder = std::make_unique<clip_graph_siglip>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_GEMMA3NV:
+            {
+                builder = std::make_unique<clip_graph_mobilenetv5>(ctx, img);
+            } break;
         case PROJECTOR_TYPE_PIXTRAL:
         case PROJECTOR_TYPE_LIGHTONOCR:
             {
@@ -831,6 +835,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         case PROJECTOR_TYPE_VOXTRAL:
         case PROJECTOR_TYPE_QWEN2A:
         case PROJECTOR_TYPE_GLMA:
+        case PROJECTOR_TYPE_MUSIC_FLAMINGO:
             {
                 builder = std::make_unique<clip_graph_whisper_enc>(ctx, img);
             } break;
@@ -850,10 +855,18 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             {
                 builder = std::make_unique<clip_graph_llava>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_LFM2A:
+            {
+                builder = std::make_unique<clip_graph_conformer>(ctx, img);
+            } break;
         case PROJECTOR_TYPE_GLM4V:
             {
                 builder = std::make_unique<clip_graph_glm4v>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_YOUTUVL:
+            {
+                builder = std::make_unique<clip_graph_youtuvl>(ctx, img);
+            } break;
         default:
             GGML_ABORT("missing cgraph builder");
     }
@@ -1154,6 +1167,14 @@ struct clip_model_loader {
                         // test model (tinygemma3) has a different value, we optionally read it
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                     } break;
+
+                case PROJECTOR_TYPE_GEMMA3NV:
+                    {
+                        // Gemma3n uses MobileNetV5 which produces 256 tokens (16x16)
+                        // Similar configuration to Gemma3
+                        hparams.n_merge = 1;  // MobileNetV5 handles resizing internally
+                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
+                    } break;
                 case PROJECTOR_TYPE_QWEN2VL:
                 case PROJECTOR_TYPE_QWEN25VL:
                 case PROJECTOR_TYPE_QWEN3VL:
@@ -1171,6 +1192,20 @@ struct clip_model_loader {
                             LOG_WRN("%s: more info: https://github.com/ggml-org/llama.cpp/issues/16842\n\n", __func__);
                         }
                     } break;
+                case PROJECTOR_TYPE_YOUTUVL:
+                    {
+                        hparams.n_merge = 2;
+                        get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
+                        get_u32(KEY_ATTN_WINDOW_SIZE, hparams.attn_window_size, true);
+                        std::vector<int> wa_layer_indexes_vec;
+                        get_arr_int(KEY_WIN_ATTN_LAYER_INDEXES, wa_layer_indexes_vec, true);
+                        for (auto & layer : wa_layer_indexes_vec) {
+                            hparams.wa_layer_indexes.insert(layer);
+                        }
+                        // support max_height * max_width = 8000 * 8000. 8000/16/2 = 250 image tokens
+                        hparams.set_limit_image_tokens(1, 62500);
+                        hparams.set_warmup_n_tokens(16*16); // avoid OOM on warmup
+                    } break;
                 case PROJECTOR_TYPE_GLM4V:
                     {
                         hparams.rope_theta = 10000.0f;
@@ -1189,6 +1224,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_QWEN2A:
                 case PROJECTOR_TYPE_GLMA:
                 case PROJECTOR_TYPE_VOXTRAL:
+                case PROJECTOR_TYPE_MUSIC_FLAMINGO:
                     {
                         bool require_stack = model.proj_type == PROJECTOR_TYPE_ULTRAVOX ||
                                              model.proj_type == PROJECTOR_TYPE_VOXTRAL ||
@@ -1204,6 +1240,15 @@ struct clip_model_loader {
                         hparams.audio_window_len   = 400;
                         hparams.audio_hop_len      = 160;
                     } break;
+                case PROJECTOR_TYPE_LFM2A:
+                    {
+                        // audio preprocessing params
+                        hparams.audio_chunk_len        = 1; // in seconds
+                        hparams.audio_sample_rate      = 16000;
+                        hparams.audio_n_fft            = 512;
+                        hparams.audio_window_len       = 400;
+                        hparams.audio_hop_len          = 160;
+                    } break;
                 default:
                     break;
             }
@@ -1229,7 +1274,14 @@ struct clip_model_loader {
                 LOG_INF("%s: has_llava_proj:     %d\n", __func__, hparams.has_llava_projector);
                 LOG_INF("%s: minicpmv_version:   %d\n", __func__, hparams.minicpmv_version);
                 LOG_INF("%s: n_merge:            %d\n", __func__, hparams.n_merge);
-                LOG_INF("%s: n_wa_pattern:       %d\n", __func__, hparams.n_wa_pattern);
+                LOG_INF("%s: n_wa_pattern: %d\n", __func__, hparams.n_wa_pattern);
+                if (!hparams.wa_layer_indexes.empty()) {
+                    LOG_INF("%s: wa_layer_indexes:  ", __func__);
+                    for (auto & layer : hparams.wa_layer_indexes) {
+                        LOG_INF("%d ", layer);
+                    }
+                    LOG_INF("\n");
+                }
                 if (hparams.image_min_pixels > 0) {
                     LOG_INF("%s: image_min_pixels:   %d%s\n", __func__, hparams.image_min_pixels, hparams.custom_image_min_tokens > 0 ? " (custom value)" : "");
                 }
@@ -1311,6 +1363,10 @@ struct clip_model_loader {
 
         model.position_embeddings = get_tensor(string_format(TN_POS_EMBD, prefix), false);
 
+        if (model.proj_type == PROJECTOR_TYPE_GEMMA3NV) {
+            hparams.n_layer = 0; // gemma3n does not use normal layer structure
+        }
+
         // layers
         model.layers.resize(hparams.n_layer);
         for (int il = 0; il < hparams.n_layer; ++il) {
@@ -1385,6 +1441,7 @@ struct clip_model_loader {
             }
         }
 
+
         switch (model.proj_type) {
             case PROJECTOR_TYPE_MLP:
             case PROJECTOR_TYPE_MLP_NORM:
@@ -1497,6 +1554,14 @@ struct clip_model_loader {
                     model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
                     model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
                 } break;
+            case PROJECTOR_TYPE_YOUTUVL:
+                {
+                    model.mm_input_norm_w = get_tensor(TN_MM_INP_NORM);        // merger.ln_q (RMS norm)
+                    model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight"));  // merger.mlp.0
+                    model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias"));
+                    model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));  // merger.mlp.2
+                    model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
+                } break;
             case PROJECTOR_TYPE_GLM4V:
                 {
                     model.projection     = get_tensor(TN_MM_PROJECTOR);
@@ -1516,11 +1581,112 @@ struct clip_model_loader {
                     model.mm_input_proj_w = get_tensor(TN_MM_INP_PROJ);
                     model.mm_soft_emb_norm_w = get_tensor(TN_MM_SOFT_EMB_N);
                 } break;
+            case PROJECTOR_TYPE_GEMMA3NV:
+                {
+                    model.mobilenet_stem_conv_w = get_tensor(TN_MNV5_STEM_CONV, false);
+                    model.mobilenet_stem_conv_b = get_tensor(TN_MNV5_STEM_BIAS, false);
+                    model.mobilenet_stem_norm_w = get_tensor(TN_MNV5_STEM_BN, false);
+
+                    model.msfa_ffn_expand_w  = get_tensor(TN_MNV5_MSFA_FFN_EXP_W, false);
+                    model.msfa_ffn_expand_bn = get_tensor(TN_MNV5_MSFA_FFN_EXP_BN, false); // Consume BN if present but likely folded
+                    model.msfa_ffn_project_w = get_tensor(TN_MNV5_MSFA_FFN_PROJ_W, false);
+                    model.msfa_ffn_project_bn = get_tensor(TN_MNV5_MSFA_FFN_PROJ_BN, false);
+
+                    model.msfa_concat_norm_w = get_tensor(TN_MNV5_MSFA_NORM, false);
+
+                    // Dynamically load blocks stage by stage
+                    for (int stage = 0; stage < 4; ++stage) {
+                        int blocks_found_in_stage = 0;
+
+                        for (int blk_idx = 0; ; ++blk_idx) {
+                            bool found_block = false;
+                            mobilenetv5_block block;
+
+                            // 1. Check for Edge Residual (S0)
+                            block.s0_conv_exp_w = get_tensor(string_format(TN_MNV5_BLK_S0_EXP_W, stage, blk_idx), false);
+                            if (block.s0_conv_exp_w) {
+                                found_block = true;
+                                block.s0_bn1_w      = get_tensor(string_format(TN_MNV5_BLK_S0_BN1_W, stage, blk_idx), false);
+                                block.s0_conv_pwl_w = get_tensor(string_format(TN_MNV5_BLK_S0_PWL_W, stage, blk_idx), false);
+                                block.s0_bn2_w      = get_tensor(string_format(TN_MNV5_BLK_S0_BN2_W, stage, blk_idx), false);
+                            }
+                            // 2. Check for UIR (Universal Inverted Residual)
+                            else {
+                                // Check for dw_start OR pw_exp (some UIR blocks skip dw_start)
+                                block.dw_start_w = get_tensor(string_format(TN_MNV5_BLK_DW_START_W, stage, blk_idx), false);
+                                block.pw_exp_w   = get_tensor(string_format(TN_MNV5_BLK_PW_EXP_W, stage, blk_idx), false);
+
+                                if (block.dw_start_w || block.pw_exp_w) {
+                                    found_block = true;
+                                    if (block.dw_start_w) {
+                                        block.dw_start_bn_w = get_tensor(string_format(TN_MNV5_BLK_DW_START_BN, stage, blk_idx), false);
+                                    }
+                                    if (block.pw_exp_w) {
+                                        block.pw_exp_bn_w   = get_tensor(string_format(TN_MNV5_BLK_PW_EXP_BN, stage, blk_idx), false);
+                                    }
+                                    block.dw_mid_w      = get_tensor(string_format(TN_MNV5_BLK_DW_MID_W, stage, blk_idx), false);
+                                    if (block.dw_mid_w) {
+                                        block.dw_mid_bn_w   = get_tensor(string_format(TN_MNV5_BLK_DW_MID_BN, stage, blk_idx), false);
+                                    }
+                                    block.pw_proj_w     = get_tensor(string_format(TN_MNV5_BLK_PW_PROJ_W, stage, blk_idx), false);
+                                    if (block.pw_proj_w) {
+                                        block.pw_proj_bn_w  = get_tensor(string_format(TN_MNV5_BLK_PW_PROJ_BN, stage, blk_idx), false);
+                                    }
+                                    block.layer_scale_w = get_tensor(string_format(TN_MNV5_BLK_LAYER_SCALE, stage, blk_idx), false);
+                                }
+                            }
+
+                            // 3. Check for Attention (MQA)
+                            // Even if UIR/Edge check failed, this might be a pure attention block
+                            ggml_tensor* attn_q_check = get_tensor(string_format(TN_MNV5_ATTN_Q_W, stage, blk_idx), false);
+                            if (attn_q_check) {
+                                found_block = true;
+                                block.attn_q_w = attn_q_check;
+                                block.attn_k_w = get_tensor(string_format(TN_MNV5_ATTN_K_W, stage, blk_idx), false);
+                                block.attn_v_w = get_tensor(string_format(TN_MNV5_ATTN_V_W, stage, blk_idx), false);
+                                block.attn_o_w = get_tensor(string_format(TN_MNV5_ATTN_O_W, stage, blk_idx), false);
+                                block.attn_k_dw_w   = get_tensor(string_format(TN_MNV5_ATTN_K_DW, stage, blk_idx), false);
+                                block.attn_k_norm_w = get_tensor(string_format(TN_MNV5_ATTN_K_NORM, stage, blk_idx), false);
+                                block.attn_v_dw_w   = get_tensor(string_format(TN_MNV5_ATTN_V_DW, stage, blk_idx), false);
+                                block.attn_v_norm_w = get_tensor(string_format(TN_MNV5_ATTN_V_NORM, stage, blk_idx), false);
+                                block.attn_norm_w   = get_tensor(string_format(TN_MNV5_ATTN_NORM, stage, blk_idx), false);
+                                // Note: Attention blocks also have layer_scale, load it if not already loaded by UIR check
+                                if (!block.layer_scale_w) {
+                                    block.layer_scale_w = get_tensor(string_format(TN_MNV5_BLK_LAYER_SCALE, stage, blk_idx), false);
+                                }
+                            }
+
+                            if (found_block) {
+                                model.mobilenet_blocks.push_back(block);
+                                blocks_found_in_stage++;
+                            } else {
+                                // End of blocks for this stage
+                                break;
+                            }
+                        }
+
+                        // Track where this stage ends in the flat vector
+                        if (blocks_found_in_stage > 0) {
+                            model.mobilenet_stage_ends.push_back(model.mobilenet_blocks.size() - 1);
+                            LOG_INF("%s: Stage %d ended at global block index %zu\n", __func__, stage, model.mobilenet_blocks.size() - 1);
+                        }
+                    }
+                    model.mm_input_proj_w = get_tensor(TN_MM_INP_PROJ);
+                    model.mm_soft_emb_norm_w = get_tensor(TN_MM_SOFT_EMB_N);
+                } break;
             case PROJECTOR_TYPE_IDEFICS3:
                 {
                     model.projection = get_tensor(TN_MM_PROJECTOR);
                 } break;
             case PROJECTOR_TYPE_LFM2:
+                {
+                    model.mm_input_norm_w = get_tensor(TN_MM_INP_NORM, false);
+                    model.mm_input_norm_b = get_tensor(TN_MM_INP_NORM_B, false);
+                    model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
+                    model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
+                    model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
+                    model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
+                } break;
             case PROJECTOR_TYPE_KIMIVL:
                 {
                     model.mm_input_norm_w = get_tensor(TN_MM_INP_NORM);
@@ -1580,6 +1746,17 @@ struct clip_model_loader {
                     model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
                     model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
                 } break;
+            case PROJECTOR_TYPE_MUSIC_FLAMINGO:
+                {
+                    model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
+                    model.conv1d_1_b = get_tensor(string_format(TN_CONV1D, 1, "bias"));
+                    model.conv1d_2_w = get_tensor(string_format(TN_CONV1D, 2, "weight"));
+                    model.conv1d_2_b = get_tensor(string_format(TN_CONV1D, 2, "bias"));
+                    model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
+                    model.mm_1_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "bias"));
+                    model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
+                    model.mm_2_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "bias"));
+                } break;
             case PROJECTOR_TYPE_INTERNVL:
                 {
                     model.mm_0_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight"));
@@ -1628,6 +1805,52 @@ struct clip_model_loader {
                     model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
                     model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
                 } break;
+            case PROJECTOR_TYPE_LFM2A:
+                {
+                    for (int i : {0, 2, 3, 5, 6}) {
+                        model.pre_encode_conv_X_w[i] = get_tensor(string_format(TN_CONV1D, i, "weight"));
+                        model.pre_encode_conv_X_b[i] = get_tensor(string_format(TN_CONV1D, i, "bias"));
+                    }
+                    model.pre_encode_out_w    = get_tensor(string_format(TN_PRE_ENCODE_OUT, "weight"));
+                    model.pre_encode_out_b    = get_tensor(string_format(TN_PRE_ENCODE_OUT, "bias"));
+
+                    model.mm_0_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 0, "weight"));
+                    model.mm_0_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 0, "bias"));
+                    model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
+                    model.mm_1_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "bias"));
+                    model.mm_3_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 3, "weight"));
+                    model.mm_3_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 3, "bias"));
+
+                    for (int il = 0; il < hparams.n_layer; ++il) {
+                        auto & layer = model.layers[il];
+
+                        layer.ff_norm_w   = get_tensor(string_format(TN_FFN_NORM,   prefix, il, "weight"));
+                        layer.ff_norm_b   = get_tensor(string_format(TN_FFN_NORM,   prefix, il, "bias"));
+                        layer.ff_norm_1_w = get_tensor(string_format(TN_FFN_NORM_1, prefix, il, "weight"));
+                        layer.ff_norm_1_b = get_tensor(string_format(TN_FFN_NORM_1, prefix, il, "bias"));
+                        layer.ff_up_1_w   = get_tensor(string_format(TN_FFN_UP_1,   prefix, il, "weight"));
+                        layer.ff_up_1_b   = get_tensor(string_format(TN_FFN_UP_1,   prefix, il, "bias"));
+                        layer.ff_down_1_w = get_tensor(string_format(TN_FFN_DOWN_1, prefix, il, "weight"));
+                        layer.ff_down_1_b = get_tensor(string_format(TN_FFN_DOWN_1, prefix, il, "bias"));
+
+                        layer.pos_bias_u = get_tensor(string_format(TN_POS_BIAS_U, prefix, il));
+                        layer.pos_bias_v = get_tensor(string_format(TN_POS_BIAS_V, prefix, il));
+
+                        layer.norm_conv_w = get_tensor(string_format(TN_NORM_CONV, prefix, il, "weight"));
+                        layer.norm_conv_b = get_tensor(string_format(TN_NORM_CONV, prefix, il, "bias"));
+
+                        layer.linear_pos_w = get_tensor(string_format(TN_LINEAR_POS, prefix, il, "weight"));
+
+                        layer.conv_norm_w  = get_tensor(string_format(TN_CONV_NORM, prefix, il, "weight"));
+                        layer.conv_norm_b  = get_tensor(string_format(TN_CONV_NORM, prefix, il, "bias"));
+                        layer.conv_dw_w    = get_tensor(string_format(TN_CONV_DW,   prefix, il, "weight"));
+                        layer.conv_dw_b    = get_tensor(string_format(TN_CONV_DW,   prefix, il, "bias"));
+                        layer.conv_pw1_w   = get_tensor(string_format(TN_CONV_PW1,  prefix, il, "weight"));
+                        layer.conv_pw1_b   = get_tensor(string_format(TN_CONV_PW1,  prefix, il, "bias"));
+                        layer.conv_pw2_w   = get_tensor(string_format(TN_CONV_PW2,  prefix, il, "weight"));
+                        layer.conv_pw2_b   = get_tensor(string_format(TN_CONV_PW2,  prefix, il, "bias"));
+                    }
+                } break;
             default:
                 GGML_ASSERT(false && "unknown projector type");
         }
@@ -1932,6 +2155,7 @@ struct clip_init_result clip_init(const char * fname, struct clip_context_params
 
     try {
         clip_model_loader loader(fname);
+        bool skip_audio = false;
 
         if (loader.has_vision) {
             ctx_vision = new clip_ctx(ctx_params);
@@ -1941,10 +2165,14 @@ struct clip_init_result clip_init(const char * fname, struct clip_context_params
                 loader.warmup(*ctx_vision);
             }
 
+            // TODO: we don't support audio for Gemma 3N, but GGUF contains audio tensors
+            // we can remove this check when we implement audio support for Gemma 3N
+            skip_audio = ctx_vision->model.proj_type == PROJECTOR_TYPE_GEMMA3NV;
+
             // clip_debug_encode(ctx_vision, 24*14, 24*14, 0.5f);
         }
 
-        if (loader.has_audio) {
+        if (loader.has_audio && !skip_audio) {
             ctx_audio = new clip_ctx(ctx_params);
             loader.load_hparams(ctx_audio->model, CLIP_MODALITY_AUDIO);
             loader.load_tensors(*ctx_audio);
@@ -2668,6 +2896,57 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
                 // res_imgs->data[0] = *res;
                 res_imgs->entries.push_back(std::move(img_f32));
             } break;
+        case PROJECTOR_TYPE_YOUTUVL:
+            {
+                const int patch_size = params.patch_size;  // typically 16
+                const int merge_size = params.n_merge;      // typically 2
+                const int align_size = patch_size * merge_size;  // 32
+
+                const int max_num_patches = params.image_max_pixels > 0 ?
+                    params.image_max_pixels / (patch_size * patch_size) : 256;
+
+                // Linear search for optimal scale to fit within max_num_patches
+                float scale = 1.0f;
+                int target_height = original_size.height;
+                int target_width = original_size.width;
+
+                auto get_scaled_image_size = [align_size](float scale, int size) -> int {
+                    float scaled_size = size * scale;
+                    // Round up to nearest multiple of align_size
+                    int aligned = static_cast<int>(std::ceil(scaled_size / align_size)) * align_size;
+                    // Ensure at least one patch
+                    return std::max(align_size, aligned);
+                };
+
+                // Linear search with 0.02 step size
+                while (scale > 0.0f) {
+                    target_height = get_scaled_image_size(scale, original_size.height);
+                    target_width = get_scaled_image_size(scale, original_size.width);
+
+                    int num_patches_h = target_height / patch_size;
+                    int num_patches_w = target_width / patch_size;
+                    int num_patches = num_patches_h * num_patches_w;
+
+                    if (num_patches > max_num_patches) {
+                        scale -= 0.02f;
+                    } else {
+                        break;
+                    }
+                }
+
+                clip_image_size new_size = {target_width, target_height};
+
+                // Resize the image
+                clip_image_u8 resized;
+                img_tool::resize(*img, resized, new_size, img_tool::RESIZE_ALGO_BILINEAR, false);
+
+                // Normalize to float32
+                clip_image_f32_ptr img_f32(clip_image_f32_init());
+                normalize_image_u8_to_f32(resized, *img_f32, params.image_mean, params.image_std);
+
+                // Add to results
+                res_imgs->entries.push_back(std::move(img_f32));
+            } break;
 
         case PROJECTOR_TYPE_IDEFICS3:
             {
@@ -2731,6 +3010,16 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
                 res_imgs->entries.push_back(std::move(img_f32));
             } break;
 
+        case PROJECTOR_TYPE_GEMMA3NV:
+            {
+                clip_image_u8 resized_image;
+                int sz = params.image_size;
+                img_tool::resize(*img, resized_image, {sz, sz}, img_tool::RESIZE_ALGO_BILINEAR, false);
+                clip_image_f32_ptr img_f32(clip_image_f32_init());
+                normalize_image_u8_to_f32(resized_image, *img_f32, params.image_mean, params.image_std);
+                res_imgs->entries.push_back(std::move(img_f32));
+            } break;
+
         case PROJECTOR_TYPE_JANUS_PRO:
             {
                 // Janus Pro preprocessing: pad to square with gray(127), resize to 384x384
@@ -2900,6 +3189,7 @@ int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 *
         case PROJECTOR_TYPE_QWEN25VL:
         case PROJECTOR_TYPE_QWEN3VL:
         case PROJECTOR_TYPE_GLM4V:
+        case PROJECTOR_TYPE_YOUTUVL:
             return (img->nx / params.patch_size) / 2;
         default:
             break;
@@ -2915,6 +3205,7 @@ int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 *
         case PROJECTOR_TYPE_QWEN25VL:
         case PROJECTOR_TYPE_QWEN3VL:
         case PROJECTOR_TYPE_GLM4V:
+        case PROJECTOR_TYPE_YOUTUVL:
             return (img->ny / params.patch_size) / 2;
         default:
             break;
@@ -2975,6 +3266,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_QWEN25VL:
         case PROJECTOR_TYPE_QWEN3VL:
         case PROJECTOR_TYPE_GLM4V:
+        case PROJECTOR_TYPE_YOUTUVL:
             {
                 // dynamic size (2 conv, so double patch size)
                 int x_patch = img->nx / (params.patch_size * 2);
@@ -2990,6 +3282,12 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
                 int scale_factor = ctx->model.hparams.n_merge;
                 n_patches /= (scale_factor * scale_factor);
             } break;
+        case PROJECTOR_TYPE_GEMMA3NV:
+            {
+                // MobileNetV5 MSFA adapter always outputs fixed 16x16 resolution
+                // regardless of input size (see architecture description)
+                n_patches = ctx->model.hparams.image_size / ctx->model.hparams.patch_size;
+            } break;
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_KIMIVL:
             {
@@ -3015,6 +3313,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_VOXTRAL:
         case PROJECTOR_TYPE_ULTRAVOX:
         case PROJECTOR_TYPE_QWEN2A:
+        case PROJECTOR_TYPE_MUSIC_FLAMINGO:
             {
                 n_patches = img->nx;
 
@@ -3047,6 +3346,10 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
             {
                 n_patches += 2; // for BOI and EOI token embeddings
             } break;
+        case PROJECTOR_TYPE_LFM2A:
+            {
+                n_patches = ((((img->nx + 1) / 2) + 1) / 2 + 1) / 2;
+            } break;
         default:
             GGML_ABORT("unsupported projector type");
     }
@@ -3097,7 +3400,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     const int pos_w = image_size_width  / patch_size;
     const int pos_h = image_size_height / patch_size;
 
-    const bool use_window_attn = hparams.n_wa_pattern > 0; // for qwen2.5vl
 
     auto get_inp_tensor = [&gf](const char * name) {
         ggml_tensor * inp = ggml_graph_get_tensor(gf, name);
@@ -3246,9 +3548,11 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                 set_input_i32("positions", positions);
             } break;
         case PROJECTOR_TYPE_QWEN25VL:
+        case PROJECTOR_TYPE_YOUTUVL:
             {
                 // pw * ph = number of tokens output by ViT after apply patch merger
                 // ipw * ipw = number of vision token been processed inside ViT
+                const bool use_window_attn = ctx->model.proj_type == PROJECTOR_TYPE_QWEN25VL ? hparams.n_wa_pattern > 0 : !hparams.wa_layer_indexes.empty();
                 const int merge_ratio = 2;
                 const int pw  = image_size_width  / patch_size / merge_ratio;
                 const int ph  = image_size_height / patch_size / merge_ratio;
@@ -3259,7 +3563,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                 std::vector<int> inv_idx(ph * pw);
 
                 if (use_window_attn) {
-                    const int attn_window_size = 112;
+                    const int attn_window_size = hparams.attn_window_size > 0 ? hparams.attn_window_size : 112;
                     const int grid_window = attn_window_size / patch_size / merge_ratio;
                     int dst = 0;
                     // [num_vision_tokens, num_vision_tokens] attention mask tensor
@@ -3376,6 +3680,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                 set_input_i32("patches", patches);
             } break;
         case PROJECTOR_TYPE_GEMMA3:
+        case PROJECTOR_TYPE_GEMMA3NV:
         case PROJECTOR_TYPE_IDEFICS3:
         case PROJECTOR_TYPE_INTERNVL:
         case PROJECTOR_TYPE_QWEN2A:
@@ -3383,6 +3688,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         case PROJECTOR_TYPE_ULTRAVOX:
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_VOXTRAL:
+        case PROJECTOR_TYPE_MUSIC_FLAMINGO:
         case PROJECTOR_TYPE_JANUS_PRO:
         case PROJECTOR_TYPE_COGVLM:
             {
@@ -3405,6 +3711,27 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                 }
                 set_input_i32("pos_w", pos_data);
             } break;
+        case PROJECTOR_TYPE_LFM2A:
+            {
+                GGML_ASSERT(imgs.entries.size() == 1);
+                const auto n_frames = clip_n_output_tokens(ctx, imgs.entries.front().get());
+
+                auto d_model = 512;
+                auto seq_len = n_frames * 2 - 1;
+                std::vector<float> pos_emb(d_model*seq_len);
+                std::vector<double> inv_freq(d_model / 2);
+                for (size_t i = 0; i < inv_freq.size(); ++i) {
+                    inv_freq[i] = std::exp(-(std::log(10000.0) / (float)d_model) * (2.0f * (float)(i)));
+                }
+                for (int64_t pos = 0; pos < seq_len; ++pos) {
+                    for (size_t i = 0; i < inv_freq.size(); ++i) {
+                        const float ang = (n_frames - pos - 1) * inv_freq[i];
+                        pos_emb[pos*d_model + 2*i + 0] = sinf(ang);  // even
+                        pos_emb[pos*d_model + 2*i + 1] = cosf(ang);  // odd
+                    }
+                }
+                set_input_f32("pos_emb", pos_emb);
+            } break;
         default:
             GGML_ABORT("Unknown projector type");
     }
@@ -3475,16 +3802,19 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_QWEN2VL:
         case PROJECTOR_TYPE_QWEN25VL:
         case PROJECTOR_TYPE_JANUS_PRO:
+        case PROJECTOR_TYPE_YOUTUVL:
             return ctx->model.mm_1_b->ne[0];
         case PROJECTOR_TYPE_QWEN3VL:
             // main path + deepstack paths
             return ctx->model.mm_1_b->ne[0] * (1 + ctx->model.n_deepstack_layers);
         case PROJECTOR_TYPE_GEMMA3:
+        case PROJECTOR_TYPE_GEMMA3NV:
             return ctx->model.mm_input_proj_w->ne[0];
         case PROJECTOR_TYPE_IDEFICS3:
             return ctx->model.projection->ne[1];
         case PROJECTOR_TYPE_ULTRAVOX:
         case PROJECTOR_TYPE_VOXTRAL:
+        case PROJECTOR_TYPE_MUSIC_FLAMINGO:
             return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_INTERNVL:
             return ctx->model.mm_3_w->ne[1];
@@ -3499,6 +3829,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_COGVLM:
             return ctx->model.mm_4h_to_h_w->ne[1];
+        case PROJECTOR_TYPE_LFM2A:
+            return ctx->model.position_embeddings->ne[0];
         case PROJECTOR_TYPE_GLM4V:
             return ctx->model.mm_ffn_down_w->ne[1];
         default:
@@ -3507,6 +3839,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
 }
 
 int clip_is_minicpmv(const struct clip_ctx * ctx) {
+    // TODO: remove this function
     if (ctx->proj_type() == PROJECTOR_TYPE_MINICPMV) {
         return ctx->model.hparams.minicpmv_version;
     }
@@ -3514,24 +3847,26 @@ int clip_is_minicpmv(const struct clip_ctx * ctx) {
 }
 
 bool clip_is_glm(const struct clip_ctx * ctx) {
+    // TODO: remove this function
     return ctx->proj_type() == PROJECTOR_TYPE_GLM_EDGE;
 }
 
 bool clip_is_mrope(const struct clip_ctx * ctx) {
-    return ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL
-        || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL
-        || ctx->proj_type() == PROJECTOR_TYPE_QWEN3VL
-        || ctx->proj_type() == PROJECTOR_TYPE_GLM4V;
+    switch (ctx->proj_type()) {
+        case PROJECTOR_TYPE_QWEN2VL:
+        case PROJECTOR_TYPE_QWEN25VL:
+        case PROJECTOR_TYPE_QWEN3VL:
+        case PROJECTOR_TYPE_GLM4V:
+            return true;
+        default:
+            return false;
+    }
 }
 
 bool clip_is_llava(const struct clip_ctx * ctx) {
     return ctx->model.hparams.has_llava_projector;
 }
 
-bool clip_is_gemma3(const struct clip_ctx * ctx) {
-    return ctx->proj_type() == PROJECTOR_TYPE_GEMMA3;
-}
-
 bool clip_has_vision_encoder(const struct clip_ctx * ctx) {
     return ctx->model.modality == CLIP_MODALITY_VISION;
 }
@@ -3541,10 +3876,16 @@ bool clip_has_audio_encoder(const struct clip_ctx * ctx) {
 }
 
 bool clip_has_whisper_encoder(const struct clip_ctx * ctx) {
-    return ctx->proj_type() == PROJECTOR_TYPE_ULTRAVOX
-        || ctx->proj_type() == PROJECTOR_TYPE_QWEN2A
-        || ctx->proj_type() == PROJECTOR_TYPE_GLMA
-        || ctx->proj_type() == PROJECTOR_TYPE_VOXTRAL;
+    switch (ctx->proj_type()) {
+        case PROJECTOR_TYPE_ULTRAVOX:
+        case PROJECTOR_TYPE_QWEN2A:
+        case PROJECTOR_TYPE_GLMA:
+        case PROJECTOR_TYPE_VOXTRAL:
+        case PROJECTOR_TYPE_MUSIC_FLAMINGO:
+            return true;
+        default:
+            return false;
+    }
 }
 
 bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) {

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

@@ -106,7 +106,8 @@ int clip_is_minicpmv(const struct clip_ctx * ctx);
 bool clip_is_glm(const struct clip_ctx * ctx);
 bool clip_is_mrope(const struct clip_ctx * ctx);
 bool clip_is_llava(const struct clip_ctx * ctx);
-bool clip_is_gemma3(const struct clip_ctx * ctx);
+// note for contributor: this clip_is_(model) pattern is deprecated
+//                       do NOT add new functions like this
 
 bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
 

llama/llama.cpp/tools/mtmd/models/conformer.cpp

@@ -0,0 +1,217 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_conformer::build() {
+    const int n_frames   = img.nx;
+    const int n_pos      = n_frames / 2;
+    const int n_pos_embd = (((((n_frames + 1) / 2) + 1) / 2 + 1) / 2) * 2 - 1;
+    GGML_ASSERT(model.position_embeddings->ne[1] >= n_pos);
+
+    ggml_tensor * pos_emb = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 512, n_pos_embd);
+    ggml_set_name(pos_emb, "pos_emb");
+    ggml_set_input(pos_emb);
+    ggml_build_forward_expand(gf, pos_emb);
+
+    ggml_tensor * inp = build_inp_raw(1);
+    cb(inp, "input", -1);
+
+    auto * cur = ggml_cont(ctx0, ggml_transpose(ctx0, inp));
+
+    // pre encode, conv subsampling
+    {
+        // layer.0 - conv2d
+        cur = ggml_conv_2d(ctx0, model.pre_encode_conv_X_w[0], cur, 2, 2, 1, 1, 1, 1);
+        cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[0]);
+        cb(cur, "conformer.pre_encode.conv.{}", 0);
+
+        // layer.1 - relu
+        cur = ggml_relu_inplace(ctx0, cur);
+
+        // layer.2 conv2d dw
+        cur = ggml_conv_2d_dw_direct(ctx0, model.pre_encode_conv_X_w[2], cur, 2, 2, 1, 1, 1, 1);
+        cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[2]);
+        cb(cur, "conformer.pre_encode.conv.{}", 2);
+
+        // layer.3 conv2d
+        cur = ggml_conv_2d_direct(ctx0, model.pre_encode_conv_X_w[3], cur, 1, 1, 0, 0, 1, 1);
+        cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[3]);
+        cb(cur, "conformer.pre_encode.conv.{}", 3);
+
+        // layer.4 - relu
+        cur = ggml_relu_inplace(ctx0, cur);
+
+        // layer.5 conv2d dw
+        cur = ggml_conv_2d_dw_direct(ctx0, model.pre_encode_conv_X_w[5], cur, 2, 2, 1, 1, 1, 1);
+        cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[5]);
+        cb(cur, "conformer.pre_encode.conv.{}", 5);
+
+        // layer.6 conv2d
+        cur = ggml_conv_2d_direct(ctx0, model.pre_encode_conv_X_w[6], cur, 1, 1, 0, 0, 1, 1);
+        cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[6]);
+        cb(cur, "conformer.pre_encode.conv.{}", 6);
+
+        // layer.7 - relu
+        cur = ggml_relu_inplace(ctx0, cur);
+
+        // flatten channel and frequency axis
+        cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 0, 2, 1, 3));
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0] * cur->ne[1], cur->ne[2]);
+
+        // calculate out
+        cur = ggml_mul_mat(ctx0, model.pre_encode_out_w, cur);
+        cur = ggml_add(ctx0, cur, model.pre_encode_out_b);
+        cb(cur, "conformer.pre_encode.out", -1);
+    }
+
+    // pos_emb
+    cb(pos_emb, "pos_emb", -1);
+
+    for (int il = 0; il < hparams.n_layer; il++) {
+        const auto & layer = model.layers[il];
+
+        auto * residual = cur;
+
+        cb(cur, "layer.in", il);
+
+        // feed_forward1
+        cur = build_norm(cur, layer.ff_norm_w, layer.ff_norm_b, NORM_TYPE_NORMAL, 1e-5, il);
+        cb(cur, "conformer.layers.{}.norm_feed_forward1", il);
+
+        cur = build_ffn(cur, layer.ff_up_w, layer.ff_up_b, nullptr, nullptr, layer.ff_down_w, layer.ff_down_b, FFN_SILU,
+                        il);
+        cb(cur, "conformer.layers.{}.feed_forward1.linear2", il);
+
+        const auto fc_factor = 0.5f;
+        residual             = ggml_add(ctx0, residual, ggml_scale(ctx0, cur, fc_factor));
+
+        // self-attention
+        {
+            cur = build_norm(residual, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, 1e-5, il);
+            cb(cur, "conformer.layers.{}.norm_self_att", il);
+
+            ggml_tensor * Qcur     = ggml_mul_mat(ctx0, layer.q_w, cur);
+            Qcur                   = ggml_add(ctx0, Qcur, layer.q_b);
+            Qcur                   = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, Qcur->ne[1]);
+            ggml_tensor * Q_bias_u = ggml_add(ctx0, Qcur, layer.pos_bias_u);
+            Q_bias_u               = ggml_permute(ctx0, Q_bias_u, 0, 2, 1, 3);
+            ggml_tensor * Q_bias_v = ggml_add(ctx0, Qcur, layer.pos_bias_v);
+            Q_bias_v               = ggml_permute(ctx0, Q_bias_v, 0, 2, 1, 3);
+
+            // TODO @ngxson : some cont can/should be removed when ggml_mul_mat support these cases
+            ggml_tensor * Kcur = ggml_mul_mat(ctx0, layer.k_w, cur);
+            Kcur               = ggml_add(ctx0, Kcur, layer.k_b);
+            Kcur               = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, Kcur->ne[1]);
+            Kcur               = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+
+            ggml_tensor * Vcur = ggml_mul_mat(ctx0, layer.v_w, cur);
+            Vcur               = ggml_add(ctx0, Vcur, layer.v_b);
+            Vcur               = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, Vcur->ne[1]);
+            Vcur               = ggml_cont(ctx0, ggml_permute(ctx0, Vcur, 1, 2, 0, 3));
+
+            // build_attn won't fit due to matrix_ac and matrix_bd separation
+            ggml_tensor * matrix_ac = ggml_mul_mat(ctx0, Q_bias_u, Kcur);
+            matrix_ac               = ggml_cont(ctx0, ggml_permute(ctx0, matrix_ac, 1, 0, 2, 3));
+            cb(matrix_ac, "conformer.layers.{}.self_attn.id3", il);
+
+            auto * p = ggml_mul_mat(ctx0, layer.linear_pos_w, pos_emb);
+            cb(p, "conformer.layers.{}.self_attn.linear_pos", il);
+            p = ggml_reshape_3d(ctx0, p, d_head, n_head, p->ne[1]);
+            p = ggml_permute(ctx0, p, 0, 2, 1, 3);
+
+            auto * matrix_bd = ggml_mul_mat(ctx0, Q_bias_v, p);
+            matrix_bd        = ggml_cont(ctx0, ggml_permute(ctx0, matrix_bd, 1, 0, 2, 3));
+
+            // rel shift
+            {
+                const auto pos_len = matrix_bd->ne[0];
+                const auto q_len   = matrix_bd->ne[1];
+                const auto h       = matrix_bd->ne[2];
+                matrix_bd          = ggml_pad(ctx0, matrix_bd, 1, 0, 0, 0);
+                matrix_bd          = ggml_roll(ctx0, matrix_bd, 1, 0, 0, 0);
+                matrix_bd          = ggml_reshape_3d(ctx0, matrix_bd, q_len, pos_len + 1, h);
+                matrix_bd          = ggml_view_3d(ctx0, matrix_bd, q_len, pos_len, h, matrix_bd->nb[1],
+                                                        matrix_bd->nb[2], matrix_bd->nb[0] * q_len);
+                matrix_bd          = ggml_cont_3d(ctx0, matrix_bd, pos_len, q_len, h);
+            }
+
+            matrix_bd     = ggml_view_3d(ctx0, matrix_bd, matrix_ac->ne[0], matrix_bd->ne[1],
+                                               matrix_bd->ne[2], matrix_bd->nb[1], matrix_bd->nb[2], 0);
+            auto * scores = ggml_add(ctx0, matrix_ac, matrix_bd);
+            scores        = ggml_scale(ctx0, scores, 1.0f / std::sqrt(d_head));
+            cb(scores, "conformer.layers.{}.self_attn.id0", il);
+
+            ggml_tensor * attn = ggml_soft_max(ctx0, scores);
+            ggml_tensor * x    = ggml_mul_mat(ctx0, attn, Vcur);
+            x                  = ggml_permute(ctx0, x, 2, 0, 1, 3);
+            x                  = ggml_cont_2d(ctx0, x, x->ne[0] * x->ne[1], x->ne[2]);
+
+            ggml_tensor * out = ggml_mul_mat(ctx0, layer.o_w, x);
+            out               = ggml_add(ctx0, out, layer.o_b);
+            cb(out, "conformer.layers.{}.self_attn.linear_out", il);
+
+            cur = out;
+        }
+
+        residual = ggml_add(ctx0, residual, cur);
+        cur      = build_norm(residual, layer.norm_conv_w, layer.norm_conv_b, NORM_TYPE_NORMAL, 1e-5, il);
+        cb(cur, "conformer.layers.{}.norm_conv", il);
+
+        // conv
+        {
+            auto * x = cur;
+            x = ggml_mul_mat(ctx0, layer.conv_pw1_w, x);
+            x = ggml_add(ctx0, x, layer.conv_pw1_b);
+            cb(x, "conformer.layers.{}.conv.pointwise_conv1", il);
+
+            // ggml_glu doesn't support sigmoid
+            // TODO @ngxson : support this ops in ggml
+            {
+                int64_t       d    = x->ne[0] / 2;
+                ggml_tensor * gate = ggml_sigmoid(ctx0, ggml_view_2d(ctx0, x, d, x->ne[1], x->nb[1], d * x->nb[0]));
+                x                  = ggml_mul(ctx0, ggml_view_2d(ctx0, x, d, x->ne[1], x->nb[1], 0), gate);
+                x                  = ggml_cont(ctx0, ggml_transpose(ctx0, x));
+            }
+
+            // use ggml_ssm_conv for f32 precision
+            x = ggml_pad(ctx0, x, 4, 0, 0, 0);
+            x = ggml_roll(ctx0, x, 4, 0, 0, 0);
+            x = ggml_pad(ctx0, x, 4, 0, 0, 0);
+            x = ggml_ssm_conv(ctx0, x, layer.conv_dw_w);
+            x = ggml_add(ctx0, x, layer.conv_dw_b);
+
+            x = ggml_add(ctx0, ggml_mul(ctx0, x, layer.conv_norm_w), layer.conv_norm_b);
+            x = ggml_silu(ctx0, x);
+
+            // pointwise_conv2
+            x = ggml_mul_mat(ctx0, layer.conv_pw2_w, x);
+            x = ggml_add(ctx0, x, layer.conv_pw2_b);
+
+            cur = x;
+        }
+
+        residual = ggml_add(ctx0, residual, cur);
+
+        cur = build_norm(residual, layer.ff_norm_1_w, layer.ff_norm_1_b, NORM_TYPE_NORMAL, 1e-5, il);
+        cb(cur, "conformer.layers.{}.norm_feed_forward2", il);
+
+        cur = build_ffn(cur, layer.ff_up_1_w, layer.ff_up_1_b, nullptr, nullptr, layer.ff_down_1_w, layer.ff_down_1_b,
+                        FFN_SILU, il);  // TODO(tarek): read activation for ffn from hparams
+        cb(cur, "conformer.layers.{}.feed_forward2.linear2", il);
+
+        residual = ggml_add(ctx0, residual, ggml_scale(ctx0, cur, fc_factor));
+        cb(residual, "conformer.layers.{}.conv.id", il);
+
+        cur = build_norm(residual, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, 1e-5, il);
+        cb(cur, "conformer.layers.{}.norm_out", il);
+    }
+
+    // audio adapter
+    cur = build_norm(cur, model.mm_0_w, model.mm_0_b, NORM_TYPE_NORMAL, 1e-5, -1);
+    cb(cur, "audio_adapter.model.{}", 0);
+    cur = build_ffn(cur, model.mm_1_w, model.mm_1_b, nullptr, nullptr, model.mm_3_w, model.mm_3_b, FFN_GELU_ERF, -1);
+
+    cb(cur, "projected", -1);
+
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

llama/llama.cpp/tools/mtmd/models/mobilenetv5.cpp

@@ -0,0 +1,451 @@
+#include "models.h"
+
+// Helpers for MobileNetV5 Blocks
+// RMS Norm 2D - normalizes over channels for each spatial position
+ggml_tensor * clip_graph_mobilenetv5::rms_norm_2d(ggml_tensor * inp, ggml_tensor * weight, float eps) {
+    // inp: [W, H, C, B]
+
+    ggml_tensor * cur = ggml_permute(ctx0, inp, 2, 1, 0, 3);
+    cur = ggml_cont(ctx0, cur);
+    cur = ggml_rms_norm(ctx0, cur, eps);
+
+    if (weight) {
+        cur = ggml_mul(ctx0, cur, weight);
+    }
+
+    cur = ggml_permute(ctx0, cur, 2, 1, 0, 3);
+    cur = ggml_cont(ctx0, cur);
+
+    return cur;
+}
+
+// Conv2dSame padding - asymmetric SAME padding like PyTorch/TF
+ggml_tensor* clip_graph_mobilenetv5::pad_same_2d(ggml_tensor* inp, int kernel_h, int kernel_w, int stride_h, int stride_w, int dilation_h, int dilation_w) {
+    const int64_t ih = inp->ne[1];  // height
+    const int64_t iw = inp->ne[0];  // width
+
+    // Calculate output size (ceil division)
+    const int64_t oh = (ih + stride_h - 1) / stride_h;
+    const int64_t ow = (iw + stride_w - 1) / stride_w;
+
+    // Calculate padding needed
+    const int64_t pad_h = std::max((int64_t)0, (oh - 1) * stride_h + (kernel_h - 1) * dilation_h + 1 - ih);
+    const int64_t pad_w = std::max((int64_t)0, (ow - 1) * stride_w + (kernel_w - 1) * dilation_w + 1 - iw);
+
+    // Split padding asymmetrically
+    const int pad_h_top = pad_h / 2;
+    const int pad_h_bottom = pad_h - pad_h_top;
+    const int pad_w_left = pad_w / 2;
+    const int pad_w_right = pad_w - pad_w_left;
+
+    // Apply padding if needed
+    // ggml_pad_ext: (ctx, tensor, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3)
+    // For [W, H, C, B]: p0=width, p1=height, p2=channels, p3=batch
+    if (pad_h > 0 || pad_w > 0) {
+        inp = ggml_pad_ext(ctx0, inp,
+            pad_w_left, pad_w_right,     // width padding (dim 0)
+            pad_h_top, pad_h_bottom,      // height padding (dim 1)
+            0, 0,                         // no channel padding (dim 2)
+            0, 0);                        // no batch padding (dim 3)
+    }
+
+    return inp;
+}
+
+
+// Edge Residual Block (Stage 0)
+ggml_tensor * clip_graph_mobilenetv5::build_edge_residual(ggml_tensor * inp, const mobilenetv5_block & block, int stride) {
+    ggml_tensor * cur = inp;
+
+    // 1. Expansion Conv (3x3)
+    if (stride == 2) {
+        // Case: Downsampling (Block 0)
+        // Replicates Conv2dSame(kernel=3, stride=2)
+        cur = pad_same_2d(cur, 3, 3, stride, stride);
+        cur = ggml_conv_2d_direct(ctx0, block.s0_conv_exp_w, cur, stride, stride, 0, 0, 1, 1);
+    } else {
+        // Case: Normal 3x3 Block (Block 1, 2)
+        // Replicates Conv2d(kernel=3, stride=1, padding=1)
+        cur = ggml_conv_2d_direct(ctx0, block.s0_conv_exp_w, cur, stride, stride, 1, 1, 1, 1);
+    }
+
+    // BN + Activation
+    if (block.s0_bn1_w) cur = rms_norm_2d(cur, block.s0_bn1_w);
+    cur = ggml_gelu(ctx0, cur);
+
+    // 2. Pointwise Linear Conv (1x1)
+    // 1x1 Convs usually have padding=0 and stride=1
+    cur = ggml_conv_2d_direct(ctx0, block.s0_conv_pwl_w, cur, 1, 1, 0, 0, 1, 1);
+    if (block.s0_bn2_w) cur = rms_norm_2d(cur, block.s0_bn2_w);
+
+    // 3. Residual Connection
+    // Only apply residual if spatial dimensions and channels match (stride 1)
+    if (stride == 1 && inp->ne[2] == cur->ne[2] && inp->ne[0] == cur->ne[0]) {
+        cur = ggml_add(ctx0, cur, inp);
+    }
+
+    return cur;
+}
+
+// Universal Inverted Residual Block (Stage 1+)
+ggml_tensor * clip_graph_mobilenetv5::build_inverted_residual(ggml_tensor * inp, const mobilenetv5_block & block, int stride) {
+    ggml_tensor * cur = inp;
+
+    // 1. Depthwise Start (Optional)
+    // NOTE: dw_start always has stride=1 (no downsampling here)
+    if (block.dw_start_w) {
+        int k = block.dw_start_w->ne[0]; // 3 or 5
+        int p = k / 2;
+        cur = ggml_conv_2d_dw(ctx0, block.dw_start_w, cur, 1, 1, p, p, 1, 1);
+        if (block.dw_start_bn_w) cur = rms_norm_2d(cur, block.dw_start_bn_w);
+    }
+
+    // 2. Pointwise Expansion (1x1)
+    if (block.pw_exp_w) {
+        // Standard 1x1 conv, pad=0, stride=1
+        cur = ggml_conv_2d_direct(ctx0, block.pw_exp_w, cur, 1, 1, 0, 0, 1, 1);
+        if (block.pw_exp_bn_w) cur = rms_norm_2d(cur, block.pw_exp_bn_w);
+        cur = ggml_gelu(ctx0, cur);
+    }
+
+    // 3. Depthwise Mid (Optional)
+    // NOTE: dw_mid is where downsampling happens (stride=2 for first block of stage)
+    if (block.dw_mid_w) {
+        int k = block.dw_mid_w->ne[0]; // 3 or 5
+
+        if (stride > 1) {
+            // Case: Stride 2 (Downsample) -> Use Asymmetric "Same" Padding
+            cur = pad_same_2d(cur, k, k, stride, stride);
+            cur = ggml_conv_2d_dw(ctx0, block.dw_mid_w, cur, stride, stride, 0, 0, 1, 1); // pad=0
+        } else {
+            // Case: Stride 1 -> Use Standard Symmetric Padding
+            int p = k / 2;
+            cur = ggml_conv_2d_dw(ctx0, block.dw_mid_w, cur, stride, stride, p, p, 1, 1);
+        }
+
+        if (block.dw_mid_bn_w) cur = rms_norm_2d(cur, block.dw_mid_bn_w);
+        cur = ggml_gelu(ctx0, cur);
+    }
+
+    // 4. Pointwise Projection (1x1)
+    if (block.pw_proj_w) {
+        cur = ggml_conv_2d_direct(ctx0, block.pw_proj_w, cur, 1, 1, 0, 0, 1, 1);
+        if (block.pw_proj_bn_w) cur = rms_norm_2d(cur, block.pw_proj_bn_w);
+    }
+
+    // Apply Layer Scaling if present
+    if (block.layer_scale_w) {
+        cur = ggml_mul(ctx0, cur, block.layer_scale_w);
+    }
+
+    // 5. Residual Connection
+    bool same_spatial = (inp->ne[0] == cur->ne[0]) && (inp->ne[1] == cur->ne[1]);
+    bool same_channel = (inp->ne[2] == cur->ne[2]);
+    if (same_spatial && same_channel) {
+        cur = ggml_add(ctx0, cur, inp);
+    }
+
+    return cur;
+}
+
+// Attention Block (MQA)
+ggml_tensor * clip_graph_mobilenetv5::build_mobilenet_attn(ggml_tensor * inp, const mobilenetv5_block & block) {
+    ggml_tensor * cur = inp;
+
+    // Norm
+    if (block.attn_norm_w) {
+        cur = rms_norm_2d(cur, block.attn_norm_w, 1e-6f);
+    }
+
+    // 1. Q Calculation
+    ggml_tensor * q = ggml_conv_2d_direct(ctx0, block.attn_q_w, cur, 1, 1, 0, 0, 1, 1);
+
+    // 2. K Calculation (Downsampled)
+    // Uses Conv2dSame(640, 640, kernel_size=(3, 3), stride=(2, 2), groups=640)
+    ggml_tensor * k_inp = cur;
+    if (block.attn_k_dw_w) {
+        int k_size = block.attn_k_dw_w->ne[0];  // Usually 3
+        k_inp = pad_same_2d(cur, k_size, k_size, 2, 2);  // Apply SAME padding
+        k_inp = ggml_conv_2d_dw(ctx0, block.attn_k_dw_w, k_inp, 2, 2, 0, 0, 1, 1);  // padding=0
+        if (block.attn_k_norm_w) {
+            k_inp = rms_norm_2d(k_inp, block.attn_k_norm_w, 1e-6f);
+        }
+    }
+    ggml_tensor * k = ggml_conv_2d_direct(ctx0, block.attn_k_w, k_inp, 1, 1, 0, 0, 1, 1);
+
+    // 3. V Calculation (Downsampled)
+    // Uses Conv2dSame(640, 640, kernel_size=(3, 3), stride=(2, 2), groups=640)
+    ggml_tensor * v_inp = cur;
+    if (block.attn_v_dw_w) {
+        int v_size = block.attn_v_dw_w->ne[0];  // Usually 3
+        v_inp = pad_same_2d(cur, v_size, v_size, 2, 2);  // Apply SAME padding
+        v_inp = ggml_conv_2d_dw(ctx0, block.attn_v_dw_w, v_inp, 2, 2, 0, 0, 1, 1);  // padding=0
+        if (block.attn_v_norm_w) {
+            v_inp = rms_norm_2d(v_inp, block.attn_v_norm_w, 1e-6f);
+        }
+    }
+    ggml_tensor * v = ggml_conv_2d_direct(ctx0, block.attn_v_w, v_inp, 1, 1, 0, 0, 1, 1);
+
+    const int W = cur->ne[0]; const int H = cur->ne[1]; const int B = cur->ne[3];
+    const int D = k->ne[2]; // Head dimension
+    const int n_head = q->ne[2] / D;
+    const int N = W * H;
+
+    // Process Q: [W, H, D*n_head, B] -> [D, N, n_head, B]
+    q = ggml_reshape_3d(ctx0, q, N, D*n_head, B);
+    q = ggml_reshape_4d(ctx0, q, N, D, n_head, B);
+    q = ggml_permute(ctx0, q, 1, 0, 2, 3); // [D, N, n_head, B]
+    q = ggml_cont(ctx0, q);
+
+    const int Wk = k->ne[0]; const int Hk = k->ne[1];
+    const int M = Wk * Hk;
+
+    // Process K: [Wk, Hk, D, B] -> [D, M, 1, B]
+    k = ggml_reshape_3d(ctx0, k, M, D, B);
+    k = ggml_reshape_4d(ctx0, k, M, D, 1, B);
+    k = ggml_permute(ctx0, k, 1, 0, 2, 3); // [D, M, 1, B]
+    k = ggml_cont(ctx0, k);
+
+    // Process V: [Wk, Hk, D, B] -> [M, D, 1, B]
+    v = ggml_reshape_3d(ctx0, v, M, D, B);
+    v = ggml_reshape_4d(ctx0, v, M, D, 1, B);
+    v = ggml_cont(ctx0, v); // [M, D, 1, B]
+
+    // Multi-Query Attention
+    float scale = 1.0f / sqrtf((float)D);
+
+    // Step 1: Compute Q @ K.T
+    ggml_tensor * scores = ggml_mul_mat(ctx0, k, q);
+
+    scores = ggml_scale(ctx0, scores, scale);
+
+    scores = ggml_soft_max(ctx0, scores);
+
+    ggml_tensor * kqv = ggml_mul_mat(ctx0, v, scores);
+
+    kqv = ggml_permute(ctx0, kqv, 1, 0, 2, 3);
+    kqv = ggml_cont(ctx0, kqv);
+
+
+    kqv = ggml_reshape_3d(ctx0, kqv, N, D * n_head, B);
+    kqv = ggml_reshape_4d(ctx0, kqv, W, H, D * n_head, B);
+    kqv = ggml_cont(ctx0, kqv);
+
+    // Output projection
+    cur = ggml_conv_2d_direct(ctx0, block.attn_o_w, kqv, 1, 1, 0, 0, 1, 1);
+
+    // Residual & Layer Scale
+    if (inp->ne[0] == cur->ne[0] && inp->ne[2] == cur->ne[2]) {
+        if (block.layer_scale_w) {
+            cur = ggml_mul(ctx0, cur, block.layer_scale_w);
+        }
+        cur = ggml_add(ctx0, cur, inp);
+    }
+
+    return cur;
+}
+
+ggml_cgraph * clip_graph_mobilenetv5::build() {
+    ggml_tensor * inp = build_inp_raw();
+
+    // 1. Stem - Conv2dSame(3, 64, kernel_size=(3, 3), stride=(2, 2))
+    ggml_tensor * cur = pad_same_2d(inp, 3, 3, 2, 2);  // Apply SAME padding
+
+    cur = ggml_conv_2d_direct(ctx0, model.mobilenet_stem_conv_w, cur, 2, 2, 0, 0, 1, 1);  // padding=0
+    if (model.mobilenet_stem_conv_b) {
+        cur = ggml_add(ctx0, cur, model.mobilenet_stem_conv_b);
+    }
+    if (model.mobilenet_stem_norm_w) cur = rms_norm_2d(cur, model.mobilenet_stem_norm_w);
+    cur = ggml_gelu(ctx0, cur);
+
+
+    // 2. Blocks
+    std::vector<ggml_tensor*> intermediate_features;
+    const int total_blocks = model.mobilenet_blocks.size();
+
+    auto is_stage_start = [&](int i) {
+        if (i == 0) return true;
+        for (int end_idx : model.mobilenet_stage_ends) {
+            if (i == end_idx + 1) return true;
+        }
+        return false;
+    };
+
+    auto is_fusion_point = [&](int i) {
+        if (model.mobilenet_stage_ends.size() >= 4) {
+                if (i == model.mobilenet_stage_ends[2]) return true; // End of Stage 2
+                if (i == model.mobilenet_stage_ends[3]) return true; // End of Stage 3
+        } else {
+            if (i == total_blocks - 1) return true;
+        }
+        return false;
+    };
+
+    for (int i = 0; i < total_blocks; i++) {
+        const auto & block = model.mobilenet_blocks[i];
+        int stride = is_stage_start(i) ? 2 : 1;
+
+        if (block.s0_conv_exp_w)      cur = build_edge_residual(cur, block, stride);
+        else if (block.attn_q_w)      cur = build_mobilenet_attn(cur, block);
+        else                          cur = build_inverted_residual(cur, block, stride);
+
+        if (is_fusion_point(i)) {
+
+            intermediate_features.push_back(cur);
+        }
+    }
+
+    // 3. Multi-Scale Fusion Adapter (MSFA)
+    if (!intermediate_features.empty()) {
+
+        // A. Reference Resolution: PyTorch implementation uses inputs[0]
+        // We assume intermediate_features[0] is the "High Resolution" target.
+        // In MobileNet designs, this is typically the feature map with the smallest stride (e.g. 32x32).
+        ggml_tensor* target_feat = intermediate_features[0];
+        int high_res_w = target_feat->ne[0];
+        int high_res_h = target_feat->ne[1];
+
+        std::vector<ggml_tensor*> resized_feats;
+
+        // B. Resize inputs to match inputs[0] (High Resolution)
+        for (auto feat : intermediate_features) {
+            int feat_w = feat->ne[0];
+            int feat_h = feat->ne[1];
+
+            // PyTorch: if feat_size < high_resolution: interpolate
+            if (feat_w < high_res_w || feat_h < high_res_h) {
+                // Calculate scale factor.
+                // Note: PyTorch 'nearest' works on arbitrary float scales.
+                // ggml_upscale generally takes integer factors or target sizes depending on helper.
+                // Assuming standard power-of-2 scaling (e.g. 16 -> 32 means scale=2).
+                int scale_w = high_res_w / feat_w;
+                // int scale_h = high_res_h / feat_h;
+
+                // Safety check for non-integer scaling if strictly replicating
+                GGML_ASSERT(high_res_w % feat_w == 0);
+
+                // Upsample (Nearest Neighbor)
+                // 2 is the scale factor
+                feat = ggml_upscale(ctx0, feat, scale_w, ggml_scale_mode::GGML_SCALE_MODE_NEAREST);
+            }
+            resized_feats.push_back(feat);
+        }
+
+        // C. Concatenate at High Resolution (Channel Dim = 2 in ggml)
+        cur = resized_feats[0];
+        for (size_t k = 1; k < resized_feats.size(); ++k) {
+            cur = ggml_concat(ctx0, cur, resized_feats[k], 2);
+        }
+
+        // D. FFN (UniversalInvertedResidual)
+        // Structure: Expand Conv -> Norm -> GELU -> Project Conv -> Norm
+
+        // 1. Expansion
+        if (model.msfa_ffn_expand_w) {
+            // 1x1 Conv
+            cur = ggml_conv_2d_direct(ctx0, model.msfa_ffn_expand_w, cur, 1, 1, 0, 0, 1, 1);
+
+            if (model.msfa_ffn_expand_bn) {
+                cur = rms_norm_2d(cur, model.msfa_ffn_expand_bn);
+            }
+
+            cur = ggml_gelu(ctx0, cur);
+
+        }
+
+        // 2. Projection (No DW because kernel_size=0)
+        if (model.msfa_ffn_project_w) {
+            // 1x1 Conv
+            cur = ggml_conv_2d_direct(ctx0, model.msfa_ffn_project_w, cur, 1, 1, 0, 0, 1, 1);
+
+            // UniversalInvertedResidual typically has a norm after projection
+            if (model.msfa_ffn_project_bn) {
+                cur = rms_norm_2d(cur, model.msfa_ffn_project_bn);
+            }
+
+        }
+
+        // E. Final Downsample to Target Resolution (Output Resolution)
+        // PyTorch: matches self.output_resolution (e.g. 16x16)
+        const int target_out_res = 16;
+        int current_w = cur->ne[0];
+
+        if (current_w > target_out_res) {
+            int s = current_w / target_out_res;
+
+            GGML_ASSERT(current_w % target_out_res == 0);
+
+            // Avg Pool: Kernel=s, Stride=s
+            cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_AVG, s, s, s, s, 0, 0);
+
+        }
+
+        // F. Final Norm
+        if (model.msfa_concat_norm_w) {
+            cur = rms_norm_2d(cur, model.msfa_concat_norm_w);
+
+        }
+    }
+
+    // 4. Gemma 3n Multimodal Projection (Embedder)
+    // Input: 'cur' is [Width, Height, Channels, Batch]
+    int W = cur->ne[0];
+    int H = cur->ne[1];
+    int C = cur->ne[2];
+    int B = cur->ne[3];
+
+    GGML_ASSERT(C == hparams.n_embd);
+
+    // 1. Permute and Flatten to [Channels, Tokens, Batch]
+    // PyTorch expects (Batch, Seq, Hidden), GGML usually processes (Hidden, Seq, Batch)
+    cur = ggml_permute(ctx0, cur, 2, 1, 0, 3); // -> [C, H, W, B]
+    cur = ggml_permute(ctx0, cur, 0, 2, 1, 3); // -> [C, W, H, B]
+    cur = ggml_cont(ctx0, cur);
+    cur = ggml_reshape_3d(ctx0, cur, C, W*H, B);
+    cur = ggml_cont(ctx0, cur);
+
+
+    // 2. FEATURE SCALING
+    // PyTorch: vision_outputs *= self.config.vision_config.hidden_size**0.5
+    const float scale_factor = sqrtf((float)C);
+    cur = ggml_scale(ctx0, cur, scale_factor);
+
+
+    // 3. SOFT EMBEDDING NORM
+    // PyTorch: self._norm(x) * self.weight
+    // We must normalize regardless, then multiply if weight exists.
+    {
+        const float eps = 1e-6f; // Gemma3n uses 1e-6
+        cur = ggml_rms_norm(ctx0, cur, eps);
+
+        if (model.mm_soft_emb_norm_w) {
+            // Weight shape is (2048,) -> Element-wise broadcast multiply
+            cur = ggml_mul(ctx0, cur, model.mm_soft_emb_norm_w);
+        }
+
+    }
+
+    // 4. PROJECTION
+    // PyTorch: embedding_projection = nn.Linear(vision_hidden, text_hidden, bias=False)
+    // Weight stored as [out_features, in_features] = [text_hidden_size, vision_hidden_size]
+    if (model.mm_input_proj_w) {
+        cur = ggml_mul_mat(ctx0, model.mm_input_proj_w, cur);
+    }
+
+    // 5. POST PROJECTION NORM
+    // PyTorch: embedding_post_projection_norm = Gemma3nRMSNorm(..., with_scale=False)
+    // with_scale=False means weight is registered as buffer with value 1.0
+    // So output = rms_norm(x) * 1.0 = rms_norm(x), magnitude ~1
+    {
+        const float eps = 1e-6f;
+        cur = ggml_rms_norm(ctx0, cur, eps);
+
+        if (model.mm_post_proj_norm_w) {
+            // If weight is loaded, multiply (should be ~1.0 anyway)
+            cur = ggml_mul(ctx0, cur, model.mm_post_proj_norm_w);
+        }
+    }
+
+    ggml_build_forward_expand(gf, cur);
+    return gf;
+}

llama/llama.cpp/tools/mtmd/models/models.h

@@ -2,6 +2,11 @@
 
 #include "../clip-graph.h"
 
+/*
+ * IMPORTANT: The mtmd module does NOT accept pull requests that are fully or predominantly AI-generated.
+ * We encourage human contributors to ensure the quality and reliability of the codebase.
+ */
+
 struct clip_graph_siglip : clip_graph {
     clip_graph_siglip(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;
@@ -22,6 +27,11 @@ struct clip_graph_qwen3vl : clip_graph {
     ggml_cgraph * build() override;
 };
 
+struct clip_graph_youtuvl : clip_graph {
+    clip_graph_youtuvl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
 struct clip_graph_minicpmv : clip_graph {
     clip_graph_minicpmv(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;
@@ -57,7 +67,45 @@ struct clip_graph_whisper_enc : clip_graph {
     ggml_cgraph * build() override;
 };
 
+struct clip_graph_conformer : clip_graph {
+    clip_graph_conformer(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
 struct clip_graph_glm4v : clip_graph {
     clip_graph_glm4v(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;
 };
+
+struct clip_graph_mobilenetv5 : clip_graph {
+    clip_graph_mobilenetv5(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+
+    ggml_tensor * rms_norm_2d(
+        ggml_tensor * inp,
+        ggml_tensor * weight,
+        float eps = 1e-6f);
+
+    ggml_tensor* pad_same_2d(
+        ggml_tensor* inp,
+        int kernel_h,
+        int kernel_w,
+        int stride_h,
+        int stride_w,
+        int dilation_h = 1,
+        int dilation_w = 1);
+
+    ggml_tensor * build_edge_residual(
+        ggml_tensor * inp,
+        const mobilenetv5_block & block,
+        int stride);
+
+    ggml_tensor * build_inverted_residual(
+        ggml_tensor * inp,
+        const mobilenetv5_block & block,
+        int stride);
+
+    ggml_tensor * build_mobilenet_attn(
+        ggml_tensor * inp,
+        const mobilenetv5_block & block);
+};

llama/llama.cpp/tools/mtmd/models/siglip.cpp

@@ -50,10 +50,15 @@ ggml_cgraph * clip_graph_siglip::build() {
         const int scale_factor = model.hparams.n_merge;
         cur = build_patch_merge_permute(cur, scale_factor);
 
-        // projection
-        cur = ggml_norm(ctx0, cur, 1e-5); // default nn.LayerNorm
-        cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);
-        cur = ggml_add(ctx0, cur, model.mm_input_norm_b);
+        // projection, in LFM2-VL input norm is optional
+        if (model.mm_input_norm_w) {
+            cur = ggml_norm(ctx0, cur, 1e-5); // default nn.LayerNorm
+            cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);
+        }
+
+        if (model.mm_input_norm_b) {
+            cur = ggml_add(ctx0, cur, model.mm_input_norm_b);
+        }
 
         cur = build_ffn(cur,
             model.mm_1_w, model.mm_1_b,

llama/llama.cpp/tools/mtmd/models/whisper-enc.cpp

@@ -86,6 +86,15 @@ ggml_cgraph * clip_graph_whisper_enc::build() {
             FFN_GELU_ERF,
             -1);
 
+    } else if (proj_type == PROJECTOR_TYPE_MUSIC_FLAMINGO) {
+        // projector
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU_ERF,
+            -1);
+
     } else if (proj_type == PROJECTOR_TYPE_GLMA) {
             cur = ggml_norm(ctx0, cur, hparams.eps);
             cur = ggml_mul(ctx0, cur, model.mm_norm_pre_w);

llama/llama.cpp/tools/mtmd/models/youtuvl.cpp

@@ -0,0 +1,179 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_youtuvl::build() {
+    GGML_ASSERT(model.class_embedding == nullptr);
+    const int batch_size       = 1;
+    const bool use_window_attn = !hparams.wa_layer_indexes.empty();
+    const int n_pos            = n_patches;
+    const int num_position_ids = n_pos * 4;
+    const int m = 2;
+    const int Wp = n_patches_x;
+    const int Hp = n_patches_y;
+    const int Hm = Hp / m;
+    const int Wm = Wp / m;
+    norm_type norm_t = NORM_TYPE_NORMAL;
+
+    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
+
+    ggml_tensor * inp = build_inp_raw();
+
+    // change conv3d to linear
+    // reshape and permute to get patches, permute from (patch_size, m, Wm, patch_size, m, Hm, C) to (C, patch_size, patch_size, m, m, Wm, Hm)
+    {
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            Wm * m * patch_size, m * patch_size, Hm, 3);
+        inp = ggml_permute(ctx0, inp, 1, 2, 3, 0);
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            m * patch_size * 3, Wm, m * patch_size, Hm);
+
+        inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            m * patch_size * 3, patch_size, m, Hm * Wm);
+
+        inp = ggml_permute(ctx0, inp, 1, 0, 2, 3);
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            patch_size, 3, patch_size, Hm * Wm * m * m);
+
+        inp = ggml_permute(ctx0, inp, 2, 0, 1, 3);
+        inp = ggml_cont_3d(
+            ctx0, inp,
+            3*patch_size* patch_size,  Hm * Wm * m * m, 1);
+    }
+    inp = ggml_mul_mat(ctx0, model.patch_embeddings_0, inp);
+
+    if (model.patch_bias) {
+        inp = ggml_add(ctx0, inp, model.patch_bias);
+    }
+
+    inp = ggml_reshape_2d(ctx0, inp, n_embd, n_patches);
+
+    ggml_tensor * inpL           = inp;
+    ggml_tensor * window_mask    = nullptr;
+    ggml_tensor * window_idx     = nullptr;
+    ggml_tensor * inv_window_idx = nullptr;
+
+    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    // pre-layernorm
+    if (model.pre_ln_w) {
+        inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, norm_t, eps, -1);
+    }
+    if (use_window_attn) {
+        inv_window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4);
+        ggml_set_name(inv_window_idx, "inv_window_idx");
+        ggml_set_input(inv_window_idx);
+        // mask for window attention
+        window_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos);
+        ggml_set_name(window_mask, "window_mask");
+        ggml_set_input(window_mask);
+
+        // if flash attn is used, we need to pad the mask and cast to f16
+        if (flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) {
+            window_mask = ggml_cast(ctx0, window_mask, GGML_TYPE_F16);
+        }
+
+        // inpL shape: [n_embd, n_patches_x * n_patches_y, batch_size]
+        GGML_ASSERT(batch_size == 1);
+        inpL = ggml_reshape_2d(ctx0, inpL, n_embd * 4, n_patches_x * n_patches_y * batch_size / 4);
+        inpL = ggml_get_rows(ctx0, inpL, inv_window_idx);
+        inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_patches_x * n_patches_y, batch_size);
+    }
+
+    // loop over layers
+    for (int il = 0; il < n_layer; il++) {
+        const auto & layer = model.layers[il];
+        const bool full_attn = use_window_attn ? hparams.wa_layer_indexes.count(il) > 0 : true;
+
+        ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states
+
+        // layernorm1
+        cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il);
+        // self-attention
+        {
+            ggml_tensor * Qcur = ggml_add(ctx0,
+                ggml_mul_mat(ctx0, layer.q_w, cur), layer.q_b);
+            ggml_tensor * Kcur = ggml_add(ctx0,
+                ggml_mul_mat(ctx0, layer.k_w, cur), layer.k_b);
+            ggml_tensor * Vcur = ggml_add(ctx0,
+                ggml_mul_mat(ctx0, layer.v_w, cur), layer.v_b);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_patches);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_patches);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_patches);
+
+            Qcur = ggml_rope_multi(
+                ctx0, Qcur, positions, nullptr,
+                d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+            Kcur = ggml_rope_multi(
+                ctx0, Kcur, positions, nullptr,
+                d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1);
+
+            ggml_tensor * attn_mask = full_attn ? nullptr : window_mask;
+
+            cur = build_attn(layer.o_w, layer.o_b,
+                Qcur, Kcur, Vcur, attn_mask, kq_scale, il);
+        }
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, inpL);
+
+        inpL = cur; // inpL = residual, cur = hidden_states
+
+        // layernorm2
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il);
+
+        // ffn
+        cur = build_ffn(cur,
+            layer.ff_up_w, layer.ff_up_b,
+            nullptr, nullptr,
+            layer.ff_down_w, layer.ff_down_b,
+            hparams.ffn_op, il);
+
+        // residual 2
+        cur = ggml_add(ctx0, inpL, cur);
+
+        inpL = cur;
+    }
+
+    ggml_tensor * embeddings = inpL;
+    if (use_window_attn) {
+        const int spatial_merge_unit = 4;
+        window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / spatial_merge_unit);
+        ggml_set_name(window_idx, "window_idx");
+        ggml_set_input(window_idx);
+        GGML_ASSERT(batch_size == 1);
+        embeddings = ggml_reshape_2d(ctx0, embeddings, n_embd * spatial_merge_unit, n_patches / spatial_merge_unit);
+        embeddings = ggml_get_rows(ctx0, embeddings, window_idx);
+        embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd, n_patches, batch_size);
+        cb(embeddings, "window_order_restored", -1);
+    }
+
+    // post-layernorm (part of Siglip2VisionTransformer, applied after encoder)
+    if (model.post_ln_w) {
+        embeddings = build_norm(embeddings, model.post_ln_w, model.post_ln_b, norm_t, eps, n_layer);
+    }
+
+    // Now apply merger (VLPatchMerger):
+    // 1. Apply RMS norm (ln_q in VLPatchMerger)
+    embeddings = build_norm(embeddings, model.mm_input_norm_w, nullptr, NORM_TYPE_RMS, 1e-6, -1);
+    cb(embeddings, "merger_normed", -1);
+
+    // 2. First reshape for spatial merge (merge 2x2 patches)
+    embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size);
+    cb(embeddings, "merger_reshaped", -1);
+
+    embeddings = build_ffn(embeddings,
+                    model.mm_0_w, model.mm_0_b,
+                    nullptr, nullptr,
+                    model.mm_1_w, model.mm_1_b,
+                    FFN_GELU,
+                    -1);
+    ggml_build_forward_expand(gf, embeddings);
+
+    return gf;
+}

llama/llama.cpp/tools/mtmd/mtmd-audio.cpp

@@ -9,207 +9,250 @@
 #include <fstream>
 #include <algorithm>
 
-// most of the code here is copied from whisper.cpp
+// some of the code here is copied from whisper.cpp
 
 constexpr bool DEBUG = false;
 
-struct mtmd_audio_mel_filters {
-    int32_t n_mel;
-    int32_t n_fft;
+void mtmd_audio_cache::fill_sin_cos_table(int n) {
+    sin_vals.resize(n);
+    cos_vals.resize(n);
+    for (int i = 0; i < n; i++) {
+        double theta = (2 * M_PI * i) / n;
+        sin_vals[i]  = sinf(theta);
+        cos_vals[i]  = cosf(theta);
+    }
+}
 
-    std::vector<float> data;
-};
+void mtmd_audio_cache::fill_hann_window(int length, bool periodic) {
+    hann_window.resize(length);
+    int offset = -1;
+    if (periodic) {
+        offset = 0;
+    }
+    for (int i = 0; i < length; i++) {
+        hann_window[i] = 0.5 * (1.0 - cosf((2.0 * M_PI * i) / (length + offset)));
+    }
+}
 
-// note: this global cache is shared among all preprocessors
-//       if we want to use multiple preprocessors at the same time,
-//       we will need to enclose it in the preprocessor class in the future
-static struct mtmd_audio_global_cache {
-    // precomputed sin/cos table for FFT
-    std::vector<float> sin_vals;
-    std::vector<float> cos_vals;
-
-    // hann window
-    std::vector<float> hann_window;
-
-    // mel filter bank
-    mtmd_audio_mel_filters filters;
-
-    void fill_sin_cos_table(int n) {
-        sin_vals.resize(n);
-        cos_vals.resize(n);
-        for (int i = 0; i < n; i++) {
-            double theta = (2 * M_PI * i) / n;
-            sin_vals[i] = sinf(theta);
-            cos_vals[i] = cosf(theta);
-        }
+void mtmd_audio_cache::fill_mel_filterbank_matrix(int   n_mel,
+                                                  int   n_fft,
+                                                  int   sample_rate,
+                                                  float fmin,
+                                                  float fmax,
+                                                  bool  slaney_area_norm,
+                                                  float scale) {
+    GGML_ASSERT(n_mel > 0 && n_fft > 1);
+    if (fmax <= 0.0f) {
+        fmax = 0.5f * sample_rate;
     }
 
-    void fill_hann_window(int length, bool periodic) {
-        hann_window.resize(length);
-        int offset = -1;
-        if (periodic) {
-            offset = 0;
-        }
-        for (int i = 0; i < length; i++) {
-            hann_window[i] = 0.5 * (1.0 - cosf((2.0 * M_PI * i) / (length + offset)));
-        }
+    // Slaney scale (matches librosa default)
+    const double min_log_hz  = 1000.0;
+    const double lin_slope   = 3 / 200.;
+    const double min_log_mel = min_log_hz * lin_slope;
+    const double log_step    = log(6.4) / 27.0;
+    auto         hz_to_mel   = [min_log_hz, lin_slope, log_step, min_log_mel](const double f_hz) -> double {
+        return (f_hz < min_log_hz) ? f_hz * lin_slope : min_log_mel + log(f_hz / min_log_hz) / log_step;
+    };
+    auto mel_to_hz = [min_log_hz, lin_slope, log_step, min_log_mel](const double m) -> double {
+        return (m < min_log_mel) ? m / lin_slope : min_log_hz * exp((m - min_log_mel) * log_step);
+    };
+
+    // infer N_fft from n_fft_bins
+    const double bin_hz_step = double(sample_rate) / double(n_fft);
+
+    // mel grid: n_mel + 2 edges
+    const double        m_lo = hz_to_mel(fmin);
+    const double        m_hi = hz_to_mel(fmax);
+    std::vector<double> mel_pts(n_mel + 2);
+    for (int i = 0; i < n_mel + 2; ++i) {
+        mel_pts[i] = m_lo + (m_hi - m_lo) * (double(i) / (n_mel + 1));
     }
 
-    // Build mel filterbank matrix [n_mel × n_fft_bins] at runtime.
-    // n_fft_bins must be (N_fft / 2 + 1). Example: if N_fft=512 -> n_fft_bins=257.
-    void fill_mel_filterbank_matrix(
-        int n_mel,
-        int n_fft,
-        int sample_rate,            // e.g. 16000
-        float fmin = 0.0f,          // e.g. 0.0
-        float fmax = -1.0f,         // e.g. sr/2; pass -1 for auto
-        bool slaney_area_norm = true,
-        float scale = 1.0f          // optional extra scaling; use 1.0f/1000.0f to mimic your code
-    ) {
-        GGML_ASSERT(n_mel > 0 && n_fft > 1);
-        if (fmax <= 0.0f) {
-            fmax = 0.5f * sample_rate;
-        }
+    // convert to Hz
+    std::vector<double> hz_pts(n_mel + 2);
+    for (int i = 0; i < n_mel + 2; ++i) {
+        hz_pts[i] = mel_to_hz(mel_pts[i]);
+    }
 
-        // Slaney scale (matches librosa default)
-        const double min_log_hz = 1000.0;
-        const double lin_slope = 3 / 200.;
-        const double min_log_mel = min_log_hz * lin_slope;
-        const double log_step = log(6.4) / 27.0;
-        auto hz_to_mel = [min_log_hz, lin_slope, log_step, min_log_mel](const double f_hz) -> double {
-            return (f_hz < min_log_hz) ? f_hz * lin_slope : min_log_mel + log(f_hz / min_log_hz) / log_step;
-        };
-        auto mel_to_hz = [min_log_hz, lin_slope, log_step, min_log_mel](const double m) -> double {
-            return (m < min_log_mel) ? m / lin_slope : min_log_hz * exp((m - min_log_mel) * log_step);
-        };
+    const int n_fft_bins = n_fft / 2 + 1;
 
-        // infer N_fft from n_fft_bins
-        const double bin_hz_step = double(sample_rate) / double(n_fft);
+    // filterbank
+    std::vector<float> out(n_mel * n_fft_bins, 0);
+    for (int m = 0; m < n_mel; ++m) {
+        const double f_left   = hz_pts[m];
+        const double f_center = hz_pts[m + 1];
+        const double f_right  = hz_pts[m + 2];
 
-        // mel grid: n_mel + 2 edges
-        const double m_lo = hz_to_mel(fmin);
-        const double m_hi = hz_to_mel(fmax);
-        std::vector<double> mel_pts(n_mel + 2);
-        for (int i = 0; i < n_mel + 2; ++i) {
-            mel_pts[i] = m_lo + (m_hi - m_lo) * (double(i) / (n_mel + 1));
-        }
+        const double denom_l = std::max(1e-30, f_center - f_left);
+        const double denom_r = std::max(1e-30, f_right - f_center);
+        const double enorm   = slaney_area_norm ? (2.0 / std::max(1e-30, f_right - f_left)) : 1.0;
 
-        // convert to Hz
-        std::vector<double> hz_pts(n_mel + 2);
-        for (int i = 0; i < n_mel + 2; ++i) {
-            hz_pts[i] = mel_to_hz(mel_pts[i]);
-        }
-
-        const int n_fft_bins = n_fft / 2 + 1;
-
-        // filterbank
-        std::vector<float> out(n_mel * n_fft_bins, 0);
-        for (int m = 0; m < n_mel; ++m) {
-            const double f_left   = hz_pts[m];
-            const double f_center = hz_pts[m + 1];
-            const double f_right  = hz_pts[m + 2];
-
-            const double denom_l = std::max(1e-30, f_center - f_left);
-            const double denom_r = std::max(1e-30, f_right  - f_center);
-            const double enorm   = slaney_area_norm ? (2.0 / std::max(1e-30, f_right - f_left)) : 1.0;
-
-            for (int k = 0; k < n_fft_bins; ++k) {
-                const double f = k * bin_hz_step;
-                double w = 0.0;
-                if (f >= f_left && f <= f_center) {
-                    w = (f - f_left) / denom_l;
-                } else if (f > f_center && f <= f_right) {
-                    w = (f_right - f) / denom_r;
-                }
-                out[size_t(m) * size_t(n_fft_bins) + size_t(k)] = float(w * enorm * scale);
+        for (int k = 0; k < n_fft_bins; ++k) {
+            const double f = k * bin_hz_step;
+            double       w = 0.0;
+            if (f >= f_left && f <= f_center) {
+                w = (f - f_left) / denom_l;
+            } else if (f > f_center && f <= f_right) {
+                w = (f_right - f) / denom_r;
             }
+            out[size_t(m) * size_t(n_fft_bins) + size_t(k)] = float(w * enorm * scale);
         }
+    }
 
-        filters.n_mel = n_mel;
-        filters.n_fft = n_fft;
-        filters.data  = std::move(out);
+    filters.n_mel = n_mel;
+    filters.n_fft = n_fft;
+    filters.data  = std::move(out);
 
-        if (DEBUG) { // debug
-            for (size_t i = 0; i < filters.data.size(); ++i) {
-                if (filters.data[i] != 0.0f) {
-                    printf("filters[%zu] = %f\n", i, filters.data[i] * 1000.0f);
-                }
+    if (DEBUG) {  // debug
+        for (size_t i = 0; i < filters.data.size(); ++i) {
+            if (filters.data[i] != 0.0f) {
+                printf("filters[%zu] = %f\n", i, filters.data[i] * 1000.0f);
             }
         }
     }
-} g_cache;
+}
 
-// naive Discrete Fourier Transform
-// input is real-valued
-// output is complex-valued
-static void dft(const float * in, int N, float * out) {
-    const int n_sin_cos_vals = g_cache.sin_vals.size();
-    const int sin_cos_step = n_sin_cos_vals / N;
+// Unified DFT implementation for both forward and inverse transforms
+// Template parameters:
+//   Inverse: false = DFT with exp(-2πi·k·n/N), no scaling
+//            true  = IDFT with exp(+2πi·k·n/N), scales by 1/N
+//   RealInput: true = input is real-valued (stride 1), avoids imaginary computations
+//              false = input is complex-valued (interleaved real/imag, stride 2)
+template <bool Inverse, bool RealInput>
+static void dft_impl(const mtmd_audio_cache & cache, const float * in, int N, float * out) {
+    const int n_sin_cos_vals = cache.sin_vals.size();
+    const int sin_cos_step   = n_sin_cos_vals / N;
+
+    constexpr float sign  = Inverse ? 1.0f : -1.0f;
+    const float     scale = Inverse ? (1.0f / N) : 1.0f;
 
     for (int k = 0; k < N; k++) {
         float re = 0;
         float im = 0;
 
         for (int n = 0; n < N; n++) {
-            int idx = (k * n * sin_cos_step) % (n_sin_cos_vals); // t = 2*M_PI*k*n/N
-            re += in[n] * g_cache.cos_vals[idx]; // cos(t)
-            im -= in[n] * g_cache.sin_vals[idx]; // sin(t)
+            int   idx     = (k * n * sin_cos_step) % n_sin_cos_vals;
+            float cos_val = cache.cos_vals[idx];
+            float sin_val = cache.sin_vals[idx];
+
+            if constexpr (RealInput) {
+                // Real input: in_im = 0, simplifies to:
+                // re += in_re * cos_val
+                // im += sign * in_re * sin_val
+                float in_re = in[n];
+                re += in_re * cos_val;
+                im += sign * in_re * sin_val;
+            } else {
+                float in_re = in[n * 2 + 0];
+                float in_im = in[n * 2 + 1];
+                // (a + bi) * (cos + sign*i*sin) = (a*cos - sign*b*sin) + (sign*a*sin + b*cos)i
+                re += in_re * cos_val - sign * in_im * sin_val;
+                im += sign * in_re * sin_val + in_im * cos_val;
+            }
         }
 
-        out[k*2 + 0] = re;
-        out[k*2 + 1] = im;
+        out[k * 2 + 0] = re * scale;
+        out[k * 2 + 1] = im * scale;
     }
 }
 
-// Cooley-Tukey FFT
-// poor man's implementation - use something better
-// input is real-valued
-// output is complex-valued
-static void fft(float * in, int N, float * out) {
-    const int n_sin_cos_vals = g_cache.sin_vals.size();
+// Cooley-Tukey FFT/IFFT unified implementation
+// Template parameters:
+//   Inverse: false = FFT with exp(-2πi·k/N), no scaling
+//            true  = IFFT with exp(+2πi·k/N), scales by 0.5 at each level
+//   RealInput: true = input is real-valued (stride 1)
+//              false = input is complex-valued (interleaved real/imag, stride 2)
+template <bool Inverse, bool RealInput>
+static void fft_impl(const mtmd_audio_cache & cache, float * in, int N, float * out) {
+    const int n_sin_cos_vals = cache.sin_vals.size();
+
     if (N == 1) {
         out[0] = in[0];
-        out[1] = 0;
+        if constexpr (RealInput) {
+            out[1] = 0.0f;
+        } else {
+            out[1] = in[1];
+        }
         return;
     }
 
     const int half_N = N / 2;
-    if (N - half_N*2 == 1) {
-        dft(in, N, out);
+    if (N - half_N * 2 == 1) {
+        // Odd N: fall back to DFT
+        dft_impl<Inverse, RealInput>(cache, in, N, out);
         return;
     }
 
-    float* even = in + N;
-    for (int i = 0; i < half_N; ++i) {
-        even[i]= in[2*i];
-    }
-    float* even_fft = out + 2 * N;
-    fft(even, half_N, even_fft);
+    // Split into even and odd
+    if constexpr (RealInput) {
+        // Real input: stride is 1, copy only real values
+        float * even = in + N;
+        for (int i = 0; i < half_N; ++i) {
+            even[i] = in[2 * i];
+        }
+        float * even_fft = out + 2 * N;
+        fft_impl<Inverse, true>(cache, even, half_N, even_fft);
 
-    float* odd = even;
-    for (int i = 0; i < half_N; ++i) {
-        odd[i] = in[2*i + 1];
+        float * odd = even;
+        for (int i = 0; i < half_N; ++i) {
+            odd[i] = in[2 * i + 1];
+        }
+        float * odd_fft = even_fft + N;
+        fft_impl<Inverse, true>(cache, odd, half_N, odd_fft);
+    } else {
+        // Complex input: stride is 2, copy complex pairs
+        float * even = in + N * 2;
+        for (int i = 0; i < half_N; ++i) {
+            even[i * 2 + 0] = in[2 * i * 2 + 0];
+            even[i * 2 + 1] = in[2 * i * 2 + 1];
+        }
+        float * even_fft = out + 2 * N;
+        fft_impl<Inverse, false>(cache, even, half_N, even_fft);
+
+        float * odd = even;
+        for (int i = 0; i < half_N; ++i) {
+            odd[i * 2 + 0] = in[(2 * i + 1) * 2 + 0];
+            odd[i * 2 + 1] = in[(2 * i + 1) * 2 + 1];
+        }
+        float * odd_fft = even_fft + N;
+        fft_impl<Inverse, false>(cache, odd, half_N, odd_fft);
     }
-    float* odd_fft = even_fft + N;
-    fft(odd, half_N, odd_fft);
+
+    float * even_fft = out + 2 * N;
+    float * odd_fft  = even_fft + N;
 
     const int sin_cos_step = n_sin_cos_vals / N;
+
+    constexpr float sign  = Inverse ? 1.0f : -1.0f;
+    constexpr float scale = Inverse ? 0.5f : 1.0f;
+
     for (int k = 0; k < half_N; k++) {
-        int idx = k * sin_cos_step; // t = 2*M_PI*k/N
-        float re =  g_cache.cos_vals[idx]; // cos(t)
-        float im = -g_cache.sin_vals[idx]; // sin(t)
+        int   idx = k * sin_cos_step;  // t = 2*M_PI*k/N
+        float re  = cache.cos_vals[idx];
+        float im  = sign * cache.sin_vals[idx];
 
-        float re_odd = odd_fft[2*k + 0];
-        float im_odd = odd_fft[2*k + 1];
+        float re_odd = odd_fft[2 * k + 0];
+        float im_odd = odd_fft[2 * k + 1];
 
-        out[2*k + 0] = even_fft[2*k + 0] + re*re_odd - im*im_odd;
-        out[2*k + 1] = even_fft[2*k + 1] + re*im_odd + im*re_odd;
+        out[2 * k + 0] = scale * (even_fft[2 * k + 0] + re * re_odd - im * im_odd);
+        out[2 * k + 1] = scale * (even_fft[2 * k + 1] + re * im_odd + im * re_odd);
 
-        out[2*(k + half_N) + 0] = even_fft[2*k + 0] - re*re_odd + im*im_odd;
-        out[2*(k + half_N) + 1] = even_fft[2*k + 1] - re*im_odd - im*re_odd;
+        out[2 * (k + half_N) + 0] = scale * (even_fft[2 * k + 0] - re * re_odd + im * im_odd);
+        out[2 * (k + half_N) + 1] = scale * (even_fft[2 * k + 1] - re * im_odd - im * re_odd);
     }
 }
 
+// Forward FFT for real input (used by mel spectrogram)
+static void fft(const mtmd_audio_cache & cache, float * in, int N, float * out) {
+    fft_impl<false, true>(cache, in, N, out);
+}
+
+// Inverse FFT for complex input
+static void ifft(const mtmd_audio_cache & cache, float * in, int N, float * out) {
+    fft_impl<true, false>(cache, in, N, out);
+}
+
 struct filter_params {
     int32_t n_mel;
     int32_t n_fft_bins;
@@ -222,20 +265,27 @@ struct filter_params {
     bool    norm_per_feature = false;
 };
 
-static void log_mel_spectrogram_worker_thread(int ith, const float * hann, const std::vector<float> & samples,
-                                              int n_samples, int frame_size, int frame_step, int n_threads,
-                                              const filter_params & params, mtmd_audio_mel & out) {
+static void log_mel_spectrogram_worker_thread(int                        ith,
+                                              const float *              hann,
+                                              const std::vector<float> & samples,
+                                              int                        n_samples,
+                                              int                        frame_size,
+                                              int                        frame_step,
+                                              int                        n_threads,
+                                              const filter_params &      params,
+                                              const mtmd_audio_cache &   cache,
+                                              mtmd_audio_mel &           out) {
     std::vector<float> fft_in(frame_size * 2, 0.0);
     std::vector<float> fft_out(frame_size * 2 * 2 * 2);
 
     int n_fft_bins = params.n_fft_bins;
     int i = ith;
 
-    const auto & filters = g_cache.filters;
+    const auto & filters = cache.filters;
 
     // make sure n_fft == 1 + (WHISPER_N_FFT / 2), bin_0 to bin_nyquist
     GGML_ASSERT(n_fft_bins == 1 + (frame_size / 2));
-    GGML_ASSERT(g_cache.sin_vals.size() == g_cache.cos_vals.size());
+    GGML_ASSERT(cache.sin_vals.size() == cache.cos_vals.size());
     // calculate FFT only when fft_in are not all zero
     for (; i < std::min(n_samples / frame_step + 1, out.n_len); i += n_threads) {
         const int offset = i * frame_step;
@@ -251,7 +301,7 @@ static void log_mel_spectrogram_worker_thread(int ith, const float * hann, const
         }
 
         // FFT
-        fft(fft_in.data(), frame_size, fft_out.data());
+        fft(cache, fft_in.data(), frame_size, fft_out.data());
 
         // Calculate modulus^2 of complex numbers
         // Use pow(fft_out[2 * j + 0], 2) + pow(fft_out[2 * j + 1], 2) causes inference quality problem? Interesting.
@@ -298,6 +348,7 @@ static bool log_mel_spectrogram(
         const int     n_samples_in,
         const int     n_threads,
         const filter_params & params,
+        const mtmd_audio_cache & cache,
         mtmd_audio_mel & out) {
     //const int64_t t_start_us = ggml_time_us();
 
@@ -305,9 +356,9 @@ static bool log_mel_spectrogram(
     int n_samples = n_samples_in;
 
     // Hann window
-    const float * hann = g_cache.hann_window.data();
-    const int frame_size = (params.n_fft_bins - 1) * 2;
-    const int frame_step = params.hop_length;
+    const float * hann       = cache.hann_window.data();
+    const int     frame_size = (params.n_fft_bins - 1) * 2;
+    const int     frame_step = params.hop_length;
 
     // Padding
     std::vector<float> samples_padded;
@@ -335,9 +386,9 @@ static bool log_mel_spectrogram(
 
     // preemphasis
     if (params.preemph) {
-        const int pad_amount = frame_size / 2;
+        const int   pad_amount = frame_size / 2;
         const float preemph = 0.97f;
-        float prev = samples_padded[pad_amount];
+        float       prev = samples_padded[pad_amount];
         for (int i = pad_amount + 1; i + pad_amount < n_samples; ++i) {
             float cur = samples_padded[i];
             samples_padded[i] = cur - preemph * prev;
@@ -372,14 +423,14 @@ static bool log_mel_spectrogram(
     {
         std::vector<std::thread> workers(n_threads - 1);
         for (int iw = 0; iw < n_threads - 1; ++iw) {
-            workers[iw] = std::thread(
-                    log_mel_spectrogram_worker_thread, iw + 1, hann, std::cref(samples_padded),
-                    n_samples, frame_size, frame_step, n_threads,
-                    std::cref(params), std::ref(out));
+            workers[iw] =
+                std::thread(log_mel_spectrogram_worker_thread, iw + 1, hann, std::cref(samples_padded), n_samples,
+                            frame_size, frame_step, n_threads, std::cref(params), std::cref(cache), std::ref(out));
         }
 
         // main thread
-        log_mel_spectrogram_worker_thread(0, hann, samples_padded, n_samples, frame_size, frame_step, n_threads, params, out);
+        log_mel_spectrogram_worker_thread(0, hann, samples_padded, n_samples, frame_size, frame_step, n_threads, params,
+                                          cache, out);
         for (int iw = 0; iw < n_threads - 1; ++iw) {
             workers[iw].join();
         }
@@ -404,7 +455,7 @@ static bool log_mel_spectrogram(
 
             for (int j = 0; j < effective_n_len; ++j) {
                 auto &value = out.data[i * out.n_len + j];
-                value = (value - mean) / mstd;
+                value        = (value - mean) / mstd;
             }
 
             // pad the rest with zeros
@@ -450,18 +501,14 @@ static bool log_mel_spectrogram(
 //
 
 void mtmd_audio_preprocessor_whisper::initialize() {
-    g_cache.fill_sin_cos_table(hparams.audio_n_fft);
-    g_cache.fill_hann_window(hparams.audio_window_len, true);
-    g_cache.fill_mel_filterbank_matrix(
-        hparams.n_mel_bins,
-        hparams.audio_n_fft,
-        hparams.audio_sample_rate);
+    cache.fill_sin_cos_table(hparams.audio_n_fft);
+    cache.fill_hann_window(hparams.audio_window_len, true);
+    cache.fill_mel_filterbank_matrix(hparams.n_mel_bins, hparams.audio_n_fft, hparams.audio_sample_rate);
 }
 
-bool mtmd_audio_preprocessor_whisper::preprocess(
-        const float * samples,
-        size_t n_samples,
-        std::vector<mtmd_audio_mel> & output) {
+bool mtmd_audio_preprocessor_whisper::preprocess(const float *                 samples,
+                                                 size_t                        n_samples,
+                                                 std::vector<mtmd_audio_mel> & output) {
     if (n_samples == 0) {
         // empty audio
         return false;
@@ -471,7 +518,7 @@ bool mtmd_audio_preprocessor_whisper::preprocess(
     // if input is too short, pad with zeros
     // this is to avoid potential issues with stage1/2 padding in log_mel_spectrogram
     // TODO: maybe handle this better
-    size_t min_samples = (size_t)hparams.audio_sample_rate * (hparams.audio_chunk_len + 1); // +1 second margin
+    size_t min_samples = (size_t) hparams.audio_sample_rate * (hparams.audio_chunk_len + 1);  // +1 second margin
     if (n_samples < min_samples) {
         smpl.resize(min_samples, 0.0f);
         std::memcpy(smpl.data(), samples, n_samples * sizeof(float));
@@ -486,22 +533,19 @@ bool mtmd_audio_preprocessor_whisper::preprocess(
     params.hop_length       = hparams.audio_hop_len;
     params.sample_rate      = hparams.audio_sample_rate;
     params.center_padding   = false;
-    params.preemph          = 0.0f; // disabled
+    params.preemph          = 0.0f;  // disabled
     params.use_natural_log  = false;
     params.norm_per_feature = false;
 
-    // make sure the global cache is initialized
-    GGML_ASSERT(!g_cache.sin_vals.empty());
-    GGML_ASSERT(!g_cache.cos_vals.empty());
-    GGML_ASSERT(!g_cache.filters.data.empty());
+    // make sure the cache is initialized
+    GGML_ASSERT(!cache.sin_vals.empty());
+    GGML_ASSERT(!cache.cos_vals.empty());
+    GGML_ASSERT(!cache.filters.data.empty());
 
     mtmd_audio_mel out_full;
-    bool ok = log_mel_spectrogram(
-                samples,
-                n_samples,
-                4, // n_threads
-                params,
-                out_full);
+    bool           ok = log_mel_spectrogram(samples, n_samples,
+                                            4,  // n_threads
+                                            params, cache, out_full);
     if (!ok) {
         return false;
     }
@@ -512,21 +556,21 @@ bool mtmd_audio_preprocessor_whisper::preprocess(
         printf("output: n_mel = %d, n_len = %d\n", out_full.n_mel, out_full.n_len);
     }
     const size_t frames_per_chunk = 3000;
-    GGML_ASSERT((size_t)out_full.n_len > frames_per_chunk);
-    for (size_t off = 0; off < (size_t)out_full.n_len; off += frames_per_chunk) {
-        int n_len = std::min(frames_per_chunk, (size_t)out_full.n_len - off);
-        if ((size_t)n_len < frames_per_chunk) {
-            break; // last uncomplete chunk will always be a padded chunk, safe to ignore
+    GGML_ASSERT((size_t) out_full.n_len > frames_per_chunk);
+    for (size_t off = 0; off < (size_t) out_full.n_len; off += frames_per_chunk) {
+        int n_len = std::min(frames_per_chunk, (size_t) out_full.n_len - off);
+        if ((size_t) n_len < frames_per_chunk) {
+            break;  // last uncomplete chunk will always be a padded chunk, safe to ignore
         }
 
         mtmd_audio_mel out_chunk;
         out_chunk.n_len     = n_len;
         out_chunk.n_mel     = out_full.n_mel;
-        out_chunk.n_len_org = out_full.n_mel; // unused
+        out_chunk.n_len_org = out_full.n_mel;  // unused
         out_chunk.data.reserve(out_chunk.n_mel * out_chunk.n_len);
 
         for (int i = 0; i < out_full.n_mel; i++) {
-            auto src = out_full.data.begin() + i*out_full.n_len + off;
+            auto src = out_full.data.begin() + i * out_full.n_len + off;
             out_chunk.data.insert(out_chunk.data.end(), src, src + frames_per_chunk);
         }
 
@@ -535,3 +579,152 @@ bool mtmd_audio_preprocessor_whisper::preprocess(
 
     return true;
 }
+
+//
+// mtmd_audio_preprocessor_conformer
+//
+
+void mtmd_audio_preprocessor_conformer::initialize() {
+    cache.fill_sin_cos_table(hparams.audio_n_fft);
+    cache.fill_hann_window(hparams.audio_window_len, true);
+    cache.fill_mel_filterbank_matrix(hparams.n_mel_bins, hparams.audio_n_fft, hparams.audio_sample_rate);
+}
+
+bool mtmd_audio_preprocessor_conformer::preprocess(const float *                 samples,
+                                                   size_t                        n_samples,
+                                                   std::vector<mtmd_audio_mel> & output) {
+    // empty audio
+    if (n_samples == 0) {
+        return false;
+    }
+
+    filter_params params;
+    params.n_mel            = hparams.n_mel_bins;
+    params.n_fft_bins       = 1 + (hparams.audio_n_fft / 2);
+    params.hann_window_size = hparams.audio_window_len;
+    params.hop_length       = hparams.audio_hop_len;
+    params.sample_rate      = hparams.audio_sample_rate;
+    params.center_padding   = true;
+    params.preemph          = 0.97f;
+    params.use_natural_log  = true;
+    params.norm_per_feature = true;
+
+    // make sure the cache is initialized
+    GGML_ASSERT(!cache.sin_vals.empty());
+    GGML_ASSERT(!cache.cos_vals.empty());
+    GGML_ASSERT(!cache.filters.data.empty());
+
+    mtmd_audio_mel out_full;
+    bool           ok = log_mel_spectrogram(samples, n_samples,
+                                            4,  // n_threads
+                                            params, cache, out_full);
+    if (!ok) {
+        return false;
+    }
+
+    output.push_back(std::move(out_full));
+    return true;
+}
+
+//
+// mtmd_audio_streaming_istft implementation
+//
+
+mtmd_audio_streaming_istft::mtmd_audio_streaming_istft(int n_fft, int hop_length) :
+    n_fft(n_fft),
+    hop_length(hop_length),
+    n_fft_bins(n_fft / 2 + 1),
+    overlap_buffer(n_fft, 0.0f),
+    window_sum_buffer(n_fft, 0.0f),
+    padding_to_remove((n_fft - hop_length) / 2),
+    ifft_in(n_fft * 2 * 4, 0.0f),  // extra space for recursive IFFT
+    ifft_out(n_fft * 2 * 4, 0.0f) {
+    cache.fill_sin_cos_table(n_fft);
+    cache.fill_hann_window(n_fft, true);
+}
+
+void mtmd_audio_streaming_istft::reset() {
+    std::fill(overlap_buffer.begin(), overlap_buffer.end(), 0.0f);
+    std::fill(window_sum_buffer.begin(), window_sum_buffer.end(), 0.0f);
+    padding_to_remove = (n_fft - hop_length) / 2;
+}
+
+std::vector<float> mtmd_audio_streaming_istft::process_frame(const float * frame_spectrum) {
+    std::vector<float> output(hop_length);
+
+    // copy frequencies
+    for (int j = 0; j < n_fft_bins; j++) {
+        ifft_in[j * 2 + 0] = frame_spectrum[j * 2 + 0];
+        ifft_in[j * 2 + 1] = frame_spectrum[j * 2 + 1];
+    }
+
+    // mirror negative frequencies
+    for (int j = 1; j < n_fft_bins - 1; j++) {
+        int mirror_idx              = n_fft - j;
+        ifft_in[mirror_idx * 2 + 0] = ifft_in[j * 2 + 0];
+        ifft_in[mirror_idx * 2 + 1] = -ifft_in[j * 2 + 1];  // conjugate
+    }
+
+    ifft(cache, ifft_in.data(), n_fft, ifft_out.data());
+
+    // update window sum and overlap buffer
+    for (int j = 0; j < n_fft; j++) {
+        window_sum_buffer[j] += cache.hann_window[j] * cache.hann_window[j];
+        overlap_buffer[j] += ifft_out[j * 2] * cache.hann_window[j];
+    }
+
+    // extract hop_length samples with normalization
+    for (int i = 0; i < hop_length; i++) {
+        if (window_sum_buffer[i] > 1e-8f) {
+            output[i] = overlap_buffer[i] / window_sum_buffer[i];
+        } else {
+            output[i] = overlap_buffer[i];
+        }
+    }
+
+    // shift buffers left by hop_length
+    std::copy(overlap_buffer.begin() + hop_length, overlap_buffer.end(), overlap_buffer.begin());
+    std::fill(overlap_buffer.end() - hop_length, overlap_buffer.end(), 0.0f);
+
+    std::copy(window_sum_buffer.begin() + hop_length, window_sum_buffer.end(), window_sum_buffer.begin());
+    std::fill(window_sum_buffer.end() - hop_length, window_sum_buffer.end(), 0.0f);
+
+    // Remove padding if needed
+    int to_remove = std::min(padding_to_remove, (int) output.size());
+    padding_to_remove -= to_remove;
+    output.erase(output.begin(), output.begin() + to_remove);
+
+    return output;
+}
+
+std::vector<float> mtmd_audio_streaming_istft::flush() {
+    std::vector<float> output;
+
+    // Extract remaining samples from overlap buffer
+    // Continue until we've extracted all meaningful samples
+    int remaining = n_fft - hop_length;
+    while (remaining > 0) {
+        int chunk_size = std::min(remaining, hop_length);
+
+        for (int i = 0; i < chunk_size; i++) {
+            float sample;
+            if (window_sum_buffer[i] > 1e-8f) {
+                sample = overlap_buffer[i] / window_sum_buffer[i];
+            } else {
+                sample = overlap_buffer[i];
+            }
+            output.push_back(sample);
+        }
+
+        // Shift buffers
+        std::copy(overlap_buffer.begin() + chunk_size, overlap_buffer.end(), overlap_buffer.begin());
+        std::fill(overlap_buffer.end() - chunk_size, overlap_buffer.end(), 0.0f);
+
+        std::copy(window_sum_buffer.begin() + chunk_size, window_sum_buffer.end(), window_sum_buffer.begin());
+        std::fill(window_sum_buffer.end() - chunk_size, window_sum_buffer.end(), 0.0f);
+
+        remaining -= chunk_size;
+    }
+
+    return output;
+}

llama/llama.cpp/tools/mtmd/mtmd-audio.h

@@ -17,6 +17,38 @@ struct mtmd_audio_mel {
     std::vector<float> data;
 };
 
+struct mtmd_audio_mel_filters {
+    int32_t n_mel;
+    int32_t n_fft;
+
+    std::vector<float> data;
+};
+
+// cache for audio processing, each processor instance owns its own cache
+struct mtmd_audio_cache {
+    std::vector<float> sin_vals;
+    std::vector<float> cos_vals;
+
+    std::vector<float> hann_window;
+
+    mtmd_audio_mel_filters filters;
+
+    void fill_sin_cos_table(int n);
+
+    void fill_hann_window(int length, bool periodic);
+
+    // Build mel filterbank matrix [n_mel × n_fft_bins] at runtime.
+    // n_fft_bins must be (N_fft / 2 + 1). Example: if N_fft=512 -> n_fft_bins=257.
+    void fill_mel_filterbank_matrix(int   n_mel,
+                                    int   n_fft,
+                                    int   sample_rate,               // e.g. 16000
+                                    float fmin             = 0.0f,   // e.g. 0.0
+                                    float fmax             = -1.0f,  // e.g. sr/2; pass -1 for auto
+                                    bool  slaney_area_norm = true,
+                                    float scale = 1.0f  // optional extra scaling
+    );
+};
+
 struct mtmd_audio_preprocessor {
     const clip_hparams & hparams;
 
@@ -31,4 +63,51 @@ struct mtmd_audio_preprocessor_whisper : mtmd_audio_preprocessor {
     mtmd_audio_preprocessor_whisper(const clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) {}
     void initialize() override;
     bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) override;
+
+  private:
+    mtmd_audio_cache cache;
+};
+
+struct mtmd_audio_preprocessor_conformer : mtmd_audio_preprocessor {
+    mtmd_audio_preprocessor_conformer(const clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) {}
+    void initialize() override;
+    bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) override;
+
+  private:
+    mtmd_audio_cache cache;
+};
+
+//
+// streaming ISTFT - converts spectrogram frames back to audio one frame at a time
+//
+struct mtmd_audio_streaming_istft {
+    mtmd_audio_streaming_istft(int n_fft, int hop_length);
+
+    // reset streaming state
+    void reset();
+
+    // process a single STFT frame (streaming)
+    // frame_spectrum: [n_fft_bins x 2] interleaved real/imag
+    // returns: up to hop_length samples
+    std::vector<float> process_frame(const float * frame_spectrum);
+
+    // flush remaining samples at end of stream
+    std::vector<float> flush();
+
+  private:
+    int n_fft;
+    int hop_length;
+    int n_fft_bins;
+
+    // Own cache for output processing
+    mtmd_audio_cache cache;
+
+    // Streaming state
+    std::vector<float> overlap_buffer;
+    std::vector<float> window_sum_buffer;
+    int                padding_to_remove;
+
+    // Working buffers for IFFT
+    std::vector<float> ifft_in;
+    std::vector<float> ifft_out;
 };

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

@@ -276,7 +276,7 @@ struct mtmd_context {
         }
 
         // set boi/eoi
-        if (proj == PROJECTOR_TYPE_GEMMA3) {
+        if (proj == PROJECTOR_TYPE_GEMMA3 || proj == PROJECTOR_TYPE_GEMMA3NV) {
             // <start_of_image> ... (image embeddings) ... <end_of_image>
             img_beg = "<start_of_image>";
             img_end = "<end_of_image>";
@@ -293,7 +293,7 @@ struct mtmd_context {
             // https://github.com/huggingface/transformers/blob/1cd110c6cb6a6237614130c470e9a902dbc1a4bd/docs/source/en/model_doc/pixtral.md
             img_end = "[IMG_END]";
 
-        } else if (proj == PROJECTOR_TYPE_QWEN2VL || proj == PROJECTOR_TYPE_QWEN25VL || proj == PROJECTOR_TYPE_QWEN3VL) {
+        } else if (proj == PROJECTOR_TYPE_QWEN2VL || proj == PROJECTOR_TYPE_QWEN25VL || proj == PROJECTOR_TYPE_QWEN3VL || proj == PROJECTOR_TYPE_YOUTUVL) {
             // <|vision_start|> ... (image embeddings) ... <|vision_end|>
             img_beg = "<|vision_start|>";
             img_end = "<|vision_end|>";
@@ -339,8 +339,13 @@ struct mtmd_context {
             case PROJECTOR_TYPE_QWEN25O:
             case PROJECTOR_TYPE_ULTRAVOX:
             case PROJECTOR_TYPE_VOXTRAL:
+            case PROJECTOR_TYPE_GLMA:
+            case PROJECTOR_TYPE_MUSIC_FLAMINGO:
                 audio_preproc = std::make_unique<mtmd_audio_preprocessor_whisper>(ctx_a);
                 break;
+            case PROJECTOR_TYPE_LFM2A:
+                audio_preproc = std::make_unique<mtmd_audio_preprocessor_conformer>(ctx_a);
+                break;
             default:
                 GGML_ABORT("unsupported audio projector type");
         }
@@ -358,6 +363,9 @@ struct mtmd_context {
             // [BEGIN_AUDIO] ... (embeddings) ...
             aud_beg = "[BEGIN_AUDIO]";
 
+        } else if (proj == PROJECTOR_TYPE_MUSIC_FLAMINGO) {
+            // <sound> ... (embeddings) ...
+            aud_beg = "<sound>";
         }
     }
 
@@ -864,10 +872,15 @@ float * mtmd_get_output_embd(mtmd_context * ctx) {
 }
 
 bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
-    if (ctx->ctx_v && clip_get_projector_type(ctx->ctx_v) == PROJECTOR_TYPE_GEMMA3) {
-        return true;
+    switch (ctx->proj_type_v()) {
+        case PROJECTOR_TYPE_QWEN2VL:
+        case PROJECTOR_TYPE_QWEN25VL:
+        case PROJECTOR_TYPE_QWEN3VL:
+        case PROJECTOR_TYPE_YOUTUVL:
+            return true;
+        default:
+            return false;
     }
-    return false;
 }
 
 bool mtmd_decode_use_mrope(mtmd_context * ctx) {

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

@@ -27,6 +27,9 @@
  * - Make sure the C API is aligned with the libllama C API (as in llama.h)
  * - Do not include model name (e.g., qwen, gemma) in the API, use generic terms instead
  * - Keep the API minimal, do not expose internal details unless necessary
+ *
+ * IMPORTANT: The mtmd module does NOT accept pull requests that are fully or predominantly AI-generated.
+ * We encourage human contributors to ensure the quality and reliability of the codebase.
  */
 
 #ifdef LLAMA_SHARED