GGML update to ec98e2002 (#13451)

* Revert "add support for NVIDIA Nemotron 3 Nano"

This reverts commit e7d2ae9d69.

* GGML update to 380b4c984

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

* update to c45f89d55

* ec98e2002

solar pro needed more adjusting - needs verification

* review comments

CMakeLists.txt

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

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggml-org/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=17f7f4baad8b3a716ee139da7bb56ae984e8c0fa
+FETCH_HEAD=ec98e2002
 
 .PHONY: help
 help:

api/types.go

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

api/types_test.go

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

app/ui/ui.go

@@ -12,13 +12,13 @@ import (
 	"log/slog"
 	"net/http"
 	"net/http/httputil"
-	"net/url"
 	"os"
 	"runtime"
 	"runtime/debug"
 	"slices"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/google/uuid"
@@ -117,40 +117,66 @@ func (s *Server) log() *slog.Logger {
 
 // ollamaProxy creates a reverse proxy handler to the Ollama server
 func (s *Server) ollamaProxy() http.Handler {
-	ollamaHost := os.Getenv("OLLAMA_HOST")
+	var (
-	if ollamaHost == "" {
+		proxy   http.Handler
-		ollamaHost = "http://127.0.0.1:11434"
+		proxyMu sync.Mutex
-	}
+	)
 
-	if !strings.HasPrefix(ollamaHost, "http://") && !strings.HasPrefix(ollamaHost, "https://") {
-		ollamaHost = "http://" + ollamaHost
-	}
-
-	target, err := url.Parse(ollamaHost)
-	if err != nil {
-		s.log().Error("failed to parse OLLAMA_HOST", "error", err, "host", ollamaHost)
 	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			http.Error(w, "failed to configure proxy", http.StatusInternalServerError)
+		proxyMu.Lock()
-		})
+		p := proxy
+		proxyMu.Unlock()
+
+		if p == nil {
+			proxyMu.Lock()
+			if proxy == nil {
+				var err error
+				for i := range 2 {
+					if i > 0 {
+						s.log().Warn("ollama server not ready, retrying", "attempt", i+1)
+						time.Sleep(1 * time.Second)
 					}
 
+					err = WaitForServer(context.Background(), 10*time.Second)
+					if err == nil {
+						break
+					}
+				}
+
+				if err != nil {
+					proxyMu.Unlock()
+					s.log().Error("ollama server not ready after retries", "error", err)
+					http.Error(w, "Ollama server is not ready", http.StatusServiceUnavailable)
+					return
+				}
+
+				target := envconfig.Host()
 				s.log().Info("configuring ollama proxy", "target", target.String())
 
-	proxy := httputil.NewSingleHostReverseProxy(target)
+				newProxy := httputil.NewSingleHostReverseProxy(target)
 
-	originalDirector := proxy.Director
+				originalDirector := newProxy.Director
-	proxy.Director = func(req *http.Request) {
+				newProxy.Director = func(req *http.Request) {
 					originalDirector(req)
 					req.Host = target.Host
 					s.log().Debug("proxying request", "method", req.Method, "path", req.URL.Path, "target", target.Host)
 				}
 
-	proxy.ErrorHandler = func(w http.ResponseWriter, r *http.Request, err error) {
+				newProxy.ErrorHandler = func(w http.ResponseWriter, r *http.Request, err error) {
 					s.log().Error("proxy error", "error", err, "path", r.URL.Path, "target", target.String())
 					http.Error(w, "proxy error: "+err.Error(), http.StatusBadGateway)
 				}
 
-	return proxy
+				proxy = newProxy
+				p = newProxy
+			} else {
+				p = proxy
+			}
+			proxyMu.Unlock()
+		}
+
+		p.ServeHTTP(w, r)
+	})
 }
 
 type errHandlerFunc func(http.ResponseWriter, *http.Request) error

convert/convert.go

@@ -202,6 +202,8 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &qwen25VLModel{}
 	case "Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration":
 		conv = &qwen3VLModel{}
+	case "Olmo3ForCausalLM":
+		conv = &olmoModel{}
 	case "BertModel":
 		conv = &bertModel{}
 	case "NomicBertModel", "NomicBertMoEModel":

convert/convert_olmo.go

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

envconfig/config.go

@@ -199,7 +199,7 @@ var (
 	// MultiUserCache optimizes prompt caching for multi-user scenarios
 	MultiUserCache = Bool("OLLAMA_MULTIUSER_CACHE")
 	// Enable the new Ollama engine
-	NewEngine = BoolWithDefault("OLLAMA_NEW_ENGINE")
+	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 	// ContextLength sets the default context length
 	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
 	// Auth enables authentication between the Ollama client and server
@@ -291,7 +291,7 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
 		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4096)"},
-		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(true), "Enable the new Ollama engine"},
+		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
 		"OLLAMA_REMOTES":           {"OLLAMA_REMOTES", Remotes(), "Allowed hosts for remote models (default \"ollama.com\")"},
 
 		// Informational

fs/ggml/ggml.go

@@ -241,18 +241,20 @@ func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
 
 func (kv KV) OllamaEngineRequired() bool {
 	return slices.Contains([]string{
+		"bert",
+		"deepseek2",
+		"deepseekocr",
 		"gemma3",
 		"gemma3n",
 		"gptoss", "gpt-oss",
 		"llama4",
 		"mistral3",
 		"mllama",
+		"nomic-bert",
+		"olmo3",
 		"qwen25vl",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
-		"deepseekocr",
-		"deepseek2",
-		"nomic-bert",
 	}, kv.Architecture())
 }
 
@@ -838,9 +840,11 @@ func (f GGML) SupportsFlashAttention() bool {
 // FlashAttention checks if the model should enable flash attention
 func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
+		"bert",
 		"gemma3",
 		"gptoss", "gpt-oss",
 		"mistral3",
+		"olmo3",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
 	}, f.KV().String("general.architecture"))

kvcache/causal.go

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

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "17f7f4baad8b3a716ee139da7bb56ae984e8c0fa";
+char const *LLAMA_COMMIT = "ec98e2002";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/.rsync-filter

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

llama/llama.cpp/common/common.cpp

@@ -1013,31 +1013,40 @@ bool tty_can_use_colors() {
 // Model utils
 //
 
-static inline void common_init_sampler_from_model(
+// TODO: move to common/sampling
+static void common_init_sampler_from_model(
     const llama_model * model,
     common_params_sampling & sparams) {
 
     const uint64_t config = sparams.user_sampling_config;
 
     auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) {
-        if (config & user_config) return;
+        if (config & user_config) {
+            return;
+        }
 
         char buf[64] = {0};
         if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
             char * end = nullptr;
             int32_t v = strtol(buf, &end, 10);
-            if (end && end != buf) dst = v;
+            if (end && end != buf) {
+                dst = v;
+            }
         }
     };
 
     auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
-        if (config & user_config) return;
+        if (config & user_config) {
+            return;
+        }
 
         char buf[128] = {0};
         if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
             char * end = nullptr;
             float v = strtof(buf, &end);
-            if (end && end != buf) dst = v;
+            if (end && end != buf) {
+                dst = v;
+            }
         }
     };
 
@@ -1065,31 +1074,125 @@ static inline void common_init_sampler_from_model(
     get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA),    sparams.mirostat_eta,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA);
 }
 
-struct common_init_result common_init_from_params(common_params & params) {
+struct common_init_result::impl {
-    common_init_result iparams;
+    impl() = default;
+    ~impl() = default;
+
+    llama_model_ptr   model;
+    llama_context_ptr context;
+
+    std::vector<llama_adapter_lora_ptr> lora;
+
+    std::vector<common_sampler_ptr> samplers;
+};
+
+common_init_result::common_init_result(common_params & params) :
+    pimpl(new impl{}) {
     auto mparams = common_model_params_to_llama(params);
+    auto cparams = common_context_params_to_llama(params);
+
+    if (params.fit_params) {
+        LOG_INF("%s: fitting params to device memory, to report bugs during this step use -fit off (or --verbose if you can't)\n", __func__);
+        llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
+            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target, params.fit_params_min_ctx,
+            params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
+    }
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
     if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+        return;
-            __func__, params.model.path.c_str());
-        return iparams;
     }
 
-    common_init_sampler_from_model(model, params.sampling);
+    pimpl->model.reset(model);
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
-    auto cparams = common_context_params_to_llama(params);
+    // updates params.sampling
+    // TODO: fix naming
+    common_init_sampler_from_model(model, params.sampling);
+
+    if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
+        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
+        params.sampling.ignore_eos = false;
+    }
+
+    // initialize once
+    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
+        if (llama_vocab_is_eog(vocab, i)) {
+            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(vocab, i).c_str(), -INFINITY);
+            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
+        }
+    }
+
+    if (params.sampling.ignore_eos) {
+        // add EOG biases to the active set of logit biases
+        params.sampling.logit_bias.insert(
+                params.sampling.logit_bias.end(),
+                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
+    }
+
+    //if (params.sampling.penalty_last_n == -1) {
+    //    LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+    //    params.sampling.penalty_last_n = llama_n_ctx(lctx);
+    //}
+
+    //if (params.sampling.dry_penalty_last_n == -1) {
+    //    LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+    //    params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
+    //}
+
+    pimpl->samplers.resize(cparams.n_seq_max);
+
+    for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
+        pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
+    }
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
-            __func__, params.model.path.c_str());
+        return;
-        llama_model_free(model);
-        return iparams;
     }
 
+    pimpl->context.reset(lctx);
+}
+
+llama_model * common_init_result::model() {
+    return pimpl->model.get();
+}
+
+llama_context * common_init_result::context() {
+    return pimpl->context.get();
+}
+
+common_sampler * common_init_result::sampler(llama_seq_id seq_id) {
+    return pimpl->samplers[seq_id].get();
+}
+
+std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
+    return pimpl->lora;
+}
+
+void common_init_result::free_context() {
+    pimpl->context.reset();
+}
+
+common_init_result_ptr common_init_from_params(common_params & params) {
+    common_init_result_ptr res(new common_init_result(params));
+
+    llama_model * model = res->model();
+    if (model == NULL) {
+        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
+        return res;
+    }
+
+    llama_context * lctx = res->context();
+    if (lctx == NULL) {
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
+        return res;
+    }
+
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
     if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
         LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
@@ -1101,10 +1204,7 @@ struct common_init_result common_init_from_params(common_params & params) {
 
         const auto cvec = common_control_vector_load(params.control_vectors);
         if (cvec.n_embd == -1) {
-            llama_free(lctx);
+            return res;
-            llama_model_free(model);
-
-            return iparams;
         }
 
         int err = llama_apply_adapter_cvec(
@@ -1115,10 +1215,7 @@ struct common_init_result common_init_from_params(common_params & params) {
                 params.control_vector_layer_start,
                 params.control_vector_layer_end);
         if (err) {
-            llama_free(lctx);
+            return res;
-            llama_model_free(model);
-
-            return iparams;
         }
     }
 
@@ -1142,10 +1239,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
 
         if (!ok) {
-            llama_free(lctx);
+            return res;
-            llama_model_free(model);
-
-            return iparams;
         }
     }
 
@@ -1155,9 +1249,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
         if (lora == nullptr) {
             LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
-            llama_free(lctx);
+            return res;
-            llama_model_free(model);
-            return iparams;
         }
 
         char buf[1024];
@@ -1166,43 +1258,13 @@ struct common_init_result common_init_from_params(common_params & params) {
         la.task_name = buf;
         llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
         la.prompt_prefix = buf;
-        iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
+        res->lora().emplace_back(std::move(lora)); // copy to list of loaded adapters
     }
 
     if (!params.lora_init_without_apply) {
         common_set_adapter_lora(lctx, params.lora_adapters);
     }
 
-    if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
-        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
-        params.sampling.ignore_eos = false;
-    }
-
-    // initialize once
-    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
-        if (llama_vocab_is_eog(vocab, i)) {
-            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
-            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
-        }
-    }
-
-    if (params.sampling.ignore_eos) {
-        // add EOG biases to the active set of logit biases
-        params.sampling.logit_bias.insert(
-                params.sampling.logit_bias.end(),
-                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
-    }
-
-    if (params.sampling.penalty_last_n == -1) {
-        LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-        params.sampling.penalty_last_n = llama_n_ctx(lctx);
-    }
-
-    if (params.sampling.dry_penalty_last_n == -1) {
-        LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-        params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
-    }
-
     if (params.warmup) {
         LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
@@ -1241,12 +1303,11 @@ struct common_init_result common_init_from_params(common_params & params) {
         llama_set_warmup(lctx, false);
     }
 
-    iparams.model.reset(model);
+    return res;
-    iparams.context.reset(lctx);
-
-    return iparams;
 }
 
+common_init_result::~common_init_result() = default;
+
 std::string get_model_endpoint() {
     const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
     // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
@@ -1255,7 +1316,9 @@ std::string get_model_endpoint() {
     std::string model_endpoint = "https://huggingface.co/";
     if (endpoint_env) {
         model_endpoint = endpoint_env;
-        if (model_endpoint.back() != '/') model_endpoint += '/';
+        if (model_endpoint.back() != '/') {
+            model_endpoint += '/';
+        }
     }
     return model_endpoint;
 }

llama/llama.cpp/common/common.h

@@ -82,7 +82,8 @@ int32_t cpu_get_num_math();
 enum llama_example {
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
-    LLAMA_EXAMPLE_MAIN,
+    LLAMA_EXAMPLE_COMPLETION,
+    LLAMA_EXAMPLE_CLI,
     LLAMA_EXAMPLE_EMBEDDING,
     LLAMA_EXAMPLE_PERPLEXITY,
     LLAMA_EXAMPLE_RETRIEVAL,
@@ -98,6 +99,7 @@ enum llama_example {
     LLAMA_EXAMPLE_TTS,
     LLAMA_EXAMPLE_DIFFUSION,
     LLAMA_EXAMPLE_FINETUNE,
+    LLAMA_EXAMPLE_FIT_PARAMS,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -194,7 +196,6 @@ struct common_params_sampling {
 
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 
-
     std::vector<enum common_sampler_type> samplers = {
         COMMON_SAMPLER_TYPE_PENALTIES,
         COMMON_SAMPLER_TYPE_DRY,
@@ -215,6 +216,10 @@ struct common_params_sampling {
     std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
     std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
+    bool has_logit_bias() const {
+        return !logit_bias.empty();
+    }
+
     // print the parameters into a string
     std::string print() const;
 };
@@ -302,8 +307,8 @@ struct lr_opt {
 struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
 
 struct common_params {
-    int32_t n_predict             =    -1; // new tokens to predict
+    int32_t n_predict             =    -1; // max. number of new tokens to predict, -1 == no limit
-    int32_t n_ctx                 =  4096; // context size
+    int32_t n_ctx                 =     0; // context size, 0 == context the model was trained with
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt
@@ -327,6 +332,9 @@ struct common_params {
     int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM (-1 - use default)
     int32_t main_gpu           = 0;                // the GPU that is used for scratch and small tensors
     float   tensor_split[128]  = {0};              // how split tensors should be distributed across GPUs
+    bool    fit_params         = true;             // whether to fit unset model/context parameters to free device memory
+    size_t  fit_params_target  = 1024 * 1024*1024; // margin per device in bytes for fitting parameters to free memory
+    int32_t fit_params_min_ctx = 4096;             // minimum context size to set when trying to reduce memory use
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 
@@ -406,6 +414,7 @@ struct common_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool no_perf           = false; // disable performance metrics
+    bool show_timings      = true;  // show timing information on CLI
     bool ctx_shift         = false; // context shift on infinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     bool kv_unified        = false; // enable unified KV cache
@@ -462,7 +471,7 @@ struct common_params {
     std::string public_path   = "";                                                                         // NOLINT
     std::string api_prefix    = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
-    bool use_jinja = false;                                                                                 // NOLINT
+    bool use_jinja = true;                                                                                  // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     int reasoning_budget = -1;
@@ -483,6 +492,7 @@ struct common_params {
 
     // router server configs
     std::string models_dir    = ""; // directory containing models for the router server
+    std::string models_preset = ""; // directory containing model presets for the router server
     int models_max = 4;             // maximum number of models to load simultaneously
     bool models_autoload = true;    // automatically load models when requested via the router server
 
@@ -666,15 +676,29 @@ bool tty_can_use_colors();
 // Model utils
 //
 
-// note: defines object's lifetime
+struct common_sampler;
-struct common_init_result {
-    llama_model_ptr   model;
-    llama_context_ptr context;
 
-    std::vector<llama_adapter_lora_ptr> lora;
+// note: defines the model, context, samplers, ets. lifetimes
+struct common_init_result {
+    common_init_result(common_params & params);
+    ~common_init_result();
+
+    llama_model * model();
+    llama_context * context();
+    common_sampler * sampler(llama_seq_id seq_id);
+
+    std::vector<llama_adapter_lora_ptr> & lora();
+
+    void free_context();
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
 };
 
-struct common_init_result     common_init_from_params(common_params & params);
+using common_init_result_ptr = std::unique_ptr<common_init_result>;
+
+common_init_result_ptr common_init_from_params(common_params & params);
 
 struct llama_model_params     common_model_params_to_llama  (      common_params & params);
 struct llama_context_params   common_context_params_to_llama(const common_params & params);

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

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

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

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

llama/llama.cpp/common/log.cpp

@@ -420,6 +420,11 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
 
+void common_log_flush(struct common_log * log) {
+    log->pause();
+    log->resume();
+}
+
 static int common_get_verbosity(enum ggml_log_level level) {
     switch (level) {
         case GGML_LOG_LEVEL_DEBUG: return LOG_LEVEL_DEBUG;

llama/llama.cpp/common/log.h

@@ -84,6 +84,7 @@ void common_log_set_file      (struct common_log * log, const char * file); // n
 void common_log_set_colors    (struct common_log * log, log_colors colors); // not thread-safe
 void common_log_set_prefix    (struct common_log * log, bool prefix);       // whether to output prefix to each log
 void common_log_set_timestamps(struct common_log * log, bool timestamps);   // whether to output timestamps in the prefix
+void common_log_flush         (struct common_log * log);                    // flush all pending log messages
 
 // helper macros for logging
 // use these to avoid computing log arguments if the verbosity of the log is higher than the threshold

llama/llama.cpp/common/sampling.cpp

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

llama/llama.cpp/common/sampling.h

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

llama/llama.cpp/include/llama.h

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

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

@@ -3,6 +3,7 @@
 #include "ggml.h" // ggml_op
 
 #include <string>
+#include <set>
 
 //
 // gguf constants (sync with gguf.py)
@@ -79,6 +80,7 @@ enum llm_arch {
     LLM_ARCH_JAIS,
     LLM_ARCH_NEMOTRON,
     LLM_ARCH_NEMOTRON_H,
+    LLM_ARCH_NEMOTRON_H_MOE,
     LLM_ARCH_EXAONE,
     LLM_ARCH_EXAONE4,
     LLM_ARCH_RWKV6,
@@ -317,6 +319,7 @@ enum llm_tensor {
     LLM_TENSOR_DENSE_3_OUT,
     LLM_TENSOR_OUTPUT,
     LLM_TENSOR_OUTPUT_NORM,
+    LLM_TENSOR_OUTPUT_NORM_LFM2, // fix for wrong tensor name
     LLM_TENSOR_ROPE_FREQS,
     LLM_TENSOR_ROPE_FACTORS_LONG,
     LLM_TENSOR_ROPE_FACTORS_SHORT,
@@ -528,6 +531,10 @@ struct LLM_TN_IMPL {
     const int bid;
     const int xid;
 
+    const std::set<llm_tensor> model_tensors;
+
+    LLM_TN_IMPL(llm_arch arch, llm_tensor tensor, const char * suffix, int bid, int xid);
+
     std::string str() const;
 
     operator std::string() const {
@@ -549,11 +556,11 @@ struct LLM_TN {
     llm_arch arch;
 
     LLM_TN_IMPL operator()(llm_tensor tensor, const char * suffix, int bid = -1, int xid = -1) const {
-        return { arch, tensor, suffix, bid, xid };
+        return LLM_TN_IMPL(arch, tensor, suffix, bid, xid);
     }
 
     LLM_TN_IMPL operator()(llm_tensor tensor, int bid = -1, int xid = -1) const {
-        return { arch, tensor, nullptr, bid, xid };
+        return LLM_TN_IMPL(arch, tensor, nullptr, bid, xid);
     }
 };
 

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

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

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

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

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

@@ -9,6 +9,7 @@
 #include "llama-model.h"
 
 #include <cinttypes>
+#include <cmath>
 #include <cstring>
 #include <limits>
 #include <stdexcept>
@@ -72,6 +73,43 @@ llama_context::llama_context(
         cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f;
     }
 
+    if (cparams.yarn_ext_factor != 0) {
+        static auto get_mscale = [](float scale, float mscale) {
+            return scale <= 1.0f ? 1.0f : (0.1f * mscale * logf(scale) + 1.0f);
+        };
+
+        const float factor = 1.0f / cparams.rope_freq_scale;
+
+        // ref: https://github.com/huggingface/transformers/blob/6d00f6b0a5679c36510f203e4226e36f517c3032/src/transformers/modeling_rope_utils.py#L336-L348
+        if (hparams.rope_yarn_log_mul != 0.0f) {
+            // note: here we assume `mscale == 1.0f`
+            // TODO: start reading the actual value of mscale and handle the case where it is not 1.0f
+                  float mscale          = 1.0f;
+            const float mscale_all_dims = hparams.rope_yarn_log_mul;
+
+            // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+            // special-case DEEPSEEK v2:
+            // https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/blob/main/config.json#L42-L43
+            if (model.arch == LLM_ARCH_DEEPSEEK2 && mscale_all_dims != 1.0f) {
+                mscale = mscale_all_dims;
+            }
+
+            cparams.yarn_attn_factor = get_mscale(factor, mscale) / get_mscale(factor, mscale_all_dims);
+
+            LLAMA_LOG_WARN("%s: setting new yarn_attn_factor = %.4f (mscale == %.1f, mscale_all_dim = %.1f)\n",
+                    __func__, cparams.yarn_attn_factor, mscale, mscale_all_dims);
+        } else {
+            cparams.yarn_attn_factor = get_mscale(factor, 1.0f);
+        }
+
+        // when YARN is applied with yarn_ext_factor != 0.0f, we need to cancel this factor:
+        // https://github.com/ggml-org/llama.cpp/blob/a81a569577cc38b32558958b048228150be63eae/ggml/src/ggml-cpu/ops.cpp#L5541-L5544
+        //
+        // ref: https://github.com/ggml-org/llama.cpp/discussions/7416
+        //      https://github.com/ggml-org/llama.cpp/pull/17945
+        cparams.yarn_attn_factor *= 1.0f / (1.0f + 0.1f * logf(factor));
+    }
+
     cparams.yarn_attn_factor *= hparams.rope_attn_factor;
 
     if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
@@ -93,14 +131,6 @@ llama_context::llama_context(
     // with causal attention, the batch size is limited by the context size
     cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
 
-    // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask
-    // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext)
-    // ref: https://github.com/ggerganov/llama.cpp/pull/5021
-    // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_memory
-    if (cparams.n_batch < GGML_KQ_MASK_PAD) {
-        LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
-        cparams.n_batch = GGML_KQ_MASK_PAD;
-    }
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
 
     cparams.op_offload = params.op_offload;
@@ -228,6 +258,7 @@ llama_context::llama_context(
 
         backend_buft.clear();
         backend_ptrs.clear();
+        backend_buf_exp_size.clear();
 
         for (auto & backend : backends) {
             auto * buft = ggml_backend_get_default_buffer_type(backend.get());
@@ -244,6 +275,7 @@ llama_context::llama_context(
 
             backend_buft.push_back(buft);
             backend_ptrs.push_back(backend.get());
+            backend_buf_exp_size.push_back(0);
         }
 
         LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size());
@@ -359,7 +391,8 @@ llama_context::llama_context(
 
         // reserve pp (prompt processing) graph first so that buffers are only allocated once
         {
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get(),
+                model.hparams.no_alloc, model.hparams.no_alloc ? backend_buf_exp_size.data() : nullptr);
             if (!gf) {
                 if (pipeline_parallel) {
                     LLAMA_LOG_WARN("%s: compute buffer allocation failed, retrying without pipeline parallelism\n", __func__);
@@ -377,7 +410,7 @@ llama_context::llama_context(
 
         // reserve with tg (token generation) graph to get the number of splits and nodes
         {
-            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get());
+            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get(), model.hparams.no_alloc);
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute tg buffers");
             }
@@ -392,7 +425,7 @@ llama_context::llama_context(
             //
             // auto * gf = graph_reserve(n_tokens, 1, n_tokens, mctx.get());
             //
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get(), model.hparams.no_alloc);
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -401,11 +434,13 @@ llama_context::llama_context(
         for (size_t i = 0; i < backend_ptrs.size(); ++i) {
             ggml_backend_t             backend = backend_ptrs[i];
             ggml_backend_buffer_type_t buft    = backend_buft[i];
-            size_t size = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+            if (!model.hparams.no_alloc) {
-            if (size > 1) {
+                backend_buf_exp_size[i] = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+            }
+            if (backend_buf_exp_size[i] > 1) {
                 LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__,
                         ggml_backend_buft_name(buft),
-                        size / 1024.0 / 1024.0);
+                        backend_buf_exp_size[i] / 1024.0 / 1024.0);
             }
         }
 
@@ -424,6 +459,23 @@ llama_context::llama_context(
 }
 
 llama_context::~llama_context() {
+    // FIXME this currently results in a use-after-free bug if the model is freed before the context
+    // if (!model.hparams.no_alloc) {
+    //     for (size_t i = 0; i < backend_ptrs.size(); ++i) {
+    //         ggml_backend_t             backend = backend_ptrs[i];
+    //         ggml_backend_buffer_type_t buft    = backend_buft[i];
+
+    //         const size_t size_exp = backend_buf_exp_size[i];
+    //         const size_t size_act = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+    //         if (size_exp == size_act) {
+    //             LLAMA_LOG_DEBUG("%s: %10s compute buffer size is %8.4f MiB, matches expectation of %8.4f MiB\n",
+    //                 __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+    //         } else {
+    //             LLAMA_LOG_WARN("%s: %10s compute buffer size of %8.4f MiB, does not match expectation of %8.4f MiB\n",
+    //                 __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+    //         }
+    //     }
+    // }
     ggml_opt_free(opt_ctx);
 }
 
@@ -1325,6 +1377,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
             // This doesn't happen often, but may be annoying in some cases (like the HellaSwag benchmark)
             LLAMA_LOG_INFO("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
+            synchronize();
             buf_output = nullptr;
             logits = nullptr;
             embd = nullptr;
@@ -1396,7 +1449,8 @@ llm_graph_result * llama_context::get_gf_res_reserve() const {
     return static_cast<llm_graph_result *>(gf_res_reserve.get());
 }
 
-ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only) {
+ggml_cgraph * llama_context::graph_reserve(
+        uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only, size_t * sizes) {
     LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
     GGML_ASSERT(n_outputs >= 1);
 
@@ -1433,8 +1487,13 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
 
     // initialize scheduler with the specified graph
     if (split_only) {
+        if (sizes) {
+            ggml_backend_sched_reserve_size(sched.get(), gf, sizes);
+        } else {
             ggml_backend_sched_split_graph(sched.get(), gf);
+        }
     } else if (!ggml_backend_sched_reserve(sched.get(), gf)) {
+        GGML_ASSERT(!sizes);
         LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
         return nullptr;
     }
@@ -2056,15 +2115,26 @@ void llama_context::perf_reset() {
 
 std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> llama_context::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> ret;
-    for (const auto & buft_size : model.memory_breakdown()) {
+    for (const auto & [buft, size] : model.memory_breakdown()) {
-        ret[buft_size.first].model += buft_size.second;
+        ret[buft].model += size;
     }
-    for (const auto & buft_size : memory->memory_breakdown()) {
+    if (memory) {
-        ret[buft_size.first].context += buft_size.second;
+        for (const auto & [buft, size] : memory->memory_breakdown()) {
+            ret[buft].context += size;
         }
+    }
+    if (model.hparams.no_alloc) {
+        for (size_t i = 0; i < backends.size(); ++i) {
+            ggml_backend_t             backend = backends[i].get();
+            ggml_backend_buffer_type_t buft    = ggml_backend_sched_get_buffer_type(sched.get(), backend);
+            ret[buft].compute += backend_buf_exp_size[i];
+        }
+    } else {
         for (const auto & backend_ptr : backends) {
             ggml_backend_t             backend = backend_ptr.get();
-        ret[ggml_backend_sched_get_buffer_type(sched.get(), backend)].compute += ggml_backend_sched_get_buffer_size(sched.get(), backend);
+            ggml_backend_buffer_type_t buft    = ggml_backend_sched_get_buffer_type(sched.get(), backend);
+            ret[buft].compute += ggml_backend_sched_get_buffer_size(sched.get(), backend);
+        }
     }
     return ret;
 }

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

@@ -26,6 +26,10 @@ struct llama_memory_breakdown_data {
     size_t model   = 0; // memory allocated for the model
     size_t context = 0; // memory allocated for the context
     size_t compute = 0; // memory allocated for temporary compute buffers
+
+    size_t total() const {
+        return model + context + compute;
+    }
 };
 
 struct llama_context {
@@ -206,7 +210,8 @@ public:
     ggml_status graph_compute(ggml_cgraph * gf, bool batched);
 
     // reserve a graph with a dummy ubatch of the specified size
-    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false);
+    ggml_cgraph * graph_reserve(
+        uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false, size_t * sizes = nullptr);
 
 private:
     llm_graph_params graph_params(
@@ -281,9 +286,10 @@ private:
 
     std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
 
-    // buffer types used for the compute buffer of each backend
+    // pointers and buffer types used for the compute buffer of each backend
     std::vector<ggml_backend_t>             backend_ptrs;
     std::vector<ggml_backend_buffer_type_t> backend_buft;
+    std::vector<size_t>                     backend_buf_exp_size; // expected buffer sizes
 
     llm_graph_result_ptr gf_res_prev;
     llm_graph_result_ptr gf_res_reserve;

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

@@ -78,7 +78,7 @@ void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
         for (int i = 0; i < n_tokens; ++i) {
             const float pos = ubatch->pos[i];
             attn_scale_data[i] = std::log(
-                std::floor((pos + 1.0f) / n_attn_temp_floor_scale) + 1.0
+                std::floor((pos + f_attn_temp_offset) / n_attn_temp_floor_scale) + 1.0
             ) * f_attn_temp_scale + 1.0;
         }
 
@@ -254,6 +254,24 @@ void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+bool llm_graph_input_rs::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_memory_recurrent_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= s_copy->ne[0] == mctx->get_n_rs();
+
+    res &= s_copy_main->ne[0]  == params.ubatch.n_seqs;
+    res &= s_copy_extra->ne[0] == mctx->get_n_rs() - params.ubatch.n_seqs;
+
+    res &= head == mctx->get_head();
+    res &= rs_z == mctx->get_rs_z();
+
+    return res;
+}
+
 void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
@@ -385,7 +403,7 @@ bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) {
   //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
 
     res &= self_kq_mask->ne[0] == mctx->get_n_kv();
-    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+    res &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
 
     return res;
 }
@@ -416,10 +434,10 @@ bool llm_graph_input_attn_kv_iswa::can_reuse(const llm_graph_params & params) {
   //res &= self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
 
     res &= self_kq_mask->ne[0] == mctx->get_base()->get_n_kv();
-    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+    res &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
 
     res &= self_kq_mask_swa->ne[0] == mctx->get_swa()->get_n_kv();
-    res &= self_kq_mask_swa->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+    res &= self_kq_mask_swa->ne[1] == params.ubatch.n_tokens;
 
     return res;
 }
@@ -452,7 +470,7 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
             }
         }
 
-        for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
+        for (int i = n_tokens; i < n_tokens; ++i) {
             for (int j = 0; j < n_enc; ++j) {
                 data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY;
             }
@@ -461,8 +479,46 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
-    inp_attn->set_input(ubatch);
+    mctx->get_attn()->set_input_k_idxs(inp_attn->self_k_idxs, ubatch);
-    inp_rs->set_input(ubatch);
+    mctx->get_attn()->set_input_v_idxs(inp_attn->self_v_idxs, ubatch);
+
+    mctx->get_attn()->set_input_kq_mask(inp_attn->self_kq_mask, ubatch, cparams.causal_attn);
+
+    const int64_t n_rs = mctx->get_recr()->get_n_rs();
+
+    if (inp_rs->s_copy) {
+        GGML_ASSERT(ggml_backend_buffer_is_host(inp_rs->s_copy->buffer));
+        int32_t * data = (int32_t *) inp_rs->s_copy->data;
+
+        // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
+        for (uint32_t i = 0; i < n_rs; ++i) {
+            data[i] = mctx->get_recr()->s_copy(i);
+        }
+    }
+}
+
+bool llm_graph_input_mem_hybrid::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_memory_hybrid_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= inp_attn->self_k_idxs->ne[0] == params.ubatch.n_tokens;
+  //res &= inp_attn->self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+    res &= inp_attn->self_kq_mask->ne[0] == mctx->get_attn()->get_n_kv();
+    res &= inp_attn->self_kq_mask->ne[1] == params.ubatch.n_tokens;
+
+    res &= inp_rs->s_copy->ne[0] == mctx->get_recr()->get_n_rs();
+
+    res &= inp_rs->s_copy_main->ne[0]  == params.ubatch.n_seqs;
+    res &= inp_rs->s_copy_extra->ne[0] == mctx->get_recr()->get_n_rs() - params.ubatch.n_seqs;
+
+    res &= inp_rs->head == mctx->get_recr()->get_head();
+    res &= inp_rs->rs_z == mctx->get_recr()->get_rs_z();
+
+    return res;
 }
 
 //
@@ -1089,6 +1145,15 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 cur = ggml_relu(ctx0, cur);
                 cb(cur, "ffn_moe_relu", il);
             } break;
+        case LLM_FFN_RELU_SQR:
+            if (gate_exps) {
+                // TODO: add support for gated squared relu
+                GGML_ABORT("fatal error: gated squared relu not implemented");
+            } else {
+                cur = ggml_relu(ctx0, cur);
+                cur = ggml_sqr(ctx0, cur);
+                cb(cur, "ffn_moe_relu_sqr", il);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -1203,7 +1268,7 @@ ggml_tensor * llm_graph_context::build_inp_pos() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
-    auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
+    auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale, hparams.f_attn_temp_offset);
 
     auto & cur = inp->attn_scale;
 
@@ -1470,13 +1535,13 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
     auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);
 
     // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
-    inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+    inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens, 1, 1);
     ggml_set_input(inp->self_kq_mask);
 
     inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
 
     if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
-        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens, 1, 1);
         ggml_set_input(inp->self_kq_mask_swa);
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
@@ -1558,7 +1623,7 @@ static std::unique_ptr<llm_graph_input_attn_kv> build_attn_inp_kv_impl(
         inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1701,7 +1766,7 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
 
     const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
 
-    inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+    inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, n_tokens, 1, 1);
     ggml_set_input(inp->cross_kq_mask);
 
     inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->cross_kq_mask, GGML_TYPE_F16) : inp->cross_kq_mask;
@@ -1767,7 +1832,7 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
         inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1781,7 +1846,7 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
         inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp->self_kq_mask_swa);
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
@@ -1841,6 +1906,9 @@ static std::unique_ptr<llm_graph_input_rs> build_rs_inp_impl(
     inp->s_copy_main  = ggml_view_1d(ctx0, inp->s_copy, n_seqs, 0);
     inp->s_copy_extra = ggml_view_1d(ctx0, inp->s_copy, n_rs - n_seqs, n_seqs * inp->s_copy->nb[0]);
 
+    inp->head = mctx_cur->get_head();
+    inp->rs_z = mctx_cur->get_rs_z();
+
     return inp;
 }
 
@@ -1912,7 +1980,7 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
     auto inp_rs   = build_rs_inp_impl     (ctx0, ubatch, mctx_cur->get_recr());
     auto inp_attn = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn());
 
-    auto inp = std::make_unique<llm_graph_input_mem_hybrid>(std::move(inp_attn), std::move(inp_rs), mctx_cur);
+    auto inp = std::make_unique<llm_graph_input_mem_hybrid>(cparams, std::move(inp_attn), std::move(inp_rs), mctx_cur);
 
     return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
 }

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

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

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

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

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

@@ -34,6 +34,7 @@ struct llama_hparams_convnext {
 
 struct llama_hparams {
     bool vocab_only;
+    bool no_alloc;
     bool rope_finetuned;
     bool use_par_res;
     bool swin_norm;
@@ -109,6 +110,7 @@ struct llama_hparams {
     float    rope_freq_base_train_swa;
     float    rope_freq_scale_train;
     float    rope_freq_scale_train_swa;
+
     uint32_t n_ctx_orig_yarn;
     float    rope_yarn_log_mul = 0.0f;
 
@@ -166,6 +168,7 @@ struct llama_hparams {
     uint32_t n_no_rope_layer_step    = 4;
     uint32_t n_attn_temp_floor_scale = 0;
     float    f_attn_temp_scale       = 0.0f;
+    float    f_attn_temp_offset      = 0.0f; // offset position index
 
     // gemma3n altup
     uint32_t n_altup      = 4; // altup_num_inputs
@@ -272,7 +275,8 @@ struct llama_hparams {
     // TODO: think of a better place for this function
     // TODO: pack the SWA params in a struct?
     static bool is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1);
+
+    bool use_mrope() const;
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
-

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

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

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

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

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

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

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

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

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

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

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

@@ -120,6 +120,8 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_16B_A1B:       return "16B.A1B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
+        case LLM_TYPE_31B_A3_5B:     return "31B.A3.5B";
+        case LLM_TYPE_80B_A3B:       return "80B.A3B";
         case LLM_TYPE_100B_A6B:      return "100B.A6B";
         case LLM_TYPE_106B_A12B:     return "106B.A12B";
         case LLM_TYPE_230B_A10B:     return "230B.A10B";
@@ -667,6 +669,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.n_swa                   = 8192;
                     hparams.n_attn_temp_floor_scale = 8192;
                     hparams.f_attn_temp_scale       = 0.1f;
+                    hparams.f_attn_temp_offset      = 1.0f;
                     hparams.set_swa_pattern(4);   // pattern: 3 chunked - 1 full
                 }
 
@@ -1634,12 +1637,19 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     // that have no expert_gating_func model parameter set
                     hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;
                 }
-                ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, false);
+
+                if (ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, 0.0f)) {
+                    // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+                    // cancel the factor from the convert script
+                    hparams.rope_yarn_log_mul /= 0.1f;
+                }
 
                 // (optional) temperature tuning - used by mistral-large
                 ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE,  hparams.f_attn_temp_scale,       false);
                 ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.n_attn_temp_floor_scale, false);
 
+                hparams.f_attn_temp_offset = 0.0f;
+
                 switch (hparams.n_layer) {
                     case 27: type = LLM_TYPE_16B; break;
                     case 60: type = LLM_TYPE_236B; break;
@@ -1680,6 +1690,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
         case LLM_ARCH_GLM4:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
                 switch (hparams.n_layer) {
                     case 40: type = LLM_TYPE_9B; break;
                     case 61: type = LLM_TYPE_32B; break;
@@ -1690,6 +1701,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             {
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,     hparams.n_ff_exp);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
 
                 // MoE parameters
                 ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert);
@@ -1788,6 +1800,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
             } break;
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             {
                 ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
                 ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
@@ -1803,7 +1816,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp,        false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp,      false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
+
                 switch (hparams.n_layer) {
+                    case 52: type = LLM_TYPE_31B_A3_5B; break; // Nemotron-H_MOE 31B
                     case 56: type = LLM_TYPE_9B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
@@ -2272,7 +2292,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
 
                 switch (hparams.n_layer) {
-                    case 80: type = LLM_TYPE_80B_A3B; break;
+                    case 48: type = LLM_TYPE_80B_A3B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -2283,7 +2303,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_FAST, hparams.yarn_beta_fast,    false);
                 ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, hparams.yarn_beta_slow,    false);
-                ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,     hparams.rope_yarn_log_mul, false);
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,   hparams.rope_yarn_log_mul, 0.0f);
+
+                hparams.f_attn_temp_offset = 0.0f;
 
                 // TODO: maybe add n_attn_temp_floor_scale as a separate KV?
                 if (hparams.f_attn_temp_scale != 0.0f) {
@@ -2293,18 +2315,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     }
                 }
 
-                // TODO: this seems to be correct with the case of mscale == mscale_all_dims == 1.0f
-                //       but may need further verification with other values
-                if (hparams.rope_yarn_log_mul != 0.0f) {
-                    float factor = 1.0f / hparams.rope_freq_scale_train;
-                    float mscale = 1.0f;
-                    float mscale_all_dims = hparams.rope_yarn_log_mul;
-                    static auto get_mscale = [](float scale, float mscale) {
-                        return scale <= 1.0f ? 1.0f : (0.1f * mscale * logf(scale) + 1.0f);
-                    };
-                    hparams.yarn_attn_factor = get_mscale(factor, mscale) / get_mscale(factor, mscale_all_dims);
-                }
-
                 switch (hparams.n_layer) {
                     case 26: type = LLM_TYPE_3B; break;
                     case 34: type = LLM_TYPE_8B; break;
@@ -3404,9 +3414,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
 
                         // optional bias tensors
-                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, 0);
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
-                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
-                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
@@ -5175,6 +5185,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
                 } break;
             case LLM_ARCH_NEMOTRON_H:
+            case LLM_ARCH_NEMOTRON_H_MOE:
                 {
                     // mamba2 Mixer SSM params
                     // NOTE: int64_t for tensor dimensions
@@ -5185,6 +5196,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_group    = hparams.ssm_n_group;
                     const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
 
+                    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+                    const int64_t n_ff_shexp = hparams.n_ff_shexp;
+
                     // embeddings
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -5234,6 +5248,19 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias",   i), {n_embd_k_gqa_i}, TENSOR_NOT_REQUIRED);
                             layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias",   i), {n_embd_v_gqa_i}, TENSOR_NOT_REQUIRED);
                             layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd},         TENSOR_NOT_REQUIRED);
+                        }  else {
+                            if (n_expert != 0) {
+                                layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, 0);
+                                layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, 0);
+
+                                // MoE branch
+                                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+                                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+                                // Shared expert branch
+                                layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
+                                layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
+
                             } else {
                                 // mlp layers
                                 layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, 0);
@@ -5242,6 +5269,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                                 layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);
                             }
                         }
+                    }
                 } break;
             case LLM_ARCH_EXAONE:
                 {
@@ -6650,9 +6678,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
         std::vector<ggml_backend_buffer_ptr> bufs;
         if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported && is_default_buft) {
+            GGML_ASSERT(!ml.no_alloc);
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                 // only the mmap region containing the tensors in the model is mapped to the backend buffer
-                // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
+                // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer,
+                //     then we could just use metal for all layers
                 // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
                 void * addr = nullptr;
                 size_t first, last; // NOLINT
@@ -6668,9 +6698,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 bufs.emplace_back(buf);
                 buf_map.emplace(idx, buf);
             }
+        } else {
+            ggml_backend_buffer_t buf;
+            if (ml.no_alloc) {
+                buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
+                for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
+                    t->buffer = buf; // set dummy buffer for weights so that the backend scheduler won't try to allocate them
+                }
+            } else {
+                buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); // real buffer
             }
-        else {
-            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
             if (buf == nullptr) {
                 throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
             }
@@ -6725,6 +6762,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         }
     }
 
+    if (ml.no_alloc) {
+        return true;
+    }
+
     // load tensor data
     for (auto & [ctx, buf_map] : ctx_buf_maps) {
         if (!ml.load_all_data(ctx, buf_map, use_mlock ? &pimpl->mlock_mmaps : NULL, params.progress_callback, params.progress_callback_user_data)) {
@@ -6767,11 +6808,20 @@ size_t llama_model::n_devices() const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_model::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, bufs] : pimpl->ctxs_bufs) {
+    for (const auto & [ctx, bufs] : pimpl->ctxs_bufs) {
+        if (hparams.no_alloc) {
+            GGML_ASSERT(bufs.size() == 1);
+            ggml_backend_buffer_t buf = bufs[0].get();
+            GGML_ASSERT(ggml_backend_buffer_get_base(buf) == nullptr);
+            ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf);
+            ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
+        } else {
             for (const auto & buf : bufs) {
+                // GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) != nullptr); // multi_buffer does not have a defined base
                 ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
             }
         }
+    }
     return ret;
 }
 
@@ -6814,6 +6864,7 @@ void llama_model::print_info() const {
     // hparams
     LLAMA_LOG_INFO("%s: arch             = %s\n",     __func__, arch_name().c_str());
     LLAMA_LOG_INFO("%s: vocab_only       = %d\n",     __func__, hparams.vocab_only);
+    LLAMA_LOG_INFO("%s: no_alloc         = %d\n",     __func__, hparams.no_alloc);
 
     if (!hparams.vocab_only) {
         LLAMA_LOG_INFO("%s: n_ctx_train      = %u\n",     __func__, hparams.n_ctx_train);
@@ -6848,6 +6899,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
         LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
+        LLAMA_LOG_INFO("%s: rope_yarn_log_mul= %.4f\n",   __func__, hparams.rope_yarn_log_mul);
         LLAMA_LOG_INFO("%s: rope_finetuned   = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
         // MRoPE (Multi-axis Rotary Position Embedding) sections
         if (const auto & s = hparams.rope_sections; s[0] || s[1] || s[2] || s[3]) {
@@ -6870,7 +6922,8 @@ void llama_model::print_info() const {
         arch == LLM_ARCH_PLAMO2 ||
         arch == LLM_ARCH_GRANITE_HYBRID ||
         arch == LLM_ARCH_QWEN3NEXT ||
-        arch == LLM_ARCH_NEMOTRON_H) {
+        arch == LLM_ARCH_NEMOTRON_H ||
+        arch == LLM_ARCH_NEMOTRON_H_MOE) {
         LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
         LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
         LLAMA_LOG_INFO("%s: ssm_d_state      = %u\n",     __func__, hparams.ssm_d_state);
@@ -6911,7 +6964,6 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
         LLAMA_LOG_INFO("%s: expert_weights_norm  = %d\n",     __func__, hparams.expert_weights_norm);
         LLAMA_LOG_INFO("%s: expert_gating_func   = %s\n",     __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
-        LLAMA_LOG_INFO("%s: rope_yarn_log_mul    = %.4f\n",   __func__, hparams.rope_yarn_log_mul);
     }
 
     if (arch == LLM_ARCH_QWEN2MOE) {
@@ -6926,7 +6978,8 @@ void llama_model::print_info() const {
     if (arch == LLM_ARCH_MINICPM ||
         arch == LLM_ARCH_GRANITE ||
         arch == LLM_ARCH_GRANITE_MOE ||
-        arch == LLM_ARCH_GRANITE_HYBRID) {
+        arch == LLM_ARCH_GRANITE_HYBRID ||
+        arch == LLM_ARCH_NEMOTRON_H_MOE) {
         LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
         LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
         LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
@@ -7107,7 +7160,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                     if (arch == LLM_ARCH_FALCON_H1) {
                         filter_attn = [&](int32_t) { return true; };
                         filter_recr = [&](int32_t) { return true; };
-                    } else if (arch == LLM_ARCH_NEMOTRON_H) {
+                    } else if (arch == LLM_ARCH_NEMOTRON_H || arch == LLM_ARCH_NEMOTRON_H_MOE) {
                         filter_attn = [&](int32_t il) {
                             return !hparams.is_recurrent(il) && hparams.n_ff(il) == 0;
                         };
@@ -7478,6 +7531,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
                 llm = std::make_unique<llm_build_nemotron>(*this, params);
             } break;
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             {
                 llm = std::make_unique<llm_build_nemotron_h>(*this, params);
             } break;
@@ -7666,6 +7720,7 @@ llama_model_params llama_model_default_params() {
         /*.check_tensors               =*/ false,
         /*.use_extra_bufts             =*/ true,
         /*.no_host                     =*/ false,
+        /*.no_alloc                    =*/ false,
     };
 
     return result;
@@ -7765,6 +7820,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ARWKV7:
         case LLM_ARCH_WAVTOKENIZER_DEC:
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             return LLAMA_ROPE_TYPE_NONE;
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values
@@ -7785,7 +7841,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_PLM:
         case LLM_ARCH_CHATGLM:
-        case LLM_ARCH_GLM4:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
         case LLM_ARCH_GRANITE_HYBRID:
@@ -7848,7 +7903,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_SMALLTHINKER:
-        case LLM_ARCH_GLM4_MOE:
         case LLM_ARCH_SEED_OSS:
         case LLM_ARCH_GROVEMOE:
         case LLM_ARCH_APERTUS:
@@ -7865,6 +7919,11 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_QWEN3VLMOE:
             return LLAMA_ROPE_TYPE_IMROPE;
 
+        case LLM_ARCH_GLM4:
+            return model->hparams.use_mrope() ? LLAMA_ROPE_TYPE_MROPE : LLAMA_ROPE_TYPE_NORM;
+        case LLM_ARCH_GLM4_MOE:
+            return model->hparams.use_mrope() ? LLAMA_ROPE_TYPE_MROPE : LLAMA_ROPE_TYPE_NEOX;
+
         // all model arches should be listed explicitly here
         case LLM_ARCH_UNKNOWN:
             GGML_ABORT("unknown architecture");

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

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

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

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

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

@@ -1884,7 +1884,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 clean_spaces = false;
             } else if (
                     tokenizer_pre == "qwen2" ||
-                    tokenizer_pre == "deepseek-r1-qwen") {
+                    tokenizer_pre == "deepseek-r1-qwen" ||
+                    tokenizer_pre == "kormo") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
                 clean_spaces = false;
             } else if (

llama/llama.cpp/src/llama.cpp

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

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

@@ -1,7 +1,5 @@
 #include "models.h"
 
-
-
 llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
@@ -20,9 +18,15 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
 
     // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
     // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
-    const float mscale      = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+    // And also: https://github.com/ggml-org/llama.cpp/pull/17945 [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+
+    // first cancel the adjustment from llama_hparams::yarn_attn_factor_adjust to get the original attn_factor
+    GGML_ASSERT(ext_factor >= 0.0f);
+    const float attn_factor_org = attn_factor * (1.0f + 0.1f * logf(1.0f / freq_scale));
+
+    // use the original attn_factor to pre-scale the kq_scale
+    const float mscale   = attn_factor_org * (1.0f + 0.1f * hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
     const float kq_scale = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k));
-    const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
 
     ggml_tensor * cur;
     ggml_tensor * inpL;

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

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

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

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

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

@@ -441,23 +441,13 @@ private:
                 ggml_tensor * cur,
                 ggml_tensor * causal_mask,
                 ggml_tensor * identity,
+                ggml_tensor * diag_mask,
                         int   il);
 
     ggml_tensor * build_layer_ffn(
                 ggml_tensor * cur,
                         int   il);
 
-    ggml_tensor * build_delta_net_recurrent(
-                ggml_tensor * q,
-                ggml_tensor * k,
-                ggml_tensor * v,
-                ggml_tensor * g,
-                ggml_tensor * beta,
-                ggml_tensor * state,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                        int   il);
-
     ggml_tensor * build_delta_net_chunking(
                 ggml_tensor * q,
                 ggml_tensor * k,
@@ -467,6 +457,16 @@ private:
                 ggml_tensor * state,
                 ggml_tensor * causal_mask,
                 ggml_tensor * identity,
+                ggml_tensor * diag_mask,
+                        int   il);
+
+    ggml_tensor * build_delta_net_autoregressive(
+                ggml_tensor * q,
+                ggml_tensor * k,
+                ggml_tensor * v,
+                ggml_tensor * g,
+                ggml_tensor * beta,
+                ggml_tensor * state,
                 int           il);
 
     ggml_tensor * build_norm_gated(

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

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

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

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

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

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

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

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

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

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

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

@@ -0,0 +1,300 @@
+#pragma once
+
+#include "ggml.h"
+#include "clip.h"
+#include "clip-impl.h"
+
+#include <vector>
+#include <unordered_set>
+#include <cstdint>
+#include <cmath>
+
+enum ffn_op_type {
+    FFN_GELU,
+    FFN_GELU_ERF,
+    FFN_SILU,
+    FFN_GELU_QUICK,
+};
+
+enum norm_type {
+    NORM_TYPE_NORMAL,
+    NORM_TYPE_RMS,
+};
+
+enum patch_merge_type {
+    PATCH_MERGE_FLAT,
+    PATCH_MERGE_SPATIAL_UNPAD,
+};
+
+struct clip_hparams {
+    int32_t image_size = 0;
+    int32_t patch_size = 0;
+    int32_t n_embd = 0;
+    int32_t n_ff = 0;
+    int32_t projection_dim = 0;
+    int32_t n_head = 0;
+    int32_t n_layer = 0;
+    // idefics3
+    int32_t image_longest_edge = 0;
+    int32_t image_min_pixels = -1;
+    int32_t image_max_pixels = -1;
+    int32_t n_merge = 0; // number of patch merges **per-side**
+
+    float image_mean[3];
+    float image_std[3];
+
+    // for models using dynamic image size, we need to have a smaller image size to warmup
+    // otherwise, user will get OOM everytime they load the model
+    int32_t warmup_image_size = 0;
+    int32_t warmup_audio_size = 3000;
+
+    ffn_op_type ffn_op = FFN_GELU;
+
+    patch_merge_type mm_patch_merge_type = PATCH_MERGE_FLAT;
+
+    float eps = 1e-6;
+    float rope_theta = 0.0;
+
+    std::vector<clip_image_size> image_res_candidates; // for llava-uhd style models
+    int32_t image_crop_resolution;
+    std::unordered_set<int32_t> vision_feature_layer;
+    int32_t attn_window_size = 0;
+    int32_t n_wa_pattern = 0;
+
+    // audio
+    int32_t n_mel_bins = 0; // whisper preprocessor
+    int32_t proj_stack_factor = 0; // ultravox
+
+    // audio-to-mel preprocessor params
+    int32_t audio_chunk_len   = -1; // in seconds
+    int32_t audio_sample_rate = -1;
+    int32_t audio_n_fft       = -1;
+    int32_t audio_window_len  = -1;
+    int32_t audio_hop_len     = -1;
+
+    // legacy
+    bool has_llava_projector = false;
+    int minicpmv_version = 0;
+    int32_t minicpmv_query_num = 0;         // MiniCPM-V query number
+
+    // custom value provided by user, can be undefined if not set
+    int32_t custom_image_min_tokens = -1;
+    int32_t custom_image_max_tokens = -1;
+
+    void set_limit_image_tokens(int n_tokens_min, int n_tokens_max) {
+        const int cur_merge = n_merge == 0 ? 1 : n_merge;
+        const int patch_area = patch_size * patch_size * cur_merge * cur_merge;
+        image_min_pixels = (custom_image_min_tokens > 0 ? custom_image_min_tokens : n_tokens_min) * patch_area;
+        image_max_pixels = (custom_image_max_tokens > 0 ? custom_image_max_tokens : n_tokens_max) * patch_area;
+        warmup_image_size = static_cast<int>(std::sqrt(image_max_pixels));
+    }
+
+    void set_warmup_n_tokens(int n_tokens) {
+        int n_tok_per_side = static_cast<int>(std::sqrt(n_tokens));
+        GGML_ASSERT(n_tok_per_side * n_tok_per_side == n_tokens && "n_tokens must be n*n");
+        const int cur_merge = n_merge == 0 ? 1 : n_merge;
+        warmup_image_size = n_tok_per_side * patch_size * cur_merge;
+        // TODO: support warmup size for custom token numbers
+    }
+};
+
+struct clip_layer {
+    // attention
+    ggml_tensor * k_w = nullptr;
+    ggml_tensor * k_b = nullptr;
+    ggml_tensor * q_w = nullptr;
+    ggml_tensor * q_b = nullptr;
+    ggml_tensor * v_w = nullptr;
+    ggml_tensor * v_b = nullptr;
+    ggml_tensor * qkv_w = nullptr;
+    ggml_tensor * qkv_b = nullptr;
+
+    ggml_tensor * o_w = nullptr;
+    ggml_tensor * o_b = nullptr;
+
+    ggml_tensor * k_norm = nullptr;
+    ggml_tensor * q_norm = nullptr;
+
+    // layernorm 1
+    ggml_tensor * ln_1_w = nullptr;
+    ggml_tensor * ln_1_b = nullptr;
+
+    ggml_tensor * ff_up_w = nullptr;
+    ggml_tensor * ff_up_b = nullptr;
+    ggml_tensor * ff_gate_w = nullptr;
+    ggml_tensor * ff_gate_b = nullptr;
+    ggml_tensor * ff_down_w = nullptr;
+    ggml_tensor * ff_down_b = nullptr;
+
+    // layernorm 2
+    ggml_tensor * ln_2_w = nullptr;
+    ggml_tensor * ln_2_b = nullptr;
+
+    // layer scale (no bias)
+    ggml_tensor * ls_1_w = nullptr;
+    ggml_tensor * ls_2_w = nullptr;
+
+    // qwen3vl deepstack merger
+    ggml_tensor * deepstack_norm_w = nullptr;
+    ggml_tensor * deepstack_norm_b = nullptr;
+    ggml_tensor * deepstack_fc1_w = nullptr;
+    ggml_tensor * deepstack_fc1_b = nullptr;
+    ggml_tensor * deepstack_fc2_w = nullptr;
+    ggml_tensor * deepstack_fc2_b = nullptr;
+
+    bool has_deepstack() const {
+        return deepstack_fc1_w != nullptr;
+    }
+};
+
+struct clip_model {
+    clip_modality modality = CLIP_MODALITY_VISION;
+    projector_type proj_type = PROJECTOR_TYPE_MLP;
+    clip_hparams hparams;
+
+    // embeddings
+    ggml_tensor * class_embedding = nullptr;
+    ggml_tensor * patch_embeddings_0 = nullptr;
+    ggml_tensor * patch_embeddings_1 = nullptr;  // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL)
+    ggml_tensor * patch_bias = nullptr;
+    ggml_tensor * position_embeddings = nullptr;
+    ggml_tensor * norm_embd_w = nullptr;
+    ggml_tensor * norm_embd_b = nullptr;
+
+    ggml_tensor * pre_ln_w = nullptr;
+    ggml_tensor * pre_ln_b = nullptr;
+
+    std::vector<clip_layer> layers;
+
+    int32_t n_deepstack_layers = 0; // used by Qwen3-VL, calculated from clip_layer
+
+    ggml_tensor * post_ln_w;
+    ggml_tensor * post_ln_b;
+
+    ggml_tensor * projection; // TODO: rename it to fc (fully connected layer)
+    ggml_tensor * mm_fc_w;
+    ggml_tensor * mm_fc_b;
+    ggml_tensor * mm_ffn_up_w = nullptr;
+    ggml_tensor * mm_ffn_up_b = nullptr;
+    ggml_tensor * mm_ffn_gate_w = nullptr;
+    ggml_tensor * mm_ffn_gate_b = nullptr;
+    ggml_tensor * mm_ffn_down_w = nullptr;
+    ggml_tensor * mm_ffn_down_b = nullptr;
+    ggml_tensor * mm_post_norm_w = nullptr;
+    ggml_tensor * mm_post_norm_b = nullptr;
+
+    // LLaVA projection
+    ggml_tensor * mm_input_norm_w = nullptr;
+    ggml_tensor * mm_input_norm_b = nullptr;
+    ggml_tensor * mm_0_w = nullptr;
+    ggml_tensor * mm_0_b = nullptr;
+    ggml_tensor * mm_2_w = nullptr;
+    ggml_tensor * mm_2_b = nullptr;
+
+    ggml_tensor * image_newline = nullptr;
+
+    // Yi type models with mlp+normalization projection
+    ggml_tensor * mm_1_w = nullptr; // Yi type models have 0, 1, 3, 4
+    ggml_tensor * mm_1_b = nullptr;
+    ggml_tensor * mm_3_w = nullptr;
+    ggml_tensor * mm_3_b = nullptr;
+    ggml_tensor * mm_4_w = nullptr;
+    ggml_tensor * mm_4_b = nullptr;
+
+    // GLMV-Edge projection
+    ggml_tensor * mm_model_adapter_conv_w = nullptr;
+    ggml_tensor * mm_model_adapter_conv_b = nullptr;
+
+    // MobileVLM projection
+    ggml_tensor * mm_model_mlp_1_w = nullptr;
+    ggml_tensor * mm_model_mlp_1_b = nullptr;
+    ggml_tensor * mm_model_mlp_3_w = nullptr;
+    ggml_tensor * mm_model_mlp_3_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_0_0_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_0_1_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_0_1_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc1_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc1_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc2_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_1_fc2_b = nullptr;
+    ggml_tensor * mm_model_block_1_block_2_0_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_2_1_w = nullptr;
+    ggml_tensor * mm_model_block_1_block_2_1_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_0_0_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_0_1_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_0_1_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc1_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc1_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc2_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_1_fc2_b = nullptr;
+    ggml_tensor * mm_model_block_2_block_2_0_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_2_1_w = nullptr;
+    ggml_tensor * mm_model_block_2_block_2_1_b = nullptr;
+
+    // MobileVLM_V2 projection
+    ggml_tensor * mm_model_mlp_0_w = nullptr;
+    ggml_tensor * mm_model_mlp_0_b = nullptr;
+    ggml_tensor * mm_model_mlp_2_w = nullptr;
+    ggml_tensor * mm_model_mlp_2_b = nullptr;
+    ggml_tensor * mm_model_peg_0_w = nullptr;
+    ggml_tensor * mm_model_peg_0_b = nullptr;
+
+    // MINICPMV projection
+    ggml_tensor * mm_model_pos_embed_k = nullptr;
+    ggml_tensor * mm_model_query = nullptr;
+    ggml_tensor * mm_model_proj = nullptr;
+    ggml_tensor * mm_model_kv_proj = nullptr;
+    ggml_tensor * mm_model_attn_q_w = nullptr;
+    ggml_tensor * mm_model_attn_q_b = nullptr;
+    ggml_tensor * mm_model_attn_k_w = nullptr;
+    ggml_tensor * mm_model_attn_k_b = nullptr;
+    ggml_tensor * mm_model_attn_v_w = nullptr;
+    ggml_tensor * mm_model_attn_v_b = nullptr;
+    ggml_tensor * mm_model_attn_o_w = nullptr;
+    ggml_tensor * mm_model_attn_o_b = nullptr;
+    ggml_tensor * mm_model_ln_q_w = nullptr;
+    ggml_tensor * mm_model_ln_q_b = nullptr;
+    ggml_tensor * mm_model_ln_kv_w = nullptr;
+    ggml_tensor * mm_model_ln_kv_b = nullptr;
+    ggml_tensor * mm_model_ln_post_w = nullptr;
+    ggml_tensor * mm_model_ln_post_b = nullptr;
+
+    // gemma3
+    ggml_tensor * mm_input_proj_w = nullptr;
+    ggml_tensor * mm_soft_emb_norm_w = nullptr;
+
+    // pixtral, glm4v
+    ggml_tensor * token_embd_img_break = nullptr;
+    ggml_tensor * mm_patch_merger_w = nullptr;
+    ggml_tensor * mm_patch_merger_b = nullptr;
+
+    // ultravox / whisper encoder
+    ggml_tensor * conv1d_1_w = nullptr;
+    ggml_tensor * conv1d_1_b = nullptr;
+    ggml_tensor * conv1d_2_w = nullptr;
+    ggml_tensor * conv1d_2_b = nullptr;
+    ggml_tensor * mm_norm_pre_w = nullptr;
+    ggml_tensor * mm_norm_pre_b = nullptr;
+    ggml_tensor * mm_norm_mid_w = nullptr;
+
+    // cogvlm
+    ggml_tensor * mm_post_fc_norm_w = nullptr;
+    ggml_tensor * mm_post_fc_norm_b = nullptr;
+    ggml_tensor * mm_h_to_4h_w = nullptr;
+    ggml_tensor * mm_gate_w = nullptr;
+    ggml_tensor * mm_4h_to_h_w = nullptr;
+    ggml_tensor * mm_boi = nullptr;
+    ggml_tensor * mm_eoi = nullptr;
+
+    bool audio_has_avgpool() const {
+        return proj_type == PROJECTOR_TYPE_QWEN2A
+            || proj_type == PROJECTOR_TYPE_VOXTRAL;
+    }
+
+    bool audio_has_stack_frames() const {
+        return proj_type == PROJECTOR_TYPE_ULTRAVOX
+            || proj_type == PROJECTOR_TYPE_VOXTRAL;
+    }
+};
+
+const clip_hparams * clip_get_hparams(const struct clip_ctx * ctx);

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

@@ -7,6 +7,8 @@
 
 // !!! Internal header, to be used by mtmd only !!!
 
+#define MTMD_INTERNAL_HEADER
+
 struct clip_ctx;
 
 struct clip_image_size {
@@ -102,7 +104,7 @@ bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct
 
 int clip_is_minicpmv(const struct clip_ctx * ctx);
 bool clip_is_glm(const struct clip_ctx * ctx);
-bool clip_is_qwen2vl(const struct clip_ctx * ctx);
+bool clip_is_mrope(const struct clip_ctx * ctx);
 bool clip_is_llava(const struct clip_ctx * ctx);
 bool clip_is_gemma3(const struct clip_ctx * ctx);
 

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

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

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

@@ -0,0 +1,120 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_glm4v::build() {
+    GGML_ASSERT(model.patch_bias != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+    GGML_ASSERT(model.class_embedding == nullptr);
+
+    const int batch_size = 1;
+
+    norm_type norm_t = NORM_TYPE_RMS;
+
+    ggml_tensor * inp_raw = build_inp_raw();
+    ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
+    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches * 4);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    GGML_ASSERT(img.nx % (patch_size * 2) == 0);
+    GGML_ASSERT(img.ny % (patch_size * 2) == 0);
+
+    // second conv dimension
+    {
+        auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+        inp = ggml_add(ctx0, inp, inp_1);
+
+        inp = ggml_permute(ctx0, inp, 1, 2, 0, 3);  // [w, h, c, b] -> [c, w, h, b]
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+        inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
+        inp = ggml_cont_3d(
+            ctx0, inp,
+            n_embd, n_patches_x * n_patches_y, batch_size);
+    }
+
+    // add patch bias
+    inp = ggml_add(ctx0, inp, model.patch_bias);
+    cb(inp, "patch_bias", -1);
+
+    // pos-conv norm
+    inp = build_norm(inp, model.norm_embd_w, model.norm_embd_b, norm_t, eps, -1);
+
+    // calculate absolute position embedding and apply
+    ggml_tensor * learned_pos_embd = resize_position_embeddings(GGML_SCALE_MODE_BICUBIC);
+    learned_pos_embd = ggml_cont_4d(
+        ctx0, learned_pos_embd,
+        n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+    learned_pos_embd = ggml_reshape_4d(
+        ctx0, learned_pos_embd,
+        n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+    learned_pos_embd = ggml_permute(ctx0, learned_pos_embd, 0, 2, 1, 3);
+    learned_pos_embd = ggml_cont_3d(
+        ctx0, learned_pos_embd,
+        n_embd, n_patches_x * n_patches_y, batch_size);
+    cb(learned_pos_embd, "learned_pos_embd", -1);
+
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        return ggml_rope_multi(
+                    ctx0, cur, positions, nullptr,
+                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION,
+                    32768, hparams.rope_theta, 1, 0, 1, 32, 1);
+    };
+
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            norm_t,
+                            hparams.ffn_op,
+                            learned_pos_embd,
+                            add_pos);
+
+    cb(cur, "vit_out", -1);
+    // cb(ggml_sum(ctx0, cur), "vit_out_sum", -1);
+
+    // GLM4V projector
+    // ref: https://github.com/huggingface/transformers/blob/40dc11cd3eb4126652aa41ef8272525affd4a636/src/transformers/models/glm4v/modeling_glm4v.py#L116-L130
+
+    // patch merger (downsample)
+    {
+        int n_merge = hparams.n_merge;
+        GGML_ASSERT(n_merge > 0);
+
+        int n_token_out = n_patches / n_merge / n_merge;
+        cur = ggml_reshape_4d(ctx0, cur, n_embd, n_merge, n_merge, n_token_out);
+        cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 2, 0, 1, 3)); // [n_merge, n_merge, n_embd, n_token_out]
+        cur = ggml_conv_2d(ctx0, model.mm_patch_merger_w, cur, n_merge, n_merge, 0, 0, 1, 1);
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[2], n_token_out); // [n_embd_out, n_token_out]
+
+        cur = ggml_add(ctx0, cur, model.mm_patch_merger_b);
+    }
+
+    // FC projector
+    {
+        cur = ggml_mul_mat(ctx0, model.projection, cur);
+        // default LayerNorm (post_projection_norm)
+        cur = build_norm(cur, model.mm_post_norm_w, model.mm_post_norm_b, NORM_TYPE_NORMAL, 1e-5, -1);
+        cur = ggml_gelu_erf(ctx0, cur);
+        cb(cur, "after_fc_proj", -1);
+    }
+
+    // FFN projector
+    {
+        cur = build_ffn(cur,
+            model.mm_ffn_up_w, model.mm_ffn_up_b,
+            model.mm_ffn_gate_w, model.mm_ffn_gate_b,
+            model.mm_ffn_down_w, model.mm_ffn_down_b,
+            hparams.ffn_op, -1);
+        cb(cur, "after_ffn_proj", -1);
+        // cb(ggml_sum(ctx0, cur), "merged_sum", -1);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

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

@@ -0,0 +1,69 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_internvl::build() {
+    GGML_ASSERT(model.class_embedding != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+
+    const int n_pos = n_patches + 1;
+    ggml_tensor * inp = build_inp();
+
+    // add CLS token
+    inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+
+    // The larger models use a different ViT, which uses RMS norm instead of layer norm
+    // ref: https://github.com/ggml-org/llama.cpp/pull/13443#issuecomment-2869786188
+    norm_type norm_t = (hparams.n_embd == 3200 && hparams.n_layer == 45)
+        ? NORM_TYPE_RMS // 6B ViT (Used by InternVL 2.5/3 - 26B, 38B, 78B)
+        : NORM_TYPE_NORMAL; // 300M ViT (Used by all smaller InternVL models)
+
+    ggml_tensor * cur = build_vit(
+                            inp, n_pos,
+                            norm_t,
+                            hparams.ffn_op,
+                            model.position_embeddings,
+                            nullptr);
+
+    // remove CLS token
+    cur = ggml_view_2d(ctx0, cur,
+        n_embd, n_patches,
+        ggml_row_size(cur->type, n_embd), 0);
+
+    // pixel shuffle
+    {
+        const int scale_factor = model.hparams.n_merge;
+        const int bsz    = 1; // batch size, always 1 for now since we don't support batching
+        const int height = n_patches_y;
+        const int width  = n_patches_x;
+        GGML_ASSERT(scale_factor > 0);
+        cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, height / scale_factor, width, bsz);
+        cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+        cur = ggml_cont_4d(ctx0, cur,
+            n_embd * scale_factor * scale_factor,
+            height / scale_factor,
+            width / scale_factor,
+            bsz);
+        cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+        // flatten to 2D
+        cur = ggml_cont_2d(ctx0, cur,
+            n_embd * scale_factor * scale_factor,
+            cur->ne[1] * cur->ne[2]);
+    }
+
+    // projector (always using GELU activation)
+    {
+        // projector LayerNorm uses pytorch's default eps = 1e-5
+        // ref: https://huggingface.co/OpenGVLab/InternVL3-8B-Instruct/blob/a34d3e4e129a5856abfd6aa6de79776484caa14e/modeling_internvl_chat.py#L79
+        cur = build_norm(cur, model.mm_0_w, model.mm_0_b, NORM_TYPE_NORMAL, 1e-5, -1);
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_3_w, model.mm_3_b,
+            FFN_GELU,
+            -1);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

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

@@ -0,0 +1,63 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_kimivl::build() {
+    // 2D input positions
+    ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_h, "pos_h");
+    ggml_set_input(pos_h);
+
+    ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_w, "pos_w");
+    ggml_set_input(pos_w);
+
+    ggml_tensor * learned_pos_embd = resize_position_embeddings();
+
+    // build ViT with 2D position embeddings
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        // first half is X axis and second half is Y axis
+        return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);
+    };
+
+    ggml_tensor * inp = build_inp();
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            learned_pos_embd,
+                            add_pos);
+
+    cb(cur, "vit_out", -1);
+
+    {
+        // patch_merger
+        const int scale_factor = model.hparams.n_merge;
+        cur = build_patch_merge_permute(cur, scale_factor);
+
+        // projection norm
+        int proj_inp_dim = cur->ne[0];
+        cur = ggml_view_2d(ctx0, cur,
+            n_embd, cur->ne[1] * scale_factor * scale_factor,
+            ggml_row_size(cur->type, n_embd), 0);
+        cur = ggml_norm(ctx0, cur, 1e-5); // default nn.LayerNorm
+        cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);
+        cur = ggml_add(ctx0, cur, model.mm_input_norm_b);
+        cur = ggml_view_2d(ctx0, cur,
+            proj_inp_dim, cur->ne[1] / scale_factor / scale_factor,
+            ggml_row_size(cur->type, proj_inp_dim), 0);
+        cb(cur, "proj_inp_normed", -1);
+
+        // projection mlp
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU,
+            -1);
+        cb(cur, "proj_out", -1);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

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

@@ -0,0 +1,96 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_llama4::build() {
+    GGML_ASSERT(model.class_embedding != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+
+    const int n_pos = n_patches + 1; // +1 for [CLS]
+
+    // 2D input positions
+    ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+    ggml_set_name(pos_h, "pos_h");
+    ggml_set_input(pos_h);
+
+    ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+    ggml_set_name(pos_w, "pos_w");
+    ggml_set_input(pos_w);
+
+    ggml_tensor * inp = build_inp_raw();
+
+    // Llama4UnfoldConvolution
+    {
+        ggml_tensor * kernel = ggml_reshape_4d(ctx0, model.patch_embeddings_0,
+                                                patch_size, patch_size, 3, n_embd);
+        inp = ggml_im2col(ctx0, kernel, inp, patch_size, patch_size, 0, 0, 1, 1, true, inp->type);
+        inp = ggml_mul_mat(ctx0, model.patch_embeddings_0, inp);
+        inp = ggml_reshape_2d(ctx0, inp, n_embd, n_patches);
+        cb(inp, "patch_conv", -1);
+    }
+
+    // add CLS token
+    inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+
+    // build ViT with 2D position embeddings
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        // first half is X axis and second half is Y axis
+        // ref: https://github.com/huggingface/transformers/blob/40a493c7ed4f19f08eadb0639cf26d49bfa5e180/src/transformers/models/llama4/modeling_llama4.py#L1312
+        // ref: https://github.com/Blaizzy/mlx-vlm/blob/a57156aa87b33cca6e5ee6cfc14dd4ef8f611be6/mlx_vlm/models/llama4/vision.py#L441
+        return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);
+    };
+    ggml_tensor * cur = build_vit(
+                            inp, n_pos,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            model.position_embeddings,
+                            add_pos);
+
+    // remove CLS token
+    cur = ggml_view_2d(ctx0, cur,
+        n_embd, n_patches,
+        ggml_row_size(cur->type, n_embd), 0);
+
+    // pixel shuffle
+    // based on Llama4VisionPixelShuffleMLP
+    // https://github.com/huggingface/transformers/blob/2932f318a20d9e54cc7aea052e040164d85de7d6/src/transformers/models/llama4/modeling_llama4.py#L1151
+    {
+        const int scale_factor = model.hparams.n_merge;
+        const int bsz = 1; // batch size, always 1 for now since we don't support batching
+        GGML_ASSERT(scale_factor > 0);
+        GGML_ASSERT(n_patches_x == n_patches_y); // llama4 only supports square images
+        cur = ggml_reshape_4d(ctx0, cur,
+            n_embd * scale_factor,
+            n_patches_x / scale_factor,
+            n_patches_y,
+            bsz);
+        cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+        cur = ggml_cont_4d(ctx0, cur,
+            n_embd * scale_factor * scale_factor,
+            n_patches_x / scale_factor,
+            n_patches_y / scale_factor,
+            bsz);
+        //cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+        // flatten to 2D
+        cur = ggml_cont_2d(ctx0, cur,
+            n_embd * scale_factor * scale_factor,
+            n_patches / scale_factor / scale_factor);
+        cb(cur, "pixel_shuffle", -1);
+    }
+
+    // based on Llama4VisionMLP2 (always uses GELU activation, no bias)
+    {
+        cur = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, cur);
+        cur = ggml_gelu(ctx0, cur);
+        cur = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, cur);
+        cur = ggml_gelu(ctx0, cur);
+        cb(cur, "adapter_mlp", -1);
+    }
+
+    // Llama4MultiModalProjector
+    cur = ggml_mul_mat(ctx0, model.mm_model_proj, cur);
+    cb(cur, "projected", -1);
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

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

@@ -0,0 +1,374 @@
+#include "models.h"
+
+// this graph is used by llava, granite and glm
+// due to having embedding_stack (used by granite), we cannot reuse build_vit
+ggml_cgraph * clip_graph_llava::build() {
+    const int batch_size = 1;
+    const int n_pos = n_patches + (model.class_embedding ? 1 : 0);
+
+    GGML_ASSERT(n_patches_x == n_patches_y && "only square images supported");
+
+    // Calculate the deepest feature layer based on hparams and projector type
+    int max_feature_layer = n_layer;
+    {
+        // Get the index of the second to last layer; this is the default for models that have a llava projector
+        int il_last = hparams.n_layer - 1;
+        int deepest_feature_layer = -1;
+
+        if (proj_type == PROJECTOR_TYPE_MINICPMV || proj_type == PROJECTOR_TYPE_GLM_EDGE) {
+            il_last += 1;
+        }
+
+        // If we set explicit vision feature layers, only go up to the deepest one
+        // NOTE: only used by granite-vision models for now
+        for (const auto & feature_layer : hparams.vision_feature_layer) {
+            if (feature_layer > deepest_feature_layer) {
+                deepest_feature_layer = feature_layer;
+            }
+        }
+        max_feature_layer = deepest_feature_layer < 0 ? il_last : deepest_feature_layer;
+    }
+
+    ggml_tensor * inp = build_inp();
+
+    // concat class_embeddings and patch_embeddings
+    if (model.class_embedding) {
+        inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+    }
+
+    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    inp = ggml_add(ctx0, inp, ggml_get_rows(ctx0, model.position_embeddings, positions));
+
+    ggml_tensor * inpL = inp;
+
+    // pre-layernorm
+    if (model.pre_ln_w) {
+        inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, NORM_TYPE_NORMAL, eps, -1);
+        cb(inpL, "pre_ln", -1);
+    }
+
+    std::vector<ggml_tensor *> embedding_stack;
+    const auto & vision_feature_layer = hparams.vision_feature_layer;
+
+    // loop over layers
+    for (int il = 0; il < max_feature_layer; il++) {
+        auto & layer = model.layers[il];
+        ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states
+
+        // If this is an embedding feature layer, save the output.
+        // NOTE: 0 index here refers to the input to the encoder.
+        if (vision_feature_layer.find(il) != vision_feature_layer.end()) {
+            embedding_stack.push_back(cur);
+        }
+
+        // layernorm1
+        cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, eps, il);
+        cb(cur, "layer_inp_normed", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = ggml_mul_mat(ctx0, layer.q_w, cur);
+            if (layer.q_b) {
+                Qcur = ggml_add(ctx0, Qcur, layer.q_b);
+            }
+
+            ggml_tensor * Kcur = ggml_mul_mat(ctx0, layer.k_w, cur);
+            if (layer.k_b) {
+                Kcur = ggml_add(ctx0, Kcur, layer.k_b);
+            }
+
+            ggml_tensor * Vcur = ggml_mul_mat(ctx0, layer.v_w, cur);
+            if (layer.v_b) {
+                Vcur = ggml_add(ctx0, Vcur, layer.v_b);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_pos);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(layer.o_w, layer.o_b,
+                Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, inpL);
+
+        inpL = cur; // inpL = residual, cur = hidden_states
+
+        cb(cur, "ffn_inp", il);
+
+        // layernorm2
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, eps, il);
+        cb(cur, "ffn_inp_normed", il);
+
+        // ffn
+        cur = build_ffn(cur,
+            layer.ff_up_w, layer.ff_up_b,
+            layer.ff_gate_w, layer.ff_gate_b,
+            layer.ff_down_w, layer.ff_down_b,
+            hparams.ffn_op, il);
+
+        cb(cur, "ffn_out", il);
+
+        // residual 2
+        cur = ggml_add(ctx0, inpL, cur);
+        cb(cur, "layer_out", il);
+
+        inpL = cur;
+    }
+
+    // post-layernorm
+    if (model.post_ln_w) {
+        inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, NORM_TYPE_NORMAL, eps, -1);
+    }
+
+    ggml_tensor * embeddings = inpL;
+
+    // process vision feature layers (used by granite)
+    {
+        // final layer is a vision feature layer
+        if (vision_feature_layer.find(max_feature_layer) != vision_feature_layer.end()) {
+            embedding_stack.push_back(inpL);
+        }
+
+        // If feature layers are explicitly set, stack them (if we have multiple)
+        if (!embedding_stack.empty()) {
+            embeddings = embedding_stack[0];
+            for (size_t i = 1; i < embedding_stack.size(); i++) {
+                embeddings = ggml_concat(ctx0, embeddings, embedding_stack[i], 0);
+            }
+        }
+    }
+
+    // llava projector (also used by granite)
+    if (hparams.has_llava_projector) {
+        embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
+
+        ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+        ggml_set_name(patches, "patches");
+        ggml_set_input(patches);
+
+        // shape [1, 576, 1024]
+        // ne is whcn, ne = [1024, 576, 1, 1]
+        embeddings = ggml_get_rows(ctx0, embeddings, patches);
+
+        // print_tensor_info(embeddings, "embeddings");
+
+        // llava projector
+        if (proj_type == PROJECTOR_TYPE_MLP) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+
+            embeddings = ggml_gelu(ctx0, embeddings);
+            if (model.mm_2_w) {
+                embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
+                embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+            }
+        }
+        else if (proj_type == PROJECTOR_TYPE_MLP_NORM) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+            // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
+            // First LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w),
+                                model.mm_1_b);
+
+            // GELU activation
+            embeddings = ggml_gelu(ctx0, embeddings);
+
+            // Second linear layer
+            embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_3_b);
+
+            // Second LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w),
+                                model.mm_4_b);
+        }
+        else if (proj_type == PROJECTOR_TYPE_LDP) {
+            // MobileVLM projector
+            int n_patch = 24;
+            ggml_tensor * mlp_1 = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, embeddings);
+            mlp_1 = ggml_add(ctx0, mlp_1, model.mm_model_mlp_1_b);
+            mlp_1 = ggml_gelu(ctx0, mlp_1);
+            ggml_tensor * mlp_3 = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, mlp_1);
+            mlp_3 = ggml_add(ctx0, mlp_3, model.mm_model_mlp_3_b);
+            // mlp_3 shape = [1, 576, 2048], ne = [2048, 576, 1, 1]
+
+            // block 1
+            ggml_tensor * block_1 = nullptr;
+            {
+                // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24]
+                mlp_3 = ggml_permute(ctx0, mlp_3, 1, 0, 2, 3);
+                mlp_3 = ggml_cont_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
+                // stride = 1, padding = 1, bias is nullptr
+                block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);
+
+                // layer norm
+                // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_0_1_w), model.mm_model_block_1_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+
+                // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // hardswish
+                ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+                // block_1_hw shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1], block_1 shape = [1, 2048], ne = [2048, 1, 1, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_2_1_w), model.mm_model_block_1_block_2_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // residual
+                block_1 = ggml_add(ctx0, mlp_3, block_1);
+            }
+
+            // block_2
+            {
+                // stride = 2
+                block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1);
+
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // layer norm
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_0_1_w), model.mm_model_block_2_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // hardswish
+                ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                // not sure the parameters is right for globalAvgPooling
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+
+                // block_1_hw shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1], block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_2_1_w), model.mm_model_block_2_block_2_1_b);
+                block_1 = ggml_reshape_3d(ctx0, block_1, block_1->ne[0], block_1->ne[1] * block_1->ne[2], block_1->ne[3]);
+                // block_1 shape = [1, 144, 2048], ne = [2048, 144, 1]
+            }
+            embeddings = block_1;
+        }
+        else if (proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            int n_patch = 24;
+            ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+            mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b);
+            mlp_0 = ggml_gelu(ctx0, mlp_0);
+            ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0);
+            mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b);
+            // mlp_2 ne = [2048, 576, 1, 1]
+            // // AVG Pool Layer 2*2, strides = 2
+            mlp_2 = ggml_permute(ctx0, mlp_2, 1, 0, 2, 3);
+            // mlp_2 ne = [576, 2048, 1, 1]
+            mlp_2 = ggml_cont_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
+            // mlp_2 ne [24, 24, 2048, 1]
+            mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
+            // weight ne = [3, 3, 2048, 1]
+            ggml_tensor * peg_0 = ggml_conv_2d_dw(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
+            peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b);
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, mlp_2);
+            peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]);
+            embeddings = peg_0;
+        }
+        else {
+            GGML_ABORT("fatal error");
+        }
+    }
+
+    // glm projector
+    else if (proj_type == PROJECTOR_TYPE_GLM_EDGE) {
+        size_t gridsz = (size_t)sqrt(embeddings->ne[1]);
+        embeddings = ggml_permute(ctx0,embeddings,1,0,2,3);
+        embeddings = ggml_cont_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]);
+        embeddings = ggml_conv_2d(ctx0, model.mm_model_adapter_conv_w, embeddings, 2, 2, 0, 0, 1, 1);
+        embeddings = ggml_reshape_3d(ctx0, embeddings,embeddings->ne[0]*embeddings->ne[1] , embeddings->ne[2], batch_size);
+        embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings, 1, 0, 2, 3));
+        embeddings = ggml_add(ctx0, embeddings, model.mm_model_adapter_conv_b);
+        // GLU
+        {
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
+            embeddings = ggml_gelu_inplace(ctx0, embeddings);
+            ggml_tensor * x = embeddings;
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, embeddings);
+            x = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w,x);
+            embeddings = ggml_swiglu_split(ctx0, embeddings, x);
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, embeddings);
+        }
+        // arrangement of BOI/EOI token embeddings
+        // note: these embeddings are not present in text model, hence we cannot process them as text tokens
+        // see: https://huggingface.co/THUDM/glm-edge-v-2b/blob/main/siglip.py#L53
+        {
+            embeddings = ggml_concat(ctx0, model.mm_boi, embeddings, 1); // BOI
+            embeddings = ggml_concat(ctx0, embeddings, model.mm_eoi, 1); // EOI
+        }
+    }
+
+    else {
+        GGML_ABORT("llava: unknown projector type");
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    return gf;
+}

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

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

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

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

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

@@ -0,0 +1,63 @@
+#pragma once
+
+#include "../clip-graph.h"
+
+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;
+};
+
+struct clip_graph_pixtral : clip_graph {
+    clip_graph_pixtral(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_qwen2vl : clip_graph {
+    clip_graph_qwen2vl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_qwen3vl : clip_graph {
+    clip_graph_qwen3vl(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;
+};
+
+struct clip_graph_internvl : clip_graph {
+    clip_graph_internvl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_llama4 : clip_graph {
+    clip_graph_llama4(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_kimivl : clip_graph {
+    clip_graph_kimivl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_cogvlm : clip_graph {
+    clip_graph_cogvlm(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_llava : clip_graph {
+    clip_graph_llava(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
+struct clip_graph_whisper_enc : clip_graph {
+    clip_graph_whisper_enc(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;
+};

llama/llama.cpp/tools/mtmd/models/pixtral.cpp

@@ -0,0 +1,86 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_pixtral::build() {
+    const int n_merge = hparams.n_merge;
+
+    // 2D input positions
+    ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_h, "pos_h");
+    ggml_set_input(pos_h);
+
+    ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_w, "pos_w");
+    ggml_set_input(pos_w);
+
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        return build_rope_2d(ctx0, cur, pos_h, pos_w, hparams.rope_theta, true);
+    };
+
+    ggml_tensor * inp = build_inp();
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            NORM_TYPE_RMS,
+                            hparams.ffn_op,
+                            nullptr, // no learned pos embd
+                            add_pos);
+
+    // mistral small 3.1 patch merger
+    // ref: https://github.com/huggingface/transformers/blob/7a3e208892c06a5e278144eaf38c8599a42f53e7/src/transformers/models/mistral3/modeling_mistral3.py#L67
+    if (model.mm_patch_merger_w) {
+        GGML_ASSERT(hparams.n_merge > 0);
+
+        cur = ggml_mul(ctx0, ggml_rms_norm(ctx0, cur, eps), model.mm_input_norm_w);
+
+        // reshape image tokens to 2D grid
+        cur = ggml_reshape_3d(ctx0, cur, n_embd, n_patches_x, n_patches_y);
+        cur = ggml_permute(ctx0, cur, 2, 0, 1, 3); // [x, y, n_embd]
+        cur = ggml_cont(ctx0, cur);
+
+        // torch.nn.functional.unfold is just an im2col under the hood
+        // we just need a dummy kernel to make it work
+        ggml_tensor * kernel = ggml_view_3d(ctx0, cur, n_merge, n_merge, cur->ne[2], 0, 0, 0);
+        cur = ggml_im2col(ctx0, kernel, cur, n_merge, n_merge, 0, 0, 1, 1, true, inp->type);
+
+        // project to n_embd
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], cur->ne[1] * cur->ne[2]);
+        cur = ggml_mul_mat(ctx0, model.mm_patch_merger_w, cur);
+    }
+
+    // LlavaMultiModalProjector (always using GELU activation)
+    {
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU,
+            -1);
+    }
+
+    // arrangement of the [IMG_BREAK] token
+    if (model.token_embd_img_break) {
+        // not efficient, but works
+        // the trick is to view the embeddings as a 3D tensor with shape [n_embd, n_patches_per_row, n_rows]
+        // and then concatenate the [IMG_BREAK] token to the end of each row, aka n_patches_per_row dimension
+        // after the concatenation, we have a tensor with shape [n_embd, n_patches_per_row + 1, n_rows]
+
+        const int p_y             = n_merge > 0 ? n_patches_y / n_merge : n_patches_y;
+        const int p_x             = n_merge > 0 ? n_patches_x / n_merge : n_patches_x;
+        const int p_total         = p_x * p_y;
+        const int n_embd_text     = cur->ne[0];
+        const int n_tokens_output = p_total + p_y - 1; // one [IMG_BREAK] per row, except the last row
+
+        ggml_tensor * tmp = ggml_reshape_3d(ctx0, cur, n_embd_text, p_x, p_y);
+        ggml_tensor * tok = ggml_new_tensor_3d(ctx0, tmp->type, n_embd_text, 1, p_y);
+        tok = ggml_scale(ctx0, tok, 0.0); // clear the tensor
+        tok = ggml_add(ctx0, tok, model.token_embd_img_break);
+        tmp = ggml_concat(ctx0, tmp, tok, 1);
+        cur = ggml_view_2d(ctx0, tmp,
+            n_embd_text, n_tokens_output,
+            ggml_row_size(tmp->type, n_embd_text), 0);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

llama/llama.cpp/tools/mtmd/models/qwen2vl.cpp

@@ -0,0 +1,183 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_qwen2vl::build() {
+    GGML_ASSERT(model.patch_bias == nullptr);
+    GGML_ASSERT(model.class_embedding == nullptr);
+
+    const int batch_size       = 1;
+    const bool use_window_attn = hparams.n_wa_pattern > 0;
+    const int n_wa_pattern     = hparams.n_wa_pattern;
+    const int n_pos            = n_patches;
+    const int num_position_ids = n_pos * 4; // m-rope requires 4 dim per position
+
+    norm_type norm_t = proj_type == PROJECTOR_TYPE_QWEN25VL
+        ? NORM_TYPE_RMS // qwen 2.5 vl
+        : NORM_TYPE_NORMAL; // qwen 2 vl
+
+    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
+
+    ggml_tensor * inp_raw = build_inp_raw();
+    ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    GGML_ASSERT(img.nx % (patch_size * 2) == 0);
+    GGML_ASSERT(img.ny % (patch_size * 2) == 0);
+
+    // second conv dimension
+    {
+        auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+        inp = ggml_add(ctx0, inp, inp_1);
+
+        inp = ggml_permute(ctx0, inp, 1, 2, 0, 3);  // [w, h, c, b] -> [c, w, h, b]
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+        inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
+        inp = ggml_cont_3d(
+            ctx0, inp,
+            n_embd, n_patches_x * n_patches_y, batch_size);
+    }
+
+    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) {
+        // handle window attention inputs
+        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 ? (il + 1) % n_wa_pattern == 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);
+        cb(cur, "ln1", 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);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // apply M-RoPE
+            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);
+
+            cb(Qcur, "Qcur_rope", il);
+            cb(Kcur, "Kcur_rope", il);
+
+            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);
+            cb(cur, "attn_out", il);
+        }
+
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, inpL);
+
+        inpL = cur; // inpL = residual, cur = hidden_states
+
+        cb(cur, "ffn_inp", il);
+
+        // layernorm2
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il);
+        cb(cur, "ffn_inp_normed", il);
+
+        // ffn
+        cur = build_ffn(cur,
+            layer.ff_up_w, layer.ff_up_b,
+            layer.ff_gate_w, layer.ff_gate_b,
+            layer.ff_down_w, layer.ff_down_b,
+            hparams.ffn_op, il);
+
+        cb(cur, "ffn_out", il);
+
+        // residual 2
+        cur = ggml_add(ctx0, inpL, cur);
+        cb(cur, "layer_out", il);
+
+        inpL = cur;
+    }
+
+    // post-layernorm
+    if (model.post_ln_w) {
+        inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, n_layer);
+    }
+
+    // multimodal projection
+    ggml_tensor * embeddings = inpL;
+    embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size);
+    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);
+
+    if (use_window_attn) {
+        window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4);
+        ggml_set_name(window_idx, "window_idx");
+        ggml_set_input(window_idx);
+
+        // embeddings shape: [n_embd, n_patches_x * n_patches_y, batch_size]
+        GGML_ASSERT(batch_size == 1);
+        embeddings = ggml_reshape_2d(ctx0, embeddings, hparams.projection_dim, n_patches_x * n_patches_y / 4);
+        embeddings = ggml_get_rows(ctx0, embeddings, window_idx);
+        embeddings = ggml_reshape_3d(ctx0, embeddings, hparams.projection_dim, n_patches_x * n_patches_y / 4, batch_size);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    return gf;
+}

llama/llama.cpp/tools/mtmd/models/qwen3vl.cpp

@@ -0,0 +1,191 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_qwen3vl::build() {
+    GGML_ASSERT(model.patch_bias != nullptr);
+    GGML_ASSERT(model.position_embeddings != nullptr);
+    GGML_ASSERT(model.class_embedding == nullptr);
+
+    const int batch_size       = 1;
+    const int n_pos            = n_patches;
+    const int num_position_ids = n_pos * 4; // m-rope requires 4 dim per position
+
+    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_raw = build_inp_raw();
+    ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    GGML_ASSERT(img.nx % (patch_size * 2) == 0);
+    GGML_ASSERT(img.ny % (patch_size * 2) == 0);
+
+    // second conv dimension
+    {
+        auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+        inp = ggml_add(ctx0, inp, inp_1);
+
+        inp = ggml_permute(ctx0, inp, 1, 2, 0, 3);  // [w, h, c, b] -> [c, w, h, b]
+        inp = ggml_cont_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+        inp = ggml_reshape_4d(
+            ctx0, inp,
+            n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+        inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
+        inp = ggml_cont_3d(
+            ctx0, inp,
+            n_embd, n_patches_x * n_patches_y, batch_size);
+    }
+
+    // add patch bias
+    if (model.patch_bias != nullptr) {
+        inp = ggml_add(ctx0, inp, model.patch_bias);
+        cb(inp, "patch_bias", -1);
+    }
+
+    // calculate absolute position embedding and apply
+    ggml_tensor * learned_pos_embd = resize_position_embeddings();
+    learned_pos_embd = ggml_cont_4d(
+        ctx0, learned_pos_embd,
+        n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
+    learned_pos_embd = ggml_reshape_4d(
+        ctx0, learned_pos_embd,
+        n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
+    learned_pos_embd = ggml_permute(ctx0, learned_pos_embd, 0, 2, 1, 3);
+    learned_pos_embd = ggml_cont_3d(
+        ctx0, learned_pos_embd,
+        n_embd, n_patches_x * n_patches_y, batch_size);
+    inp = ggml_add(ctx0, inp, learned_pos_embd);
+    cb(inp, "inp_pos_emb", -1);
+
+    ggml_tensor * inpL = inp;
+
+    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);
+    }
+
+    // deepstack features (stack along the feature dimension), [n_embd * len(deepstack_layers), n_patches_x * n_patches_y, batch_size]
+    ggml_tensor * deepstack_features = nullptr;
+    const int merge_factor = hparams.n_merge > 0 ? hparams.n_merge * hparams.n_merge : 4; // default 2x2=4 for qwen3vl
+
+    // loop over layers
+    for (int il = 0; il < n_layer; il++) {
+        auto & layer = model.layers[il];
+
+        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);
+        cb(cur, "ln1", il);
+
+        // self-attention
+        {
+            cur = ggml_mul_mat(ctx0, layer.qkv_w, cur);
+            cur = ggml_add(ctx0, cur, layer.qkv_b);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos,
+                    /* nb1    */ ggml_row_size(cur->type, d_head),
+                    /* nb2    */ cur->nb[1],
+                    /* offset */ 0);
+
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos,
+                    /* nb1    */ ggml_row_size(cur->type, d_head),
+                    /* nb2    */ cur->nb[1],
+                    /* offset */ ggml_row_size(cur->type, n_embd));
+
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos,
+                    /* nb1    */ ggml_row_size(cur->type, d_head),
+                    /* nb2    */ cur->nb[1],
+                    /* offset */ ggml_row_size(cur->type, 2 * n_embd));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // apply M-RoPE
+            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);
+
+            cb(Qcur, "Qcur_rope", il);
+            cb(Kcur, "Kcur_rope", il);
+
+            cur = build_attn(layer.o_w, layer.o_b,
+                Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, inpL);
+
+        inpL = cur; // inpL = residual, cur = hidden_states
+
+        cb(cur, "ffn_inp", il);
+
+        // layernorm2
+        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il);
+        cb(cur, "ffn_inp_normed", il);
+
+        // ffn
+        cur = build_ffn(cur,
+            layer.ff_up_w, layer.ff_up_b,
+            layer.ff_gate_w, layer.ff_gate_b,
+            layer.ff_down_w, layer.ff_down_b,
+            hparams.ffn_op, il);
+
+        cb(cur, "ffn_out", il);
+
+        // residual 2
+        cur = ggml_add(ctx0, inpL, cur);
+        cb(cur, "layer_out", il);
+
+        if (layer.has_deepstack()) {
+            ggml_tensor * feat = ggml_reshape_3d(ctx0, cur, n_embd * merge_factor, n_pos / merge_factor, batch_size);
+            feat = build_norm(feat, layer.deepstack_norm_w, layer.deepstack_norm_b, norm_t, eps, il);
+            feat = build_ffn(feat,
+                layer.deepstack_fc1_w, layer.deepstack_fc1_b,
+                nullptr, nullptr,
+                layer.deepstack_fc2_w, layer.deepstack_fc2_b,
+                ffn_op_type::FFN_GELU, il);
+
+            if(!deepstack_features) {
+                deepstack_features = feat;
+            } else {
+                // concat along the feature dimension
+                deepstack_features = ggml_concat(ctx0, deepstack_features, feat, 0);
+            }
+        }
+
+        inpL = cur;
+    }
+
+    // post-layernorm
+    if (model.post_ln_w) {
+        inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, n_layer);
+    }
+
+    // multimodal projection
+    ggml_tensor * embeddings = inpL;
+    embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size);
+
+    embeddings = build_ffn(embeddings,
+        model.mm_0_w, model.mm_0_b,
+        nullptr, nullptr,
+        model.mm_1_w, model.mm_1_b,
+        ffn_op_type::FFN_GELU, -1);
+
+    embeddings = ggml_concat(ctx0, embeddings, deepstack_features, 0); // concat along the feature dimension
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    return gf;
+}

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

@@ -0,0 +1,81 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_siglip::build() {
+    ggml_tensor * inp = build_inp();
+
+    ggml_tensor * learned_pos_embd = model.position_embeddings;
+    if (proj_type == PROJECTOR_TYPE_LFM2) {
+        learned_pos_embd = resize_position_embeddings();
+    }
+
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            learned_pos_embd,
+                            nullptr);
+
+    if (proj_type == PROJECTOR_TYPE_GEMMA3) {
+        const int batch_size = 1;
+        GGML_ASSERT(n_patches_x == n_patches_y);
+        const int patches_per_image = n_patches_x;
+        const int kernel_size = hparams.n_merge;
+
+        cur = ggml_transpose(ctx0, cur);
+        cur = ggml_cont_4d(ctx0, cur, patches_per_image, patches_per_image, n_embd, batch_size);
+
+        // doing a pool2d to reduce the number of output tokens
+        cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0);
+        cur = ggml_reshape_3d(ctx0, cur, cur->ne[0] * cur->ne[0], n_embd, batch_size);
+        cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+
+        // apply norm before projection
+        cur = ggml_rms_norm(ctx0, cur, eps);
+        cur = ggml_mul(ctx0, cur, model.mm_soft_emb_norm_w);
+
+        // apply projection
+        cur = ggml_mul_mat(ctx0,
+            ggml_cont(ctx0, ggml_transpose(ctx0, model.mm_input_proj_w)),
+            cur);
+
+    } else if (proj_type == PROJECTOR_TYPE_IDEFICS3) {
+        // pixel_shuffle
+        // https://github.com/huggingface/transformers/blob/0a950e0bbe1ed58d5401a6b547af19f15f0c195e/src/transformers/models/idefics3/modeling_idefics3.py#L578
+        const int scale_factor = model.hparams.n_merge;
+        cur = build_patch_merge_permute(cur, scale_factor);
+        cur = ggml_mul_mat(ctx0, model.projection, cur);
+
+    } else if (proj_type == PROJECTOR_TYPE_LFM2) {
+        // pixel unshuffle block
+        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);
+
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU,
+            -1);
+
+    } else if (proj_type == PROJECTOR_TYPE_JANUS_PRO) {
+        cur = build_ffn(cur,
+            model.mm_0_w, model.mm_0_b,
+            nullptr, nullptr,
+            model.mm_1_w, model.mm_1_b,
+            hparams.ffn_op,
+            -1);
+
+    } else {
+        GGML_ABORT("SigLIP: Unsupported projector type");
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

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

@@ -0,0 +1,106 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_whisper_enc::build() {
+    const int n_frames = img.nx;
+    const int n_pos    = n_frames / 2;
+    GGML_ASSERT(model.position_embeddings->ne[1] >= n_pos);
+
+    ggml_tensor * inp = build_inp_raw(1);
+
+    // conv1d block
+    {
+        // convolution + gelu
+        ggml_tensor * cur = ggml_conv_1d_ph(ctx0, model.conv1d_1_w, inp, 1, 1);
+        cur = ggml_add(ctx0, cur, model.conv1d_1_b);
+
+        cur = ggml_gelu_erf(ctx0, cur);
+
+        cur = ggml_conv_1d_ph(ctx0, model.conv1d_2_w, cur, 2, 1);
+        cur = ggml_add(ctx0, cur, model.conv1d_2_b);
+
+        cur = ggml_gelu_erf(ctx0, cur);
+        // transpose
+        inp = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
+        cb(inp, "after_conv1d", -1);
+    }
+
+    // sanity check (only check one layer, but it should be the same for all)
+    GGML_ASSERT(model.layers[0].ln_1_w && model.layers[0].ln_1_b);
+    GGML_ASSERT(model.layers[0].ln_2_w && model.layers[0].ln_2_b);
+    GGML_ASSERT(model.layers[0].q_b);
+    GGML_ASSERT(model.layers[0].v_b);
+    GGML_ASSERT(!model.layers[0].k_b); // no bias for k
+
+    ggml_tensor * pos_embd_selected = ggml_view_2d(
+        ctx0, model.position_embeddings,
+        model.position_embeddings->ne[0], n_pos,
+        model.position_embeddings->nb[1], 0
+    );
+    ggml_tensor * cur = build_vit(
+                            inp, n_pos,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            pos_embd_selected,
+                            nullptr);
+
+    cb(cur, "after_transformer", -1);
+
+    if (model.audio_has_stack_frames()) {
+        // StackAudioFrames
+        // https://huggingface.co/fixie-ai/ultravox-v0_5-llama-3_2-1b/blob/main/ultravox_model.py
+        cur = build_stack(cur, hparams.proj_stack_factor, n_embd);
+        cb(cur, "after_stacked", -1);
+    }
+
+    if (proj_type == PROJECTOR_TYPE_ULTRAVOX) {
+        // UltravoxProjector
+        // pre-norm
+        cur = ggml_rms_norm(ctx0, cur, 1e-6);
+        cur = ggml_mul(ctx0, cur, model.mm_norm_pre_w);
+
+        // ffn in
+        cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
+
+        // swiglu
+        // see SwiGLU in ultravox_model.py, the second half passed through is silu, not the first half
+        cur = ggml_swiglu_swapped(ctx0, cur);
+
+        // mid-norm
+        cur = ggml_rms_norm(ctx0, cur, 1e-6);
+        cur = ggml_mul(ctx0, cur, model.mm_norm_mid_w);
+
+        // ffn out
+        cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);
+
+    } else if (proj_type == PROJECTOR_TYPE_QWEN2A) {
+        // projector
+        cur = ggml_mul_mat(ctx0, model.mm_fc_w, cur);
+        cur = ggml_add(ctx0, cur, model.mm_fc_b);
+
+    } else if (proj_type == PROJECTOR_TYPE_VOXTRAL) {
+        // 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);
+            cur = ggml_add(ctx0, cur, model.mm_norm_pre_b);
+            cur = build_stack(cur, hparams.proj_stack_factor, n_embd);
+            cur = build_ffn(cur, model.mm_1_w, model.mm_1_b, nullptr, nullptr, model.mm_2_w, model.mm_2_b, hparams.ffn_op, 0);
+            cur = ggml_concat(ctx0, model.mm_boi, cur, 1);
+            cur = ggml_concat(ctx0, cur, model.mm_eoi, 1);
+    } else {
+        GGML_ABORT("%s: unknown projector type", __func__);
+    }
+
+    cb(cur, "projected", -1);
+
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

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

@@ -1,23 +1,15 @@
 #pragma once
 
 #include "ggml.h"
+#include "clip-model.h"
 
 #include <cstdint>
 #include <vector>
 #include <string>
 
-#define WHISPER_ASSERT GGML_ASSERT
+#define MTMD_INTERNAL_HEADER
 
-#define WHISPER_SAMPLE_RATE 16000
+struct mtmd_audio_mel {
-#define WHISPER_N_FFT       400
-#define WHISPER_HOP_LENGTH  160
-#define WHISPER_CHUNK_SIZE  30
-
-#define COMMON_SAMPLE_RATE 16000
-
-namespace whisper_preprocessor {
-
-struct whisper_mel {
     int n_len;
     int n_len_org;
     int n_mel;
@@ -25,23 +17,18 @@ struct whisper_mel {
     std::vector<float> data;
 };
 
-struct whisper_filters {
+struct mtmd_audio_preprocessor {
-    int32_t n_mel;
+    const clip_hparams & hparams;
-    int32_t n_fft;
 
-    std::vector<float> data;
+    mtmd_audio_preprocessor(const clip_ctx * ctx): hparams(*clip_get_hparams(ctx)) {}
+
+    virtual ~mtmd_audio_preprocessor() = default;
+    virtual void initialize() = 0; // NOT thread-safe
+    virtual bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) = 0;
 };
 
-bool preprocess_audio(
+struct mtmd_audio_preprocessor_whisper : mtmd_audio_preprocessor {
-        const float * samples,
+    mtmd_audio_preprocessor_whisper(const clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) {}
-        size_t n_samples,
+    void initialize() override;
-        const whisper_filters & filters,
+    bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) override;
-        std::vector<whisper_mel> & output);
+};
-
-} // namespace whisper_preprocessor
-
-namespace whisper_precalc_filters {
-
-whisper_preprocessor::whisper_filters get_128_bins();
-
-} // namespace whisper_precalc_filters

llama/llama.cpp/tools/mtmd/mtmd-helper.cpp

@@ -32,6 +32,10 @@
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb/stb_image.h"
 
+#ifdef MTMD_INTERNAL_HEADER
+#error "mtmd-helper is a public library outside of mtmd. it must not include internal headers"
+#endif
+
 //
 // internal logging functions
 //

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

@@ -161,8 +161,7 @@ struct mtmd_context {
     // string template for slice image delimiters with row/col (idefics3)
     std::string sli_img_start_tmpl;
 
-    // for whisper, we pre-calculate the mel filter bank
+    std::unique_ptr<mtmd_audio_preprocessor> audio_preproc;
-    whisper_preprocessor::whisper_filters w_filters;
 
     // TODO @ngxson : add timings
 
@@ -228,7 +227,7 @@ struct mtmd_context {
 
     void init_vision() {
         GGML_ASSERT(ctx_v != nullptr);
-        use_mrope = clip_is_qwen2vl(ctx_v);
+        use_mrope = clip_is_mrope(ctx_v);
 
         projector_type proj = clip_get_projector_type(ctx_v);
         int minicpmv_version = clip_is_minicpmv(ctx_v);
@@ -320,6 +319,10 @@ struct mtmd_context {
             img_beg = "<|image_start|>";
             img_end = "<|image_end|>";
 
+        } else if (proj == PROJECTOR_TYPE_GLM4V) {
+            img_beg = "<|begin_of_image|>";
+            img_end = "<|end_of_image|>";
+
         }
     }
 
@@ -327,14 +330,25 @@ struct mtmd_context {
         GGML_ASSERT(ctx_a != nullptr);
         projector_type proj = clip_get_projector_type(ctx_a);
 
-        if (clip_has_whisper_encoder(ctx_a)) {
-            // TODO @ngxson : check if model n_mel is 128 or 80
-            w_filters = whisper_precalc_filters::get_128_bins();
-        }
-
         LOG_WRN("%s: audio input is in experimental stage and may have reduced quality:\n"
                 "    https://github.com/ggml-org/llama.cpp/discussions/13759\n", __func__);
 
+        // set preprocessor
+        switch (proj) {
+            case PROJECTOR_TYPE_QWEN2A:
+            case PROJECTOR_TYPE_QWEN25O:
+            case PROJECTOR_TYPE_ULTRAVOX:
+            case PROJECTOR_TYPE_VOXTRAL:
+                audio_preproc = std::make_unique<mtmd_audio_preprocessor_whisper>(ctx_a);
+                break;
+            default:
+                GGML_ABORT("unsupported audio projector type");
+        }
+
+        // initialize audio preprocessor
+        audio_preproc->initialize();
+
+        // set special tokens
         if (proj == PROJECTOR_TYPE_QWEN2A) {
             // <|audio_bos|> ... (embeddings) ... <|audio_eos|>
             aud_beg = "<|audio_bos|>";
@@ -663,11 +677,10 @@ struct mtmd_tokenizer {
             }
 
             // preprocess audio
-            GGML_ASSERT(ctx->w_filters.n_mel); // make sure we have filter preloaded
+            std::vector<mtmd_audio_mel> mel_spec_chunks;
-            std::vector<whisper_preprocessor::whisper_mel> mel_spec_chunks;
             const float * samples = (const float *)bitmap->data.data();
             size_t n_samples = bitmap->data.size() / sizeof(float);
-            bool ok = whisper_preprocessor::preprocess_audio(samples, n_samples, ctx->w_filters, mel_spec_chunks);
+            bool ok = ctx->audio_preproc->preprocess(samples, n_samples, mel_spec_chunks);
             if (!ok) {
                 LOG_ERR("Unable to preprocess audio\n");
                 return 2;
@@ -873,8 +886,7 @@ int mtmd_get_audio_bitrate(mtmd_context * ctx) {
     if (!ctx->ctx_a) {
         return -1;
     }
-    // for now, we assume that all audio models have the same bitrate
+    return clip_get_hparams(ctx->ctx_a)->audio_sample_rate;
-    return 16000; // 16kHz
 }
 
 //

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

@@ -22,6 +22,11 @@
  *          Issues related to API usage may receive lower priority support.
  *
  * For the usage, see an example in mtmd-cli.cpp
+ *
+ * For contributors:
+ * - 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
  */
 
 #ifdef LLAMA_SHARED

llama/llama.go

@@ -42,6 +42,7 @@ import (
 	_ "github.com/ollama/ollama/llama/llama.cpp/common"
 	_ "github.com/ollama/ollama/llama/llama.cpp/src"
 	_ "github.com/ollama/ollama/llama/llama.cpp/tools/mtmd"
+	_ "github.com/ollama/ollama/llama/llama.cpp/tools/mtmd/models"
 	"github.com/ollama/ollama/ml"
 	ggml "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
 )

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

@@ -23,10 +23,10 @@ problem.
  8 files changed, 21 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index 08681f35e..afde2f0b7 100644
+index 8547ecc84..9f37ca70c 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
-@@ -113,7 +113,6 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
+@@ -112,7 +112,6 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
      if (buffer->iface.free_buffer != NULL) {
          buffer->iface.free_buffer(buffer);
      }
@@ -34,7 +34,7 @@ index 08681f35e..afde2f0b7 100644
  }
  
  size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
-@@ -586,6 +585,7 @@ static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer)
+@@ -591,6 +590,7 @@ static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer)
  
      free(ctx->buffers);
      free(ctx);
@@ -42,7 +42,7 @@ index 08681f35e..afde2f0b7 100644
  }
  
  static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-@@ -2106,6 +2106,11 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+@@ -2125,6 +2125,11 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
  static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      GGML_ASSERT(buffer);
      ggml_aligned_free(buffer->context, buffer->size);
@@ -54,7 +54,7 @@ index 08681f35e..afde2f0b7 100644
  }
  
  static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
-@@ -2158,7 +2163,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
+@@ -2177,7 +2182,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
  };
  
  static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
@@ -64,7 +64,7 @@ index 08681f35e..afde2f0b7 100644
      /* .init_tensor     = */ NULL, // no initialization required
      /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
 diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
-index 81288464c..866758782 100644
+index da624c587..efc63e092 100644
 --- a/ggml/src/ggml-cann/ggml-cann.cpp
 +++ b/ggml/src/ggml-cann/ggml-cann.cpp
 @@ -831,6 +831,7 @@ static bool ggml_backend_buffer_is_cann(ggml_backend_buffer_t buffer) {
@@ -84,7 +84,7 @@ index 81288464c..866758782 100644
  
  /**
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 279679a4e..5145c1e88 100644
+index ab0f6fe9c..6519af435 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -583,6 +583,7 @@ struct ggml_backend_cuda_buffer_context {
@@ -156,10 +156,10 @@ index 18a45d2d9..89041805e 100644
  
  static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
-index 7449a9160..e69a1ff5f 100644
+index e996d98be..84b679315 100644
 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp
 +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
-@@ -355,6 +355,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
+@@ -356,6 +356,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
      ggml_sycl_set_device(ctx->device);
  
      delete ctx;
@@ -167,7 +167,7 @@ index 7449a9160..e69a1ff5f 100644
  }
  catch (sycl::exception const &exc) {
    std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-@@ -816,6 +817,7 @@ struct ggml_backend_sycl_split_buffer_context {
+@@ -817,6 +818,7 @@ struct ggml_backend_sycl_split_buffer_context {
  static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
      delete ctx;
@@ -175,7 +175,7 @@ index 7449a9160..e69a1ff5f 100644
  }
  
  static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1158,6 +1160,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
+@@ -1159,6 +1161,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
  
  static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_sycl_host_free(buffer->context);
@@ -184,10 +184,10 @@ index 7449a9160..e69a1ff5f 100644
  
  static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index c6f5809cc..c801d2fd2 100644
+index 34ec09d40..120191ca0 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -12271,6 +12271,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -12365,6 +12365,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
      ggml_vk_destroy_buffer(ctx->dev_buffer);
      delete ctx;
@@ -195,7 +195,7 @@ index c6f5809cc..c801d2fd2 100644
  }
  
  static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -12414,6 +12415,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
+@@ -12508,6 +12509,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
  static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      VK_LOG_MEMORY("ggml_backend_vk_host_buffer_free_buffer()");
      ggml_vk_host_free(vk_instance.devices[0], buffer->context);

llama/patches/0002-pretokenizer.patch

@@ -10,7 +10,7 @@ logs instead of throwing an error
  1 file changed, 3 insertions(+), 11 deletions(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index e2cca66e4..8246a0a14 100644
+index 7b01a2edf..63250cdf1 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -1825,16 +1825,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
@@ -31,7 +31,7 @@ index e2cca66e4..8246a0a14 100644
                  pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -2014,7 +2005,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+@@ -2015,7 +2006,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                  pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2;
                  clean_spaces = false;
              } else {

llama/patches/0003-clip-unicode.patch

@@ -10,7 +10,7 @@ filesystems for paths that include wide characters
  1 file changed, 39 insertions(+)
 
 diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
-index 3ed08a0fe..6be1470ad 100644
+index 35e3aef0a..84a3796b5 100644
 --- a/tools/mtmd/clip.cpp
 +++ b/tools/mtmd/clip.cpp
 @@ -24,6 +24,19 @@
@@ -32,8 +32,8 @@ index 3ed08a0fe..6be1470ad 100644
 +
  struct clip_logger_state g_logger_state = {clip_log_callback_default, NULL};
  
- enum ffn_op_type {
+ //#define CLIP_DEBUG_FUNCTIONS
-@@ -3257,7 +3270,29 @@ struct clip_model_loader {
+@@ -1619,7 +1632,29 @@ struct clip_model_loader {
          {
              std::vector<uint8_t> read_buf;
  
@@ -63,7 +63,7 @@ index 3ed08a0fe..6be1470ad 100644
              if (!fin) {
                  throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
              }
-@@ -3284,7 +3319,11 @@ struct clip_model_loader {
+@@ -1646,7 +1681,11 @@ struct clip_model_loader {
                      ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
                  }
              }

llama/patches/0004-solar-pro.patch

@@ -6,7 +6,7 @@ Subject: [PATCH] solar-pro
 adds support for the Solar Pro architecture
 ---
  src/CMakeLists.txt         |   1 +
- src/llama-arch.cpp         |  21 +++++
+ src/llama-arch.cpp         |  20 +++++
  src/llama-arch.h           |   3 +
  src/llama-hparams.cpp      |   8 ++
  src/llama-hparams.h        |   5 ++
@@ -15,7 +15,7 @@ adds support for the Solar Pro architecture
  src/llama-model.h          |   3 +
  src/models/models.h        |   5 ++
  src/models/solar.cpp       | 158 +++++++++++++++++++++++++++++++++++++
- 10 files changed, 253 insertions(+), 1 deletion(-)
+ 10 files changed, 252 insertions(+), 1 deletion(-)
  create mode 100644 src/models/solar.cpp
 
 diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
@@ -31,10 +31,10 @@ index 4192af7c0..bd44d73e7 100644
              models/starcoder.cpp
              models/starcoder2.cpp
 diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
-index 64ad1b776..a5fe4f66c 100644
+index 8caf80afc..2ce8ffec0 100644
 --- a/src/llama-arch.cpp
 +++ b/src/llama-arch.cpp
-@@ -85,6 +85,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+@@ -87,6 +87,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
      { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
      { LLM_ARCH_GRANITE_HYBRID,   "granitehybrid"    },
      { LLM_ARCH_CHAMELEON,        "chameleon"        },
@@ -42,7 +42,7 @@ index 64ad1b776..a5fe4f66c 100644
      { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
      { LLM_ARCH_PLM,              "plm"              },
      { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
-@@ -206,6 +207,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+@@ -208,6 +209,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
      { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
      { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
      { LLM_KV_ATTENTION_TEMPERATURE_SCALE,            "%s.attention.temperature_scale"            },
@@ -50,32 +50,38 @@ index 64ad1b776..a5fe4f66c 100644
      { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
      { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
  
-@@ -2025,6 +2027,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+@@ -339,6 +341,7 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
-             { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+     { LLM_TENSOR_ATTN_QKV,                               "blk.%d.attn_qkv" },
-         },
+     { LLM_TENSOR_LAYER_OUT_NORM,                         "blk.%d.layer_output_norm" },
-     },
+     { LLM_TENSOR_ATTN_OUT_NORM,                          "blk.%d.attn_output_norm" },
-+    {
-+        LLM_ARCH_SOLAR,
-+        {
-+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
-+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
-+            { LLM_TENSOR_OUTPUT,          "output" },
-+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
-+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
-+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
-+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
-+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
-+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
-+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
-+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
-+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
 +    { LLM_TENSOR_BSKCN_TV,                               "bskcn_tv" },
-+        },
+     { LLM_TENSOR_POS_EMBD,                               "position_embd" },
-+    },
+     { LLM_TENSOR_FFN_ACT,                                "blk.%d.ffn.act" },
-     {
+     { LLM_TENSOR_TOKEN_EMBD_NORM,                        "token_embd_norm" },
-         LLM_ARCH_WAVTOKENIZER_DEC,
+@@ -2176,6 +2179,22 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
-         {
+             return {
-@@ -2710,6 +2730,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
+                 LLM_TENSOR_TOKEN_EMBD,
+             };
++        case LLM_ARCH_SOLAR:
++            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_OUT,
++                LLM_TENSOR_FFN_NORM,
++                LLM_TENSOR_FFN_GATE,
++                LLM_TENSOR_FFN_DOWN,
++                LLM_TENSOR_FFN_UP,
++                LLM_TENSOR_BSKCN_TV,
++            };
+         default:
+             GGML_ABORT("unknown architecture for tensor mapping");
+     }
+@@ -2344,6 +2363,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
      {LLM_TENSOR_LAUREL_POST_NORM,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      // this tensor is loaded for T5, but never used
      {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
@@ -84,10 +90,10 @@ index 64ad1b776..a5fe4f66c 100644
      {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 diff --git a/src/llama-arch.h b/src/llama-arch.h
-index e11318002..ec9e3a6df 100644
+index 6cbf9b1f8..14d461c76 100644
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
-@@ -89,6 +89,7 @@ enum llm_arch {
+@@ -91,6 +91,7 @@ enum llm_arch {
      LLM_ARCH_GRANITE_MOE,
      LLM_ARCH_GRANITE_HYBRID,
      LLM_ARCH_CHAMELEON,
@@ -95,7 +101,7 @@ index e11318002..ec9e3a6df 100644
      LLM_ARCH_WAVTOKENIZER_DEC,
      LLM_ARCH_PLM,
      LLM_ARCH_BAILINGMOE,
-@@ -210,6 +211,7 @@ enum llm_kv {
+@@ -212,6 +213,7 @@ enum llm_kv {
      LLM_KV_ATTENTION_OUTPUT_SCALE,
      LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
      LLM_KV_ATTENTION_TEMPERATURE_SCALE,
@@ -103,7 +109,7 @@ index e11318002..ec9e3a6df 100644
      LLM_KV_ATTENTION_KEY_LENGTH_MLA,
      LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
  
-@@ -462,6 +464,7 @@ enum llm_tensor {
+@@ -465,6 +467,7 @@ enum llm_tensor {
      LLM_TENSOR_ENC_OUTPUT_NORM,
      LLM_TENSOR_CLS,
      LLM_TENSOR_CLS_OUT,
@@ -112,10 +118,10 @@ index e11318002..ec9e3a6df 100644
      LLM_TENSOR_CONVNEXT_DW,
      LLM_TENSOR_CONVNEXT_NORM,
 diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
-index 8cdbaf69f..41127bf91 100644
+index fe1fa4341..aabff2f06 100644
 --- a/src/llama-hparams.cpp
 +++ b/src/llama-hparams.cpp
-@@ -161,6 +161,14 @@ uint32_t llama_hparams::n_pos_per_embd() const {
+@@ -163,6 +163,14 @@ uint32_t llama_hparams::n_pos_per_embd() const {
      return rope_type == LLAMA_ROPE_TYPE_MROPE || rope_type == LLAMA_ROPE_TYPE_IMROPE ? 4 : 1;
  }
  
@@ -131,10 +137,10 @@ index 8cdbaf69f..41127bf91 100644
      if (il < n_layer) {
          return swa_layers[il];
 diff --git a/src/llama-hparams.h b/src/llama-hparams.h
-index 6eff334a5..a778fc3cf 100644
+index f6e95b5d2..c6e673276 100644
 --- a/src/llama-hparams.h
 +++ b/src/llama-hparams.h
-@@ -64,6 +64,8 @@ struct llama_hparams {
+@@ -65,6 +65,8 @@ struct llama_hparams {
      std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
      std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
  
@@ -143,7 +149,7 @@ index 6eff334a5..a778fc3cf 100644
      uint32_t n_layer_dense_lead = 0;
      uint32_t n_lora_q           = 0;
      uint32_t n_lora_kv          = 0;
-@@ -256,6 +258,9 @@ struct llama_hparams {
+@@ -259,6 +261,9 @@ struct llama_hparams {
  
      uint32_t n_pos_per_embd() const;
  
@@ -154,7 +160,7 @@ index 6eff334a5..a778fc3cf 100644
  
      bool has_kv(uint32_t il) const;
 diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
-index aa3a65f87..ee303bd58 100644
+index ca2ea2461..8916a6242 100644
 --- a/src/llama-model-loader.cpp
 +++ b/src/llama-model-loader.cpp
 @@ -466,7 +466,7 @@ namespace GGUFMeta {
@@ -167,10 +173,10 @@ index aa3a65f87..ee303bd58 100644
  llama_model_loader::llama_model_loader(
          const std::string & fname,
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index 04fccc979..3c503b424 100644
+index ae8207ee1..00cd579e0 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1975,6 +1975,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -1995,6 +1995,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                      default: type = LLM_TYPE_UNKNOWN;
                 }
              } break;
@@ -192,7 +198,7 @@ index 04fccc979..3c503b424 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
-@@ -5401,6 +5416,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -5429,6 +5444,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
  
                          layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
  
@@ -227,7 +233,7 @@ index 04fccc979..3c503b424 100644
                          layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
                          layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
                          layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
-@@ -7480,6 +7523,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
+@@ -7534,6 +7577,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
              {
                  llm = std::make_unique<llm_build_chameleon>(*this, params);
              } break;
@@ -238,7 +244,7 @@ index 04fccc979..3c503b424 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params);
-@@ -7743,6 +7790,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+@@ -7798,6 +7845,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
          case LLM_ARCH_GRANITE_MOE:
          case LLM_ARCH_GRANITE_HYBRID:
          case LLM_ARCH_CHAMELEON:
@@ -247,7 +253,7 @@ index 04fccc979..3c503b424 100644
          case LLM_ARCH_NEO_BERT:
          case LLM_ARCH_SMOLLM3:
 diff --git a/src/llama-model.h b/src/llama-model.h
-index f8342cf2c..cbf4e1bfa 100644
+index c6eb95318..b378b23ec 100644
 --- a/src/llama-model.h
 +++ b/src/llama-model.h
 @@ -76,6 +76,7 @@ enum llm_type {
@@ -258,7 +264,7 @@ index f8342cf2c..cbf4e1bfa 100644
      LLM_TYPE_26B,
      LLM_TYPE_27B,
      LLM_TYPE_30B,
-@@ -404,6 +405,8 @@ struct llama_layer {
+@@ -405,6 +406,8 @@ struct llama_layer {
      struct ggml_tensor * ffn_act_beta    = nullptr;
      struct ggml_tensor * ffn_act_eps     = nullptr;
  
@@ -268,7 +274,7 @@ index f8342cf2c..cbf4e1bfa 100644
  
      struct llama_layer_convnext convnext;
 diff --git a/src/models/models.h b/src/models/models.h
-index 6494f5450..e0aec822c 100644
+index ffb36acc6..6d84a185d 100644
 --- a/src/models/models.h
 +++ b/src/models/models.h
 @@ -515,6 +515,11 @@ struct llm_build_smollm3 : public llm_graph_context {

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

@@ -12,7 +12,7 @@ regex
  2 files changed, 22 insertions(+), 1 deletion(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 8246a0a14..dfba7778b 100644
+index 63250cdf1..dd86a1745 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -299,7 +299,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {

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

@@ -8,10 +8,10 @@ Subject: [PATCH] maintain ordering for rules for grammar
  1 file changed, 1 insertion(+), 1 deletion(-)
 
 diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp
-index c3b4e5d9d..6be552826 100644
+index 2f67c74d7..acf00e2d2 100644
 --- a/common/json-schema-to-grammar.cpp
 +++ b/common/json-schema-to-grammar.cpp
-@@ -310,7 +310,7 @@ private:
+@@ -311,7 +311,7 @@ private:
      friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
      std::function<json(const std::string &)> _fetch_json;
      bool _dotall;

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

@@ -53,7 +53,7 @@ index b165d8bdc..f91d4faba 100644
  }
  
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index dfba7778b..f72f321b9 100644
+index dd86a1745..d63ce9c84 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -1781,9 +1781,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {

llama/patches/0011-ollama-debug-tensor.patch

@@ -8,7 +8,7 @@ Subject: [PATCH] ollama debug tensor
  1 file changed, 6 insertions(+)
 
 diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
-index b468b115a..bb65985b4 100644
+index a59b51893..53891a91f 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.c
 +++ b/ggml/src/ggml-cpu/ggml-cpu.c
 @@ -15,6 +15,8 @@
@@ -20,7 +20,7 @@ index b468b115a..bb65985b4 100644
  #if defined(_MSC_VER) || defined(__MINGW32__)
  #include <malloc.h> // using malloc.h with MSC/MINGW
  #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
-@@ -2928,6 +2930,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
+@@ -2945,6 +2947,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
  
          ggml_compute_forward(&params, node);
  

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

@@ -11,10 +11,10 @@ Subject: [PATCH] graph memory reporting on failure
  4 files changed, 40 insertions(+), 3 deletions(-)
 
 diff --git a/ggml/include/ggml-alloc.h b/ggml/include/ggml-alloc.h
-index 2cb150fd2..7ab3f0192 100644
+index 78aa059dd..7fa8403b3 100644
 --- a/ggml/include/ggml-alloc.h
 +++ b/ggml/include/ggml-alloc.h
-@@ -65,6 +65,7 @@ GGML_API bool ggml_gallocr_reserve_n(
+@@ -72,6 +72,7 @@ GGML_API bool ggml_gallocr_reserve_n(
  GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
  
  GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
@@ -23,10 +23,10 @@ index 2cb150fd2..7ab3f0192 100644
  // Utils
  // Create a buffer and allocate all the tensors in a ggml_context
 diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
-index f1b740785..c54ff98bf 100644
+index 4ed5f3577..a7ebe5dcd 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
-@@ -318,6 +318,7 @@ extern "C" {
+@@ -319,6 +319,7 @@ extern "C" {
  
      GGML_API ggml_backend_buffer_type_t ggml_backend_sched_get_buffer_type(ggml_backend_sched_t sched, ggml_backend_t backend);
      GGML_API size_t                     ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
@@ -35,10 +35,10 @@ index f1b740785..c54ff98bf 100644
      GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
      GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
-index a5995fdc2..dbfd8b5b2 100644
+index 41419b617..73b39bfea 100644
 --- a/ggml/src/ggml-alloc.c
 +++ b/ggml/src/ggml-alloc.c
-@@ -494,6 +494,7 @@ struct node_alloc {
+@@ -485,6 +485,7 @@ struct node_alloc {
  struct ggml_gallocr {
      ggml_backend_buffer_type_t * bufts; // [n_buffers]
      struct vbuffer ** buffers; // [n_buffers]
@@ -46,7 +46,7 @@ index a5995fdc2..dbfd8b5b2 100644
      struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
      int n_buffers;
  
-@@ -517,6 +518,9 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
+@@ -508,6 +509,9 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
      galloc->buffers = calloc(n_bufs, sizeof(struct vbuffer *));
      GGML_ASSERT(galloc->buffers != NULL);
  
@@ -56,7 +56,7 @@ index a5995fdc2..dbfd8b5b2 100644
      galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
      GGML_ASSERT(galloc->buf_tallocs != NULL);
  
-@@ -584,6 +588,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
+@@ -575,6 +579,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
      ggml_hash_set_free(&galloc->hash_set);
      free(galloc->hash_values);
      free(galloc->bufts);
@@ -64,7 +64,7 @@ index a5995fdc2..dbfd8b5b2 100644
      free(galloc->buffers);
      free(galloc->buf_tallocs);
      free(galloc->node_allocs);
-@@ -899,6 +904,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
+@@ -904,6 +909,8 @@ static bool ggml_gallocr_reserve_n_impl(
          }
      }
  
@@ -73,9 +73,9 @@ index a5995fdc2..dbfd8b5b2 100644
      // reallocate buffers if needed
      for (int i = 0; i < galloc->n_buffers; i++) {
          // if the buffer type is used multiple times, we reuse the same buffer
-@@ -933,14 +940,19 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
+@@ -940,15 +947,20 @@ static bool ggml_gallocr_reserve_n_impl(
- #endif
+                 galloc->buffers[i] = NULL;
-             ggml_vbuffer_free(galloc->buffers[i]);
+             } else {
                  galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
 -                if (galloc->buffers[i] == NULL) {
 +                if (galloc->buffers[i]) {
@@ -86,6 +86,7 @@ index a5995fdc2..dbfd8b5b2 100644
 +                    galloc->buffer_sizes[i] = new_size;
 +                    success = false;
                  }
+             }
 +        } else {
 +            galloc->buffer_sizes[i] = ggml_vbuffer_size(galloc->buffers[i]);
          }
@@ -95,8 +96,8 @@ index a5995fdc2..dbfd8b5b2 100644
 +    return success;
  }
  
- bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
+ void ggml_gallocr_reserve_n_size(
-@@ -1095,6 +1107,22 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
+@@ -1118,6 +1130,22 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
      return ggml_vbuffer_size(galloc->buffers[buffer_id]);
  }
  
@@ -120,10 +121,10 @@ index a5995fdc2..dbfd8b5b2 100644
  
  static void free_buffers(ggml_backend_buffer_t ** buffers, const size_t * n_buffers) {
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index afde2f0b7..dbf8486a0 100644
+index 9f37ca70c..1459d16dd 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
-@@ -1840,6 +1840,13 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
+@@ -1859,6 +1859,13 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
      return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
  }
  

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

@@ -10,7 +10,7 @@ Subject: [PATCH] ggml: Export GPU UUIDs
  3 files changed, 63 insertions(+), 6 deletions(-)
 
 diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
-index c54ff98bf..229bf387b 100644
+index a7ebe5dcd..03557bb31 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
 @@ -158,6 +158,7 @@ extern "C" {
@@ -22,7 +22,7 @@ index c54ff98bf..229bf387b 100644
          size_t memory_total;
          // device type
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 5145c1e88..f641c1016 100644
+index 6519af435..c9d3a2b03 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -189,6 +189,51 @@ static int ggml_cuda_parse_id(char devName[]) {
@@ -136,7 +136,7 @@ index 5145c1e88..f641c1016 100644
      props->type        = ggml_backend_cuda_device_get_type(dev);
      props->device_id   = ctx->pci_bus_id.empty() ? nullptr : ctx->pci_bus_id.c_str();
      ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
-@@ -4833,6 +4887,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+@@ -4834,6 +4888,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                  cudaDeviceProp prop;
                  CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                  dev_ctx->description = prop.name;

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

@@ -10,7 +10,7 @@ Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
  2 files changed, 13 insertions(+)
 
 diff --git a/tools/mtmd/mtmd.cpp b/tools/mtmd/mtmd.cpp
-index d06fa42e6..0f5712e21 100644
+index 2638fe4fc..c4e905a4e 100644
 --- a/tools/mtmd/mtmd.cpp
 +++ b/tools/mtmd/mtmd.cpp
 @@ -87,6 +87,16 @@ enum mtmd_slice_tmpl {
@@ -31,10 +31,10 @@ index d06fa42e6..0f5712e21 100644
      return "<__media__>";
  }
 diff --git a/tools/mtmd/mtmd.h b/tools/mtmd/mtmd.h
-index b3df24c29..a6a1af3b8 100644
+index 9f7e861e9..72cec1937 100644
 --- a/tools/mtmd/mtmd.h
 +++ b/tools/mtmd/mtmd.h
-@@ -75,6 +75,9 @@ typedef struct mtmd_input_chunk  mtmd_input_chunk;
+@@ -80,6 +80,9 @@ typedef struct mtmd_input_chunk  mtmd_input_chunk;
  typedef struct mtmd_input_chunks mtmd_input_chunks;
  typedef struct mtmd_input_text   mtmd_input_text;
  

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

@@ -8,10 +8,10 @@ Subject: [PATCH] no power throttling win32 with gnuc
  1 file changed, 1 insertion(+), 1 deletion(-)
 
 diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
-index bb65985b4..47089a62e 100644
+index 53891a91f..8d4851312 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.c
 +++ b/ggml/src/ggml-cpu/ggml-cpu.c
-@@ -2464,7 +2464,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
+@@ -2479,7 +2479,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
          // Newer Windows 11 versions aggresively park (offline) CPU cores and often place
          // all our threads onto the first 4 cores which results in terrible performance with
          // n_threads > 4

llama/patches/0018-ggml-Add-batch-size-hint.patch

@@ -20,7 +20,7 @@ consistent performance.
  8 files changed, 58 insertions(+), 32 deletions(-)
 
 diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
-index 229bf387b..2763f2bd6 100644
+index 03557bb31..93c95602d 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
 @@ -98,7 +98,7 @@ extern "C" {
@@ -40,8 +40,8 @@ index 229bf387b..2763f2bd6 100644
 +    GGML_API void                 ggml_backend_sched_set_batch_size(ggml_backend_sched_t sched, int batch_size);
 +
      // Initialize backend buffers from a measure graph
+     GGML_API void                 ggml_backend_sched_reserve_size(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph, size_t * sizes);
      GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); // returns success
- 
 diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h
 index 6792ba986..0f5b03cef 100644
 --- a/ggml/src/ggml-backend-impl.h
@@ -58,10 +58,10 @@ index 6792ba986..0f5b03cef 100644
          // (optional) event synchronization
          // record an event on this stream
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index dbf8486a0..312ca873c 100644
+index 1459d16dd..498186a7c 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
-@@ -348,14 +348,14 @@ enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_ba
+@@ -353,14 +353,14 @@ enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_ba
  }
  
  enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
@@ -79,7 +79,7 @@ index dbf8486a0..312ca873c 100644
  }
  
  bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-@@ -722,6 +722,8 @@ struct ggml_backend_sched {
+@@ -727,6 +727,8 @@ struct ggml_backend_sched {
  
      bool op_offload;
  
@@ -88,7 +88,7 @@ index dbf8486a0..312ca873c 100644
      int debug;
  
      // used for debugging graph reallocations [GGML_SCHED_DEBUG_REALLOC]
-@@ -820,7 +822,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
+@@ -825,7 +827,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
          if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
              int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
              // check if a backend with higher prio wants to offload the op
@@ -97,7 +97,7 @@ index dbf8486a0..312ca873c 100644
                  for (int b = 0; b < src_backend_id; b++) {
                      if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
                          SET_CAUSE(tensor, "1.off");
-@@ -1572,7 +1574,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
+@@ -1577,7 +1579,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
          }
  
          if (!sched->callback_eval) {
@@ -106,7 +106,7 @@ index dbf8486a0..312ca873c 100644
              if (ec != GGML_STATUS_SUCCESS) {
                  return ec;
              }
-@@ -1594,7 +1596,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
+@@ -1599,7 +1601,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
  
                  struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);
  
@@ -115,7 +115,7 @@ index dbf8486a0..312ca873c 100644
                  if (ec != GGML_STATUS_SUCCESS) {
                      return ec;
                  }
-@@ -1684,6 +1686,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
+@@ -1689,6 +1691,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
  
      sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
      sched->op_offload = op_offload;
@@ -123,7 +123,7 @@ index dbf8486a0..312ca873c 100644
  
      ggml_backend_sched_reset(sched);
  
-@@ -1715,6 +1718,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+@@ -1720,6 +1723,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
      free(sched);
  }
  
@@ -156,7 +156,7 @@ index 5b888cdd8..88d088952 100644
  
  static struct ggml_backend_i blas_backend_i = {
 diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp
-index 3191faaa4..32f14c811 100644
+index f4713a421..92ba577a5 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.cpp
 +++ b/ggml/src/ggml-cpu/ggml-cpu.cpp
 @@ -164,7 +164,7 @@ static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backe
@@ -178,7 +178,7 @@ index 3191faaa4..32f14c811 100644
  
  static const struct ggml_backend_i ggml_backend_cpu_i = {
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index f641c1016..17062697b 100644
+index c9d3a2b03..25548629d 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -2901,7 +2901,7 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
@@ -278,10 +278,10 @@ index 8fc1c2fb5..ba95b4acc 100644
  
  static void ggml_backend_metal_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index c801d2fd2..b2c0d0cee 100644
+index 120191ca0..5349bce24 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -13006,7 +13006,7 @@ static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const stru
+@@ -13099,7 +13099,7 @@ static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const stru
      return num_adds;
  }
  
@@ -290,7 +290,7 @@ index c801d2fd2..b2c0d0cee 100644
      VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
      ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
  
-@@ -13241,6 +13241,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
+@@ -13334,6 +13334,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
      return GGML_STATUS_SUCCESS;
  
      UNUSED(backend);

llama/patches/0019-fix-mtmd-audio.cpp-build-on-windows.patch

@@ -8,7 +8,7 @@ Subject: [PATCH] fix mtmd-audio.cpp build on windows
  1 file changed, 1 insertion(+), 1 deletion(-)
 
 diff --git a/tools/mtmd/mtmd-audio.cpp b/tools/mtmd/mtmd-audio.cpp
-index 4d053895c..84bdc2777 100644
+index f68829a61..2024d3d37 100644
 --- a/tools/mtmd/mtmd-audio.cpp
 +++ b/tools/mtmd/mtmd-audio.cpp
 @@ -1,6 +1,6 @@

llama/patches/0020-ggml-No-alloc-mode.patch

@@ -10,13 +10,13 @@ must be recreated with no-alloc set to false before loading data.
 ---
  ggml/include/ggml-backend.h     |   1 +
  ggml/src/ggml-backend-impl.h    |  16 +++
- ggml/src/ggml-backend.cpp       |  72 +++++++++-
+ ggml/src/ggml-backend.cpp       |  75 ++++++++++-
  ggml/src/ggml-cuda/common.cuh   |  62 ++++++++-
  ggml/src/ggml-cuda/ggml-cuda.cu | 224 ++++++++++++++++++++++++++------
- 5 files changed, 331 insertions(+), 44 deletions(-)
+ 5 files changed, 333 insertions(+), 45 deletions(-)
 
 diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
-index 2763f2bd6..b3b5b356a 100644
+index 93c95602d..dbbb61d9c 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
 @@ -305,6 +305,7 @@ extern "C" {
@@ -75,13 +75,19 @@ index 0f5b03cef..7bdf9d81f 100644
  
      struct ggml_backend {
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index 312ca873c..4092dfe8a 100644
+index 498186a7c..7746e8b92 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
-@@ -41,6 +41,19 @@ ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t
+@@ -36,11 +36,25 @@ const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
-         return ggml_backend_buffer_init(buft, {}, NULL, 0);
  }
  
+ ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+-    GGML_ASSERT(buft);
+     if (size == 0) {
+         // return a dummy buffer for zero-sized allocations
+         return ggml_backend_buffer_init(buft, {}, NULL, 0);
+     }
++
 +    if (buft->no_alloc) {
 +        ggml_backend_buffer_t buf;
 +
@@ -95,10 +101,11 @@ index 312ca873c..4092dfe8a 100644
 +        return buf;
 +    }
 +
-     GGML_ASSERT(buft);
++    GGML_ASSERT(buft);
      return buft->iface.alloc_buffer(buft, size);
  }
-@@ -95,7 +108,8 @@ ggml_backend_buffer_t ggml_backend_buffer_init(
+ 
+@@ -94,7 +108,8 @@ ggml_backend_buffer_t ggml_backend_buffer_init(
          /* .buft      = */ buft,
          /* .context   = */ context,
          /* .size      = */ size,
@@ -108,7 +115,7 @@ index 312ca873c..4092dfe8a 100644
      };
  
      return buffer;
-@@ -127,6 +141,12 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+@@ -126,6 +141,12 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
          return NULL;
      }
  
@@ -118,10 +125,10 @@ index 312ca873c..4092dfe8a 100644
 +        return (void *)ggml_backend_buffer_get_alignment(buffer);
 +    }
 +
-     void * base = buffer->iface.get_base(buffer);
+     // FIXME JG: a multi_buffer has a non-zero size, according to the above comment get_base is not optional,
- 
+     //     I don't know whether the above comment is correct
-     GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
+     if (!buffer->iface.get_base) {
-@@ -731,6 +751,12 @@ struct ggml_backend_sched {
+@@ -736,6 +757,12 @@ struct ggml_backend_sched {
      int debug_realloc;
      int debug_graph_size;
      int debug_prev_graph_size;
@@ -134,7 +141,7 @@ index 312ca873c..4092dfe8a 100644
  };
  
  #define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
-@@ -1630,6 +1656,17 @@ ggml_backend_sched_t ggml_backend_sched_new(
+@@ -1635,6 +1662,17 @@ ggml_backend_sched_t ggml_backend_sched_new(
          size_t graph_size,
          bool parallel,
          bool op_offload) {
@@ -152,7 +159,7 @@ index 312ca873c..4092dfe8a 100644
      GGML_ASSERT(n_backends > 0);
      GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
      GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
-@@ -1682,11 +1719,14 @@ ggml_backend_sched_t ggml_backend_sched_new(
+@@ -1687,11 +1725,14 @@ ggml_backend_sched_t ggml_backend_sched_new(
                  sched->events[b][c] = ggml_backend_event_new(backends[b]->device);
              }
          }
@@ -167,7 +174,7 @@ index 312ca873c..4092dfe8a 100644
  
      ggml_backend_sched_reset(sched);
  
-@@ -1701,6 +1741,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+@@ -1706,6 +1747,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
          for (int c = 0; c < sched->n_copies; c++) {
              ggml_backend_event_free(sched->events[b][c]);
          }
@@ -178,7 +185,7 @@ index 312ca873c..4092dfe8a 100644
      }
      ggml_gallocr_free(sched->galloc);
      ggml_free(sched->ctx);
-@@ -1746,6 +1790,24 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
+@@ -1765,6 +1810,24 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
          return false;
      }
  
@@ -203,7 +210,7 @@ index 312ca873c..4092dfe8a 100644
      ggml_backend_sched_reset(sched);
  
      return true;
-@@ -1851,7 +1913,13 @@ size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched,
+@@ -1870,7 +1933,13 @@ size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched,
      int backend_index = ggml_backend_sched_backend_id(sched, backend);
      GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
  
@@ -219,7 +226,7 @@ index 312ca873c..4092dfe8a 100644
  
  void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
 diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
-index c4529f5d9..8b0fb5d42 100644
+index 9fcb2f9fd..e800ee8f6 100644
 --- a/ggml/src/ggml-cuda/common.cuh
 +++ b/ggml/src/ggml-cuda/common.cuh
 @@ -37,6 +37,41 @@
@@ -264,7 +271,7 @@ index c4529f5d9..8b0fb5d42 100644
  #define STRINGIZE_IMPL(...) #__VA_ARGS__
  #define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
  
-@@ -938,6 +973,9 @@ struct ggml_cuda_pool {
+@@ -941,6 +976,9 @@ struct ggml_cuda_pool {
  
      virtual void * alloc(size_t size, size_t * actual_size) = 0;
      virtual void free(void * ptr, size_t size) = 0;
@@ -274,7 +281,7 @@ index c4529f5d9..8b0fb5d42 100644
  };
  
  template<typename T>
-@@ -1229,11 +1267,15 @@ struct ggml_backend_cuda_context {
+@@ -1232,11 +1270,15 @@ struct ggml_backend_cuda_context {
      // pool
      std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS];
  
@@ -292,7 +299,7 @@ index c4529f5d9..8b0fb5d42 100644
          }
          return *pools[device][curr_stream_no];
      }
-@@ -1241,6 +1283,22 @@ struct ggml_backend_cuda_context {
+@@ -1244,6 +1286,22 @@ struct ggml_backend_cuda_context {
      ggml_cuda_pool & pool() {
          return pool(device);
      }
@@ -316,7 +323,7 @@ index c4529f5d9..8b0fb5d42 100644
  
  struct ggml_cuda_mm_fusion_args_host {
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 17062697b..ede1d089a 100644
+index 25548629d..eeaae3fe4 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -365,6 +365,8 @@ const ggml_cuda_device_info & ggml_cuda_info() {

llama/patches/0021-decode-disable-output_all.patch

@@ -8,10 +8,10 @@ Subject: [PATCH] decode: disable output_all
  1 file changed, 1 insertion(+), 2 deletions(-)
 
 diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index 417140071..87f407f99 100644
+index 8786d4ee3..9e6998272 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
-@@ -999,8 +999,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
+@@ -1051,8 +1051,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
      const int64_t n_vocab = vocab.n_tokens();
      const int64_t n_embd  = hparams.n_embd_inp();
  

llama/patches/0022-ggml-Enable-resetting-backend-devices.patch

@@ -16,7 +16,7 @@ unused then it can be reset to free these data structures.
  6 files changed, 32 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
-index b3b5b356a..69223c488 100644
+index dbbb61d9c..92ca32a4b 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
 @@ -178,6 +178,7 @@ extern "C" {
@@ -43,10 +43,10 @@ index 7bdf9d81f..21b35ac5c 100644
  
      struct ggml_backend_device {
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index 4092dfe8a..a1a19fe51 100644
+index 7746e8b92..189e97170 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
-@@ -526,6 +526,14 @@ ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * par
+@@ -532,6 +532,14 @@ ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * par
      return device->iface.init_backend(device, params);
  }
  
@@ -62,7 +62,7 @@ index 4092dfe8a..a1a19fe51 100644
      GGML_ASSERT(device);
      return device->iface.get_buffer_type(device);
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index ede1d089a..ec63cadab 100644
+index eeaae3fe4..6852d2e20 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -113,6 +113,11 @@ int ggml_cuda_get_device() {
@@ -89,7 +89,7 @@ index ede1d089a..ec63cadab 100644
  
      bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
  #ifdef GGML_CUDA_NO_PEER_COPY
-@@ -4907,6 +4915,11 @@ static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, g
+@@ -4908,6 +4916,11 @@ static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, g
      CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
  }
  
@@ -101,7 +101,7 @@ index ede1d089a..ec63cadab 100644
  static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
      /* .get_name                = */ ggml_backend_cuda_device_get_name,
      /* .get_description         = */ ggml_backend_cuda_device_get_description,
-@@ -4923,6 +4936,7 @@ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
+@@ -4924,6 +4937,7 @@ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
      /* .event_new               = */ ggml_backend_cuda_device_event_new,
      /* .event_free              = */ ggml_backend_cuda_device_event_free,
      /* .event_synchronize       = */ ggml_backend_cuda_device_event_synchronize,
@@ -110,10 +110,10 @@ index ede1d089a..ec63cadab 100644
  
  // backend reg
 diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
-index b7d6edf7f..b987d7aeb 100644
+index 951a88d56..4e162258d 100644
 --- a/ggml/src/ggml-cuda/vendors/hip.h
 +++ b/ggml/src/ggml-cuda/vendors/hip.h
-@@ -45,6 +45,7 @@
+@@ -49,6 +49,7 @@
  #define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
  #define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
  #define cudaDeviceProp hipDeviceProp_t
@@ -122,10 +122,10 @@ index b7d6edf7f..b987d7aeb 100644
  #define cudaError_t hipError_t
  #define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
 diff --git a/src/llama.cpp b/src/llama.cpp
-index ab2e9868a..74c49e651 100644
+index f69964b6d..759152b76 100644
 --- a/src/llama.cpp
 +++ b/src/llama.cpp
-@@ -270,10 +270,12 @@ static struct llama_model * llama_model_load_from_file_impl(
+@@ -921,10 +921,12 @@ static struct llama_model * llama_model_load_from_file_impl(
      for (auto * dev : model->devices) {
          ggml_backend_dev_props props;
          ggml_backend_dev_get_props(dev, &props);

llama/patches/0024-GPU-discovery-enhancements.patch

@@ -28,7 +28,7 @@ fix vulkan PCI ID and ID handling
  create mode 100644 ggml/src/mem_nvml.cpp
 
 diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
-index 69223c488..6510e0cba 100644
+index 92ca32a4b..6ad583f09 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
 @@ -169,6 +169,12 @@ extern "C" {
@@ -58,7 +58,7 @@ index d55aed348..99ae293cc 100644
  
  set_target_properties(ggml-base PROPERTIES
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index ec63cadab..cd71902df 100644
+index 6852d2e20..48cdb1dcf 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -267,6 +267,16 @@ static ggml_cuda_device_info ggml_cuda_init() {
@@ -159,7 +159,7 @@ index ec63cadab..cd71902df 100644
      bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
  #ifdef GGML_CUDA_NO_PEER_COPY
      bool events = false;
-@@ -5046,6 +5102,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+@@ -5047,6 +5103,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
          std::lock_guard<std::mutex> lock(mutex);
          if (!initialized) {
              ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
@@ -167,7 +167,7 @@ index ec63cadab..cd71902df 100644
  
              for (int i = 0; i < ggml_cuda_info().device_count; i++) {
                  ggml_backend_cuda_device_context * dev_ctx = new ggml_backend_cuda_device_context;
-@@ -5061,6 +5118,14 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+@@ -5062,6 +5119,14 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                  snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.0", prop.pciDomainID, prop.pciBusID, prop.pciDeviceID);
                  dev_ctx->pci_bus_id = pci_bus_id;
  
@@ -183,7 +183,7 @@ index ec63cadab..cd71902df 100644
                      /* .iface   = */ ggml_backend_cuda_device_interface,
                      /* .reg     = */ &reg,
 diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
-index b987d7aeb..5ad5623ae 100644
+index 4e162258d..d89e35a8e 100644
 --- a/ggml/src/ggml-cuda/vendors/hip.h
 +++ b/ggml/src/ggml-cuda/vendors/hip.h
 @@ -5,6 +5,8 @@
@@ -195,7 +195,7 @@ index b987d7aeb..5ad5623ae 100644
  
  #if defined(GGML_HIP_ROCWMMA_FATTN)
  #include <rocwmma/rocwmma-version.hpp>
-@@ -47,6 +49,7 @@
+@@ -51,6 +53,7 @@
  #define cudaDeviceProp hipDeviceProp_t
  #define cudaDeviceReset hipDeviceReset
  #define cudaDeviceSynchronize hipDeviceSynchronize
@@ -243,7 +243,7 @@ index ba95b4acc..f6f8f7a10 100644
          /* .async                 = */ true,
          /* .host_buffer           = */ false,
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index b2c0d0cee..d9f4d34f5 100644
+index 5349bce24..d43d46d1d 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 @@ -236,6 +236,7 @@ class vk_memory_logger;
@@ -254,7 +254,7 @@ index b2c0d0cee..d9f4d34f5 100644
  
  static constexpr uint32_t mul_mat_vec_max_cols = 8;
  static constexpr uint32_t p021_max_gqa_ratio = 8;
-@@ -12256,6 +12257,29 @@ static void ggml_vk_get_device_description(int device, char * description, size_
+@@ -12350,6 +12351,29 @@ static void ggml_vk_get_device_description(int device, char * description, size_
      snprintf(description, description_size, "%s", props.deviceName.data());
  }
  
@@ -284,7 +284,7 @@ index b2c0d0cee..d9f4d34f5 100644
  // backend interface
  
  #define UNUSED GGML_UNUSED
-@@ -13535,15 +13559,72 @@ void ggml_backend_vk_get_device_description(int device, char * description, size
+@@ -13628,15 +13652,72 @@ void ggml_backend_vk_get_device_description(int device, char * description, size
      ggml_vk_get_device_description(dev_idx, description, description_size);
  }
  
@@ -361,7 +361,7 @@ index b2c0d0cee..d9f4d34f5 100644
  
      if (membudget_supported) {
          memprops.pNext = &budgetprops;
-@@ -13595,8 +13676,13 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
+@@ -13688,8 +13769,13 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
          }
      }
  
@@ -376,7 +376,7 @@ index b2c0d0cee..d9f4d34f5 100644
      }
  
      vk::PhysicalDeviceProperties2 props = {};
-@@ -13613,19 +13699,24 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
+@@ -13706,19 +13792,24 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
  
      char pci_bus_id[16] = {};
      snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.%x", pci_domain, pci_bus, pci_device, pci_function);
@@ -410,7 +410,7 @@ index b2c0d0cee..d9f4d34f5 100644
  
  static const char * ggml_backend_vk_device_get_name(ggml_backend_dev_t dev) {
      ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
-@@ -13637,9 +13728,14 @@ static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t de
+@@ -13730,9 +13821,14 @@ static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t de
      return ctx->description.c_str();
  }
  
@@ -426,7 +426,7 @@ index b2c0d0cee..d9f4d34f5 100644
  }
  
  static ggml_backend_buffer_type_t ggml_backend_vk_device_get_buffer_type(ggml_backend_dev_t dev) {
-@@ -13663,8 +13759,9 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
+@@ -13756,8 +13852,9 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
  
      props->name        = ggml_backend_vk_device_get_name(dev);
      props->description = ggml_backend_vk_device_get_description(dev);
@@ -437,7 +437,7 @@ index b2c0d0cee..d9f4d34f5 100644
      ggml_backend_vk_device_get_memory(dev, &props->memory_free, &props->memory_total);
      props->caps = {
          /* .async                 = */ false,
-@@ -13672,6 +13769,13 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
+@@ -13765,6 +13862,13 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
          /* .buffer_from_host_ptr  = */ false,
          /* .events                = */ false,
      };
@@ -451,7 +451,7 @@ index b2c0d0cee..d9f4d34f5 100644
  }
  
  static ggml_backend_t ggml_backend_vk_device_init(ggml_backend_dev_t dev, const char * params) {
-@@ -14236,6 +14340,8 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+@@ -14331,6 +14435,8 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
          static std::mutex mutex;
          std::lock_guard<std::mutex> lock(mutex);
          if (!initialized) {
@@ -460,7 +460,7 @@ index b2c0d0cee..d9f4d34f5 100644
              for (int i = 0; i < ggml_backend_vk_get_device_count(); i++) {
                  ggml_backend_vk_device_context * ctx = new ggml_backend_vk_device_context;
                  char desc[256];
-@@ -14244,12 +14350,41 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+@@ -14339,12 +14445,41 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
                  ctx->name = GGML_VK_NAME + std::to_string(i);
                  ctx->description = desc;
                  ctx->is_integrated_gpu = ggml_backend_vk_get_device_type(i) == vk::PhysicalDeviceType::eIntegratedGpu;

llama/patches/0028-Add-memory-detection-using-DXGI-PDH.patch

@@ -38,7 +38,7 @@ index 1c07e767a..0da3e065b 100644
  #ifdef __cplusplus
  }
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index d9f4d34f5..8a83427fb 100644
+index d43d46d1d..df79f9f79 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 @@ -74,6 +74,7 @@ DispatchLoaderDynamic & ggml_vk_default_dispatcher();
@@ -49,7 +49,7 @@ index d9f4d34f5..8a83427fb 100644
  
  typedef struct VkPhysicalDeviceShaderBfloat16FeaturesKHR {
      VkStructureType                       sType;
-@@ -13576,6 +13577,7 @@ struct ggml_backend_vk_device_context {
+@@ -13669,6 +13670,7 @@ struct ggml_backend_vk_device_context {
      std::string pci_id;
      std::string id;
      std::string uuid;
@@ -57,7 +57,7 @@ index d9f4d34f5..8a83427fb 100644
      int major;
      int minor;
      int driver_major;
-@@ -13594,6 +13596,20 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
+@@ -13687,6 +13689,20 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
      
      vk::PhysicalDeviceProperties2 props2;
      vkdev.getProperties2(&props2);
@@ -78,7 +78,7 @@ index d9f4d34f5..8a83427fb 100644
  
      if (!is_integrated_gpu)
      {
-@@ -13625,7 +13641,6 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
+@@ -13718,7 +13734,6 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
      }
      // else fallback to memory budget if supported
  
@@ -86,7 +86,7 @@ index d9f4d34f5..8a83427fb 100644
      if (membudget_supported) {
          memprops.pNext = &budgetprops;
      }
-@@ -14357,7 +14372,6 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+@@ -14452,7 +14467,6 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
                      /* .reg     = */ reg,
                      /* .context = */ ctx,
                  });
@@ -94,7 +94,7 @@ index d9f4d34f5..8a83427fb 100644
                  // Gather additional information about the device
                  int dev_idx = vk_instance.device_indices[i];
                  vk::PhysicalDeviceProperties props1;
-@@ -14380,6 +14394,14 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+@@ -14475,6 +14489,14 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
                      }
                  }
                  ctx->uuid = oss.str();

llama/patches/0029-ggml-cuda-skip-large-batches.patch

@@ -10,10 +10,10 @@ fallback to cpu
  1 file changed, 3 insertions(+)
 
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index cd71902df..d69d62193 100644
+index 48cdb1dcf..3102d7ea7 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -4632,6 +4632,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
+@@ -4633,6 +4633,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                  if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
                      return false;
                  }

llama/patches/0030-win-exit-instead-of-abort.patch

@@ -8,7 +8,7 @@ Subject: [PATCH] win: exit instead of abort
  1 file changed, 6 insertions(+), 1 deletion(-)
 
 diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
-index 530ff7b95..fc0196eb7 100644
+index eb3ae72ea..c9242a15a 100644
 --- a/ggml/src/ggml.c
 +++ b/ggml/src/ggml.c
 @@ -250,8 +250,13 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) {

llama/patches/0031-fix-bakllava-regression.patch

@@ -9,10 +9,10 @@ Rever to prior logic of assuming an empty projector type is mlp
  1 file changed, 4 insertions(+)
 
 diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
-index 6be1470ad..2a325c726 100644
+index 84a3796b5..d3a37842d 100644
 --- a/tools/mtmd/clip.cpp
 +++ b/tools/mtmd/clip.cpp
-@@ -2649,6 +2649,10 @@ struct clip_model_loader {
+@@ -960,6 +960,10 @@ struct clip_model_loader {
              if (proj_type.empty()) {
                  if (modality == CLIP_MODALITY_VISION) {
                      get_string(KEY_VISION_PROJ_TYPE, proj_type, false);

llama/sampling_ext.cpp

@@ -72,7 +72,7 @@ struct llama_vocab * llama_load_vocab_from_file(const char * fname) {
     try {
         const auto kv = LLM_KV(LLM_ARCH_UNKNOWN);
         std::vector<std::string> splits = {};
-        llama_model_loader ml(std::string(fname), splits, false, false, nullptr, nullptr);
+        llama_model_loader ml(std::string(fname), splits, false, false, false, nullptr, nullptr);
         vocab->load(ml, kv);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());

llm/server.go

@@ -143,7 +143,7 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 	var llamaModel *llama.Model
 	var textProcessor model.TextProcessor
 	var err error
-	if envconfig.NewEngine(true) || f.KV().OllamaEngineRequired() {
+	if envconfig.NewEngine() || f.KV().OllamaEngineRequired() {
 		if len(projectors) == 0 {
 			textProcessor, err = model.NewTextProcessor(modelPath)
 		} else {

ml/backend.go

@@ -54,10 +54,6 @@ type CacheConfig struct {
 	// MaskDType specifies the data type for generating the mask. If unset it will
 	// default to DTypeF32.
 	MaskDType DType
-
-	// MaskBatchPadding specifies the multiple for the batch size dimension in the mask.
-	// Any position that does not correspond to an actual token will be filled with -Inf.
-	MaskBatchPadding int
 }
 
 // BackendParams controls how the backend loads and executes models

ml/backend/ggml/ggml.go

@@ -685,7 +685,7 @@ func (b *Backend) NewContextSize(n int) ml.Context {
 
 func (b *Backend) CacheConfig() ml.CacheConfig {
 	if b.flashAttention == ml.FlashAttentionEnabled {
-		return ml.CacheConfig{CachePadding: 256, MaskDType: ml.DTypeF16, MaskBatchPadding: C.GGML_KQ_MASK_PAD}
+		return ml.CacheConfig{CachePadding: 256, MaskDType: ml.DTypeF16}
 	} else {
 		return ml.CacheConfig{CachePadding: 256, PermutedV: true}
 	}
@@ -1534,7 +1534,8 @@ func (t *Tensor) RoPE(ctx ml.Context, positions ml.Tensor, ropeDim int, ropeBase
 			unsafe.SliceData(mropeSections),
 			C.int(opts.Type),
 			cmp.Or(C.int(opts.YaRN.OriginalContextLength), 128<<10),
-			C.float(ropeBase), C.float(ropeScale),
+			C.float(ropeBase),
+			C.float(ropeScale),
 			C.float(opts.YaRN.ExtrapolationFactor),
 			cmp.Or(C.float(opts.YaRN.AttentionFactor), 1),
 			cmp.Or(C.float(opts.YaRN.BetaFast), 32),
@@ -1546,9 +1547,11 @@ func (t *Tensor) RoPE(ctx ml.Context, positions ml.Tensor, ropeDim int, ropeBase
 			dequant,
 			positions.(*Tensor).t,
 			opts.Factors.(*Tensor).t,
-			C.int(ropeDim), C.int(opts.Type),
+			C.int(ropeDim),
+			C.int(opts.Type),
 			cmp.Or(C.int(opts.YaRN.OriginalContextLength), 128<<10),
-			C.float(ropeBase), C.float(ropeScale),
+			C.float(ropeBase),
+			C.float(ropeScale),
 			C.float(opts.YaRN.ExtrapolationFactor),
 			cmp.Or(C.float(opts.YaRN.AttentionFactor), 1),
 			cmp.Or(C.float(opts.YaRN.BetaFast), 32),
@@ -1657,11 +1660,6 @@ func (t *Tensor) ScaledDotProductAttention(ctx ml.Context, key, value, mask, sin
 		}
 
 		if mask != nil {
-			padSize := int(pad(C.size_t(mask.Dim(1)), C.size_t(cacheConfig.MaskBatchPadding))) - mask.Dim(1)
-			if padSize > 0 {
-				mask = mask.Pad(ctx, 0, padSize, 0, 0)
-			}
-
 			if mask.DType() != cacheConfig.MaskDType {
 				mask = mask.Cast(ctx, cacheConfig.MaskDType)
 			}

ml/backend/ggml/ggml/include/ggml-alloc.h

@@ -53,7 +53,14 @@ GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
 // call with a worst-case graph to avoid buffer reallocations
 // not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
 // returns false if the buffer allocation failed
+// ggml_gallocr_resrve_n_size writes the buffer sizes per galloc buffer that would be allocated by ggml_gallocr_reserve_n to sizes
 GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+GGML_API void ggml_gallocr_reserve_n_size(
+    ggml_gallocr_t galloc,
+    struct ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids,
+    size_t * sizes);
 GGML_API bool ggml_gallocr_reserve_n(
     ggml_gallocr_t galloc,
     struct ggml_cgraph * graph,
@@ -69,6 +76,8 @@ GGML_API size_t ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, in
 
 // Utils
 // Create a buffer and allocate all the tensors in a ggml_context
+// ggml_backend_alloc_ctx_tensors_from_buft_size returns the size of the buffer that would be allocated by ggml_backend_alloc_ctx_tensors_from_buft
+GGML_API size_t                       ggml_backend_alloc_ctx_tensors_from_buft_size(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
 GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
 GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);