Merge branch 'master' into xsn/lora_keep_track

* mtmd: fix use_non_causal being reported incorrectly

* move clip_is_mrope to mtmd_decode_use_mrope

* fix sloppy code ggml_cpy

CONTRIBUTING.md

@@ -20,7 +20,7 @@ If AI is used to generate any portion of the code, contributors must adhere to t
 1. Explicitly disclose the manner in which AI was employed.
 2. Perform a comprehensive manual review prior to submitting the pull request.
 3. Be prepared to explain every line of code they submitted when asked about it by a maintainer.
-4. Using AI to respond to human reviewers is strictly prohibited.
+4. Using AI to write pull request descriptions or to respond to human reviewers is strictly prohibited.
 
 For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 

include/llama-cpp.h

@@ -21,7 +21,9 @@ struct llama_sampler_deleter {
 };
 
 struct llama_adapter_lora_deleter {
-    void operator()(llama_adapter_lora * adapter) { llama_adapter_lora_free(adapter); }
+    void operator()(llama_adapter_lora *) {
+        // llama_adapter_lora_free is deprecated
+    }
 };
 
 typedef std::unique_ptr<llama_model, llama_model_deleter> llama_model_ptr;

include/llama.h

@@ -646,7 +646,8 @@ extern "C" {
 
     // Manually free a LoRA adapter
     // NOTE: loaded adapters will be free when the associated model is deleted
-    LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter);
+    LLAMA_API DEPRECATED(void llama_adapter_lora_free(struct llama_adapter_lora * adapter),
+            "adapters are now freed together with the associated model");
 
     // Get the invocation tokens if the current lora is an alora
     LLAMA_API uint64_t            llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter);

src/llama-adapter.cpp

@@ -146,11 +146,9 @@ llama_adapter_lora_weight * llama_adapter_lora::get_weight(ggml_tensor * w) {
     return nullptr;
 }
 
-static void llama_adapter_lora_init_impl(const char * path_lora, llama_adapter_lora & adapter) {
+static void llama_adapter_lora_init_impl(llama_model & model, const char * path_lora, llama_adapter_lora & adapter) {
     LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
 
-    llama_model & model = adapter.model;
-
     ggml_context * ctx_init;
     gguf_init_params meta_gguf_params = {
         /* .no_alloc = */ true,
@@ -413,17 +411,17 @@ static void llama_adapter_lora_init_impl(const char * path_lora, llama_adapter_l
         }
     }
 
-    // update number of nodes used
-    model.n_lora_nodes += adapter.get_n_nodes();
+    // register adapter with model
+    model.loras.insert(&adapter);
 
     LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
 }
 
 llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * path_lora) {
-    llama_adapter_lora * adapter = new llama_adapter_lora(*model);
+    llama_adapter_lora * adapter = new llama_adapter_lora();
 
     try {
-        llama_adapter_lora_init_impl(path_lora, *adapter);
+        llama_adapter_lora_init_impl(*model, path_lora, *adapter);
         return adapter;
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
@@ -473,12 +471,8 @@ int32_t llama_adapter_meta_val_str_by_index(const llama_adapter_lora * adapter,
     return snprintf(buf, buf_size, "%s", it->second.c_str());
 }
 
-void llama_adapter_lora_free(llama_adapter_lora * adapter) {
-    // update number of nodes used
-    GGML_ASSERT(adapter->model.n_lora_nodes >= adapter->get_n_nodes());
-    adapter->model.n_lora_nodes -= adapter->get_n_nodes();
-
-    delete adapter;
+void llama_adapter_lora_free(llama_adapter_lora *) {
+    // deprecated: adapters are freed by llama_model's destructor
 }
 
 uint64_t llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter) {

src/llama-adapter.h

@@ -59,8 +59,6 @@ struct llama_adapter_lora_weight {
 };
 
 struct llama_adapter_lora {
-    llama_model & model;
-
     // map tensor name to lora_a_b
     std::unordered_map<std::string, llama_adapter_lora_weight> ab_map;
 
@@ -75,7 +73,7 @@ struct llama_adapter_lora {
     // activated lora (aLoRA)
     std::vector<llama_token> alora_invocation_tokens;
 
-    llama_adapter_lora(llama_model & model) : model(model) {}
+    llama_adapter_lora() = default;
     ~llama_adapter_lora() = default;
 
     llama_adapter_lora_weight * get_weight(ggml_tensor * w);

src/llama-context.cpp

@@ -1955,7 +1955,9 @@ uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
         return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
     }
     uint32_t res = std::max<uint32_t>(1024u, 8u*model.n_tensors());
-    res += model.n_lora_nodes;
+    for (const auto & lora : model.loras) {
+        res += lora->get_n_nodes();
+    }
     return res;
 }
 

src/llama-model.cpp

@@ -468,7 +468,11 @@ llama_model::llama_model(const llama_model_params & params) : params(params), pi
     pimpl->has_tensor_overrides = params.tensor_buft_overrides && params.tensor_buft_overrides[0].pattern;
 }
 
-llama_model::~llama_model() = default;
+llama_model::~llama_model() {
+    for (auto * lora : loras) {
+        delete lora;
+    }
+}
 
 void llama_model::load_stats(llama_model_loader & ml) {
     pimpl->n_elements = ml.n_elements;

src/llama-model.h

@@ -11,6 +11,7 @@
 #include <memory>
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include <vector>
 
 struct llama_cparams;
@@ -476,8 +477,8 @@ struct llama_model {
     // for quantize-stats only
     std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
 
-    // for keeping track of extra nodes used by lora adapters
-    uint32_t n_lora_nodes = 0;
+    // for keeping track of associated LoRA adapters
+    std::unordered_set<llama_adapter_lora *> loras;
 
     int64_t t_load_us  = 0;
     int64_t t_start_us = 0;

src/models/gemma3n-iswa.cpp

@@ -258,12 +258,12 @@ ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
         res->add_input(std::move(inp));
     } else {
         // Vision embedding path: use padding token (ID=0) embedding
+        // TODO: verify if this is the correct behavior in transformers implementation
         const int64_t embd_size = model.tok_embd_per_layer->ne[0];  // n_embd_altup * n_layer
 
-        // Extract and dequantize padding token embedding (column 0)
-        ggml_tensor * padding_q = ggml_view_1d(ctx0, model.tok_embd_per_layer, embd_size, 0);
-        ggml_tensor * padding_f32 = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, embd_size);
-        inp_per_layer = ggml_cpy(ctx0, padding_q, padding_f32);
+        // Extract and dequantize padding token embedding (row 0)
+        ggml_tensor * padding = ggml_view_1d(ctx0, model.tok_embd_per_layer, embd_size, 0);
+        inp_per_layer = ggml_cast(ctx0, padding, GGML_TYPE_F32);
 
         // Reshape to [n_embd_altup, n_layer, 1]
         inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, 1);

tools/mtmd/clip.cpp

@@ -3808,18 +3808,6 @@ bool clip_is_glm(const struct clip_ctx * ctx) {
     return ctx->proj_type() == PROJECTOR_TYPE_GLM_EDGE;
 }
 
-bool clip_is_mrope(const struct clip_ctx * ctx) {
-    switch (ctx->proj_type()) {
-        case PROJECTOR_TYPE_QWEN2VL:
-        case PROJECTOR_TYPE_QWEN25VL:
-        case PROJECTOR_TYPE_QWEN3VL:
-        case PROJECTOR_TYPE_GLM4V:
-            return true;
-        default:
-            return false;
-    }
-}
-
 bool clip_is_llava(const struct clip_ctx * ctx) {
     return ctx->model.hparams.has_llava_projector;
 }

tools/mtmd/clip.h

@@ -104,7 +104,6 @@ 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_mrope(const struct clip_ctx * ctx);
 bool clip_is_llava(const struct clip_ctx * ctx);
 // note for contributor: this clip_is_(model) pattern is deprecated
 //                       do NOT add new functions like this

tools/mtmd/mtmd.cpp

@@ -146,8 +146,6 @@ struct mtmd_context {
     bool        tok_row_end_trail = false;
     bool        ov_img_first      = false;
 
-    bool use_mrope = false; // for Qwen2VL, we need to use M-RoPE
-
     // string template for slice image delimiters with row/col (idefics3)
     std::string sli_img_start_tmpl;
 
@@ -217,7 +215,6 @@ struct mtmd_context {
 
     void init_vision() {
         GGML_ASSERT(ctx_v != nullptr);
-        use_mrope = clip_is_mrope(ctx_v);
 
         projector_type proj = clip_get_projector_type(ctx_v);
         int minicpmv_version = clip_is_minicpmv(ctx_v);
@@ -627,7 +624,7 @@ struct mtmd_tokenizer {
                 }
 
                 mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
-                if (ctx->use_mrope) {
+                if (mtmd_decode_use_mrope(ctx)) {
                     // for Qwen2VL, we need this information for M-RoPE decoding positions
                     image_tokens->nx = clip_n_output_tokens_x(ctx->ctx_v, batch_f32.entries[0].get());
                     image_tokens->ny = clip_n_output_tokens_y(ctx->ctx_v, batch_f32.entries[0].get());
@@ -863,10 +860,7 @@ float * mtmd_get_output_embd(mtmd_context * ctx) {
 
 bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
     switch (ctx->proj_type_v()) {
-        case PROJECTOR_TYPE_QWEN2VL:
-        case PROJECTOR_TYPE_QWEN25VL:
-        case PROJECTOR_TYPE_QWEN3VL:
-        case PROJECTOR_TYPE_YOUTUVL:
+        case PROJECTOR_TYPE_GEMMA3:
             return true;
         default:
             return false;
@@ -874,7 +868,15 @@ bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
 }
 
 bool mtmd_decode_use_mrope(mtmd_context * ctx) {
-    return ctx->use_mrope;
+    switch (ctx->proj_type_v()) {
+        case PROJECTOR_TYPE_QWEN2VL:
+        case PROJECTOR_TYPE_QWEN25VL:
+        case PROJECTOR_TYPE_QWEN3VL:
+        case PROJECTOR_TYPE_GLM4V:
+            return true;
+        default:
+            return false;
+    }
 }
 
 bool mtmd_support_vision(mtmd_context * ctx) {