Merge commit 'a61c8bc3bfae4f86b8205535bcea73f476b28c2c' into concedo_experimental

# Conflicts:
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-webgpu/ggml-webgpu.cpp
#	scripts/pr2wt.sh
#	src/llama-model.cpp
#	tools/CMakeLists.txt
#	tools/mtmd/CMakeLists.txt
#	tools/mtmd/clip.cpp
#	tools/mtmd/clip.h

src/llama-model.cpp

@@ -2561,6 +2561,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         pimpl->gpu_buft_list.emplace(dev, std::move(buft_list));
     }
 
+    ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    if (cpu_dev == nullptr) {
+        throw std::runtime_error(format("%s: no CPU backend found", __func__));
+    }
+
     // calculate the split points
     bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + n_devices(), [](float x) { return x == 0.0f; });
     std::vector<float> splits(n_devices());
@@ -2571,6 +2576,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             size_t total;
             size_t free;
             ggml_backend_dev_memory(dev, &free, &total);
+
+            // devices can return 0 bytes for free and total memory if they do not
+            // have any to report. in this case, we will use the host memory as a fallback
+            // fixes: https://github.com/ggml-org/llama.cpp/issues/18577
+            if (free == 0 && total == 0) {
+                ggml_backend_dev_memory(cpu_dev, &free, &total);
+            }
             splits[i] = free;
         }
     } else {
@@ -2587,7 +2599,6 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         splits[i] /= split_sum;
     }
 
-    ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
     int i_gpu_start = std::max(int(hparams.n_layer) + 1 - n_gpu_layers, 0);
 
     #if defined(GGML_USE_CLBLAST)
@@ -2596,9 +2607,6 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     i_gpu_start = std::max((int64_t) hparams.n_layer, (int64_t) 0);
     #endif
 
-    if (cpu_dev == nullptr) {
-        throw std::runtime_error(format("%s: no CPU backend found", __func__));
-    }
     const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, int(n_layer) + 1);
     auto get_layer_buft_list = [&](int il) -> llama_model::impl::layer_dev {
         const bool is_swa = il < int(hparams.n_layer) && hparams.is_swa(il);

src/llama-sampling.cpp

@@ -2142,7 +2142,7 @@ struct llama_sampler_xtc {
     const uint32_t seed;
     uint32_t       seed_cur;
 
-    std::mt19937    rng;
+    std::mt19937   rng;
 };
 
 static const char * llama_sampler_xtc_name(const struct llama_sampler * /*smpl*/) {

src/llama.cpp

@@ -135,8 +135,20 @@ static std::vector<llama_device_memory_data> llama_get_device_memory_data(
         }
     }
     for (size_t i = 0; i < ret.size(); i++) {
-        size_t free, total;
+        size_t free;
+        size_t total;
         ggml_backend_dev_memory(model->devices[i], &free, &total);
+
+        // devices can return 0 bytes for free and total memory if they do not
+        // have any to report. in this case, we will use the host memory as a fallback
+        // fixes: https://github.com/ggml-org/llama.cpp/issues/18577
+        if (free == 0 && total == 0) {
+            ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+            if (cpu_dev == nullptr) {
+                throw std::runtime_error(format("%s: no CPU backend found", __func__));
+            }
+            ggml_backend_dev_memory(cpu_dev, &free, &total);
+        }
         ret[i].free  = free;
         ret[i].total = total;
     }

src/models/gemma3n-iswa.cpp

@@ -255,10 +255,20 @@ ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
         inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
         inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_altup));
         cb(inp_per_layer, "inp_per_layer_selected", -1);
+        res->add_input(std::move(inp));
     } else {
-        GGML_ABORT("TODO: support embd input");
+        // Vision embedding path: use padding token (ID=0) embedding
+        const int64_t embd_size = model.tok_embd_per_layer->ne[0];  // n_embd_altup * n_layer
+
+        // Extract and dequantize padding token embedding (column 0)
+        ggml_tensor * padding_q = ggml_view_1d(ctx0, model.tok_embd_per_layer, embd_size, 0);
+        ggml_tensor * padding_f32 = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, embd_size);
+        inp_per_layer = ggml_cpy(ctx0, padding_q, padding_f32);
+
+        // Reshape to [n_embd_altup, n_layer, 1]
+        inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, 1);
+        cb(inp_per_layer, "inp_per_layer_vision", -1);
     }
-    res->add_input(std::move(inp));
     return inp_per_layer;
 }
 
@@ -276,7 +286,7 @@ ggml_tensor * llm_build_gemma3n_iswa::project_per_layer_inputs(ggml_tensor * inp
                                               -1);  // [n_embd_altup, n_layer, n_tokens]
     cb(per_layer_proj, "per_layer_proj", -1);
 
-    inp_per_layer = ggml_add(ctx0, inp_per_layer, per_layer_proj);
+    inp_per_layer = ggml_add(ctx0, per_layer_proj, inp_per_layer);
     inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);
     cb(inp_per_layer, "inp_per_layer", -1);