Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.github/workflows/release.yml
#	app/llama.cpp
#	common/download.cpp
#	docs/backend/SYCL.md
#	docs/ops.md
#	docs/ops/SYCL.csv
#	ggml/CMakeLists.txt
#	ggml/src/CMakeLists.txt
#	ggml/src/ggml-cpu/CMakeLists.txt
#	ggml/src/ggml-sycl/CMakeLists.txt
#	ggml/src/ggml-sycl/backend.hpp
#	ggml/src/ggml-sycl/common.cpp
#	ggml/src/ggml-sycl/common.hpp
#	ggml/src/ggml-sycl/convert.cpp
#	ggml/src/ggml-sycl/dequantize.hpp
#	ggml/src/ggml-sycl/dmmv.cpp
#	ggml/src/ggml-sycl/dpct/helper.hpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	ggml/src/ggml-sycl/mmvq.cpp
#	ggml/src/ggml-sycl/outprod.cpp
#	ggml/src/ggml-sycl/vecdotq.hpp
#	tools/server/README.md

src/llama-context.cpp

@@ -1392,7 +1392,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
     const auto & hparams = model.hparams;
 
     // eagle3/DFlash: features as encoder input, and non-draft paths fall back to model's input dim
-    const int64_t n_embd = hparams.n_embd_inp();
+    const int64_t n_embd = hparams.n_embd_inp_enc();
     const int64_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0

src/llama-hparams.cpp

@@ -104,6 +104,10 @@ uint32_t llama_hparams::n_embd_inp() const {
     return n_embd_inp;
 }
 
+uint32_t llama_hparams::n_embd_inp_enc() const {
+    return n_embd_inp_enc_impl > 0 ? n_embd_inp_enc_impl : n_embd_inp();
+}
+
 uint32_t llama_hparams::n_embd_out() const {
     return n_embd_out_impl > 0 ? n_embd_out_impl : n_embd;
 }

src/llama-hparams.h

@@ -189,6 +189,10 @@ struct llama_hparams {
     // input embedding dimension (0 = use n_embd)
     uint32_t n_embd_inp_impl = 0;
 
+    // encoder input embedding dimension (0 = use n_embd_inp())
+    // e.g. the eagle3 encoder fuses target_layers * target_hidden features
+    uint32_t n_embd_inp_enc_impl = 0;
+
     // output embedding dimension (0 = use n_embd)
     uint32_t n_embd_out_impl = 0;
 
@@ -305,6 +309,9 @@ struct llama_hparams {
     // dimension of main + auxiliary input embeddings
     uint32_t n_embd_inp() const;
 
+    // dimension of the encoder input embeddings
+    uint32_t n_embd_inp_enc() const;
+
     // dimension of output embeddings
     uint32_t n_embd_out() const;
 

src/llama.cpp

@@ -273,7 +273,7 @@ static bool llama_prepare_model_devices(const llama_model_params & params, llama
     }
 
     // if using single GPU mode, remove all except the main GPU
-    if (params.split_mode == LLAMA_SPLIT_MODE_NONE) {
+    if (params.split_mode == LLAMA_SPLIT_MODE_NONE && !model->devices.empty()) {
         if (params.main_gpu < 0) {
             model->devices.clear();
         } else {

src/models/eagle3.cpp

@@ -19,7 +19,7 @@ void llama_model_eagle3::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_TARGET_HIDDEN_SIZE, n_embd_tgt);
     LLAMA_LOG_INFO("%s: EAGLE3 n_embd_tgt = %u (draft n_embd = %u)\n", __func__, n_embd_tgt, hparams.n_embd);
 
-    hparams.n_embd_inp_impl = (uint32_t) target_layer_ids.size() * n_embd_tgt;
+    hparams.n_embd_inp_enc_impl = (uint32_t) target_layer_ids.size() * n_embd_tgt;
 
     // eagle3 norm_before_residual (optional, default false)
     // compatible with Readhat eagle3 speculator model
@@ -34,7 +34,7 @@ void llama_model_eagle3::load_arch_hparams(llama_model_loader & ml) {
 void llama_model_eagle3::load_arch_tensors(llama_model_loader &) {
     LLAMA_LOAD_LOCALS;
 
-    const int64_t n_embd_inp = hparams.n_embd_inp();
+    const int64_t n_embd_inp = hparams.n_embd_inp_enc();
     const int64_t n_embd_attn_input = 2 * n_embd;
 
     // Get vocab size from the d2t tensor in the GGUF file (optional - only needed if eagle3 has different vocab_size than target)
@@ -109,8 +109,8 @@ ggml_tensor * llama_model_eagle3::graph<true>::build_inp_embd_enc() const {
 
     // Input: Target model features (3 layers concatenated: low, mid, high)
     // Data will be provided via ubatch->embd in encode_eagle3_features()
-    auto inp_target = std::make_unique<llm_graph_input_embd>(hparams.n_embd_inp());
-    inp_target->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32,hparams.n_embd_inp(), n_tokens);
+    auto inp_target = std::make_unique<llm_graph_input_embd>(hparams.n_embd_inp_enc());
+    inp_target->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp_enc(), n_tokens);
     ggml_set_input(inp_target->embd);
 
     cur = inp_target->embd;