fix eagle3 logits sync bug & remove ggml_set_sync()

ggml/include/ggml.h

@@ -629,7 +629,6 @@ extern "C" {
         GGML_TENSOR_FLAG_OUTPUT =  2, // ...is an output for the GGML compute graph
         GGML_TENSOR_FLAG_PARAM  =  4, // ...contains trainable parameters
         GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
-        GGML_TENSOR_FLAG_SYNC   = 16, // ...forces a new split/sync point in the scheduler (e.g. for EAGLE3 decoder)
     };
 
     enum ggml_tri_type {
@@ -854,7 +853,6 @@ extern "C" {
     GGML_API void ggml_set_output(struct ggml_tensor * tensor);
     GGML_API void ggml_set_param(struct ggml_tensor * tensor);
     GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
-    GGML_API void ggml_set_sync(struct ggml_tensor * tensor);  // force sync point in scheduler
 
     //
     // operations on tensors with backpropagation

ggml/src/ggml-backend.cpp

@@ -1202,11 +1202,6 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                 }
             }
 
-            // check if this node requires a sync point (e.g. for EAGLE3 parallel path fix)
-            if (node->flags & GGML_TENSOR_FLAG_SYNC) {
-                need_new_split = true;
-            }
-
             if (node_backend_id != cur_backend_id || need_new_split) {
                 split->i_end = i;
                 i_split++;
@@ -1581,15 +1576,6 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
             if (ec != GGML_STATUS_SUCCESS) {
                 return ec;
             }
-            
-            // If any node in this split has SYNC flag, synchronize after compute
-            // This ensures the sync node is complete before next split (e.g. for EAGLE3 parallel path sync fix)
-            for (int j = 0; j < split->graph.n_nodes; j++) {
-                if (split->graph.nodes[j]->flags & GGML_TENSOR_FLAG_SYNC) {
-                    ggml_backend_synchronize(split_backend);
-                    break;
-                }
-            }
         } else {
             // similar to ggml_backend_compare_graph_backend
             for (int j0 = 0; j0 < split->graph.n_nodes; j0++) {

ggml/src/ggml.c

@@ -7451,10 +7451,6 @@ void ggml_set_loss(struct ggml_tensor * tensor) {
     tensor->flags |= GGML_TENSOR_FLAG_LOSS;
 }
 
-void ggml_set_sync(struct ggml_tensor * tensor) {
-    tensor->flags |= GGML_TENSOR_FLAG_SYNC;
-}
-
 ////////////////////////////////////////////////////////////////////////////////
 
 void ggml_quantize_init(enum ggml_type type) {

src/llama-context.cpp

@@ -1261,7 +1261,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
                     // Read only the last token's draft logits
                     eagle3_draft_logits.resize(draft_vocab_size);
                     const size_t last_offset = last_idx * draft_vocab_size * sizeof(float);
-                    ggml_backend_tensor_get(t_logits, eagle3_draft_logits.data(), last_offset, draft_vocab_size * sizeof(float));
+                    ggml_backend_tensor_get_async(backend_res, t_logits, eagle3_draft_logits.data(), last_offset, draft_vocab_size * sizeof(float));
+                    synchronize();
                     
                     
                     // Map only the last token's draft logits to target vocab

src/models/eagle3.cpp

@@ -63,10 +63,6 @@ llm_build_eagle3_decode::llm_build_eagle3_decode(const llama_model & model, cons
                 LLM_NORM_RMS, 0);
         cb(input_embeds_normed, "input_layernorm", -1);
 
-        // Force a sync point between the two parallel RMS_NORM paths
-        // This prevents buffer reuse issues on GPU (EAGLE3 GPU fix)
-        ggml_set_sync(input_embeds_normed);
-
         // Apply hidden_norm to g_embeddings
         ggml_tensor * g_embeddings_normed = build_norm(g_embeddings,
                 model.layers[0].eagle3_hidden_norm, NULL,