fix: force ALL Phase2 nodes+leafs to GPU (node_XXX ADD ops and views were on CPU)

Author: Alex7MV

src/llama-context.cpp

@@ -2081,16 +2081,10 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
                             // --- CAPTURE PATH ---
                             // 1. Reset + force All Phase 2 ops/srcs to GPU, then alloc
                             ggml_backend_sched_reset(sched.get());
-                            for (int i = 0; i < ggml_graph_n_nodes(phase2_gf); i++) {
-                                ggml_tensor * t = ggml_graph_node(phase2_gf, i);
-                                auto [tid, il] = h2_hijack.match_name(t->name);
-                                if (tid >= 0) ggml_backend_sched_set_tensor_backend(sched.get(), t, gpu);
-                            }
-                            for (int i = 0; i < ggml_graph_n_leafs(phase2_gf); i++) {
-                                ggml_tensor * t = ggml_graph_leaf(phase2_gf, i);
-                                auto [tid, il] = h2_hijack.match_name(t->name);
-                                if (tid >= 0) ggml_backend_sched_set_tensor_backend(sched.get(), t, gpu);
-                            }
+                            for (int i = 0; i < ggml_graph_n_nodes(phase2_gf); i++)
+                                ggml_backend_sched_set_tensor_backend(sched.get(), ggml_graph_node(phase2_gf, i), gpu);
+                            for (int i = 0; i < ggml_graph_n_leafs(phase2_gf); i++)
+                                ggml_backend_sched_set_tensor_backend(sched.get(), ggml_graph_leaf(phase2_gf, i), gpu);
                             force_idxs_to_cpu();
                             if (!ggml_backend_sched_alloc_graph(sched.get(), phase2_gf)) {
                                 ret = GGML_STATUS_ALLOC_FAILED; return nullptr;