fix: force only matched (ffn_moe_*) tensors to GPU, not all nodes

Forcing ALL 4790 nodes to GPU breaks MUL_MAT_ID expert ID handling
in ggml_backend_sched_compute_splits (assertion: id >= 0 && id < n_expert).
The non-matched tensors (attention, norms, masks, expert IDs) stay on
CPU where the scheduler and set_input functions can handle them correctly.

Remove node dump diagnostics (no longer needed).

Author: Alex7MV

src/llama-context.cpp

@@ -2038,10 +2038,18 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
                         phase2_gf = model.build_graph(gparams, nullptr, nullptr, &moe_weight_cache);
                         if (!phase2_gf) { ret = GGML_STATUS_FAILED; return nullptr; }
                         ggml_backend_sched_reset(sched.get());
-                        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);
+                        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);
+                        }
                         force_idxs_to_cpu();
                         if (!ggml_backend_sched_alloc_graph(sched.get(), phase2_gf)) { ret = GGML_STATUS_ALLOC_FAILED; return nullptr; }
 
@@ -2090,24 +2098,24 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
 
                         if (do_cuda && !h2_hijack.captured) {
                             // --- CAPTURE PATH ---
-                            // 1. Reset + force All Phase 2 ops/srcs to GPU, then alloc
+                            // 1. Reset + force matched MoE tensors to GPU
                             ggml_backend_sched_reset(sched.get());
-                            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_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);
+                            }
                             force_idxs_to_cpu();
                             if (!ggml_backend_sched_alloc_graph(sched.get(), phase2_gf)) {
                                 ret = GGML_STATUS_ALLOC_FAILED; return nullptr;
                             }
-                            #ifdef GGML_USE_CUDA
-                            fprintf(stderr, "phase2_dump: %d nodes, %d leafs\n",
-                                ggml_graph_n_nodes(phase2_gf), ggml_graph_n_leafs(phase2_gf));
-                            for (int i = 0; i < ggml_graph_n_nodes(phase2_gf); i++) {
-                                ggml_tensor * t = ggml_graph_node(phase2_gf, i);
-                                ggml_backend_t be_t = ggml_backend_sched_get_tensor_backend(sched.get(), t);
-                                fprintf(stderr, "phase2[node %d]: %s op=%d cuda=%d\n", i, t->name, (int)t->op,
-                                    be_t ? ggml_backend_is_cuda(be_t) : -1);
-                            }
-                            #endif
                             if (res->t_logits) ggml_backend_sched_set_tensor_backend(sched.get(), res->t_logits, be);
                             if (res->t_embd)   ggml_backend_sched_set_tensor_backend(sched.get(), res->t_embd,   be);