fix: use ggml_graph_cpy to keep hash-set/use-counts in Phase 2 capture (#huanglawsded)

Author: Alex7MV

src/llama-context.cpp

@@ -1884,12 +1884,6 @@ 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; }
 
-                        // Inputs must be written and CPU-assigned before capture.
-                        // alloc_graph needs the final shapes and host-buffer inputs,
-                        // and the persistent graph will share these tensor objects.
-                        force_idxs_to_cpu_p2();
-                        res->set_inputs(&ubatch);
-
                         // Persist the Phase 2 graph structure across res->reset().
                         // Tensor objects remain shared with the freshly built graph, so
                         // framework-owned backend metadata is preserved exactly.

src/moe-hijacker.cpp

@@ -161,16 +161,12 @@ void capture_phase2_graph(phase2_graph_cache & cache, ggml_cgraph * src_gf) {
         return;
     }
 
+    // ggml_graph_cpy copies nodes/leafs AND the visited_hash_set + use_counts,
+    // which the scheduler's allocator and compute rely on.
+    ggml_graph_cpy(src_gf, cache.persistent_gf);
     cache.persistent_gf->order = src_gf->order;
     cache.persistent_gf->uid   = src_gf->uid;
 
-    for (int i = 0; i < n_leafs; i++) {
-        cache.persistent_gf->leafs[cache.persistent_gf->n_leafs++] = ggml_graph_leaf(src_gf, i);
-    }
-    for (int i = 0; i < n_nodes; i++) {
-        cache.persistent_gf->nodes[cache.persistent_gf->n_nodes++] = ggml_graph_node(src_gf, i);
-    }
-
     cache.valid = true;
     fprintf(stderr, "capture_phase2_graph: %d nodes + %d leafs (shallow, size=%d)\n", n_nodes, n_leafs, gsize);
 }