fix: limit cascade to VIEW(op=37) and ADD(op=2) only (1 level, no do/while)

The iterative cascade propagated through l_out → norm → routing chain,
forcing ARGSORT to GPU which crashed with 'invalid argument'. Now only
forces the specific VIEW and ADD nodes that directly consume matched
tensors (the expert output combination chain), not arbitrary ops.

Author: Alex7MV

src/llama-context.cpp

@@ -2047,24 +2047,18 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
                             if (tid >= 0)
                                 ggml_backend_sched_set_tensor_backend(sched.get(), t, gpu);
                         }
-                        // Cascade: force nodes whose src tensors match ffn_moe_*
-                        bool changed;
-                        do {
-                            changed = false;
-                            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);
-                                if (be_t && ggml_backend_is_cuda(be_t)) continue;
-                                for (int s = 0; s < GGML_MAX_SRC && t->src[s]; s++) {
-                                    auto [sid, sil] = h2_hijack.match_name(t->src[s]->name);
-                                    if (sid >= 0) {
-                                        ggml_backend_sched_set_tensor_backend(sched.get(), t, gpu);
-                                        changed = true;
-                                        break;
-                                    }
+                        // Cascade: force VIEW (op=37) and ADD (op=2) consumers of matched tensors
+                        for (int i = 0; i < ggml_graph_n_nodes(phase2_gf); i++) {
+                            ggml_tensor * t = ggml_graph_node(phase2_gf, i);
+                            if (t->op != 37 && t->op != 2) continue;
+                            for (int s = 0; s < GGML_MAX_SRC && t->src[s]; s++) {
+                                auto [sid, sil] = h2_hijack.match_name(t->src[s]->name);
+                                if (sid >= 0) {
+                                    ggml_backend_sched_set_tensor_backend(sched.get(), t, gpu);
+                                    break;
                                 }
                             }
-                        } while (changed);
+                        }
                         force_idxs_to_cpu();
                         if (!ggml_backend_sched_alloc_graph(sched.get(), phase2_gf)) { ret = GGML_STATUS_ALLOC_FAILED; return nullptr; }
 
@@ -2125,25 +2119,19 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
                                 auto [tid, il] = h2_hijack.match_name(t->name);
                                 if (tid >= 0) ggml_backend_sched_set_tensor_backend(sched.get(), t, gpu);
                             }
-                            // Cascade: force nodes whose src tensors match ffn_moe_*
-                            // (catches VIEW and ADD nodes that consume matched tensors)
-                            bool changed;
-                            do {
-                                changed = false;
-                                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);
-                                    if (be_t && ggml_backend_is_cuda(be_t)) continue;
-                                    for (int s = 0; s < GGML_MAX_SRC && t->src[s]; s++) {
-                                        auto [sid, sil] = h2_hijack.match_name(t->src[s]->name);
-                                        if (sid >= 0) {
-                                            ggml_backend_sched_set_tensor_backend(sched.get(), t, gpu);
-                                            changed = true;
-                                            break;
-                                        }
+                            // Cascade: force VIEW (op=37) and ADD (op=2) nodes that consume
+                            // matched tensors (catches expert output combination chain only)
+                            for (int i = 0; i < ggml_graph_n_nodes(phase2_gf); i++) {
+                                ggml_tensor * t = ggml_graph_node(phase2_gf, i);
+                                if (t->op != 37 && t->op != 2) continue;
+                                for (int s = 0; s < GGML_MAX_SRC && t->src[s]; s++) {
+                                    auto [sid, sil] = h2_hijack.match_name(t->src[s]->name);
+                                    if (sid >= 0) {
+                                        ggml_backend_sched_set_tensor_backend(sched.get(), t, gpu);
+                                        break;
                                     }
                                 }
-                            } while (changed);
+                            }
                             force_idxs_to_cpu();
                             if (!ggml_backend_sched_alloc_graph(sched.get(), phase2_gf)) {
                                 ret = GGML_STATUS_ALLOC_FAILED; return nullptr;