fix: use C++ heap (new/delete) instead of ggml_context for deep-copy (#huanglawsded)

- Replaced ggml_dup_tensor+ggml_context with copy-constructor new ggml_tensor(*src)
- Avoids ggml context corruption that caused SIGSEGV in split_graph
- All fields (name, op, data, extra, flags, op_params, ne, nb, view_offs, buffer)
  copied via default memberwise copy; src[] and view_src fixed up separately
- release() deletes individual tensors + graph buffer

Author: Alex7MV

src/moe-hijacker.cpp

@@ -190,9 +190,9 @@ void cascade_force_moe_consumers(
 // ---- Phase 2 Graph Deep-Copy ----
 
 void phase2_graph_cache::release() {
-    if (persistent_gf) { delete[] (char *)persistent_gf; persistent_gf = nullptr; }
-    if (persistent_ctx) { ggml_free(persistent_ctx); persistent_ctx = nullptr; }
+    for (auto * t : persistent_tensors) delete t;
     persistent_tensors.clear();
+    if (persistent_gf) { delete[] (char *)persistent_gf; persistent_gf = nullptr; }
     valid = false;
 }
 
@@ -206,28 +206,12 @@ void deep_copy_phase2_graph(
     int n_leafs = ggml_graph_n_leafs(src_gf);
     int total_t = n_nodes + n_leafs;
 
-    const size_t tensor_bytes = (size_t)total_t * 1024;
-    struct ggml_init_params params = { tensor_bytes, nullptr, true };
-    cache.persistent_ctx = ggml_init(params);
-    if (!cache.persistent_ctx) {
-        fprintf(stderr, "deep_copy_phase2_graph: ggml_init(%zu) failed\n", tensor_bytes);
-        return;
-    }
-
     std::unordered_map<const ggml_tensor *, ggml_tensor *> map;
 
     auto dup_tensor = [&](const ggml_tensor * src) -> ggml_tensor * {
-        auto * dst = ggml_dup_tensor(cache.persistent_ctx, src);
-        if (!dst) return nullptr;
-        strncpy(dst->name, src->name, GGML_MAX_NAME - 1);
-        dst->name[GGML_MAX_NAME - 1] = '\0';
-        dst->op       = src->op;
-        dst->data     = src->data;
-        dst->flags    = src->flags;
-        dst->view_offs = src->view_offs;
-        dst->extra    = src->extra;
-        memcpy(dst->op_params, src->op_params, sizeof(dst->op_params));
-        for (int d = 0; d < GGML_MAX_DIMS; d++) dst->nb[d] = src->nb[d];
+        auto * dst = new ggml_tensor(*src);
+        for (int s = 0; s < GGML_MAX_SRC; s++) dst->src[s] = src->src[s];
+        dst->view_src = src->view_src;
         map[src] = dst;
         cache.persistent_tensors.push_back(dst);
         return dst;
@@ -257,11 +241,8 @@ void deep_copy_phase2_graph(
     cache.persistent_gf->n_leafs = 0;
     cache.persistent_gf->nodes   = (ggml_tensor **)(buf + sizeof(ggml_cgraph));
     cache.persistent_gf->leafs   = cache.persistent_gf->nodes + total_t;
-    cache.persistent_gf->grads     = nullptr;
-    cache.persistent_gf->grad_accs = nullptr;
-    cache.persistent_gf->use_counts = nullptr;
-    cache.persistent_gf->order      = GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT;
-    cache.persistent_gf->uid        = 0;
+    cache.persistent_gf->order   = src_gf->order;
+    cache.persistent_gf->uid     = 0;
 
     for (int i = 0; i < n_leafs; i++) {
         auto it = map.find(ggml_graph_leaf(src_gf, i));
@@ -273,8 +254,7 @@ void deep_copy_phase2_graph(
     }
 
     cache.valid = true;
-    fprintf(stderr, "deep_copy_phase2_graph: copied %d nodes + %d leafs (%zu KB ctx, %zu B graph)\n",
-            n_nodes, n_leafs, tensor_bytes / 1024, graph_bytes);
+    fprintf(stderr, "deep_copy_phase2_graph: copied %d nodes + %d leafs\n", n_nodes, n_leafs);
 }
 
 } // namespace moe

src/moe-hijacker.h

@@ -14,7 +14,6 @@
 
 struct phase2_hijack;
 struct ggml_cgraph;
-struct ggml_context;
 struct ggml_tensor;
 
 namespace moe {
@@ -38,7 +37,6 @@ void cascade_force_moe_consumers(
 struct phase2_graph_cache {
     bool valid = false;
     ggml_cgraph * persistent_gf = nullptr;
-    struct ggml_context * persistent_ctx = nullptr;
     std::vector<ggml_tensor *> persistent_tensors;
 
     void capture() { valid = true; }