Fix grammar-based sampling issue in server (#2566)

* ggml-alloc: Don't try to re-use buffers of external tensors

They might be weights that came from another context, so we
have no control over them (and they might be re-used elsewhere
so writing to them would be a bad idea).

* ggml-alloc: >= when checking for out-of-bounds

Co-authored-by: slaren <slarengh@gmail.com>

---------

Co-authored-by: slaren <slarengh@gmail.com>

examples/server/server.cpp

@@ -196,6 +196,7 @@ struct llama_server_context
     llama_context *ctx = nullptr;
     gpt_params params;
 
+    grammar_parser::parse_state parsed_grammar;
     llama_grammar *grammar = nullptr;
 
     bool truncated = false;
@@ -241,10 +242,13 @@ struct llama_server_context
         stopped_limit = false;
         stopping_word = "";
         multibyte_pending = 0;
-        grammar = nullptr;
-
         n_remain = 0;
         n_past = 0;
+
+        if (grammar != nullptr) {
+            llama_grammar_free(grammar);
+            grammar = nullptr;
+        }
     }
 
     bool loadModel(const gpt_params &params_)
@@ -265,8 +269,6 @@ struct llama_server_context
     bool loadGrammar()
     {
         if (!params.grammar.empty()) {
-            grammar_parser::parse_state parsed_grammar;
-
             parsed_grammar = grammar_parser::parse(params.grammar.c_str());
             // will be empty (default) if there are parse errors
             if (parsed_grammar.rules.empty()) {

ggml-alloc.c

@@ -394,6 +394,14 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     if (parent == NULL) {
                         break;
                     }
+
+                    // if the node's data is external, then we cannot re-use it
+                    if ((char *) parent->data < (char *) alloc->data ||
+                        (char *) parent->data >= ((char *) alloc->data + alloc->size)) {
+                        AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
+                        continue;
+                    }
+
                     struct hash_node * p_hn = hash_get(ht, parent);
                     if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
                         if (ggml_is_view(parent)) {