happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

grammar: prevent stack overflow with nullable symbol loop 

Fix a potential stack overflow in llama_grammar_advance_stack that
could occur when processing grammars with nullable symbols that lead
to infinite derivations of empty strings. The fix introduces cycle
detection by tracking visited stacks to prevent infinite recursion.

rg-edit regexp: llama_grammar_advance_stack
rg-edit extra-args: -A20
rg-edit directive: """Rewrite: fix the following segfault:

[..]
⚫ Testing segfault. Grammar:
            root ::= ( [x]* )*

            root ::= ( [x]* )*

Segmentation fault         build/bin/test-grammar-integration"""

gptel-context Value:
(("~/devel/ai/llama.cpp/src/llama-grammar.cpp")
 ("~/devel/ai/llama.cpp/tests/test-grammar-integration.cpp")
 ("~/devel/ai/llama.cpp/grammars/./list.gbnf")
 ("~/devel/ai/llama.cpp/grammars/./json_arr.gbnf")
 ("~/devel/ai/llama.cpp/grammars/./json.gbnf")
 ("~/devel/ai/llama.cpp/grammars/./japanese.gbnf")
 ("~/devel/ai/llama.cpp/grammars/./english.gbnf")
 ("~/devel/ai/llama.cpp/grammars/./chess.gbnf")
 ("~/devel/ai/llama.cpp/grammars/./c.gbnf")
 ("~/devel/ai/llama.cpp/grammars/./arithmetic.gbnf")
 ("~/devel/ai/llama.cpp/grammars/./README.md"))
This commit is contained in:
Andrea Arcangeli 2026-01-02 17:42:31 +01:00
parent 17d8309fe0
commit e289f380bf
1 changed files with 18 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
src/llama-grammar.cpp
View File

@ -830,7 +830,13 @@ static bool llama_grammar_match_token(
static void llama_grammar_advance_stack(
const llama_grammar_rules & rules,
const llama_grammar_stack & stack,
llama_grammar_stacks & new_stacks) {
llama_grammar_stacks & new_stacks,
llama_grammar_stacks & seen_stacks) {
if (std::find(seen_stacks.begin(), seen_stacks.end(), stack) != seen_stacks.end()) {
return;
}
seen_stacks.push_back(stack);
if (stack.empty()) {
if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) {
new_stacks.emplace_back(stack);
@ -855,7 +861,7 @@ static void llama_grammar_advance_stack(
// if alternate is nonempty, add to stack
new_stack.push_back(subpos);
}
llama_grammar_advance_stack(rules, new_stack, new_stacks);
llama_grammar_advance_stack(rules, new_stack, new_stacks, seen_stacks);
while (!llama_grammar_is_end_of_sequence(subpos)) {
// scan to end of alternate def
subpos++;
@ -989,7 +995,8 @@ static void llama_grammar_accept_chr(
if (!llama_grammar_is_end_of_sequence(match.second)) {
new_stack.push_back(match.second);
}
llama_grammar_advance_stack(grammar.rules, new_stack, new_stacks);
llama_grammar_stacks seen_stacks;
llama_grammar_advance_stack(grammar.rules, new_stack, new_stacks, seen_stacks);
}
}
@ -1065,7 +1072,8 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
stack_after.push_back(stack_pos_after);
}
llama_grammar_stacks next_stacks;
llama_grammar_advance_stack(rules, stack_after, next_stacks);
llama_grammar_stacks seen_stacks;
llama_grammar_advance_stack(rules, stack_after, next_stacks, seen_stacks);
auto next_rejects = llama_grammar_reject_candidates(rules, next_stacks, next_candidates);
for (const auto & tok : next_rejects) {
@ -1116,7 +1124,8 @@ struct llama_grammar * llama_grammar_init_impl(
// if alternate is nonempty, add to stack
stack.push_back(pos);
}
llama_grammar_advance_stack(vec_rules, stack, stacks);
llama_grammar_stacks seen_stacks;
llama_grammar_advance_stack(vec_rules, stack, stacks, seen_stacks);
while (!llama_grammar_is_end_of_sequence(pos)) {
// scan to end of alternate def
pos++;
@ -1209,7 +1218,8 @@ struct llama_grammar * llama_grammar_init_impl(
// if alternate is nonempty, add to stack
stack.push_back(pos);
}
llama_grammar_advance_stack(vec_rules, stack, stacks);
llama_grammar_stacks seen_stacks;
llama_grammar_advance_stack(vec_rules, stack, stacks, seen_stacks);
while (!llama_grammar_is_end_of_sequence(pos)) {
// scan to end of alternate def
pos++;
@ -1428,7 +1438,8 @@ void llama_grammar_accept_token(struct llama_grammar & grammar, llama_token toke
if (!llama_grammar_is_end_of_sequence(pos + 1)) {
new_stack.push_back(pos + 1);
}
llama_grammar_advance_stack(grammar.rules, new_stack, stacks_new);
llama_grammar_stacks seen_stacks;
llama_grammar_advance_stack(grammar.rules, new_stack, stacks_new, seen_stacks);
}
} else {
llama_grammar_stacks current_stacks = {stack};