diff --git a/src/llama-grammar.cpp b/src/llama-grammar.cpp
index c685346b630..8c823820b2b 100644
--- a/src/llama-grammar.cpp
+++ b/src/llama-grammar.cpp
@@ -7,6 +7,8 @@
 #include <cmath>
 #include <algorithm>
 #include <cstdint>
+#include <functional>
+#include <iterator>
 #include <set>
 #include <stdexcept>
 
@@ -1335,14 +1337,27 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
     };
 
     // redirect elements in stacks to point to new rules
-    for (size_t is = 0; is < result->stacks.size(); is++) {
-        for (size_t ie = 0; ie < result->stacks[is].size(); ie++) {
-            for (size_t ir0 = 0; ir0 < grammar.rules.size(); ir0++) {
-                for (size_t ir1 = 0; ir1 < grammar.rules[ir0].size(); ir1++) {
-                    if (grammar.stacks[is][ie] == &grammar.rules[ir0][ir1]) {
-                        result->stacks[is][ie] =  &result->rules[ir0][ir1];
-                    }
-                }
+    // find the rule of each element with a binary search on the rule start addresses: a search of all elements is too slow for large grammars (the server clones the grammar on each speculative step)
+    using elem_ptr = const llama_grammar_element *;
+    const std::less<elem_ptr> ptr_less;
+    std::vector<std::pair<elem_ptr, size_t>> starts;
+    starts.reserve(grammar.rules.size());
+    for (size_t ir = 0; ir < grammar.rules.size(); ir++) {
+        if (!grammar.rules[ir].empty()) {
+            starts.emplace_back(grammar.rules[ir].data(), ir);
+        }
+    }
+    std::sort(starts.begin(), starts.end(), [&](const auto & a, const auto & b) { return ptr_less(a.first, b.first); });
+    for (auto & stack : result->stacks) {
+        for (auto & pe : stack) {
+            auto it = std::upper_bound(starts.begin(), starts.end(), pe, [&](elem_ptr v, const auto & e) { return ptr_less(v, e.first); });
+            if (it == starts.begin()) {
+                continue;
+            }
+            const size_t ir = std::prev(it)->second;
+            const auto & rule = grammar.rules[ir];
+            if (ptr_less(pe, rule.data() + rule.size())) {
+                pe = &result->rules[ir][pe - rule.data()];
             }
         }
     }
diff --git a/tests/test-grammar-integration.cpp b/tests/test-grammar-integration.cpp
index eb4b7c78f50..1574ab5b087 100644
--- a/tests/test-grammar-integration.cpp
+++ b/tests/test-grammar-integration.cpp
@@ -188,6 +188,48 @@ static void test(const std::string & test_desc, const std::string & grammar_str,
 static void test_grammar(const std::string & test_desc, const std::string & grammar_str, const std::vector<std::string> & passing_strings, const std::vector<std::string> & failing_strings) {
     test(test_desc + ". Grammar: " + grammar_str, grammar_str, passing_strings, failing_strings);
 }
+
+static bool stacks_point_into_rules(llama_grammar * grammar) {
+    const auto & rules = llama_grammar_get_rules(grammar);
+    for (const auto & stack : llama_grammar_get_stacks(grammar)) {
+        for (const llama_grammar_element * pe : stack) {
+            bool found = false;
+            for (const auto & rule : rules) {
+                found = found || (!rule.empty() && pe >= rule.data() && pe < rule.data() + rule.size());
+            }
+            if (!found) {
+                return false;
+            }
+        }
+    }
+    return true;
+}
+
+static void test_clone() {
+    fprintf(stderr, "⚫ Testing grammar clone at each position of an input\n");
+    const std::string grammar_str = json_schema_to_grammar(json::parse(R"""({
+        "type": "object",
+        "properties": {
+            "name": {"type": "string"},
+            "tags": {"type": "array", "items": {"type": "string"}},
+            "n": {"type": "integer"}
+        },
+        "required": ["name", "n"]
+    })"""), true);
+    const std::string input = R"""({"name": "a", "n": 42, "tags": ["x", "y"]})""";
+    for (size_t k = 0; k <= input.size(); k++) {
+        auto * grammar = build_grammar(grammar_str);
+        for (const auto & in : parse_tokens(input.substr(0, k))) {
+            llama_grammar_accept_token(*grammar, in.token, in.piece);
+        }
+        auto * clone = llama_grammar_clone_impl(*grammar);
+        llama_grammar_free_impl(grammar);
+        assert(stacks_point_into_rules(clone));
+        assert(match_string(input.substr(k), clone));
+        llama_grammar_free_impl(clone);
+    }
+    fprintf(stdout, "  ✅︎\n");
+}
 static void test_schema(const std::string & test_desc, const std::string & schema_str, const std::vector<std::string> & passing_strings, const std::vector<std::string> & failing_strings) {
     test(test_desc + ". Schema: " + schema_str, json_schema_to_grammar(json::parse(schema_str), true), passing_strings, failing_strings);
 }
@@ -1488,6 +1530,7 @@ int main() {
     test_failure_missing_root_symbol();
     test_custom_root_symbol_check();
     test_json_schema();
+    test_clone();
     fprintf(stdout, "All tests passed.\n");
     return 0;
 }