llama : adds llama-grammar memoization stacks (#4218)

examples/gbnf-validator/gbnf-validator.cpp

@@ -15,10 +15,11 @@ static bool llama_grammar_validate(struct llama_grammar * grammar, const std::st
           llama_grammar_stacks & stacks_cur = llama_grammar_get_stacks(grammar);
 
     size_t pos = 0;
+    llama_grammar_stacks_cache stacks_cache;
     for (const auto & cpt : cpts) {
         const llama_grammar_stacks stacks_prev = llama_grammar_get_stacks(grammar); // copy
 
-        llama_grammar_accept(rules, stacks_prev, cpt, stacks_cur);
+        llama_grammar_accept(rules, stacks_prev, cpt, stacks_cur, stacks_cache);
 
         if (stacks_cur.empty()) {
             error_pos = pos;

src/llama-grammar.cpp

@@ -682,6 +682,101 @@ static bool llama_grammar_match_partial_char(
     return !is_positive_char;
 }
 
+// transforms a grammar pushdown stack into N possible stacks, all ending
+// at a character range (terminal element)
+// additionally memorizes the stack to its possible stacks by mapping
+// < llama_grammar_stack, llama_grammar_stacks >
+
+static void llama_grammar_advance_stack_memo(
+        const llama_grammar_rules  & rules,
+        const llama_grammar_stack  & stack,
+              llama_grammar_stacks & new_stacks,
+              llama_grammar_stacks_cache & stacks_cache);
+
+static void llama_grammar_advance_stack_memo_impl(
+        const llama_grammar_rules  & rules,
+        const llama_grammar_stack  & stack,
+              llama_grammar_stacks & new_stacks,
+              llama_grammar_stacks_cache & stacks_cache) {
+    if (stack.empty()) {
+        if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) {
+            new_stacks.emplace_back(stack);
+        }
+        return;
+    }
+
+    const llama_grammar_element * pos = stack.back();
+
+    switch (pos->type) {
+        case LLAMA_GRETYPE_RULE_REF: {
+            const size_t                  rule_id = static_cast<size_t>(pos->value);
+            const llama_grammar_element * subpos  = rules[rule_id].data();
+            do {
+                // init new stack without the top (pos)
+                llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
+                if (!llama_grammar_is_end_of_sequence(pos + 1)) {
+                    // if this rule ref is followed by another element, add that to stack
+                    new_stack.push_back(pos + 1);
+                }
+                if (!llama_grammar_is_end_of_sequence(subpos)) {
+                    // if alternate is nonempty, add to stack
+                    new_stack.push_back(subpos);
+                }
+                llama_grammar_advance_stack_memo(rules, new_stack, new_stacks, stacks_cache);
+                while (!llama_grammar_is_end_of_sequence(subpos)) {
+                    // scan to end of alternate def
+                    subpos++;
+                }
+                if (subpos->type == LLAMA_GRETYPE_ALT) {
+                    // there's another alternate def of this rule to process
+                    subpos++;
+                } else {
+                    break;
+                }
+            } while (true);
+            break;
+        }
+        case LLAMA_GRETYPE_CHAR:
+        case LLAMA_GRETYPE_CHAR_NOT:
+        case LLAMA_GRETYPE_CHAR_ANY:
+            if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) {
+                // only add the stack if it's not a duplicate of one we already have
+                new_stacks.emplace_back(stack);
+            }
+            break;
+        default:
+            // end of alternate (LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT) or middle of char range
+            // (LLAMA_GRETYPE_CHAR_ALT, LLAMA_GRETYPE_CHAR_RNG_UPPER); stack should never be left on
+            // those
+            GGML_ABORT("fatal error");
+    }
+}
+
+static void llama_grammar_advance_stack_memo(
+        const llama_grammar_rules  & rules,
+        const llama_grammar_stack  & stack,
+              llama_grammar_stacks & new_stacks,
+              llama_grammar_stacks_cache & stacks_cache) {
+
+    llama_grammar_stacks advanced_stacks;
+    // Look if stack is already in memory
+    auto it = stacks_cache.find(stack);
+    if (it != stacks_cache.end()) {
+           advanced_stacks = it->second;
+    } else {
+        // Advance stacks with memorization 
+        llama_grammar_advance_stack_memo_impl(rules, stack, advanced_stacks, stacks_cache);
+        stacks_cache.insert(make_pair(stack, advanced_stacks));
+    }
+    // Add the advanced stacks to new_stacks avoiding duplicates
+    for (const auto & new_stack : advanced_stacks) {
+        if (std::find(new_stacks.begin(), new_stacks.end(), new_stack) == new_stacks.end()) {
+            new_stacks.emplace_back(new_stack);
+        }
+    }
+
+}
+
 // transforms a grammar pushdown stack into N possible stacks, all ending
 // at a character range (terminal element)
 static void llama_grammar_advance_stack(
@@ -826,7 +921,8 @@ void llama_grammar_accept(
         const llama_grammar_rules  & rules,
         const llama_grammar_stacks & stacks,
         const uint32_t               chr,
-              llama_grammar_stacks & stacks_new) {
+              llama_grammar_stacks & stacks_new,
+              llama_grammar_stacks_cache & stacks_cache) {
     stacks_new.clear();
     stacks_new.reserve(stacks.size());
 
@@ -844,7 +940,7 @@ void llama_grammar_accept(
             if (!llama_grammar_is_end_of_sequence(pos)) {
                 new_stack.push_back(pos);
             }
-            llama_grammar_advance_stack(rules, new_stack, stacks_new);
+            llama_grammar_advance_stack_memo(rules, new_stack, stacks_new, stacks_cache);
         }
     }
 }
@@ -938,14 +1034,15 @@ struct llama_grammar * llama_grammar_init_impl(
 
     // loop over alternates of start rule to build initial stacks
     llama_grammar_stacks stacks;
+    llama_grammar_stacks_cache stacks_cache;
     pos = vec_rules[start_rule_index].data();
     do {
         llama_grammar_stack stack;
         if (!llama_grammar_is_end_of_sequence(pos)) {
             // if alternate is nonempty, add to stack
             stack.push_back(pos);
         }
-        llama_grammar_advance_stack(vec_rules, stack, stacks);
+        llama_grammar_advance_stack_memo(vec_rules, stack, stacks, stacks_cache);
         while (!llama_grammar_is_end_of_sequence(pos)) {
             // scan to end of alternate def
             pos++;
@@ -1016,14 +1113,15 @@ struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab,
 
     // loop over alternates of start rule to build initial stacks
     llama_grammar_stacks stacks;
+    llama_grammar_stacks_cache stacks_cache;
     pos = vec_rules[start_rule_index].data();
     do {
         llama_grammar_stack stack;
         if (!llama_grammar_is_end_of_sequence(pos)) {
             // if alternate is nonempty, add to stack
             stack.push_back(pos);
         }
-        llama_grammar_advance_stack(vec_rules, stack, stacks);
+        llama_grammar_advance_stack_memo(vec_rules, stack, stacks, stacks_cache);
         while (!llama_grammar_is_end_of_sequence(pos)) {
             // scan to end of alternate def
             pos++;
@@ -1127,9 +1225,10 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
     const auto & code_points = decoded.first;
 
     llama_grammar_stacks stacks_new;
+    llama_grammar_stacks_cache stacks_cache;
 
     for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
-        llama_grammar_accept(grammar.rules, grammar.stacks, *it, stacks_new);
+        llama_grammar_accept(grammar.rules, grammar.stacks, *it, stacks_new, stacks_cache);
         grammar.stacks = std::move(stacks_new);
     }
 

src/llama-grammar.h

@@ -3,6 +3,7 @@
 #include "llama-impl.h"
 
 #include <map>
+#include <unordered_map>
 
 struct llama_vocab;
 
@@ -61,6 +62,18 @@ using llama_grammar_candidates = std::vector<llama_grammar_candidate>;
 const llama_grammar_rules  & llama_grammar_get_rules (const struct llama_grammar * grammar);
       llama_grammar_stacks & llama_grammar_get_stacks(      struct llama_grammar * grammar);
 
+struct VectorPointerHash {
+    size_t operator()(const llama_grammar_stack & v) const {
+        size_t seed = v.size();
+        for (const auto* ptr : v) {
+            seed ^= std::hash<const llama_grammar_element*>()(ptr) + 0x9e3779b9 + (seed << 6) + (seed >> 2);
+        }
+        return seed;
+    }
+};
+
+using llama_grammar_stacks_cache = std::unordered_map<llama_grammar_stack, llama_grammar_stacks, VectorPointerHash>;
+
 // takes a set of possible pushdown stacks on a grammar, which are required to
 // be positioned at a character range (see `llama_grammar_advance_stack`), and
 // produces the N possible stacks if the given char is accepted at those
@@ -69,7 +82,8 @@ void llama_grammar_accept(
         const llama_grammar_rules  & rules,
         const llama_grammar_stacks & stacks,
                           uint32_t   chr,
-              llama_grammar_stacks & stacks_new);
+              llama_grammar_stacks & stacks_new,
+              llama_grammar_stacks_cache & stacks_cache);
 
 std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_stack(
         const llama_grammar_rules      & rules,

tests/test-grammar-integration.cpp

@@ -35,10 +35,11 @@ static bool match_string(const std::string & input, llama_grammar * grammar) {
     const llama_grammar_rules  & rules      = llama_grammar_get_rules (grammar);
           llama_grammar_stacks & stacks_cur = llama_grammar_get_stacks(grammar);
 
+    llama_grammar_stacks_cache stacks_cache;
     for (const auto & cpt : cpts) {
         const llama_grammar_stacks stacks_prev = llama_grammar_get_stacks(grammar); // copy
 
-        llama_grammar_accept(rules, stacks_prev, cpt, stacks_cur);
+        llama_grammar_accept(rules, stacks_prev, cpt, stacks_cur, stacks_cache);
 
         if (stacks_cur.empty()) {
             // no stacks means that the grammar failed to match at this point