sampling : support for llguidance grammars (#10224)

* initial porting of previous LLG patch

* update for new APIs

* build: integrate llguidance as an external project

* use '%llguidance' as marker to enable llg lark syntax

* add some docs

* clarify docs

* code style fixes

* remove llguidance.h from .gitignore

* fix tests when llg is enabled

* pass vocab not model to llama_sampler_init_llg()

* copy test-grammar-integration.cpp to test-llguidance.cpp

* clang fmt

* fix ref-count bug

* build and run test

* gbnf -> lark syntax

* conditionally include llguidance test based on LLAMA_LLGUIDANCE flag

* rename llguidance test file to test-grammar-llguidance.cpp

* add gh action for llg test

* align tests with LLG grammar syntax and JSON Schema spec

* llama_tokenizer() in fact requires valid utf8

* update llg

* format file

* add $LLGUIDANCE_LOG_LEVEL support

* fix whitespace

* fix warning

* include <cmath> for INFINITY

* add final newline

* fail llama_sampler_init_llg() at runtime

* Link gbnf_to_lark.py script; fix links; refer to llg docs for lexemes

* simplify #includes

* improve doc string for LLAMA_LLGUIDANCE

* typo in merge

* bump llguidance to 0.6.12

.github/workflows/build.yml

@@ -302,6 +302,36 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
+  ubuntu-latest-llguidance:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_LLGUIDANCE=ON
+          cmake --build . --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
   ubuntu-latest-cmake-rpc:
     runs-on: ubuntu-latest
 

CMakeLists.txt

@@ -80,6 +80,7 @@ option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
 
 # 3rd party libs
 option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
+option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)

common/CMakeLists.txt

@@ -65,6 +65,7 @@ add_library(${TARGET} STATIC
     console.h
     json-schema-to-grammar.cpp
     json.hpp
+    llguidance.cpp
     log.cpp
     log.h
     minja.hpp
@@ -91,6 +92,33 @@ if (LLAMA_CURL)
     set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
 endif ()
 
+if (LLAMA_LLGUIDANCE)
+    include(ExternalProject)
+    set(LLGUIDANCE_SRC ${CMAKE_BINARY_DIR}/llguidance/source)
+    set(LLGUIDANCE_PATH ${LLGUIDANCE_SRC}/target/release)
+    ExternalProject_Add(llguidance_ext
+        GIT_REPOSITORY https://github.com/guidance-ai/llguidance
+        # v0.6.12:
+        GIT_TAG ced1c9023d47ec194fa977932d35ce65c2ebfc09
+        PREFIX ${CMAKE_BINARY_DIR}/llguidance
+        SOURCE_DIR ${LLGUIDANCE_SRC}
+        BUILD_IN_SOURCE TRUE
+        CONFIGURE_COMMAND ""
+        BUILD_COMMAND cargo build --release
+        INSTALL_COMMAND ""
+        BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/libllguidance.a ${LLGUIDANCE_PATH}/llguidance.h
+        UPDATE_COMMAND ""
+    )
+    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_LLGUIDANCE)
+
+    add_library(llguidance STATIC IMPORTED)
+    set_target_properties(llguidance PROPERTIES IMPORTED_LOCATION ${LLGUIDANCE_PATH}/libllguidance.a)
+    add_dependencies(llguidance llguidance_ext)
+
+    target_include_directories(${TARGET} PRIVATE ${LLGUIDANCE_PATH})
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance)
+endif ()
+
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 target_link_libraries     (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)

common/json-schema-to-grammar.cpp

@@ -991,7 +991,14 @@ public:
     }
 };
 
-std::string json_schema_to_grammar(const json & schema) {
+std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
+#ifdef LLAMA_USE_LLGUIDANCE
+    if (!force_gbnf) {
+        return "%llguidance {}\nstart: %json " + schema.dump();
+    }
+#else
+    (void)force_gbnf;
+#endif // LLAMA_USE_LLGUIDANCE
     return build_grammar([&](const common_grammar_builder & callbacks) {
         auto copy = schema;
         callbacks.resolve_refs(copy);

common/json-schema-to-grammar.h

@@ -5,7 +5,8 @@
 #define JSON_ASSERT GGML_ASSERT
 #include "json.hpp"
 
-std::string json_schema_to_grammar(const nlohmann::ordered_json & schema);
+std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
+                                   bool force_gbnf = false);
 
 struct common_grammar_builder {
     std::function<std::string(const std::string &, const std::string &)> add_rule;

common/llguidance.cpp

@@ -0,0 +1,270 @@
+#include "sampling.h"
+#include "log.h"
+
+#ifdef LLAMA_USE_LLGUIDANCE
+
+#    include "llguidance.h"
+#    include <cmath>
+
+struct llama_sampler_llg {
+    const llama_vocab * vocab;
+    std::string         grammar_kind;
+    std::string         grammar_data;
+    LlgTokenizer *      tokenizer;
+    LlgConstraint *     grammar;
+    LlgMaskResult       llg_res;
+    bool                has_llg_res;
+};
+
+static LlgConstraint * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
+                                             const char * grammar_data) {
+    LlgConstraintInit cinit;
+    llg_constraint_init_set_defaults(&cinit, tokenizer);
+    const char * log_level = getenv("LLGUIDANCE_LOG_LEVEL");
+    if (log_level && *log_level) {
+        cinit.log_stderr_level = atoi(log_level);
+    }
+    auto c = llg_new_constraint_any(&cinit, grammar_kind, grammar_data);
+    if (llg_get_error(c)) {
+        LOG_ERR("llg error: %s\n", llg_get_error(c));
+        llg_free_constraint(c);
+        return nullptr;
+    }
+    return c;
+}
+
+static const char * llama_sampler_llg_name(const llama_sampler * /*smpl*/) {
+    return "llguidance";
+}
+
+static void llama_sampler_llg_accept_impl(llama_sampler * smpl, llama_token token) {
+    auto * ctx = (llama_sampler_llg *) smpl->ctx;
+    if (ctx->grammar) {
+        LlgCommitResult res;
+        llg_commit_token(ctx->grammar, token, &res);
+        ctx->has_llg_res = false;
+    }
+}
+
+static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_llg *) smpl->ctx;
+    if (ctx->grammar) {
+        if (!ctx->has_llg_res) {
+            if (llg_compute_mask(ctx->grammar, &ctx->llg_res) == 0) {
+                ctx->has_llg_res = true;
+            } else {
+                LOG_ERR("llg error: %s\n", llg_get_error(ctx->grammar));
+                llg_free_constraint(ctx->grammar);
+                ctx->grammar = nullptr;
+            }
+        }
+        if (ctx->has_llg_res) {
+            if (ctx->llg_res.is_stop) {
+                for (size_t i = 0; i < cur_p->size; ++i) {
+                    if (!llama_vocab_is_eog(ctx->vocab, cur_p->data[i].id)) {
+                        cur_p->data[i].logit = -INFINITY;
+                    }
+                }
+            } else {
+                const uint32_t * mask = ctx->llg_res.sample_mask;
+                for (size_t i = 0; i < cur_p->size; ++i) {
+                    auto token = cur_p->data[i].id;
+                    if ((mask[token / 32] & (1 << (token % 32))) == 0) {
+                        cur_p->data[i].logit = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+}
+
+static void llama_sampler_llg_reset(llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_llg *) smpl->ctx;
+    if (!ctx->grammar) {
+        return;
+    }
+
+    auto * grammar_new = llama_sampler_llg_new(ctx->tokenizer, ctx->grammar_kind.c_str(), ctx->grammar_data.c_str());
+    llg_free_constraint(ctx->grammar);
+    ctx->grammar     = grammar_new;
+    ctx->has_llg_res = false;
+}
+
+static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_llg *) smpl->ctx;
+
+    auto * result = llama_sampler_init_llg(ctx->vocab, nullptr, nullptr);
+
+    // copy the state
+    {
+        auto * result_ctx = (llama_sampler_llg *) result->ctx;
+
+        if (ctx->grammar) {
+            result_ctx->grammar_kind = ctx->grammar_kind;
+            result_ctx->grammar_data = ctx->grammar_data;
+            result_ctx->grammar      = llg_clone_constraint(ctx->grammar);
+            result_ctx->tokenizer    = llg_clone_tokenizer(ctx->tokenizer);
+        }
+    }
+
+    return result;
+}
+
+static void llama_sampler_llg_free(llama_sampler * smpl) {
+    const auto * ctx = (llama_sampler_llg *) smpl->ctx;
+
+    if (ctx->grammar) {
+        llg_free_constraint(ctx->grammar);
+        llg_free_tokenizer(ctx->tokenizer);
+    }
+
+    delete ctx;
+}
+
+static llama_sampler_i llama_sampler_llg_i = {
+    /* .name   = */ llama_sampler_llg_name,
+    /* .accept = */ llama_sampler_llg_accept_impl,
+    /* .apply  = */ llama_sampler_llg_apply,
+    /* .reset  = */ llama_sampler_llg_reset,
+    /* .clone  = */ llama_sampler_llg_clone,
+    /* .free   = */ llama_sampler_llg_free,
+};
+
+static size_t llama_sampler_llg_tokenize_fn(const void * user_data, const uint8_t * bytes, size_t bytes_len,
+                                            uint32_t * output_tokens, size_t output_tokens_len) {
+    const llama_vocab * vocab = (const llama_vocab *) user_data;
+    int                 r     = 0;
+    try {
+        r = llama_tokenize(vocab, (const char *) bytes, bytes_len, (int32_t *) output_tokens, output_tokens_len, false,
+                           true);
+    } catch (const std::exception & e) {
+        GGML_ABORT("llama_tokenize failed: %s\n", e.what());
+    }
+    if (r < 0) {
+        return -r;
+    }
+    return r;
+}
+
+static LlgTokenizer * llama_sampler_llg_new_tokenizer(const llama_vocab * vocab) {
+    // TODO store the tokenizer in the vocab somehow
+    static const llama_vocab * vocab_cache;
+    static LlgTokenizer *      tokenizer_cache;
+
+    if (vocab_cache == vocab) {
+        return llg_clone_tokenizer(tokenizer_cache);
+    }
+
+    auto tok_eos = llama_vocab_eot(vocab);
+    if (tok_eos == LLAMA_TOKEN_NULL) {
+        tok_eos = llama_vocab_eos(vocab);
+    }
+
+    size_t vocab_size = llama_vocab_n_tokens(vocab);
+
+    auto token_lens       = new uint32_t[vocab_size];
+    // we typically have ~7 bytes per token; let's go on the safe side here
+    auto token_bytes_size = vocab_size * 16 + 1024 * 1024;
+    auto token_bytes      = new uint8_t[token_bytes_size];
+
+    size_t offset = 0;
+    for (size_t i = 0; i < vocab_size; i++) {
+        size_t max_token = 1024;
+        if (token_bytes_size - offset < max_token) {
+            GGML_ABORT("token_bytes buffer too small\n");
+        }
+
+        llama_token token = i;
+        auto        dp    = (char *) token_bytes + offset;
+        auto        size  = llama_detokenize(vocab, &token, 1, dp, max_token, false, false);
+        if (size < 0) {
+            GGML_ABORT("llama_detokenize failed\n");
+        }
+        if (size == 0) {
+            size = llama_detokenize(vocab, &token, 1, dp + 1, max_token - 1, false, true);
+            if (size < 0) {
+                GGML_ABORT("llama_detokenize failed\n");
+            }
+            if (size != 0) {
+                *dp = '\xff';  // special token prefix marker
+                size += 1;
+            }
+        }
+
+        token_lens[i] = size;
+        offset += size;
+    }
+
+    LlgTokenizerInit tinit = {
+        /* .vocab_size                         = */ (uint32_t) vocab_size,
+        /* .tok_eos                            = */ (uint32_t) tok_eos,
+        /* .token_lens                         = */ token_lens,
+        /* .token_bytes                        = */ token_bytes,
+        /* .tokenizer_json                     = */ nullptr,
+        /* .tokenize_assumes_string            = */ true,
+        /* .tokenize_fn                        = */ llama_sampler_llg_tokenize_fn,
+        /* .use_approximate_greedy_tokenize_fn = */ false,
+        /* .tokenize_user_data                 = */ vocab,
+    };
+
+    char           error_buffer[1024];
+    LlgTokenizer * tokenizer = llg_new_tokenizer(&tinit, error_buffer, sizeof(error_buffer));
+
+    delete[] token_bytes;
+    delete[] token_lens;
+
+    if (tokenizer == nullptr) {
+        LOG_ERR("llg tokenizer error: %s\n", error_buffer);
+        return tokenizer;
+    }
+
+    if (tokenizer_cache) {
+        llg_free_tokenizer(tokenizer_cache);
+    }
+    vocab_cache     = vocab;
+    tokenizer_cache = tokenizer;
+
+    return llg_clone_tokenizer(tokenizer_cache);
+}
+
+llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * grammar_kind,
+                                       const char * grammar_data) {
+    auto * ctx = new llama_sampler_llg;
+
+    if (grammar_kind != nullptr && grammar_kind[0] != '\0') {
+        auto tokenizer = llama_sampler_llg_new_tokenizer(vocab);
+        *ctx           = {
+            /* .vocab        = */ vocab,
+            /* .grammar_kind = */ grammar_kind,
+            /* .grammar_data = */ grammar_data,
+            /* .tokenizer    = */ tokenizer,
+            /* .grammar      = */ llama_sampler_llg_new(tokenizer, grammar_kind, grammar_data),
+            /* .llg_res      = */ {},
+            /* .has_llg_res  = */ false,
+        };
+    } else {
+        *ctx = {
+            /* .vocab        = */ vocab,
+            /* .grammar_kind = */ {},
+            /* .grammar_data = */ {},
+            /* .tokenizer    = */ nullptr,
+            /* .grammar      = */ nullptr,
+            /* .llg_res      = */ {},
+            /* .has_llg_res  = */ false,
+        };
+    }
+
+    return new llama_sampler{
+        /* .iface = */ &llama_sampler_llg_i,
+        /* .ctx   = */ ctx,
+    };
+}
+
+#else
+
+llama_sampler * llama_sampler_init_llg(const llama_vocab *, const char *, const char *) {
+    LOG_WRN("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
+    return nullptr;
+}
+
+#endif  // LLAMA_USE_LLGUIDANCE

common/sampling.cpp

@@ -156,13 +156,25 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
     for (const auto & str : params.grammar_trigger_words) {
         trigger_words.push_back(str.word.c_str());
     }
+
+    struct llama_sampler * grmr;
+    if (params.grammar.compare(0, 11, "%llguidance") == 0) {
+#ifdef LLAMA_USE_LLGUIDANCE
+        grmr = llama_sampler_init_llg(vocab, "lark", params.grammar.c_str());
+#else
+        GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
+#endif // LLAMA_USE_LLGUIDANCE
+    } else {
+        grmr = params.grammar_lazy
+             ? llama_sampler_init_grammar_lazy(vocab, params.grammar.c_str(), "root",
+                                               trigger_words.data(), trigger_words.size(),
+                                               params.grammar_trigger_tokens.data(), params.grammar_trigger_tokens.size())
+             :      llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
+    }
+
     auto * result = new common_sampler {
         /* .params = */ params,
-        /* .grmr   = */ params.grammar_lazy
+        /* .grmr   = */ grmr,
-            ? llama_sampler_init_grammar_lazy(vocab, params.grammar.c_str(), "root",
-                                              trigger_words.data(), trigger_words.size(),
-                                              params.grammar_trigger_tokens.data(), params.grammar_trigger_tokens.size())
-            :      llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root"),
         /* .chain  = */ llama_sampler_chain_init(lparams),
         /* .prev   = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
         /* .cur    = */ {},

common/sampling.h

@@ -102,3 +102,6 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr);
 
 std::vector<enum common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
 std::vector<enum common_sampler_type> common_sampler_types_from_chars(const std::string & chars);
+
+llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab,
+                const char * grammar_kind, const char * grammar_data);

docs/llguidance.md

@@ -0,0 +1,51 @@
+# LLGuidance Support in llama.cpp
+
+[LLGuidance](https://github.com/guidance-ai/llguidance) is a library for constrained decoding (also called constrained sampling or structured outputs) for Large Language Models (LLMs). Initially developed as the backend for the [Guidance](https://github.com/guidance-ai/guidance) library, it can also be used independently.
+
+LLGuidance supports JSON Schemas and arbitrary context-free grammars (CFGs) written in a [variant](https://github.com/guidance-ai/llguidance/blob/main/docs/syntax.md) of Lark syntax. It is [very fast](https://github.com/guidance-ai/jsonschemabench/tree/main/maskbench) and has [excellent](https://github.com/guidance-ai/llguidance/blob/main/docs/json_schema.md) JSON Schema coverage but requires the Rust compiler, which complicates the llama.cpp build process.
+
+## Building
+
+To enable LLGuidance support, build llama.cpp with the `LLAMA_LLGUIDANCE` option:
+
+```sh
+cmake -B build -DLLAMA_LLGUIDANCE=ON
+make -C build -j
+```
+
+This requires the Rust compiler and the `cargo` tool to be [installed](https://www.rust-lang.org/tools/install).
+
+## Interface
+
+There are no new command-line arguments or modifications to `common_params`. When enabled, grammars starting with `%llguidance` are passed to LLGuidance instead of the [current](../grammars/README.md) llama.cpp grammars. Additionally, JSON Schema requests (e.g., using the `-j` argument in `llama-cli`) are also passed to LLGuidance.
+
+For your existing GBNF grammars, you can use [gbnf_to_lark.py script](https://github.com/guidance-ai/llguidance/blob/main/scripts/gbnf_to_lark.py) to convert them to LLGuidance Lark-like format.
+
+## Performance
+
+Computing a "token mask" (i.e., the set of allowed tokens) for a llama3 tokenizer with 128k tokens takes, on average, 50μs of single-core CPU time for the [JSON Schema Bench](https://github.com/guidance-ai/jsonschemabench). The p99 time is 0.5ms, and the p100 time is 20ms. These results are due to the lexer/parser split and several [optimizations](https://github.com/guidance-ai/llguidance/blob/main/docs/optimizations.md).
+
+## JSON Schema
+
+LLGuidance adheres closely to the JSON Schema specification. For example:
+
+- `additionalProperties` defaults to `true`, unlike current grammars, though you can set `"additionalProperties": false` if needed.
+- any whitespace is allowed.
+- The definition order in the `"properties": {}` object is maintained, regardless of whether properties are required (current grammars always puts required properties first).
+
+Unsupported schemas result in an error message—no keywords are silently ignored.
+
+## Why Not Reuse GBNF Format?
+
+GBNF lacks the concept of a lexer.
+
+Most programming languages, including JSON, use a two-step process: a lexer (built with regular expressions) converts a byte stream into lexemes, which are then processed by a CFG parser. This approach is faster because lexers are cheaper to evaluate, and there is ~10x fewer lexemes than bytes.
+LLM tokens often align with lexemes, so the parser is engaged in under 0.5% of tokens, with the lexer handling the rest.
+
+However, the user has to provide the distinction between lexemes and CFG symbols. In [Lark](https://github.com/lark-parser/lark), lexeme names are uppercase, while CFG symbols are lowercase.
+The [gbnf_to_lark.py script](https://github.com/guidance-ai/llguidance/blob/main/scripts/gbnf_to_lark.py) can often take care of this automatically.
+See [LLGuidance syntax docs](https://github.com/guidance-ai/llguidance/blob/main/docs/syntax.md#terminals-vs-rules) for more details.
+
+## Error Handling
+
+Errors are currently printed to `stderr`, and generation continues. Improved error handling may be added in the future.

tests/CMakeLists.txt

@@ -86,6 +86,9 @@ llama_test(test-tokenizer-0 NAME test-tokenizer-0-qwen2             ARGS ${CMAKE
 llama_test(test-tokenizer-0 NAME test-tokenizer-0-refact            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
 llama_test(test-tokenizer-0 NAME test-tokenizer-0-starcoder         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
 
+if (LLAMA_LLGUIDANCE)
+    llama_target_and_test(test-grammar-llguidance.cpp ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf)
+endif ()
 
 if (NOT WIN32)
     # these tests are disabled on Windows because they use internal functions not exported with LLAMA_API

tests/test-grammar-integration.cpp

@@ -129,7 +129,7 @@ static void test_grammar(const std::string & test_desc, const std::string & gram
     test(test_desc + ". Grammar: " + grammar_str, grammar_str, passing_strings, failing_strings);
 }
 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)), passing_strings, failing_strings);
+    test(test_desc + ". Schema: " + schema_str, json_schema_to_grammar(json::parse(schema_str), true), passing_strings, failing_strings);
 }
 
 static void test_simple_grammar() {

tests/test-json-schema-to-grammar.cpp

@@ -1246,7 +1246,7 @@ int main() {
 
     test_all("C++", [](const TestCase & tc) {
         try {
-            tc.verify(json_schema_to_grammar(nlohmann::ordered_json::parse(tc.schema)));
+            tc.verify(json_schema_to_grammar(nlohmann::ordered_json::parse(tc.schema), true));
             tc.verify_status(SUCCESS);
         } catch (const std::runtime_error & ex) {
             fprintf(stderr, "Error: %s\n", ex.what());