cmake : add sanitizer flags for llama.cpp (ggml-org#11279)

* cmake : add sanitizer flags for llama.cpp

ggml-ci

* tests : fix compile warnings

ggml-ci

* cmake : move sanitizer flags to llama_add_compile_flags

ggml-ci

* cmake : move llama.cpp compile flags to top level lists

ggml-ci

* cmake : apply only sanitizer flags at top level

ggml-ci

* tests : fix gguf context use in same_tensor_data

* gguf-test: tensor data comparison

* dummy : trigger ggml-ci

* unicode : silence gcc warnings

ggml-ci

* ci : use sanitizer builds only in Debug mode

ggml-ci

* cmake : add status messages [no ci]

---------

Co-authored-by: Johannes Gäßler <[email protected]>

Author: ggerganov

.github/workflows/build.yml

@@ -234,7 +234,7 @@ jobs:
     strategy:
       matrix:
         sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
+        build_type: [Debug]
 
     steps:
       - name: Clone

CMakeLists.txt

@@ -83,11 +83,8 @@ include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
 
 # override ggml options
-set(GGML_SANITIZE_THREAD    ${LLAMA_SANITIZE_THREAD})
-set(GGML_SANITIZE_ADDRESS   ${LLAMA_SANITIZE_ADDRESS})
-set(GGML_SANITIZE_UNDEFINED ${LLAMA_SANITIZE_UNDEFINED})
-set(GGML_ALL_WARNINGS       ${LLAMA_ALL_WARNINGS})
-set(GGML_FATAL_WARNINGS     ${LLAMA_FATAL_WARNINGS})
+set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
+set(GGML_FATAL_WARNINGS ${LLAMA_FATAL_WARNINGS})
 
 # change the default for these ggml options
 if (NOT DEFINED GGML_LLAMAFILE)
@@ -117,16 +114,62 @@ llama_option_depr(WARNING     LLAMA_SYCL                GGML_SYCL)
 llama_option_depr(WARNING     LLAMA_SYCL_F16            GGML_SYCL_F16)
 llama_option_depr(WARNING     LLAMA_CANN                GGML_CANN)
 
+if (NOT MSVC)
+    if (LLAMA_SANITIZE_THREAD)
+        message(STATUS "Using -fsanitize=thread")
+
+        add_compile_options(-fsanitize=thread)
+        link_libraries     (-fsanitize=thread)
+    endif()
+
+    if (LLAMA_SANITIZE_ADDRESS)
+        message(STATUS "Using -fsanitize=address")
+
+        add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
+        link_libraries     (-fsanitize=address)
+    endif()
+
+    if (LLAMA_SANITIZE_UNDEFINED)
+        message(STATUS "Using -fsanitize=undefined")
+
+        add_compile_options(-fsanitize=undefined)
+        link_libraries     (-fsanitize=undefined)
+    endif()
+endif()
+
 #
-# build the library
+# 3rd-party
 #
 
 if (NOT TARGET ggml)
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
+
+#
+# build the library
+#
+
 add_subdirectory(src)
 
+#
+# utils, programs, examples and tests
+#
+
+if (LLAMA_BUILD_COMMON)
+    add_subdirectory(common)
+endif()
+
+if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
+    include(CTest)
+    add_subdirectory(tests)
+endif()
+
+if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+    add_subdirectory(pocs)
+endif()
+
 #
 # install
 #
@@ -200,21 +243,3 @@ configure_file(cmake/llama.pc.in
 
 install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
         DESTINATION lib/pkgconfig)
-
-#
-# utils, programs, examples and tests
-#
-
-if (LLAMA_BUILD_COMMON)
-    add_subdirectory(common)
-endif()
-
-if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
-    include(CTest)
-    add_subdirectory(tests)
-endif()
-
-if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES)
-    add_subdirectory(examples)
-    add_subdirectory(pocs)
-endif()

ggml/src/gguf.cpp

@@ -648,6 +648,10 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
             ok = ok && data != nullptr;
 
+            if (ok) {
+                ggml_set_name(data, "GGUF tensor data binary blob");
+            }
+
             // read the binary blob with the tensor data
             ok = ok && gr.read(data->data, ctx->size);
 

src/unicode.cpp

@@ -7,18 +7,17 @@
 
 #include <algorithm>
 #include <cassert>
+#include <codecvt>
 #include <cstddef>
 #include <cstdint>
+#include <locale>
 #include <map>
 #include <regex>
 #include <stdexcept>
 #include <string>
 #include <unordered_map>
-#include <unordered_set>
 #include <utility>
 #include <vector>
-#include <locale>
-#include <codecvt>
 
 size_t unicode_len_utf8(char src) {
     const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };

tests/CMakeLists.txt

@@ -1,3 +1,5 @@
+llama_add_compile_flags()
+
 function(llama_test target)
     include(CMakeParseArguments)
     set(options)

tests/test-gguf.cpp

@@ -48,7 +48,7 @@ enum handcrafted_file_type {
     HANDCRAFTED_DATA_CUSTOM_ALIGN          = 810 + offset_has_data,
 };
 
-std::string handcrafted_file_type_name(const enum handcrafted_file_type hft) {
+static std::string handcrafted_file_type_name(const enum handcrafted_file_type hft) {
     switch (hft) {
         case HANDCRAFTED_HEADER_BAD_MAGIC:           return "HEADER_BAD_MAGIC";
         case HANDCRAFTED_HEADER_BAD_VERSION_1:       return "HEADER_BAD_VERSION_1";
@@ -99,7 +99,7 @@ static bool expect_context_not_null(const enum handcrafted_file_type hft) {
 
 typedef std::pair<enum ggml_type, std::array<int64_t, GGML_MAX_DIMS>> tensor_config_t;
 
-std::vector<tensor_config_t> get_tensor_configs(std::mt19937 & rng) {
+static std::vector<tensor_config_t> get_tensor_configs(std::mt19937 & rng) {
     std::vector<tensor_config_t> tensor_configs;
     tensor_configs.reserve(100);
 
@@ -122,7 +122,7 @@ std::vector<tensor_config_t> get_tensor_configs(std::mt19937 & rng) {
     return tensor_configs;
 }
 
-std::vector<std::pair<enum gguf_type, enum gguf_type>> get_kv_types(std::mt19937 rng) {
+static std::vector<std::pair<enum gguf_type, enum gguf_type>> get_kv_types(std::mt19937 rng) {
     std::vector<std::pair<enum gguf_type, enum gguf_type>> kv_types;
     kv_types.reserve(100);
 
@@ -626,8 +626,6 @@ static bool handcrafted_check_tensor_data(const gguf_context * gguf_ctx, const u
 
     bool ok = true;
 
-    const uint32_t alignment = GGUF_DEFAULT_ALIGNMENT;
-
     for (int i = 0; i < int(tensor_configs.size()); ++i) {
         const ggml_type                          type  = tensor_configs[i].first;
         const std::array<int64_t, GGML_MAX_DIMS> shape = tensor_configs[i].second;
@@ -866,13 +864,13 @@ static struct random_gguf_context_result get_random_gguf_context(ggml_backend_t
                     case GGUF_TYPE_COUNT:
                     default: {
                         GGML_ABORT("fatal error");
-                    } break;
+                    }
                 }
             } break;
             case GGUF_TYPE_COUNT:
             default: {
                 GGML_ABORT("fatal error");
-            } break;
+            }
         }
     }
 
@@ -938,7 +936,7 @@ static bool all_kv_in_other(const gguf_context * ctx, const gguf_context * other
         }
 
         if (type == GGUF_TYPE_ARRAY) {
-            const int arr_n = gguf_get_arr_n(ctx, id);
+            const size_t arr_n = gguf_get_arr_n(ctx, id);
             if (arr_n != gguf_get_arr_n(other, idx_other)) {
                 ok = false;
                 continue;
@@ -953,7 +951,7 @@ static bool all_kv_in_other(const gguf_context * ctx, const gguf_context * other
             if (type_arr == GGUF_TYPE_BOOL) {
                 const int8_t * data       = reinterpret_cast<const int8_t *>(gguf_get_arr_data(ctx,   id));
                 const int8_t * data_other = reinterpret_cast<const int8_t *>(gguf_get_arr_data(other, idx_other));
-                for (int arr_i = 0; arr_i < arr_n; ++arr_i) {
+                for (size_t arr_i = 0; arr_i < arr_n; ++arr_i) {
                     if (bool(data[arr_i]) != bool(data_other[arr_i])) {
                         ok = false;
                     }
@@ -962,7 +960,7 @@ static bool all_kv_in_other(const gguf_context * ctx, const gguf_context * other
             }
 
             if (type_arr == GGUF_TYPE_STRING) {
-                for (int arr_i = 0; arr_i < arr_n; ++arr_i) {
+                for (size_t arr_i = 0; arr_i < arr_n; ++arr_i) {
                     const std::string str       = gguf_get_arr_str(ctx,   id,       arr_i);
                     const std::string str_other = gguf_get_arr_str(other, idx_other, arr_i);
                     if (str != str_other) {
@@ -1033,6 +1031,12 @@ static bool same_tensor_data(const struct ggml_context * orig, const struct ggml
 
     struct ggml_tensor * t_orig = ggml_get_first_tensor(orig);
     struct ggml_tensor * t_read = ggml_get_first_tensor(read);
+
+    if (std::string(t_read->name) != "GGUF tensor data binary blob") {
+        return false;
+    }
+    t_read = ggml_get_next_tensor(read, t_read);
+
     while (t_orig) {
         if (!t_read) {
             ok = false;
@@ -1051,13 +1055,13 @@ static bool same_tensor_data(const struct ggml_context * orig, const struct ggml
         }
 
         t_orig = ggml_get_next_tensor(orig, t_orig);
-        t_read = ggml_get_next_tensor(orig, t_read);
+        t_read = ggml_get_next_tensor(read, t_read);
     }
     if (t_read) {
         ok = false;
     }
 
-    return true;
+    return ok;
 }
 
 static std::pair<int, int> test_roundtrip(ggml_backend_dev_t dev, const unsigned int seed, const bool only_meta) {

tests/test-sampling.cpp

@@ -144,7 +144,6 @@ static void test_penalties(
 
     sampler_tester tester(probs, probs_expected);
 
-    const size_t n_vocab = probs.size();
     auto * sampler = llama_sampler_init_penalties(last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence);
 
     for (size_t i = 0; i < last_tokens.size(); i++) {