LLM inference in C/C++
C++ ggml
server: add component tests (#30312)
Xuan-Son Nguyen62a6f74bab74
diff --git a/CODEOWNERS b/CODEOWNERSindex df6905252c..b9762685b5 100644--- a/CODEOWNERS+++ b/CODEOWNERS@@ -97,6 +97,7 @@/src/models/ @CISC/tests/ @ggerganov/tests/test-chat.* @pwilkin+/tests/test-server-component.cpp @ggml-org/llama-server/tools/batched-bench/ @ggerganov/tools/cli/ @ngxson/tools/completion/ @ggerganovdiff --git a/skills/code-review/SKILL.md b/skills/code-review/SKILL.mdindex 7550085f21..9f14b4269b 100644--- a/skills/code-review/SKILL.md+++ b/skills/code-review/SKILL.md@@ -132,6 +132,7 @@ Public API changes carry a higher bar than internal ones (`CONTRIBUTING.md`). Re- Only add tests that bring meaningful results. Too-trivial tests just bloat the suite and CI.- No time-sensitive tests (timing thresholds, sleeps, races against wall-clock); they are flaky on CI.- Think twice about tests that significantly increase CI run time (expensive computation, large inputs, or long sleep/wait delays) or download large amounts of data from the internet (big models, datasets). Flag them and ask whether a smaller model/input or an existing fixture would do.+- Server: small tests of internal components (helpers, `server_tokens`, queues, format converters, ...) belong in `tests/test-server-component.cpp`. The end-to-end pytest suite in `tools/server/tests` is costly in CI time; only add to it when the test requires end-to-end behavior with a real model.## General (always)diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txtindex 5f9bd037f3..a732b6d655 100644--- a/tests/CMakeLists.txt+++ b/tests/CMakeLists.txt@@ -163,6 +163,9 @@ if (NOT WIN32 OR NOT BUILD_SHARED_LIBS)llama_build_and_test(test-chat.cpp WORKING_DIRECTORY ${PROJECT_SOURCE_DIR})target_include_directories(test-chat PRIVATE ${PROJECT_SOURCE_DIR}/tools/server)target_link_libraries(test-chat PRIVATE server-context)+ llama_build_and_test(test-server-component.cpp)+ target_include_directories(test-server-component PRIVATE ${PROJECT_SOURCE_DIR}/tools/server)+ target_link_libraries(test-server-component PRIVATE server-context)llama_build_and_test(test-json-schema-to-grammar.cpp)if (NOT GGML_BACKEND_DL)diff --git a/tests/test-server-component.cpp b/tests/test-server-component.cppnew file mode 100644index 0000000000..d2d103a45a--- /dev/null+++ b/tests/test-server-component.cpp@@ -0,0 +1,393 @@+#include "testing.h"++#include "../tools/server/server-common.h"++#include <cstdlib>+#include <string>+#include <thread>+#include <vector>++// unit tests for server internal components, no model required+// NOTE: end-to-end tests live in tools/server/tests+// see skills/code-review/SKILL.md for more++//+// server_tokens (text only, has_mtmd = false)+//++static void test_server_tokens_basic(testing & t) {+ server_tokens st(llama_tokens{1, 2, 3}, false);+ t.assert_equal("size", (size_t) 3, st.size());+ t.assert_true("not empty", !st.empty());+ t.assert_equal("operator[]", 2, st[1]);++ st.push_back(4);+ t.assert_equal("push_back", (size_t) 4, st.size());+ t.assert_true("get_tokens", st.get_tokens() == llama_tokens({1, 2, 3, 4}));++ st.insert({5, 6});+ t.assert_true("insert appends", st.get_tokens() == llama_tokens({1, 2, 3, 4, 5, 6}));++ st.set_token(0, 9);+ t.assert_equal("set_token", 9, st[0]);++ bool thrown = false;+ try {+ st.push_back(LLAMA_TOKEN_NULL);+ } catch (const std::exception &) {+ thrown = true;+ }+ t.assert_true("push_back(LLAMA_TOKEN_NULL) throws", thrown);+ t.assert_equal("size unchanged after throw", (size_t) 6, st.size());++ st.clear();+ t.assert_true("clear", st.empty());+}++static void test_server_tokens_pos(testing & t) {+ server_tokens st(llama_tokens{1, 2, 3, 4, 5}, false);+ t.assert_equal("pos_next()", (llama_pos) 5, st.pos_next());+ t.assert_equal("pos_next(0)", (llama_pos) 0, st.pos_next(0));+ t.assert_equal("pos_next(3)", (llama_pos) 3, st.pos_next(3));++ t.assert_equal("size_up_to_pos(0)", (size_t) 0, st.size_up_to_pos(0));+ t.assert_equal("size_up_to_pos(3)", (size_t) 3, st.size_up_to_pos(3));+ t.assert_equal("size_up_to_pos clamps", (size_t) 5, st.size_up_to_pos(100));+}++static void test_server_tokens_keep_first(testing & t) {+ server_tokens st(llama_tokens{1, 2, 3, 4, 5}, false);+ st.keep_first(5);+ t.assert_equal("keep all", (size_t) 5, st.size());+ st.keep_first(2);+ t.assert_true("keep 2", st.get_tokens() == llama_tokens({1, 2}));+ st.keep_first(0);+ t.assert_true("keep 0", st.empty());+}++static void test_server_tokens_common_prefix(testing & t) {+ server_tokens a(llama_tokens{1, 2, 3, 4}, false);+ server_tokens empty(llama_tokens{}, false);++ t.assert_equal("both empty", (size_t) 0, empty.get_common_prefix(empty));+ t.assert_equal("one empty", (size_t) 0, a.get_common_prefix(empty));+ t.assert_equal("identical", (size_t) 4, a.get_common_prefix(a));+ t.assert_equal("diverge at 2", (size_t) 2, a.get_common_prefix(server_tokens(llama_tokens{1, 2, 9, 4}, false)));+ t.assert_equal("diverge at 0", (size_t) 0, a.get_common_prefix(server_tokens(llama_tokens{9, 2, 3, 4}, false)));+ t.assert_equal("b is prefix", (size_t) 2, a.get_common_prefix(server_tokens(llama_tokens{1, 2}, false)));+ t.assert_equal("a is prefix", (size_t) 4, a.get_common_prefix(server_tokens(llama_tokens{1, 2, 3, 4, 5, 6}, false)));+}++static void test_server_tokens_serialize(testing & t) {+ // same packing as slot save/restore in server-context.cpp+ auto pack = [](const std::vector<char> & data) {+ llama_tokens packed(data.size() / sizeof(llama_token));+ memcpy(packed.data(), data.data(), data.size());+ return packed;+ };++ t.test("roundtrip", [&](testing & t) {+ server_tokens st(llama_tokens{10, 20, 30}, false);+ std::vector<char> data = st.serialize();+ t.assert_equal("aligned to llama_token", (size_t) 0, data.size() % sizeof(llama_token));++ server_tokens restored = server_tokens::deserialize(pack(data), false);+ t.assert_true("tokens", restored.get_tokens() == st.get_tokens());+ });++ t.test("roundtrip empty", [&](testing & t) {+ server_tokens st(llama_tokens{}, false);+ server_tokens restored = server_tokens::deserialize(pack(st.serialize()), false);+ t.assert_true("empty", restored.empty());+ });++ t.test("legacy plain token list", [&](testing & t) {+ server_tokens restored = server_tokens::deserialize(llama_tokens{7, 8, 9}, false);+ t.assert_true("tokens", restored.get_tokens() == llama_tokens({7, 8, 9}));+ });++ auto expect_throw = [](testing & t, const std::string & msg, const llama_tokens & packed) {+ bool thrown = false;+ try {+ server_tokens::deserialize(packed, false);+ } catch (const std::exception &) {+ thrown = true;+ }+ t.assert_true(msg, thrown);+ };++ t.test("corrupted", [&](testing & t) {+ llama_tokens good = pack(server_tokens(llama_tokens{10, 20, 30}, false).serialize());++ llama_tokens bad_version = good;+ bad_version[1] = 999;+ expect_throw(t, "bad version", bad_version);++ llama_tokens truncated = good;+ truncated.pop_back();+ expect_throw(t, "truncated", truncated);++ llama_tokens trailing = good;+ trailing.push_back(42);+ expect_throw(t, "trailing data", trailing);++ llama_tokens huge_count = good;+ huge_count[2] = 0x7fffffff;+ expect_throw(t, "huge token count", huge_count);++ expect_throw(t, "marker only", llama_tokens{LLAMA_TOKEN_NULL});+ });+}++static void test_server_tokens_clone(testing & t) {+ server_tokens st(llama_tokens{1, 2, 3}, false);+ server_tokens cl = st.clone();+ cl.push_back(4);+ t.assert_equal("original untouched", (size_t) 3, st.size());+ t.assert_equal("clone grew", (size_t) 4, cl.size());+ t.assert_true("get_text_tokens", cl.get_text_tokens() == llama_tokens({1, 2, 3, 4}));+}++static void test_server_tokens(testing & t) {+ t.test("basic", test_server_tokens_basic);+ t.test("pos", test_server_tokens_pos);+ t.test("keep_first", test_server_tokens_keep_first);+ t.test("common_prefix", test_server_tokens_common_prefix);+ t.test("serialize", test_server_tokens_serialize);+ t.test("clone", test_server_tokens_clone);+}++//+// string / json helpers+//++static void test_utf8(testing & t) {+ t.test("validate_utf8", [](testing & t) {+ t.assert_equal("empty", (size_t) 0, validate_utf8(""));+ t.assert_equal("ascii", (size_t) 3, validate_utf8("abc"));+ t.assert_equal("full 2-byte", (size_t) 3, validate_utf8("a\xC3\xA9"));+ t.assert_equal("full 3-byte", (size_t) 4, validate_utf8("a\xE2\x82\xAC"));+ t.assert_equal("full 4-byte", (size_t) 5, validate_utf8("a\xF0\x9F\x98\x80"));+ t.assert_equal("cut 2-byte (1/2)", (size_t) 1, validate_utf8("a\xC3"));+ t.assert_equal("cut 3-byte (1/3)", (size_t) 1, validate_utf8("a\xE2"));+ t.assert_equal("cut 3-byte (2/3)", (size_t) 1, validate_utf8("a\xE2\x82"));+ t.assert_equal("cut 4-byte (1/4)", (size_t) 1, validate_utf8("a\xF0"));+ t.assert_equal("cut 4-byte (2/4)", (size_t) 1, validate_utf8("a\xF0\x9F"));+ t.assert_equal("cut 4-byte (3/4)", (size_t) 1, validate_utf8("a\xF0\x9F\x98"));+ t.assert_equal("cut only", (size_t) 0, validate_utf8("\xF0\x9F\x98"));+ t.assert_equal("full then cut", (size_t) 4, validate_utf8("\xF0\x9F\x98\x80\xC3"));+ });++ t.test("is_valid_utf8", [](testing & t) {+ t.assert_true("empty", is_valid_utf8(""));+ t.assert_true("ascii", is_valid_utf8("hello"));+ t.assert_true("multi-byte", is_valid_utf8("\xC3\xA9\xE2\x82\xAC\xF0\x9F\x98\x80"));+ t.assert_true("truncated", !is_valid_utf8("\xE2\x82"));+ t.assert_true("bad cont. byte", !is_valid_utf8("\xC3\x41"));+ t.assert_true("lone cont. byte", !is_valid_utf8("\x80"));+ t.assert_true("invalid lead", !is_valid_utf8("\xFF"));+ });+}++static void test_sse(testing & t) {+ t.test("oai", [](testing & t) {+ t.assert_equal("single", std::string("data: {\"a\":1}\n\n"), format_oai_sse(json{{"a", 1}}));+ t.assert_equal("array", std::string("data: {\"a\":1}\n\ndata: {\"b\":2}\n\n"),+ format_oai_sse(json::array({json{{"a", 1}}, json{{"b", 2}}})));+ t.assert_equal("invalid utf8 replaced", std::string("data: \"x\xEF\xBF\xBD\"\n\n"), format_oai_sse(json("x\xC3")));+ });++ t.test("oai_resp", [](testing & t) {+ json ev = {{"event", "response.created"}, {"data", {{"id", "r1"}}}};+ t.assert_equal("single", std::string("event: response.created\ndata: {\"id\":\"r1\"}\n\n"), format_oai_resp_sse(ev));+ t.assert_equal("array", std::string("event: response.created\ndata: {\"id\":\"r1\"}\n\nevent: response.created\ndata: {\"id\":\"r1\"}\n\n"),+ format_oai_resp_sse(json::array({ev, ev})));+ });++ t.test("anthropic", [](testing & t) {+ json ev = {{"event", "message_start"}, {"data", {{"type", "message_start"}}}};+ t.assert_equal("with event", std::string("event: message_start\ndata: {\"type\":\"message_start\"}\n\n"), format_anthropic_sse(ev));+ t.assert_equal("data only", std::string("data: {\"type\":\"ping\"}\n\n"), format_anthropic_sse(json{{"type", "ping"}}));+ });+}++static void test_json_helpers(testing & t) {+ t.test("json_is_array_of_numbers", [](testing & t) {+ t.assert_true("ints", json_is_array_of_numbers(json::array({1, 2, 3})));+ t.assert_true("empty array", json_is_array_of_numbers(json::array()));+ t.assert_true("float", !json_is_array_of_numbers(json::array({1, 2.5})));+ t.assert_true("string", !json_is_array_of_numbers(json::array({1, "a"})));+ t.assert_true("not an array", !json_is_array_of_numbers(json(1)));+ });++ t.test("json_is_array_of_mixed_numbers_strings", [](testing & t) {+ t.assert_true("mixed", json_is_array_of_mixed_numbers_strings(json::array({"a", 1})));+ t.assert_true("numbers only", !json_is_array_of_mixed_numbers_strings(json::array({1, 2})));+ t.assert_true("strings only", !json_is_array_of_mixed_numbers_strings(json::array({"a", "b"})));+ t.assert_true("empty", !json_is_array_of_mixed_numbers_strings(json::array()));+ });++ t.test("json_is_array_and_contains_numbers", [](testing & t) {+ t.assert_true("contains", json_is_array_and_contains_numbers(json::array({"a", 1})));+ t.assert_true("strings only", !json_is_array_and_contains_numbers(json::array({"a"})));+ t.assert_true("not an array", !json_is_array_and_contains_numbers(json(1)));+ });++ t.test("json_get_nested_values", [](testing & t) {+ json js = {{"a", {{"b", {{"c", 1}}}}}, {"x", 2}};+ json res = json_get_nested_values({"a/b/c", "x", "a/b", "a/missing", "x/y"}, js);+ t.assert_equal("nested", 1, res.at("a/b/c").get<int>());+ t.assert_equal("top", 2, res.at("x").get<int>());+ t.assert_true("object", res.at("a/b") == json{{"c", 1}});+ t.assert_true("missing", !res.contains("a/missing"));+ t.assert_true("not obj", !res.contains("x/y"));+ t.assert_equal("count", (size_t) 3, res.size());+ });+}++static void test_lora_helpers(testing & t) {+ // only pointer identity is used, never dereferenced+ auto * p0 = reinterpret_cast<llama_adapter_lora *>(0x10);+ auto * p1 = reinterpret_cast<llama_adapter_lora *>(0x20);++ auto make = [](llama_adapter_lora * ptr, float scale) {+ common_adapter_lora_info info;+ info.scale = scale;+ info.ptr = ptr;+ return info;+ };++ t.test("parse_lora_request", [](testing & t) {+ auto res = parse_lora_request(json::array({+ json{{"id", 0}, {"scale", 0.5}},+ json{{"id", 2}, {"scale", 1.0}},+ json{{"id", 0}, {"scale", 0.25}}, // last one wins+ }));+ t.assert_equal("count", (size_t) 2, res.size());+ t.assert_equal("id 0", 0.25f, res.at(0));+ t.assert_equal("id 2", 1.0f, res.at(2));++ auto defaults = parse_lora_request(json::array({json::object()}));+ t.assert_equal("default id/scale", 0.0f, defaults.at(-1));+ });++ t.test("are_lora_equal", [&](testing & t) {+ t.assert_true("both empty", are_lora_equal({}, {}));+ t.assert_true("same", are_lora_equal({make(p0, 1.0f)}, {make(p0, 1.0f)}));+ t.assert_true("diff scale", !are_lora_equal({make(p0, 1.0f)}, {make(p0, 0.5f)}));+ t.assert_true("diff ptr", !are_lora_equal({make(p0, 1.0f)}, {make(p1, 1.0f)}));+ t.assert_true("diff size", !are_lora_equal({make(p0, 1.0f)}, {}));+ });++ t.test("lora_get_enabled_ids", [&](testing & t) {+ auto ids = lora_get_enabled_ids({make(p0, 0.0f), make(p1, 0.5f), make(p0, -1.0f), make(p1, 2.0f)});+ t.assert_true("ids", ids == std::vector<size_t>({1, 3}));+ });+}++//+// server_pipe+//++static void test_server_pipe(testing & t) {+ auto stop_now = []() { return true; };++ t.test("drain then eof", [](testing & t) {+ server_pipe<int> p;+ t.assert_true("write 1", p.write(1));+ t.assert_true("write 2", p.write(2));+ p.close_write();++ int v = 0;+ t.assert_true("read 1", p.read(v, nullptr) && v == 1);+ t.assert_true("read 2", p.read(v, nullptr) && v == 2);+ t.assert_true("eof", !p.read(v, nullptr));+ });++ t.test("broken pipe", [](testing & t) {+ server_pipe<int> p;+ p.close_read();+ t.assert_true("write fails", !p.write(1));+ });++ t.test("max_size drops oldest", [](testing & t) {+ server_pipe<int> p;+ p.max_size = 2;+ p.write(1);+ p.write(2);+ p.write(3);+ p.close_write();++ int v = 0;+ t.assert_true("read 2", p.read(v, nullptr) && v == 2);+ t.assert_true("read 3", p.read(v, nullptr) && v == 3);+ t.assert_true("eof", !p.read(v, nullptr));+ });++ t.test("queued data wins over should_stop", [&](testing & t) {+ server_pipe<int> p;+ p.write(1);+ int v = 0;+ t.assert_true("read", p.read(v, stop_now) && v == 1);+ });++ t.test("should_stop closes the pipe", [&](testing & t) {+ server_pipe<int> p;+ int v = 0;+ t.assert_true("read stopped", !p.read(v, stop_now));+ t.assert_true("write fails", !p.write(1));+ });++ t.test("should_stop as deadline keeps the pipe open", [&](testing & t) {+ server_pipe<int> p;+ int v = 0;+ t.assert_true("read stopped", !p.read(v, stop_now, /* close_on_stop */ false));+ t.assert_true("write ok", p.write(1));+ t.assert_true("read again", p.read(v, nullptr) && v == 1);+ });++ t.test("cross thread", [](testing & t) {+ constexpr int n = 1000;+ server_pipe<int> p;+ std::thread producer([&]() {+ for (int i = 0; i < n; i++) {+ p.write(int(i));+ }+ p.close_write();+ });++ int v = 0;+ int count = 0;+ bool in_order = true;+ while (p.read(v, nullptr)) {+ in_order &= v == count;+ count++;+ }+ producer.join();++ t.assert_equal("count", n, count);+ t.assert_true("in order", in_order);+ });+}++int main(int argc, char ** argv) {+ testing t;++ const char * verbose = getenv("LLAMA_TEST_VERBOSE");+ if (verbose) {+ t.verbose = std::string(verbose) == "1";+ }++ if (argc > 1) {+ t.set_filter(argv[1]);+ }++ t.test("server_tokens", test_server_tokens);+ t.test("utf8", test_utf8);+ t.test("sse", test_sse);+ t.test("json_helpers", test_json_helpers);+ t.test("lora_helpers", test_lora_helpers);+ t.test("server_pipe", test_server_pipe);++ return t.summary();+}