iamlemec · snowyu · Feb 6, 2024
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -26,6 +26,8 @@ option(GGML_FMA                    "ggml: enable FMA"
 option(GGML_CUBLAS                 "ggml: use cuBLAS"                                    OFF)
 option(GGML_METAL                  "ggml: use Metal"                                     OFF)
 
+option(BERT_BUILD_TESTS            "bert: Build tests"                                    ON)
+
 #
 # Compile flags
 #
@@ -94,3 +96,8 @@ install(TARGETS ggml LIBRARY DESTINATION bert_cpp)
 
 # add bert
 add_subdirectory(src)
+
+if (BERT_BUILD_TESTS)
+    include(CTest)
+    add_subdirectory(tests)
+endif ()
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
@@ -0,0 +1,24 @@
+include_directories(${CMAKE_SOURCE_DIR}/src)
+
+# add_executable(test_tokenizer test_tokenizer.cpp)
+# target_link_libraries(test_tokenizer PRIVATE bert ggml)
+
+set(TEST_MODEL_NAME "bge-large-zh-v1.5")
+
+function(bert_build_executable source)
+    get_filename_component(TEST_TARGET ${source} NAME_WE)
+    add_executable(${TEST_TARGET} ${source})
+    install(TARGETS ${TEST_TARGET} RUNTIME)
+    target_link_libraries(${TEST_TARGET} PRIVATE bert ggml)
+endfunction()
+
+function(bert_test_executable name source)
+    get_filename_component(TEST_TARGET ${source} NAME_WE)
+    add_test(NAME "Generate_HF_tokens" COMMAND python3 ${CMAKE_CURRENT_SOURCE_DIR}/test_hf_tokenizer.py ${TEST_MODEL_NAME})
+    add_test(NAME ${name} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
+    set_property(TEST ${name} PROPERTY LABELS "main")
+endfunction()
+
+
+bert_build_executable(test_tokenizer.cpp)
+bert_test_executable (test_tokenizer test_tokenizer.cpp -m ${CMAKE_CURRENT_SOURCE_DIR}/../models/${TEST_MODEL_NAME}/bge-large-zh-v1.5-q4_1.gguf)
diff --git a/tests/test.sh b/tests/test.sh
@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+
+set -e
+
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
+MODEL_NAME=${1:-bge-large-zh-v1.5}
+MODEL_DIR=$(realpath "$SCRIPT_DIR/../models/$MODEL_NAME")
+
+if [ ! -d "$MODEL_DIR" ]; then
+  python3 $SCRIPT_DIR/../models/download-repo.py $MODEL_NAME
+fi
+
+if [ ! -d "$MODEL_DIR/ggml-model-q4_1.gguf" ]; then
+  $SCRIPT_DIR/../models/run_conversions.sh $MODEL_NAME q4_1
+fi
+
+python3 $SCRIPT_DIR/test_hf_tokenizer.py $MODEL_DIR
+
+$SCRIPT_DIR/../build/bin/test_tokenizer -m $MODEL_DIR/ggml-model-q4_1.gguf
diff --git a/tests/test_hf_tokenizer.py b/tests/test_hf_tokenizer.py
@@ -0,0 +1,39 @@
+from ast import arg
+from transformers import AutoTokenizer, AutoModel
+import argparse
+import os
+
+SCRIPT_PATH=os.path.dirname(os.path.realpath(__file__))
+
+def main(args):
+    # tokenizer_name = "sentence-transformers/multi-qa-MiniLM-L6-cos-v1"
+    if "/" in args.model_name:
+        tokenizer_name = args.model_name
+    elif "MiniLM" in args.model_name:
+        tokenizer_name = f"sentence-transformers/{args.model_name}"
+    elif "bge-" in args.model_name:
+        tokenizer_name = f"BAAI/{args.model_name}"
+    else:
+        raise ValueError(f"Unknown model name: {args.model_name}")
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+
+    with open(SCRIPT_PATH + "/test_prompts.txt", "r", encoding="utf-8") as f:
+        inps = f.readlines()
+        inps = list(map(lambda x: x.strip(), inps))
+
+    print("Using tokenizer:", tokenizer_name)
+    output = []
+    for inp in inps:
+        oup = tokenizer(inp, return_tensors="pt").input_ids[0].tolist()
+        output.append(",".join([str(x) for x in oup]))
+        for token in oup:
+            print(f"{token} <--> {tokenizer.decode([token])}")
+
+    with open(SCRIPT_PATH + "/hf_tokenized_ids.txt", "w", encoding="utf-8") as f:
+        f.write("\n".join(output))
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Download original repo files')
+    parser.add_argument('model_name', type=str, help='Name of the repo')
+    args = parser.parse_args()
+    main(args)
diff --git a/tests/test_prompts.txt b/tests/test_prompts.txt
@@ -0,0 +1,8 @@
+hello world
+i'm going to the store to buy 3 apples and a banana! you're welcome to come along if you'd like. the time is 2:30 p.m. and it's partly cloudy outside. i'll be back soon, so don't go anywhere.
+"5 2 + 3 * 4 -"; int stack[1000], top = -1; int calculate(int a, int b, char operator) { return operator == \'+\' ? a + b : operator == \'-\' ? a - b : operator == \'*\' ? a * b : a / b; } void push(int x) { stack[++top] = x; } int pop() { return stack[top--]; } int evaluatepostfix(char* expression) { for (int i = 0; expression[i]; i++) { if (isdigit(expression[i])) push(expression[i] - \'0\'); else { int a = pop(), b = pop(); push(calculate(b, a, expression[i])); } } return pop(); } int result = evaluatepostfix(input);
+你好，世界！
+こんにちは、世界！
+1231 2431431
+你好我是gpt
+然而，分音符号（diaeresis）和变音符号（umlaut）在一些情况下也可以被泛称为 "accent"，这是因为它们都是附加在字母上的符号，用于改变字母的原始发音。
diff --git a/tests/test_tokenizer.cpp b/tests/test_tokenizer.cpp
@@ -0,0 +1,193 @@
+#ifdef NDEBUG
+#undef NDEBUG
+#endif
+
+#include "bert.h"
+#include "ggml.h"
+
+#include <cassert>
+#include <unistd.h>
+#include <map>
+#include <algorithm>
+#include <stdio.h>
+#include <string>
+#include <vector>
+#include <fstream>
+#include <sstream>
+#define ANSI_COLOR_RED     "\x1b[31m"
+#define ANSI_COLOR_RESET   "\x1b[0m"
+#define ANSI_COLOR_GREEN   "\x1b[32m"
+
+
+std::vector<std::string> txt2list(const std::string& filename) {
+    std::ifstream file(filename);
+    std::vector<std::string> all_lines;
+
+    if (!file.is_open()) {
+        printf("can not open file: %s\n", filename.c_str());
+        return all_lines;
+    }
+
+    std::string line;
+    while (std::getline(file, line)) {
+        all_lines.push_back(line);
+    }
+
+    file.close();
+    return all_lines;
+}
+
+std::vector<std::vector<int>> read_expected_tokenids(const std::string& filename) {
+    std::ifstream file(filename);
+    std::vector<std::vector<int>> all_numbers;
+
+    if (!file.is_open()) {
+        printf("can not open file: %s\n", filename.c_str());
+        return all_numbers;
+    }
+
+
+    std::string line;
+    while (std::getline(file, line)) {
+        std::vector<int> line_numbers;
+        std::istringstream iss(line);
+        std::string number_str;
+
+        while (std::getline(iss, number_str, ',')) {
+            line_numbers.push_back(std::stoi(number_str));
+        }
+
+        all_numbers.push_back(line_numbers);
+    }
+
+    file.close();
+    return all_numbers;
+}
+
+void tokenizer_test(bert_ctx * ctx, const std::string& input, const bert_tokens& expected) {
+    int N = bert_n_max_tokens(ctx);
+    bert_tokens result = bert_tokenize(ctx, input, N);
+    int n_tokens;
+
+    if (result != expected) {
+        printf("tokenizer test failed: '%.*s'\n", 16000, input.data());
+
+        printf("[");
+        for (auto& tok : result) {
+            printf("%d, ", tok);
+        }
+        printf("]\n");
+
+        for (size_t i = 0; i < result.size(); i++) {
+            bert_token a = expected[std::min(i, expected.size()-1)];
+            bert_token b = result[i];
+            const char *color_start = (a == b) ? ANSI_COLOR_GREEN : ANSI_COLOR_RED;
+            const char *color_end = ANSI_COLOR_RESET;
+
+            printf("%s%d -> %s : %d -> %s%s\n", color_start, a, bert_vocab_id_to_token(ctx, a), b, bert_vocab_id_to_token(ctx, b), color_end);
+        }
+    } else {
+        printf("Success '%.*s...'\n", 16, input.data());
+    }
+    assert(result == expected);
+  }
+
+
+struct bert_params
+{
+    int32_t n_threads = 6;
+    const char* model = "models/all-MiniLM-L6-v2/ggml-model-q4_0.bin";
+    const char* prompt = "test prompt";
+    int32_t batch_size = 32;
+    bool use_cpu = false;
+};
+
+void bert_print_usage(char **argv, const bert_params &params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model);
+    fprintf(stderr, "                        batch size to use when executing model\n");
+    fprintf(stderr, "  -c, --cpu             use CPU backend (default: use CUDA if available)\n");
+    fprintf(stderr, "\n");
+}
+
+bool bert_params_parse(int argc, char **argv, bert_params &params) {
+    for (int i = 1; i < argc; i++)
+    {
+        std::string arg = argv[i];
+
+        if (arg == "-m" || arg == "--model") {
+            params.model = argv[++i];
+        } else if (arg == "-c" || arg == "--cpu") {
+            params.use_cpu = true;
+        } else if (arg == "-h" || arg == "--help") {
+            bert_print_usage(argv, params);
+            exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            bert_print_usage(argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+
+    bert_params params;
+    params.model = "models/all-MiniLM-L6-v2/ggml-model-q4_0.bin";
+
+    if (bert_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+
+    bert_ctx * bctx;
+
+    // load the model
+    {
+        if ((bctx = bert_load_from_file(params.model, params.use_cpu)) == nullptr) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model);
+            return 1;
+        }
+    }
+    std::string dir = params.model;
+    std::size_t i = dir.rfind("/models/");
+    if (i != std::string::npos) {
+        dir.resize(i);
+    } else {
+        dir = ".";
+    }
+
+    auto expected = read_expected_tokenids(dir + "/tests/hf_tokenized_ids.txt");
+    auto prompts = txt2list(dir + "/tests/test_prompts.txt");
+
+    if (expected.size() == 0 || prompts.size() == 0) {
+        printf("failed to read test data\n");
+        return 1;
+    }
+
+    if (expected.size() != prompts.size()) {
+        printf("test data size mismatch\n");
+        return 1;
+    }
+
+    // tokenizer tests:
+    for (size_t i = 0; i < prompts.size(); i++) {
+        tokenizer_test(bctx, prompts[i], expected[i]);
+    }
+
+    // tokenizer_test(bctx, "1231 2431431", {101, 13138, 2487, 22884, 16932, 21486, 102});
+    // tokenizer_test(bctx, "Québec", {101, 5447, 102});
+    // tokenizer_test(bctx, "syömme \t  täällä tänään", {101, 25353, 5358, 4168, 11937, 25425, 9092, 14634, 102});
+    // tokenizer_test(bctx, "I'm going to the store to buy 3 apples and a banana! You're welcome to come along if you'd like. The time is 2:30 p.m. and it's partly cloudy outside. I'll be back soon, so don't go anywhere.", {101, 1045, 1005, 1049, 2183, 2000, 1996, 3573, 2000, 4965, 1017, 18108, 1998, 1037, 15212, 999, 2017, 1005, 2128, 6160, 2000, 2272, 2247, 2065, 2017, 1005, 1040, 2066, 1012, 1996, 2051, 2003, 1016, 1024, 2382, 1052, 1012, 1049, 1012, 1998, 2009, 1005, 1055, 6576, 24706, 2648, 1012, 1045, 1005, 2222, 2022, 2067, 2574, 1010, 2061, 2123, 1005, 1056, 2175, 5973, 1012, 102});
+    // tokenizer_test(bctx, "\"5 2 + 3 * 4 -\"; int stack[1000], top = -1; int calculate(int a, int b, char operator) { return operator == '+' ? a + b : operator == '-' ? a - b : operator == '*' ? a * b : a / b; } void push(int x) { stack[++top] = x; } int pop() { return stack[top--]; } int evaluatePostfix(char* expression) { for (int i = 0; expression[i]; i++) { if (isdigit(expression[i])) push(expression[i] - '0'); else { int a = pop(), b = pop(); push(calculate(b, a, expression[i])); } } return pop(); } int result = evaluatePostfix(input);", {101, 1000, 1019, 1016, 1009, 1017, 1008, 1018, 1011, 1000, 1025, 20014, 9991, 1031, 6694, 1033, 1010, 2327, 1027, 1011, 1015, 1025, 20014, 18422, 1006, 20014, 1037, 1010, 20014, 1038, 1010, 25869, 6872, 1007, 1063, 2709, 6872, 1027, 1027, 1005, 1009, 1005, 1029, 1037, 1009, 1038, 1024, 6872, 1027, 1027, 1005, 1011, 1005, 1029, 1037, 1011, 1038, 1024, 6872, 1027, 1027, 1005, 1008, 1005, 1029, 1037, 1008, 1038, 1024, 1037, 1013, 1038, 1025, 1065, 11675, 5245, 1006, 20014, 1060, 1007, 1063, 9991, 1031, 1009, 1009, 2327, 1033, 1027, 1060, 1025, 1065, 20014, 3769, 1006, 1007, 1063, 2709, 9991, 1031, 2327, 1011, 1011, 1033, 1025, 1065, 20014, 16157, 19894, 8873, 2595, 1006, 25869, 1008, 3670, 1007, 1063, 2005, 1006, 20014, 1045, 1027, 1014, 1025, 3670, 1031, 1045, 1033, 1025, 1045, 1009, 1009, 1007, 1063, 2065, 1006, 2003, 4305, 23806, 1006, 3670, 1031, 1045, 1033, 1007, 1007, 5245, 1006, 3670, 1031, 1045, 1033, 1011, 1005, 1014, 1005, 1007, 1025, 2842, 1063, 20014, 1037, 1027, 3769, 1006, 1007, 1010, 1038, 1027, 3769, 1006, 1007, 1025, 5245, 1006, 18422, 1006, 1038, 1010, 1037, 1010, 3670, 1031, 1045, 1033, 1007, 1007, 1025, 1065, 1065, 2709, 3769, 1006, 1007, 1025, 1065, 20014, 2765, 1027, 16157, 19894, 8873, 2595, 1006, 7953, 1007, 1025, 102});
+
+    // tokenizer_test(bctx, "Hello world!", {101, 7592, 2088,  999,  102});
+    // tokenizer_test(bctx, "你好，世界！", {101, 100, 100, 1989, 1745, 100, 1986, 102});
+    // tokenizer_test(bctx, "こんにちは、世界！", {101,  1655, 30217, 30194, 30188, 30198, 1635, 1745, 100, 1986, 102});
+}