forked from QwenLM/qwen.cpp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.cpp
225 lines (200 loc) · 8.52 KB
/
main.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
#include "qwen.h"
#include <iomanip>
#include <iostream>
enum InferenceMode {
INFERENCE_MODE_CHAT,
INFERENCE_MODE_GENERATE,
};
static inline auto to_inference_mode(const std::string &s) -> InferenceMode {
static std::unordered_map<std::string, InferenceMode> m{{"chat", INFERENCE_MODE_CHAT},
{"generate", INFERENCE_MODE_GENERATE}};
return m.at(s);
}
struct Args {
std::string model_path = "qwen-ggml.bin";
std::string tiktoken_path = "qwen.tiktoken";
InferenceMode mode = INFERENCE_MODE_CHAT;
std::string prompt = "你好";
int max_length = 2048;
int max_context_length = 512;
bool interactive = false;
int top_k = 0;
float top_p = 0.5;
float temp = 0.95;
float repeat_penalty = 1.0;
int num_threads = 0;
bool verbose = false;
};
static auto usage(const std::string &prog) -> void {
std::cout << "Usage: " << prog << " [options]\n"
<< "\n"
<< "options:\n"
<< " -h, --help show this help message and exit\n"
<< " -m, --model PATH model path (default: qwen-ggml.bin)\n"
<< " --mode inference mode chose from {chat, generate} (default: chat)\n"
<< " -p, --prompt PROMPT prompt to start generation with (default: 你好)\n"
<< " -i, --interactive run in interactive mode\n"
<< " -l, --max_length N max total length including prompt and output (default: 2048)\n"
<< " -c, --max_context_length N\n"
<< " max context length (default: 512)\n"
<< " --top_k N top-k sampling (default: 0)\n"
<< " --top_p N top-p sampling (default: 0.7)\n"
<< " --temp N temperature (default: 0.95)\n"
<< " --repeat_penalty N penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled)\n"
<< " -t, --threads N number of threads for inference\n"
<< " -v, --verbose display verbose output including config/system/performance info\n";
}
static auto parse_args(const std::vector<std::string> &argv) -> Args {
Args args;
for (size_t i = 1; i < argv.size(); i++) {
const std::string &arg = argv[i];
if (arg == "-h" || arg == "--help") {
usage(argv[0]);
exit(EXIT_SUCCESS);
} else if (arg == "-m" || arg == "--model") {
args.model_path = argv[++i];
} else if (arg == "--tiktoken") {
args.tiktoken_path = argv[++i];
} else if (arg == "--mode") {
args.mode = to_inference_mode(argv[++i]);
} else if (arg == "-p" || arg == "--prompt") {
args.prompt = argv[++i];
} else if (arg == "-i" || arg == "--interactive") {
args.interactive = true;
} else if (arg == "-l" || arg == "--max_length") {
args.max_length = std::stoi(argv[++i]);
} else if (arg == "-c" || arg == "--max_context_length") {
args.max_context_length = std::stoi(argv[++i]);
} else if (arg == "--top_k") {
args.top_k = std::stoi(argv[++i]);
} else if (arg == "--top_p") {
args.top_p = std::stof(argv[++i]);
} else if (arg == "--temp") {
args.temp = std::stof(argv[++i]);
} else if (arg == "--repeat_penalty") {
args.repeat_penalty = std::stof(argv[++i]);
} else if (arg == "-t" || arg == "--threads") {
args.num_threads = std::stoi(argv[++i]);
} else if (arg == "-v" || arg == "--verbose") {
args.verbose = true;
} else {
std::cerr << "Unknown argument: " << arg << std::endl;
usage(argv[0]);
exit(EXIT_FAILURE);
}
}
return args;
}
static auto parse_args(int argc, char **argv) -> Args {
std::vector<std::string> argv_vec;
argv_vec.reserve(argc);
for (int i = 0; i < argc; i++) {
argv_vec.emplace_back(argv[i]);
}
return parse_args(argv_vec);
}
static auto get_utf8_line(std::string &line) -> bool {
return !!std::getline(std::cin, line);
}
static auto chat(Args &args) -> void {
ggml_time_init();
int64_t start_load_us = ggml_time_us();
qwen::Pipeline pipeline(args.model_path, args.tiktoken_path);
int64_t end_load_us = ggml_time_us();
std::string model_name = "qwen";
auto text_streamer = std::make_shared<qwen::TextStreamer>(std::cout, pipeline.tokenizer.get());
auto perf_streamer = std::make_shared<qwen::PerfStreamer>();
auto streamer = std::make_shared<qwen::StreamerGroup>(
std::vector<std::shared_ptr<qwen::BaseStreamer>>{text_streamer, perf_streamer});
qwen::GenerationConfig gen_config(args.max_length, args.max_context_length, args.temp > 0, args.top_k,
args.top_p, args.temp, args.repeat_penalty, args.num_threads);
if (args.verbose) {
std::cout << "system info: | "
<< "AVX = " << ggml_cpu_has_avx() << " | "
<< "AVX2 = " << ggml_cpu_has_avx2() << " | "
<< "AVX512 = " << ggml_cpu_has_avx512() << " | "
<< "AVX512_VBMI = " << ggml_cpu_has_avx512_vbmi() << " | "
<< "AVX512_VNNI = " << ggml_cpu_has_avx512_vnni() << " | "
<< "FMA = " << ggml_cpu_has_fma() << " | "
<< "NEON = " << ggml_cpu_has_neon() << " | "
<< "ARM_FMA = " << ggml_cpu_has_arm_fma() << " | "
<< "F16C = " << ggml_cpu_has_f16c() << " | "
<< "FP16_VA = " << ggml_cpu_has_fp16_va() << " | "
<< "WASM_SIMD = " << ggml_cpu_has_wasm_simd() << " | "
<< "BLAS = " << ggml_cpu_has_blas() << " | "
<< "SSE3 = " << ggml_cpu_has_sse3() << " | "
<< "VSX = " << ggml_cpu_has_vsx() << " |\n";
std::cout << "inference config: | "
<< "max_length = " << args.max_length << " | "
<< "max_context_length = " << args.max_context_length << " | "
<< "top_k = " << args.top_k << " | "
<< "top_p = " << args.top_p << " | "
<< "temperature = " << args.temp << " | "
<< "num_threads = " << args.num_threads << " |\n";
std::cout << "loaded qwen model from " << args.model_path
<< " within: " << (end_load_us - start_load_us) / 1000.f << " ms\n";
std::cout << std::endl;
}
if (args.mode != INFERENCE_MODE_CHAT && args.interactive) {
std::cerr << "interactive demo is only supported for chat mode, falling back to non-interactive one\n";
args.interactive = false;
}
if (args.interactive) {
std::cout << R"( _____ )" << '\n'
<< R"(| _ | )" << '\n'
<< R"(| | | | __ __ ___ _ __ ___ _ __ _ __ )" << '\n'
<< R"(| | | | \ \ /\ / / / _ \ | '_ \ / __| | '_ \ | '_ \ )" << '\n'
<< R"(\ \/' / \ V V / | __/ | | | | _ | (__ | |_) | | |_) |)" << '\n'
<< R"( \_/\_\ \_/\_/ \___| |_| |_| (_) \___| | .__/ | .__/ )" << '\n'
<< R"( | | | | )" << '\n'
<< R"( |_| |_| )" << '\n'
<< '\n';
std::cout
<< "Welcome to Qwen.cpp! Ask whatever you want. Type 'clear' to clear context. Type 'stop' to exit.\n"
<< "\n";
std::vector<std::string> history;
while (1) {
std::cout << std::setw(model_name.size()) << std::left << "Prompt"
<< " > " << std::flush;
std::string prompt;
if (!get_utf8_line(prompt) || prompt == "stop") {
break;
}
if (prompt.empty()) {
continue;
}
if (prompt == "clear") {
history.clear();
continue;
}
history.emplace_back(std::move(prompt));
std::cout << model_name << " > ";
std::string output = pipeline.chat(history, gen_config, streamer.get());
history.emplace_back(std::move(output));
if (args.verbose) {
std::cout << "\n" << perf_streamer->to_string() << "\n\n";
}
perf_streamer->reset();
}
std::cout << "Bye\n";
} else {
if (args.mode == INFERENCE_MODE_CHAT) {
pipeline.chat({args.prompt}, gen_config, streamer.get());
} else {
pipeline.generate(args.prompt, gen_config, streamer.get());
}
if (args.verbose) {
std::cout << "\n" << perf_streamer->to_string() << "\n\n";
}
}
}
int main(int argc, char **argv) {
try {
Args args = parse_args(argc, argv);
chat(args);
} catch (std::exception &e) {
std::cerr << e.what() << std::endl;
exit(EXIT_FAILURE);
}
return 0;
}