/
main.rs
190 lines (170 loc) · 6.1 KB
/
main.rs
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
use clap::{crate_version, Arg, ArgAction, Command};
use once_cell::sync::OnceCell;
use serde::{Deserialize, Serialize};
const DEFAULT_CTX_SIZE: &str = "4096";
static CTX_SIZE: OnceCell<usize> = OnceCell::new();
fn main() -> Result<(), String> {
let matches = Command::new("llama-simple")
.version(crate_version!())
.arg(
Arg::new("prompt")
.short('p')
.long("prompt")
.value_name("PROMPT")
.help("Sets the prompt string, including system message if required.")
.required(true),
)
.arg(
Arg::new("model_alias")
.short('m')
.long("model-alias")
.value_name("ALIAS")
.help("Sets the model alias")
.default_value("default"),
)
.arg(
Arg::new("ctx_size")
.short('c')
.long("ctx-size")
.value_parser(clap::value_parser!(u32))
.value_name("CTX_SIZE")
.help("Sets the prompt context size")
.default_value(DEFAULT_CTX_SIZE),
)
.arg(
Arg::new("n_predict")
.short('n')
.long("n-predict")
.value_parser(clap::value_parser!(u32))
.value_name("N_PRDICT")
.help("Number of tokens to predict")
.default_value("1024"),
)
.arg(
Arg::new("n_gpu_layers")
.short('g')
.long("n-gpu-layers")
.value_parser(clap::value_parser!(u32))
.value_name("N_GPU_LAYERS")
.help("Number of layers to run on the GPU")
.default_value("100"),
)
.arg(
Arg::new("batch_size")
.short('b')
.long("batch-size")
.value_parser(clap::value_parser!(u32))
.value_name("BATCH_SIZE")
.help("Batch size for prompt processing")
.default_value("4096"),
)
.arg(
Arg::new("reverse_prompt")
.short('r')
.long("reverse-prompt")
.value_name("REVERSE_PROMPT")
.help("Halt generation at PROMPT, return control."),
)
.arg(
Arg::new("log_enable")
.long("log-enable")
.value_name("LOG_ENABLE")
.help("Enable trace logs")
.action(ArgAction::SetTrue),
)
.get_matches();
// model alias
let model_name = matches
.get_one::<String>("model_alias")
.unwrap()
.to_string();
// prompt
let prompt = matches.get_one::<String>("prompt").unwrap().to_string();
// create an `Options` instance
let mut options = Options::default();
// prompt context size
let ctx_size = matches.get_one::<u32>("ctx_size").unwrap();
CTX_SIZE
.set(*ctx_size as usize * 6)
.expect("Fail to parse prompt context size");
println!("[INFO] prompt context size: {size}", size = ctx_size);
// number of tokens to predict
let n_predict = matches.get_one::<u32>("n_predict").unwrap();
println!("[INFO] Number of tokens to predict: {n}", n = n_predict);
options.n_predict = *n_predict as u64;
// n_gpu_layers
let n_gpu_layers = matches.get_one::<u32>("n_gpu_layers").unwrap();
println!(
"[INFO] Number of layers to run on the GPU: {n}",
n = n_gpu_layers
);
options.n_gpu_layers = *n_gpu_layers as u64;
// batch size
let batch_size = matches.get_one::<u32>("batch_size").unwrap();
println!(
"[INFO] Batch size for prompt processing: {size}",
size = batch_size
);
options.batch_size = *batch_size as u64;
// reverse_prompt
if let Some(reverse_prompt) = matches.get_one::<String>("reverse_prompt") {
println!("[INFO] Reverse prompt: {prompt}", prompt = &reverse_prompt);
options.reverse_prompt = Some(reverse_prompt.to_string());
}
// log
let log_enable = matches.get_flag("log_enable");
println!("[INFO] Log enable: {enable}", enable = log_enable);
options.log_enable = log_enable;
// load the model into wasi-nn
let graph = wasmedge_wasi_nn::GraphBuilder::new(
wasmedge_wasi_nn::GraphEncoding::Ggml,
wasmedge_wasi_nn::ExecutionTarget::AUTO,
)
.build_from_cache(&model_name)
.expect("Failed to load the model");
// initialize the execution context
let mut context = graph
.init_execution_context()
.expect("Failed to init context");
// set metadata
let metadata = serde_json::to_string(&options).expect("Fail to serialize options");
context
.set_input(
1,
wasmedge_wasi_nn::TensorType::U8,
&[1],
metadata.as_bytes(),
)
.expect("Fail to set metadata");
// set input tensor
let tensor_data = prompt.as_bytes().to_vec();
context
.set_input(0, wasmedge_wasi_nn::TensorType::U8, &[1], &tensor_data)
.expect("Failed to set prompt as the input tensor");
// execute the inference
context.compute().expect("Failed to complete inference");
// retrieve the output
let mut output_buffer = vec![0u8; *CTX_SIZE.get().unwrap()];
let mut output_size = context
.get_output(0, &mut output_buffer)
.expect("Failed to get output tensor");
output_size = std::cmp::min(*CTX_SIZE.get().unwrap(), output_size);
let output = String::from_utf8_lossy(&output_buffer[..output_size]).to_string();
println!("\n[Answer]:\n\n{}", output);
Ok(())
}
#[derive(Debug, Default, Deserialize, Serialize)]
struct Options {
#[serde(rename = "enable-log")]
log_enable: bool,
#[serde(rename = "ctx-size")]
ctx_size: u64,
#[serde(rename = "n-predict")]
n_predict: u64,
#[serde(rename = "n-gpu-layers")]
n_gpu_layers: u64,
#[serde(rename = "batch-size")]
batch_size: u64,
#[serde(skip_serializing_if = "Option::is_none", rename = "reverse-prompt")]
reverse_prompt: Option<String>,
}