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
|
#![allow(dead_code)]
// TODO: Add an offline mode.
#[cfg(feature = "mkl")]
extern crate intel_mkl_src;
use anyhow::{Error as E, Result};
use candle::{DType, Device, Tensor, D};
use candle_hub::{api::sync::Api, Repo, RepoType};
use candle_nn::VarBuilder;
use clap::Parser;
use rand::{distributions::Distribution, SeedableRng};
use tokenizers::Tokenizer;
mod model;
use model::{Config, Falcon};
#[cfg(feature = "mkl")]
const DTYPE: DType = DType::F32;
#[cfg(not(feature = "mkl"))]
const DTYPE: DType = DType::BF16;
struct TextGeneration {
model: Falcon,
rng: rand::rngs::StdRng,
device: Device,
temperature: Option<f64>,
tokenizer: Tokenizer,
}
impl TextGeneration {
fn new(
model: Falcon,
tokenizer: Tokenizer,
seed: u64,
temperature: Option<f64>,
device: &Device,
) -> Self {
Self {
model,
tokenizer,
rng: rand::rngs::StdRng::seed_from_u64(seed),
temperature,
device: device.clone(),
}
}
fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
println!("starting the inference loop");
let mut tokens = self
.tokenizer
.encode(prompt, true)
.map_err(E::msg)?
.get_ids()
.to_vec();
let mut new_tokens = vec![];
let start_gen = std::time::Instant::now();
for index in 0..sample_len {
let start_gen = std::time::Instant::now();
let context_size = if self.model.config().use_cache && index > 0 {
1
} else {
tokens.len()
};
let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
let logits = self.model.forward(&input)?;
let logits = logits.squeeze(0)?.to_dtype(DType::F32)?;
let next_token = if let Some(temperature) = self.temperature {
let prs = (&logits / temperature)?.softmax(D::Minus1)?;
let logits_v: Vec<f32> = prs.to_vec1()?;
let distr = rand::distributions::WeightedIndex::new(&logits_v)?;
distr.sample(&mut self.rng) as u32
} else {
let logits_v: Vec<f32> = logits.to_vec1()?;
logits_v
.iter()
.enumerate()
.max_by(|(_, u), (_, v)| u.total_cmp(v))
.map(|(i, _)| i as u32)
.unwrap()
};
tokens.push(next_token);
new_tokens.push(next_token);
println!("> {:?}", start_gen.elapsed());
println!(
"{} token: {} '{}'",
index + 1,
next_token,
self.tokenizer
.decode(vec![next_token], true)
.map_err(E::msg)?
);
}
let dt = start_gen.elapsed();
println!(
"{sample_len} tokens generated ({} token/s)\n----\n{}\n----",
sample_len as f64 / dt.as_secs_f64(),
self.tokenizer.decode(new_tokens, true).map_err(E::msg)?
);
Ok(())
}
}
#[derive(Parser, Debug)]
#[command(author, version, about, long_about = None)]
struct Args {
/// Run on CPU rather than on GPU.
#[arg(long)]
cpu: bool,
#[arg(long)]
prompt: String,
/// The temperature used to generate samples.
#[arg(long)]
temperature: Option<f64>,
/// The seed to use when generating random samples.
#[arg(long, default_value_t = 299792458)]
seed: u64,
/// The length of the sample to generate (in tokens).
#[arg(long, default_value_t = 100)]
sample_len: usize,
#[arg(long, default_value = "tiiuae/falcon-7b")]
model_id: String,
#[arg(long, default_value = "refs/pr/43")]
revision: String,
}
fn main() -> Result<()> {
let args = Args::parse();
let device = if args.cpu {
Device::Cpu
} else {
Device::new_cuda(0)?
};
let start = std::time::Instant::now();
let api = Api::new()?;
let repo = Repo::with_revision(args.model_id, RepoType::Model, args.revision);
let tokenizer_filename = api.get(&repo, "tokenizer.json")?;
let mut filenames = vec![];
for rfilename in [
"model-00001-of-00002.safetensors",
"model-00002-of-00002.safetensors",
] {
let filename = api.get(&repo, rfilename)?;
filenames.push(filename);
}
println!("retrieved the files in {:?}", start.elapsed());
let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
let start = std::time::Instant::now();
let weights = filenames
.iter()
.map(|f| Ok(unsafe { candle::safetensors::MmapedFile::new(f)? }))
.collect::<Result<Vec<_>>>()?;
let weights = weights
.iter()
.map(|f| Ok(f.deserialize()?))
.collect::<Result<Vec<_>>>()?;
let vb = VarBuilder::from_safetensors(weights, DTYPE, &device);
let config = Config::falcon7b();
config.validate()?;
let model = Falcon::load(vb, config)?;
println!("loaded the model in {:?}", start.elapsed());
let mut pipeline = TextGeneration::new(model, tokenizer, args.seed, args.temperature, &device);
pipeline.run(&args.prompt, args.sample_len)?;
Ok(())
}
|
