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
|
// An implementation of LLaMA https://github.com/facebookresearch/llama
//
// This is based on nanoGPT in a similar way to:
// https://github.com/Lightning-AI/lit-llama/blob/main/lit_llama/model.py
//
// The tokenizer config can be retrieved from:
// https://huggingface.co/hf-internal-testing/llama-tokenizer/raw/main/tokenizer.json
#[cfg(feature = "accelerate")]
extern crate accelerate_src;
#[cfg(feature = "mkl")]
extern crate intel_mkl_src;
use anyhow::{bail, Error as E, Result};
use clap::{Parser, ValueEnum};
use candle::{DType, Tensor};
use candle_nn::VarBuilder;
use candle_transformers::generation::LogitsProcessor;
use hf_hub::{api::sync::Api, Repo, RepoType};
use std::io::Write;
use candle_transformers::models::llama as model;
use model::{Llama, LlamaConfig};
const EOS_TOKEN: &str = "</s>";
const DEFAULT_PROMPT: &str = "My favorite theorem is ";
#[derive(Clone, Debug, Copy, PartialEq, Eq, ValueEnum)]
enum Which {
V1,
V2,
#[value(name = "solar-10.7b")]
Solar10_7B,
}
#[derive(Parser, Debug)]
#[command(author, version, about, long_about = None)]
struct Args {
/// Run on CPU rather than on GPU.
#[arg(long)]
cpu: bool,
/// The temperature used to generate samples.
#[arg(long)]
temperature: Option<f64>,
/// Nucleus sampling probability cutoff.
#[arg(long)]
top_p: 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,
/// Disable the key-value cache.
#[arg(long)]
no_kv_cache: bool,
/// The initial prompt.
#[arg(long)]
prompt: Option<String>,
/// Use different dtype than f16
#[arg(long)]
dtype: Option<String>,
/// Enable tracing (generates a trace-timestamp.json file).
#[arg(long)]
tracing: bool,
#[arg(long)]
model_id: Option<String>,
#[arg(long)]
revision: Option<String>,
/// The model size to use.
#[arg(long, default_value = "v2")]
which: Which,
#[arg(long)]
use_flash_attn: bool,
/// Penalty to be applied for repeating tokens, 1. means no penalty.
#[arg(long, default_value_t = 1.0)]
repeat_penalty: f32,
/// The context size to consider for the repeat penalty.
#[arg(long, default_value_t = 64)]
repeat_last_n: usize,
}
fn main() -> Result<()> {
use tokenizers::Tokenizer;
use tracing_chrome::ChromeLayerBuilder;
use tracing_subscriber::prelude::*;
let args = Args::parse();
let _guard = if args.tracing {
let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
tracing_subscriber::registry().with(chrome_layer).init();
Some(guard)
} else {
None
};
let device = candle_examples::device(args.cpu)?;
let dtype = match args.dtype.as_deref() {
Some("f16") => DType::F16,
Some("bf16") => DType::BF16,
Some("f32") => DType::F32,
Some(dtype) => bail!("Unsupported dtype {dtype}"),
None => DType::F16,
};
let (llama, tokenizer_filename, cache) = {
let api = Api::new()?;
let model_id = args.model_id.unwrap_or_else(|| match args.which {
Which::V1 => "Narsil/amall-7b".to_string(),
Which::V2 => "meta-llama/Llama-2-7b-hf".to_string(),
Which::Solar10_7B => "upstage/SOLAR-10.7B-v1.0".to_string(),
});
println!("loading the model weights from {model_id}");
let revision = args.revision.unwrap_or("main".to_string());
let api = api.repo(Repo::with_revision(model_id, RepoType::Model, revision));
let tokenizer_filename = api.get("tokenizer.json")?;
let config_filename = api.get("config.json")?;
let config: LlamaConfig = serde_json::from_slice(&std::fs::read(config_filename)?)?;
let config = config.into_config(args.use_flash_attn);
let filenames =
candle_examples::hub_load_safetensors(&api, "model.safetensors.index.json")?;
println!("building the model");
let cache = model::Cache::new(!args.no_kv_cache, dtype, &config, &device)?;
let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
(Llama::load(vb, &cache, &config)?, tokenizer_filename, cache)
};
let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
let eos_token_id = tokenizer.token_to_id(EOS_TOKEN);
let prompt = args.prompt.as_ref().map_or(DEFAULT_PROMPT, |p| p.as_str());
let mut tokens = tokenizer
.encode(prompt, true)
.map_err(E::msg)?
.get_ids()
.to_vec();
println!("starting the inference loop");
print!("{prompt}");
let mut logits_processor = LogitsProcessor::new(args.seed, args.temperature, args.top_p);
let start_gen = std::time::Instant::now();
let mut index_pos = 0;
let mut token_generated = 0;
for index in 0..args.sample_len {
let context_size = if cache.use_kv_cache && index > 0 {
1
} else {
tokens.len()
};
let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
let input = Tensor::new(ctxt, &device)?.unsqueeze(0)?;
let logits = llama.forward(&input, index_pos)?;
let logits = logits.squeeze(0)?;
let logits = if args.repeat_penalty == 1. {
logits
} else {
let start_at = tokens.len().saturating_sub(args.repeat_last_n);
candle_transformers::utils::apply_repeat_penalty(
&logits,
args.repeat_penalty,
&tokens[start_at..],
)?
};
index_pos += ctxt.len();
let next_token = logits_processor.sample(&logits)?;
token_generated += 1;
tokens.push(next_token);
// Extracting the last token as a string is complicated, here we just apply some simple
// heuristics as it seems to work well enough for this example. See the following for more
// details:
// https://github.com/huggingface/tokenizers/issues/1141#issuecomment-1562644141
if let Some(text) = tokenizer.id_to_token(next_token) {
let text = text.replace('▁', " ").replace("<0x0A>", "\n");
print!("{text}");
std::io::stdout().flush()?;
}
if Some(next_token) == eos_token_id {
break;
}
}
let dt = start_gen.elapsed();
println!(
"\n\n{} tokens generated ({} token/s)\n",
token_generated,
token_generated as f64 / dt.as_secs_f64(),
);
Ok(())
}
|
