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
|
use super::*;
use candle::{Device, Result, Tensor};
use memmap2::MmapOptions;
use safetensors::SafeTensors;
use std::fs::File;
use std::path::PathBuf;
pub struct VarBuilder<'a> {
routing: HashMap<String, usize>,
safetensors: Vec<SafeTensors<'a>>,
device: Device,
}
impl<'a> VarBuilder<'a> {
pub fn new(safetensors: Vec<SafeTensors<'a>>, device: Device) -> Self {
let mut routing = HashMap::new();
for (index, sf) in safetensors.iter().enumerate() {
for k in sf.names() {
routing.insert(k.to_string(), index);
}
}
Self {
safetensors,
device,
routing,
}
}
pub fn get(&self, tensor_name: &str) -> Result<Tensor> {
// Unwrap or 0 just to let the proper error flow.
let index = self.routing.get(tensor_name).unwrap_or(&0);
let view = self.safetensors[*index].tensor(tensor_name).unwrap();
candle::safetensors::convert(view, &self.device)?.to_dtype(DTYPE)
}
}
impl Linear {
fn load(prefix: &str, vb: &VarBuilder) -> Result<Self> {
let weight = vb.get(&format!("{prefix}.weight"))?;
Ok(Self::new(weight))
}
fn load_multi(prefixes: &[&str], vb: &VarBuilder) -> Result<Self> {
let weights: Vec<_> = prefixes
.iter()
.map(|p| vb.get(&format!("{p}.weight")).unwrap())
.collect();
let weight = Tensor::cat(&weights, 0)?;
Ok(Self::new(weight))
}
}
impl RmsNorm {
fn load(prefix: &str, vb: &VarBuilder) -> Result<Self> {
let scale = vb.get(&format!("{prefix}.weight"))?;
Ok(Self::new(scale))
}
}
impl CausalSelfAttention {
fn load(prefix: &str, vb: &VarBuilder, cache: &Cache, config: &Config) -> Result<Self> {
let c_attn = Linear::load_multi(
&[
&format!("{prefix}.q_proj"),
&format!("{prefix}.k_proj"),
&format!("{prefix}.v_proj"),
],
vb,
)?;
let o_proj = Linear::load(&format!("{prefix}.o_proj"), vb)?;
Ok(Self::new(c_attn, o_proj, config.n_head, cache))
}
}
impl Mlp {
fn load(prefix: &str, vb: &VarBuilder) -> Result<Self> {
let c_fc1 = Linear::load(&format!("{prefix}.gate_proj"), vb)?;
let c_fc2 = Linear::load(&format!("{prefix}.up_proj"), vb)?;
let c_proj = Linear::load(&format!("{prefix}.down_proj"), vb)?;
Ok(Self::new(c_fc1, c_fc2, c_proj))
}
}
impl Block {
fn load(prefix: &str, vb: &VarBuilder, cache: &Cache, config: &Config) -> Result<Self> {
let attn = CausalSelfAttention::load(&format!("{prefix}.self_attn"), vb, cache, config)?;
let mlp = Mlp::load(&format!("{prefix}.mlp"), vb)?;
let input_layernorm = RmsNorm::load(&format!("{prefix}.input_layernorm"), vb)?;
let post_attention_layernorm =
RmsNorm::load(&format!("{prefix}.post_attention_layernorm"), vb)?;
Ok(Self::new(
input_layernorm,
attn,
post_attention_layernorm,
mlp,
))
}
}
impl Llama {
pub fn load(
device: &Device,
filenames: &[PathBuf],
cache: &Cache,
config: &Config,
) -> Result<Self> {
let handles: Vec<_> = filenames
.iter()
.map(|f| {
let file = File::open(f).unwrap();
unsafe { MmapOptions::new().map(&file).unwrap() }
})
.collect();
let tensors: Vec<_> = handles
.iter()
.map(|h| {
let tensors = SafeTensors::deserialize(h).unwrap();
tensors
})
.collect();
let vb = VarBuilder::new(tensors, device.clone());
let embedding = vb.get("model.embed_tokens.weight")?;
let wte = Embedding::new(embedding);
let lm_head = Linear::load("lm_head", &vb)?;
let norm = RmsNorm::load("model.norm", &vb)?;
let blocks: Vec<_> = (0..config.n_layer)
.map(|i| Block::load(&format!("model.layers.{i}"), &vb, cache, config).unwrap())
.collect();
Ok(Self::new(wte, blocks, norm, lm_head))
}
}
|
