candding 0.1.0

Pure candle embeddings for Rust with no ONNX: dense, sparse, multi-vector and rerank models today, image models on the roadmap
Documentation 
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

use candle_core::DType;
use candle_core::Device;
use candle_nn::Linear;
use candle_nn::Module;
use candle_nn::VarBuilder;
use candle_nn::VarMap;

use super::*;
use crate::layers::attention::scaled_dot_product;
use crate::layers::mask::AttentionMask;
use crate::layers::rope::RotaryEmbedding;
use crate::models::xlm_roberta_flash::attention::XlmRobertaFlashAttention;

/// Compares `forward` with the same `Wqkv` narrowed by hand and run through `layers`' public primitives.
/// Asymmetric blocks and tokens, since a symmetric block would hide a query-key swap as a transpose.
#[test]
fn the_fused_qkv_projection_splits_in_query_key_value_row_order() {
  let hidden = 4;
  let heads = 2;
  let head_dim = 2;
  let config = XlmRobertaFlashConfig {
    hidden_size: hidden,
    num_attention_heads: heads,
    ..tiny_config()
  };

  let mut varmap = VarMap::new();
  let vb = VarBuilder::from_varmap(&varmap, DType::F32, &Device::Cpu);
  let attention = XlmRobertaFlashAttention::load(&config, vb.pp("mixer")).unwrap();

  #[rustfmt::skip]
  let q_block = [
    0.7f32, -0.3, 0.2, 0.5,
    -0.1, 0.9, -0.4, 0.2,
    0.3, 0.1, 0.6, -0.2,
    -0.5, 0.4, 0.1, 0.8,
  ];
  #[rustfmt::skip]
  let k_block = [
    0.2f32, 0.6, -0.3, 0.1,
    0.9, -0.2, 0.4, -0.1,
    -0.4, 0.3, 0.7, 0.2,
    0.1, -0.5, 0.2, 0.6,
  ];
  #[rustfmt::skip]
  let v_block = [
    1.1f32, 0.2, -0.3, 0.4,
    -0.2, 1.3, 0.1, -0.4,
    0.3, -0.1, 1.2, 0.2,
    -0.4, 0.3, -0.2, 1.4,
  ];
  let mut rows = Vec::with_capacity(3 * hidden * hidden);
  rows.extend_from_slice(&q_block);
  rows.extend_from_slice(&k_block);
  rows.extend_from_slice(&v_block);
  let weight = Tensor::from_vec(rows, (3 * hidden, hidden), &Device::Cpu).unwrap();
  let bias = Tensor::from_vec(
    vec![
      0.1f32, -0.2, 0.3, 0.05, 0.4, 0.1, -0.3, 0.2, -0.15, 0.25, 0.05, -0.1,
    ],
    3 * hidden,
    &Device::Cpu,
  )
  .unwrap();
  varmap.set_one("mixer.Wqkv.weight", weight.clone()).unwrap();
  varmap.set_one("mixer.Wqkv.bias", bias.clone()).unwrap();
  varmap
    .set_one(
      "mixer.out_proj.weight",
      Tensor::eye(hidden, DType::F32, &Device::Cpu).unwrap(),
    )
    .unwrap();
  varmap
    .set_one(
      "mixer.out_proj.bias",
      Tensor::zeros(hidden, DType::F32, &Device::Cpu).unwrap(),
    )
    .unwrap();

  let hidden_input = Tensor::from_vec(
    vec![
      1.0f32, 0.5, -0.5, 0.2, -0.3, 0.8, 0.1, -0.6, 0.4, -0.2, 0.9, 0.3,
    ],
    (1, 3, hidden),
    &Device::Cpu,
  )
  .unwrap();
  let mask = AttentionMask::none();
  let rope = RotaryEmbedding::new(head_dim, 8, 10_000.0, DType::F32, &Device::Cpu).unwrap();

  let actual = attention
    .forward(&hidden_input, &mask, Some(&rope))
    .unwrap();

  let slice = |t: &Tensor, third: usize| t.narrow(0, third * hidden, hidden).unwrap();
  let q_lin = Linear::new(slice(&weight, 0), Some(slice(&bias, 0)));
  let k_lin = Linear::new(slice(&weight, 1), Some(slice(&bias, 1)));
  let v_lin = Linear::new(slice(&weight, 2), Some(slice(&bias, 2)));
  let split = |xs: Tensor| {
    xs.reshape((1, 3, heads, head_dim))
      .unwrap()
      .transpose(1, 2)
      .unwrap()
      .contiguous()
      .unwrap()
  };
  let q = split(q_lin.forward(&hidden_input).unwrap());
  let k = split(k_lin.forward(&hidden_input).unwrap());
  let v = split(v_lin.forward(&hidden_input).unwrap());
  let (q, k) = rope.apply(&q, &k).unwrap();
  let scale = 1.0 / (head_dim as f64).sqrt();
  let expected = scaled_dot_product(&q, &k, &v, &mask, scale)
    .unwrap()
    .transpose(1, 2)
    .unwrap()
    .reshape((1, 3, hidden))
    .unwrap();

  let diff = (actual - expected)
    .unwrap()
    .abs()
    .unwrap()
    .flatten_all()
    .unwrap()
    .max(0)
    .unwrap()
    .to_scalar::<f32>()
    .unwrap();
  assert!(
    diff < 1e-6,
    "max abs diff between the module and the independently-narrowed shadow computation: {diff}"
  );
}