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
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

use candle_core::Device;

use super::*;

fn t(data: &[f32], shape: (usize, usize, usize, usize)) -> Tensor {
  Tensor::from_slice(data, shape, &Device::Cpu).unwrap()
}

fn max_diff(a: &Tensor, b: &Tensor) -> f32 {
  (a - b)
    .unwrap()
    .abs()
    .unwrap()
    .flatten_all()
    .unwrap()
    .max(0)
    .unwrap()
    .to_scalar::<f32>()
    .unwrap()
}

fn values(t: &Tensor) -> Vec<f32> {
  t.flatten_all().unwrap().to_vec1::<f32>().unwrap()
}

/// One query against keys scoring `0` and `ln 3`: softmax is `[0.25, 0.75]`.
#[test]
fn attention_matches_hand_computation() {
  let q = t(&[1.0, 0.0], (1, 1, 1, 2));
  let k = t(&[0.0, 0.0, 3f32.ln(), 0.0], (1, 1, 2, 2));
  let v = t(&[1.0, 0.0, 0.0, 1.0], (1, 1, 2, 2));
  let out = values(&scaled_dot_product(&q, &k, &v, &AttentionMask::none(), 1.0).unwrap());
  assert!(
    (out[0] - 0.25).abs() < 1e-6 && (out[1] - 0.75).abs() < 1e-6,
    "{out:?}"
  );
}

#[test]
fn masked_keys_get_no_weight() {
  let q = t(&[1.0, 0.0], (1, 1, 1, 2));
  let k = t(&[0.0, 0.0, 3f32.ln(), 0.0], (1, 1, 2, 2));
  let v = t(&[1.0, 0.0, 0.0, 1.0], (1, 1, 2, 2));
  let attention_mask = Tensor::new(&[[1u32, 0]], &Device::Cpu).unwrap();
  let mask = AttentionMask::padding(&attention_mask, &[1], 1, 2, DType::F32).unwrap();
  let out = values(&scaled_dot_product(&q, &k, &v, &mask, 1.0).unwrap());
  assert!(
    (out[0] - 1.0).abs() < 1e-6 && out[1].abs() < 1e-6,
    "{out:?}"
  );
}

/// Scores `[0, 2·ln 3]` at scale 0.5 plus `[0, ln 3]` give weights `[0.1, 0.9]`; adding before the scale gives `[0.1612, 0.8388]`.
/// Pins the matmul path only: candle's Metal kernel applies its own order, which the Metal golden runs cover.
#[test]
fn the_additive_term_is_added_after_the_scale() {
  let q = t(&[1.0, 0.0, 0.0, 0.0], (1, 1, 2, 2));
  let k = t(&[0.0, 0.0, 2.0 * 3f32.ln(), 0.0], (1, 1, 2, 2));
  let v = t(&[1.0, 0.0, 0.0, 1.0], (1, 1, 2, 2));
  let bias = t(&[0.0, 3f32.ln(), 0.0, 0.0], (1, 1, 2, 2));
  let attention_mask = Tensor::new(&[[1u32, 1]], &Device::Cpu).unwrap();
  let mask =
    AttentionMask::padding_with_bias(&attention_mask, &[2], &bias, 1, 2, DType::F32).unwrap();
  let out = values(&scaled_dot_product(&q, &k, &v, &mask, 0.5).unwrap());
  assert!(
    (out[0] - 0.1).abs() < 1e-6 && (out[1] - 0.9).abs() < 1e-6,
    "{out:?}"
  );
}

#[test]
fn repeat_kv_tiles_each_kv_head_for_its_group() {
  let k = t(&[1.0, 2.0], (1, 2, 1, 1));
  assert_eq!(values(&repeat_kv(&k, 2).unwrap()), vec![1.0, 1.0, 2.0, 2.0]);
  assert_eq!(repeat_kv(&k, 1).unwrap().dims(), &[1, 2, 1, 1]);
}

#[test]
fn metal_sdpa_applies_only_for_metal_and_listed_head_sizes() {
  assert!(!metal_sdpa_applies(&Device::Cpu, 128, DType::F32));
  assert!(!metal_sdpa_applies(&Device::Cpu, 256, DType::F16));
}

/// EmbeddingGemma's 256-wide heads: the kernel takes them in a half dtype and cannot build an F32 pipeline for them.
#[test]
fn the_kernel_takes_wide_heads_only_in_a_half_dtype() {
  assert!(sdpa_accepts(256, DType::F16) && sdpa_accepts(256, DType::BF16));
  assert!(!sdpa_accepts(256, DType::F32));
  assert!(sdpa_accepts(128, DType::F32) && sdpa_accepts(64, DType::F32));
  assert!(!sdpa_accepts(512, DType::BF16) && !sdpa_accepts(40, DType::F16));
}

#[test]
fn grouped_query_matches_explicitly_repeated_heads() {
  let q = Tensor::randn(0f32, 1.0, (2, 4, 3, 8), &Device::Cpu).unwrap();
  let k = Tensor::randn(0f32, 1.0, (2, 2, 3, 8), &Device::Cpu).unwrap();
  let v = Tensor::randn(0f32, 1.0, (2, 2, 3, 8), &Device::Cpu).unwrap();
  let grouped = scaled_dot_product(&q, &k, &v, &AttentionMask::none(), 0.5).unwrap();
  let full = scaled_dot_product(
    &q,
    &repeat_kv(&k, 2).unwrap(),
    &repeat_kv(&v, 2).unwrap(),
    &AttentionMask::none(),
    0.5,
  )
  .unwrap();
  assert!(max_diff(&grouped, &full) < 1e-6);
  assert_eq!(grouped.dims(), &[2, 4, 3, 8]);
}

#[test]
fn scaled_dot_product_rejects_head_counts_that_do_not_divide() {
  let q = Tensor::randn(0f32, 1.0, (1, 4, 2, 8), &Device::Cpu).unwrap();
  let k = Tensor::randn(0f32, 1.0, (1, 3, 2, 8), &Device::Cpu).unwrap();
  let v = Tensor::randn(0f32, 1.0, (1, 3, 2, 8), &Device::Cpu).unwrap();
  let err = scaled_dot_product(&q, &k, &v, &AttentionMask::none(), 1.0).unwrap_err();
  let message = err.to_string();
  assert!(
    message.contains("4 query heads") && message.contains("3 kv heads"),
    "{message}"
  );
}

/// Values are one-hot per position, so the output shows which keys each position could see.
#[test]
fn causal_attention_sees_only_the_past() {
  let q = Tensor::randn(0f32, 1.0, (1, 1, 3, 4), &Device::Cpu).unwrap();
  let k = Tensor::randn(0f32, 1.0, (1, 1, 3, 4), &Device::Cpu).unwrap();
  let v = t(
    &[1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0],
    (1, 1, 3, 4),
  );
  let mask = AttentionMask::causal(3, 4, DType::F32, &Device::Cpu).unwrap();
  let rows = scaled_dot_product(&q, &k, &v, &mask, 0.5)
    .unwrap()
    .squeeze(0)
    .unwrap()
    .squeeze(0)
    .unwrap()
    .to_vec2::<f32>()
    .unwrap();
  assert!(
    (rows[0][0] - 1.0).abs() < 1e-6 && rows[0][1].abs() < 1e-6 && rows[0][2].abs() < 1e-6,
    "position 0 attends to key 0 only: {:?}",
    rows[0]
  );
  assert!(
    rows[1][2].abs() < 1e-6,
    "position 1 must not see key 2: {:?}",
    rows[1]
  );
}

/// Rows `0..H` are every head's query, `H..2H` the key and `2H..3H` the value; a per-head read would keep the shape.
#[test]
fn the_fused_projection_splits_into_query_key_value_along_its_rows() {
  let fused = Tensor::from_slice(
    &[1f32, 1.0, 1.0, 1.0, 2.0, 2.0, 2.0, 2.0, 3.0, 3.0, 3.0, 3.0],
    (6, 2),
    &Device::Cpu,
  )
  .unwrap();
  let (q, k, v) = split_fused_qkv(&fused, 2).unwrap();
  assert_eq!(
    q.to_vec2::<f32>().unwrap(),
    vec![vec![1.0, 1.0], vec![1.0, 1.0]]
  );
  assert_eq!(
    k.to_vec2::<f32>().unwrap(),
    vec![vec![2.0, 2.0], vec![2.0, 2.0]]
  );
  assert_eq!(
    v.to_vec2::<f32>().unwrap(),
    vec![vec![3.0, 3.0], vec![3.0, 3.0]]
  );
}

mod weights;