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
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
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224

use candle_core::Device;
use candle_core::Tensor;

use super::*;

#[test]
fn padding_mask_zeroes_real_tokens_and_floors_padding() {
  let attention_mask = Tensor::new(&[[1u32, 1, 0], [1, 1, 1]], &Device::Cpu).unwrap();
  let mask = AttentionMask::padding(&attention_mask, &[2, 3], 4, 64, DType::F16).unwrap();
  let additive = mask.additive().expect("padded rows produce a mask");
  assert_eq!(additive.dims(), &[2, 1, 1, 3]);
  assert_eq!(additive.dtype(), DType::F16);
  let values = additive
    .to_dtype(DType::F32)
    .unwrap()
    .flatten_all()
    .unwrap()
    .to_vec1::<f32>()
    .unwrap();
  assert_eq!(values, vec![0.0, 0.0, -65504.0, 0.0, 0.0, 0.0]);
  assert!(!mask.is_causal());
}

#[test]
fn padding_mask_is_none_without_padding() {
  let attention_mask = Tensor::new(&[[1u32, 1], [1, 1]], &Device::Cpu).unwrap();
  let mask = AttentionMask::padding(&attention_mask, &[2, 2], 2, 64, DType::F32).unwrap();
  assert!(mask.additive().is_none() && !mask.is_causal());
}

#[test]
fn causal_mask_is_upper_triangular_on_cpu() {
  let mask = AttentionMask::causal(3, 128, DType::F32, &Device::Cpu).unwrap();
  let values = mask
    .additive()
    .unwrap()
    .squeeze(0)
    .unwrap()
    .squeeze(0)
    .unwrap()
    .to_vec2::<f32>()
    .unwrap();
  let floor = f32::MIN;
  assert_eq!(
    values,
    vec![
      vec![0.0, floor, floor],
      vec![0.0, 0.0, floor],
      vec![0.0, 0.0, 0.0]
    ]
  );
  assert!(
    !mask.is_causal(),
    "the flag is only raised for the Metal kernel"
  );
}

/// Too few lengths would leave rows unmasked; a `debug_assert` says so in a test build and nothing in release.
#[test]
fn a_lengths_slice_that_does_not_match_the_batch_is_refused() {
  let attention_mask = Tensor::new(&[[1u32, 1, 0], [1, 1, 1]], &Device::Cpu).unwrap();
  let bias = Tensor::zeros((1, 4, 3, 3), DType::F32, &Device::Cpu).unwrap();
  for lengths in [&[3usize][..], &[3, 3, 3][..], &[][..]] {
    for built in [
      AttentionMask::padding(&attention_mask, lengths, 4, 64, DType::F32),
      AttentionMask::padding_with_bias(&attention_mask, lengths, &bias, 4, 64, DType::F32),
      AttentionMask::sliding(&attention_mask, lengths, 2, 4, 64, DType::F32),
    ] {
      let err = built
        .err()
        .unwrap_or_else(|| panic!("{} lengths for 2 rows must be refused", lengths.len()));
      assert!(
        err.to_string().contains("for a batch of 2 rows"),
        "{lengths:?}: {err}"
      );
    }
  }
  assert!(AttentionMask::padding(&attention_mask, &[2, 3], 4, 64, DType::F32).is_ok());
}

#[test]
fn causal_mask_materializes_for_a_head_size_the_kernel_rejects() {
  let mask = AttentionMask::causal(3, 40, DType::F32, &Device::Cpu).unwrap();
  assert!(mask.additive().is_some());
}

/// Window 3 over six positions; row 1 has two padded keys.
#[test]
fn sliding_mask_combines_the_window_with_padding() {
  let attention_mask =
    Tensor::new(&[[1u32, 1, 1, 1, 1, 1], [1, 1, 1, 1, 0, 0]], &Device::Cpu).unwrap();
  let mask = AttentionMask::sliding(&attention_mask, &[6, 4], 3, 2, 64, DType::F32).unwrap();
  let additive = mask
    .additive()
    .expect("a window narrower than the sequence produces a mask");
  assert_eq!(additive.dims(), &[2, 1, 6, 6]);
  let rows = additive.squeeze(1).unwrap().to_vec3::<f32>().unwrap();
  let floor = f32::MIN;
  let (row0, row1) = (&rows[0], &rows[1]);
  for (i, (cols0, cols1)) in row0.iter().zip(row1.iter()).enumerate() {
    for (j, (&value0, &value1)) in cols0.iter().zip(cols1.iter()).enumerate() {
      let in_window = i.abs_diff(j) < 3;
      assert_eq!(value0 == 0.0, in_window, "row 0 ({i}, {j})");
      let visible1 = (in_window && j < 4) || i == j;
      assert_eq!(value1 == 0.0, visible1, "row 1 ({i}, {j})");
      if !visible1 {
        assert_eq!(value1, floor);
      }
    }
  }
  assert!(!mask.is_causal());
}

#[test]
fn sliding_mask_is_none_when_the_window_covers_everything() {
  let attention_mask = Tensor::new(&[[1u32, 1, 1]], &Device::Cpu).unwrap();
  let mask = AttentionMask::sliding(&attention_mask, &[3], 3, 1, 64, DType::F32).unwrap();
  assert!(mask.additive().is_none());
  let padded = Tensor::new(&[[1u32, 1, 0]], &Device::Cpu).unwrap();
  let mask = AttentionMask::sliding(&padded, &[2], 3, 1, 64, DType::F16).unwrap();
  let additive = mask.additive().expect("padding still needs a mask");
  assert_eq!(additive.dtype(), DType::F16);
  let rows = additive
    .to_dtype(DType::F32)
    .unwrap()
    .squeeze(1)
    .unwrap()
    .to_vec3::<f32>()
    .unwrap();
  assert_eq!(rows[0][0][2], -65504.0);
  assert_eq!(rows[0][0][1], 0.0);
}

/// Without its own key the row would be entirely floored, and candle's Metal kernel divides by a zero softmax sum there.
#[test]
fn sliding_mask_leaves_no_row_entirely_floored() {
  let attention_mask = Tensor::new(&[[1u32, 1, 0, 0, 0, 0, 0, 0]], &Device::Cpu).unwrap();
  let mask = AttentionMask::sliding(&attention_mask, &[2], 2, 1, 64, DType::F32).unwrap();
  let rows = mask
    .additive()
    .expect("a padded row produces a mask")
    .squeeze(1)
    .unwrap()
    .to_vec3::<f32>()
    .unwrap();
  for (i, row) in rows[0].iter().enumerate() {
    let open: Vec<usize> = row
      .iter()
      .enumerate()
      .filter(|(_, value)| **value == 0.0)
      .map(|(j, _)| j)
      .collect();
    assert!(!open.is_empty(), "row {i} has no visible key");
    assert!(
      open.contains(&i),
      "row {i} cannot see its own key: {open:?}"
    );
  }
  assert_eq!(
    rows[0][7],
    vec![
      f32::MIN,
      f32::MIN,
      f32::MIN,
      f32::MIN,
      f32::MIN,
      f32::MIN,
      f32::MIN,
      0.0
    ]
  );
}

/// candle's fused Metal kernel assumes only the last mask dimension is contiguous, so each constructor hands it a broadcast
/// view; every dtype is pinned because a cast after the broadcast would re-materialize the copy.
#[cfg(feature = "metal")]
#[test]
fn every_mask_reaches_the_metal_kernel_as_a_broadcast_view() {
  let device = Device::new_metal(0).expect("a metal device");
  let padded = Tensor::new(&[[1u32, 1, 0], [1, 1, 1]], &device).unwrap();
  let full = Tensor::new(&[[1u32, 1, 1], [1, 1, 1]], &device).unwrap();
  for dtype in [DType::F32, DType::F16, DType::BF16] {
    let bias = Tensor::zeros((1, 4, 3, 3), DType::F32, &device).unwrap();
    for (name, built, strides) in [
      (
        "padding",
        AttentionMask::padding(&padded, &[2, 3], 4, 64, dtype),
        [3usize, 0, 0, 1],
      ),
      (
        "padding_with_bias, unpadded",
        AttentionMask::padding_with_bias(&full, &[3, 3], &bias, 4, 64, dtype),
        [0, 9, 3, 1],
      ),
      (
        "sliding",
        AttentionMask::sliding(&padded, &[2, 3], 2, 4, 64, dtype),
        [9, 0, 3, 1],
      ),
    ] {
      let additive = built
        .unwrap()
        .additive()
        .unwrap_or_else(|| panic!("{name} in {dtype:?} produces an additive term"))
        .clone();
      assert_eq!(additive.dims(), &[2, 4, 3, 3], "{name} in {dtype:?}");
      assert_eq!(additive.dtype(), dtype, "{name} in {dtype:?}");
      assert_eq!(additive.stride(), &strides, "{name} in {dtype:?}");
      assert!(
        !additive.is_contiguous(),
        "{name} in {dtype:?} was materialized"
      );
    }
  }
}

#[test]
fn mask_floor_per_dtype() {
  assert_eq!(mask_min(DType::F16), -65504.0);
  assert_eq!(mask_min(DType::F32), f32::MIN as f64);
  assert!(mask_min(DType::BF16) < -3.0e38);
}

mod bias;