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
|
#[cfg(feature = "mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]
extern crate accelerate_src;
use candle::{test_device, test_utils::to_vec3_round, Device, Result, Tensor};
fn softmax(device: &Device) -> Result<()> {
let data = &[[[3f32, 1., 4.], [1., 5., 9.]], [[2., 1., 7.], [8., 2., 8.]]];
let tensor = Tensor::new(data, device)?;
let t0 = candle_nn::ops::softmax(&tensor.log()?, 0)?;
let t1 = candle_nn::ops::softmax(&tensor.log()?, 1)?;
let t2 = candle_nn::ops::softmax(&tensor.log()?, 2)?;
assert_eq!(
to_vec3_round(&t0, 4)?,
&[
// 3/5, 1/2, 4/11
[[0.6, 0.5, 0.3636], [0.1111, 0.7143, 0.5294]],
// 2/5, 1/2, 7/11
[[0.4, 0.5, 0.6364], [0.8889, 0.2857, 0.4706]]
]
);
assert_eq!(
to_vec3_round(&t1, 4)?,
&[
// 3/4, 1/6, 4/13
[[0.75, 0.1667, 0.3077], [0.25, 0.8333, 0.6923]],
// 2/10, 1/3, 7/15
[[0.2, 0.3333, 0.4667], [0.8, 0.6667, 0.5333]]
]
);
assert_eq!(
to_vec3_round(&t2, 4)?,
&[
// (3, 1, 4) / 8, (1, 5, 9) / 15
[[0.375, 0.125, 0.5], [0.0667, 0.3333, 0.6]],
// (2, 1, 7) / 10, (8, 2, 8) / 18
[[0.2, 0.1, 0.7], [0.4444, 0.1111, 0.4444]]
]
);
let t2 = candle_nn::ops::softmax_last_dim(&tensor.log()?)?;
assert_eq!(
to_vec3_round(&t2, 4)?,
&[
// (3, 1, 4) / 8, (1, 5, 9) / 15
[[0.375, 0.125, 0.5], [0.0667, 0.3333, 0.6]],
// (2, 1, 7) / 10, (8, 2, 8) / 18
[[0.2, 0.1, 0.7], [0.4444, 0.1111, 0.4444]]
]
);
Ok(())
}
fn rms_norm(device: &Device) -> Result<()> {
let data = &[[[3f32, 1., 4.], [1., 5., 9.]], [[2., 1., 7.], [8., 2., 8.]]];
let tensor = Tensor::new(data, device)?;
let alpha = Tensor::new(&[1f32, 2f32, 3f32], device)?;
let t = candle_nn::ops::rms_norm(&tensor, &alpha, 1e-5)?;
assert_eq!(
to_vec3_round(&t, 4)?,
&[
[[1.019, 0.6794, 4.0762], [0.1674, 1.6744, 4.521]],
[[0.4714, 0.4714, 4.9497], [1.206, 0.603, 3.6181]]
]
);
let t2 = candle_nn::ops::rms_norm_slow(&tensor, &alpha, 1e-5)?;
assert_eq!(
to_vec3_round(&t2, 4)?,
&[
[[1.019, 0.6794, 4.0762], [0.1674, 1.6744, 4.521]],
[[0.4714, 0.4714, 4.9497], [1.206, 0.603, 3.6181]]
]
);
let diff = (t - t2)?.abs()?.sum_all()?.to_vec0::<f32>()?;
assert!(diff < 1e-5);
Ok(())
}
#[test]
fn softmax_numerical_stability() -> Result<()> {
let dev = &Device::Cpu;
let xs = Tensor::new(&[1234f32, 0.], dev)?;
let softmax = candle_nn::ops::softmax(&xs, 0)?;
assert_eq!(softmax.to_vec1::<f32>()?, &[1f32, 0.]);
Ok(())
}
fn ropei(device: &Device) -> Result<()> {
use rand::{rngs::StdRng, Rng, SeedableRng};
let (b_size, num_head, seq_len, head_dim) = (2, 5, 10, 16);
let el_count = b_size * num_head * seq_len * head_dim;
let mut rng = StdRng::seed_from_u64(299792458);
let src: Vec<f32> = (0..el_count).map(|_| rng.gen::<f32>()).collect();
let cos: Vec<f32> = (0..seq_len * head_dim / 2)
.map(|_| rng.gen::<f32>())
.collect();
let sin: Vec<f32> = (0..seq_len * head_dim / 2)
.map(|_| rng.gen::<f32>())
.collect();
let src = Tensor::from_vec(src, (b_size, num_head, seq_len, head_dim), device)?;
let cos = Tensor::from_vec(cos, (seq_len, head_dim / 2), device)?;
let sin = Tensor::from_vec(sin, (seq_len, head_dim / 2), device)?;
let rope1 = candle_nn::rotary_emb::rope_i(&src, &cos, &sin)?;
let rope2 = candle_nn::rotary_emb::rope_i_slow(&src, &cos, &sin)?;
let sum_diff = (rope1 - rope2)?.abs()?.sum_all()?.to_vec0::<f32>()?;
if device.is_cpu() {
assert_eq!(sum_diff, 0.);
} else {
assert!(sum_diff < 1e-4);
}
Ok(())
}
fn rope(device: &Device) -> Result<()> {
use rand::{rngs::StdRng, Rng, SeedableRng};
let (b_size, num_head, seq_len, head_dim) = (2, 5, 10, 16);
let el_count = b_size * num_head * seq_len * head_dim;
let mut rng = StdRng::seed_from_u64(299792458);
let src: Vec<f32> = (0..el_count).map(|_| rng.gen::<f32>()).collect();
let cos: Vec<f32> = (0..seq_len * head_dim / 2)
.map(|_| rng.gen::<f32>())
.collect();
let sin: Vec<f32> = (0..seq_len * head_dim / 2)
.map(|_| rng.gen::<f32>())
.collect();
let src = Tensor::from_vec(src, (b_size, num_head, seq_len, head_dim), device)?;
let cos = Tensor::from_vec(cos, (seq_len, head_dim / 2), device)?;
let sin = Tensor::from_vec(sin, (seq_len, head_dim / 2), device)?;
let rope1 = candle_nn::rotary_emb::rope(&src, &cos, &sin)?;
let rope2 = candle_nn::rotary_emb::rope_slow(&src, &cos, &sin)?;
let sum_diff = (rope1 - rope2)?.abs()?.sum_all()?.to_vec0::<f32>()?;
if device.is_cpu() {
assert_eq!(sum_diff, 0.);
} else {
assert!(sum_diff < 1e-4);
}
Ok(())
}
fn rope_thd(device: &Device) -> Result<()> {
use rand::{rngs::StdRng, Rng, SeedableRng};
let (b_size, num_head, seq_len, head_dim) = (2, 5, 10, 16);
let el_count = b_size * num_head * seq_len * head_dim;
let mut rng = StdRng::seed_from_u64(299792458);
let src: Vec<f32> = (0..el_count).map(|_| rng.gen::<f32>()).collect();
let cos: Vec<f32> = (0..seq_len * head_dim / 2)
.map(|_| rng.gen::<f32>())
.collect();
let sin: Vec<f32> = (0..seq_len * head_dim / 2)
.map(|_| rng.gen::<f32>())
.collect();
let src = Tensor::from_vec(src, (b_size, num_head, seq_len, head_dim), device)?;
let cos = Tensor::from_vec(cos, (seq_len, head_dim / 2), device)?;
let sin = Tensor::from_vec(sin, (seq_len, head_dim / 2), device)?;
let rope1 = {
let src = src.transpose(1, 2)?.contiguous()?;
candle_nn::rotary_emb::rope_thd(&src, &cos, &sin)?.transpose(1, 2)?
};
let rope2 = candle_nn::rotary_emb::rope_slow(&src, &cos, &sin)?;
let sum_diff = (rope1 - rope2)?.abs()?.sum_all()?.to_vec0::<f32>()?;
if device.is_cpu() {
assert_eq!(sum_diff, 0.);
} else {
assert!(sum_diff < 1e-4);
}
Ok(())
}
test_device!(ropei, ropei_cpu, ropei_gpu, ropei_metal);
test_device!(rope, rope_cpu, rope_gpu, rope_metal);
test_device!(rope_thd, rope_thd_cpu, rope_thd_gpu, rope_thd_metal);
test_device!(softmax, softmax_cpu, softmax_gpu, softmax_metal);
test_device!(rms_norm, rms_norm_cpu, rms_norm_gpu, rms_norm_metal);
|
