numr 0.5.1

High-performance numerical computing with multi-backend GPU acceleration (CPU/CUDA/WebGPU)
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

//! Normalization operations for WebGPU runtime

use crate::error::Result;
use crate::ops::NormalizationOps;
use crate::runtime::wgpu::WgpuClient;
use crate::runtime::wgpu::WgpuRuntime;
use crate::runtime::wgpu::ops::native::{
    native_fused_add_layer_norm, native_fused_add_layer_norm_bwd, native_fused_add_rms_norm,
    native_fused_add_rms_norm_bwd, native_group_norm, native_layer_norm, native_rms_norm,
};
use crate::tensor::Tensor;

impl NormalizationOps<WgpuRuntime> for WgpuClient {
    fn rms_norm(
        &self,
        a: &Tensor<WgpuRuntime>,
        weight: &Tensor<WgpuRuntime>,
        eps: f32,
    ) -> Result<Tensor<WgpuRuntime>> {
        native_rms_norm(self, a, weight, eps)
    }

    fn layer_norm(
        &self,
        a: &Tensor<WgpuRuntime>,
        weight: &Tensor<WgpuRuntime>,
        bias: &Tensor<WgpuRuntime>,
        eps: f32,
    ) -> Result<Tensor<WgpuRuntime>> {
        native_layer_norm(self, a, weight, bias, eps)
    }

    fn group_norm(
        &self,
        input: &Tensor<WgpuRuntime>,
        weight: &Tensor<WgpuRuntime>,
        bias: &Tensor<WgpuRuntime>,
        num_groups: usize,
        eps: f32,
    ) -> Result<Tensor<WgpuRuntime>> {
        native_group_norm(self, input, weight, bias, num_groups, eps)
    }

    fn fused_add_rms_norm(
        &self,
        x: &Tensor<WgpuRuntime>,
        residual: &Tensor<WgpuRuntime>,
        weight: &Tensor<WgpuRuntime>,
        eps: f32,
    ) -> Result<(Tensor<WgpuRuntime>, Tensor<WgpuRuntime>)> {
        native_fused_add_rms_norm(self, x, residual, weight, eps)
    }

    fn fused_add_layer_norm(
        &self,
        x: &Tensor<WgpuRuntime>,
        residual: &Tensor<WgpuRuntime>,
        weight: &Tensor<WgpuRuntime>,
        bias: &Tensor<WgpuRuntime>,
        eps: f32,
    ) -> Result<(Tensor<WgpuRuntime>, Tensor<WgpuRuntime>)> {
        native_fused_add_layer_norm(self, x, residual, weight, bias, eps)
    }

    fn fused_add_rms_norm_bwd(
        &self,
        grad: &Tensor<WgpuRuntime>,
        pre_norm: &Tensor<WgpuRuntime>,
        weight: &Tensor<WgpuRuntime>,
        eps: f32,
    ) -> Result<(Tensor<WgpuRuntime>, Tensor<WgpuRuntime>)> {
        native_fused_add_rms_norm_bwd(self, grad, pre_norm, weight, eps)
    }

    fn fused_add_layer_norm_bwd(
        &self,
        grad: &Tensor<WgpuRuntime>,
        pre_norm: &Tensor<WgpuRuntime>,
        weight: &Tensor<WgpuRuntime>,
        bias: &Tensor<WgpuRuntime>,
        eps: f32,
    ) -> Result<(
        Tensor<WgpuRuntime>,
        Tensor<WgpuRuntime>,
        Tensor<WgpuRuntime>,
    )> {
        native_fused_add_layer_norm_bwd(self, grad, pre_norm, weight, bias, eps)
    }
}