numr 0.5.2

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
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
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252

//! Fused SwiGLU activation with gradient support
//!
//! SwiGLU(gate, up) = silu(gate) * up
//!
//! Fused version saves one intermediate tensor vs composing var_silu + var_mul:
//! - Composed: stores gate, silu(gate), up (3 tensors)
//! - Fused: stores gate, up, output (3 tensors but recomputes sigmoid in backward)
//!
//! More importantly, the fused backward computes gradients in fewer ops.

use crate::autograd::Var;
use crate::autograd::var_ops::var_mul;
use crate::dtype::DType;
use crate::error::Result;
use crate::ops::{ActivationOps, BinaryOps, ScalarOps, TensorOps};
use crate::runtime::{Runtime, RuntimeClient};
use std::sync::Arc;

/// Fused SwiGLU: output = silu(gate) * up
///
/// # Arguments
/// * `gate` - Gate input (will have silu applied)
/// * `up` - Up projection (multiplied element-wise with activated gate)
/// * `client` - Runtime client
///
/// # Returns
/// The SwiGLU output with autograd tracking.
pub fn var_swiglu<R, C>(gate: &Var<R>, up: &Var<R>, client: &C) -> Result<Var<R>>
where
    R: Runtime<DType = DType>,
    C: RuntimeClient<R> + TensorOps<R> + ActivationOps<R> + ScalarOps<R> + BinaryOps<R>,
    R::Client: TensorOps<R> + ActivationOps<R> + ScalarOps<R> + BinaryOps<R>,
{
    // Forward: output = silu(gate) * up (fused single-pass kernel)
    let silu_gate = client.silu(gate.tensor())?;
    let output = client.silu_mul(gate.tensor(), up.tensor())?;

    if gate.requires_grad() || up.requires_grad() {
        let grad_fn = SwiGLUBackward::<R>::new(
            gate.id(),
            up.id(),
            gate.tensor().clone(),
            up.tensor().clone(),
            silu_gate,
            gate.grad_fn().cloned(),
            up.grad_fn().cloned(),
        );
        Ok(Var::from_op(output, Arc::new(grad_fn)))
    } else {
        Ok(Var::new(output, false))
    }
}

/// Backward for fused SwiGLU: output = silu(gate) * up
///
/// Gradients:
/// - d_gate = grad_output * up * silu'(gate)
///   = grad_output * up * (sigmoid(gate) * (1 + gate - silu(gate)))
/// - d_up   = grad_output * silu(gate)
pub struct SwiGLUBackward<R: Runtime> {
    input_ids: [crate::tensor::TensorId; 2],
    saved_gate: crate::tensor::Tensor<R>,
    saved_up: crate::tensor::Tensor<R>,
    saved_silu_gate: crate::tensor::Tensor<R>,
    gate_grad_fn: Option<Arc<dyn crate::autograd::GradFn<R>>>,
    up_grad_fn: Option<Arc<dyn crate::autograd::GradFn<R>>>,
}

impl<R: Runtime> SwiGLUBackward<R> {
    pub fn new(
        gate_id: crate::tensor::TensorId,
        up_id: crate::tensor::TensorId,
        gate: crate::tensor::Tensor<R>,
        up: crate::tensor::Tensor<R>,
        silu_gate: crate::tensor::Tensor<R>,
        gate_grad_fn: Option<Arc<dyn crate::autograd::GradFn<R>>>,
        up_grad_fn: Option<Arc<dyn crate::autograd::GradFn<R>>>,
    ) -> Self {
        Self {
            input_ids: [gate_id, up_id],
            saved_gate: gate,
            saved_up: up,
            saved_silu_gate: silu_gate,
            gate_grad_fn,
            up_grad_fn,
        }
    }
}

impl<R: Runtime<DType = DType>> crate::autograd::GradFn<R> for SwiGLUBackward<R>
where
    R::Client: TensorOps<R> + ActivationOps<R> + ScalarOps<R> + BinaryOps<R>,
{
    fn backward(
        &self,
        grad_output: &crate::tensor::Tensor<R>,
    ) -> Result<Vec<Option<crate::tensor::Tensor<R>>>> {
        let client = R::default_client(grad_output.device());

        // d_up = grad_output * silu(gate)
        let d_up = client.mul(grad_output, &self.saved_silu_gate)?;

        // d_gate = grad_output * up * silu'(gate)
        // silu'(x) = sigmoid(x) * (1 + x - silu(x))
        let sigmoid_gate = client.sigmoid(&self.saved_gate)?;
        let one_plus_gate = client.add_scalar(&self.saved_gate, 1.0)?;
        let one_plus_gate_minus_silu = client.sub(&one_plus_gate, &self.saved_silu_gate)?;
        let silu_deriv = client.mul(&sigmoid_gate, &one_plus_gate_minus_silu)?;
        let grad_times_up = client.mul(grad_output, &self.saved_up)?;
        let d_gate = client.mul(&grad_times_up, &silu_deriv)?;

        Ok(vec![Some(d_gate), Some(d_up)])
    }

    fn backward_var(&self, grad_output: &Var<R>) -> Result<Vec<Option<Var<R>>>>
    where
        R::Client: RuntimeClient<R> + TensorOps<R> + ActivationOps<R> + ScalarOps<R> + BinaryOps<R>,
    {
        let client = R::default_client(grad_output.tensor().device());

        // d_up = grad_output * silu(gate) [silu_gate is constant w.r.t. higher-order]
        let silu_var = Var::new(self.saved_silu_gate.clone(), false);
        let d_up = var_mul(grad_output, &silu_var, &client)?;

        // d_gate = grad_output * up * silu'(gate)
        let sigmoid_gate = client.sigmoid(&self.saved_gate)?;
        let one_plus_gate = client.add_scalar(&self.saved_gate, 1.0)?;
        let one_plus_gate_minus_silu = client.sub(&one_plus_gate, &self.saved_silu_gate)?;
        let silu_deriv = client.mul(&sigmoid_gate, &one_plus_gate_minus_silu)?;
        let silu_deriv_var = Var::new(silu_deriv, false);

        let up_var = Var::new(self.saved_up.clone(), false);
        let grad_times_up = var_mul(grad_output, &up_var, &client)?;
        let d_gate = var_mul(&grad_times_up, &silu_deriv_var, &client)?;

        Ok(vec![Some(d_gate), Some(d_up)])
    }

    fn inputs(&self) -> &[crate::tensor::TensorId] {
        &self.input_ids
    }

    fn input_grad_fns(&self) -> Vec<Option<Arc<dyn crate::autograd::GradFn<R>>>> {
        vec![self.gate_grad_fn.clone(), self.up_grad_fn.clone()]
    }

    fn saved_tensors(&self) -> &[crate::tensor::Tensor<R>] {
        std::slice::from_ref(&self.saved_gate)
    }

    fn name(&self) -> &'static str {
        "SwiGLUBackward"
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::autograd::backward;
    use crate::runtime::cpu::{CpuDevice, CpuRuntime};
    use crate::tensor::Tensor;

    #[test]
    fn test_swiglu_forward() {
        let device = CpuDevice::new();
        let client = CpuRuntime::default_client(&device);

        let gate = Var::new(
            Tensor::<CpuRuntime>::from_slice(&[0.0f32, 1.0, -1.0], &[3], &device),
            false,
        );
        let up = Var::new(
            Tensor::<CpuRuntime>::from_slice(&[1.0f32, 2.0, 3.0], &[3], &device),
            false,
        );

        let output = var_swiglu(&gate, &up, &client).unwrap();
        let data: Vec<f32> = output.tensor().to_vec();

        // silu(0) * 1 = 0 * 0.5 * 1 = 0
        assert!(data[0].abs() < 1e-6);
        // silu(1) * 2 = 0.7311 * 2 ≈ 1.4621
        let silu_1 = 1.0 / (1.0 + (-1.0f32).exp());
        assert!((data[1] - silu_1 * 2.0).abs() < 1e-4);
        // silu(-1) * 3 = -0.2689 * 3 ≈ -0.8067
        let silu_neg1 = -1.0 / (1.0 + 1.0f32.exp());
        assert!((data[2] - silu_neg1 * 3.0).abs() < 1e-4);
    }

    #[test]
    fn test_swiglu_backward_gate() {
        let device = CpuDevice::new();
        let client = CpuRuntime::default_client(&device);

        let gate = Var::new(
            Tensor::<CpuRuntime>::from_slice(&[1.0f32, -1.0], &[2], &device),
            true,
        );
        let up = Var::new(
            Tensor::<CpuRuntime>::from_slice(&[2.0f32, 3.0], &[2], &device),
            true,
        );

        let output = var_swiglu(&gate, &up, &client).unwrap();
        let loss = crate::autograd::var_sum(&output, &[], false, &client).unwrap();
        let grads = backward(&loss, &client).unwrap();

        let d_gate: Vec<f32> = grads.get(gate.id()).unwrap().to_vec();
        let d_up: Vec<f32> = grads.get(up.id()).unwrap().to_vec();

        // Verify d_up = silu(gate)
        for (i, &g) in [1.0f32, -1.0].iter().enumerate() {
            let expected_d_up = g * (1.0 / (1.0 + (-g).exp()));
            assert!(
                (d_up[i] - expected_d_up).abs() < 1e-5,
                "d_up[{i}]: got {}, expected {expected_d_up}",
                d_up[i]
            );
        }

        // Verify d_gate = up * silu'(gate)
        for (i, (&g, &u)) in [1.0f32, -1.0].iter().zip([2.0f32, 3.0].iter()).enumerate() {
            let sig = 1.0 / (1.0 + (-g).exp());
            let silu_g = g * sig;
            let silu_deriv = sig * (1.0 + g - silu_g);
            let expected = u * silu_deriv;
            assert!(
                (d_gate[i] - expected).abs() < 1e-4,
                "d_gate[{i}]: got {}, expected {expected}",
                d_gate[i]
            );
        }
    }

    #[test]
    fn test_swiglu_no_grad() {
        let device = CpuDevice::new();
        let client = CpuRuntime::default_client(&device);

        let gate = Var::new(
            Tensor::<CpuRuntime>::from_slice(&[1.0f32], &[1], &device),
            false,
        );
        let up = Var::new(
            Tensor::<CpuRuntime>::from_slice(&[2.0f32], &[1], &device),
            false,
        );

        let output = var_swiglu(&gate, &up, &client).unwrap();
        assert!(!output.requires_grad());
    }
}