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
253
254
255
256
257
258
259
260
261
use crate::autograd::{self, Variable};
use crate::tensor::{Device, Result};
use super::init;
use super::parameter::Parameter;
use super::Module;
/// Fully connected layer: `y = x @ W^T + b`.
///
/// Weight shape: `[out_features, in_features]`.
/// Bias shape: `[out_features]` (optional).
///
/// Input shape: `[batch, in_features]`.
/// Output shape: `[batch, out_features]`.
///
/// ```ignore
/// let layer = Linear::new(4, 2)?;
/// let x = Variable::new(Tensor::randn(&[8, 4], opts)?, false);
/// let y = layer.forward(&x)?;
/// assert_eq!(y.shape(), vec![8, 2]);
/// ```
pub struct Linear {
pub weight: Parameter,
pub bias: Option<Parameter>,
}
impl Linear {
/// Create a linear layer on CPU with bias.
pub fn new(in_features: i64, out_features: i64) -> Result<Self> {
Self::on_device(in_features, out_features, Device::CPU)
}
/// Create a linear layer on a specific device with bias.
pub fn on_device(in_features: i64, out_features: i64, device: Device) -> Result<Self> {
let w = init::kaiming_uniform(&[out_features, in_features], in_features, 5.0_f64.sqrt(), device)?;
let b = init::uniform_bias(in_features, &[out_features], device)?;
Ok(Linear {
weight: Parameter::new(w, "weight"),
bias: Some(Parameter::new(b, "bias")),
})
}
/// Create a linear layer without bias on CPU.
pub fn no_bias(in_features: i64, out_features: i64) -> Result<Self> {
Self::no_bias_on_device(in_features, out_features, Device::CPU)
}
/// Create a linear layer without bias on a specific device.
pub fn no_bias_on_device(in_features: i64, out_features: i64, device: Device) -> Result<Self> {
let w = init::kaiming_uniform(&[out_features, in_features], in_features, 5.0_f64.sqrt(), device)?;
Ok(Linear {
weight: Parameter::new(w, "weight"),
bias: None,
})
}
/// Build a `Linear` around an externally-owned weight `Parameter`,
/// enabling weight tying between this layer and another module that
/// already holds the same `Parameter`.
///
/// `Parameter` is `Clone`; a clone shares the underlying `Variable`
/// (and therefore the C++ tensor) by `Rc`. Gradients from every path
/// that touches the shared weight accumulate on the same leaf tensor,
/// exactly like PyTorch's `decoder.weight = embeddings.word_embeddings.weight`
/// pattern. `Graph::named_parameters()` deduplicates by pointer
/// identity, so a tied weight surfaces exactly once under whichever
/// node is visited first.
///
/// Pass `bias = Some(Parameter::new(...))` for the common
/// MLM / LM-head case (BERT, RoBERTa, DistilBERT ship a fresh
/// per-vocab decoder bias alongside the tied weight); pass `None` for
/// GPT-2-style heads with no bias.
///
/// Shape contract matches [`Linear::on_device`]: `weight.data()` must
/// have shape `[out_features, in_features]`. No device transfer
/// happens here — both `weight` and `bias` must already live on the
/// device the graph runs on.
///
/// ```ignore
/// use flodl::{Embedding, Linear, Parameter, Tensor, TensorOptions};
///
/// let embed = Embedding::new(vocab_size, hidden)?;
/// let tied = embed.weight.clone(); // shared Rc
/// let bias = Parameter::new(
/// Tensor::zeros(&[vocab_size], opts)?,
/// "bias",
/// );
/// let decoder = Linear::from_shared_weight(tied, Some(bias));
/// ```
pub fn from_shared_weight(weight: Parameter, bias: Option<Parameter>) -> Self {
Linear { weight, bias }
}
}
impl Module for Linear {
fn name(&self) -> &str { "linear" }
fn forward(&self, input: &Variable) -> Result<Variable> {
autograd::linear(
input,
&self.weight.variable,
self.bias.as_ref().map(|b| &b.variable),
)
}
fn parameters(&self) -> Vec<Parameter> {
let mut params = vec![self.weight.clone()];
if let Some(ref b) = self.bias {
params.push(b.clone());
}
params
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::tensor::{Tensor, test_device, test_opts};
#[test]
fn test_linear_forward_shape() {
let dev = test_device();
let layer = Linear::on_device(4, 2, dev).unwrap();
let x = Variable::new(Tensor::randn(&[8, 4], test_opts()).unwrap(), false);
let y = layer.forward(&x).unwrap();
assert_eq!(y.shape(), vec![8, 2]);
}
#[test]
fn test_linear_parameters_with_bias() {
let layer = Linear::on_device(4, 2, test_device()).unwrap();
let params = layer.parameters();
assert_eq!(params.len(), 2);
assert_eq!(params[0].variable.shape(), vec![2, 4]); // weight
assert_eq!(params[1].variable.shape(), vec![2]); // bias
}
#[test]
fn test_linear_no_bias() {
let layer = Linear::no_bias_on_device(4, 2, test_device()).unwrap();
let params = layer.parameters();
assert_eq!(params.len(), 1);
assert!(layer.bias.is_none());
let x = Variable::new(Tensor::randn(&[3, 4], test_opts()).unwrap(), false);
let y = layer.forward(&x).unwrap();
assert_eq!(y.shape(), vec![3, 2]);
}
#[test]
fn test_linear_gradient_flow() {
let dev = test_device();
let layer = Linear::on_device(3, 2, dev).unwrap();
let x = Variable::new(Tensor::randn(&[4, 3], test_opts()).unwrap(), false);
let y = layer.forward(&x).unwrap();
let loss = y.sum().unwrap();
loss.backward().unwrap();
let params = layer.parameters();
assert!(params[0].variable.grad().is_some(), "weight should have gradient");
assert!(params[1].variable.grad().is_some(), "bias should have gradient");
}
#[test]
fn test_linear_on_device() {
let dev = test_device();
let layer = Linear::on_device(4, 2, dev).unwrap();
assert_eq!(layer.weight.variable.device(), dev);
if let Some(ref b) = layer.bias {
assert_eq!(b.variable.device(), dev);
}
}
#[test]
fn test_linear_name() {
let layer = Linear::new(4, 2).unwrap();
assert_eq!(layer.name(), "linear");
}
/// Two `Linear`s built from the same `Parameter` share the underlying
/// `Variable` (pointer identity) and therefore the same C++ leaf
/// tensor. Backward through both paths must accumulate onto that one
/// tensor, and the accumulated gradient must be visible from either
/// handle.
#[test]
fn test_from_shared_weight_shares_rc_and_gradient() {
use std::rc::Rc;
// Shared weight: [out=2, in=3].
let shared = Parameter::new(
Tensor::randn(&[2, 3], test_opts()).unwrap(),
"weight",
);
let layer_a = Linear::from_shared_weight(shared.clone(), None);
let layer_b = Linear::from_shared_weight(shared.clone(), None);
// Rc pointer identity across both layers and the original handle.
assert!(Rc::ptr_eq(&layer_a.weight.variable.inner, &layer_b.weight.variable.inner));
assert!(Rc::ptr_eq(&shared.variable.inner, &layer_a.weight.variable.inner));
// Two distinct inputs, two forward paths, single scalar loss.
let x1 = Variable::new(Tensor::randn(&[4, 3], test_opts()).unwrap(), false);
let x2 = Variable::new(Tensor::randn(&[5, 3], test_opts()).unwrap(), false);
let y1 = layer_a.forward(&x1).unwrap();
let y2 = layer_b.forward(&x2).unwrap();
let loss = y1.sum().unwrap().add(&y2.sum().unwrap()).unwrap();
loss.backward().unwrap();
// The one leaf accumulates gradient from both paths; either
// handle sees it.
let g_a = layer_a.weight.variable.grad().expect("layer_a sees gradient");
let g_b = layer_b.weight.variable.grad().expect("layer_b sees gradient");
assert_eq!(g_a.shape(), vec![2, 3]);
assert_eq!(g_b.shape(), vec![2, 3]);
// Gradient value should equal sum over batch rows of each input
// (since d/dW sum(x @ W^T) = sum_rows(x) broadcast to out axis).
let expected = {
let s1 = x1.data().sum_dim(0, false).unwrap(); // [3]
let s2 = x2.data().sum_dim(0, false).unwrap(); // [3]
let row = s1.add(&s2).unwrap(); // [3]
// Broadcast to [2, 3] by stacking the same row twice.
row.unsqueeze(0).unwrap()
.expand(&[2, 3]).unwrap()
.contiguous().unwrap()
};
let got = g_a.to_f32_vec().unwrap();
let want = expected.to_f32_vec().unwrap();
for (g, w) in got.iter().zip(want.iter()) {
assert!((g - w).abs() < 1e-4, "grad mismatch: got {g}, want {w}");
}
}
/// `Graph::named_parameters()` deduplicates shared parameters by
/// `Rc::as_ptr` identity (flodl/src/graph/graph.rs). This test pins
/// the weight-tying contract: a shared decoder/embedding weight
/// surfaces exactly once, under the first-visited tag.
#[test]
fn test_from_shared_weight_dedups_in_graph_named_parameters() {
use crate::graph::FlowBuilder;
let shared = Parameter::new(
Tensor::randn(&[3, 4], test_opts()).unwrap(),
"weight",
);
let layer_a = Linear::from_shared_weight(shared.clone(), None);
let layer_b = Linear::from_shared_weight(shared.clone(), None);
let graph = FlowBuilder::from(layer_a)
.tag("embeddings")
.through(layer_b)
.tag("decoder")
.build()
.unwrap();
let named = graph.named_parameters();
assert_eq!(named.len(), 1, "shared weight should be listed once");
assert_eq!(named[0].0, "embeddings/weight", "first-visited tag wins");
}
}