rumus 0.2.0

A native-Rust deep learning framework with explicit memory safety and hardware acceleration
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

//! End-to-end "Tiny GPT" test.
//!
//! Architecture: Token Embedding + Positional Embedding + TransformerBlock + Linear Head.
//! Task: Next-token prediction on a repeating sequence [0,1,2,3,0,1,2,3,...].
//! Proves that attention, LayerNorm, GELU, bmm, softmax, broadcasting,
//! cross-entropy, and AdamW all compose correctly through the autograd engine.

use rumus::nn::{
    cross_entropy_loss, Embedding, LayerNorm, Linear, Module, TransformerBlock,
};
use rumus::optim::AdamW;
use rumus::tensor::Tensor;
use rumus::train::Trainer;


// ---------------------------------------------------------------------------
// TinyGPT model
// ---------------------------------------------------------------------------

#[allow(dead_code)]
struct TinyGPT {
    token_emb: Embedding,
    pos_emb: Embedding,
    block: TransformerBlock,
    ln_final: LayerNorm,
    head: Linear,
    seq_len: usize,
}

impl TinyGPT {
    fn new(vocab_size: usize, seq_len: usize, d_model: usize, num_heads: usize) -> Self {
        Self {
            token_emb: Embedding::new(vocab_size, d_model),
            pos_emb: Embedding::new(seq_len, d_model),
            block: TransformerBlock::new(d_model, num_heads),
            ln_final: LayerNorm::new(d_model, 1e-5),
            head: Linear::new(d_model, vocab_size, false),
            seq_len,
        }
    }

    fn forward(&self, token_ids: &Tensor, mask: &Tensor) -> Tensor {
        let b = token_ids.shape()[0];
        let s = token_ids.shape()[1];
        let d = self.token_emb.embed_dim;

        // Token embeddings: [B, S] → [B, S, D]
        let tok = self.token_emb.forward(token_ids);

        // Positional embeddings: [S] → [S, D] → broadcast add to [B, S, D]
        let pos_ids = Tensor::new((0..s).map(|i| i as f32).collect(), vec![s]);
        let pos = self.pos_emb.forward(&pos_ids); // [S, D]
        let x = tok.broadcast_add(&pos);

        // Transformer block
        let x = self.block.forward(&x, Some(mask));

        // Final LayerNorm + vocab projection
        let x = self.ln_final.forward(&x);
        let flat = x.reshape_tracked(vec![b * s, d]);
        let logits = self.head.forward(&flat);
        logits.reshape_tracked(vec![b, s, self.head.weight.tensor.shape()[1]])
    }

    fn parameters(&self) -> Vec<rumus::nn::Parameter> {
        let mut p = self.token_emb.parameters();
        p.extend(self.pos_emb.parameters());
        p.extend(self.block.parameters());
        p.extend(self.ln_final.parameters());
        p.extend(self.head.parameters());
        p
    }
}

// ---------------------------------------------------------------------------
// Data generation
// ---------------------------------------------------------------------------

/// Generate a causal mask: upper triangle = -1e9, lower+diagonal = 0.
fn causal_mask(seq_len: usize) -> Tensor {
    let mut data = vec![0.0f32; seq_len * seq_len];
    for i in 0..seq_len {
        for j in (i + 1)..seq_len {
            data[i * seq_len + j] = -1e9;
        }
    }
    Tensor::new(data, vec![seq_len, seq_len])
}

/// Generate repeating sequences: `[0, 1, 2, 3, 0, 1, ...]`.
/// Returns (input, target) where target is input shifted left by 1.
fn make_repeating_data(batch: usize, seq_len: usize, vocab: usize) -> (Tensor, Tensor) {
    let mut input_data = Vec::with_capacity(batch * seq_len);
    let mut target_data = Vec::with_capacity(batch * seq_len);
    for b in 0..batch {
        for s in 0..seq_len {
            let token = ((b + s) % vocab) as f32;
            let next = ((b + s + 1) % vocab) as f32;
            input_data.push(token);
            target_data.push(next);
        }
    }
    (
        Tensor::new(input_data, vec![batch, seq_len]),
        Tensor::new(target_data, vec![batch * seq_len]),
    )
}

// ---------------------------------------------------------------------------
// CPU Test
// ---------------------------------------------------------------------------

#[test]
fn test_tiny_gpt_cpu() {
    let vocab = 4;
    let seq_len = 6;
    let d_model = 32;
    let num_heads = 1;
    let batch = 8;

    let (inputs, targets) = make_repeating_data(batch, seq_len, vocab);
    let mask = causal_mask(seq_len);

    let model = TinyGPT::new(vocab, seq_len, d_model, num_heads);
    let optimizer = AdamW::new(model.parameters(), 0.001);
    let mut trainer = Trainer::new(optimizer);

    let mut final_loss = f32::MAX;
    for _epoch in 0..200 {
        trainer.reset_epoch();
        let loss = trainer
            .train_step(|| {
                let logits = model.forward(&inputs, &mask); // [B, S, vocab]
                let logits_flat = logits.reshape_tracked(vec![batch * seq_len, vocab]);
                cross_entropy_loss(&logits_flat, &targets)
            })
            .expect("train_step failed");
        final_loss = loss;
    }

    assert!(
        final_loss < 1.0,
        "Tiny GPT did not converge: final loss = {:.4}",
        final_loss,
    );
}

// ---------------------------------------------------------------------------
// GPU Test
// ---------------------------------------------------------------------------

#[cfg(feature = "gpu")]
#[test]
fn test_tiny_gpt_gpu() {
    use rumus::backend::gpu::context::GpuContext;
    use rumus::nn::ModuleToGpu;

    if !GpuContext::is_available() {
        eprintln!("Skipping GPU Transformer test: no GPU available");
        return;
    }

    let vocab = 4;
    let seq_len = 6;
    let d_model = 32;
    let num_heads = 1;
    let batch = 8;

    let (inputs, targets) = make_repeating_data(batch, seq_len, vocab);
    let mask = causal_mask(seq_len);

    let model = TinyGPT::new(vocab, seq_len, d_model, num_heads);

    // Push everything to GPU.
    for p in model.parameters() { p.tensor.to_gpu(); }
    inputs.to_gpu();
    targets.to_gpu();
    mask.to_gpu();

    let optimizer = AdamW::new(model.parameters(), 0.001);
    let mut trainer = Trainer::new(optimizer);

    let ctx = GpuContext::get().unwrap();
    let mut final_loss = f32::MAX;
    for epoch in 0..200 {
        trainer.reset_epoch();
        let loss = trainer
            .train_step(|| {
                let logits = model.forward(&inputs, &mask);
                let logits_flat = logits.reshape_tracked(vec![batch * seq_len, vocab]);
                cross_entropy_loss(&logits_flat, &targets)
            })
            .expect("train_step failed");
        final_loss = loss;

        // Buffer pool leak check mid-training.
        if epoch == 10 {
            let count = ctx.pool.cached_count();
            assert!(
                count > 0,
                "BufferPool empty at epoch 10 — buffers not recycling",
            );
        }
    }

    assert!(
        final_loss < 1.0,
        "GPU Tiny GPT did not converge: final loss = {:.4}",
        final_loss,
    );
}