File size: 18,107 Bytes
1269259 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 | // SPF Smart Gateway - Transformer Decoder
// Copyright 2026 Joseph Stone - All Rights Reserved
//
// N stacked decoder layers: masked self-attention → cross-attention → FFN
// Each sublayer has residual connection + layer norm (pre-norm architecture).
// Causal masking prevents attending to future tokens.
// Cross-attention allows decoder to attend to encoder output.
//
// Depends on: tensor.rs, attention.rs, ffn.rs (Layers 0-1)
use crate::tensor::Tensor;
use crate::attention::{AttentionCache, AttentionConfig, MultiHeadAttention};
use crate::ffn::{FfnCache, FfnConfig, FeedForward};
// ============================================================================
// ACTIVATION CACHE (for backward pass — P2-C)
// ============================================================================
/// Cached activations from one decoder layer forward pass (causal/self-only mode)
pub struct DecoderLayerCache {
/// Input before LN1 (for layer_norm_backward)
pub ln1_input: Tensor,
/// Self-attention cache (Q, K, V, attn_weights, input, scale)
pub self_attn_cache: AttentionCache,
/// Input before LN3 (before FFN — ln2 unused in causal path)
pub ln3_input: Tensor,
/// FFN cache (input, hidden_pre_gelu)
pub ffn_cache: FfnCache,
/// Original layer input (for residual backward)
pub residual_input: Tensor,
}
// ============================================================================
// DECODER CONFIGURATION
// ============================================================================
/// Configuration for the decoder stack
#[derive(Debug, Clone)]
pub struct DecoderConfig {
/// Number of decoder layers
pub n_layers: usize,
/// Model dimension
pub d_model: usize,
/// Number of attention heads
pub n_heads: usize,
/// Feed-forward hidden dimension
pub d_ff: usize,
/// Maximum sequence length
pub max_seq_len: usize,
/// Layer norm epsilon
pub ln_eps: f32,
}
impl DecoderConfig {
/// SPF Writer default: matches encoder config
pub fn spf_writer() -> Self {
Self {
n_layers: 6,
d_model: 256,
n_heads: 8,
d_ff: 1024,
max_seq_len: 2048,
ln_eps: 1e-5,
}
}
/// Smaller config for testing
pub fn small() -> Self {
Self {
n_layers: 2,
d_model: 64,
n_heads: 4,
d_ff: 256,
max_seq_len: 512,
ln_eps: 1e-5,
}
}
}
// ============================================================================
// SINGLE DECODER LAYER
// ============================================================================
/// One decoder layer with three sublayers:
///
/// 1. Masked self-attention (causal — can't see future tokens)
/// 2. Cross-attention (attends to encoder output)
/// 3. Feed-forward network
///
/// Pre-norm architecture:
/// x → LN → MaskedSelfAttn → +residual
/// → LN → CrossAttn → +residual
/// → LN → FFN → +residual
pub struct DecoderLayer {
/// Masked multi-head self-attention (causal)
pub self_attn: MultiHeadAttention,
/// Cross-attention to encoder output
pub cross_attn: MultiHeadAttention,
/// Feed-forward network
pub ffn: FeedForward,
/// Layer norm before self-attention
pub ln1_gamma: Tensor,
pub ln1_beta: Tensor,
/// Layer norm before cross-attention
pub ln2_gamma: Tensor,
pub ln2_beta: Tensor,
/// Layer norm before FFN
pub ln3_gamma: Tensor,
pub ln3_beta: Tensor,
/// Epsilon for layer norm
pub ln_eps: f32,
}
impl DecoderLayer {
/// Initialize a single decoder layer
pub fn new(d_model: usize, n_heads: usize, d_ff: usize, ln_eps: f32, seed: u64) -> Self {
let self_attn_config = AttentionConfig {
d_model,
n_heads,
causal: true, // Decoder self-attention is causal
};
let cross_attn_config = AttentionConfig {
d_model,
n_heads,
causal: false, // Cross-attention is bidirectional over encoder output
};
let ffn_config = FfnConfig { d_model, d_ff };
Self {
self_attn: MultiHeadAttention::new(self_attn_config, seed),
cross_attn: MultiHeadAttention::new(cross_attn_config, seed + 50),
ffn: FeedForward::new(ffn_config, seed + 100),
ln1_gamma: Tensor::ones(&[d_model]),
ln1_beta: Tensor::zeros(&[d_model]),
ln2_gamma: Tensor::ones(&[d_model]),
ln2_beta: Tensor::zeros(&[d_model]),
ln3_gamma: Tensor::ones(&[d_model]),
ln3_beta: Tensor::zeros(&[d_model]),
ln_eps,
}
}
/// Forward pass with encoder output for cross-attention
/// x: decoder input [batch, dec_seq, d_model]
/// encoder_output: [batch, enc_seq, d_model]
pub fn forward(&self, x: &Tensor, encoder_output: &Tensor) -> Result<Tensor, String> {
// 1. Masked self-attention with residual
let normed = x.layer_norm(&self.ln1_gamma, &self.ln1_beta, self.ln_eps)?;
let self_attn_out = self.self_attn.forward(&normed)?;
let x = x.add(&self_attn_out)?;
// 2. Cross-attention with residual
let normed = x.layer_norm(&self.ln2_gamma, &self.ln2_beta, self.ln_eps)?;
let cross_attn_out = self.cross_attn.forward_cross(&normed, encoder_output)?;
let x = x.add(&cross_attn_out)?;
// 3. FFN with residual
let normed = x.layer_norm(&self.ln3_gamma, &self.ln3_beta, self.ln_eps)?;
let ffn_out = self.ffn.forward(&normed)?;
x.add(&ffn_out)
}
/// Forward pass without cross-attention (decoder-only / causal LM mode)
/// Used when running as a causal language model without encoder.
/// The Researcher transformer uses this mode.
pub fn forward_self_only(&self, x: &Tensor) -> Result<Tensor, String> {
// 1. Masked self-attention with residual
let normed = x.layer_norm(&self.ln1_gamma, &self.ln1_beta, self.ln_eps)?;
let self_attn_out = self.self_attn.forward(&normed)?;
let x = x.add(&self_attn_out)?;
// Skip cross-attention (no encoder output)
// 2. FFN with residual (use ln3 — ln2 unused in this path)
let normed = x.layer_norm(&self.ln3_gamma, &self.ln3_beta, self.ln_eps)?;
let ffn_out = self.ffn.forward(&normed)?;
x.add(&ffn_out)
}
/// Forward pass without cross-attention, with cached activations for backward.
/// Output is IDENTICAL to forward_self_only(). Cache is additional data only.
pub fn forward_self_only_with_cache(&self, x: &Tensor) -> Result<(Tensor, DecoderLayerCache), String> {
let residual_input = x.clone();
// 1. Masked self-attention with residual
let ln1_input = x.clone();
let normed = x.layer_norm(&self.ln1_gamma, &self.ln1_beta, self.ln_eps)?;
let (self_attn_out, self_attn_cache) = self.self_attn.forward_with_cache(&normed)?;
let x = x.add(&self_attn_out)?;
// Skip cross-attention (no encoder output)
// 2. FFN with residual (use ln3 — ln2 unused in this path)
let ln3_input = x.clone();
let normed = x.layer_norm(&self.ln3_gamma, &self.ln3_beta, self.ln_eps)?;
let (ffn_out, ffn_cache) = self.ffn.forward_with_cache(&normed)?;
let output = x.add(&ffn_out)?;
let cache = DecoderLayerCache {
ln1_input,
self_attn_cache,
ln3_input,
ffn_cache,
residual_input,
};
Ok((output, cache))
}
/// Total parameters in this layer
pub fn num_params(&self) -> usize {
let d = self.ln1_gamma.numel();
// self_attn + cross_attn + ffn + 3 layer norms × (gamma + beta)
self.self_attn.num_params() + self.cross_attn.num_params() + self.ffn.num_params() + 6 * d
}
/// Collect all weight tensors
pub fn weights(&self) -> Vec<&Tensor> {
let mut w = self.self_attn.weights();
w.extend(self.cross_attn.weights());
w.extend(self.ffn.weights());
w.extend([
&self.ln1_gamma, &self.ln1_beta,
&self.ln2_gamma, &self.ln2_beta,
&self.ln3_gamma, &self.ln3_beta,
]);
w
}
/// Collect all weight tensors mutably
pub fn weights_mut(&mut self) -> Vec<&mut Tensor> {
let mut w = self.self_attn.weights_mut();
w.extend(self.cross_attn.weights_mut());
w.extend(self.ffn.weights_mut());
w.extend([
&mut self.ln1_gamma, &mut self.ln1_beta,
&mut self.ln2_gamma, &mut self.ln2_beta,
&mut self.ln3_gamma, &mut self.ln3_beta,
]);
w
}
}
// ============================================================================
// DECODER STACK
// ============================================================================
/// Full decoder: positional encoding + N decoder layers + final layer norm.
pub struct Decoder {
pub config: DecoderConfig,
/// Positional encoding table [max_seq_len, d_model]
pub pos_encoding: Tensor,
/// Stack of decoder layers
pub layers: Vec<DecoderLayer>,
/// Final layer norm
pub final_ln_gamma: Tensor,
pub final_ln_beta: Tensor,
}
impl Decoder {
/// Initialize decoder with given config
pub fn new(config: DecoderConfig, seed: u64) -> Self {
// Reuse encoder's sinusoidal positional encoding
let pos_encoding = crate::encoder::sinusoidal_positional_encoding(
config.max_seq_len,
config.d_model,
);
let layers: Vec<DecoderLayer> = (0..config.n_layers)
.map(|i| {
DecoderLayer::new(
config.d_model,
config.n_heads,
config.d_ff,
config.ln_eps,
seed + (i as u64) * 2000,
)
})
.collect();
Self {
final_ln_gamma: Tensor::ones(&[config.d_model]),
final_ln_beta: Tensor::zeros(&[config.d_model]),
pos_encoding,
layers,
config,
}
}
/// Forward pass with encoder output (encoder-decoder mode)
/// embeddings: [batch, dec_seq, d_model]
/// encoder_output: [batch, enc_seq, d_model]
pub fn forward(
&self,
embeddings: &Tensor,
encoder_output: &Tensor,
) -> Result<Tensor, String> {
let seq_len = embeddings.shape[1];
if seq_len > self.config.max_seq_len {
return Err(format!(
"Sequence length {} exceeds max {}",
seq_len, self.config.max_seq_len
));
}
let mut x = self.add_positional_encoding(embeddings)?;
for layer in &self.layers {
x = layer.forward(&x, encoder_output)?;
}
x.layer_norm(&self.final_ln_gamma, &self.final_ln_beta, self.config.ln_eps)
}
/// Forward pass without encoder (decoder-only / causal LM mode)
/// Used by the Researcher transformer and for autoregressive generation
pub fn forward_causal(&self, embeddings: &Tensor) -> Result<Tensor, String> {
let seq_len = embeddings.shape[1];
if seq_len > self.config.max_seq_len {
return Err(format!(
"Sequence length {} exceeds max {}",
seq_len, self.config.max_seq_len
));
}
let mut x = self.add_positional_encoding(embeddings)?;
for layer in &self.layers {
x = layer.forward_self_only(&x)?;
}
x.layer_norm(&self.final_ln_gamma, &self.final_ln_beta, self.config.ln_eps)
}
/// Forward pass without encoder, with cached activations for backward.
/// Output is IDENTICAL to forward_causal(). Caches are additional data only.
pub fn forward_causal_with_cache(&self, embeddings: &Tensor) -> Result<(Tensor, Vec<DecoderLayerCache>), String> {
let seq_len = embeddings.shape[1];
if seq_len > self.config.max_seq_len {
return Err(format!(
"Sequence length {} exceeds max {}",
seq_len, self.config.max_seq_len
));
}
let mut x = self.add_positional_encoding(embeddings)?;
let mut layer_caches = Vec::with_capacity(self.layers.len());
for layer in &self.layers {
let (out, cache) = layer.forward_self_only_with_cache(&x)?;
x = out;
layer_caches.push(cache);
}
let output = x.layer_norm(&self.final_ln_gamma, &self.final_ln_beta, self.config.ln_eps)?;
Ok((output, layer_caches))
}
/// Add sinusoidal positional encoding to embeddings
fn add_positional_encoding(&self, embeddings: &Tensor) -> Result<Tensor, String> {
let batch = embeddings.shape[0];
let seq_len = embeddings.shape[1];
let d_model = embeddings.shape[2];
let pos_enc = self.pos_encoding.slice(0, seq_len)?;
let mut data = embeddings.data.clone();
for b in 0..batch {
for s in 0..seq_len {
for d in 0..d_model {
data[(b * seq_len + s) * d_model + d] += pos_enc.data[s * d_model + d];
}
}
}
Tensor::from_data(data, embeddings.shape.clone())
}
/// Total parameters in the decoder
pub fn num_params(&self) -> usize {
let layer_params: usize = self.layers.iter().map(|l| l.num_params()).sum();
layer_params + 2 * self.config.d_model // final LN
}
/// Collect all weight tensors
pub fn weights(&self) -> Vec<&Tensor> {
let mut w: Vec<&Tensor> = Vec::new();
for layer in &self.layers {
w.extend(layer.weights());
}
w.push(&self.final_ln_gamma);
w.push(&self.final_ln_beta);
w
}
/// Collect all weight tensors mutably
pub fn weights_mut(&mut self) -> Vec<&mut Tensor> {
let mut w: Vec<&mut Tensor> = Vec::new();
for layer in &mut self.layers {
w.extend(layer.weights_mut());
}
w.push(&mut self.final_ln_gamma);
w.push(&mut self.final_ln_beta);
w
}
}
// ============================================================================
// TESTS
// ============================================================================
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_decoder_layer_with_encoder() {
let layer = DecoderLayer::new(64, 4, 256, 1e-5, 42);
let dec_input = Tensor::randn(&[1, 4, 64], 99); // decoder: 4 tokens
let enc_output = Tensor::randn(&[1, 8, 64], 100); // encoder: 8 tokens
let out = layer.forward(&dec_input, &enc_output).unwrap();
assert_eq!(out.shape, vec![1, 4, 64]); // follows decoder seq_len
assert!(out.data.iter().all(|v| v.is_finite()));
}
#[test]
fn test_decoder_layer_self_only() {
let layer = DecoderLayer::new(64, 4, 256, 1e-5, 42);
let x = Tensor::randn(&[1, 6, 64], 99);
let out = layer.forward_self_only(&x).unwrap();
assert_eq!(out.shape, vec![1, 6, 64]);
assert!(out.data.iter().all(|v| v.is_finite()));
}
#[test]
fn test_decoder_full_forward() {
let config = DecoderConfig::small();
let decoder = Decoder::new(config, 42);
let dec_emb = Tensor::randn(&[1, 4, 64], 99);
let enc_out = Tensor::randn(&[1, 8, 64], 100);
let out = decoder.forward(&dec_emb, &enc_out).unwrap();
assert_eq!(out.shape, vec![1, 4, 64]);
assert!(out.data.iter().all(|v| v.is_finite()));
}
#[test]
fn test_decoder_causal_forward() {
let config = DecoderConfig::small();
let decoder = Decoder::new(config, 42);
let x = Tensor::randn(&[1, 6, 64], 99);
let out = decoder.forward_causal(&x).unwrap();
assert_eq!(out.shape, vec![1, 6, 64]);
assert!(out.data.iter().all(|v| v.is_finite()));
}
#[test]
fn test_decoder_seq_exceeds_max() {
let config = DecoderConfig { max_seq_len: 10, ..DecoderConfig::small() };
let decoder = Decoder::new(config, 42);
let x = Tensor::randn(&[1, 20, 64], 99);
let enc = Tensor::randn(&[1, 5, 64], 100);
assert!(decoder.forward(&x, &enc).is_err());
}
#[test]
fn test_decoder_causal_seq_exceeds_max() {
let config = DecoderConfig { max_seq_len: 10, ..DecoderConfig::small() };
let decoder = Decoder::new(config, 42);
let x = Tensor::randn(&[1, 20, 64], 99);
assert!(decoder.forward_causal(&x).is_err());
}
#[test]
fn test_decoder_num_params() {
let config = DecoderConfig::small(); // 2 layers, d=64, ff=256
let decoder = Decoder::new(config, 42);
let params = decoder.num_params();
// Each layer: self_attn(16640) + cross_attn(16640) + ffn(33088) + 6×64(LN) = 66752
// 2 layers + final LN: 2×66752 + 128 = 133632
assert_eq!(params, 133632);
}
#[test]
fn test_decoder_batch() {
let config = DecoderConfig::small();
let decoder = Decoder::new(config, 42);
let x = Tensor::randn(&[3, 4, 64], 99); // batch=3
let enc = Tensor::randn(&[3, 6, 64], 100);
let out = decoder.forward(&x, &enc).unwrap();
assert_eq!(out.shape, vec![3, 4, 64]);
}
#[test]
fn test_decoder_weights_count() {
let config = DecoderConfig::small();
let decoder = Decoder::new(config, 42);
let weights = decoder.weights();
// Each layer: 8(self_attn) + 8(cross_attn) + 4(ffn) + 6(LN) = 26
// 2 layers + 2 final LN = 54
assert_eq!(weights.len(), 54);
}
#[test]
fn test_decoder_layer_params() {
let layer = DecoderLayer::new(64, 4, 256, 1e-5, 42);
// self_attn: 16640, cross_attn: 16640, ffn: 33088, 6×64 LN: 384
assert_eq!(layer.num_params(), 66752);
}
}
|