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
//! GH-177: Encoder-decoder forward pass for T5/Whisper
//!
//! Provides `encode()` and `decode()` methods on OwnedQuantizedModel
//! for encoder-decoder architectures (T5, Whisper).
//!
//! The encoder runs bidirectional attention (no causal mask).
//! The decoder runs causal self-attention + cross-attention
//! (Q from decoder, K/V from encoder output).
use crate::error::Result;
use crate::gguf::ops;
use crate::gguf::OwnedQuantizedModel;
/// Encoder output: hidden states from the encoder.
///
/// Stored and passed to the decoder for cross-attention.
/// For T5-small (6 layers, hidden=512), this is ~2KB per token.
#[derive(Debug, Clone)]
pub struct EncoderOutput {
/// Hidden states: [seq_len, hidden_dim]
pub hidden_states: Vec<f32>,
/// Sequence length of the encoder input
pub seq_len: usize,
/// Hidden dimension
pub hidden_dim: usize,
}
impl OwnedQuantizedModel {
/// GH-177: Run the encoder forward pass.
///
/// Processes input tokens through the encoder layers using
/// bidirectional attention (each position attends to all others).
///
/// Each encoder layer: LayerNorm → self-attn (bidirectional) →
/// residual → LayerNorm → FFN (GELU) → residual.
pub fn encode(&self, input_tokens: &[u32]) -> Result<EncoderOutput> {
if !self.config.is_encoder_decoder() {
return Err(crate::error::RealizarError::UnsupportedOperation {
operation: "encode".to_string(),
reason: format!(
"encode() requires encoder-decoder architecture, \
got '{}'",
self.config.architecture
),
});
}
let hidden_dim = self.config.hidden_dim;
let seq_len = input_tokens.len();
// Token embedding lookup
let mut hidden = Vec::with_capacity(seq_len * hidden_dim);
for &token_id in input_tokens {
let start = token_id as usize * hidden_dim;
let end = start + hidden_dim;
if end > self.token_embedding.len() {
return Err(crate::error::RealizarError::InferenceError(format!(
"Token ID {} out of range (embedding size: {})",
token_id,
self.token_embedding.len() / hidden_dim
)));
}
hidden.extend_from_slice(&self.token_embedding[start..end]);
}
// GH-177 Item 2: Run through encoder layers with bidirectional attention.
// Uses encoder_layers if populated, falls back to self.layers for
// models where encoder/decoder share the same layer format.
let encoder_layers = if self.encoder_layers.is_empty() {
&self.layers
} else {
&self.encoder_layers
};
let head_dim = self.config.head_dim();
let num_heads = self.config.num_heads;
let num_kv_heads = self.config.num_kv_heads;
let scale = 1.0 / (head_dim as f32).sqrt();
let tile_size = 64;
for layer in encoder_layers.iter() {
// Pre-attention LayerNorm (T5 uses LayerNorm, not RMSNorm)
let use_rmsnorm = self.config.constraints.uses_rmsnorm();
let normed = if use_rmsnorm {
ops::rms_norm(&hidden, &layer.attn_norm_weight, self.config.eps)
} else {
ops::layer_norm(
&hidden,
&layer.attn_norm_weight,
layer.attn_norm_bias.as_deref(),
self.config.eps,
)
};
// QKV projection for all positions
let mut all_attn_out = vec![0.0f32; seq_len * hidden_dim];
// Project Q, K, V for entire sequence
let q_dim = num_heads * head_dim;
let kv_dim_per = num_kv_heads * head_dim;
let mut full_q = Vec::with_capacity(seq_len * q_dim);
let mut full_k = Vec::with_capacity(seq_len * kv_dim_per);
let mut full_v = Vec::with_capacity(seq_len * kv_dim_per);
for pos in 0..seq_len {
let pos_hidden = &normed[pos * hidden_dim..(pos + 1) * hidden_dim];
let mut qkv = self.qkv_matmul(pos_hidden, &layer.qkv_weight)?;
if let Some(ref bias) = layer.qkv_bias {
ops::add_bias(&mut qkv, bias);
}
let q = &qkv[0..q_dim];
let k = &qkv[q_dim..q_dim + kv_dim_per];
let v = &qkv[q_dim + kv_dim_per..q_dim + 2 * kv_dim_per];
full_q.extend_from_slice(q);
full_k.extend_from_slice(k);
full_v.extend_from_slice(v);
}
// Note: T5 uses relative position bias, not RoPE.
// For now, skip positional encoding (relative bias not yet impl).
// This is correct for models that don't use any positional encoding
// in the attention computation itself (e.g., absolute positions
// added at embedding time).
// Multi-head bidirectional attention
let group_size = if num_kv_heads > 0 {
num_heads / num_kv_heads
} else {
1
};
for h in 0..num_heads {
let kv_h = h / group_size;
// Extract per-head Q, K, V across all positions
let mut h_q = Vec::with_capacity(seq_len * head_dim);
let mut h_k = Vec::with_capacity(seq_len * head_dim);
let mut h_v = Vec::with_capacity(seq_len * head_dim);
for pos in 0..seq_len {
let q_start = pos * q_dim + h * head_dim;
h_q.extend_from_slice(&full_q[q_start..q_start + head_dim]);
let k_start = pos * kv_dim_per + kv_h * head_dim;
h_k.extend_from_slice(&full_k[k_start..k_start + head_dim]);
let v_start = pos * kv_dim_per + kv_h * head_dim;
h_v.extend_from_slice(&full_v[v_start..v_start + head_dim]);
}
// Bidirectional attention (all positions attend to all)
let head_out = self.tiled_bidirectional_attention(
&h_q, &h_k, &h_v, seq_len, head_dim, scale, tile_size,
)?;
// Scatter back to multi-head output
for pos in 0..seq_len {
let src = &head_out[pos * head_dim..(pos + 1) * head_dim];
let dst_start = pos * hidden_dim + h * head_dim;
all_attn_out[dst_start..dst_start + head_dim].copy_from_slice(src);
}
}
// Output projection + residual
for pos in 0..seq_len {
let attn_pos = &all_attn_out[pos * hidden_dim..(pos + 1) * hidden_dim];
let mut proj = self.fused_matmul(attn_pos, &layer.attn_output_weight)?;
if let Some(ref bias) = layer.attn_output_bias {
ops::add_bias(&mut proj, bias);
}
for i in 0..hidden_dim {
hidden[pos * hidden_dim + i] += proj[i];
}
}
// FFN block: LayerNorm → up → GELU → down → residual
for pos in 0..seq_len {
let pos_hidden = &hidden[pos * hidden_dim..(pos + 1) * hidden_dim];
let ffn_input = if let Some(ref ffn_norm) = layer.ffn_norm_weight {
if use_rmsnorm {
ops::rms_norm(pos_hidden, ffn_norm, self.config.eps)
} else {
ops::layer_norm(
pos_hidden,
ffn_norm,
layer.ffn_norm_bias.as_deref(),
self.config.eps,
)
}
} else {
pos_hidden.to_vec()
};
// T5 uses GELU FFN (no gate weight)
let mut ffn_hidden = self.fused_matmul(&ffn_input, &layer.ffn_up_weight)?;
if let Some(ref bias) = layer.ffn_up_bias {
ops::add_bias(&mut ffn_hidden, bias);
}
ops::gelu(&mut ffn_hidden);
let mut ffn_out = self.fused_matmul(&ffn_hidden, &layer.ffn_down_weight)?;
if let Some(ref bias) = layer.ffn_down_bias {
ops::add_bias(&mut ffn_out, bias);
}
// Residual
for i in 0..hidden_dim {
hidden[pos * hidden_dim + i] += ffn_out[i];
}
}
}
// Encoder final norm
let use_rmsnorm = self.config.constraints.uses_rmsnorm();
let enc_norm_weight = self
.encoder_output_norm_weight
.as_deref()
.unwrap_or(&self.output_norm_weight);
let enc_norm_bias = self
.encoder_output_norm_bias
.as_deref()
.or(self.output_norm_bias.as_deref());
let mut normed_hidden = Vec::with_capacity(seq_len * hidden_dim);
for pos in 0..seq_len {
let pos_h = &hidden[pos * hidden_dim..(pos + 1) * hidden_dim];
let normed = if use_rmsnorm {
ops::rms_norm(pos_h, enc_norm_weight, self.config.eps)
} else {
ops::layer_norm(pos_h, enc_norm_weight, enc_norm_bias, self.config.eps)
};
normed_hidden.extend_from_slice(&normed);
}
Ok(EncoderOutput {
hidden_states: normed_hidden,
seq_len,
hidden_dim,
})
}
}
include!("encoder_decoder_decode.rs");