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
// src/granite_conformer/shaw_attn.cpp - shared Shaw block-local attention.
//
// See shaw_attn.h for the contract. Both Granite families' encoders
// inline this op identically; this file is the single source of truth.
#include "shaw_attn.h"
#include "transcribe-log.h"
#include "ggml.h"
#include <cmath>
#include <cstdio>
namespace transcribe::granite_conformer {
namespace {
ggml_tensor * layer_norm(ggml_context * ctx, ggml_tensor * x,
ggml_tensor * gamma, ggml_tensor * beta,
float eps) {
ggml_tensor * y = ggml_norm(ctx, x, eps);
y = ggml_mul(ctx, y, gamma);
if (beta != nullptr) y = ggml_add(ctx, y, beta);
return y;
}
} // namespace
ggml_tensor * shaw_block_attn(ggml_context * ctx,
ggml_tensor * x,
ggml_tensor * zero_pad,
ggml_tensor * dists,
ggml_tensor * pad_mask_3d,
const ShawAttnWeights & w,
int n_heads,
int head_dim,
int context_size,
int num_blocks,
int T_enc,
float layer_norm_eps) {
const int64_t inner_dim = static_cast<int64_t>(n_heads) * head_dim;
const int64_t T_pad = static_cast<int64_t>(context_size) * num_blocks;
const float scale = 1.0f / std::sqrt(static_cast<float>(head_dim));
// Offline batch: x ne=[d_model, T_enc, B]. The block-local attention
// treats each (time-block) independently, so B utterances fold into the
// block axis: num_blocks_eff = num_blocks * B. Memory order is
// [.., T_pad(=ctx*num_blocks), B], so a single reshape splits the time
// axis into (ctx, num_blocks) with B outermost == (ctx, num_blocks*B).
// B == 1 (AR granite, single-shot) keeps num_blocks_eff == num_blocks and
// is byte-identical to the pre-batch graph. The caller tiles the per-block
// pad mask across B and sizes zero_pad with the batch.
const int64_t B = x->ne[2];
const int64_t num_blocks_eff = static_cast<int64_t>(num_blocks) * B;
// Pre-LayerNorm on the T_enc input.
ggml_tensor * h = layer_norm(ctx, x, w.norm_attn_w, w.norm_attn_b,
layer_norm_eps);
// Zero-pad along the time axis when T_enc is not block-aligned. The
// pad buffer is a caller-owned graph input; nullptr is a programmer
// error.
if (T_pad > T_enc) {
if (zero_pad == nullptr) {
log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
"granite_conformer::shaw_block_attn: "
"zero_pad is null but T_pad > T_enc");
return nullptr;
}
h = ggml_concat(ctx, h, zero_pad, /*dim=*/1);
}
// h ne = [d_model, T_pad]
// QKV projections. q has no bias; kv is fused (2*inner_dim) with no
// bias; out has bias.
ggml_tensor * q = ggml_mul_mat(ctx, w.attn_q_w, h); // [inner_dim, T_pad]
ggml_tensor * kv = ggml_mul_mat(ctx, w.attn_kv_w, h); // [2*inner_dim, T_pad]
// Split kv into k, v along ne[0]. The mul_mat output is contiguous, so
// views with a strided ne[0] are safe. ne[2] carries the utterance batch
// (1 for single-shot, == B otherwise).
ggml_tensor * k = ggml_view_3d(ctx, kv, inner_dim, kv->ne[1], kv->ne[2],
kv->nb[1], kv->nb[2], /*offset=*/0);
ggml_tensor * v = ggml_view_3d(ctx, kv, inner_dim, kv->ne[1], kv->ne[2],
kv->nb[1], kv->nb[2],
inner_dim * ggml_element_size(kv));
k = ggml_cont(ctx, k);
v = ggml_cont(ctx, v);
// Reshape into block-local form. Target ne layout
// [head_dim, context_size, n_heads, num_blocks_eff] so that
// ggml_mul_mat(k, q) does per-(head, block, utterance) batched attention.
auto reshape_qkv = [&](ggml_tensor * t) -> ggml_tensor * {
ggml_tensor * r = ggml_reshape_4d(ctx, t,
head_dim, n_heads,
context_size, num_blocks_eff);
r = ggml_cont(ctx, ggml_permute(ctx, r, 0, 2, 1, 3));
return r;
};
q = reshape_qkv(q);
k = reshape_qkv(k);
v = reshape_qkv(v);
// ----- Shaw positional bias -----
// rel_pos_emb is [head_dim, 2*max_pos_emb+1]. attention_dists is
// [context_size, context_size] int32. Materialise the per-(c, r)
// lookup as a [head_dim, context_size*context_size] tensor, then
// reshape to [head_dim, r=context_size, c=context_size].
ggml_tensor * dists_flat = ggml_reshape_1d(
ctx, dists, static_cast<int64_t>(context_size) * context_size);
ggml_tensor * rel_lookup = ggml_get_rows(ctx, w.attn_rel_pos_emb,
dists_flat);
rel_lookup = ggml_reshape_3d(ctx, rel_lookup,
head_dim, context_size, context_size);
// pos_attn[h, b, c, r] = sum_d q[h, b, c, d] * rel_lookup[c, r, d].
// Permute q to put c on ne[2] for batched mul_mat against
// rel_lookup; rel_lookup's ne[3]=1 broadcasts across num_blocks.
ggml_tensor * q_perm = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3));
ggml_tensor * pos_attn = ggml_mul_mat(ctx, rel_lookup, q_perm);
// ----- QK^T -----
// ggml_mul_mat(K, Q) with K, Q both [head_dim, context_size,
// n_heads, num_blocks] -> [context_size=k, context_size=q,
// n_heads, num_blocks].
ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
// Align pos_attn's axis order to kq's before the add.
pos_attn = ggml_cont(ctx, ggml_permute(ctx, pos_attn, 0, 2, 1, 3));
// scores = (kq + pos_attn) * scale + pad_mask
ggml_tensor * scores = ggml_add(ctx, kq, pos_attn);
scores = ggml_scale(ctx, scores, scale);
// Broadcast the per-block additive mask across n_heads by giving it
// ne[2]=1 in a 4D view. pad_mask_3d is [ctx, ctx, num_blocks_eff] — the
// caller tiles the per-block pattern across the batch.
ggml_tensor * pad_mask_4d = ggml_reshape_4d(ctx, pad_mask_3d,
context_size, context_size,
1, num_blocks_eff);
scores = ggml_add(ctx, scores, pad_mask_4d);
// Softmax over the k axis (ne[0]).
ggml_tensor * attn = ggml_soft_max(ctx, scores);
// out = V^T @ attn → [head_dim, q, n_heads, num_blocks].
ggml_tensor * v_t = ggml_cont(ctx, ggml_permute(ctx, v, 1, 0, 2, 3));
ggml_tensor * out = ggml_mul_mat(ctx, v_t, attn);
// Reshape back to [inner_dim, T_pad, B] (B==1 collapses to the
// pre-batch [inner_dim, T_pad]).
out = ggml_cont(ctx, ggml_permute(ctx, out, 0, 2, 1, 3));
out = ggml_reshape_3d(ctx, out, inner_dim, T_pad, B);
// Slice off pad rows BEFORE out_proj. This mirrors the reference's
// `out = self.to_out(out[:, :num_features, :])` order; it keeps
// pad-row garbage from leaking through the post-attention depthwise
// conv.
if (T_pad > T_enc) {
out = ggml_view_3d(ctx, out, inner_dim, T_enc, B,
out->nb[1], out->nb[2], 0);
out = ggml_cont(ctx, out);
}
// Final out_proj (with bias). Operates at T_enc.
out = ggml_mul_mat(ctx, w.attn_out_w, out);
if (w.attn_out_b != nullptr) out = ggml_add(ctx, out, w.attn_out_b);
// out ne = [d_model, T_enc]
return out;
}
} // namespace transcribe::granite_conformer