Repository navigation
Expand file tree
/
Copy pathinstructsam_detr_encoder.cpp
More file actions
248 lines (214 loc) · 10.6 KB
/
Copy pathinstructsam_detr_encoder.cpp
File metadata and controls
248 lines (214 loc) · 10.6 KB
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
#include "sam3/instructsam_detr_encoder.h"
#include "ggml-alloc.h"
#include <cmath>
#include <cstring>
#include <limits>
#include <stdexcept>
#include <string>
#include <vector>
namespace sam3 {
namespace {
constexpr int32_t kModelDim = 256;
constexpr int32_t kHeads = 8;
constexpr int32_t kHeadDim = kModelDim / kHeads;
constexpr int32_t kLayers = 6;
constexpr int32_t kFfnDim = 2048;
constexpr float kLayerNormEps = 1e-5f;
ggml_tensor * require_tensor(const GgufModel & model, const std::string & name) {
ggml_tensor * t = model.find_weight(name);
if (t == nullptr) throw std::runtime_error("detr_encoder: missing tensor: " + name);
return t;
}
ggml_tensor * ensure_f32(ggml_context * ctx, ggml_tensor * t) {
if (t->type == GGML_TYPE_F32) return t;
return ggml_cast(ctx, t, GGML_TYPE_F32);
}
ggml_tensor * layer_norm(
ggml_context * ctx, ggml_tensor * x,
ggml_tensor * weight, ggml_tensor * bias
) {
ggml_tensor * y = ggml_norm(ctx, ensure_f32(ctx, x), kLayerNormEps);
y = ggml_mul(ctx, y, ensure_f32(ctx, weight));
y = ggml_add(ctx, y, ensure_f32(ctx, bias));
return y;
}
ggml_tensor * linear(
ggml_context * ctx,
ggml_tensor * weight, ggml_tensor * bias, ggml_tensor * x
) {
ggml_tensor * y = ggml_mul_mat(ctx, weight, x);
y = ensure_f32(ctx, y);
if (bias != nullptr) y = ggml_add(ctx, y, ensure_f32(ctx, bias));
return y;
}
ggml_tensor * mha(
ggml_context * ctx,
ggml_tensor * q, ggml_tensor * k, ggml_tensor * v,
int32_t q_len, int32_t kv_len, ggml_tensor * mask
) {
ggml_tensor * qh = ggml_permute(
ctx, ggml_cont_3d(ctx, q, kHeadDim, kHeads, q_len), 0, 2, 1, 3);
ggml_tensor * kh = ggml_permute(
ctx, ggml_cont_3d(ctx, k, kHeadDim, kHeads, kv_len), 0, 2, 1, 3);
ggml_tensor * vh = ggml_cont_3d(
ctx,
ggml_permute(ctx, ggml_cont_3d(ctx, v, kHeadDim, kHeads, kv_len), 1, 2, 0, 3),
kv_len, kHeadDim, kHeads);
ggml_tensor * scores = ensure_f32(ctx, ggml_mul_mat(ctx, kh, qh));
ggml_tensor * probs = ggml_soft_max_ext(
ctx, scores, mask,
1.0f / std::sqrt(static_cast<float>(kHeadDim)),
0.0f);
ggml_tensor * attn = ensure_f32(ctx, ggml_mul_mat(ctx, vh, probs));
ggml_tensor * merged = ggml_permute(ctx, attn, 0, 2, 1, 3);
return ggml_cont_2d(ctx, merged, kModelDim, q_len);
}
std::string layer_prefix(int layer) {
return "transformer.encoder.layers." + std::to_string(layer);
}
// Build one encoder layer's forward graph. PRE-NORM structure.
ggml_tensor * build_layer(
ggml_context * ctx, const GgufModel & model, int layer,
ggml_tensor * hidden, // [hw, kModelDim]
ggml_tensor * vision_pos, // [hw, kModelDim]
ggml_tensor * text_mem, // [text_seq, kModelDim]
int32_t hw, int32_t text_seq,
ggml_tensor * text_mask // [text_seq, hw] f16 additive mask
) {
const std::string p = layer_prefix(layer);
// ── Self-attention with vision_pos ───────────────────────────────────
ggml_tensor * normed1 = layer_norm(ctx, hidden,
require_tensor(model, p + ".layer_norm1.weight"),
require_tensor(model, p + ".layer_norm1.bias"));
ggml_tensor * qk = ggml_add(ctx, normed1, vision_pos);
ggml_tensor * self_out = mha(ctx,
linear(ctx, require_tensor(model, p + ".self_attn.q_proj.weight"),
require_tensor(model, p + ".self_attn.q_proj.bias"), qk),
linear(ctx, require_tensor(model, p + ".self_attn.k_proj.weight"),
require_tensor(model, p + ".self_attn.k_proj.bias"), qk),
linear(ctx, require_tensor(model, p + ".self_attn.v_proj.weight"),
require_tensor(model, p + ".self_attn.v_proj.bias"), normed1),
hw, hw, nullptr);
self_out = linear(ctx,
require_tensor(model, p + ".self_attn.o_proj.weight"),
require_tensor(model, p + ".self_attn.o_proj.bias"), self_out);
hidden = ggml_add(ctx, hidden, self_out);
// ── Cross-attention: vision queries attend to text ───────────────────
ggml_tensor * normed2 = layer_norm(ctx, hidden,
require_tensor(model, p + ".layer_norm2.weight"),
require_tensor(model, p + ".layer_norm2.bias"));
ggml_tensor * cross_out = mha(ctx,
linear(ctx, require_tensor(model, p + ".cross_attn.q_proj.weight"),
require_tensor(model, p + ".cross_attn.q_proj.bias"), normed2),
linear(ctx, require_tensor(model, p + ".cross_attn.k_proj.weight"),
require_tensor(model, p + ".cross_attn.k_proj.bias"), text_mem),
linear(ctx, require_tensor(model, p + ".cross_attn.v_proj.weight"),
require_tensor(model, p + ".cross_attn.v_proj.bias"), text_mem),
hw, text_seq, text_mask);
cross_out = linear(ctx,
require_tensor(model, p + ".cross_attn.o_proj.weight"),
require_tensor(model, p + ".cross_attn.o_proj.bias"), cross_out);
hidden = ggml_add(ctx, hidden, cross_out);
// ── MLP ──────────────────────────────────────────────────────────────
ggml_tensor * normed3 = layer_norm(ctx, hidden,
require_tensor(model, p + ".layer_norm3.weight"),
require_tensor(model, p + ".layer_norm3.bias"));
ggml_tensor * ff = linear(ctx,
require_tensor(model, p + ".mlp.fc1.weight"),
require_tensor(model, p + ".mlp.fc1.bias"), normed3);
ff = ggml_relu(ctx, ff);
ff = linear(ctx,
require_tensor(model, p + ".mlp.fc2.weight"),
require_tensor(model, p + ".mlp.fc2.bias"), ff);
hidden = ggml_add(ctx, hidden, ff);
return hidden;
}
} // namespace
InstructsamDetrEncoder::InstructsamDetrEncoder(const GgufModel & model) : model_(model) {}
DetrEncoderOutput InstructsamDetrEncoder::run(
const std::vector<float> & vision_features,
const std::vector<int64_t> & vision_shape,
const std::vector<float> & vision_pos,
const std::vector<int64_t> & vision_pos_shape,
const std::vector<float> & text_features,
const std::vector<int64_t> & text_shape,
const std::vector<float> & text_mask,
const std::vector<int64_t> & text_mask_shape
) const {
if (vision_shape.size() != 2 || vision_shape[1] != kModelDim)
throw std::runtime_error("encoder: vision must be [hw, 256]");
if (vision_pos_shape != vision_shape)
throw std::runtime_error("encoder: vision_pos shape must match vision");
if (text_shape.size() != 2 || text_shape[1] != kModelDim)
throw std::runtime_error("encoder: text must be [seq, 256]");
if (text_mask_shape.size() != 1 || text_mask_shape[0] != text_shape[0])
throw std::runtime_error("encoder: text_mask must be [seq]");
const int32_t hw = static_cast<int32_t>(vision_shape[0]);
const int32_t text_seq = static_cast<int32_t>(text_shape[0]);
// Text mask → f16 additive attention mask [text_seq, hw]
std::vector<ggml_fp16_t> text_mask_f16(static_cast<size_t>(text_seq * hw));
const float neg_inf = -std::numeric_limits<float>::infinity();
for (int32_t q = 0; q < hw; ++q) {
for (int32_t k = 0; k < text_seq; ++k) {
const float valid = text_mask[static_cast<size_t>(k)];
const float m = (valid > 0.5f) ? 0.0f : neg_inf;
text_mask_f16[static_cast<size_t>(q * text_seq + k)] = ggml_fp32_to_fp16(m);
}
}
ggml_backend_t backend = model_.backend();
ggml_backend_t cpu_backend = nullptr;
ggml_backend_t backends[2] = { backend, nullptr };
int n_backends = 1;
if (ggml_backend_dev_type(ggml_backend_get_device(backend)) != GGML_BACKEND_DEVICE_TYPE_CPU) {
cpu_backend = model_.cpu_backend();
if (cpu_backend != nullptr) { backends[1] = cpu_backend; n_backends = 2; }
}
DetrEncoderOutput out;
out.num_layers = kLayers;
out.vision_seq = hw;
out.hidden_dim = kModelDim;
out.hs.resize(static_cast<size_t>(kLayers));
std::vector<float> current_hidden = vision_features;
for (int layer = 0; layer < kLayers; ++layer) {
const size_t graph_size = 32768;
const size_t ctx_size = ggml_tensor_overhead() * graph_size +
ggml_graph_overhead_custom(graph_size, false);
std::vector<uint8_t> ctx_buf(ctx_size);
ggml_context * ctx = ggml_init({ctx_size, ctx_buf.data(), true});
if (!ctx) throw std::runtime_error("ggml_init failed");
ggml_backend_sched_t sched = ggml_backend_sched_new(backends, nullptr, n_backends, graph_size, false, true);
if (!sched) { ggml_free(ctx); throw std::runtime_error("sched_new failed"); }
ggml_cgraph * gf = ggml_new_graph_custom(ctx, graph_size, false);
ggml_tensor * hidden_t = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, kModelDim, hw);
ggml_tensor * vpos_t = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, kModelDim, hw);
ggml_tensor * text_t = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, kModelDim, text_seq);
ggml_tensor * tmask_t = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, text_seq, hw);
for (ggml_tensor * t : {hidden_t, vpos_t, text_t, tmask_t}) {
ggml_backend_sched_set_tensor_backend(sched, t, backend);
}
ggml_tensor * layer_out = build_layer(
ctx, model_, layer,
hidden_t, vpos_t, text_t,
hw, text_seq, tmask_t);
ggml_tensor * capture = ggml_cont(ctx, layer_out);
ggml_build_forward_expand(gf, capture);
ggml_backend_sched_alloc_graph(sched, gf);
ggml_backend_tensor_set(hidden_t, current_hidden.data(), 0, current_hidden.size() * sizeof(float));
ggml_backend_tensor_set(vpos_t, vision_pos.data(), 0, vision_pos.size() * sizeof(float));
ggml_backend_tensor_set(text_t, text_features.data(), 0, text_features.size() * sizeof(float));
ggml_backend_tensor_set(tmask_t, text_mask_f16.data(), 0, text_mask_f16.size() * sizeof(ggml_fp16_t));
if (ggml_backend_sched_graph_compute(sched, gf) != GGML_STATUS_SUCCESS) {
ggml_backend_sched_free(sched); ggml_free(ctx);
throw std::runtime_error("encoder layer compute failed");
}
out.hs[static_cast<size_t>(layer)].resize(static_cast<size_t>(hw * kModelDim));
ggml_backend_tensor_get(capture, out.hs[static_cast<size_t>(layer)].data(), 0,
out.hs[static_cast<size_t>(layer)].size() * sizeof(float));
current_hidden = out.hs[static_cast<size_t>(layer)];
ggml_backend_sched_free(sched);
ggml_free(ctx);
}
out.last_hidden_state = out.hs.back();
return out;
}
} // namespace sam3