Skip to content

Commit 4f02d47

Browse files
CISCCISC
authored
model : refactor bias tensor variable names (ggml-org#22079)
* refactor bias tensor variable names * use create_tensor_qkv for jina-bert-v2
1 parent 23b8cc4 commit 4f02d47

55 files changed

Lines changed: 104 additions & 117 deletions

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

‎src/llama-graph.cpp‎

Lines changed: 9 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -1077,9 +1077,9 @@ llm_graph_qkv llm_graph_context::build_qkv(
10771077
// fused QKV path
10781078
ggml_tensor * qkv = build_lora_mm(layer.wqkv, cur, layer.wqkv_s);
10791079
cb(qkv, "wqkv", il);
1080-
if (layer.bqkv) {
1081-
qkv = ggml_add(ctx0, qkv, layer.bqkv);
1082-
cb(qkv, "bqkv", il);
1080+
if (layer.wqkv_b) {
1081+
qkv = ggml_add(ctx0, qkv, layer.wqkv_b);
1082+
cb(qkv, "wqkv_b", il);
10831083
}
10841084
if (hparams.f_clamp_kqv > 0.0f) {
10851085
qkv = ggml_clamp(ctx0, qkv, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
@@ -1097,8 +1097,8 @@ llm_graph_qkv llm_graph_context::build_qkv(
10971097
// separate Q/K/V path
10981098
Qcur = build_lora_mm(layer.wq, cur, layer.wq_s);
10991099
cb(Qcur, "Qcur", il);
1100-
if (layer.bq) {
1101-
Qcur = ggml_add(ctx0, Qcur, layer.bq);
1100+
if (layer.wq_b) {
1101+
Qcur = ggml_add(ctx0, Qcur, layer.wq_b);
11021102
cb(Qcur, "Qcur", il);
11031103
}
11041104
if (hparams.f_clamp_kqv > 0.0f) {
@@ -1107,8 +1107,8 @@ llm_graph_qkv llm_graph_context::build_qkv(
11071107
}
11081108
Kcur = build_lora_mm(layer.wk, cur, layer.wk_s);
11091109
cb(Kcur, "Kcur", il);
1110-
if (layer.bk) {
1111-
Kcur = ggml_add(ctx0, Kcur, layer.bk);
1110+
if (layer.wk_b) {
1111+
Kcur = ggml_add(ctx0, Kcur, layer.wk_b);
11121112
cb(Kcur, "Kcur", il);
11131113
}
11141114
if (hparams.f_clamp_kqv > 0.0f) {
@@ -1117,8 +1117,8 @@ llm_graph_qkv llm_graph_context::build_qkv(
11171117
}
11181118
Vcur = build_lora_mm(layer.wv, cur, layer.wv_s);
11191119
cb(Vcur, "Vcur", il);
1120-
if (layer.bv) {
1121-
Vcur = ggml_add(ctx0, Vcur, layer.bv);
1120+
if (layer.wv_b) {
1121+
Vcur = ggml_add(ctx0, Vcur, layer.wv_b);
11221122
cb(Vcur, "Vcur", il);
11231123
}
11241124
if (hparams.f_clamp_kqv > 0.0f) {

‎src/llama-model.cpp‎

Lines changed: 40 additions & 48 deletions
Large diffs are not rendered by default.

‎src/llama-model.h‎

Lines changed: 2 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -246,6 +246,8 @@ struct llama_layer {
246246
struct ggml_tensor * wkv_b = nullptr;
247247
struct ggml_tensor * wk_b = nullptr;
248248
struct ggml_tensor * wv_b = nullptr;
249+
struct ggml_tensor * wqkv_b = nullptr;
250+
struct ggml_tensor * wo_b = nullptr;
249251
struct ggml_tensor * wq_cross = nullptr;
250252
struct ggml_tensor * wk_cross = nullptr;
251253
struct ggml_tensor * wv_cross = nullptr;
@@ -256,13 +258,6 @@ struct llama_layer {
256258
struct ggml_tensor * wo_enc = nullptr;
257259
struct ggml_tensor * wqkv_gate = nullptr;
258260

259-
// attention bias
260-
struct ggml_tensor * bq = nullptr;
261-
struct ggml_tensor * bk = nullptr;
262-
struct ggml_tensor * bv = nullptr;
263-
struct ggml_tensor * bo = nullptr;
264-
struct ggml_tensor * bqkv = nullptr;
265-
266261
// relative position bias
267262
struct ggml_tensor * attn_rel_b = nullptr;
268263
struct ggml_tensor * attn_rel_b_enc = nullptr;

‎src/models/apertus.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -50,7 +50,7 @@ llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_
5050
cb(Vcur, "Vcur_pos", il);
5151

5252
cur = build_attn(inp_attn,
53-
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
53+
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
5454
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
5555
cb(cur, "attn_out", il);
5656
}

‎src/models/arcee.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -55,7 +55,7 @@ llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_para
5555
cb(Vcur, "Vcur", il);
5656

5757
cur = build_attn(inp_attn,
58-
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
58+
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
5959
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
6060
cb(cur, "attn_out", il);
6161
}

‎src/models/bailingmoe.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -48,7 +48,7 @@ llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_
4848
cb(Vcur, "Vcur", il);
4949

5050
cur = build_attn(inp_attn,
51-
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
51+
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
5252
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
5353
}
5454

‎src/models/bailingmoe2.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -48,7 +48,7 @@ llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const ll
4848
cb(Vcur, "Vcur", il);
4949

5050
cur = build_attn(inp_attn,
51-
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
51+
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
5252
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
5353
}
5454

‎src/models/bert.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -72,7 +72,7 @@ llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params
7272
cb(Vcur, "Vcur", il);
7373

7474
cur = build_attn(inp_attn,
75-
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
75+
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
7676
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
7777
cb(cur, "kqv_out", il);
7878
}

‎src/models/bitnet.cpp‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -57,8 +57,8 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
5757
cb(cur, "attn_sub_norm", il);
5858

5959
cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
60-
if (model.layers[il].bo) {
61-
cur = ggml_add(ctx0, cur, model.layers[il].bo);
60+
if (model.layers[il].wo_b) {
61+
cur = ggml_add(ctx0, cur, model.layers[il].wo_b);
6262
}
6363
cb(cur, "attn_out", il);
6464
}

‎src/models/bloom.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -33,7 +33,7 @@ llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_para
3333
n_embd_head, n_head, n_head_kv, il);
3434

3535
cur = build_attn(inp_attn,
36-
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
36+
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
3737
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
3838
}
3939

0 commit comments

Comments
 (0)