Skip to content

Commit 0699a44

Browse files
ggerganovggerganov
committed
context : remove redundant virtual, protected -> private
ggml-ci
1 parent 6378112 commit 0699a44

File tree

4 files changed

+79
-66
lines changed

4 files changed

+79
-66
lines changed

‎src/llama-context.cpp‎

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1842,6 +1842,14 @@ ggml_tensor * llama_context::build_attn(
18421842
return cur;
18431843
}
18441844

1845+
ggml_tensor * llama_context::build_inp_self_k_shift(
1846+
ggml_context * ctx0) {
1847+
GGML_UNUSED(ctx0);
1848+
1849+
LLAMA_LOG_ERROR("%s: not implemented\n", __func__);
1850+
return nullptr;
1851+
}
1852+
18451853
void llama_context::build_kv_self_shift(
18461854
ggml_context * ctx0,
18471855
ggml_cgraph * gf) {

‎src/llama-context.h‎

Lines changed: 71 additions & 55 deletions
Original file line numberDiff line numberDiff line change
@@ -140,6 +140,7 @@ struct llama_context : public llama_graph_i {
140140

141141
virtual void input_set(const llama_ubatch & ubatch);
142142

143+
private:
143144
struct {
144145
// base input tensors
145146
ggml_tensor * tokens; // I32 [n_batch]
@@ -155,6 +156,7 @@ struct llama_context : public llama_graph_i {
155156
ggml_tensor * kq_mask_cnv; // [n_tokens, n_batch]
156157
} inp;
157158

159+
protected:
158160
//
159161
// output
160162
//
@@ -192,71 +194,71 @@ struct llama_context : public llama_graph_i {
192194
// graph build
193195
//
194196

195-
virtual void build_cb(
197+
void build_cb(
196198
ggml_tensor * cur,
197199
const char * name,
198200
const llama_ubatch & ubatch,
199201
int il) override;
200202

201203
// apply control vector for layer il
202-
virtual ggml_tensor * build_cvec(
204+
ggml_tensor * build_cvec(
203205
ggml_context * ctx0,
204206
ggml_tensor * cur,
205207
int il) override;
206208

207209
// do mat_mul, while optionally apply lora
208-
virtual ggml_tensor * build_lora_mm(
210+
ggml_tensor * build_lora_mm(
209211
ggml_context * ctx0,
210212
ggml_tensor * w,
211213
ggml_tensor * cur) override;
212214

213215
// do mat_mul_id, while optionally apply lora
214-
virtual ggml_tensor * build_lora_mm_id(
216+
ggml_tensor * build_lora_mm_id(
215217
ggml_context * ctx0,
216218
ggml_tensor * w, // struct ggml_tensor * as
217219
ggml_tensor * cur, // struct ggml_tensor * b
218220
ggml_tensor * ids) override;
219221

220-
virtual ggml_tensor * build_rope_factors(int il) override;
222+
ggml_tensor * build_rope_factors(int il) override;
221223

222-
virtual ggml_tensor * build_rope_shift(
224+
ggml_tensor * build_rope_shift(
223225
ggml_context * ctx0,
224226
ggml_tensor * cur,
225227
ggml_tensor * shift,
226228
ggml_tensor * factors,
227229
ggml_backend_buffer * bbuf) override;
228230

229-
virtual ggml_tensor * build_inp_embd(
231+
ggml_tensor * build_inp_embd(
230232
ggml_context * ctx0,
231233
ggml_tensor * tok_embd,
232234
const llama_ubatch & ubatch) override;
233235

234-
virtual ggml_tensor * build_inp_pos(
236+
ggml_tensor * build_inp_pos(
235237
ggml_context * ctx0,
236238
int32_t n_tokens) override;
237239

238-
virtual ggml_tensor * build_inp_pos_bucket(
240+
ggml_tensor * build_inp_pos_bucket(
239241
ggml_context * ctx0,
240242
int32_t n_tokens) override;
241243

242-
virtual ggml_tensor * build_inp_out_ids(
244+
ggml_tensor * build_inp_out_ids(
243245
ggml_context * ctx0) override;
244246

245-
virtual ggml_tensor * build_inp_mean(
247+
ggml_tensor * build_inp_mean(
246248
ggml_context * ctx0,
247249
int32_t n_tokens) override;
248250

249-
virtual ggml_tensor * build_inp_cls(
251+
ggml_tensor * build_inp_cls(
250252
ggml_context * ctx0,
251253
int32_t n_tokens) override;
252254

253-
virtual void build_attn_inp(
255+
void build_attn_inp(
254256
ggml_context * ctx0,
255257
int32_t n_tokens,
256258
bool causal,
257259
bool swa) override;
258260

259-
virtual ggml_tensor * build_attn(
261+
ggml_tensor * build_attn(
260262
ggml_context * ctx0,
261263
ggml_cgraph * gf,
262264
ggml_tensor * wo,
@@ -270,6 +272,9 @@ struct llama_context : public llama_graph_i {
270272
int il) override;
271273

272274
protected:
275+
virtual ggml_tensor * build_inp_self_k_shift(
276+
ggml_context * ctx0);
277+
273278
virtual void build_kv_self_shift(
274279
ggml_context * ctx0,
275280
ggml_cgraph * gf);
@@ -288,6 +293,7 @@ struct llama_context : public llama_graph_i {
288293
virtual void perf_reset();
289294

290295
protected:
296+
// TODO: become private
291297
mutable int64_t t_start_us = 0;
292298
mutable int64_t t_load_us = 0;
293299
mutable int64_t t_p_eval_us = 0;
@@ -346,6 +352,7 @@ struct llama_context : public llama_graph_i {
346352
//
347353
// members
348354
//
355+
// TODO: become private / move to llama_graph_i
349356

350357
const llama_model & model;
351358

@@ -412,24 +419,25 @@ class llama_context_kv_self : public llama_context {
412419
virtual ~llama_context_kv_self();
413420

414421
protected:
415-
virtual void reserve() override;
422+
void reserve() override;
416423

417424
public:
418-
virtual llama_kv_cache * get_kv_self() override;
419-
virtual const llama_kv_cache * get_kv_self() const override;
425+
llama_kv_cache * get_kv_self() override;
426+
const llama_kv_cache * get_kv_self() const override;
420427

421-
virtual void kv_self_update() override;
428+
void kv_self_update() override;
422429

423-
virtual int encode(llama_batch & inp_batch) override;
424-
virtual int decode(llama_batch & inp_batch) override;
430+
int encode(llama_batch & inp_batch) override;
431+
int decode(llama_batch & inp_batch) override;
425432

426433
protected:
427434
//
428435
// input
429436
//
430437

431-
virtual void input_set(const llama_ubatch & ubatch) override;
438+
void input_set(const llama_ubatch & ubatch) override;
432439

440+
private:
433441
struct {
434442
ggml_tensor * self_pos_bucket; // I32 [n_kv, n_batch]
435443
ggml_tensor * self_kq_mask; // F32 [n_kv, n_batch]
@@ -443,26 +451,24 @@ class llama_context_kv_self : public llama_context {
443451
// graph
444452
//
445453

446-
virtual ggml_cgraph * graph_init() override;
454+
ggml_cgraph * graph_init() override;
447455

448456
public:
449457
//
450458
// graph build
451459
//
452460

453-
virtual ggml_tensor * build_inp_self_k_shift(ggml_context * ctx0) override;
454-
455-
virtual ggml_tensor * build_inp_pos_bucket(
461+
ggml_tensor * build_inp_pos_bucket(
456462
ggml_context * ctx0,
457463
int32_t n_tokens) override;
458464

459-
virtual void build_attn_inp(
465+
void build_attn_inp(
460466
ggml_context * ctx0,
461467
int32_t n_tokens,
462468
bool causal,
463469
bool swa) override;
464470

465-
virtual ggml_tensor * build_attn(
471+
ggml_tensor * build_attn(
466472
ggml_context * ctx0,
467473
ggml_cgraph * gf,
468474
ggml_tensor * wo,
@@ -476,16 +482,22 @@ class llama_context_kv_self : public llama_context {
476482
int il) override;
477483

478484
protected:
479-
virtual void build_kv_self_shift(
485+
ggml_tensor * build_inp_self_k_shift(ggml_context * ctx0) override;
486+
487+
void build_kv_self_shift(
480488
ggml_context * ctx0,
481489
ggml_cgraph * gf) override;
482490

483491
// find holes from the beginning of the KV cache and fill them by moving data from the end of the cache
484-
virtual void build_kv_self_defrag(
492+
void build_kv_self_defrag(
485493
ggml_context * ctx0,
486494
ggml_cgraph * gf) override;
487495

496+
// =======================================================
488497
// === encoder-decoder ===
498+
//
499+
// TODO: this is temporary here, it will be moved
500+
//
489501

490502
// whether we are computing encoder output or decoder output
491503
bool is_encoding = false;
@@ -497,23 +509,25 @@ class llama_context_kv_self : public llama_context {
497509
struct ggml_tensor * inp_embd_enc; // F32 [n_embd, n_outputs_enc]
498510
struct ggml_tensor * inp_kq_mask_cross; // F32 [n_outputs_enc, n_batch]
499511

500-
virtual ggml_tensor * build_inp_embd_enc(
512+
ggml_tensor * build_inp_embd_enc(
501513
ggml_context * ctx0) override;
502514

503-
virtual ggml_tensor * build_inp_kq_mask_cross(
515+
ggml_tensor * build_inp_kq_mask_cross(
504516
ggml_context * ctx0,
505517
int32_t n_tokens) override;
518+
// ======================================================
506519

507520
//
508521
// state save/load
509522
//
510523

511-
virtual size_t state_get_data(llama_io_write_i & io) override;
512-
virtual size_t state_set_data(llama_io_read_i & io) override;
524+
size_t state_get_data(llama_io_write_i & io) override;
525+
size_t state_set_data(llama_io_read_i & io) override;
513526

514-
virtual size_t state_seq_get_data(llama_io_write_i & io, llama_seq_id seq_id) override;
515-
virtual size_t state_seq_set_data(llama_io_read_i & io, llama_seq_id seq_id) override;
527+
size_t state_seq_get_data(llama_io_write_i & io, llama_seq_id seq_id) override;
528+
size_t state_seq_set_data(llama_io_read_i & io, llama_seq_id seq_id) override;
516529

530+
private:
517531
//
518532
// members
519533
//
@@ -532,24 +546,25 @@ class llama_context_recurrent : public llama_context {
532546
virtual ~llama_context_recurrent();
533547

534548
protected:
535-
virtual void reserve() override;
549+
void reserve() override;
536550

537551
public:
538-
virtual llama_kv_cache * get_kv_self() override;
539-
virtual const llama_kv_cache * get_kv_self() const override;
552+
llama_kv_cache * get_kv_self() override;
553+
const llama_kv_cache * get_kv_self() const override;
540554

541-
virtual void kv_self_update() override;
555+
void kv_self_update() override;
542556

543-
virtual int encode(llama_batch & inp_batch) override;
544-
virtual int decode(llama_batch & inp_batch) override;
557+
int encode(llama_batch & inp_batch) override;
558+
int decode(llama_batch & inp_batch) override;
545559

546560
protected:
547561
//
548562
// input
549563
//
550564

551-
virtual void input_set(const llama_ubatch & ubatch) override;
565+
void input_set(const llama_ubatch & ubatch) override;
552566

567+
private:
553568
struct {
554569
ggml_tensor * s_copy; // I32 [kv_size]
555570
ggml_tensor * s_mask; // F32 [1, n_kv]
@@ -559,20 +574,20 @@ class llama_context_recurrent : public llama_context {
559574
// graph
560575
//
561576

562-
virtual ggml_cgraph * graph_init() override;
577+
ggml_cgraph * graph_init() override;
563578

564579
public:
565580
//
566581
// graph build
567582
//
568583

569-
virtual ggml_tensor * build_inp_s_copy(
584+
ggml_tensor * build_inp_s_copy(
570585
ggml_context * ctx0) override;
571586

572-
virtual ggml_tensor * build_inp_s_mask(
587+
ggml_tensor * build_inp_s_mask(
573588
ggml_context * ctx0) override;
574589

575-
virtual ggml_tensor * build_copy_mask_state(
590+
ggml_tensor * build_copy_mask_state(
576591
ggml_context * ctx0,
577592
ggml_cgraph * gf,
578593
ggml_tensor * s,
@@ -581,7 +596,7 @@ class llama_context_recurrent : public llama_context {
581596
int32_t n_state,
582597
int32_t n_seqs) override;
583598

584-
virtual ggml_tensor * build_mamba_layer(
599+
ggml_tensor * build_mamba_layer(
585600
ggml_context * ctx0,
586601
ggml_cgraph * gf,
587602
ggml_tensor * cur,
@@ -590,21 +605,21 @@ class llama_context_recurrent : public llama_context {
590605
const llama_ubatch & ubatch,
591606
int il) override;
592607

593-
virtual ggml_tensor * build_rwkv_token_shift_load(
608+
ggml_tensor * build_rwkv_token_shift_load(
594609
ggml_context * ctx0,
595610
ggml_cgraph * gf,
596611
ggml_tensor * state_copy,
597612
ggml_tensor * state_mask,
598613
const llama_ubatch & ubatch,
599614
int il) override;
600615

601-
virtual ggml_tensor * build_rwkv_token_shift_store(
616+
ggml_tensor * build_rwkv_token_shift_store(
602617
ggml_context * ctx0,
603618
ggml_tensor * token_shift,
604619
const llama_ubatch & ubatch,
605620
int il) override;
606621

607-
virtual ggml_tensor * build_rwkv6_time_mix(
622+
ggml_tensor * build_rwkv6_time_mix(
608623
ggml_context * ctx0,
609624
ggml_cgraph * gf,
610625
ggml_tensor * cur,
@@ -619,12 +634,13 @@ class llama_context_recurrent : public llama_context {
619634
// state save/load
620635
//
621636

622-
virtual size_t state_get_data(llama_io_write_i & io) override;
623-
virtual size_t state_set_data(llama_io_read_i & io) override;
637+
size_t state_get_data(llama_io_write_i & io) override;
638+
size_t state_set_data(llama_io_read_i & io) override;
624639

625-
virtual size_t state_seq_get_data(llama_io_write_i & io, llama_seq_id seq_id) override;
626-
virtual size_t state_seq_set_data(llama_io_read_i & io, llama_seq_id seq_id) override;
640+
size_t state_seq_get_data(llama_io_write_i & io, llama_seq_id seq_id) override;
641+
size_t state_seq_set_data(llama_io_read_i & io, llama_seq_id seq_id) override;
627642

643+
private:
628644
//
629645
// members
630646
//
@@ -646,7 +662,7 @@ class llama_context_enc_dec : public llama_context {
646662

647663
virtual ~llama_context_enc_dec();
648664

649-
protected:
665+
private:
650666
llama_context_kv_self ctx_dec;
651667
};
652668

‎src/llama-graph.cpp‎

Lines changed: 0 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -32,14 +32,6 @@ ggml_tensor * llama_graph_i::build_attn(
3232
return nullptr;
3333
}
3434

35-
ggml_tensor * llama_graph_i::build_inp_self_k_shift(
36-
ggml_context * ctx0) {
37-
GGML_UNUSED(ctx0);
38-
39-
LLAMA_LOG_ERROR("%s: not implemented\n", __func__);
40-
return nullptr;
41-
}
42-
4335
ggml_tensor * llama_graph_i::build_inp_embd_enc(
4436
ggml_context * ctx0) {
4537
GGML_UNUSED(ctx0);

0 commit comments

Comments
 (0)