llama.cpp/tools/server/server-context.cpp at master · rioscode/llama.cpp · GitHub

History

5558 lines (4576 loc) · 221 KB

Raw

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

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

307

308

309

310

311

312

313

314

315

316

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

336

337

338

339

340

341

342

343

344

345

346

347

348

349

350

351

352

353

354

355

356

357

358

359

360

361

362

363

364

365

366

367

368

369

370

371

372

373

374

375

376

377

378

379

380

381

382

383

384

385

386

387

388

389

390

391

392

393

394

395

396

397

398

399

400

401

402

403

404

405

406

407

408

409

410

411

412

413

414

415

416

417

418

419

420

421

422

423

424

425

426

427

428

429

430

431

432

433

434

435

436

437

438

439

440

441

442

443

444

445

446

447

448

449

450

451

452

453

454

455

456

457

458

459

460

461

462

463

464

465

466

467

468

469

470

471

472

473

474

475

476

477

478

479

480

481

482

483

484

485

486

487

488

489

490

491

492

493

494

495

496

497

498

499

500

501

502

503

504

505

506

507

508

509

510

511

512

513

514

515

516

517

518

519

520

521

522

523

524

525

526

527

528

529

530

531

532

533

534

535

536

537

538

539

540

541

542

543

544

545

546

547

548

549

550

551

552

553

554

555

556

557

558

559

560

561

562

563

564

565

566

567

568

569

570

571

572

573

574

575

576

577

578

579

580

581

582

583

584

585

586

587

588

589

590

591

592

593

594

595

596

597

598

599

600

601

602

603

604

605

606

607

608

609

610

611

612

613

614

615

616

617

618

619

620

621

622

623

624

625

626

627

628

629

630

631

632

633

634

635

636

637

638

639

640

641

642

643

644

645

646

647

648

649

650

651

652

653

654

655

656

657

658

659

660

661

662

663

664

665

666

667

668

669

670

671

672

673

674

675

676

677

678

679

680

681

682

683

684

685

686

687

688

689

690

691

692

693

694

695

696

697

698

699

700

701

702

703

704

705

706

707

708

709

710

711

712

713

714

715

716

717

718

719

720

721

722

723

724

725

726

727

728

729

730

731

732

733

734

735

736

737

738

739

740

741

742

743

744

745

746

747

748

749

750

751

752

753

754

755

756

757

758

759

760

761

762

763

764

765

766

767

768

769

770

771

772

773

774

775

776

777

778

779

780

781

782

783

784

785

786

787

788

789

790

791

792

793

794

795

796

797

798

799

800

801

802

803

804

805

806

807

808

809

810

811

812

813

814

815

816

817

818

819

820

821

822

823

824

825

826

827

828

829

830

831

832

833

834

835

836

837

838

839

840

841

842

843

844

845

846

847

848

849

850

851

852

853

854

855

856

857

858

859

860

861

862

863

864

865

866

867

868

869

870

871

872

873

874

875

876

877

878

879

880

881

882

883

884

885

886

887

888

889

890

891

892

893

894

895

896

897

898

899

900

901

902

903

904

905

906

907

908

909

910

911

912

913

914

915

916

917

918

919

920

921

922

923

924

925

926

927

928

929

930

931

932

933

934

935

936

937

938

939

940

941

942

943

944

945

946

947

948

949

950

951

952

953

954

955

956

957

958

959

960

961

962

963

964

965

966

967

968

969

970

971

972

973

974

975

976

977

978

979

980

981

982

983

984

985

986

987

988

989

990

991

992

993

994

995

996

997

998

999

1000

#include "server-context.h"

#include "server-chat.h"

#include "server-common.h"

#include "server-http.h"

#include "server-task.h"

#include "server-queue.h"

#include "server-schema.h"

#include "server-stream.h"

#include "build-info.h"

#include "common.h"

#include "fit.h"

#include "llama.h"

#include "log.h"

#include "sampling.h"

#include "speculative.h"

#include "mtmd.h"

#include "mtmd-helper.h"

#include <algorithm>

#include <cstddef>

#include <cinttypes>

#include <exception>

#include <memory>

#include <filesystem>

#include <random>

#include <utility>

#include <fstream>

// fix problem with std::min and std::max

#if defined(_WIN32)

#define WIN32_LEAN_AND_MEAN

#ifndef NOMINMAX

# define NOMINMAX

#endif

#include <windows.h>

#endif

constexpr int HTTP_POLLING_SECONDS = 1;

static common_speculative_output_limits server_output_limits(const common_params & params) {

if (params.embedding ||

(params.pooling_type != LLAMA_POOLING_TYPE_UNSPECIFIED && params.pooling_type != LLAMA_POOLING_TYPE_NONE)) {

return { params.n_batch, 1 };

}

auto result = common_speculative_get_output_limits(

params.n_batch, params.n_parallel, common_speculative_n_max(&params.speculative));

result.total = std::max<int32_t>(1, result.total);

result.per_seq = std::max<int32_t>(1, result.per_seq);

return result;

}

// synthetic draft verification for benchmarking - accept draft tokens at random instead of by match with the target

// on replay the draft was already accepted before a context checkpoint restore, so repeat the same decisions

static std::vector<llama_token> server_sample_and_accept_synth(

common_sampler * smpl,

llama_context * ctx,

const std::vector<int32_t> & idxs,

const llama_tokens & draft,

const std::vector<double> & synth_probs,

std::mt19937 & rng,

bool is_replay) {

GGML_ASSERT(idxs.size() == draft.size() + 1);

GGML_ASSERT(synth_probs.size() >= draft.size());

std::vector<llama_token> result;

result.reserve(idxs.size());

const llama_vocab * vocab = llama_model_get_vocab(llama_get_model(ctx));

std::uniform_real_distribution<double> dist(0.0, 1.0);

for (size_t i = 0; i < draft.size(); ++i) {

const llama_token id = common_sampler_sample(smpl, ctx, idxs[i]);

const bool accept = is_replay || dist(rng) < synth_probs[i];

// do not accept a drafted EOG token - it would end the generation early

// on replay the last token is from the target and can be EOG, so skip this check

if (accept && (is_replay || !llama_vocab_is_eog(vocab, draft[i]))) {

// synthetic draft tokens do not advance grammar or reasoning state

// the last replay token is from the target and must advance both

const bool is_replay_target = is_replay && i + 1 == draft.size();

common_sampler_accept(smpl, draft[i], is_replay_target);

result.push_back(draft[i]);

continue;

}

common_sampler_accept(smpl, id, true);

result.push_back(id);

return result;

}

const llama_token id = common_sampler_sample(smpl, ctx, idxs[draft.size()]);

common_sampler_accept(smpl, id, true);

result.push_back(id);

return result;

}

// state diagram: https://github.com/ggml-org/llama.cpp/pull/9283

enum slot_state {

SLOT_STATE_IDLE,

SLOT_STATE_WAIT_OTHER, // after assigning a task, but waiting for parent slot to process prompt

SLOT_STATE_STARTED, // after assigning a task and about to process prompt

SLOT_STATE_PROCESSING_PROMPT,

SLOT_STATE_DONE_PROMPT,

SLOT_STATE_GENERATING,

};

struct server_slot; // forward declaration

struct server_batch {

llama_batch batch;

bool batch_rendered = false;

struct token {

int32_t id_slot;

llama_token token;

llama_pos pos;

bool output;

bool is_prompt; // for stats tracking

};

std::vector<token> tokens;

int32_t n_tokens_alloc = 0;

int32_t n_embd = 0;

// track if given slot can be batched with slots already in the batch

server_slot * slot_batched = nullptr;

// in embd mode, we temporarily swap out the tokens arr and restore it on clear()

bool has_embd = false;

llama_token * tokens_ptr = nullptr;

std::vector<float> embd;

float alora_scale = -1.0f;

size_t alora_disabled_id = 0;

server_batch() {

batch.pos = nullptr; // sentinel: uninitialized batch

}

~server_batch() {

if (batch.pos != nullptr) {

clear();

llama_batch_free(batch);

}

void init(int32_t n_tokens_alloc, int32_t n_embd) {

this->n_tokens_alloc = n_tokens_alloc;

this->n_embd = n_embd;

batch = llama_batch_init(n_tokens_alloc, 0, 1);

tokens_ptr = batch.token;

tokens.reserve(n_tokens_alloc);

}

bool add(int32_t id_slot, llama_token token, llama_pos pos, bool output, bool is_prompt) {

GGML_ASSERT(!has_embd); // cannot mix tokens + embd in same batch

GGML_ASSERT(batch.pos != nullptr);

if ((int32_t)tokens.size() >= n_tokens_alloc) {

return false;

}

tokens.push_back({ id_slot, token, pos, output, is_prompt });

return true;

}

bool add(int32_t id_slot, const std::vector<float> & embd_in, llama_pos pos, bool output, bool is_prompt) {

GGML_ASSERT(batch.pos != nullptr);

if ((int32_t)tokens.size() >= n_tokens_alloc) {

return false;

}

tokens.push_back({ id_slot, LLAMA_TOKEN_NULL, pos, output, is_prompt });

has_embd = true;

embd.insert(embd.end(), embd_in.begin(), embd_in.end());

return true;

}

void clear() {

tokens.clear();

embd.clear();

common_batch_clear(batch);

slot_batched = nullptr;

alora_scale = -1.0f;

alora_disabled_id = 0;

batch_rendered = false;

has_embd = false;

if (batch.token == nullptr) {

batch.token = tokens_ptr;

batch.embd = nullptr;

}

int32_t size() const {

return (int32_t)tokens.size();

}

void set_output(int32_t idx, bool output) {

GGML_ASSERT(idx >= 0 && idx < (int32_t)tokens.size());

tokens[idx].output = output;

}

void render() {

GGML_ASSERT(!batch_rendered);

GGML_ASSERT(batch.pos != nullptr);

common_batch_clear(batch);

for (int32_t i = 0; i < size(); i++) {

const auto & t = tokens[i];

common_batch_add(batch, t.token, t.pos, { t.id_slot }, t.output);

}

if (has_embd) {

batch.token = nullptr; // will be restored on clear()

batch.embd = embd.data();

}

batch_rendered = true;

}

llama_batch get_view(int32_t off, int32_t n_tokens) const {

GGML_ASSERT(batch.pos != nullptr);

GGML_ASSERT(batch_rendered);

GGML_ASSERT(off >= 0 && off < size());

GGML_ASSERT(n_tokens > 0 && off + n_tokens <= size());

auto * token = batch.token ? batch.token + off : nullptr;

auto * embd = batch.embd ? batch.embd + off * n_embd : nullptr;

llama_batch view = {

n_tokens,

token,

embd,

batch.pos + off,

batch.n_seq_id + off,

batch.seq_id + off,

batch.logits + off,

};

return view;

}

};

struct server_slot {

int id;

llama_context * ctx_tgt = nullptr;

llama_context * ctx_dft = nullptr;

common_memory mem;

// multimodal

mtmd_context * mctx = nullptr;

mtmd::batch_ptr mbatch = nullptr;

// speculative decoding

common_speculative * spec;

llama_tokens spec_draft;

llama_tokens spec_prompt;

std::vector<int32_t> spec_i_batch;

common_prompt_checkpoint spec_ckpt;

bool spec_is_replay = false;

std::mt19937 spec_synth_rng;

// TODO: move members that belong to the task (such as `generated_text`, `has_new_line`) to task_results_state

// see https://github.com/ggml-org/llama.cpp/pull/18283#issuecomment-3710175837

std::unique_ptr<const server_task> task;

std::unique_ptr<const server_task> task_prev; // used for debugging

// used to determine the slot that has been used the longest

int64_t t_last_used = -1;

// generation props

int32_t n_ctx = 0; // context size per slot

int32_t n_keep = 0;

int32_t i_batch = -1;

// effective generation limit for the current task, -1 means unlimited

int32_t n_predict_max = -1;

size_t last_nl_pos = 0;

std::string generated_text;

std::string debug_generated_text;

llama_tokens generated_tokens;

size_t n_sent_text = 0; // number of sent text character (i.e. handle partial UTF-8 on streaming)

std::vector<completion_token_output> generated_token_probs;

bool has_next_token = true;

bool has_new_line = false;

bool truncated = false;

stop_type stop;

std::string stopping_word;

// state

slot_state state = SLOT_STATE_IDLE;

server_prompt prompt;

bool prompt_save(server_prompt_cache & prompt_cache) const {

if (prompt.tokens.size() == 0) {

return false;

}

const size_t cur_size_tgt = llama_state_seq_get_size_ext(ctx_tgt, id, LLAMA_STATE_SEQ_FLAGS_NONE);

const size_t cur_size_dft = ctx_dft ? llama_state_seq_get_size_ext(ctx_dft, id, LLAMA_STATE_SEQ_FLAGS_NONE) : 0;

const size_t cur_size = cur_size_tgt + cur_size_dft;

SRV_TRC(" - saving prompt with length %d, total state size = %.3f MiB (draft: %.3f MiB)\n",

(int) prompt.tokens.size(), cur_size / (1024.0 * 1024.0), cur_size_dft / (1024.0 * 1024.0));

auto * cur = prompt_cache.alloc(prompt, cur_size_tgt, cur_size_dft);

if (cur == nullptr) {

return false;

}

llama_state_seq_get_data_ext(ctx_tgt, cur->data.main.data(), cur_size_tgt, id, LLAMA_STATE_SEQ_FLAGS_NONE);

if (ctx_dft) {

llama_state_seq_get_data_ext(ctx_dft, cur->data.drft.data(), cur_size_dft, id, LLAMA_STATE_SEQ_FLAGS_NONE);

}

return true;

}

bool prompt_load(server_prompt_cache & prompt_cache, const server_tokens & tokens) {

bool res = prompt_cache.load(prompt, tokens, ctx_tgt, ctx_dft, id);

if (!res) {

SLT_WRN(*this, "%s", "failed to load prompt from cache\n");

}

return res;

}

void prompt_clear() {

SLT_TRC(*this, "clearing prompt with %zu tokens\n", prompt.tokens.size());

mem.seq_rm(id, -1, -1);

prompt.clear();

}

std::vector<common_adapter_lora_info> lora;

int32_t alora_invocation_start = -1;

// sampling

json json_schema;

common_sampler_ptr smpl;

llama_token sampled; // in speculative mode, this is the last accepted token

// for TTS models, this is the embd generated from prev step, decode this to generate next hidden state

// corresponding to one token position (size = n_embd)

std::vector<float> inp_embd;

server_slot_stats stats;

// accepted tokens per draft position

// not in server_slot_stats to avoid copying to every task result

std::vector<uint64_t> n_accepted_per_pos;

std::function<void(int /* id_slot */)> callback_on_release;

std::function<void(const server_slot &)> callback_on_reset; // called before reset()

// this is for printing timings with slot progress, not part of metrics

int64_t t_print_last = 0;

int32_t n_gen_last = 0;

void reset() {

SLT_DBG(*this, "%s", "\n");

spec_is_replay = false;

last_nl_pos = 0;

generated_text = "";

has_new_line = false;

truncated = false;

stop = STOP_TYPE_NONE;

stopping_word = "";

n_sent_text = 0;

if (can_speculate()) {

spec_draft.clear();

spec_i_batch.clear();

spec_ckpt.clear();

}

generated_tokens.clear();

generated_token_probs.clear();

json_schema = json();

task_prev = std::move(task);

task.reset();

// note: callback_on_reset() must have run before this, see release()

stats = {};

n_accepted_per_pos.clear();

n_predict_max = -1;

llama_set_sampler(ctx_tgt, id, nullptr);

// clear alora start

alora_invocation_start = -1;

// clear multimodal state

mbatch.reset();

}

void init_sampler() const {

common_sampler_reset(smpl.get());

if (!task->need_sampling()) {

return;

}

const int64_t t_start = ggml_time_us();

int n_text = 0;

for (int i = 0; i < (int) prompt.tokens.size(); i++) {

const llama_token id = prompt.tokens[i];

if (id != LLAMA_TOKEN_NULL) {

common_sampler_accept(smpl.get(), id, false);

n_text++;

}

SLT_TRC(*this, "init sampler, took %0.2f ms, tokens: text = %d, total = %d\n",

(ggml_time_us() - t_start) / 1000.0, n_text, (int) prompt.tokens.size());

}

bool need_embd() const {

GGML_ASSERT(task);

return task->need_embd();

}

// if the context does not have a memory module then all embeddings have to be computed within a single ubatch

// also we cannot split if the pooling would require any past tokens

// (MTP supports splitting — uses task->need_embd() not need_embd())

bool can_split() const {

GGML_ASSERT(task);

return

!task->need_embd() ||

(llama_get_memory(ctx_tgt) && llama_pooling_type(ctx_tgt) == LLAMA_POOLING_TYPE_LAST);

}

bool can_batch_with(server_slot & other_slot) const {

GGML_ASSERT(task);

return task->type == other_slot.task->type

&& inp_embd.size() == other_slot.inp_embd.size()

&& are_lora_equal(lora, other_slot.lora);

}

// returns -1 if the generation is limitless

int32_t n_remaining() const {

return n_predict_max == -1 ? -1 : n_predict_max - (int32_t) stats.n_gen;

}

bool has_budget() const {

return n_predict_max == -1 || n_remaining() > 0;

}

bool is_processing() const {

return state != SLOT_STATE_IDLE;

}

bool can_speculate() const {

return !!spec;

}

void add_token(const completion_token_output & token) {

if (!is_processing()) {

SLT_WRN(*this, "%s", "slot is not processing\n");

return;

}

generated_token_probs.push_back(token);

}

int get_n_draft_max() const {

GGML_ASSERT(task);

if (!can_speculate()) {

return 0;

}

// determine the max draft that fits the current slot state

// note: slot.prompt is not yet expanded with the `id` token sampled above

// also, need to leave space for 1 extra token to allow context shifts

int n_draft_max = n_ctx - prompt.n_tokens() - 2;

if (n_remaining() > 0) {

n_draft_max = std::min(n_draft_max, n_remaining() - 1);

}

SLT_DBG(*this, "max possible draft: %d\n", n_draft_max);

return n_draft_max;

}

// add sampled token of this slot to the batch, optionally add the speculative draft tokens if any

void handle_last_sampled_token(server_batch & batch) {

bool add_ok = true;

if (spec_draft.empty()) {

// no speculative decoding

i_batch = batch.size();

if (!inp_embd.empty()) {

add_ok &= batch.add(id, inp_embd, prompt.tokens.pos_next(), true, false);

} else {

add_ok &= batch.add(id, sampled, prompt.tokens.pos_next(), true, false);

}

SLT_DBG(*this, "slot decode token, id=%d, n_ctx = %d, n_tokens = %d, truncated = %d\n",

sampled, n_ctx, prompt.n_tokens(), truncated);

} else {

SLT_DBG(*this, "generate_draft: id=%d, #tokens=%zu, #draft=%zu, pos_next=%d\n",

sampled, prompt.tokens.size(), spec_draft.size(), prompt.tokens.pos_next());

GGML_ASSERT(spec_i_batch.empty());

spec_i_batch.push_back(batch.size());

for (size_t i = 0; i < spec_draft.size(); i++) {

spec_i_batch.push_back(batch.size() + i + 1);

}

auto pos0 = prompt.tokens.pos_next();

add_ok &= batch.add(id, sampled, pos0++, true, false);

for (auto token : spec_draft) {

add_ok &= batch.add(this->id, token, pos0++, true, false);

}

GGML_ASSERT(add_ok && "batch must be large enough to hold the sampled and draft tokens");

prompt.tokens.push_back(sampled);

prompt.tokens.insert(spec_draft);

}

void release() {

if (is_processing()) {

GGML_ASSERT(task);

SLT_INF(*this, "stop processing: n_tokens = %d, truncated = %d\n", prompt.n_tokens(), truncated);

t_last_used = ggml_time_us();

state = SLOT_STATE_IDLE;

// do not keep context of the child slots - the parent's context is enough

if (task->is_child()) {

prompt_clear();

}

callback_on_reset(*this);

reset();

callback_on_release(id);

}

size_t find_stopping_strings(const std::string & text, const size_t last_token_size, bool is_full_stop) {

GGML_ASSERT(task);

size_t stop_pos = std::string::npos;

for (const std::string & word : task->params.antiprompt) {

size_t pos;

if (is_full_stop) {

const size_t tmp = word.size() + last_token_size;

const size_t from_pos = text.size() > tmp ? text.size() - tmp : 0;

pos = text.find(word, from_pos);

} else {

// otherwise, partial stop

pos = string_find_partial_stop(text, word);

}

if (pos != std::string::npos && (stop_pos == std::string::npos || pos < stop_pos)) {

if (is_full_stop) {

stop = STOP_TYPE_WORD;

stopping_word = word;

has_next_token = false;

}

stop_pos = pos;

}

return stop_pos;

}

void print_timings_tg() {

if (stats.n_gen < 100) {

return;

}

const int64_t t_now = ggml_time_us();

if (t_now - t_print_last < 3*1000*1000) {

return;

}

const double n_gen_second = stats.n_gen_tps();

const double n_gen_second_win = 1e6 / (t_now - t_print_last) * (stats.n_gen - n_gen_last);

t_print_last = t_now;

n_gen_last = stats.n_gen;

SLT_INF(*this, "n_gen = %6d, tg = %6.2f t/s, tg_3s = %6.2f t/s\n", (int) stats.n_gen, n_gen_second, n_gen_second_win);

}

void print_timings_pp() const {

const double t_prompt_total = stats.t_prompt_ms();

if (t_prompt_total < 3000.0) {

return;

}

const double n_prompt_second = stats.n_prompt_tps();

const double f_progress = task->n_tokens() > 0 ? (double) prompt.n_tokens() / task->n_tokens() : 0.0;

SLT_INF(*this, "prompt processing, n_tokens = %6d, progress = %.2f, t = %6.2f s / %.2f tokens per second\n",

(int) stats.n_prompt_processed, f_progress, t_prompt_total / 1e3, n_prompt_second);

}

void print_timings() const {

const double t_prompt_total = stats.t_prompt_ms();

const double t_gen_total = stats.t_gen_ms();

const double t_prompt = stats.t_prompt_per_token_ms();

const double n_prompt_second = stats.n_prompt_tps();

const double t_gen = stats.t_gen_per_token_ms();

const double n_gen_second = stats.n_gen_tps();

SLT_INF(*this,

"prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",

t_prompt_total, (int) stats.n_prompt_processed, t_prompt, n_prompt_second);

SLT_INF(*this,

" eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",

t_gen_total, (int) stats.n_gen, t_gen, n_gen_second);

SLT_INF(*this,

" total time = %10.2f ms / %5d tokens\n",

t_prompt_total + t_gen_total, (int) (stats.n_prompt_processed + stats.n_gen));

SLT_INF(*this,

" graphs reused = %10d\n",

llama_perf_context(ctx_tgt).n_reused);

const int32_t n_draft_total = stats.n_draft_tokens;

const int32_t n_draft_accepted = stats.n_draft_accepted;

const int32_t n_draft_verif_steps = stats.n_draft_verif_steps;

if (n_draft_total > 0) {

const float draft_ratio = (float) n_draft_accepted / n_draft_total;

const double mean_acc_len = n_draft_verif_steps > 0 ? 1.0 + (double) n_draft_accepted / (double) n_draft_verif_steps : 1.0;

std::string acceptance_rates_per_pos;

if (n_draft_verif_steps > 0) {

for (size_t i = 0; i < n_accepted_per_pos.size(); ++i) {

if (i > 0) {

acceptance_rates_per_pos += ", ";

}

acceptance_rates_per_pos += string_format("%.3f", (double) n_accepted_per_pos[i] / (double) n_draft_verif_steps);

}

SLT_INF(*this,

"draft acceptance = %0.5f (%5d accepted / %5d generated), mean len = %5.2f\n",

draft_ratio, n_draft_accepted, n_draft_total, mean_acc_len);

SLT_TRC(*this,

" acc per pos = (%s)\n", acceptance_rates_per_pos.c_str());

}

common_speculative_print_stats(spec);

}

json to_json(bool only_metrics = false) const {

json res;

res = {

{"id", id},

{"n_ctx", n_ctx},

{"speculative", can_speculate()},

{"is_processing", is_processing()},

};

const auto & ptask = task ? task : task_prev;

if (ptask) {

res["id_task"] = ptask->id;

res["n_prompt_tokens"] = (int32_t) prompt.tokens.size();

res["n_prompt_tokens_processed"] = stats.n_prompt_processed;

res["n_prompt_tokens_cache"] = stats.n_prompt_cached;

res["params"] = ptask->params.to_json(only_metrics);

res["next_token"] = json::array({

{

{"has_next_token", has_next_token},

{"has_new_line", has_new_line},

{"n_remain", n_remaining()},

{"n_decoded", stats.n_gen},

}

});

if (!only_metrics) {

res["prompt"] = ptask->tokens.detokenize(ctx_tgt, true);

res["generated"] = generated_text.empty() ? debug_generated_text : generated_text;

}

return res;

}

void copy_state_to(server_slot & other) const {

GGML_ASSERT(state == SLOT_STATE_DONE_PROMPT);

mem.seq_rm(other.id, -1, -1);

mem.seq_cp(id, other.id, -1, -1);

other.i_batch = i_batch;

other.stats = stats;

other.prompt = prompt.clone();

other.init_sampler();

}

};

// returns 0 on success

// caller need to update prompt.tokens after a successful call to keep track of the processing progress

// note: this is not a member of server_slot because we want to run it inside yield_to_queue

// slot is passed as const to avoid accidental modification of the slot state

// some pointers are allowed to be used, they are not used by to_json()

static int process_mtmd_chunk(const server_slot & slot, mtmd::batch_ptr & mbatch, size_t idx, size_t & n_tokens_out) {

GGML_ASSERT(slot.mctx);

const auto & mctx = slot.mctx;

const auto & input_tokens = slot.task->tokens;

const auto & chunk = input_tokens.find_chunk(idx);

int32_t res = 0;

auto try_decode = [&]() -> int32_t {

if (mbatch) {

float * embd = mtmd_batch_get_output_embd(mbatch.get(), chunk.get());

if (embd) {

void * cb_data = slot.spec;

static auto cb = [](llama_batch batch, void * user_data) {

common_speculative * spec = static_cast<common_speculative *>(user_data);

if (!common_speculative_process(spec, batch)) {

return 1;

}

return 0;

};

llama_pos new_n_past; // unused for now

res = mtmd_helper_decode_image_chunk(

mctx,

slot.ctx_tgt,

chunk.get(),

embd,

slot.prompt.tokens.pos_next(),

slot.id,

llama_n_batch(slot.ctx_tgt),

&new_n_past,

cb,

cb_data

);

if (res != 0) {

SLT_ERR(slot, "failed to decode mtmd chunk, idx = %zu, res = %d\n", idx, res);

return -1;

}

n_tokens_out = mtmd_input_chunk_get_n_tokens(chunk.get());

return 0; // success

}

return 1; // (non-error) need to create & encode batch

};

// if the batch is already exist, try searching & encode

res = try_decode();

if (res == 0) {

return 0;

}

if (res < 0) {

// fatal error

return res;

}

// otherwise, the batch is either uninitialized or is used up

// we need to create & encode a new batch

mbatch.reset(mtmd_batch_init(mctx));

res = mtmd_batch_add_chunk(mbatch.get(), chunk.get());

GGML_ASSERT(res == 0); // we should never have an empty batch

// try batching as much as possible

int n_added = 1;

size_t idx_cur = idx;

while (res == 0) {

auto [next_chunk, next_idx] = input_tokens.find_next_media_chunk(idx_cur);

if (next_chunk == nullptr) {

break;

}

res = mtmd_batch_add_chunk(mbatch.get(), next_chunk->get());

n_added += (res == 0 ? 1 : 0);

idx_cur = next_idx;

SLT_DBG(slot, "try adding media chunk idx = %zu to batch, res = %d\n", next_idx, res);

// if res != 0, batch is full or chunk is not compatible -> this loop breaks

}

// TODO @ngxson : move this log line to debug when it become more stable

SLT_TRC(slot, "encoding mtmd batch from idx = %zu, n_chunks = %d\n", idx, n_added);

res = mtmd_batch_encode(mbatch.get());

if (res != 0) {

SLT_ERR(slot, "failed to encode mtmd batch for chunk idx = %zu, res = %d\n", idx, res);

return -1;

}

return try_decode();

}

// server_context_impl (private implementation)

struct server_context_impl {

friend struct server_context;

public:

// only use these pointers outside of this class:

// - when not in sleeping state

// - and, with thread-safe APIs (e.g., tokenizer calls)

llama_model * model_tgt = nullptr;

mtmd_context * mctx = nullptr;

// note: video_params.ffmpeg_bin_dir points into params_base, which outlives this struct

mtmd_helper_init_opt init_opt = mtmd_helper_init_opt_default();

const llama_vocab * vocab = nullptr;

server_queue queue_tasks;

server_response queue_results;

// note: chat_params must not be refreshed upon existing sleeping state

server_chat_params chat_params;

server_state_callback_t callback_state = [](server_state, json) -> void {};

server_context_impl() {

mtmd_helper_log_set(common_log_default_callback, nullptr);

}

~server_context_impl() {

if (!sleeping) {

// destroy() is already called when entering sleeping state

// we don't call it again here to avoid double free

destroy();

}

server_metrics get_metrics() const {

return metrics;

}

void reset_metrics_bucket() {

metrics.reset_bucket();

}

private:

// note: accessing these fields outside of this class is not thread-safe

// use server_context methods instead

common_params params_base;

// note: keep these alive - they determine the lifetime of the model, context, etc.

common_init_result_ptr llama_init;

llama_context * ctx_tgt = nullptr;

server_batch batch;

llama_model * model_dft = nullptr;

llama_context * ctx_dft = nullptr;

common_speculative_init_result_ptr spec_init;

common_context_seq_rm_type ctx_tgt_seq_rm_type = COMMON_CONTEXT_SEQ_RM_TYPE_NO;

common_context_seq_rm_type ctx_dft_seq_rm_type = COMMON_CONTEXT_SEQ_RM_TYPE_NO;

common_speculative_ptr spec;

bool add_bos_token = true;

int32_t n_ctx; // total context for all clients / slots

// set to llama_model_n_swa(model)

// if swa_full is enabled, this is set to 0 to simulate a non-SWA model

int32_t n_swa;

// slots / clients

std::vector<server_slot> slots;

int trace = 0; // env: LLAMA_TRACE

int slots_debug = 0; // env: LLAMA_SERVER_SLOTS_DEBUG

int slots_n_diff = 0; // env: LLAMA_SERVER_SLOTS_N_DIFF

int n_empty_consecutive = 0;

std::unique_ptr<server_prompt_cache> prompt_cache;

server_metrics metrics;

// queued prompt stats - llama_decode() is async, so the timing is only valid after a sync

// note: kept out of server_metrics, which is copied as-is into the task result

int64_t t_decode_start = 0; // start of the last submitted decode

int64_t t_prompt_start = 0; // start of the oldest queued prompt decode

uint64_t n_prompt_queued = 0;

json json_ui_settings = json::object();

// Necessary similarity of prompt for slot selection

float slot_prompt_similarity = 0.0f;

std::string model_name; // name of the loaded model, to be used by API

std::set<std::string> model_aliases; // additional names for the model

std::set<std::string> model_tags; // informational tags

bool sleeping = false;

int64_t t_last_load_progress_ms = 0;

void destroy() {

spec.reset();

spec_init.reset();

ctx_dft = nullptr;

model_dft = nullptr;

llama_init.reset();

ctx_tgt = nullptr;

model_tgt = nullptr;

mtmd_free(mctx);

mctx = nullptr;

}

void handle_sleeping_state(bool new_state) {

GGML_ASSERT(sleeping != new_state);

if (new_state) {

if (callback_state) {

callback_state(SERVER_STATE_SLEEPING, {});

// note: for sleeping == false, event is emitted by load_model()

}

SRV_INF("%s", "server is entering sleeping state\n");

destroy();

} else {

SRV_INF("%s", "server is exiting sleeping state\n");

if (!load_model(params_base)) {

GGML_ABORT("failed to reload model after sleeping");

}

sleeping = new_state;

}

struct load_progress_data {

server_context_impl * ctx;

std::string stage;

std::vector<std::string> stages;

int64_t t_last_load_progress_ms = 0;

load_progress_data(server_context_impl * ctx, const std::string & stage) : ctx(ctx), stage(stage) {}

};

static bool load_progress_callback(float progress, void * user_data) {

auto * d = static_cast<load_progress_data *>(user_data);

GGML_ASSERT(d);

// always emit the first and final sample; throttle the rest to one per 200ms

{

auto & t_last = d->t_last_load_progress_ms;

const int64_t t_now = ggml_time_ms();

const bool first = t_last == 0;

const bool done = progress >= 1.0f;

const bool throttled = !first && !done && (t_now - t_last) < 200;

if (throttled) {

return true;

}

t_last = t_now;

}

if (d->ctx->callback_state) {

d->ctx->callback_state(SERVER_STATE_LOADING, {

{"stages", d->stages},

{"current", d->stage},

{"value", progress},

});

}

View remainder of file in raw view

llama.cpp/tools/server/server-context.cpp at master · rioscode/llama.cpp · GitHub

Navigation Menu

Expand file tree

Breadcrumbs

server-context.cpp

Latest commit

History

Breadcrumbs

server-context.cpp

File metadata and controls

Footer

Footer navigation