TransferQueue/transfer_queue/controller.py at dev · TransferQueue/TransferQueue · GitHub

This repository was archived by the owner on Jan 21, 2026. It is now read-only.

History

1422 lines (1147 loc) · 57.6 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

# Licensed under the Apache License, Version 2.0 (the "License");

# you may not use this file except in compliance with the License.

# You may obtain a copy of the License at

# http://www.apache.org/licenses/LICENSE-2.0

# Unless required by applicable law or agreed to in writing, software

# distributed under the License is distributed on an "AS IS" BASIS,

# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

# See the License for the specific language governing permissions and

# limitations under the License.

import copy

import logging

import os

import time

from collections import defaultdict

from dataclasses import dataclass, field

from itertools import groupby

from operator import itemgetter

from threading import Lock, Thread

from typing import Any, Optional

from uuid import uuid4

import ray

import torch

import zmq

from ray.util import get_node_ip_address

from transfer_queue.metadata import (

BatchMeta,

FieldMeta,

SampleMeta,

)

from transfer_queue.sampler import BaseSampler, SequentialSampler

from transfer_queue.utils.perf_utils import IntervalPerfMonitor

from transfer_queue.utils.utils import (

ProductionStatus,

TransferQueueRole,

)

from transfer_queue.utils.zmq_utils import (

ZMQMessage,

ZMQRequestType,

ZMQServerInfo,

create_zmq_socket,

get_free_port,

)

logger = logging.getLogger(__name__)

logger.setLevel(os.getenv("TQ_LOGGING_LEVEL", logging.WARNING))

# Ensure logger has a handler (for Ray Actor subprocess)

if not logger.hasHandlers():

handler = logging.StreamHandler()

handler.setFormatter(logging.Formatter("%(asctime)s - %(levelname)s - %(name)s - %(message)s"))

logger.addHandler(handler)

TQ_CONTROLLER_GET_METADATA_TIMEOUT = int(os.environ.get("TQ_CONTROLLER_GET_METADATA_TIMEOUT", 1))

TQ_CONTROLLER_GET_METADATA_CHECK_INTERVAL = int(os.environ.get("TQ_CONTROLLER_GET_METADATA_CHECK_INTERVAL", 0.2))

TQ_CONTROLLER_CONNECTION_CHECK_INTERVAL = int(os.environ.get("TQ_CONTROLLER_CONNECTION_CHECK_INTERVAL", 2))

TQ_INIT_SAMPLE_NUM = int(os.environ.get("TQ_INIT_SAMPLE_NUM", 1)) # Initial number of samples

TQ_INIT_FIELD_NUM = int(os.environ.get("TQ_INIT_FIELD_NUM", 1))

# Expansion configuration - Unified approach using minimum expansion sizes

TQ_SAMPLE_MIN_EXPANSION_SIZE = int(

os.environ.get("TQ_SAMPLE_MIN_EXPANSION_SIZE", 1)

) # Minimum expansion size for samples (rows)

TQ_FIELD_MIN_EXPANSION_SIZE = int(

os.environ.get("TQ_FIELD_MIN_EXPANSION_SIZE", 1)

) # Minimum expansion size for fields (columns)

class PartitionIndexManager:

"""

Manages the mapping relationship between partitions and global indexes,

responsible for index allocation and reuse.

"""

def __init__(self):

# Records the set of global_indexes used by each partition

self.partition_to_indexes = defaultdict(set)

# Reusable global_index pool - stored using list

self.reusable_indexes = []

# Global index counter for allocating new indexes

self.global_index_counter = 0

# Track all active indexes

self.allocated_indexes = set()

def allocate_indexes(self, partition_id, count=1) -> list:

"""

Allocate global_indexes for the specified partition.

Prioritizes obtaining from reusable pool, allocates new indexes when insufficient.

Args:

partition_id: Partition ID

count: Number of indexes needed

Returns:

list: List of allocated global_indexes

"""

if count <= 0:

raise ValueError(f"Number of indexes needed must be larger than 0, but got {count}")

indexes = []

# Get indexes from reusable pool

if self.reusable_indexes:

# Calculate number of indexes needed from reusable pool

num_reuse = min(count, len(self.reusable_indexes))

# Use slice operation to get multiple elements at once (FIFO principle)

indexes.extend(self.reusable_indexes[:num_reuse])

del self.reusable_indexes[:num_reuse]

# If reusable pool doesn't have enough indexes, allocate new ones

if len(indexes) < count:

# Ensure newly allocated indexes don't conflict with existing ones

needed = count - len(indexes)

# Batch allocate consecutive index ranges

start_index = self.global_index_counter

end_index = start_index + needed

# Directly generate consecutive index list

new_indexes = list(range(start_index, end_index))

# Batch update status

self.allocated_indexes.update(new_indexes)

self.global_index_counter = end_index

indexes.extend(new_indexes)

# Record partition-index relationship

self.partition_to_indexes[partition_id].update(indexes)

return indexes

def release_partition(self, partition_id) -> list[int]:

"""

Release all global_indexes of the specified partition, adding them to reusable pool.

Args:

partition_id: Partition ID

Returns:

list: List of released global_indexes

"""

if partition_id in self.partition_to_indexes:

indexes = self.partition_to_indexes.pop(partition_id)

# Add released indexes to reusable pool

self.reusable_indexes.extend(indexes)

# Remove these indexes from allocated_indexes

for idx in indexes:

self.allocated_indexes.discard(idx)

return list(indexes)

return []

def release_indexes(self, partition_id: str, indexes_to_release: list[int]):

"""

Release specific global_indexes for a partition, adding them to reusable pool.

Args:

partition_id: Partition ID

indexes_to_release: List of specific indexes to release

"""

if partition_id not in self.partition_to_indexes:

return []

partition_indexes = self.partition_to_indexes[partition_id]

if not set(indexes_to_release).issubset(partition_indexes):

raise ValueError("Some indexes to release do not belong to the specified partition.")

partition_indexes.difference_update(indexes_to_release)

self.reusable_indexes.extend(indexes_to_release)

self.allocated_indexes.difference_update(indexes_to_release)

# If partition has no more indexes, remove it from the mapping

if not partition_indexes:

self.partition_to_indexes.pop(partition_id, None)

def get_indexes_for_partition(self, partition_id) -> set[int]:

"""

Get all global_indexes for the specified partition.

Args:

partition_id: Partition ID

Returns:

set: Set of global_indexes for this partition

"""

return self.partition_to_indexes.get(partition_id, set()).copy()

@dataclass

class DataPartitionStatus:

"""

Robust status information for a data partition with dynamic expansion support.

This class tracks the production and consumption status of data within a specific

partition (e.g., "train@global_batch_0", "inference@kv_cache_1") with full support

for dynamic row and column expansion.

"""

partition_id: str

created_at: float = field(default_factory=time.time)

# Production status tensor - dynamically expandable

# Values: 0 = not produced, 1 = ready for consumption

production_status: Optional[torch.Tensor] = torch.zeros(TQ_INIT_SAMPLE_NUM, TQ_INIT_FIELD_NUM, dtype=torch.int8)

# Consumption status per task - task_name -> consumption_tensor

# Each tensor tracks which samples have been consumed by that task

consumption_status: dict[str, torch.Tensor] = field(default_factory=dict)

# Sample metadata

global_indexes: set[int] = field(

default_factory=set

) # set of global indexes that have been added to this partition

# Field metadata

field_name_mapping: dict[str, int] = field(default_factory=dict) # field_name -> column_index

field_dtypes: dict[int, dict[str, Any]] = field(default_factory=dict) # global_idx -> {field: dtype}

field_shapes: dict[int, dict[str, Any]] = field(default_factory=dict) # global_idx -> {field: shape}

# Threading lock for concurrency control; only for preventing mask operation error when expanding production_status.

# No need to strictly lock for every read/write operation since freshness is not critical.

data_status_lock: Lock = field(default_factory=Lock)

# Dynamic configuration - these are computed from the current state

@property

def total_samples_num(self) -> int:

"""Current number of samples in the partition."""

return len(self.global_indexes)

@property

def total_fields_num(self) -> int:

"""Current number of fields (columns) in the partition."""

return len(self.field_name_mapping)

@property

def allocated_fields_num(self) -> int:

"""Current number of allocated columns in the tensor."""

return self.production_status.shape[1] if self.production_status is not None else 0

@property

def allocated_samples_num(self) -> int:

"""Current number of allocated rows in the tensor."""

return self.production_status.shape[0] if self.production_status is not None else 0

# ==================== Dynamic Expansion Methods ====================

def ensure_samples_capacity(self, required_samples: int) -> bool:

"""

Ensure the production status tensor has enough rows for the required samples.

Dynamically expands if needed using unified minimum expansion size.

Args:

required_samples: Minimum number of samples needed

"""

current_sample_space = self.allocated_samples_num

if required_samples > current_sample_space:

# Expand rows using minimum expansion size for predictable memory usage

expansion_needed = required_samples - current_sample_space

min_expansion = max(TQ_SAMPLE_MIN_EXPANSION_SIZE, expansion_needed)

new_samples = current_sample_space + min_expansion

new_fields = self.production_status.shape[1]

expanded_tensor = torch.zeros(new_samples, new_fields, dtype=torch.int8)

expanded_tensor[:current_sample_space, :] = self.production_status

self.production_status = expanded_tensor

# Update consumption tensors for all tasks

for task_name, consumption_tensor in self.consumption_status.items():

expanded_consumption = torch.zeros(new_samples, dtype=torch.int8)

expanded_consumption[:current_sample_space] = consumption_tensor

self.consumption_status[task_name] = expanded_consumption

logger.debug(

f"Expanded partition {self.partition_id} from {current_sample_space} "

f"to {new_samples} samples (added {min_expansion} samples)"

)

def ensure_fields_capacity(self, required_fields: int):

"""

Ensure the production status tensor has enough columns for the required fields.

Dynamically expands if needed using unified minimum expansion size.

Args:

required_fields: Minimum number of fields needed

"""

if self.production_status is None:

# Will be initialized when samples are added

return

current_fields = self.production_status.shape[1]

if required_fields > current_fields:

# Expand columns using minimum expansion size for predictable memory usage

expansion_needed = required_fields - current_fields

min_expansion = max(TQ_FIELD_MIN_EXPANSION_SIZE, expansion_needed)

new_fields = current_fields + min_expansion

new_samples = self.production_status.shape[0]

expanded_tensor = torch.zeros(new_samples, new_fields, dtype=torch.int8)

expanded_tensor[:, :current_fields] = self.production_status

self.production_status = expanded_tensor

logger.debug(

f"Expanded partition {self.partition_id} from {current_fields} "

f"to {new_fields} fields (added {min_expansion} fields)"

)

# ==================== Production Status Interface ====================

def update_production_status(

self,

global_indices: list[int],

field_names: list[str],

dtypes: Optional[dict[int, dict[str, Any]]],

shapes: Optional[dict[int, dict[str, Any]]],

) -> bool:

"""

Update production status for specific samples and fields.

Handles dynamic expansion of both samples and fields.

Args:

global_indices: List of sample indices to update

field_names: List of field names to mark as produced

dtypes: Optional per-sample field dtype information

shapes: Optional per-sample field shape information

Returns:

True if update was successful, False on error

"""

try:

# Determine required capacity

max_sample_idx = max(global_indices) if global_indices else -1

required_samples = max_sample_idx + 1

# Ensure we have enough rows

with self.data_status_lock:

self.ensure_samples_capacity(required_samples)

# Register new fields if needed

new_fields = [field for field in field_names if field not in self.field_name_mapping]

if new_fields:

# Add new fields to mapping

for field in new_fields:

self.field_name_mapping[field] = len(self.field_name_mapping)

required_fields = len(self.field_name_mapping)

with self.data_status_lock:

self.ensure_fields_capacity(required_fields)

# Update production status

if self.production_status is not None and global_indices and field_names:

field_indices = [self.field_name_mapping.get(field) for field in field_names]

self.production_status[torch.tensor(global_indices)[:, None], torch.tensor(field_indices)] = 1

# Update field metadata

self._update_field_metadata(global_indices, dtypes, shapes)

# Save these global_indexes

self.global_indexes.update(global_indices)

return True

except Exception as e:

logger.error(f"Error updating production status for partition {self.partition_id}: {e}")

return False

def _update_field_metadata(

self,

global_indices: list[int],

dtypes: Optional[dict[int, dict[str, Any]]],

shapes: Optional[dict[int, dict[str, Any]]],

"""Update field dtype and shape metadata."""

if not global_indices:

return

assert len(global_indices) == len(dtypes), "`global_indices` and `dtypes` length mismatch."

assert len(global_indices) == len(shapes), "`global_indices` and `shapes` length mismatch."

dtype_value = itemgetter(*global_indices)(dtypes) if dtypes else None

shape_value = itemgetter(*global_indices)(shapes) if shapes else None

if not isinstance(dtype_value, tuple):

dtype_value = (dtype_value,)

if not isinstance(shape_value, tuple):

shape_value = (shape_value,)

for i, global_idx in enumerate(global_indices):

if global_idx not in self.field_dtypes:

self.field_dtypes[global_idx] = {}

if global_idx not in self.field_shapes:

self.field_shapes[global_idx] = {}

if dtype_value is not None:

self.field_dtypes[global_idx].update(dtype_value[i])

if shape_value is not None:

self.field_shapes[global_idx].update(shape_value[i])

# ==================== Consumption Status Interface ====================

def get_consumption_status(self, task_name: str) -> torch.Tensor:

"""

Get or create consumption status for a specific task.

Handles dynamic expansion when new samples are added.

Args:

task_name: Name of the consumer task

Returns:

Consumption status tensor for the specified task

"""

if task_name not in self.consumption_status:

if self.production_status is not None:

self.consumption_status[task_name] = torch.zeros(self.allocated_samples_num, dtype=torch.int8)

else:

self.consumption_status[task_name] = torch.zeros(0, dtype=torch.int8)

return self.consumption_status[task_name]

def mark_consumed(self, task_name: str, global_indices: list[int]):

"""

Mark specific samples as consumed by a task.

Args:

task_name: Name of the consumer task

global_indices: List of sample indices to mark as consumed

"""

try:

consumption_status = self.get_consumption_status(task_name)

if consumption_status.numel() > 0 and global_indices:

consumption_status[global_indices] = 1

except Exception as e:

logger.error(

f"Error marking samples consumed for partition {self.partition_id}, task {task_name}: {e}. "

f"Target global_indices {global_indices}, but current consumption_status has "

f"shape {consumption_status.shape}"

)

def get_production_status_for_fields(self, field_names: list[str]) -> bool:

"""

Check if all samples for specified fields are fully produced and ready.

Args:

field_names: List of field names to check production status for

Returns:

bool: True if all samples have been produced for all specified fields, False otherwise

"""

if self.production_status is None or field_names is None or len(field_names) == 0:

return False

# Check if all requested fields are registered

for field_name in field_names:

if field_name not in self.field_name_mapping:

return False

# Create column mask for requested fields

col_mask = torch.zeros(self.allocated_fields_num, dtype=torch.bool)

field_indices = [self.field_name_mapping[field] for field in field_names]

if field_indices:

col_mask[field_indices] = True

# Get production status for requested fields

relevant_status = self.production_status[:, col_mask]

# Check if all samples have all requested fields produced (all values are 1)

all_fields_produced = torch.all(relevant_status == 1).item()

return all_fields_produced

# ==================== Data Scanning and Query Methods ====================

def scan_data_status(self, field_names: list[str], task_name: str) -> list[int]:

"""

Scan data status to find samples ready for consumption.

This replaces the original _scan_data_status functionality.

Args:

field_names: List of required field names

task_name: Name of the consumer task

Returns:

List of sample indices that are ready for consumption

"""

if self.production_status is None:

return []

# Check if all requested fields are registered

for field_name in field_names:

if field_name not in self.field_name_mapping:

return []

with self.data_status_lock:

row_mask = torch.ones(self.allocated_samples_num, dtype=torch.bool)

# Apply consumption filter (exclude already consumed samples)

consumption_status = self.get_consumption_status(task_name)

if consumption_status is not None:

unconsumed_mask = consumption_status == 0

row_mask &= unconsumed_mask

# Create column mask for requested fields

col_mask = torch.zeros(self.allocated_fields_num, dtype=torch.bool)

field_indices = [self.field_name_mapping[field] for field in field_names]

if field_indices:

col_mask[field_indices] = True

# Filter production status by masks

relevant_status = self.production_status[row_mask][:, col_mask]

# Check if all required fields are ready for each sample

all_fields_ready = torch.all(relevant_status, dim=1)

ready_indices_in_filtered = torch.nonzero(all_fields_ready, as_tuple=False).flatten()

# Map back to original sample indices

all_indices = torch.where(row_mask)[0]

ready_sample_indices = all_indices[ready_indices_in_filtered].tolist()

return ready_sample_indices

# ==================== Field Metadata Methods ====================

def get_field_dtype(self, global_index: int, field_name: str) -> Optional[Any]:

"""Get dtype for a specific sample and field."""

return self.field_dtypes.get(global_index, {}).get(field_name)

def get_field_shape(self, global_index: int, field_name: str) -> Optional[Any]:

"""Get shape for a specific sample and field."""

return self.field_shapes.get(global_index, {}).get(field_name)

# ==================== Statistics and Monitoring ====================

def get_statistics(self) -> dict[str, Any]:

"""Get detailed statistics for this partition."""

stats = {

"partition_id": self.partition_id,

"created_at": self.created_at,

"total_samples_num": self.total_samples_num,

"total_fields_num": self.total_fields_num,

"allocated_samples_num": self.allocated_samples_num,

"allocated_fields_num": self.allocated_fields_num,

"registered_tasks": list(self.consumption_status.keys()),

}

if self.production_status is not None:

produced_samples = torch.any(self.production_status == 1, dim=1).sum().item()

stats["produced_samples"] = produced_samples

stats["production_progress"] = (

produced_samples / self.total_samples_num if self.total_samples_num > 0 else 0

)

# Field-wise production statistics

field_stats = {}

for field_name, field_idx in self.field_name_mapping.items():

field_produced = (self.production_status[:, field_idx] == 1).sum().item()

field_stats[field_name] = {

"produced_samples": field_produced,

"production_progress": field_produced / self.total_samples_num if self.total_samples_num > 0 else 0,

}

stats["field_statistics"] = field_stats

# Consumption statistics per task

consumption_stats = {}

for task_name, consumption_tensor in self.consumption_status.items():

consumed_samples = (consumption_tensor == 1).sum().item()

consumption_stats[task_name] = {

"consumed_samples": consumed_samples,

"consumption_progress": consumed_samples / self.total_samples_num if self.total_samples_num > 0 else 0,

}

stats["consumption_statistics"] = consumption_stats

return stats

# ==================== Serialization ====================

def to_snapshot(self):

"""

Get a snapshot of partition status information.

Returns:

DataPartitionStatus object without threading.Lock()

"""

def _perform_copy():

cls = self.__class__

snapshot = cls.__new__(cls)

for name, value in self.__dict__.items():

if name == "data_status_lock":

continue

if isinstance(value, torch.Tensor):

new_val = value.clone().detach()

else:

new_val = copy.deepcopy(value)

setattr(snapshot, name, new_val)

return snapshot

lock_obj = getattr(self, "data_status_lock", None)

if lock_obj:

with lock_obj:

return _perform_copy()

else:

return _perform_copy()

def clear_data(self, indexes_to_release: list[int], clear_consumption: bool = True):

"""Clear all production and optionally consumption data for given global_indexes."""

try:

if self.production_status is not None:

self.production_status[indexes_to_release, :] = 0

if clear_consumption:

for consumption_tensor in self.consumption_status.values():

consumption_tensor[indexes_to_release] = 0

self.global_indexes.difference_update(indexes_to_release)

except Exception as e:

logger.error(

f"Error clearing data for partition {self.partition_id}: {e}. "

f"Attempted to clear global_indexes: {indexes_to_release}"

)

@ray.remote(num_cpus=1)

class TransferQueueController:

"""

TransferQueue Controller with partition-based data management.

This refactored controller manages data through dynamic partitions instead of

fixed global batches. Each partition represents a logical data container

(e.g., "train@global_batch_0", "inference@kv_cache_1") that can be created

on-demand and managed independently.

Key improvements:

- Dynamic partition creation on-demand

- No dependency on training-specific parameters (global_batch_size, etc.)

- Support for diverse use cases (KV cache migration, model resharding, etc.)

- Flexible data organization through partition-based addressing

"""

def __init__(

self, sampler: BaseSampler | type[BaseSampler] = SequentialSampler, polling_mode: bool = False

) -> None:

"""Initialize the TransferQueue Controller.

Args:

sampler: Sampler instance or sampler class to use for data sampling.

- If a BaseSampler instance is provided, it will be used directly

- If a BaseSampler subclass is provided, it will be instantiated

- Defaults to SequentialSampler for simple sequential sampling

- Example: sampler=GRPOGroupNSampler() (instance)

- Example: sampler=GRPOGroupNSampler (class)

polling_mode: Whether to use polling mode for TransferQueue controller.

- If False, the controller will raise an error when no enough data is available.

- If True, the controller will return an empty BatchMeta when no enough data is available.

The user side is responsible for handling this empty case (retrying later).

"""

if isinstance(sampler, BaseSampler):

self.sampler = sampler

elif isinstance(sampler, type) and issubclass(sampler, BaseSampler):

self.sampler = sampler()

else:

raise TypeError(

f"sampler {getattr(sampler, '__name__', repr(sampler))} must be an instance or subclass of BaseSampler"

)

self.controller_id = f"TQ_CONTROLLER_{uuid4().hex[:8]}"

self.polling_mode = polling_mode

# Initialize ZMQ sockets for communication

self._init_zmq_socket()

# Partition management

self.partitions: dict[str, DataPartitionStatus] = {} # partition_id -> DataPartitionStatus

# Partition-GlobalIndex management

self.index_manager = PartitionIndexManager() # partition_id -> global_indexes

# Connected storage managers tracking

self._connected_storage_managers: set[str] = set()

# Start background processing threads

self._start_process_handshake()

self._start_process_update_data_status()

self._start_process_request()

logger.info(f"TransferQueue Controller {self.controller_id} initialized")

# ==================== Partition Management API ====================

def create_partition(self, partition_id: str) -> bool:

"""

Create a new data partition.

Note: Partitions now dynamically expand as needed, so initial capacity is not required.

Args:

partition_id: Unique identifier for the partition (e.g., "train@global_batch_0")

Returns:

True if partition was created successfully, False if it already exists

"""

if partition_id in self.partitions:

logger.warning(f"Partition {partition_id} already exists")

return False

self.partitions[partition_id] = DataPartitionStatus(partition_id=partition_id)

logger.info(f"Created partition {partition_id}")

return True

def _get_partition(self, partition_id: str) -> Optional[DataPartitionStatus]:

"""

Get partition status information.

Args:

partition_id: ID of the partition to retrieve

Returns:

DataPartitionStatus object if partition exists, None otherwise

"""

return self.partitions.get(partition_id)

def get_partition_snapshot(self, partition_id: str) -> Optional[DataPartitionStatus]:

"""

Get a copy of partition status information, without threading.Lock().

Args:

partition_id: ID of the partition to retrieve

Returns:

DataPartitionStatus object if partition exists, None otherwise

"""

partition = self._get_partition(partition_id)

if partition is None:

return None

return partition.to_snapshot()

def list_partitions(self) -> list[str]:

"""

List all available partition IDs.

Returns:

List of partition IDs

"""

return list(self.partitions.keys())

# ==================== Partition Index Management API ====================

def get_partition_index_range(self, partition: DataPartitionStatus) -> set:

"""

Get all indexes for a specific partition.

Args:

partition: Partition identifier

Returns:

Set of indexes allocated to the partition

"""

return self.index_manager.get_indexes_for_partition(partition)

# ==================== Data Production API ====================

# TODO: Modify dtypes & shapes to be required

def update_production_status(

self,

partition_id: str,

global_indexes: list[int],

field_names: list[str],

dtypes: Optional[dict[int, dict[str, Any]]],

shapes: Optional[dict[int, dict[str, Any]]],

) -> bool:

"""

Update production status for specific samples and fields in a partition.

Delegates to the partition's own update_production_status method.

Args:

partition_id: ID of the partition

global_indexes: List of sample indices to update

field_names: List of field names to mark as produced

dtypes: Optional per-sample field dtype information

shapes: Optional per-sample field shape information

Returns:

True if update was successful, False otherwise

"""

partition = self._get_partition(partition_id)

if not partition:

logger.error(f"Partition {partition_id} not found")

return False

success = partition.update_production_status(global_indexes, field_names, dtypes, shapes)

if success:

logger.debug(

f"[{self.controller_id}]: Updated production status for partition {partition_id}: "

f"samples={global_indexes}, fields={field_names}"

)

return success

# ==================== Data Consumption API ====================

def get_consumption_status(self, partition_id: str, task_name: str) -> Optional[torch.Tensor]:

"""

Get or create consumption status for a specific task and partition.

Delegates to the partition's own method.

Args:

partition_id: ID of the partition

task_name: Name of the consumer task

Returns:

Consumption status tensor if partition exists, None otherwise

"""

partition = self._get_partition(partition_id)

if not partition:

return None

return partition.get_consumption_status(task_name)

def get_production_status(self, partition_id: str, data_fields: list[str]) -> bool:

"""

Check if all samples for specified fields are fully produced in a partition.

Args:

partition_id: ID of the partition

data_fields: List of field names to check production status for

Returns:

bool: True if all samples have been produced for all specified fields, False otherwise

"""

partition = self._get_partition(partition_id)

if not partition:

return False

return partition.get_production_status_for_fields(data_fields)

def get_metadata(

self,

data_fields: list[str],

partition_id: str,

mode: str = "fetch",

task_name: str | None = None,

batch_size: int | None = None,

sampling_config: Optional[dict[str, Any]] = None,

*args,

**kwargs,

) -> BatchMeta:

"""

Retrieve metadata with support for three modes.

Args:

data_fields: List of field names to include in metadata

partition_id: Partition id for which to retrieve metadata

mode: Operation mode - 'insert', 'fetch', or 'force_fetch'

- mode="insert": Create metadata for new samples (for data insertion)

- mode="fetch": Get metadata from ready samples using the configured sampler

- mode="force_fetch": Get metadata for unconsumed samples without sampling

(excludes already consumed samples)

task_name: Name of the consumer task (required for fetch modes)

batch_size: Number of samples to retrieve

*args: Additional positional arguments

**kwargs: Additional keyword arguments

Returns:

BatchMeta object containing the requested metadata

Raises:

TimeoutError: If waiting for sufficient data times out in fetch mode

"""

if partition_id not in self.partitions:

self.create_partition(partition_id)

if mode == "insert":

if data_fields:

# First put_data call, get_metadata in insert mode

batch_global_indexes = self.index_manager.allocate_indexes(partition_id, count=batch_size)

else:

# clear metadata call passes empty data_fields

batch_global_indexes = self.index_manager.get_indexes_for_partition(partition_id)

return self.generate_batch_meta(partition_id, batch_global_indexes, data_fields, mode)

assert task_name is not None

if mode == "fetch":

# Find ready samples within current data partition and package into BatchMeta when reading

start_time = time.time()

while True:

# ready_for_consume_indexes: samples where all required fields are produced

# (production status is ready) and not yet consumed

ready_for_consume_indexes = self.scan_data_status(partition_id, data_fields, task_name)

if len(ready_for_consume_indexes) < batch_size:

if self.polling_mode:

logger.debug(

f"[{self.controller_id}]: Not enough data for task {task_name} in partition {partition_id}."

f" Required: {batch_size}, Available: {len(ready_for_consume_indexes)}."

f" Returning None due to polling mode."

)

return BatchMeta.empty()

if time.time() - start_time > TQ_CONTROLLER_GET_METADATA_TIMEOUT:

raise TimeoutError(

f"Timeout while waiting for sufficient data for task {task_name}. "

f"Required: {batch_size}, Available: {len(ready_for_consume_indexes)}"

)

logger.warning(

f"[{self.controller_id}]: Insufficient data for task {task_name}. Required: {batch_size} "

f"samples with fields {data_fields} in partition {partition_id}, but only have "

f"{len(ready_for_consume_indexes)} samples meeting the criteria. "

f"Retrying in {TQ_CONTROLLER_GET_METADATA_CHECK_INTERVAL}s..."

)

time.sleep(TQ_CONTROLLER_GET_METADATA_CHECK_INTERVAL)

else:

break

batch_global_indexes, consumed_indexes = self.sampler(

ready_for_consume_indexes,

batch_size,

**(sampling_config or {}),

)

# Check if we got valid results from the sampler

if len(batch_global_indexes) != batch_size:

raise RuntimeError(

f"Sampler returned insufficient samples. Please check the sampler logic. "

f"Expected: {batch_size}, before sampling: {len(ready_for_consume_indexes)}, "

f"after sampling: {len(batch_global_indexes)}"

)

elif mode == "force_fetch":

global_indexes_range = self.index_manager.get_indexes_for_partition(partition_id)

consumer_status = self.get_consumption_status(partition_id, task_name)

not_consumed_idx = [i for i in global_indexes_range if consumer_status[i] == 0]

batch_global_indexes = not_consumed_idx

consumed_indexes = []

# Package into metadata

metadata = self.generate_batch_meta(partition_id, batch_global_indexes, data_fields, mode)

# Mark samples as consumed if in fetch mode

if mode == "fetch" and consumed_indexes:

partition = self.partitions[partition_id]

partition.mark_consumed(task_name, consumed_indexes)

return metadata

def scan_data_status(

self,

partition_id: str,

data_fields: list[str],

task_name: str,

) -> list[int]:

"""

Find samples that are ready for consumption in a specific partition.

Delegates scanning functionality to the partition's own method.

Args:

partition_id: ID of the partition

data_fields: List of required field names

task_name: Name of the consumer task

Returns:

List of global indices that are ready for consumption

"""

partition = self._get_partition(partition_id)

if not partition:

return []

# Use partition's own scanning method

ready_sample_indices = partition.scan_data_status(data_fields, task_name)

return ready_sample_indices

# ==================== Metadata Generation API ====================

def generate_batch_meta(

self,

partition_id: str,

View remainder of file in raw view

TransferQueue/transfer_queue/controller.py at dev · TransferQueue/TransferQueue · GitHub

Uh oh!

Expand file tree

Breadcrumbs

controller.py

Latest commit

History

Breadcrumbs

controller.py

File metadata and controls