v1.3.0

Major Features and Improvements

Train/Eval/Predict/Export

• AOTInductor
  • Optimize AOTI performance with export knobs for fp32 unbacked floats, sample-input autotune, and TF32 from export config in #498
  • AOTInductor support match models in #509
• Cache feature-permute order during export to avoid per-forward H2D sync in #527
• Support keep_checkpoint_max with async checkpoint pruning in #528
• [WIP] Support model delta tracker to track incremental embedding rows in #546

Model

• HSTUMatch:
  • Optimize HSTUMatch with STUStack, UIHPreprocessor, and block-suffix candidates to support efficient triton hstu ops in #505 #506 #510 #519 #520 #526
  • Support scalar item-tower export view for HSTUMatch in #518
• DlrmHSTU / UltraHSTU
  • Support stu.scaling_seqlen and drop autotune assert strip in #500
  • Thread contextual_seq_len from preprocessor to STULayer with proto sentinel, truncation total_uih_len, and AOTI-friendly SLA builder in #501
  • Bundle a repackaged Triton 3.7.1 wheel with ptxas 12.9.86 to fix the HSTU WGMMA shared-memory OOB on sm_90/H20 in #551
• SID:
  • Add SidRqkmeans model with FAISS-trained residual K-Means in #538 #539
  • Add SidRqvae model with STE/Gumbel-Softmax and CLIP support in #545 #552

Embedding

• Add FeatureGroupConfig.embedding_name_suffix to break embedding sharing across groups in #504
• Add planner support to pick DynamicEmbedding HYBRID vs CACHING from topology budgets in #508
• Add zch_to_dynamicemb_convert tool to migrate ZCH checkpoints to DynamicEmbedding in #511

Dataset

• Add Kafka event-time driven checkpointing from message timestamp in #541
• Support dataloader checkpoint resume across forked workers and epochs in #544
• Batch and retry Kafka offsets_for_times to survive broker timeouts in #547

Optimizer and Metrics

• Add Adadelta and RMSprop sparse and dense optimizers in #499
• Add NormalizedEntropy metric for binary classification in #507

Upgrade

• Upgrade PyTorch to v2.12.1, TorchRec to v1.7.0, FBGEMM to v1.7.0, support NumPy 2, and add CUDA 13.0 TensorRT image support in #551
• Allow pandas >= 3 in #556

Note

For TorchEasyRec 1.3.x, you should use Docker image version 1.3.

• For the GPU version (CUDA 13.0) with tensorrt:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.3-cu130
  • PyTorch: v2.12.1 CUDA: v13.0 FBGEMM: v1.7.0 TorchRec: v1.7.0 TensorRT: v10.16 Python: v3.11
  • Supported GPUs: sm_75 / 80 / 86 / 90 / 100 / 120, It supports Turing (T4), Ampere/Ada (A10/A30/A100/L4/L20), Hopper (H100/H200), Blackwell (B100/B200), and other GPUs with CC 7.5-12.0.
  • Driver requirement: NVIDIA driver R580 or later is recommended. For older drivers, set LD_LIBRARY_PATH=/usr/local/cuda-13.0/compat.
• For the GPU version (CUDA 12.9):
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.3-cu129
  • PyTorch: v2.12.1 CUDA: v12.9 FBGEMM: v1.7.0 TorchRec: v1.7.0 Python: v3.11
  • Supported GPUs: sm_75 / 80 / 86 / 90 / 100 / 120. It supports Turing (T4), Ampere/Ada (A10/A30/A100/L4/L20), Hopper (H100/H200), Blackwell (B100/B200), and other GPUs with CC 7.5-12.0.
• For the GPU version (CUDA 12.6):
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.3-cu126
  • PyTorch: v2.12.1 CUDA: v12.6 FBGEMM: v1.7.0 TorchRec: v1.7.0 Python: v3.11
  • Supported GPUs: sm_60 / 70 / 75 / 80 / 86 / 90. It supports Pascal (P100), Volta (V100), Turing (T4), Ampere/Ada (A10/A30/A100/L4/L20), Hopper (H100), and other GPUs with CC 6.0-9.0. It does not support Blackwell GPUs.
• For the CPU version:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.3-cpu
  • PyTorch: v2.12.1 FBGEMM: v1.7.0 TorchRec: v1.7.0 Python: v3.11

Bug Fixes and Other Changes

• [bugfix] update dynemb doc by @eric-gecheng in #496
• [bugfix] add cc<7.0 dist.barrier() at TrainPipelineSparseDist.progress() entry by @tiankongdeguiji in #497
• [ci] H20 unit test lane: workflow, Hopper-specific fixes, --scope filter by @tiankongdeguiji in #494
• [bugfix] replace deprecated torch._check_is_size with torch._check by @tiankongdeguiji in #503
• [bugfix] aoti: backport pytorch/pytorch#178147 int_array dedup key by @tiankongdeguiji in #512
• [fix] sampler: emit scalar negs for sequence-positive item-side fields by @tiankongdeguiji in #514
• [bugfix] code_review.yml: replace PR's .claude/ instead of nesting (GHSA-f9x3-9rgg-92p7) by @tiankongdeguiji in #515
• [chore] bump dynamicemb to 0.1.0+20260519.e0c1fbb by @tiankongdeguiji in #516
• fix(pepnet): fix domain output overwrite in _select_domain_task_output by @zhpjunfei in #524
• [fix] H20 CI: cutlass WGMMA bwd + unified-AOT int_array races; add test-skip infra by @tiankongdeguiji in #531
• [ci][feat] PPU: add unittest CI lane + framework adaptations + DLC/ODPS tutorial by @tiankongdeguiji in #529
• [chore] standardize OSS host on oss-accelerate by @tiankongdeguiji in #533
• [bugfix] RTP export: keep dense-embedding modules as FX leaves so AutoDis/MLP params restore by @tiankongdeguiji in #534
• [doc] PAI deployment docs: ENABLE_AOT export + serving processor params by @tiankongdeguiji in #536
• [bugfix] ODPS reader: pin pyodps==0.12.5.1 (retry connection resets) + log session/table context on errors by @tiankongdeguiji in #537
• [bugfix] fix flaky TRT test: align engine TF32 with baseline, run strict FP32 by @tiankongdeguiji in #535
• [bugfix] DlrmHSTU predict: keep candidate split key in original order for descending timestamps by @tiankongdeguiji in #540
• [doc] fix grouped sequence_fields default rule; drop epoch/wall-clock note from train.md by @tiankongdeguiji in #542
• [ci] code review: disable Remote Control in headless Claude run by @tiankongdeguiji in #548
• [bugfix] fix flaky TRT predict test by making the mock model confident by @tiankongdeguiji in #549
• [ci] resolve externals symlink for DSW authZ mount allowlist by @tiankongdeguiji in #550
• [ci] cut GPU + H20 unit-test CI toward 1 hour (cache persistence, derandomize, GPU scoping) by @tiankongdeguiji in #553
• [bugfix] fail fast when graphlearn sampler server dies by @tiankongdeguiji in #554

New Contributors

• @zhpjunfei made their first contribution in #524
• @WhiteSwan1 made their first contribution in #538

Full Changelog: v1.2.0...v1.3.0

v1.2.0

Major Features and Improvements

Train/Eval/Predict/Export

• Enhance HSTU export in #443
• Support unified one-stage AOTI export with torch.export compatibility fixes in #475
• Support generic --additional_export_config JSON for export in #481
• Reduce AOTI compile memory usage by releasing verify-forward activations before compile in #491

Model

• DlrmHSTU:
  • Add CUTLASS kernel backend for HSTU attention in #465
  • Add concat_contextual_features option in #459
  • Support scaling_seqlen in HSTU attention stack in #480
  • Support per-task loss weight in FusionSubTaskConfig in #453
• ULTRA-HSTU:
  • Add Semi-Local Attention and selective activation rematerialization in #486
  • Add mid-stack attention truncation in #488
  • Add Mixture of Transducers in #492
• Add label smoothing support to BinaryCrossEntropy loss in #455

Embedding

• Update DynamicEmbedding to use align_to_table_size in #460
• Integrate DynamicEmbedding table fusion in #466

Feature

• Add CombineFeature support in #447
• Support TokenizeFeature as token-level sequence input in #470

Dataset

• Add start.timestamp.ms support to KafkaDataset in #446
• Add heartbeat thread to prevent Kafka MAX_POLL_EXCEEDED in #471

Optimizer

• Add CosineAnnealingLR and CosineAnnealingWarmRestartsLR schedules in #454

Upgrade

• Upgrade PyTorch to v2.11, TorchRec to v1.6.0, and FBGEMM to v1.6.0 in #479

Note

For TorchEasyRec 1.2.x, you should use Docker image version 1.2.

• For the GPU version (CUDA 12.9) with tensorrt:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.2-cu129
  • PyTorch: v2.11 CUDA: v12.9 FBGEMM: v1.6.0 TorchRec: v1.6.0 Python: v3.11
  • Supported GPUs: sm_75 / 80 / 86 / 90 / 100 / 120. It supports Turing (T4), Ampere/Ada (A10/A30/A100/L4/L20), Hopper (H100/H200/H20), Blackwell (B100/B200), and other GPUs with CC 7.5-12.0.
• For the GPU version (CUDA 12.6) with tensorrt:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.2-cu126
  • PyTorch: v2.11 CUDA: v12.6 FBGEMM: v1.6.0 TorchRec: v1.6.0 Python: v3.11
  • Supported GPUs: sm_70 / 75 / 80 / 86 / 90. It supports Volta (V100), Turing (T4), Ampere/Ada (A10/A30/A100/L4/L20), Hopper (H100/H20), and other GPUs with CC 7.0-9.0. It does not support Blackwell GPUs.
• For the CPU version:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.2-cpu
  • PyTorch: v2.11 FBGEMM: v1.6.0 TorchRec: v1.6.0 Python: v3.11

Bug Fixes and Other Changes

• [bugfix] fix readthedocs build failure by @tiankongdeguiji in #439
• [bugfix] fix list-to-integer comparison in embedding sequence encoder validation by @tiankongdeguiji in #429
• [bugfix] remove redundant .data access in pe_mtl_loss by @tiankongdeguiji in #430
• [bugfix] clarify sample_weight fallback value in match and rank models by @tiankongdeguiji in #435
• [feat] bump up pyfg to 1.0.2 by @tiankongdeguiji in #427
• [bugfix] fix data_config mutation during model export by @tiankongdeguiji in #441
• [feat] replace claude-code-action with direct claude -p for code review by @tiankongdeguiji in #444
• [bugfix] rename loop variable to avoid shadowing builtin input() by @tiankongdeguiji in #438
• [bugfix] replace deprecated torch.autograd.Variable in optimizer test by @tiankongdeguiji in #437
• [bugfix] fix unclosed file handle in benchmark by @tiankongdeguiji in #431
• [bugfix] fix potential socket resource leak in get_free_port by @tiankongdeguiji in #432
• [bugfix] strengthen doc reviewer to cross-reference existing user-facing docs by @tiankongdeguiji in #449
• [bugfix] fix contextual_seq_len not passed from preprocessor to STULayer by @tiankongdeguiji in #450
• filter non grad when adding to summaries by @eric-gecheng in #448
• [bugfix] fix flaky TRT test by adding allow_tf32 to predict() by @tiankongdeguiji in #456
• [bugfix] suppress false-positive range validation warnings for dynamicemb features by @tiankongdeguiji in #458
• [bugfix] fix sequence feature default_value inconsistency by @tiankongdeguiji in #461
• [docs] add FAQ for Triton v3.6.0 WGMMA crash on Hopper GPUs by @tiankongdeguiji in #452
• [bugfix] ensure predict threads are joined on exception by @tiankongdeguiji in #433
• [bugfix] fix ZCH finetune from checkpoint with different world size by @tiankongdeguiji in #467
• [bugfix] accept ChunkedArray in Parquet/Odps/Csv writers and ensure TDM writer close by @tiankongdeguiji in #469
• [bugfix] fix NameError on sampled when TDMSampler is combined with sample_mask by @tiankongdeguiji in #468
• [doc] fix dynamicemb pip install command by @tiankongdeguiji in #473
• [bugfix] fix fbgemm int32 overflow during embedding quantization by @tiankongdeguiji in #472
• [chore] bump pyfg to 1.0.4 by @tiankongdeguiji in #482
• [bugfix] fix two-stage AOTI predict hang under multi-thread workers by @tiankongdeguiji in #484
• [bugfix] share Dim across grouped-sequence tensors in legacy AOT export by @tiankongdeguiji in #485
• [bugfix] add CombineFeature to SINGLE_INPUT_FEATURE_CLASSES by @tiankongdeguiji in #487
• [bump] pyfg 1.0.4 -> 1.0.5; doc updates and TokenizeFeature fix by @tiankongdeguiji in #489

Full Changelog: v1.1.0...v1.2.0

v1.1.0

Major Features and Improvements

Train/Eval/Predict/Export

• Support grad clipping for dense params in #424
• Refactor AOTInductor export with split model in #394
• Support env configs for reproducibility in #361
• Support ignore restore optimizer option for train_eval in #389

Model

• Add PE-LTR model in #381
• Add WuKong model in #372
• Add PEPNet model in #402
• Improve DlrmHSTU model in #352 #395 #393 #359
  • support num_class > 1
  • support descending order sequence
  • support jagged label
  • support time_bucket_increments in PositionEncoder
• DLRM and WuKong model support only one sparse group in #385

Embedding

• Add AdmissionStrategy support for DynamicEmbedding in #362
• Add storage estimate for dynamic embedding kv counter in #391

Feature

• Support sequence cross features in #375
• Support convert compatible feature configs of EasyRec in #392

Dataset

• Add Kafka dataset with checkpoint support in #401 #408 #413
• Add checkpointable Parquet dataset in #410
• Add checkpointable ODPS dataset in #409
• Support input fields str in #412

Optimizer

• Support initial_accumulator_value for FusedSparseAdagradOptimizer & add additional optimizer configuration options in #382

Upgrade

• Upgrade torchrec to v1.5.0 in #405

Note

For TorchEasyRec 1.1.x, you should use Docker image version 1.1.

• For the GPU version (CUDA 12.9) with tensorrt:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.1-cu129
  • PyTorch: v2.10 CUDA: v12.9 FBGEMM: v1.5.0 TorchRec: v1.5.0 Python: v3.11
• For the GPU version (CUDA 12.6) w/o tensorrt:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.1-cu126
  • PyTorch: v2.10 CUDA: v12.6 FBGEMM: v1.5.0 TorchRec: v1.5.0 Python: v3.11
• For the CPU version:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.1-cpu
  • PyTorch: v2.10 FBGEMM: v1.5.0 TorchRec: v1.5.0 Python: v3.11

Bug Fixes and Other Changes

• [bugfix] fix session_status check error of OdpsWriter by @tiankongdeguiji in #351
• [feat] add MTGR style DlrmHSTU config doc by @tiankongdeguiji in #353
• [bugfix] sampler add raise parse kv error by @chengaofei in #355
• [bugfix] make fsspec disable by default by @tiankongdeguiji in #356
• [feat] add faq doc - The "kv" feature key contains ":" character by @yanzhen1233 in #357
• [feat] bump up pyfg to 1.0.0 by @tiankongdeguiji in #363
• [feat] update local tutorial tdm doc with FG_DAG mode by @asdfasdfsdfas in #365
• [bugfix] fix input names of custom sequence feature when use pyfg 1.0.0 by @tiankongdeguiji in #368
• [feat] update dlc tutorial doc with oss mount by @asdfasdfsdfas in #373
• [bugfix] fix dynamicemb is_sparse for custom/lookup/match feature by @tiankongdeguiji in #376
• [bugfix] fix torch.full error of apply_split_helper for uvm embedding kernel by @tiankongdeguiji in #383
• [feat] use cpu npoc_per_node_is_1 by @chengaofei in #384
• [bugfix] use KVCounter initialization fix of dynamicemb by @tiankongdeguiji in #387
• [feat] add warning for default_value in vocab_list or vocab_dict by @tiankongdeguiji in #388
• [bugfix] make sequence related config optional by @tiankongdeguiji in #390
• [bugfix] fix occasional failure in test_add_timestamp_positional_embeddings_triton by @tiankongdeguiji in #396
• [feat] update custom feature doc by @tiankongdeguiji in #400
• [feat] add odps test quota & fix ecs ram role in test by @tiankongdeguiji in #403
• [bugfix] fix ecs ram role timeout in benchmark by @tiankongdeguiji in #411
• [feat] add ai code review by @tiankongdeguiji in #414
• [feat] refine readme and update doc with new features by @tiankongdeguiji in #416
• [bugfix] fix introduction doc & refine kafka dataset doc by @tiankongdeguiji in #417
• [bugfix] fix search tool hang in readthedoc and doc build warnings by @tiankongdeguiji in #418
• [bugfix] fix kafka dataset test with embed schema by @tiankongdeguiji in #419
• [bugfix] correct typo TRAGET_REPEAT_INTERLEAVE_KEY -> TARGET_REPEAT_INTERLEAVE_KEY by @hobostay in #420
• [feat] add ArrowRecordBatch flink udf usage doc by @tiankongdeguiji in #421
• [feat] unify proto configs w/o using colon by @tiankongdeguiji in #423
• [bugfix] fix ecs ram role error log by @tiankongdeguiji in #426
• [bugfix] fix INPUT_TILE_3_ONLINE=1 and add docs by @tiankongdeguiji in #428

New Contributors

• @asdfasdfsdfas made their first contribution in #365
• @hobostay made their first contribution in #420

Full Changelog: v1.0.0...v1.1.0

v1.0.0

Major Features and Improvements

Train/Eval/Predict/Export

• Support training with dynamic batch size by sample cost in #343
• Support logging train metrics in #310
• Support predicting checkpoint in #320 #322 #324
• [EXPERIMENTAL] Support exporting with AOTInductor in #239 #274
• Support exporting with TensorRT in #318
• Support exporting the best model in #294
• Support exporting to RTP in #298 #307 #329 #332 #339
• Support AdamW optimizer and label smoothing in #297
• Support setting an optimizer for a subset of parameters in #297
• Support PanguDFS in #311 #348 #349 #350

Embedding

• Support dynamic embedding in #279 #281 #283 #286 #289 #316
• Support initialize dynamic embeddings from tables in #282 #288
• MLPEmbedding support feature value_dim > 1 in #331

Model

• Optimize and refactor DlrmHSTU preprocessor to support MTGR style preprocessing in #290 #296 #300 #314
• Decouple contextual feature dimension from sequence id embedding dimension in DlrmHSTU in #302
• DlrmHSTU support uih and contextual share embedding in #337
• DlrmHSTU support global average loss option in #334
• Add TMA support for hstu attn in #336
• Optimize gpu memory usage of GAUC metric in #312

Feature

• Support kv dot product feature in #276
• Support bool mask feature in #285
• Support farm hash in #295

Upgrade

• Upgrade pytorch to v2.9 and torchrec to v1.4.0 in #345

Note

For TorchEasyRec 1.0.x, you should use Docker image version 1.0.

• For the GPU version (CUDA 12.6):
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.0-cu126
  • PyTorch: v2.9 CUDA: v12.6 FBGEMM: v1.4.0 TorchRec: v1.4.0 Python: v3.11
  • We drop support for the 470 GPU driver version. If you still want to use the 470 GPU driver version, you can set LD_LIBRARY_PATH=/usr/local/cuda-12.6/compat
• For the CPU version:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:1.0-cpu
  • PyTorch: v2.9 FBGEMM: v1.4.0 TorchRec: v1.4.0 Python: v3.11

Bug Fixes and Other Changes

• [feat] make bash as default shell by @tiankongdeguiji in #273
• [feat] add benchmark odps quota and skip trt test when trt not avaiable by @tiankongdeguiji in #278
• [feat] add rdma addons into dockerfile by @tiankongdeguiji in #280
• [feat] clean up fg_encoded in docs by @tiankongdeguiji in #287
• support create tzrec config based on pyfg json by @chengaofei in #284
• [bugfix] fix finetune checkpoint path runtime error print when path not exist by @tiankongdeguiji in #291
• [feat] refactor export model by @tiankongdeguiji in #293
• fix sequence raw feature pyfg sub_type not effective by @chengaofei in #292
• [feat] optimize hstu triton op warning by @tiankongdeguiji in #301
• [bugfix] improve create init ckpt for dynamic embedding when certain id_feature in the config lack embedding_dim by @tiankongdeguiji in #303
• [feat] bump up tzrec version to 0.9.7 by @tiankongdeguiji in #305
• [bugfix] fix dlrm hstu gauc and l2 loss support by @tiankongdeguiji in #306
• [bugfix] fix content encoder with additional_content_features and target_enrich_features by @tiankongdeguiji in #304
• [bugfix] fix create dynamic embedding ckpt when raw feature in config by @tiankongdeguiji in #308
• Add evaluation metrics documentation by @yanzhen1233 in #309
• [bugfix] fix fsspec ci test by @tiankongdeguiji in #317
• add train_metric docs by @chengaofei in #315
• Update custom development model documentation by @yanzhen1233 in #313
• [feat] Adapt integration test config to local cuda device count by @eric-gecheng in #319
• [bugfix] fix dlrm hstu preprocessor doc by @tiankongdeguiji in #321
• [feat] add dlrm hstu demo data by @tiankongdeguiji in #326
• [bugfix] avoid jit convert error when using large number by @eric-gecheng in #328
• [feat] improve prune_unused_param_and_buffer when export model by @tiankongdeguiji in #327
• [feat] refactor ops directory to fix import triton error by @tiankongdeguiji in #335
• [feat]add assert to avoid using ckpt predict for two tower models by @eric-gecheng in #333
• [feat] upgrade 2025 dingtalk qrcode by @tiankongdeguiji in #340
• [bugfix] always lazy init predict checkpoint writer by @tiankongdeguiji in #341
• Feature/dynamic routing support zero init by @eric-gecheng in #342
• [bugfix] fix parse batch empty MapArray error in NegativeSampler by @tiankongdeguiji in #344
• [feat] add doc for dynamic batch by @tiankongdeguiji in #346
• [bugfix] fix array type in feature doc by @tiankongdeguiji in #347

Full Changelog: v0.9.0...v1.0.0

v0.9.0

Major Features and Improvements

Train/Eval/Export

• Support get null value for int/float dtype features when use negative sampler #202
• Support freeze embedding parameters #206
• Add mixed_precision bf16/fp16 and gradient accumulation support #220
• Add fp16 embedding dtype support #221
• Add TrainPipelineBase to support model w/o sparse parameters #222
• Add EmbeddingCollection quant support #265
• Optimize sequence emb inference speed #266

Model

• Add DlrmHSTU model #224 #227 #231 #232 #237 #250 #257
• Add DCN_V1 model #235
• Add DCN_V2 and xDeepFM model #242
• Add WideAndDeep model and wide init_fn #212
• Add sequence self_attention encoder #251
• Add binary focal loss #208
• Add xauc and grouped xauc #252
• Add feature selection for DSSM_V2 #219
• Add use_ln option for MLP module #223

Feature

• Add const input for feature #210
• Expr feature support value_dim #216
• Support feature only used as fg dag intermediate result (stub_type=true) #218

Dataset

• Sampler support odps schema #267

Upgrade

• Upgrade pytorch to v2.8 and torchrec to v1.3.0 #241

Python

• Support py312 #271 #272

Note

For TorchEasyRec 0.9.x, you should use Docker image version 0.9.

• For the GPU version (CUDA 12.6):
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:0.9-cu126
  • PyTorch: v2.8 CUDA: v12.6 FBGEMM: v1.3.0 TorchRec: v1.3.0 Python: v3.11
  • We drop support for the 470 GPU driver version. If you still want to use the 470 GPU driver version, you can set LD_LIBRARY_PATH=/usr/local/cuda-12.6/compat
• For the CPU version:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:0.9-cpu
  • PyTorch: v2.8 FBGEMM: v1.3.0 TorchRec: v1.3.0 Python: v3.11

Bug Fixes and Other Changes

• [feat] prune mem of one shard > mem of one device for DynamicProgrammingProposer by @tiankongdeguiji in #194
• [feat] use oss accelerate endpoint by @tiankongdeguiji in #203
• [bugfix] fix force_base_data_group when export model by @tiankongdeguiji in #200
• [bugfix] fix int32 and double column type of negative sampler table by @tiankongdeguiji in #204
• [bugfix] remove redundant print in sampler by @tiankongdeguiji in #207
• [bugfix] fix hitrate hang on OdpsWriter & refactor broadcast_object and gather_object pg by @tiankongdeguiji in #205
• [feat] upgrade pyfg to 0.6.9 and refine expr/overlap feature doc by @tiankongdeguiji in #199
• create fg.json if exist will error by @chengaofei in #211
• [bugfix] add missing wide_and_deep doc index by @tiankongdeguiji in #214
• Feature/fix dense embedding export in dssmv2 by @eric-gecheng in #213
• [bugfix] support remove bucketizer for sequence feature and add tests by @tiankongdeguiji in #215
• [feat] add error and warning for restore_model when model path not exists by @tiankongdeguiji in #217
• [bugfix] prevent unittest nightly timeout by @tiankongdeguiji in #225
• [bugfix] fix tzrec optimizer not update params by @tiankongdeguiji in #226
• [bugfix] fix pyfg oss accelerate url by @tiankongdeguiji in #228
• rocket launching train failed by @chengaofei in #229
• [bugfix] add ops init py & add build wheel ci test by @tiankongdeguiji in #234
• increase dlrm and rocket_launching benchmark by @chengaofei in #233
• [bugfix] fix value cannot be converted to type int32 without overflow in trt test by @tiankongdeguiji in #243
• [feat] increase benchmark timeout by @tiankongdeguiji in #244
• [bugfix] fix string id support for tdm sampler by @tiankongdeguiji in #245
• [bugfix] fix tdm user defined attr delim and optimize attrs of TDMSampler by @tiankongdeguiji in #248
• [bugfix] fix HardNegativeSampler with string id by @tiankongdeguiji in #249
• [bugfix] fix pyre check by @tiankongdeguiji in #246
• [bugfix] fix save checkpoint at epoch 0 when save_checkpoint_epochs > 1 by @tiankongdeguiji in #253
• [bugfix] refine distinguish sparse module in create_train_pipeline to fix continue training failure by @tiankongdeguiji in #254
• [bugfix] fix tdm retrieval nccl hang when use odps writer by @tiankongdeguiji in #255
• [feat] refactor ec_list to ec_dict for TDMEmbedding export by @tiankongdeguiji in #258
• [bugfix] fix hard negative sampler with zero hard negative indices by @tiankongdeguiji in #259
• [bugfix] fix sampler with string id memory leak by @tiankongdeguiji in #260
• [feat] bump up pyfg to 0.7.3 by @tiankongdeguiji in #263
• support odps three schema by @chengaofei in #264
• [bugfix]nccl timeout by @eric-gecheng in #262
• [bugfix] fix quant ec doc by @tiankongdeguiji in #269
• [bugfix] fix odps dataset test by @tiankongdeguiji in #270

Full Changelog: v0.8.0...v0.9.0

v0.8.0

Major Features and Improvements

Train/Eval/Export

• Support eval and save checkpoint by epoch #116
• Support export fp32/fp16/int8/int4/int2 ebc embedding quant model #137
• Enhance export efficiency by restoring state dict directly instead of copying and gathering #177
• Add faiss gpu support for evaluation #170
• Enhance optimizer state loading for changed plans with plan checkpoint #185
• Support tensorboard log for model parameters #181
• Add restore ckpt check for continue train #180
• Add allow_tf32 flag and global embedding param constraint #188

Model

• Add MIND model #119 #123 #157 #172
• Add RocketLaunching model #129
• Add DLRM model #148
• Add MaskNet #179 #187
• Add dice activation and support bn for sequence mlp #107
• Add regression and multi-classification metric #149
• Optimize distributed GAUC memory use #127
• Add SequenceEmbeddingGroup and support jagged forward #152
• Support max sequence length setting for sequence encoder #184
• Support hard negative sampler #195
• Optimize HSTU training and sampling process and add triton ops (WIP) #93 #154

Feature

• Support custom feature and custom sequence feature #144
• Weighted id feature support map dtype #190
• Dump parsed inputs support weighted id and multi-value sequence feature #191

Dataset

• Support dataset shuffle #114
• Optimize performance of ParquetDataset and Rebalance parquet files dynamically #125 #126
• Add odps read session refresh to extend odps session expired time #132
• Add more alibaba cloud credentials for odps dataset #115
• Add odps_data_compression (ZSTD) config for OdpsDataset #146
• Always lazy init odps writer #178

Upgrade

• Upgrade pytorch to v2.7 and torchrec to v1.2.0 #197

Note

For TorchEasyRec 0.8.x, you should use Docker image version 0.8.

• For the GPU version (CUDA 12.6):
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:0.8-cu126
  • PyTorch: v2.7 CUDA: v12.6 FBGEMM: v1.2.0 TorchRec: v1.2.0 Python: v3.11
  • We drop support for the 470 GPU driver version. If you still want to use the 470 GPU driver version, you can set LD_LIBRARY_PATH=/usr/local/cuda-12.6/compat
• For the CPU version:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:0.8-cpu
  • PyTorch: v2.7 FBGEMM: v1.2.0 TorchRec: v1.2.0 Python: v3.11

Bug Fixes and Other Changes

• [bugfix] fix cpu docker image build without trt by @tiankongdeguiji in #100
• add_dssm_recall_benchmark by @chengaofei in #101
• [feat] support ignore unused features in negative sampler by @tiankongdeguiji in #102
• [bugfix] fix multi-val sequence embedding nan when pooling_type = mean by @tiankongdeguiji in #104
• [feat] upgrade ruff to 2025 code style by @tiankongdeguiji in #105
• [bugfix] fix correctness of kjt.lengths when ShardedEmbeddingBag’s pooling_type is mean and shard_type is row_wise by @tiankongdeguiji in #106
• [feat] upload feature assets to odps and fix remove_bucketizer in create_fg_json by @tiankongdeguiji in #103
• [bugfix] fix multi-value sequence raw feature by @tiankongdeguiji in #109
• [bugfix] fix dice init params by @tiankongdeguiji in #108
• [feat] docker support rtx gpu by @tiankongdeguiji in #111
• [bugfix] fix create_fg_json invalid option force_update_resource by @chengaofei in #110
• [feat] clean fg_encoded config by @tiankongdeguiji in #112
• [bugfix] prevent redundant file uploading to odps when use create_fg_json by @tiankongdeguiji in #113
• [bugfix] fix modify feature group config in training by @chengaofei in #118
• [feat] refactor weighted id feature with pyfg 0.4.5 encoded format by @tiankongdeguiji in #117
• [bugfix] fix feature.keys() none error when all features in embedding group are zch by @tiankongdeguiji in #120
• [feat] refine pyarrow type to odps table type convert by @tiankongdeguiji in #121
• [bugfix] bump up pyfg version 0.4.8 to fix sequence_length in config < true sequence length in data by @tiankongdeguiji in #130
• Support non null string list by @yanzhen1233 in #128
• [bugfix] fix mc-abc divisor none error when use mean pooling by @tiankongdeguiji in #133
• [bugfix] fix feature permute when use mc-ebc and mean pooling by @tiankongdeguiji in #134
• support feature_groups select features by @chengaofei in #135
• [bugfix] fix export model with zch by @tiankongdeguiji in #136
• bugfix_export_input_tile_is_2 by @chengaofei in #138
• fix the feature bug: has_dag by @yjjinjie in #140
• [bugfix] add missing dataset utils test by @tiankongdeguiji in #139
• [bugfix] add ArrowInvalid retry for refresh odps session by @tiankongdeguiji in #141
• [feat] remove redundant side_inputs warn when fg_mode=FG_NONE by @tiankongdeguiji in #142
• [bugfix]fix autodis parameter init in dist mode by @eric-gecheng in #143
• [bugfix] fix mlp embedding param init bug by @eric-gecheng in #145
• [bugfix] revert emb_impl call by @tiankongdeguiji in #147
• [bugfix] fix ple typo by @tiankongdeguiji in #150
• [feat] add kernel config and BaseModule by @tiankongdeguiji in #151
• [feat] refactor label_name to label tensor in loss and metric impl by @tiankongdeguiji in #153
• [feat] update maxcompute vpc endpoint and quota doc by @tiankongdeguiji in #155
• [feat] add auto rebalance doc for ParquetDataset by @tiankongdeguiji in #156
• [feat] add odps dataset ci test by @tiankongdeguiji in #159
• [feat] add nightly build wheel and doc by @tiankongdeguiji in #160
• [feat] add benchmark and nightly test by @tiankongdeguiji in #161
• [bugfix] fix build nightly wheel by @tiankongdeguiji in #162
• [bugfix] fix regression metric by @tiankongdeguiji in #163
• [bugfix] fix fork repo cpu ci by @tiankongdeguiji in #166
• fix loop logic in hitrate.py by @eric-gecheng in #165
• [bugfix] fix combo feature value and length mismatch when input data with only one separator by @tiankongdeguiji in #168
• [bugfix] fix mtl weight always equal to 1 after div by mean by @tiankongdeguiji in #171
• [feat] optimze dssm and mtl with weight benchmark by @tiankongdeguiji in #173
• [bugfix] fix clear_variational_dropout and visualize flag of feature selection by @tiankongdeguiji in #174
• [feat] add fg value_type config and make num_buckets default value_dtype as string by @tiankongdeguiji in https://github.com/alibaba/...

v0.7.0

Major Features and Improvements

Train/Eval/Export

• Support train/eval/export on cpu #27
• Support TRT export (Beta) #30 #32 #41 #43 #58 #59 #89
• Support AOT export (WIP) #79

Model

• Optimize TDM gen tree speed #33
• TDM Support string id #72
• Rank and Match models support sample weight #50 #57 #63 #65
• Add zero collision hash embedding #60
• Add intervention methods for multi-target learning #49
• Add Autodis and MLP embedding for raw features #73 #75
• Add task space for multi-target learning loss #82
• Add dual augmented two-tower match model #83
• Add HSTU (WIP) #55

Feature

• pyfg support CPU without avx512 #20
• ExprFeature support l2_norm|dot|euclid_dist #35
• Add fg bucketize only mode & refactor fg_encoded to fg_mode #62
• Make default bucketize value configurable #94
• Support multi-value sequence #96
• Support vocab file #97

Dataset

• Enhance stability for credential of OdpsDataset #45
• Add complex type and credential support for sampler when use odps dataset #52
• Support CsvDataset with null columns #56
• Negative sampler support string id #70

Config

• Support easyrec config convert to tzrec config #37 #39 #51 #90

Upgrade

• Release official dlc image #26
• Upgrade pytorch to v2.6 torchrec to v1.1.0 #99

Note

For TorchEasyRec 0.7.x, you should use Docker image version 0.7.

• For the GPU version (CUDA 12.4):
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:0.7-cu124
  • PyTorch: v2.6 CUDA: v12.4 FBGEMM: v1.1.0 TorchRec: v1.1.0 Python: v3.11
• For the CPU version:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:0.7-cpu
  • PyTorch: v2.6 FBGEMM: v1.1.0 TorchRec: v1.1.0 Python: v3.11

Bug Fixes and Other Changes

• [bugfix] remove redundant sequence key in feature input names when fg_mode is DAG by @tiankongdeguiji in #21
• fix quota_name for add feature info by @chengaofei in #22
• update config delete drop feature config by @chengaofei in #23
• [feat] make docker compat with gpu driver 470 by @tiankongdeguiji in #24
• [bugfix] fix dlc tutorial doc by @tiankongdeguiji in #25
• [bugfix] fix dbmtl model doc by @tiankongdeguiji in #28
• [feat] add pai dlc and dsw dependency in docker by @tiankongdeguiji in #29
• [feat] update easyrec dinggroup qrcode by @tiankongdeguiji in #31
• [feat] update pyfg doc to 0.3.5 by @tiankongdeguiji in #34
• [bugfix] fix fg arrow handler with sample mask by @tiankongdeguiji in #38
• [feat] add unique test work dir by @tiankongdeguiji in #40
• [bugfix] add id field of negative sampler to selected columns by @tiankongdeguiji in #42
• [bugfix] prevent predict hang when subthread or subproc exception by @tiankongdeguiji in #44
• [bugfix] input_tile=3: make dataparser to get user feats before creat… by @yjjinjie in #46
• [bugfix] fix sequence feature doc by @tiankongdeguiji in #48
• [feat] optimize is_user_feat of Feature when use dag by @tiankongdeguiji in #53
• [bugfix] refine sample weight compatibility & refine label dtype check & relax predict pipeline check & fix num_rows < num_workers when use OdpsDataset by @tiankongdeguiji in #54
• [feat] add doc for training with maxcompute tables on DLC by @yanzhen1233 in #47
• create fg will use resource name by @chengaofei in #64
• [bugfix] fix is_sparse of LookupFeature and MatchFeature when use vocab_dict by @tiankongdeguiji in #66
• [bugfix] fix odps quota in hitrate.py & refine error info of CsvReader and ParquetReader by @tiankongdeguiji in #67
• [bugfix] fix mtl model label in ut by @tiankongdeguiji in #68
• [bugfix] fix calculate_shard_storages to handle optimizer correctly by @tiankongdeguiji in #69
• [feat] add LOG_LEVEL environ variable by @tiankongdeguiji in #71
• [bugfix] fix predict when num_workers = 0 by @tiankongdeguiji in #74
• [bugfix] fix duplicate server launch error in odps sampler test by @tiankongdeguiji in #76
• [feat] refactor batch_size to tile_size in Batch dataclass by @tiankongdeguiji in #77
• [feat]add total_loss to the plogger and summary_writer by @eric-gecheng in #78
• [bugfix] fix weighted feature when INPUT_TILE=2 by @tiankongdeguiji in #80
• [bugfix] fix negative sample table with multiple partitions by @tiankongdeguiji in #81
• [bugfix] readme typo by @eric-gecheng in #85
• [doc] fix task space doc error by @chengaofei in #86
• [bugfix] add div_no_nan and prevent divide by zero loss weight by @tiankongdeguiji in #88
• [feat] remove sample weight and labels when export by @tiankongdeguiji in #91
• [feat] configure the shell to be bash by default in docker environments by @tiankongdeguiji in #92
• [doc] creat fg json doc add upload fg json to mc method by @chengaofei in #95

New Contributors

• @yjjinjie made their first contribution in #30
• @eric-gecheng made their first contribution in #50
• @yanzhen1233 made their first contribution in #47
• @Dave-AdamsWANG made their first contribution in #49
• @chengmengli06 made their first contribution in #79
• @iWelkin-coder made their first contribution in #55

Full Changelog: v0.6.0...v0.7.0

v0.6.0

We are excited to announce the release of TorchEasyRec 0.6.0, the first public release for TorchEasyRec.

Major Features and Improvements

• High-performance training, evaluation, and prediction with GPUs.
• Supported a variety of input data types, including MaxCompute Table, OSS files, CSV files, Parquet files doc here.
• Supported a variety of feature types, including IdFeature, RawFeature, ComboFeature, LookupFeature, MatchFeature, ExprFeature, OverlapFeature, TokenizeFeature, SequenceIdFeature, SequenceRawFeature, and SequenceFeature. The feature generation operations is also efficient and robust doc here.
• Supported a variety of models, including DSSM, TDM, DeepFM, MultiTower, DIN, MMoE, DBMTL, PLE. It is also easy to implement customized models.
• Supported a variety of loss, including binary_cross_entropy, softmax_cross_entropy, l2_loss, jrc_loss doc here.
• Supported VariationalDropout feature selection.
• Easy to deploy a TorchEasyRec model as a high-performance inference service using the TorchEasyRec Processor.

Bug Fixes and Other Changes

• [bugfix] fix train_eval may hang when use OdpsDataset and set is_orderby_partition=true by @tiankongdeguiji in
• [bugfix] fix offline predict input tile model with sequence by @tiankongdeguiji in #14

Note

For TorchEasyRec 0.6.x, you should use Docker image version 0.6.

• For the GPU version (CUDA 12.1):
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:0.6-cu121
• For the CPU version:
  • mybigpai-public-registry.cn-beijing.cr.aliyuncs.com/easyrec/tzrec-devel:0.6-cpu

New Contributors

• @tiankongdeguiji made their first contribution in #1
• @jjbbong made their first contribution in #3
• @chengaofei made their first contribution in #4

Full Changelog: https://github.com/alibaba/TorchEasyRec/commits/v0.6.0