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