Proto commits in alibaba/TorchEasyRec

These commits are when the Protocol Buffers files have changed: (only the last 100 relevant commits are shown)

2026-06-24

[feat] add model delta tracker (#546)

The documentation is generated from this commit.

2026-06-11

[feat] SID: add SidRqkmeans model (FAISS-trained residual K-Means) (#539) Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

2026-06-09

[feat] Kafka: event-time driven checkpointing from message timestamp (#541) Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

2026-05-29

Merge remote-tracking branch 'origin/master' into feat/ultra-hstu-fp8 # Conflicts: # tzrec/version.py

[fix] ULTRA-HSTU FP8: narrow the arch gate to SM90 + SM120-mode2 The previous "SM90+" gate was too permissive: - SM100 (Blackwell datacenter) has no FP8 kernel in the wheel; the dispatcher routes there via _sm100.hstu_varlen_fwd_100 which doesn't even take quant_mode (cuda_hstu_attention.py:399-403). - SM120 (Blackwell RTX) only handles quant_mode==2 (per-block, fwd-only, cuda_hstu_attention.py:282); for any other mode the wheel silently falls into the sm80 bf16/fp16 branch (line 308's `or major_version == 12`) -- the user gets non-FP8 attention with no warning. Tighten _assert_fp8_capable to accept exactly (sm90, any mode) or (sm120, mode=2), and reject everything else loudly. Pass fp8_quant_mode into the helper so it can mode-check on sm120. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

[feat] ULTRA-HSTU FP8: extend support from SM90 to SM90+ (Blackwell) Relax the FP8 capability gate from "exactly SM90 (Hopper)" to "SM90+" so the same fp8_quant_mode>=0 path also runs on sm100 (Blackwell) and sm120 (Blackwell RTX). The wheel dispatches to its per-arch FP8 kernel internally (sm120/Blackwell RTX is forward-only and supports only quant_mode=2; that constraint surfaces from the wheel's own check, not tzrec's). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-27

[feat] support keep_checkpoint_max with async checkpoint pruning (#528) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

[feat] ULTRA-HSTU: add STU.fp8_quant_mode proto field Add an int32 `fp8_quant_mode` (default -1) to the STU message. -1 keeps attention in bf16/fp16; 0..5 select an FP8 mode forwarded to the CUTLASS (SM90/Hopper) kernel. Mirrors the wheel's quant_mode int and the existing scaling_seqlen=-1 sentinel style. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-19

[refactor] HSTUMatch with STUStack + UIHPreprocessor + block-suffix candidates (#506) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-14

[refactor] MatchTowerWoEG accepts feature_groups (plural); DSSMTower takes EmbeddingGroup (#510) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-13

[feat] metrics: NormalizedEntropy for binary classification (#507) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-12

[refactor] dataset/sampler: dynamic expand_factor + build_sampler_input + block-suffix combine (#505) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-11

[feat] FeatureGroupConfig.embedding_name_suffix to break embedding sharing across groups (#504) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-09

[bugfix] thread contextual_seq_len from preprocessor to STULayer (proto sentinel + truncation total_uih_len + AOTI-friendly SLA builder) (#501) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-08

[feat] stu.scaling_seqlen + drop autotune assert strip (#500) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-07

[feat] Adadelta + RMSprop sparse and dense optimizers (#499) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

[perf] AOTI export knobs: fp32 unbacked floats, sample-input autotune, TF32 from export_config (#498) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-01

[feat] ULTRA-HSTU Mixture of Transducers (MoT) (#492) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-04-29

[feat] ULTRA-HSTU mid-stack attention truncation (#488) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-04-27

[feat] Semi-Local Attention + selective activation rematerialization (#486) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-04-11

[feat] support TokenizeFeature as token-level sequence input (#470) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-08

[feat] add CUTLASS kernel backend for HSTU attention (#465) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

[feat] integrate dynamicemb table fusion (wheel 20260407.97b80bf) (#466) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-01

[feat] add concat_contextual_features option to DlrmHSTU (#459) When enabled, all contextual features are concatenated on the channel dimension and projected as a single token instead of N separate tokens, reducing HSTU attention cost by shortening sequence length. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-03-31

[feat] add per-task loss weight to FusionSubTaskConfig (#453) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

[feat] add CosineAnnealingLR and CosineAnnealingWarmRestartsLR schedules (#454) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-03-30

[feat] add CombineFeature support (#447) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

[feat] add label_smoothing support to BinaryCrossEntropy loss (#455) Label smoothing helps with noisy click labels and improves generalization in ranking models. Smooths hard binary labels using the standard formula: label * (1 - eps) + 0.5 * eps, consistent with PyTorch CrossEntropyLoss. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-03-16

[feat] add grad clipping for dense params (#424)

2026-03-05

[feat] support input fields str (#412)

2026-03-03

[feat] support pepnet (#402)

2026-02-11

[feat] add kafka dataset (#401)

2026-01-30

[feat] dlrm hstu support sequence timestamp is descending order (#395)

2026-01-27

[bugfix] make sequence related config optional (#390)

2026-01-24

[feat] dlrm and wukong support only one sparse group (#385)

2026-01-21

[feat] support sequence cross features (#375)

2026-01-20

[feat] support pe ltr in train wrapper (#381)

2026-01-16

[feat] support initial_accumulator_value for FusedSparseAdagradOptimizer & add additional optimizer configuration options (#382)

2026-01-14

[feat] add wukong model (#372)

2025-12-30

[feat] add AdmissionStrategy support for DynamicEmbedding (#362)

2025-12-25

[feat] add time_bucket_increments for DlrmHSTU PositionEncoder (#359)

[feat] add time_bucket_increments for DlrmHSTU PositionEncoder (#359)

2025-12-16

[feat] mtl tower of dlrm hstu support num_class > 1 (#352)

2025-12-09

[feat] support dynamic batch size with sample cost (#343)

2025-12-04

[feat] mind dynamic routing support zero init (#342)

2025-12-01

[feat] add TMA support for hstu attn & rms_norm test (#336)

2025-11-27

[feat] add global average loss option for DlrmHSTU (#334)

2025-11-17

[feat] TensorRT export (#318)

2025-11-10

[feat] refactor dlrm hstu preprocess modules (#314)

2025-11-07

[feat] support log training metric (#310)

2025-10-28

[feat] support adamw optimizer and part optimizer and label soomthing (#297)

2025-10-22

[feat] export best model (#294)

[feat] add contextual_feature_to_pooling to DlrmHSTU preprocessors (#296)

2025-10-17

[feat] make dlrmhstu watchtime feature optional (#290)

2025-10-13

[feat] support bool mask feature (#285)

2025-10-11

[feat] add dynamicemb doc (#283)

[feat] add tool to initialize dynamic embeddings from tables (#282)

2025-10-10

[bugfix] revert dynamic embedding hash bucket size and remove unused evict_strategy (#281)

2025-10-09

[feat] add dynamic embedding support (#279)

2025-09-18

[feat] add kv dot product feature (#276)

2025-08-21

[feat] support xauc and grouped xauc (#252)

[feat] add sequence self attention encoder (#251)

2025-08-15

[feat] add dcnv2 and xdeepfm net (#242)

2025-08-08

[feat] dcn_v1 (#235)

2025-08-07

[feat] refine hstu ops & add triton tests for dlrm hstu (#231)

2025-08-06

[feat] add hstu rank model (#227)

2025-07-30

[feat] oss dlrm hstu modules (#224)

[feat] add use_ln option for MLP module & fix parse encoded sequence feature error msg (#223)

2025-07-29

[feat] add fp16 embedding dtype and fix weight decay mode (#221)

[feat] add mixed_precision bf16/fp16 and gradient accumulation support (#220)

2025-07-25

[feat] support feature only used as fg dag intermediate result (stub_type=true) (#218)

2025-07-24

[feat] expr feature support value_dim & bump up pyfg to 0.7.1 (#216)

2025-07-22

[feat] add wide and deep model and wide_init_fn (#212)

2025-07-16

[feat] upgrade pyfg to 0.6.9 and refine expr/overlap feature doc (#199)

2025-07-04

[feat] support freeze embedding parameters (#206)

2025-07-03

[feat] add binary focal loss (#208)

2025-05-26

[feat] add allow_tf32 flag and global embedding param constraint (#188)

[feat] add masknet for dbmtl and refine masknet logic (#187)

2025-05-22

[feat] add max sequence length for sequence encoder (#184)

2025-05-16

[feat] write tensorboard log for model parameters (#181)

2025-05-14

[feat] masknet (#179)

2025-05-07

[feat] add fg value_type config and make num_buckets default value_dtype as string (#175)

2025-05-01

[feat] optimize mind model (#157) - optimize creation and scaling for the routing_logit tensor - optimize the iteration of dynamic routing, capturing gradient after iteration - adjust MindUserTower's MLP modules, the inner layers and output layers are extracted separately - add bias hyper-parameter for the MLP module. For sequence feature, bias is not used

2025-04-01

[feat] add kernel config and BaseModule (#151)

2025-03-27

[feat] add regression and multi-classification metric (#149)

2025-03-24

[feat] support dlrm model (#148)

2025-03-21

[feat] add custom feature and custom sequence feature (#144)

2025-03-20

[feat] Optimize HSTU training and sampling process (#93)

2025-03-19

[feat] add odps_data_compression config (#146)

2025-03-12

[feat] add rocket launching model (#129)

2025-02-28

[feat] add mind model (#119)

2025-02-25

[feat] eval and save checkpoint by epoch (#116)

2025-02-24

[feat] support dataset shuffle (#114)

2025-02-05

[feat] add vocab file for features (#97)

2025-01-21

[feat] make default bucketize value configurable (#94)

2025-01-13

[feat] add hstu (#55)

2025-01-08

[feat] add dual augmented two-tower model (#83)

[feat] add task space for mtl loss (#82)

2024-12-30

[feat] refactor embedding group input tile and dense embedding collection (#75) * refactor embedding group input tile and dense embedding collection * fix tests * refactor proto and add tests * refactor proto and add tests * fix tests * fix tests * fix tests * add docs

2024-12-27

[feat] support Autodis and MLP embedding for raw features (#73)