Industrial recommendation rankers keep scaling in width and depth to billions of parameters, but this increasingly incurs costs in parameter growth, distributed communication overhead, and training inefficiency. We introduce Chain Dimension Scaling (CDS), a scaling strategy orthogonal to width and depth. CDS repeatedly applies a single feature-interaction block with parameters shared across iterations, so the number of iterations (the chain length) becomes a new scaling dimension that adds no parameters. A lightweight intra-chain residual connection stabilizes optimization and improves quality. CDS enables two regimes. In the compression regime, shallow blocks with a longer chain match deeper baselines with fewer parameters and faster training; in the quality-scaling regime, increasing chain length under a fixed parameter budget yields consistent offline and online gains, suggesting the chain dimension adds modeling capacity beyond parameter reuse. We evaluate CDS in a large-scale multi-task short-video ranking system in production. In the compression regime, CDS stays within 0.03% of baseline offline quality while reducing trainable parameters by 26.6% and improving training throughput by 6.4%, with neutral-to-positive online participation. In the quality-scaling regime, quality improves monotonically with chain length. Wall-clock analysis shows the efficiency gain comes mainly not from reduced FLOPs but from lower cross-device synchronization overhead and reduced optimizer update cost. These results position parameter-shared iterative computation as a practical, communication-efficient scaling axis for industrial ranking systems.
Figures & tables
Figure 1: The ratio of trainable parameters (left) and training-step wall-clock time (right) across the major components of the production ranker. The stacked feature-interaction module dominates both, motivating it as the target for the new scaling axis.
Figure 2: Architecture comparison betweendepth scaling and CDS.
Figure 3: Pareto comparison of compression strategies. Naive reduces the number of independently parameterized blocks without a residual connection; CDS+IR is CDS with the intra-iteration residual. Dot size is proportional to trainable parameter count. Left: offline AUC gain over the production baseline (bps). Right: online engagement delta. CDS dominates the naive variant at every parameter budget on both metrics.
Figure 4: Comparison of standard Depth Scaling vs. Chain Dimension Scaling (CDS). The left panel shows the relative training speed and parameter footprint. The middle and right panels show the corresponding offline AUC gains and online engagement gains relative to the L=1 baseline. CDS ( L=3,C=2 ) retains the efficiency of a 3-layer model while achieving quality metrics competitive with a 6-layer model.
Figure 5: Residual design ablation within CDS. Baseline : production model with L independently parameterized blocks (no CDS). Naive : CDS without residual (Eq. 2 applied as written). Intra : CDS with the intra-iteration residual (Eq. 3 ). Both : CDS with intra and inter residuals combined. All CDS variants use L′=L/2 and C=2 . Left and middle: absolute AUC on two representative classification tasks. Right: online engagement delta with 95% confidence intervals from live A/B tests.
Figure 6: Wall-clock decomposition of the per-step training time for the production baseline vs. launched CDS, in milliseconds. (a) Phase-level breakdown: Forward and Backward are the interaction-block forward and backward passes; Optimizer is the parameter update step; Overall is the full training step (the remainder is occupied by ranker components unchanged by CDS: input feature layer, embedding lookups, shared projection, task towers). (b) Operation-level breakdown: compute and memory operations (grey/blue) vs. communication operations (orange/red: All Gather for parameter unshard, Reduce Scatter for gradient aggregation, both introduced by sharded data-parallel training).
Communication-efficient distributed training reduces the data exchanged among devices so models can be trained faster and over limited-bandwidth networks.
Modern industrial Deep Learning Recommendation Models typically extract user preferences through the analysis of sequential interaction histories, subsequently generating predictions based on these derived interests. The inherent heterogeneity in data characteristics frequently result in substantial under-utilization of computational resources during large-scale training, primarily due to computational bubbles caused by severe stragglers and slow blocking communications. This paper introduces FreeScale, a solution designed to (1) mitigate the straggler problem through meticulously load balanced input samples (2) minimize the blocking communication by overlapping prioritized embedding communications with computations (3) resolve the GPU resource competition during computation and communication overlapping by communicating through SM-Free techniques. Empirical evaluation demonstrates that FreeScale achieves up to 90.3% reduction in computational bubbles when applied to real-world workloads running on 256 H100 GPUs.
Modern industrial recommender systems use a deep ranking model to score N candidates against the same user and context features. Standard implementations broadcast context features early in the forward pass, redundantly computing context-only operations N times per request. We present a rank-aware decomposition applicable to the dominant interaction mechanisms in modern recommender architectures-Factorization Machine (FM) pairwise products, Deep Cross Network (DCNv2) cross layers, self-attention, and fully connected (FC) projection layers-built on a single algebraic principle: any linear or bilinear operation over a rank-partitioned input admits an exact block decomposition that moves context-only computation from once-per-candidate to once-per-request, identity-equivalent to the original model. Closed-form analysis and controlled ablation verify that savings scale quadratically with the number of context features. Applied to a production DLRM-style ranker without any architectural change, the decomposition increases per-pod throughput by 87.5% (a 47% reduction in peak pod count) at identical model predictions. The identity-equivalent decomposition applies only at the first layer of cross networks and self-attention, since each layer mixes ranks in its output. To extend savings across depth, we further introduce rDCN, an architectural variant of DCNv2 that maintains rank discipline across depth and matches DCNv2 accuracy within training noise at 67% fewer total FLOPs, and sketch an analogous architectural variant for self-attention.
Recent advances in Large Language Models (LLMs) have inspired a surge of scaling research in industrial search, advertising, and recommendation systems. However, existing approaches focus mainly on architectural improvements, overlooking the critical synergy between data and architecture design. We observe that scaling model parameters alone exhibits diminishing returns, and that the performance degradation caused by complex heterogeneous data distributions is often irrecoverable through model design alone. In this paper, we propose UniScale, a novel co-design framework that jointly optimizes data and architecture to unlock the full potential of model scaling. UniScale includes two core parts: (1) ES3 (Entire-Space Sample System), a high-quality data scaling system that expands the training signal beyond conventional sampling strategies through intra-domain expansion with hierarchical label attribution and cross-domain searchification; and (2) HHSFT (Heterogeneous Hierarchical Sample Fusion Transformer), a novel architecture that effectively models the complex heterogeneous distribution of scaled data via Heterogeneous Hierarchical Feature Interaction and Entire Space User Interest Fusion, thereby surpassing the performance ceiling of structure-only model tuning. Extensive experiments on large-scale industrial datasets demonstrate that UniScale achieves significant improvements and exhibits clear scaling trends. Online A/B tests on a real-world E-commerce search platform confirm that UniScale consistently outperforms strong production baselines, achieving 1.70% and 2.04% increases in user purchase and Gross Merchandise Value (GMV).
Liren Yu, Caiyuan Li, Feiyi Dong +5
Taobao & Tmall Group of Alibaba Hangzhou, China · Taobao & Tmall Group of Alibaba Beijing, China