mlsanityData to 7 Oct 2026

When KV Meets Embeddings: Dynamic GPU Memory Allocation for Accelerating Generative Recommender Serving

arXiv 2605.04450 ↗First posted 6 May 2026cs.DC cs.IR cs.LGcited by 0

On the map: Efficient LLM inference and serving → KV cache optimization

Abstract (arXiv)

Generative Recommender (GR) inference places embedding hot caches (EMB) and KV caches in direct competition for limited GPU HBM: allocating more memory to one improves its efficiency but degrades the other. Existing systems optimize them in isolation, overlooking that the optimal EMB-KV allocation ratio can shift by up to 0.35 across workload regimes, leaving 20-30% latency improvement unrealized. While online reallocation is required to close this gap, naive approaches introduce H2D refill traffic on the critical path, causing P99 SLO violations. To address this, we present RACER, which jointly manages HBM allocation and request routing at runtime through two key components: (1) Adaptive Memory Allocation, a three-layer PPO-based controller (frozen base policy, online residual adapter, and burst-aware recovery controller) that achieves 32,μ s decision latency while staying within 0.024-0.029 of the offline-optimal ratio; and (2) EMB-KV-Aware Scheduling, which routes requests by jointly considering KV residency, embedding locality, and node load to avoid routing inefficiencies under heterogeneous allocations. Evaluations on three production-scale datasets over a 32-node A100 cluster show that RACER reduces P99 latency by 24-38% over the best static policy and achieves 93.5-99.6% SLO satisfaction across Steady, Trend, and Burst workloads, significantly outperforming state-of-the-art baselines without sacrificing throughput.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in KV cache optimization, one of 12 directions in Efficient LLM inference and serving. Builds on 26 of its 56 references that are in the map. First posted on arXiv 6 May 2026.

Builds on

26 of 56 references.

Closest ancestors · 3

Same direction, newest first

  1. 2025xGR: Efficient Generative Recommendation Serving at Scale12 Dec
  2. 2024Fast State Restoration in LLM Serving with HCache7 Oct
  3. 2023Efficient Memory Management for Large Language Model Serving with PagedAttention12 Sep

Other references · 23

  1. 2017Proximal Policy Optimization Algorithmscited by 14,441
  2. 2016XGBoost: A Scalable Tree Boosting Systemcited by 5,120
  3. 2019Deep Learning Recommendation Model for Personalization and Recommendation Systemscited by 327
  4. 2024Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendationscited by 184
  5. 2019The Architectural Implications of Facebook's DNN-based Personalized Recommendationcited by 83
  6. 2021Software-Hardware Co-design for Fast and Scalable Training of Deep Learning Recommendation Modelscited by 78
  7. 2025MTGR: Industrial-Scale Generative Recommendation Framework in Meituancited by 47
  8. 2025LONGER: Scaling Up Long Sequence Modeling in Industrial Recommenderscited by 46
  9. 2021FBGEMM: Enabling High-Performance Low-Precision Deep Learning Inferencecited by 40
  10. 2025OneRec-V2 Technical Reportcited by 28
  11. 2021Persia: An Open, Hybrid System Scaling Deep Learning-based Recommenders up to 100 Trillion Parameterscited by 17
  12. 2022AutoShard: Automated Embedding Table Sharding for Recommender Systemscited by 17
11 more
  1. 2025Towards Large-scale Generative Rankingcited by 10
  2. 2022RecD: Deduplication for End-to-End Deep Learning Recommendation Model Training Infrastructurecited by 9
  3. 2025TBGRecall: A Generative Retrieval Model for E-commerce Recommendation Scenarioscited by 5
  4. 2024Accelerating Communication in Deep Learning Recommendation Model Training with Dual-Level Adaptive Lossy Compressioncited by 4
  5. 2025Toward Robust and Efficient ML-Based GPU Caching for Modern Inferencecited by 3
  6. 2024Hierarchical Resource Partitioning on Modern GPUs: A Reinforcement Learning Approachcited by 2
  7. 2025FLAME: A Serving System Optimized for Large-Scale Generative Recommendation with Efficiencycited by 2
  8. 2026RelayGR: Scaling Long-Sequence Generative Recommendation via Cross-Stage Relay-Race Inferencecited by 2
  9. 2026GEMs: Breaking the Long-Sequence Barrier in Generative Recommendation with a Multi-Stream Decodercited by 2
  10. 2025Near-Zero-Overhead Freshness for Recommendation Systems via Inference-Side Model Updatescited by 1
  11. 2025Embedding Samples Dispatching for Recommendation Model Training in Edge Environmentscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Enabling High-Bandwidth Flash for Generative Recommendation Serving with Write-Aware KV Cache Policy
  2. 2025NEZHA: A Zero-sacrifice and Hyperspeed Decoding Architecture for Generative Recommendations
  3. 2026HybridGen: Efficient LLM Generative Inference via CPU-GPU Hybrid Computing
  4. 2026CollectiveKV: Decoupling and Sharing Collaborative Information in Sequential Recommendation
  5. 2025KernelEvolve: Scaling Agentic Kernel Coding for Heterogeneous AI Accelerators at Meta
  6. 2026Beyond Static RAG: An Adaptive, Tri-Metric Routing Framework for Efficient Long-Context Inference on Commodity GPUs
  7. 2025On-device Semantic Selection Made Low Latency and Memory Efficient with Monolithic Forwarding
  8. 2025Optimizing LLM Inference: Fluid-Guided Online Scheduling with Memory Constraints