ELDR reduces median TPOT by 5.9-13.9% in PD-disaggregated MoE serving via expert signatures from prefill, K-means partitioning, and locality-band routing with KV-co-indexed signature cache.
hub Canonical reference
Splitwise: Efficient generative llm inference using phase splitting
Canonical reference. 83% of citing Pith papers cite this work as background.
22
Pith papers citing it
Background
83% of classified citations
abstract
Recent innovations in generative large language models (LLMs) have made their applications and use-cases ubiquitous. This has led to large-scale deployments of these models, using complex, expensive, and power-hungry AI accelerators, most commonly GPUs. These developments make LLM inference efficiency an important challenge. Based on our extensive characterization, we find that there are two main phases during an LLM inference request: a compute-intensive prompt computation, and a memory-intensive token generation, each with distinct latency, throughput, memory, and power characteristics. Despite state-of-the-art batching and scheduling, the token generation phase underutilizes compute resources. Specifically, unlike compute-intensive prompt computation phases, token generation phases do not require the compute capability of the latest GPUs, and can be run with lower power and cost. With Splitwise, we propose splitting the two phases of a LLM inference request on to separate machines. This allows us to use hardware that is well-suited for each phase, and provision resources independently per phase. However, splitting an inference request across machines requires state transfer from the machine running prompt computation over to the machine generating tokens. We implement and optimize this state transfer using the fast back-plane interconnects available in today's GPU clusters. We use the Splitwise technique to design LLM inference clusters using the same or different types of machines for the prompt computation and token generation phases. Our clusters are optimized for three key objectives: throughput, cost, and power. In particular, we show that we can achieve 1.4x higher throughput at 20% lower cost than current designs. Alternatively, we can achieve 2.35x more throughput with the same cost and power budgets.
hub tools
citation-role summary
background 5
method 1
citation-polarity summary
representative citing papers
Presents a distribution-aware scheduling framework for LLM inference that reduces P99 TTLT by 35-50% and TTFT by 34-47% versus SRPT with perfect length knowledge using statistical signals instead of predictions.
Disaggregated inference is modeled as three games whose price of anarchy rises at GPU saturation; an adaptive controller reduces the empirical PoA-hat by up to 3.1x on real clusters at modest throughput cost.
ViBE co-optimizes expert placement with measured GPU performance variability in MoE inference to cut execution-time imbalance, delivering 14% better SLO attainment and up to 45% lower P90 TTFT.
Proposes the Intelligent Computing Architecture (ICA) as a six-layer framework with dual probabilistic-deterministic planes and three Amdahl-style heuristics to unify design of LLM-based systems.
Power capping is illusory in LLM decode as memory-bound operation leaves power headroom untouched on 700 W GPUs, while SM clock locking saves up to 32% energy and three DVFS classes appear across attention types.
Sparse prefix caching via dynamic programming for optimal checkpoint placement under overlap distributions improves the Pareto frontier for recurrent and hybrid LLM serving on shared-prefix data.
KV-cache eviction, prompt compression, recurrent state bounding, and agent memory consolidation are unified as one rate-distortion problem with a shared lower bound, shared failure mode, and transferable mechanisms.
LLM serving should triage by five-resource analytical floors and wall ordering, not grid search; on 16×H20, TP16 is capacity-capped at ~70 while EP+DP attention reaches ~644 concurrent 8K requests.
OmniPilot combines conformal quantile regression with OOD detection to rank LLM serving configurations on mixed GPUs, reporting 6.2% MAPE throughput prediction and 95% top-1 accuracy on 460 benchmark runs while abstaining on unsupported cases.
PersistentKV is a native block-table decode attention engine with page-aware workqueue scheduling that improves decode throughput 1.04-1.40x versus FlashInfer on RTX 3060 for selected long-context GQA workloads.
Chakra introduces a standardized graph-based execution trace representation for distributed ML workloads along with supporting tools to enable benchmarking, analysis, generation, and co-design across simulators and hardware.
Dooly reduces LLM inference profiling GPU-hours by 56.4% across 12 models while keeping simulation MAPE under 5% for TTFT and 8% for TPOT by making profiling configuration-agnostic and redundancy-aware.
SplitZip introduces a fast lossless KV cache compressor for disaggregated LLM inference that achieves 613 GB/s compression throughput on BF16 tensors and up to 1.32x end-to-end speedup.
MemExplorer optimizes heterogeneous memory systems for agentic LLM inference on NPUs and reports up to 2.3x higher energy efficiency than baselines under fixed power budgets.
Integrated solar-compute-radiator panels enable orbital satellites to achieve over 100 kW of AI inference compute per metric ton launched, supporting thousands of simultaneous large language model sessions.
Organizes the heterogeneous LLM prefill-decode design space along four axes and extracts three boundary decisions with guidance on precision, KV representation, and ownership.
Mix-Quant quantizes prefilling to NVFP4 and keeps BF16 for decoding in agentic LLMs, achieving up to 3x prefilling speedup while largely preserving task performance on long-context and agentic benchmarks.
ServeGen characterizes production LLM inference workloads across model types and generates realistic per-client composed workloads that reduce under-provisioning by 50% in a production validation.
Workload-aware optimizations for LLM serving in AML and fraud detection yield substantial gains in throughput, latency, and GPU utilization on synthetic compliance prompts.
CompPow makes the case that component-aware power management inside GPUs can yield 10% higher energy efficiency and 5% better performance for ML workloads.
The paper surveys techniques to speed up and reduce the resource needs of LLM inference, organized by data-level, model-level, and system-level changes, with comparative experiments on representative methods.
citing papers explorer
-
ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving
ELDR reduces median TPOT by 5.9-13.9% in PD-disaggregated MoE serving via expert signatures from prefill, K-means partitioning, and locality-band routing with KV-co-indexed signature cache.
-
Beyond Prediction: Tail-Aware Scheduling for LLM Inference
Presents a distribution-aware scheduling framework for LLM inference that reduces P99 TTLT by 35-50% and TTFT by 34-47% versus SRPT with perfect length knowledge using statistical signals instead of predictions.
-
The Price of Anarchy in Disaggregated Inference
Disaggregated inference is modeled as three games whose price of anarchy rises at GPU saturation; an adaptive controller reduces the empirical PoA-hat by up to 3.1x on real clusters at modest throughput cost.
-
ViBE: Co-Optimizing Workload Skew and Hardware Variability for MoE Serving
ViBE co-optimizes expert placement with measured GPU performance variability in MoE inference to cut execution-time imbalance, delivering 14% better SLO attainment and up to 45% lower P90 TTFT.
-
Model-Native Computing Architecture: Envisioning Future System Architecture Through the Lens of Computer Architecture
Proposes the Intelligent Computing Architecture (ICA) as a six-layer framework with dual probabilistic-deterministic planes and three Amdahl-style heuristics to unify design of LLM-based systems.
-
The Illusion of Power Capping in LLM Decode: A Phase-Aware Energy Characterisation Across Attention Architectures
Power capping is illusory in LLM decode as memory-bound operation leaves power headroom untouched on 700 W GPUs, while SM clock locking saves up to 32% energy and three DVFS classes appear across attention types.
-
Sparse Prefix Caching for Hybrid and Recurrent LLM Serving
Sparse prefix caching via dynamic programming for optimal checkpoint placement under overlap distributions improves the Pareto frontier for recurrent and hybrid LLM serving on shared-prefix data.
-
What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents
KV-cache eviction, prompt compression, recurrent state bounding, and agent memory consolidation are unified as one rate-distortion problem with a shared lower bound, shared failure mode, and transferable mechanisms.
-
Think Before You Grid-Search: Floor-First Triage for LLM Serving
LLM serving should triage by five-resource analytical floors and wall ordering, not grid search; on 16×H20, TP16 is capacity-capped at ~70 while EP+DP attention reaches ~644 concurrent 8K requests.
-
OmniPilot: An Uncertainty-Aware LLM Inference Advisor for Heterogeneous GPU Clusters
OmniPilot combines conformal quantile regression with OOD detection to rank LLM serving configurations on mixed GPUs, reporting 6.2% MAPE throughput prediction and 95% top-1 accuracy on 460 benchmark runs while abstaining on unsupported cases.
-
PersistentKV: Page-Aware Decode Scheduling for Long-Context LLM Serving on Commodity GPUs
PersistentKV is a native block-table decode attention engine with page-aware workqueue scheduling that improves decode throughput 1.04-1.40x versus FlashInfer on RTX 3060 for selected long-context GQA workloads.
-
MLCommons Chakra: Advancing Performance Benchmarking and Co-design using Standardized Execution Traces
Chakra introduces a standardized graph-based execution trace representation for distributed ML workloads along with supporting tools to enable benchmarking, analysis, generation, and co-design across simulators and hardware.
-
Dooly: Configuration-Agnostic, Redundancy-Aware Profiling for LLM Inference Simulation
Dooly reduces LLM inference profiling GPU-hours by 56.4% across 12 models while keeping simulation MAPE under 5% for TTFT and 8% for TPOT by making profiling configuration-agnostic and redundancy-aware.
-
SplitZip: Ultra Fast Lossless KV Compression for Disaggregated LLM Serving
SplitZip introduces a fast lossless KV cache compressor for disaggregated LLM inference that achieves 613 GB/s compression throughput on BF16 tensors and up to 1.32x end-to-end speedup.
-
MemExplorer: Navigating the Heterogeneous Memory Design Space for Agentic Inference NPUs
MemExplorer optimizes heterogeneous memory systems for agentic LLM inference on NPUs and reports up to 2.3x higher energy efficiency than baselines under fixed power budgets.
-
Reduced-Mass Orbital AI Inference via Integrated Solar, Compute, and Radiator Panels
Integrated solar-compute-radiator panels enable orbital satellites to achieve over 100 kW of AI inference compute per metric ton launched, supporting thousands of simultaneous large language model sessions.
-
Demystifying the Design Space and Best Practices for Heterogeneous LLM Inference and Serving
Organizes the heterogeneous LLM prefill-decode design space along four axes and extracts three boundary decisions with guidance on precision, KV representation, and ownership.
-
Mix-Quant: Quantized Prefilling, Precise Decoding for Agentic LLMs
Mix-Quant quantizes prefilling to NVFP4 and keeps BF16 for decoding in agentic LLMs, achieving up to 3x prefilling speedup while largely preserving task performance on long-context and agentic benchmarks.
-
ServeGen: Workload Characterization and Generation of Large Language Model Serving in Production
ServeGen characterizes production LLM inference workloads across model types and generates realistic per-client composed workloads that reduce under-provisioning by 50% in a production validation.
-
Rethinking LLMOps for Fraud and AML: Building a Compliance-Grade LLM Serving Stack
Workload-aware optimizations for LLM serving in AML and fraud detection yield substantial gains in throughput, latency, and GPU utilization on synthetic compliance prompts.
-
CompPow: A Case for Component-level GPU Power Management
CompPow makes the case that component-aware power management inside GPUs can yield 10% higher energy efficiency and 5% better performance for ML workloads.
-
A Survey on Efficient Inference for Large Language Models
The paper surveys techniques to speed up and reduce the resource needs of LLM inference, organized by data-level, model-level, and system-level changes, with comparative experiments on representative methods.