arXiv AI
Aug 14
TEMPO: Makespan-Aware Expert-Parallel Load Balancing Across Memory- and Compute-Bound Regimes
arXiv:2608. 13057v1 Announce Type: cross Abstract: In expert-parallel (EP) MoE serving, every layer synchronizes at the slowest GPU.
TableTune: Measured, Self-Correcting Draft-Length Tables for Dynamic Speculative Decoding
TableTune is a system that automatically measures and corrects draft-length tables for speculative decoding in large language model serving. It profiles decoding performance across batch sizes and draft lengths, generates an optimized table in minutes, and a controller plugin continuously updates the table based on live traffic. Experiments on H100 GPUs show that TableTune consistently achieves at least 94% of the best fixed draft-length throughput, outperforms other speculative control methods, and can recover up to 96% of performance lost by stale tables, delivering significant speedups in real deployments.
Related stories
arXiv AI
Aug 20
Cacheable by Design? Training Mixture-of-Experts Routers for Locality Against the Edge Memory-Bandwidth Wall: A Pre-Registered Negative Result with a Systems Measurement Study
The paper investigates whether training Mixture-of-Experts (MoE) routers can improve memory‑bandwidth locality on consumer GPUs. Using a new zero‑surgery telemetry tool, the authors measure that a large Qwen3‑235B model is bottlenecked by disk‑based expert access, and that an LRU cache can serve a majority of requests. They pre‑register experiments training 137 M‑parameter MoE models with locality‑aware losses, finding that while cache misses can drop up to 60 % (99 % static‑pin hit rate), every configuration fails to meet a strict 1 % perplexity threshold, indicating a tight coupling between cache efficiency and model quality.
arXiv Machine Learning
6d ago
Speculative Decoding for Mixture-of-Experts Models: Expert Bytes, Kernel Bandwidth and the Best Draft Length
The paper investigates speculative decoding for mixture‑of‑experts (MoE) language models, focusing on how deduplication of expert reads can reduce byte traffic and improve decoding speed. By modeling step time as a function of bytes requested and kernel bandwidth, the authors derive conditions under which speculation outperforms plain decoding and identify the optimal draft length as a staircase function of acceptance rate. Experiments on llama.cpp Metal, MLX‑LM, and llama.cpp CUDA show that deduplication is only faster for very short draft lengths (5–8 tokens) or not at all, and the predicted crossover points closely match measured savings across multiple configurations.
arXiv Machine Learning
Sep 15
Minibatch persistency, eight years later: what batch reuse costs in steps and joules, and what it saves in data
arXiv:2609.13922v1 Announce Type: new Abstract: Minibatch persistency reuses data instead of reading it: rather than drawing a fresh minibatch at every optimizer step, it takes K consecutive steps on...
arXiv AI
Oct 27
Prefetch-Steered Routing: Deterministic Cache-Aware Expert Selection for Offloaded Mixture-of-Experts Decoding
Prefetch‑Steered Routing (PSR) is a deterministic, cache‑aware method for selecting experts in Mixture‑of‑Experts models that limits physical weight copies to a logical cache of at most *n* experts per layer and token. On several large‑scale models (OLMoE‑1B‑7B, Moonlight‑16B‑A3B, Qwen3.6‑35B‑A3B) PSR achieves 2–2.3× faster decoding than exact routing with LRU and outperforms the Cache‑Prior baseline by 25–30% in many benchmarks, while maintaining comparable accuracy. The approach also shows that speculative block‑level expert selection can be faster than autoregressive decoding when GPU memory is constrained.
arXiv Machine Learning
Oct 16
When the KV Cache Is Slower Than the GPU: Scaling Remote KV Tiers for Context-Parallel Prefill
The paper investigates the performance impact of using a remote KV cache tier for long‑context language model serving. In a disaggregated deployment of GLM‑5.3 on 8xB300 nodes, enabling the remote tier initially increased the 16k‑token p99 latency from 3.96 s to 17.3 s, and synchronous loading caused severe queue instability. By sharding the KV store across context‑parallel ranks, partitioning the remote tier, adding asynchronous loading, and fixing client bugs, the authors reduced the p99 latency from 83.32 s to 0.67 s and increased prefill throughput by 14–34 %. whyItMatters":"The study demonstrates that careful sharding and asynchronous handling of remote KV caches can dramatically improve latency and throughput in large‑scale LLM serving systems."
arXiv Machine Learning
Sep 1
Ceiling-Clipped Acceptance Histograms Indicate Stranded Speed-up in Block-Diffusion Speculative Decoding
arXiv:2608.30427v1 Announce Type: cross Abstract: Speculative decoding speeds up generation with an efficient draft model (drafter) that proposes tokens for a target model to verify in one pass, pres...
arXiv Machine Learning
Jul 24
Windowed-MTP: Removing the Full-Context Draft-KV Tax at Million-Token Context
arXiv:2607. 21535v1 Announce Type: new Abstract: Speculative decoding accelerates autoregressive generation by having a cheap draft propose tokens that a target verifies in parallel.
arXiv AI
Sep 16
Calibrate, Then Route: A Measured Study of Learned Request Routing for Disaggregated LLM Serving
The paper evaluates a learned request‑routing policy for disaggregated large‑language‑model serving, where compute‑heavy prefill and memory‑heavy decode stages run on separate GPU pools. Using a discrete‑event simulator and real NVIDIA A40 GPUs, the calibrated router—leveraging prompt length, predicted output length, KV‑cache pressure, and SLO class—outperforms round‑robin, least‑loaded, and length‑based heuristics, achieving the highest mean goodput (0.864) and lowest variance across three mixed, bursty arrival traces. Hardware calibration proves critical, providing a 4.5‑point goodput boost and roughly 40 % of the tail‑latency advantage, and the learned router can match round‑robin performance with one fewer GPU in certain scenarios.
arXiv AI
Sep 17
The Inference Engineering Pareto Atlas: Which Optimizations Dominate the Cost, Quality, and Latency Frontier?
The paper presents a Pareto atlas of LLM inference optimizations, mapping cost, quality, and latency trade‑offs for Qwen2.5‑7B‑Instruct on L4, A100, and H100 GPUs. Using 54 measured configurations and a calibrated simulator, it identifies 18 of 36 setups on the Pareto frontier, showing that combined methods outperform single ones. Quality tests reveal that AWQ 4bit and FP8 weights offer significant latency reductions while largely preserving accuracy, but naive FP8 KV caching fails to answer any questions correctly.
arXiv AI
3d ago
Only Accepted Tokens Count: An Energy Accounting of Speculative Decoding Under Latency Targets
The paper investigates how speculative decoding—where a draft model proposes multiple future tokens that are then verified by the target model—affects energy consumption in large language model serving under latency constraints. It finds that energy savings depend heavily on the acceptance rate of drafted tokens, which varies by model architecture, drafting scheme, and the type of text (e.g., conversation vs. code, English vs. Romansh). The study shows that while speculation can reduce energy per token at small batch sizes, at larger batches the verification cost dominates, potentially turning speculation into an energy loss, especially when acceptance rates are low.
arXiv AI
Oct 1
Working Around the Compute Ceiling: Byte-Exact Memory in Galahad Makes LLM Reading a One-Time Cost LLM Reading a One-Time Cost
The paper introduces Galahad, a memory layer that stores a transformer language model’s key‑value state for blocks of text, allowing subsequent requests to reuse previously computed attention rather than recomputing it. On seven real‑world datasets, 98.7% of prompt tokens were already read, and with Galahad the model could attend to an entire 97,000‑token corpus, achieving 98–100% recall on a 100‑fact test while reducing inference time and energy consumption dramatically. The approach was validated across 30 models and all runtimes, demonstrating that stateful inference can replace stateless serving without loss of accuracy.