DrainSinkhorn is a verifier‑gated active‑packing layer that improves batched entropic optimal transport (EOT) by eliminating finished problems from subsequent Sinkhorn updates. It combines candidate‑axis packing, a one‑sided screen, verifier‑gated retirement, and physical compaction, while keeping the EOT objective, per‑instance map, and stopping rule unchanged. The method achieves state‑of‑the‑art execution speedups—up to 4.11× faster on MetroPT‑3 and 3.80× on ImageNet‑32 feature couplings—across multiple backends and tolerance settings. whyItMatters":"The technique delivers significant runtime reductions for heterogeneous batched‑EOT workloads, enabling faster and more efficient optimal transport computations in practical machine‑learning pipelines."
arXiv:2607. 24741v1 Announce Type: cross Abstract: Dynamic applications, including optimal-transport Flow Matching, repeatedly solve related entropic optimal transport problems, yet conventional distributed Sinkhorn processes frames sequentially and synchronizes after every iteration.
The paper introduces an error‑diffusion (ED) optimizer that replaces the full‑precision latent copy of binary‑weight language models with a small 2‑8‑bit residual buffer, reducing training state per weight from 97 bits to 3‑9 bits—up to 19× less. ED accumulates unrealized weight updates in the buffer and flips the binary weight when the accumulated value crosses a threshold aligned with the quantization grid, enabling efficient training on large transformers with minimal loss compared to latent‑weight baselines. The method scales to models with up to 19 M parameters on enwik8 and 70 M on wikitext‑103, maintaining performance within a few percent of reference while allowing on‑device adaptation within a 275 MB memory budget.
Dynamic applications, including optimal-transport Flow Matching, repeatedly solve related entropic optimal transport problems, yet conventional distributed Sinkhorn processes frames sequentially and synchronizes after every iteration. We present TemporalSinkhorn, a parallel-in-time executor that batches future candidates and their repairs without making output accuracy speculative.
arXiv:2606. 19989v1 Announce Type: cross Abstract: Modern LLM training breaks a core assumption behind offline batch samplers: the true training cost of a sample is only observable after preprocessing, augmentation, templating, tokenization, and multimodal visual-token expansion.
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.
ScaleRLT introduces a hybrid learner log‑probability method that reconciles differences between dense and deferred rollout computations in looped language models like Ouro. By using a stop‑gradient replay of the rollout schedule, the approach achieves bit‑for‑bit reproducibility across FP32 and bf16 real‑weight cells while maintaining the same gradient accuracy as dense implementations. The technique boosts rollout throughput by up to 1.56× on a single GPU and significantly speeds up multi‑node GRPO updates, leading to substantial accuracy gains on GSM8K in short training runs.
arXiv:2608. 13987v1 Announce Type: new Abstract: Nanbeige4.
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.
Dynamic Expert Quantization (DynaExq) is a runtime-aware mixed-precision serving system designed for single‑GPU Mixture‑of‑Experts (MoE) inference under a hard high‑bandwidth memory (HBM) envelope. It treats the problem as an online, budget‑constrained precision allocation task, keeping the most frequently used experts at higher precision while relegating the rest to low‑precision fallbacks. By estimating expert hotness from router traces and asynchronously promoting or demoting experts, DynaExq maintains a fully materialized expert set during the forward pass, improving accuracy and throughput compared to static post‑training quantization and offloading/prefetch baselines. whyItMatters":"DynaExq enables efficient deployment of large MoE models on memory‑limited GPUs by dynamically allocating precision based on runtime expert usage, thereby reducing memory footprint and latency while boosting accuracy and throughput."
arXiv:2610. 23796v1 Announce Type: new Abstract: Transformer prefill reuses weights across tokens and is dominated by matrix multiplication and activation data movement; single-token decode has limited reuse and is dominated by weight and KV-cache reads.
Serving a 235B-parameter Mixture-of-Experts (MoE) model on a single 8 GB GPU is bottlenecked not by compute but by memory bandwidth: decode must stream each token's active experts from whichever tier holds them, and on consumer hardware most experts sit on an SSD far slower than RAM. We quantify this bandwidth wall on Qwen3-235B (Q4_K_M, 134 GB): measured decode is 0.
NeMo-DCR is a delta‑compressed refit method that sends only the changes in a trillion‑parameter policy model, achieving bit‑exact synchronization across AWS regions. By using fixed affine mappings, XOR masks, and a relay‑tree object storage, it avoids full checkpoint transfers and cross‑cluster collectives, reducing refit times from 87.5 min to 150 s for a 1 T model at 3% change rate. The approach scales to 30 B–1 T models, delivering 12–40× speedups over traditional full‑checkpoint refits.