The paper presents an end‑to‑end MXFP8 reinforcement learning implementation on Blackwell GPUs, evaluating its benefits and limitations with Qwen3-8B and Qwen3-30B-A3B models. It finds that matching FP8 precision between sampler and learner improves consistency, that speed gains appear only for large micro‑batches, and that several weight‑synchronization correctness conditions can fail silently. The study also introduces seven explicit checks to detect such failures during training.
By Wengee Zhang
Low-precision arithmetic can make reinforcement learning (RL) for large language models cheaper, but RL requires the sampler and the learner to agree. Several frameworks now run FP8 on both sides. We...
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.
By Pierre-Olivier Bonin
arXiv:2608. 13057v1 Announce Type: cross Abstract: In expert-parallel (EP) MoE serving, every layer synchronizes at the slowest GPU.
By Jie Li, Chenxin Jia, Jinliang Shen, Cunzhuang Liu, Ruiyi Ding, Jianwen Xian, Kang He, Chengru Song
arXiv:2607. 14427v1 Announce Type: new Abstract: A depth-recurrent transformer applies a weight-tied core a variable number of times, and prior work has shown that training with a randomized recursion count yields one checkpoint usable across a range of inference depths.
By Joe Logan
The paper identifies a hidden mismatch in reinforcement‑learning fine‑tuning of large language models: when training uses fp32 master weights but rollouts are served from a bf16 copy produced by round‑to‑nearest conversion, most per‑step updates are too small to affect the served weights, causing a lag of 50–150 steps. The authors propose Fair Cast, a one‑line change that exports the master weights with an unbiased rounding rule (stochastic rounding or error feedback) and computes the loss on the served copy, eliminating the lag without adding hyperparameters and incurring only a modest computational cost. Across multiple models and optimizers, Fair Cast consistently outperforms round‑to‑nearest, improves checkpoint consistency, and reveals the implicit regularization effect of the lag.
By Andrew Gao, William Gao
The paper presents a method for reliably scheduling operator execution in real‑time inference on small devices, achieving exact peak‑memory predictions and certified optimality. By profiling the generated program and modeling its memory usage, the authors can permute statements to match the modeled peak, guaranteeing bitwise‑identical results on CPUs and characterizing GPU limitations. They also provide efficient lower‑bound traps and analyze selective recomputation, showing it is NP‑hard in general but can reduce peak memory on certain architectures.
By Ivan Fomin, Ivan Koryabkin, Boris Sheludko
arXiv:2607. 26515v1 Announce Type: new Abstract: We present, to our knowledge, the first end-to-end FP4 RL post-training, in which both the rollout and training policies, including their forward and backward passes, operate at 4-bit precision.
By Hei Yi Mak, Shadan Golestan, Hoang Le, Mehran Taghian Jazi, Yunke Peng, Yaoyuan Wang, Yao Wang, Junsong Wang, Tianchi Hu, Fengchen He, Guipeng Hu, Tanzila Rahman, Anandharaju Durai Raju
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...
By Ephrem Wu
arXiv:2610. 11475v1 Announce Type: new Abstract: Population gradient flow is a common tool for reasoning about how neural networks adapt, including after pretraining.
By Ruoyu Zhao, Mingxuan Zhang, Jianbo Dai, Jiaqi Wu, Chenyu Zhu, Tong Che
arXiv:2607. 14568v1 Announce Type: cross Abstract: A companion study ran a 35B mixture-of-experts model on a 2011 NVIDIA Tesla C2075 (Fermi, sm_20, 6GB) as a GPU-prefill/CPU-decode hybrid, because the 4-bit model did not fit in device memory (arXiv:2606.
By A. C. Opus, J. Q. Lu
The paper investigates how identical model weights can yield different next‑token distributions when run by different serving programs, due to variations in precision, kernel implementation, and batching. It introduces the concept of kernel contracts that bound how much a serving kernel may deviate from a reference kernel on each traffic slice, and proves non‑asymptotic bounds linking logit differences to divergence metrics, reward changes, policy‑gradient bias, and greedy‑token flips. Experiments on Qwen2.5‑7B‑Instruct confirm that these bounds hold across prompts and that precision or engine changes can alter token outputs, especially on undecided prompts, highlighting the need for per‑slice contract checks.
By Lan Wu, Bruce Changlong Xu