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.
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...
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. 13057v1 Announce Type: cross Abstract: In expert-parallel (EP) MoE serving, every layer synchronizes at the slowest GPU.
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.
arXiv:2608.23816v1 Announce Type: new Abstract: Quantized fine-tuning (QLoRA) saves memory but not time. It dequantizes every 4-bit weight on the fly, so it trains more slowly than fp16 LoRA. We pres...
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:2606. 26997v1 Announce Type: cross Abstract: Large language model (LLM) post-training for reasoning increasingly relies on reinforcement learning with verifiable rewards (RLVR), where models learn from ground-truth feedback on mathematical, logical, and scientific tasks.
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.
DUET is a token‑budget allocation method for reinforcement learning with verifiable rewards that jointly optimizes which prompts receive rollouts and how long each rollout runs. By using a lightweight surrogate to assign rollouts and a marker‑gated abort rule to stop them, DUET improves reasoning quality and reduces training time compared to full‑budget GRPO and other baseline methods. Experiments on Qwen3‑1.7B and other LLMs show that DUET outperforms baselines even with only 50% of the token budget, achieving up to 2.51× speedup while maintaining or improving performance.
arXiv:2504. 15610v4 Announce Type: replace Abstract: Fine-tuning a 7B language model for specialized advising is attractive in resource-constrained settings, but multi-epoch runs routinely exceed the wall-clock limits of the free-tier GPUs (Kaggle, Colab) such users rely on.
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:2607. 09375v1 Announce Type: new Abstract: We present Mach-Mind-4-Flash, a 35B-parameter Mixture-of-Experts (MoE) agentic model with 3B activated parameters.