Accelerating Disaggregated RL for Visual Generative LLMs with Diffusion-Based Parallelism and Trainer-Assisted Generation

Reinforcement learning (RL) has become a dominant post-training paradigm, driving the emergence of high-performance RL systems such as veRL for autoregressive large language models (LLMs). In parallel, diffusion-oriented RL algorithms, e.g., DanceGRPO and FlowGRPO, have rapidly expanded the scope of RL from language reasoning to diffusion-based visual and flow-based generation. However, efficient RL systems for diffusion generative LLMs remain underexplored. Existing implementations, e.g., veRL-Omni, still rely on colocated execution, which simplifies synchronization but couples rollout and training resources, limits heterogeneous deployment, and constrains independent scaling. To this end, we introduce DigenRL, a disaggregated RL framework for diffusion-based generative LLMs that supports flexible resource allocation, accommodates heterogeneous GPUs, and facilitates efficient task scheduling. To maximally reduce the execution bubbles in the disaggregated architecture, we propose: 1) a generation-axis pipeline (GAP) and time-step parallelism (TSP) in the diffusion architecture to enable finer-grained pipelining between rollout and training; 2) an elastic trainer-assisted generation (TAG) approach to enable the trainer GPU resources to dynamically assist in executing rollout generations; and 3) a tightly one-step constrained asynchronous strategy to further utilize the tail bubble in the pipeline. Extensive experiments are conducted on three hardware testbeds with 16-32 GPUs using HunyuanVideo-13B, Wan2.1-14B, FLUX.1-12B, and QwenImage-20B generative models. Experimental results show that DigenRL achieves 1.56-2.10x throughput improvements over state-of-the-art diffusion RL systems, veRL-Omni and GenRL.

GNMR: Runtime Stability Control for Low-Precision Large Language Model Training

Training stability is a key bottleneck in low-precision language model training: efficient low-cost paths can still produce short-lived numerical risks at a small set of operators. We formulate this as runtime stability control and present Gradient Norm-to-Mean Ratio (GNMR), a lightweight controller that compares each recoverable unit's current gradient norm with its historical mean. Together with $Δ$-GNMR for abrupt short-window increases, GNMR maps local risk signals to bounded recovery actions under a hard $\mathrm{maxO}$ budget and a short lock interval, without changing the numerical format, kernel, or backend recipe. Across activation-quantization stress, DeepSeek-style recipe-level training, and LLaMA-2 13B fine-tuning, GNMR preserves high-fidelity quality with sparse, budgeted recovery. These results support GNMR as a backend-agnostic controller to improve low-precision training stability while preserving low-cost execution.

TurboGR: An Accelerated Training System for Large-Scale Generative Recommendation

Generative recommendation (GR) has emerged as a promising paradigm that replaces fragmented, scenario-specific architectures with unified Transformer-based models, exhibiting scaling-law behavior where recommendation quality improves systematically with increased model capacity and training data. However, deploying GR at scale on Ascend NPUs faces fundamental system-level challenges. These challenges are further exacerbated on Ascend NPUs due to the absence of high-performance implementations for jagged operators and the architectural mismatch between irregular sparse primitives and NPU's dense-computation-optimized design. In this paper, we present \model, an Ascend-affinity training system for generative recommendation that systematically addresses these bottlenecks through three core innovations: (i) Ascend-affinity jagged acceleration, including fusion operators that eliminate padding redundancy and dynamic load balancing that reduces inter-device imbalance from 47\% to 2.4\%; (ii) distributed communication optimization, comprising hierarchical sparse parallelism, semi-asynchronous training with proven convergence guarantees, and fine-grained pipeline orchestration that sustains 94\% NPU utilization; and (iii) negative sampling optimization via asynchronous offloading, jaggedness-aware FP16 quantization, and intra-batch logit sharing that expand the effective negative space without additional embedding lookups. Evaluated on the KuaiRand-27K dataset, \model supports training at up to 0.2B parameters and achieves 54.71\% MFU with near-linear scalability (0.97).

HiFloat4 Format for Language Model Pre-training on Ascend NPUs

Large foundation models have become central to modern machine learning, with performance scaling predictably with model size and data. However, training and deploying such models incur substantial computational and memory costs, motivating the development of low-precision training techniques. Recent work has demonstrated that 4-bit floating-point (FP4) formats--such as MXFP4 and NVFP4--can be successfully applied to linear GEMM operations in large language models (LLMs), achieving up to 4x improvements in compute throughput and memory efficiency compared to higher-precision baselines. In this work, we investigate the recently proposed HiFloat4 FP4 format for Huawei Ascend NPUs and systematically compare it with MXFP4 in large-scale training settings. All experiments are conducted on Ascend NPU clusters, with linear and expert GEMM operations performed entirely in FP4 precision. We evaluate both dense architectures (e.g., Pangu and LLaMA-style models) and mixture-of-experts (MoE) models, where both standard linear layers and expert-specific GEMMs operate in FP4. Furthermore, we explore stabilization techniques tailored to FP4 training that significantly reduce numerical degradation, maintaining relative error within 1% of full-precision baselines while preserving the efficiency benefits of 4-bit computation. Our results provide a comprehensive empirical study of FP4 training on NPUs and highlight the practical trade-offs between FP4 formats in large-scale dense and MoE models.

AgentCollab: A Self-Evaluation-Driven Collaboration Paradigm for Efficient LLM Agents

Autonomous agents powered by large language models (LLMs) perform complex tasks through long-horizon reasoning and tool interaction, where a fundamental trade-off arises between execution efficiency and reasoning robustness. Models at different capability-cost levels offer complementary advantages: lower-cost models enable fast execution but may struggle on difficult reasoning segments, while stronger models provide more robust reasoning at higher computational cost. We present AgentCollab, a self-driven collaborative inference framework that dynamically coordinates models with different reasoning capacities during agent execution. Instead of relying on external routing modules, the framework uses the agent's own self-reflection signal to determine whether the current reasoning trajectory is making meaningful progress, and escalates control to a stronger reasoning tier only when necessary. To further stabilize long-horizon execution, we introduce a difficulty-aware cumulative escalation strategy that allocates additional reasoning budget based on recent failure signals. In our experiments, we instantiate this framework using a two-level small-large model setting. Experiments on diverse multi-step agent benchmarks show that AgentCollab consistently improves the accuracy-efficiency Pareto frontier of LLM agents.

RelayGR: Scaling Long-Sequence Generative Recommendation via Cross-Stage Relay-Race Inference

Real-time recommender systems execute multi-stage cascades (retrieval, pre-processing, fine-grained ranking) under strict tail-latency SLOs, leaving only tens of milliseconds for ranking. Generative recommendation (GR) models can improve quality by consuming long user-behavior sequences, but in production their online sequence length is tightly capped by the ranking-stage P99 budget. We observe that the majority of GR tokens encode user behaviors that are independent of the item candidates, suggesting an opportunity to pre-infer a user-behavior prefix once and reuse it during ranking rather than recomputing it on the critical path. Realizing this idea at industrial scale is non-trivial: the prefix cache must survive across multiple pipeline stages before the final ranking instance is determined, the user population implies cache footprints far beyond a single device, and indiscriminate pre-inference would overload shared resources under high QPS. We present RelayGR, a production system that enables in-HBM relay-race inference for GR. RelayGR selectively pre-infers long-term user prefixes, keeps their KV caches resident in HBM over the request lifecycle, and ensures the subsequent ranking can consume them without remote fetches. RelayGR combines three techniques: 1) a sequence-aware trigger that admits only at-risk requests under a bounded cache footprint and pre-inference load, 2) an affinity-aware router that co-locates cache production and consumption by routing both the auxiliary pre-infer signal and the ranking request to the same instance, and 3) a memory-aware expander that uses server-local DRAM to capture short-term cross-request reuse while avoiding redundant reloads. We implement RelayGR on Huawei Ascend NPUs and evaluate it with real queries. Under a fixed P99 SLO, RelayGR supports up to 1.5$\times$ longer sequences and improves SLO-compliant throughput by up to 3.6$\times$.

Towards Universal Offline Black-Box Optimization via Learning Language Model Embeddings

The pursuit of universal black-box optimization (BBO) algorithms is a longstanding goal. However, unlike domains such as language or vision, where scaling structured data has driven generalization, progress in offline BBO remains hindered by the lack of unified representations for heterogeneous numerical spaces. Thus, existing offline BBO approaches are constrained to single-task and fixed-dimensional settings, failing to achieve cross-domain universal optimization. Recent advances in language models (LMs) offer a promising path forward: their embeddings capture latent relationships in a unifying way, enabling universal optimization across different data types possible. In this paper, we discuss multiple potential approaches, including an end-to-end learning framework in the form of next-token prediction, as well as prioritizing the learning of latent spaces with strong representational capabilities. To validate the effectiveness of these methods, we collect offline BBO tasks and data from open-source academic works for training. Experiments demonstrate the universality and effectiveness of our proposed methods. Our findings suggest that unifying language model priors and learning string embedding space can overcome traditional barriers in universal BBO, paving the way for general-purpose BBO algorithms. The code is provided at https://github.com/lamda-bbo/universal-offline-bbo.

Pangu Pro MoE: Mixture of Grouped Experts for Efficient Sparsity

The surgence of Mixture of Experts (MoE) in Large Language Models promises a small price of execution cost for a much larger model parameter count and learning capacity, because only a small fraction of parameters are activated for each input token. However, it is commonly observed that some experts are activated far more often than others, leading to system inefficiency when running the experts on different devices in parallel. Therefore, we introduce Mixture of Grouped Experts (MoGE), which groups the experts during selection and balances the expert workload better than MoE in nature. It constrains tokens to activate an equal number of experts within each predefined expert group. When a model execution is distributed on multiple devices, this architectural design ensures a balanced computational load across devices, significantly enhancing throughput, particularly for the inference phase. Further, we build Pangu Pro MoE on Ascend NPUs, a sparse model based on MoGE with 72 billion total parameters, 16 billion of which are activated for each token. The configuration of Pangu Pro MoE is optimized for Ascend 300I Duo and 800I A2 through extensive system simulation studies. Our experiments indicate that MoGE indeed leads to better expert load balancing and more efficient execution for both model training and inference on Ascend NPUs. The inference performance of Pangu Pro MoE achieves 1148 tokens/s per card and can be further improved to 1528 tokens/s per card by speculative acceleration, outperforming comparable 32B and 72B Dense models. Furthermore, we achieve an excellent cost-to-performance ratio for model inference on Ascend 300I Duo. Our studies show that Ascend NPUs are capable of training Pangu Pro MoE with massive parallelization to make it a leading model within the sub-100B total parameter class, outperforming prominent open-source models like GLM-Z1-32B and Qwen3-32B.

TrimR: Verifier-based Training-Free Thinking Compression for Efficient Test-Time Scaling

Large Reasoning Models (LRMs) demonstrate exceptional capability in tackling complex mathematical, logical, and coding tasks by leveraging extended Chain-of-Thought (CoT) reasoning. Test-time scaling methods, such as prolonging CoT with explicit token-level exploration, can push LRMs' accuracy boundaries, but they incur significant decoding overhead. A key inefficiency source is LRMs often generate redundant thinking CoTs, which demonstrate clear structured overthinking and underthinking patterns. Inspired by human cognitive reasoning processes and numerical optimization theories, we propose TrimR, a verifier-based, training-free, efficient framework for dynamic CoT compression to trim reasoning and enhance test-time scaling, explicitly tailored for production-level deployment. Our method employs a lightweight, pretrained, instruction-tuned verifier to detect and truncate redundant intermediate thoughts of LRMs without any LRM or verifier fine-tuning. We present both the core algorithm and asynchronous online system engineered for high-throughput industrial applications. Empirical evaluations on Ascend NPUs and vLLM show that our framework delivers substantial gains in inference efficiency under large-batch workloads. In particular, on the four MATH500, AIME24, AIME25, and GPQA benchmarks, the reasoning runtime of Pangu-R-38B, QwQ-32B, and DeepSeek-R1-Distill-Qwen-32B is improved by up to 70% with negligible impact on accuracy.

Pangu Ultra MoE: How to Train Your Big MoE on Ascend NPUs

Sparse large language models (LLMs) with Mixture of Experts (MoE) and close to a trillion parameters are dominating the realm of most capable language models. However, the massive model scale poses significant challenges for the underlying software and hardware systems. In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs. The key goals are better usage of the computing resources under the dynamic sparse model structures and materializing the expected performance gain on the actual hardware. To select model configurations suitable for Ascend NPUs without repeatedly running the expensive experiments, we leverage simulation to compare the trade-off of various model hyperparameters. This study led to Pangu Ultra MoE, a sparse LLM with 718 billion parameters, and we conducted experiments on the model to verify the simulation results. On the system side, we dig into Expert Parallelism to optimize the communication between NPU devices to reduce the synchronization overhead. We also optimize the memory efficiency within the devices to further reduce the parameter and activation management overhead. In the end, we achieve an MFU of 30.0% when training Pangu Ultra MoE, with performance comparable to that of DeepSeek R1, on 6K Ascend NPUs, and demonstrate that the Ascend system is capable of harnessing all the training stages of the state-of-the-art language models. Extensive experiments indicate that our recipe can lead to efficient training of large-scale sparse language models with MoE. We also study the behaviors of such models for future reference.