Junhao Hu

Yang Liu, ZhaoKai Luo, HuaYi Jin et al.

As the input length of large language model (LLM) serving continues to grow, the KV cache has become a dominant bottleneck in AI infrastructure. It limits GPU memory capacity, serving concurrency, cache reuse, and distributed scalability. Several important problems, including position-independent KV cache, prefix KV cache compression, hot/cold KV cache separation, and distributed KV cache management, all depend on how the KV cache is represented and managed. However, existing serving systems largely rely on a monolithic KV cache abstraction, where the KV cache is treated as a homogeneous sequence of token-level memory blocks and managed with similar policies across attention heads and serving scenarios. We observe that KV cache utility is highly structured across KV heads: different heads exhibit different functional roles, attention distances, and runtime importance. Therefore, a full KV cache is not always necessary for every head, token range, or serving scenario. We present RedKnot, a head-aware KV cache management system for LLM serving. RedKnot breaks the conventional monolithic KV cache abstraction by decomposing the KV cache along KV heads, whose importance and effective attention ranges vary significantly across serving scenarios. This head-level decomposition turns the KV cache from a monolithic tensor abstraction into a structured memory object, enabling RedKnot to uniformly support position-independent KV reuse, prefix KV compression, hot/cold KV separation, and distributed KV placement while preserving output fidelity and improving resource efficiency, without requiring model retraining or fine-tuning. RedKnot establishes a new foundation for AI infrastructure by transforming the KV cache from a monolithic, passive runtime artifact into a dynamic, model-aware runtime substrate for scalable LLM serving.

Xiaomi LLM-Core Team, Dong Zhang, Gang Wang et al.

Existing audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe the emergence of few-shot learning capabilities across a diverse set of audio tasks. We develop a systematic evaluation of these capabilities and find that MiMo-Audio-7B-Base achieves SOTA performance on both speech intelligence and audio understanding benchmarks among open-source models. Beyond standard metrics, MiMo-Audio-7B-Base generalizes to tasks absent from its training data, such as voice conversion, style transfer, and speech editing. MiMo-Audio-7B-Base also demonstrates powerful speech continuation capabilities, capable of generating highly realistic talk shows, recitations, livestreaming and debates. At the post-training stage, we curate a diverse instruction-tuning corpus and introduce thinking mechanisms into both audio understanding and generation. MiMo-Audio-7B-Instruct achieves open-source SOTA on audio understanding benchmarks (MMSU, MMAU, MMAR, MMAU-Pro), spoken dialogue benchmarks (Big Bench Audio, MultiChallenge Audio) and instruct-TTS evaluations, approaching or surpassing closed-source models. Model checkpoints and full evaluation suite are available at https://github.com/XiaomiMiMo/MiMo-Audio.

Shiju Zhao, Junhao Hu, Jiaqi Zheng et al.

The Key-Value (KV) cache of Large Language Models (LLMs) is prefix-based, making it highly inefficient for processing contexts retrieved in arbitrary order. Position-Independent Caching (PIC) has been proposed to enable KV reuse without positional constraints; however, existing approaches often incur substantial accuracy degradation, limiting their practical adoption. To address this issue, we propose native PIC by reintroducing the encoder to prevalent decoder-only LLMs and explicitly training it to support PIC. We further develop COMB, a PIC-aware caching system that integrates seamlessly with existing inference frameworks. Experimental results show that COMB reduces Time-to-First-Token (TTFT) by 51-94% and increases throughput by 3$\times$ with comparable accuracy. Furthermore, the quality improvement when using DeepSeek-V2-Lite-Chat demonstrates the applicability of COMB to other types of decoder-only LLMs. Our code is available at https://github.com/shijuzhao/Comb.

Junhao Hu, Shirin Shoushtari, Zihao Zou et al.

Deep model-based architectures (DMBAs) are widely used in imaging inverse problems to integrate physical measurement models and learned image priors. Plug-and-play priors (PnP) and deep equilibrium models (DEQ) are two DMBA frameworks that have received significant attention. The key difference between the two is that the image prior in DEQ is trained by using a specific measurement model, while that in PnP is trained as a general image denoiser. This difference is behind a common assumption that PnP is more robust to changes in the measurement models compared to DEQ. This paper investigates the robustness of DEQ priors to changes in the measurement models. Our results on two imaging inverse problems suggest that DEQ priors trained under mismatched measurement models outperform image denoisers.

Tiancheng Hu, Jin Qin, Zheng Wang et al.

Disaggregation maps parts of an AI workload to different types of GPUs, offering a path to utilize modern heterogeneous GPU clusters. However, existing solutions operate at a coarse granularity and are tightly coupled to specific model architectures, leaving much room for performance improvement. This paper presents Tessera, the first kernel disaggregation system to improve performance and cost efficiency on heterogeneous GPUs for large model inference. Our key insight is that kernels within a single application exhibit diverse resource demands, making them the most suitable granularity for aligning computation with hardware capabilities. Tessera integrates offline analysis with online adaptation by extracting precise inter-kernel dependencies from PTX to ensure correctness, overlapping communication with computation through a pipelined execution model, and employing workload-aware scheduling with lightweight runtime adaptation. Extensive evaluations across five heterogeneous GPUs and four model architectures, scaling up to 16 GPUs, show that Tessera improves serving throughput and cost efficiency by up to 2.3x and 1.6x, respectively, compared to existing disaggregation methods, while generalizing to model architectures where prior approaches do not apply. Surprisingly, a heterogeneous GPU pair under Tessera can even exceed the throughput of two homogeneous high-end GPUs at a lower cost.

Junhao Hu, Fangze Li, Mingtao Xu et al.

Large language models (LLMs) demonstrate strong capabilities across a wide range of complex tasks and are increasingly deployed at scale, placing significant demands on inference efficiency. Prior work typically decomposes inference into prefill and decode stages, with the decode stage dominating total latency. To reduce time and memory complexity in the decode stage, a line of work introduces sparse-attention algorithms. In this paper, we show, both empirically and theoretically, that sparse attention can paradoxically increase end-to-end complexity: information loss often induces significantly longer sequences, a phenomenon we term ``Less is Less'' (Lil). To mitigate the Lil problem, we propose an early-stopping algorithm that detects the threshold where information loss exceeds information gain during sparse decoding. Our early-stopping algorithm reduces token consumption by up to 90% with a marginal accuracy degradation of less than 2% across reasoning-intensive benchmarks.

Yuanqing Wang, Yuchen Zhang, Hao Lin et al.

Modern large language model (LLM) training is inherently dynamic: resource fluctuations, RLHF phase shifts, and cluster elasticity continually reshape the optimal parallelism layout, posing a significant challenge to existing training frameworks built around a static execution model. We present DynaTrain, a distributed training system for sub-second, online reconfiguration across arbitrary multi-dimensional parallelism. At its core, we propose a Virtual Parameter Space (VPS) abstraction that unifies all distributed training states under one logical coordinate space, turning any parallelism configuration into a deterministic mapping and collapsing complex transition into manageable geometric intersections. On top of VPS, a state routing-and-transition layer executes rank-local transfers under a memory-aware, deadlock-free schedule, and an Elastic Device Manager overlaps new-world construction with ongoing training to mask topology-change cost. On dense and MoE models up to 235B parameters, DynaTrain reconfigures a 70B dense model in under 2s and a 235B MoE model in 4.36s, outperforming state-of-the-art checkpoint-based and elastic systems by up to three orders of magnitude while preserving correctness.

Junhao Hu, Wenrui Huang, Weidong Wang et al.

Large Language Models (LLMs) show great capabilities in a wide range of applications, but serving them efficiently becomes increasingly challenging as requests (prompts) become more complex. Context caching improves serving performance by reusing Key-Value (KV) vectors, the intermediate representations of tokens that are repeated across requests. However, existing context caching requires exact prefix matches across requests, limiting reuse cases in settings such as few-shot learning and retrieval-augmented generation, where immutable content (e.g., documents) remains unchanged across requests but is preceded by varying prefixes. Position-Independent Caching (PIC) addresses this issue by enabling modular reuse of the KV vectors regardless of prefixes. We formalize PIC and advance prior work by introducing EPIC, a serving system incorporating our new LegoLink algorithm, which mitigates the inappropriate "attention sink" effect at every document beginning, to maintain accuracy with minimal computation. Experiments show that EPIC achieves up to 8x improvements in Time-To-First-Token (TTFT) and 7x throughput gains over existing systems, with negligible or no accuracy loss.

Junhao Hu, Wenrui Huang, Weidong Wang et al.

Large Language Models (LLMs) have demonstrated strong capabilities across various domains, with recent advancements in challenging reasoning tasks such as mathematics and programming. However, solving reasoning tasks often requires an LLM to generate long sequences, incurring $O(N)$ time and memory complexities per token, where $N$ is the current sequence length. To reduce complexities, existing sparsity-based algorithms propose to retain Key-Value (KV) vectors, the intermediate representations of only the most critical tokens. However, these algorithms struggle with the "impossible trinity" of accuracy, time, and memory. For example, the state-of-the-art algorithm, Quest, achieves high accuracy with $O(L)$ time but $O(N)$ memory ($L$ is the cache budget, $L \ll N$). To address the "impossible trinity", in this paper, we identify a new attention pattern during the decode stage of reasoning tasks, where milestone tokens (analogous to lemmas in mathematical proofs) emerge, are utilized, and then become unimportant afterward. Based on this pattern, we propose a new algorithm RaaS that identifies milestone tokens and retains their KV vectors until they are no longer needed, achieving high accuracy with $O(L)$ time and $O(L)$ memory complexities.

Zixiao Huang, Junhao Hu, Hao Lin et al.

The rapid scaling of large language models (LLMs) has significantly increased GPU memory pressure, which is further aggravated by training optimization techniques such as virtual pipeline and recomputation that disrupt tensor lifespans and introduce considerable memory fragmentation. Default GPU memory allocators of popular deep learning frameworks like PyTorch use online strategies without knowledge of tensor lifespans, which can waste up to 43\% of memory and cause out-of-memory errors, rendering optimization techniques ineffective or even unusable. To address this, we introduce STWeaver, a GPU memory allocator for deep learning frameworks that reduces fragmentation by exploiting the spatial and temporal regularity in memory allocation behaviors of training workloads. STWeaver introduces a novel paradigm that combines offline planning with online allocation. The offline planning leverages spatio-temporal regularities to generate a near-optimal allocation plan, while the online allocation handles complex and dynamic models such as Mixture-of-Experts (MoE). Built as a pluggable PyTorch allocator, STWeaver reduces fragmentation ratio on average by 79.2\% (up to 100\%) across both dense and sparse models, with negligible overhead. This enables more efficient, high-throughput training configurations and improves performance by up to 32.5\%.

Shiju Zhao, Junhao Hu, Rongxiao Huang et al.

The context caching technique is employed to accelerate the Multimodal Large Language Model (MLLM) inference by prevailing serving platforms currently. However, this approach merely reuses the Key-Value (KV) cache of the initial sequence of prompt, resulting in full KV cache recomputation even if the prefix differs slightly. This becomes particularly inefficient in the context of interleaved text and images, as well as multimodal retrieval-augmented generation. This paper proposes position-independent caching as a more effective approach for multimodal information management. We have designed and implemented a caching system, named MPIC, to address both system-level and algorithm-level challenges. MPIC stores the KV cache on local disks when receiving multimodal data, and calculates and loads the KV cache in parallel during inference. To mitigate accuracy degradation, we have incorporated the integrated reuse and recompute mechanism within the system. The experimental results demonstrate that MPIC can achieve up to 54\% reduction in response time and 2$\times$ improvement in throughput compared to existing context caching systems, while maintaining negligible or no accuracy loss.

Chao Yu, Yuanqing Wang, Zhen Guo et al.

Reinforcement learning (RL) has demonstrated immense potential in advancing artificial general intelligence, agentic intelligence, and embodied intelligence. However, the inherent heterogeneity and dynamicity of RL workflows often lead to low hardware utilization and slow training on existing systems. In this paper, we present RLinf, a high-performance RL training system based on our key observation that the major roadblock to efficient RL training lies in system flexibility. To maximize flexibility and efficiency, RLinf is built atop a novel RL system design paradigm called macro-to-micro flow transformation (M2Flow), which automatically breaks down high-level, easy-to-compose RL workflows at both the temporal and spatial dimensions, and recomposes them into optimized execution flows. Supported by RLinf worker's adaptive communication capability, we devise context switching and elastic pipelining to realize M2Flow transformation, and a profiling-guided scheduling policy to generate optimal execution plans. Extensive evaluations on both reasoning RL and embodied RL tasks demonstrate that RLinf consistently outperforms state-of-the-art systems, achieving 1.1x-2.13x speedup in end-to-end training throughput.