Serving long-context LLMs is costly because attention computation grows linearly with context length. Dynamic sparse attention algorithms (DSAs) mitigate this by attending only to the key-value (KV) cache of critical tokens. However, with DSAs, the main performance bottleneck shifts from HBM bandwidth to HBM capacity: KV caches for unselected tokens must remain in HBM for low-latency decoding, constraining parallel batch size and stalling further throughput gains. Offloading these underutilized KV caches to DRAM could free HBM capacity, allowing larger parallel batch sizes. Yet, achieving such hierarchical HBM-DRAM storage raises new challenges, including fragmented KV cache access, HBM cache contention, and high HBM demands of hybrid batching, that remain unresolved in prior work. This paper proposes SparseServe, an LLM serving system that unlocks the parallel potential of DSAs through efficient hierarchical HBM-DRAM management. SparseServe introduces three key innovations to address the challenges mentioned above: (1) fragmentation-aware KV cache transfer, which accelerates HBM-DRAM data movement through GPU-direct loading (FlashH2D) and CPU-assisted saving (FlashD2H); (2) working-set-aware batch size control that adjusts batch sizes based on real-time working set estimation to minimize HBM cache thrashing; (3) layer-segmented prefill that bounds HBM use during prefill to a single layer, enabling efficient execution even for long prompts. Extensive experimental results demonstrate that SparseServe achieves up to 9.26x lower mean time-to-first-token (TTFT) latency and up to 3.14x higher token generation throughput compared to state-of-the-art LLM serving systems.

Recent breakthroughs in reasoning language models have significantly advanced text-based reasoning. On the other hand, Multi-modal Large Language Models (MLLMs) still lag behind, hindered by their outdated internal LLMs. Upgrading these is often prohibitively expensive, as it requires complete vision-language alignment retraining which is costly. To address this issue, we introduce Perception-Reasoning Decoupling, which modularizes the MLLM's reasoning component and makes it easily replaceable. This approach redefines the MLLM's role to convert multi-modal inputs into detailed textual outputs that can be processed by any powerful, external, text-only LLM reasoners. To align the MLLM's perceptual output with the final reasoning task, we propose a novel reinforcement learning algorithm called Visual Perception Optimization (VPO). VPO rewards the MLLM based on the correctness of answers generated by the external reasoner to produce faithful and query-relevant captions. Together, this decoupling pipeline and VPO form our Reasoning-Aligned PerceptIon Decoupling (RAPID) approach. Empirical results show that RAPID achieves significant performance gains on multi-modal reasoning benchmarks. Crucially, RAPID enables a novel inference-time scaling paradigm: Once trained with VPO, the MLLM can be paired with any state-of-the-art LLM reasoner for consistent performance improvement without retraining.

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.

Recent advancements in multimodal slow-thinking systems have demonstrated remarkable performance across diverse visual reasoning tasks. However, their capabilities in text-rich image reasoning tasks remain understudied due to the lack of a systematic benchmark. To address this gap, we propose OCR-Reasoning, a comprehensive benchmark designed to systematically assess Multimodal Large Language Models on text-rich image reasoning tasks. The benchmark comprises 1,069 human-annotated examples spanning 6 core reasoning abilities and 18 practical reasoning tasks in text-rich visual scenarios. Furthermore, unlike other text-rich image understanding benchmarks that only annotate the final answers, OCR-Reasoning also annotates the reasoning process simultaneously. With the annotated reasoning process and the final answers, OCR-Reasoning evaluates not only the final answers generated by models but also their reasoning processes, enabling a holistic analysis of their problem-solving abilities. Leveraging this benchmark, we conducted a comprehensive evaluation of state-of-the-art MLLMs. Our results demonstrate the limitations of existing methodologies. Notably, even state-of-the-art MLLMs exhibit substantial difficulties, with none achieving accuracy surpassing 50\% across OCR-Reasoning, indicating that the challenges of text-rich image reasoning are an urgent issue to be addressed. The benchmark and evaluation scripts are available at this https URL

Text-driven diffusion models have unlocked unprecedented abilities in image generation, whereas their video counterpart still lags behind due to the excessive training cost of temporal modeling. Besides the training burden, the generated videos also suffer from appearance inconsistency and structural flickers, especially in long video synthesis. To address these challenges, we design a \emph{training-free} framework called \textbf{ControlVideo} to enable natural and efficient text-to-video generation. ControlVideo, adapted from ControlNet, leverages coarsely structural consistency from input motion sequences, and introduces three modules to improve video generation. Firstly, to ensure appearance coherence between frames, ControlVideo adds fully cross-frame interaction in self-attention modules. Secondly, to mitigate the flicker effect, it introduces an interleaved-frame smoother that employs frame interpolation on alternated frames. Finally, to produce long videos efficiently, it utilizes a hierarchical sampler that separately synthesizes each short clip with holistic coherency. Empowered with these modules, ControlVideo outperforms the state-of-the-arts on extensive motion-prompt pairs quantitatively and qualitatively. Notably, thanks to the efficient designs, it generates both short and long videos within several minutes using one NVIDIA 2080Ti. Code is available at this https URL.