The paper introduces CHARM+, a Cross‑Hardware Attention framework that reorganizes sequential communications into multiple levels to enable intra‑level parallelism and overlap computation with communication. It decouples cross‑hardware from intra‑hardware communication, removing synchronization overhead caused by bandwidth imbalance. Experiments on dual‑NUMA PCIe and dual‑node NVLink systems show consistent speedups over state‑of‑the‑art methods, achieving 1.3×/1.7× and 2.8×/2.9× improvements for forward and backward passes respectively.
arXiv:2512. 20968v2 Announce Type: replace-cross Abstract: Distributed attention is essential for scaling large language models (LLMs) to long contexts, yet existing methods either have limited parallelism or incur high communication costs.
arXiv:2610.07098v1 Announce Type: cross Abstract: Large transformer-based models increasingly depend on multi-GPU execution, which requires frequent collective communication among GPUs. Existing comm...
AMMA is a multi‑chiplet, memory‑centric architecture designed to reduce latency in large‑context attention for language models. It replaces GPU compute dies with HBM‑PNM cubes, doubling memory bandwidth and employing a logic‑die microarchitecture, two‑level hybrid parallelism, and a reordered collective flow to minimize communication overhead. Design‑space exploration and evaluations show AMMA achieves 15.5× lower attention latency and 6.9× lower energy consumption compared with the NVIDIA H100.
arXiv:2504. 17584v2 Announce Type: replace-cross Abstract: Attention-FC Disaggregated (AFD) LLM inference systems offload memory-bound Attention operations to memory-rich accelerators (e.
The paper introduces QUILT, a new sparse‑attention execution mechanism that jointly processes neighboring queries to reuse shared key‑value (KV) entries, thereby reducing redundant memory traffic and computation. QUILT employs Shift‑and‑Compare Set Decomposition (SCSD) to convert irregular set operations into regular data‑parallel primitives and pipelines this with attention computation to hide overhead. The approach includes cascaded sharing across multiple granularities and a tile‑aware strategy that balances sharing with hardware utilization, achieving up to 55.1% lower kernel latency and 36.8% faster time‑to‑first‑token on LongBench benchmarks with negligible accuracy loss.
arXiv:2606. 30460v1 Announce Type: new Abstract: In this paper, we aim to combine the advantages of existing sequence parallelism paradigms and overcomes their drawbacks, the most serious of which is the incapability to correctly compute causal attention on the hybrid-context packed sequences, in a stronger sequence parallelism framework.
arXiv:2511. 04791v2 Announce Type: replace Abstract: Modern LLM serving systems must sustain high throughput while meeting strict latency SLOs across two distinct inference phases: compute-intensive prefill and memory-bound decode phases.
arXiv:2606. 30460v2 Announce Type: replace Abstract: In this paper, we aim to combine the advantages of existing sequence parallelism paradigms and overcomes their drawbacks, the most serious of which is the incapability to correctly compute causal attention on the hybrid-context packed sequences, in a stronger sequence parallelism framework.
arXiv:2609.13585v1 Announce Type: cross Abstract: Communication has become a bottleneck in distributed training and inference of large models. Overlapping communication with computation at the granul...
arXiv:2602. 21788v2 Announce Type: replace-cross Abstract: Scaling long-context capabilities is crucial for Large Language Models (LLMs).
The paper introduces BASP, a batch‑aware sequence parallelism method that partitions GPUs into disjoint groups based on micro‑batch size to reduce all‑to‑all communication. By localizing communication, BASP improves training efficiency for long‑context LLMs. Experiments on NVIDIA A100 clusters show up to 1.17‑1.31× faster end‑to‑end training on Llama and Qwen models while maintaining the same accuracy and memory usage.
arXiv:2606. 16352v1 Announce Type: cross Abstract: Computation integrity of remote large language model (LLM) serving can be questionable.