Flux Attention: Context-Aware Hybrid Attention for Efficient LLMs Inference

Quantong Qiu; Zhiyi Hong; Yi Yang; Haitian Wang; Kebin Liu; Qingqing Dang; Juntao Li; Min Zhang

cs.LGApr 8, 2026

Flux Attention: Context-Aware Hybrid Attention for Efficient LLMs Inference

Authors: Quantong Qiu, Zhiyi Hong, Yi Yang, Haitian Wang, Kebin Liu, Qingqing Dang, Juntao Li, Min Zhang

Organizations: School of Computer Science and Technology, Soochow University · Baidu Inc, China

Abstract

The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8 $\times$ A800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to $2.8\times$ and $2.0\times$ in the prefill and decode stages.

Figures & tables

Panel of Figure 1 — Figure 1 : Impact of sparsity on performance and decoding efficiency. (a) Certain tasks suffer performance collapse beyond a specific threshold; this sweep uses Streaming Sparse Attention (SSA) with the configuration of Appendix F.4 . (b) Layer-level sparsity achieves substantial decoding speedup, while head-level sparsity yields marginal speedup.

Figure 2 : Overview of our dynamic layer-level routing architecture. The model incorporates a Layer Router that assigns each layer to either FA or SA based on the input query $x_{Q}$ .

Method	S-Doc QA		M-Doc QA		Summ		In-Context			Synthetic		Code		Avg.
Method	Qasper	MF-en	HotQA	2Wiki	Gov.	M.News	TREC	TQA	SAMS	PCount	PRe	RB-P	Lcc	Perf.	$\mathbf{\Omega_{\mathrm{\bf MSR}}}$
Qwen3-4B backbone model
Qwen3-4B	35.21	52.16	44.81	32.15	33.47	23.45	70.67	88.22	39.74	2.33	96.84	50.84	57.93	48.45	-
+ DuoAttention	35.83	49.84	47.09	32.24	33.32	23.70	69.33	85.87	39.75	4.50	94.57	50.56	57.43	48.22	0.50
+ PruLong	34.15	50.78	44.48	32.89	32.96	23.53	67.67	88.69	39.55	3.17	90.17	49.00	54.07	47.16	0.50
+ TriangleMix	35.55	52.02	45.37	31.76	33.32	23.70	69.00	88.20	39.74	3.83	91.51	48.58	56.38	47.72	0.50

Table 1 : Performance on LongBench-E [ 1 ] . We report average performance (Perf.) and

\Omega_{\mathrm{MSR}}

per task category. The 1st and the 2nd performance in each comparison group are highlighted with bold font and underlined , respectively. Gray-shaded rows denote the sparse-decode configuration.

Panel of Figure 3 — Figure 3 : Speedup comparison across different context lengths. The dotted line represents the dense baseline performance (1.0x).

Figure 4 : Overview of the layer-wise routing activation frequencies in Llama-3.1-8B-Instruct. Dark blue indicates layers consistently routed to FA across all six tasks in LongBench-E, whereas light blue denotes layers consistently routed to SA.

Figure 5 : Comparison of performance and test-time

\Omega_{\mathrm{MSR}}

among different training sparsity target

\boldsymbol{t}

settings. The bar chart denotes the performance and the line chart denotes

\Omega_{\mathrm{MSR}}

in each task.

Figure 6 : Performance trajectories during continued training with a frozen Layer Router. The backbone effectively adapts its representations to the established sparse pathways, demonstrating steady improvement over time.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Selection unit	Selection	Primary phase	Deployment objective
Flux Attention (ours)	Whole layer	Learned, input-conditioned FA/SA	Prefill routing, decode reuse	Decode-regular execution, KV memory
Elastic Attention [ 42 ]	Head	Learned, test-time adaptive	Prefill/TTFT	Adaptive sparsity ratio
DuoAttention [ 48 ]	Head	Offline static allocation	Decode	Hybrid long-context attention
PruLong [ 4 ]	Head	Offline static allocation	Decode	Pruned long-context attention
TriangleMix [ 15 ]	Block pattern	Fixed triangle topology	Prefill	Training-free sparse backend
XAttention [ 53 ]	Block	Fixed block selection score	Prefill	Prefill acceleration

Appendix

Table 3: Positioning of Flux Attention against representative dynamic and fine-grained attention-selection methods. The axes are the unit at which computation is selected, whether the selection is learned or fixed, the inference phase the method primarily targets, and the deployment objective. Flux Attention is complementary to fine-grained operators: it chooses the coarse per-layer execution mode, while a method such as vAttention can supply fine-grained sparsity inside the routed sparse branch.

Source	Task family it covers	Regime exposed to the router
ChatQA2-Long-SFT-data [ 52 ]	Single-document and multi-document QA over long inputs	Retrieval-intensive
MuSiQue [ 44 ]	Multi-hop question answering	Retrieval-intensive
CoLT-132K [ 24 ]	Repository-level code completion	Context-holistic
GovReport [ 17 ]	Long-form summarization of government reports	Context-holistic
XSum [ 34 ]	Abstractive summarization	Context-holistic

Appendix

Table 4: Training mixture of the Layer Router (approximately 0.74B tokens over 1K–64K sequences). The mixture is chosen to expose the router to both retrieval-intensive and context-holistic regimes; it does not import the evaluation benchmark suites.

Method	Routing / selection unit	Training or calibration in our comparison	Deployment focus
FluxAttn (ours)	Whole layer, input-conditioned FA/SA	Router trained on the shared data mixture; backbone frozen	Decode-regular layer execution
PruLong	Offline head selection	Same data and environment; original method hyperparameters	Optimized block-sparse decode
DuoAttention	Head allocation	Same data and environment; original method hyperparameters	Hybrid long-context attention
TriangleMix / TA	Triangle Attention operator	Training-free	Sparse-attention backend

Appendix

Table 5: Baseline comparison controls. All trainable baselines are trained in the same environment and on the same data mixture as Flux Attention while retaining their original method-specific hyperparameters; TriangleMix’s Triangle Attention (TA) is training-free and serves as a sparse-attention backend for Flux rather than as a routed competitor.

Item	Setting
GPU	1 $\times$ NVIDIA A800 80GB (PCIe); batched-serving results in Table 14 on 8 $\times$ NVIDIA H20
Precision	BF16
Batch size	1 for profiling; 8/16/24/32 for the serving experiment
Dense attention path	FlashAttention-2 with the same PyTorch/CUDA environment as the sparse and hybrid paths
Sparse attention kernel	Block-Sparse-Attention, block size 64, sink size 128, chunk size 16384 ( Section F.4 )
Warm-up / profiling	10 warm-up iterations, 50 profiling iterations, one token per step

Appendix

Table 6: Systems setup for the latency and memory measurements reported in Figure 3 and Table 15 . All compared methods share the hardware, precision, input/output shapes, warm-up, repetition, and timing protocol; the attention execution path is the controlled difference.

Model & Training
Hyperparameter	Value
Base Model	Qwen, Llama
Sequence length	65536
Precision	bfloat16
Global Batch Size	48
Training Steps	300

Appendix

Table 7: Hyperparameters: General configuration.

Method	S-Doc QA		M-Doc QA		Summ		In-Context			Synthetic		Code		Avg.
Method	Qasper	MF-en	HotQA	2Wiki	Gov.	M.News	TREC	TQA	SAMS	PCount	PRe	RB-P	Lcc	Perf.	$\mathbf{\Omega_{\mathrm{\bf MSR}}}$
Qwen3-4B backbone model
Qwen3-4B	35.21	52.16	44.81	32.15	33.47	23.45	70.67	88.22	39.74	2.33	96.84	50.84	57.93	48.45	-
+ XAttention	32.80	50.36	44.54	33.17	33.79	23.73	69.00	87.44	39.90	3.42	74.59	49.42	59.62	46.44	-
+ LycheeDecode	34.12	52.63	44.74	31.53	31.16	23.24	68.67	87.69	39.12	4.00	96.87	49.17	57.29	47.83	0.50
+ DuoAttention	35.83	49.84	47.09	32.24	33.32	23.70	69.33	85.87	39.75	4.50	94.57	50.56	57.43	48.22	0.50

Appendix

Table 8: Performance on LongBench-E [ 1 ] . We report average performance (Perf.) and

\Omega_{\mathrm{MSR}}

per task category. The 1st and the 2nd performance in each comparison group are highlighted with bold font and underlined , respectively. Gray-shaded rows denote the sparse-decode configuration.

Parameter	Configuration Details
Context Windows	8k, 16k, 32k, 64k, 128k, 256k
Sample Size	50 samples per task-length pair
Evaluation Tasks	Retrieval (NIAH): Single_{1-3} , MultiKey_{1-3} , MultiQuery , MultiValue QA & Extraction: QA_{1,2} , FWE

Appendix

Table 9: Detailed configuration for the RULER benchmark evaluation. We evaluate across exponentially increasing context windows up to 256k tokens.

Backbone: Qwen3-4B
Models	NS-3	QA-2	NMK-3	NS-2	FWE	QA-1	NMQ	NMV	NMK-2	NS-1	NMK-1	Avg.
Qwen3-4B	95.02	35.59	37.54	80.00	84.47	40.14	66.19	65.91	47.39	100.00	73.05	66.00
+ LycheeDecode	72.33	30.67	5.00	64.67	71.33	27.33	55.58	50.58	25.00	100.00	57.67	50.92
+ TriangleMix	95.00	34.67	34.33	82.67	83.11	35.67	60.42	60.00	43.67	99.33	72.33	63.74
+ DuoAttention	97.08	32.12	21.45	84.75	81.38	27.17	66.67	63.41	30.66	99.28	63.50	60.67
+ PruLong	96.34	33.80	22.44	81.51	84.87	22.93	64.87	57.87	27.04	100.00	71.03	60.25

Appendix

Table 10: Model performance on RULER tasks. Bold and underline indicate the best and second-best performance within each backbone group, respectively.NS denotes Single-needle in a haystack (NIAH) tasks, NMK stands for Multi-key NIAH, NMV is Multi-value NIAH, and NMQ refers to Multi-query NIAH. QA represents standard Question Answering tasks, and FWE stands for Frequent Words Extraction. Numeric suffixes (e.g., -1, -2, -3) indicate different subsets or difficulty levels within the corresponding task categories.

Method	Sparsity	CWE	VT	Mean
Backbone (FA)	0%	39.90	80.60	60.25
FluxAttn	$\approx$ 51% dynamic	43.07	82.60	62.84
DuoAttention	50% static	42.10	82.87	62.49

Appendix

Table 11: Common Words Extraction (CWE) and Variable Tracking (VT), the two RULER tasks omitted from Table 10 , evaluated on Llama-3.1-8B-Instruct under the same six context lengths

\{8\text{K},16\text{K},32\text{K},64\text{K},128\text{K},256\text{K}\}

. Including them preserves the main conclusion: FluxAttn remains ahead of both the dense backbone and the static head-level baseline at approximately matched sparsity. This added evaluation covers Llama-3.1-8B-Instruct only.

Sparsity	XAttention (XA)				Triangle Attention (TA)
Sparsity	Doc QA	Multi-Hop	Summ.	Code	Doc QA	Multi-Hop	Summ.	Code
0.0	100.00	100.00	100.00	100.00	100.00	100.00	100.00	100.00
0.2	102.78	96.26	99.97	98.54	101.76	99.10	94.86	95.63
0.4	101.70	92.01	101.14	102.91	100.16	95.26	92.84	97.13
0.6	96.87	92.37	101.04	104.54	97.98	93.21	92.32	97.01
0.7	95.44	90.16	95.85	104.02	93.21	81.86	90.57	101.11

Appendix

Table 12: Sensitivity of the Figure 1 (a) observation to the sparse operator, on Llama-3.1-8B-Instruct. Values are performance normalized to the dense configuration (100.00) and are averaged over the LongBench sub-tasks grouped below. Both XAttention (XA) and Triangle Attention (TA) reproduce the qualitative trend of Figure 1 (a)—retrieval-heavy QA degrades sharply at high sparsity while summarization and code remain tolerant—but the onset and magnitude of the degradation are backend-dependent, which is why the router conditions its allocation on the chosen sparse branch instead of assuming a universal threshold.

Task	Original	Seed 456	Mean $\pm$ Std
Qasper	45.25	45.38	45.32 $\pm$ 0.07
MF-en	54.42	54.42	54.42 $\pm$ 0.00
HotQA	54.54	54.54	54.54 $\pm$ 0.00
2Wiki	41.34	41.34	41.34 $\pm$ 0.00
Gov.	34.54	34.71	34.63 $\pm$ 0.09
M.News	26.16	26.22	26.19 $\pm$ 0.03

Appendix

Table 13: LongBench-E per-task results for two independently trained FA-SSA routers on Llama-3.1-8B-Instruct. Router-training stochasticity shifts the task averages by at most 0.5 points, and the aggregate difference between the two runs is 0.08 points.

Batch size	FluxAttn (tok/s)	FA2 (tok/s)	FluxAttn / FA2
8	194	147	1.32 $\times$
16	196	95	2.07 $\times$
24	280	OOM	—
32	OOM	OOM	—

Appendix

Table 14: Decode throughput under batched serving on Llama-3.1-8B-Instruct at a 32K context, measured on a node of 8

\times

NVIDIA H20 GPUs. Throughput is aggregate tokens/s across the served batch. OOM means the configuration exhausts GPU memory on this hardware; we report it rather than extrapolating throughput. The relative advantage does not disappear once heterogeneous requests are co-scheduled, and FluxAttn reaches a larger feasible batch before running out of memory.

Metric	8K	16K	32K	64K	128K	256K
Acc. Dense / full	92.88	92.83	89.46	70.79	80.12	72.34
Acc. PruLong / head	86.96	76.55	70.65	54.52	48.18	30.00
Acc. FluxAttn / layer	90.11	79.39	79.22	56.08	62.94	59.39
$\Omega_{\mathrm{MSR}}$ (PruLong / FluxAttn)	.50 / .53	.50 / .50	.50 / .53	.50 / .53	.50 / .50	.50 / .50
Prefill speedup, PruLong	1.06 $\times$	1.14 $\times$	1.25 $\times$	1.42 $\times$	1.60 $\times$	1.76 $\times$
Prefill speedup, FluxAttn	1.22 $\times$	1.18 $\times$	1.36 $\times$	1.59 $\times$	1.67 $\times$	1.77 $\times$

Appendix

Table 15: Quality–latency–memory trade-off at matched sparsity on Llama-3.1-8B-Instruct. Accuracy is the RULER score per context length; speedups are relative to the dense FA2 baseline at the same context length, with prefill and decode reported separately because they are different phases. FluxAttn and PruLong operate at approximately matched model sparsity (

\Omega_{\mathrm{MSR}}\approx 0.50

), so the comparison isolates routing granularity rather than the amount of computation retained. “KV saving” is the reduction in persistent decode-time KV memory relative to dense. The accuracy entries coincide with the corresponding rows of Table 2 , where the FluxAttn configuration is the sparse-decode one. Raw latencies and peak memory are given in Table 16 , and the systems setup in Table 6 .

Ctx	Method / granularity	Acc.	$\mathbf{\Omega_{\mathrm{MSR}}}$	Prefill ms (speedup)	Decode ms/tok (speedup)	Persistent KV GiB (saving)	Peak GiB
8K	Dense / full	92.88	0	985.0 (1.00 $\times$ )	25.17 (1.00 $\times$ )	0.89 (0.0%)	16.95
	PruLong / head	86.96	0.500	926.2 (1.06 $\times$ )	21.88 (1.15 $\times$ )	0.89 (0.0%)	17.03
	FluxAtt / layer	90.11	0.531	807.4 (1.22 $\times$ )	21.15 (1.19 $\times$ )	0.75 (16.1%)	16.89
16K	Dense / full	92.83	0	2280.1 (1.00 $\times$ )	16.31 (1.00 $\times$ )	1.81 (0.0%)	18.96
	PruLong / head	76.55	0.500	2004.3 (1.14 $\times$ )	13.26 (1.23 $\times$ )	1.81 (0.0%)	19.09
	FluxAtt / layer	79.39	0.500	1932.3 (1.18 $\times$ )	12.55 (1.30 $\times$ )	1.28 (28.9%)	18.29

Appendix

Table 16: Full quality–latency–memory Pareto comparison at matched granularity controls on Llama-3.1-8B-Instruct; the compact view is Table 15 . All methods are measured under the same precision and profiling procedure (Appendix F.5 ); speedups are relative to the dense FA2 baseline at the same context length. “Persistent KV” is the retained decode-time KV state and “Peak” the peak GPU memory. At 128K and 256K, where

\Omega_{\mathrm{MSR}}

is exactly matched at

0.50

, FluxAtt improves accuracy over PruLong by 14.76 and 29.39 points, respectively, while reducing decode latency by 19.2% and 32.0% relative to PruLong. Prefill and decode columns report different inference phases and must not be conflated: head-level sparsity is competitive on prefill while remaining behind on per-token decode.

Method	Granularity	Sparsity	Performance		Speedup
Method	Granularity	Sparsity	LongBench	RULER	Speedup
Elastic Attention	Head-wise	68%	53.35	72.85	2.6 $\times$
Flux Attention (Ours)	Layer-wise	51%	52.28	76.75	2.5 $\times$

Appendix

Table 17: Performance and efficiency comparison between Elastic Attention and Flux Attention on Llama-3.1-8B-Instruct. Flux Attention operates at a substantially lower sparsity level (51% vs. 68%), i.e., it retains more attention computation, and still achieves a comparable prefill/TTFT speedup (2.5

\times

vs. 2.6

\times

) thanks to its hardware-friendly layer-wise routing. The reported speedups are prefill speedups; the decode-phase comparison is given in Table 15 . Retaining more attention computation allows Flux Attention to deliver superior performance on the rigorous RULER benchmark while maintaining competitive results on LongBench.

Model	Method	Performance
Model	Method	LongBench	RULER
Qwen3-4B	Dense (Full Attention)	48.45	66.00
	Elastic Attention	48.08	61.81
	Flux Attention (Ours)	48.72	66.95
Qwen3-8B	Dense (Full Attention)	52.16	75.74
	Elastic Attention	51.51	71.74

Appendix

Table 18: Comprehensive performance comparison among Dense Attention, Elastic Attention, and our Flux Attention across various foundation models. Flux Attention effectively mitigates the severe performance degradation seen in Elastic Attention on the RULER benchmark while maintaining highly competitive results on LongBench.

Figure 7: Evolution of sparsity levels across training steps under different data distributions. Left: Training on a well-balanced dataset, where the router successfully disentangles tasks into distinct sparsity levels. Right: Training on an unbalanced dataset dominated by context-holistic tasks, leading to homogenized routing.

Figure 8: Impact of pooling window size on downstream performance and routing sparsity (

\Omega_{\mathrm{MSR}}

). We evaluate varying truncation budgets (

L\in\{50,100,200,400,800,\text{Full}\}

), retaining only the sequence boundaries (prefix and suffix). Increasing the pooling size beyond 100 tokens introduces context noise, which disrupts the routing mechanism. Consequently, the router misclassifies task features and assigns excessive sparsity to retrieval-intensive tasks, thereby degrading the overall performance.

Task	Spearman $\rho$	Mean FA
Qasper	$+0.027$	35.2%
MultiFieldQA	$-0.250$	44.2%
HotpotQA	$-0.205$	38.4%
2WikiMQA	$-0.167$	39.6%
GovReport	$-0.293$	57.5%
MultiNews	$-0.314$	57.7%

Appendix

Table 19: Corrected correlation between the task-independent offline entropy score

E_{\ell}

(Appendix D.1 ) and the learned soft FA probability of each layer, on Llama-3.1-8B-Instruct.

E_{\ell}

is computed once from a trace-normalized hidden-state Gram matrix using its top-128 eigenvalues on a profiling split of 26 disjoint task-balanced LongBench-E samples; routing is collected over 20 samples from each of 13 LongBench-E tasks (260 samples total) across 32 layers. Correlations are weak to moderate and task-dependent, so the learned router departs from a static entropy ranking rather than reproducing it; entropy is not used as a routing label or supervision target. The macro-average Pearson correlation is

r=-0.108

, and the macro-average Spearman correlation is stable to the eigenvalue-truncation threshold (

-0.231

-0.234

-0.284

-0.215

for

K=16,32,64,128

Figure 9: Router latency analysis. The router incurs negligible overhead (avg. 0.20 ms). Our design ensures length-invariant stability, maintaining constant speed from 512 to 1M tokens.

Panel of Figure 10 — Figure 10: Decomposition of Training Objectives for Flux Attention. We visualize the training dynamics of the Layer Router, separating the total loss into (a) the primary language modeling objective and (b) the sparsity regularization term. Subfigures (c) and (d) illustrate the task-level differentiation in sparsity allocation ( $\Omega_{\mathrm{MSR}}$ ) and adaptive coefficients ( $\lambda$ ), demonstrating how the model automatically distinguishes between context-holistic and retrieval-intensive tasks.

Panel of Figure 11 — Figure 11: Comparison on a long-context reading comprehension task. Our model accurately extracts and verifies the severity statistics of outdated cooking methods in Africa compared to global figures, while all baselines consistently fall for the same unsupported distractor regarding carbon markets.

Panel of Figure 12 — Figure 12: Qualitative comparison on identifying the core argument in a philosophical legal text. Our model successfully synthesizes the text to identify the underlying argumentative strategy (refutation via analogy), whereas baselines are easily distracted by literal sentences from the title and opening hook.

Panel of Figure 13 — Figure 13: Qualitative comparison on extracting technical methodology from a machine learning paper. Our model accurately identifies the specific bounding box encoding strategy, whereas all baselines suffer from hallucination, confidently generating plausible but incorrect architectural details (Fourier embeddings) not supported by the text.

Explore similar work

You Only Index Once: Cross-Layer Sparse Attention with Shared Routing
Jun 4, 2026Yutao Sun, Yanqi Zhang, Li Dong +2Dynamic Sparse Attention Efficient Long-Context Inference
Microsoft Research · Tsinghua University
SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference
Jun 3, 2026Yaosheng Fu, Guangxuan Xiao, Xin Dong +2Dynamic Sparse Attention LLM Inference Optimization
1NVIDIA · 2Thinking Machines Lab · 3ByteDance Seed +1
An Efficient Hybrid Sparse Attention with CPU-GPU Parallelism for Long-Context Inference
May 8, 2026Feiyu Yao, Zhixiong Niu, Xiaqing Li +3Dynamic Sparse Attention Efficient Long-Context Inference
Beijing University of Technology Beijing, China · Microsoft Research Beijing, China · Beijing Jiaotong University Beijing, China

Jun 4, 2026, cs.CLJ/K move · Enter open · S save

You Only Index Once: Cross-Layer Sparse Attention with Shared Routing

Yutao Sun, Yanqi Zhang, Li Dong, Jianyong Wang+1

Microsoft Research · Tsinghua University

Dynamic Sparse Attention Efficient Long-Context Inference Bottlenecks

Open paper AskPDF arXiv

Long-context inference in modern LLMs is increasingly constrained by decoding efficiency, especially in reasoning-heavy settings where models generate long intermediate chains of thought. Existing sparse attention methods often face a practical efficiency-quality trade-off. Structured block sparse methods typically provide stronger acceleration but incur noticeable quality loss, while token sparse methods are usually more accurate yet deliver limited end-to-end speedup because top-k routing over the full cache remains expensive. In this work, we propose cross-layer sparse attention (CLSA), which is built on top of KV-sharing architectures such as YOCO. The core idea is to share not only the KV cache across cross-decoder layers, but also the routing index. A single indexer computes token-level top-k selection once and reuses the resulting index across layers, thereby preserving the fine-grained selectivity of token sparse attention while amortizing the routing overhead. The resulting architecture improves all major inference bottlenecks jointly, including pre-filling, KV-cache storage, and long-context decoding. Experiments across short-context and long-context benchmarks show that CLSA is both accurate and efficient, achieving up to 7.6x decoding speedup and 17.1x overall throughput improvement at 128K context. These results suggest a more complete architectural solution for long-context LLMs that jointly advances model quality and inference efficiency.

Abstract

Explore similar work

You Only Index Once: Cross-Layer Sparse Attention with Shared Routing

SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference

An Efficient Hybrid Sparse Attention with CPU-GPU Parallelism for Long-Context Inference