Local Inference of LLMs on Long Contexts: A Data Systems Perspective

Transformer-based large language models face severe scalability challenges in long-context generation due to the computational and memory costs of full-context attention. Under practical computation and memory constraints, many inference-efficient long-context methods improve efficiency by adopting bounded-context or segment-level execution only during inference, while continuing to train models under full-context attention, resulting in a mismatch between training and inference execution and state-transition semantics. Based on this insight, we propose a training-inference consistent segment-level generation framework, in which training and inference follow the same segment-level forward execution semantics. During training, consistency with inference is enforced by restricting gradient propagation to KV states carried over from the immediately preceding segment, while permitting head-specific access to past KV states during the forward pass without involving them in gradient propagation. Across long-context benchmarks, our approach achieves performance comparable to full-context attention, while achieving competitive latency-memory trade-offs against strong inference-efficient baselines, and substantially improving scalability at very long context lengths (e.g., approximately 6x lower peak prefill memory at 128K compared to full-context attention with FlashAttention).

Language models struggle to generalize beyond pretraining context lengths, limiting long-horizon reasoning and retrieval. Continued pretraining on long-context data can help but is expensive due to the quadratic scaling of Attention. We observe that most tokens do not require (Global) Attention over the entire sequence and can rely on local context. Based on this, we propose L2A (Learning To Attend), a layer that enables conditional (token-wise) long-range memory access by deciding when to invoke global attention. We evaluate L2A on Qwen 2.5 and Qwen 3 models, extending their effective context length from 32K to 128K tokens. L2A matches the performance of standard long-context training to within 3% while skipping Global Attention for $\sim$80% of tokens, outperforming prior baselines. We also design custom Triton kernels to efficiently implement this token-wise conditional Attention on GPUs, achieving up to $\sim$2$\times$ improvements in training throughput and time-to-first-token over FlashAttention. Moreover, L2A enables post-training pruning of highly sparse Global Attention layers, reducing KV cache memory by up to 50% with negligible performance loss. Our code is released under Apache 2.0 at https://github.com/awslabs/hybrid-model-factory/tree/main/examples/research/L2A.

Full self-attention in large language models scales as O(N^2), limiting long-context analysis to 65k tokens and requiring costly GPU clusters. The Reduced Interaction Sampling (RIS) inference engine addresses this as a model-agnostic architecture. Without modifying weights, RIS introduces sparse stochastic attention geometries that operate under extreme pruning (up to 99% under 1% density) while preserving factual retrieval. While foundational graph-theoretic RIS formulations establish an asymptotic O(N log N) bound, this work evaluates a practical fixed-density regime on unaccelerated commodity CPU systems. We validate RIS on Qwen2-1.5B-Instruct across two regimes. At 32,768 tokens (where native dense attention serves as the upper bound), RIS-Stochastic at 1% density and 70 ensemble seeds achieves 75.00% accuracy, outperforming the native dense baseline (71.88%), while 5% density at 10 seeds matches it (71.88%), demonstrating that multi-seed sparsity filters sequence-level noise. Crucially, a controlled ablation evaluating a single static mask at matched budget (N=1, d=0.50) yields only 53.12% accuracy, proving that multi-seed topological diversity, rather than raw edge density, drives contextual recovery. Under tight budgets, RIS-Structural reaches 65.62% accuracy at 1% density with 10 seeds (recovering 50.0% of the gap over the 59.38% zero-context floor) and 68.75% (75.0% recovery) with 30 seeds. At 65,536 tokens, where dense attention triggers out-of-memory faults, RIS yields retrieval gains of up to 14.06 percentage points over the zero-context floor (51.56%), confirmed as marginally significant under McNemar's paired test (p = 0.078 < 0.10). All evaluations run on commodity CPU servers (16-128 GB RAM), demonstrating that long-context LLM inference is feasible on standard academic hardware without GPU acceleration.