Untangling Component Imbalance in Hybrid Linear Attention Conversion Methods
This addresses a critical flaw in post-training linearisation methods for Transformers, which is important for researchers and practitioners aiming to scale models efficiently, though it is incremental as it builds on existing hybrid approaches.
The paper tackled the problem of existing hybrid linear attention conversion methods inadvertently bypassing the linear component by relying too heavily on sliding-window softmax, and proposed three solutions (inference-time hybridisation, HedgeCATs, and SSD) that maintain computational efficiency while recovering most base model performance and ensuring genuine linear attention adoption.
Transformers' quadratic computational complexity limits their scalability despite remarkable performance. While linear attention reduces this to linear complexity, pre-training such models from scratch remains, in most cases, prohibitively expensive. Recent post-training linearisation methods convert pre-trained Transformers to linear models efficiently, often using hybrid approaches that combine linear attention with sliding-window softmax. We identify a critical flaw: existing hybrid methods inadvertently bypass the linear component, relying almost entirely on SWA. Component-level diagnostics reveal this previously undetected behaviour stems from overlooked evaluation practices on common-sense benchmarks. We propose three solutions to ensure balanced component usage: (i) inference-time hybridisation of linear-only conversions with sliding-window softmax; (ii) HedgeCATs, combining attention-weight transfer with targeted LoRA fine-tuning; and (iii) Scheduled Sliding-window Dropout (SSD), which stochastically suppresses the softmax branch during training to prevent component collapse. Our methods maintain computational efficiency while recovering most base model performance and ensuring genuine linear attention adoption, restoring the validity of performance attributions in hybrid conversions.