llama.cpp Release b11712 Fixes Sliding Window Attention Test Failures
Release b11712 of llama.cpp introduces a per-tensor filler table to the random model generator, resolving intermittent test failures in models utilizing sliding window attention. This update ensures that model weights for specific architectures like Muse Spark and Mamba are correctly generated without causing unexpected numerical amplification. This fix improves the reliability of the test suite across different architectures, ensuring that continuous integration checks run smoothly on both CPU and CUDA builds. Developers relying on random model generators for validation will see more stable and accurate test results. The update consolidates tensor generation logic by drawing raw weights once per block and mapping them through a lookup table keyed on tensor names. Over a thousand test seeds, this reduced the worst-case Normalized Mean Squared Error (NMSE) from 1.36e-05 down to 2.48e-06.
## BACKGROUND
Sliding window attention is an optimization technique in large language models where each token can only attend to a fixed local range of preceding tokens, reducing memory overhead during long-context processing. llama.cpp is a popular framework for running large language models locally on various hardware accelerators, relying on rigorous testing suites to verify model conversion and execution correctness.