Search papers, labs, and topics across Lattice.
This paper investigates the effectiveness of RoPE-aligned Q/K rotations in the context of dynamic 4-bit quantization for attention heads in neural networks. The authors derive a rotation angle that minimizes channel variance under a pooled-covariance approach, but find that this method does not enhance accuracy compared to traditional full-head mixing, leading to increased perplexity across multiple checkpoints. Ultimately, the study reveals that optimality in a structured surrogate does not guarantee reduced quantization error when there is a misalignment with the quantizer's scale-setting statistic.
Despite promising theoretical foundations, RoPE-aligned rotations fail to improve quantization accuracy, highlighting a critical misalignment in current methods.
Rotation-based post-training quantisation commonly applies an orthogonal transform across an entire attention head to reduce outlier-induced error. RoPE instead partitions each head into two-dimensional frequency pairs, raising the question of whether a transform respecting this decomposition can improve on full-head mixing. Prior work has established the per-pair rotations that commute with RoPE. We state the converse result that, for distinct frequencies, no other single-head orthogonal map commutes with RoPE. For the head-shared parameterisation used in our experiments, we then derive the rotation angle that minimises the larger channel variance under a pooled-covariance, position-averaged surrogate and verify that the implementation attains its analytic minimum. The evaluated head-shared pairwise configuration does not improve accuracy in the tested dynamic W4A4KV4 setting. Across four checkpoints, replacing the full-head Hadamard with this configuration increases perplexity at both short and long context lengths. Composing the pairwise rotation with the Hadamard satisfies the selected $\pm0.05$-PPL interval criterion under the default estimator. Estimating the shared angle from K alone improves pairwise-only on every checkpoint but does not close its gap to full-head mixing. The analytic objective controls a position-averaged second moment of a pooled calibration covariance, whereas the dynamic quantiser sets its step from a tokenwise group range. The pairwise transform also has only two-channel mixing support. Along a controlled interpolation from two-channel to full-head mixing, K range, relative quantisation error, and perplexity degradation decrease as support increases. These results show that optimality for a structured surrogate need not reduce quantisation error when the surrogate and mixing support are misaligned with the quantiser's scale-setting statistic.
