Depth Registers Unlock W4A4 on SwiGLU: A Reader/Generator Decomposition
We study post-training W4A4 quantization in a controlled 300M-parameter SwiGLU decoder-only language model trained on 5B tokens of FineWeb-Edu, and ask which input-activation sites dominate the error. Naive round-to-nearest W4A4 collapses validation perplexity from FP16 23.6 to 1727. A simple residual-axis training-time intervention -- Depth Registers with a register-magnitude hinge loss (DR+sink) -- reduces this to 119 (about 14x) at matched FP16 PPL and matched zero-shot capacity, and composes with SmoothQuant to 39.9 PPL. The residual ~2 PPL gap to FP16 is the diagnostic core. We decompose W4A4 damage by input-activation site: the five trainable linears in a SwiGLU block split into residual-axis readers (qkv, w1, w3) and block-internal generators (o_proj, w2). Elementary norm arguments show residual-axis magnitude control bounds readers tightly but leaves w2's bilinear input bounded only by the trivial product of factor bounds; empirically, DR+sink collapses reader kurtosis while leaving generators essentially unchanged, and the reader-rescued W4A4 residue is flat at ~0.28 nats across three matched checkpoints with Delta-remove(w2) dominating. We present DR+sink as a training-time probe rather than a deployment proposal: a post-hoc alternative (Per-Linear QuaRot) nearly matches it on the reader axis. Full QuaRot -- adding online per-head value Hadamard plus online w2-input rotation -- does not close the gap either, directly testing the prediction that orthogonal rotation cannot bound the bilinear SwiGLU tail. Claims are specific to our 300M, 5B-token, single-seed setting, and our experiments do not isolate the partition from the hinge.
Code (0)
등록된 구현이 없습니다.
Similar Papers 제목 키워드 기반
Psycho-linguistic Experiment on Universal Semantic Components of Verbal Humor: System Description and Annotation
Objective criteria for universal semantic components that distinguish a humorous utterance from a non-humorous one are presently under debate. In this article, we give an in-depth observation of our system of self-paced …
Confidence-Adaptive SwiGLU for Mixture-of-Experts
SwiGLU has become a standard gated activation in modern Transformer MLPs, yet its gate sharpness -- the smoothness and selectivity of the gating function -- is typically fixed throughout training. In this work, we propos…
Latent Reward Registers for Diffusion Preference Alignment
Aligning diffusion models with human preferences usually relies on a sparse terminal reward evaluated on the final generated samples, which creates a severe temporal credit-assignment problem across the denoising process…
Reinforcement LearningAbstractive Text Summarization by Incorporating Reader Comments
In neural abstractive summarization field, conventional sequence-to-sequence based models often suffer from summarizing the wrong aspect of the document with respect to the main aspect. To tackle this problem, we propose…
Abstractive Text SummarizationReader-Aware SummarizationText SummarizationTile-Level Activation Overlap for Efficient LLM Inference
SwiGLU is the dominant MLP activation in modern large language models, yet its intermediate tensor materialization costs 9-37% of MLP execution time. We present two complementary CUTLASS-based SM90 kernels that fuse SwiG…