Abstract
Large Language Model training is increasingly constrained by memory capacity. Activation Checkpointing reduces activation memory by avoiding storing selected activations during the forward pass and rematerializing them during backpropagation, thereby trading memory savings for additional computation. Although this technique is essential for scaling model training, rematerialization can be prohibitively expensive, particularly for Transformer linear layers. Approximate computation could reduce this overhead, but conventional approaches either introduce errors that accumulate over the backward pass or fail to exploit information available from the initial forward pass. We introduce a novel approximate rematerialization method that can yield unbiased gradient estimates: using information collected during the initial forward pass, the method selectively recomputes only the most important layer components, either through greedy selection or randomized sketching. Ablation studies show that the resulting gradients closely match the exact gradients. When applied to feed-forward layers, attention layers, or both, our method can recompute only two-thirds of the layer components in a Llama-488M model, while increasing perplexity by just 1.7%, 1.2%, and 2.5%, respectively.
More in this topicMemory-efficient training reduces the memory needed to optimize neural networks by compressing, recomputing, or relocating training state.
See all 100 papers → May 2, 2026
Training large language models (LLMs) is highly memory-intensive, as training must store not only weights and optimizer states but also intermediate activations for backpropagation. While existing memory-efficient methods largely focus on gradients and optimizer states, activation compression is less well established due to the lack of LLM-tailored theory and guarantees. In this work, we develop a theoretical framework showing that activation compression is safe for linear operators when activation compression is unbiased, but problematic for nonlinear ones. We further derive gradient variance bound and establish convergence guarantees for applying activation compression to all linear operators under the standard
L-smoothness assumption, showing that it does not change the convergence rate. Guided by the theory, we propose an activation-gradient co-compression method that reuses low-rank activation factors to compress linear-layer gradients without extra computation or additional gradient error. We conduct extensive experiments on Qwen and LLaMA models using a pretraining benchmark and multiple fine-tuning benchmarks to validate our theory and demonstrate competitive performance of our method in both accuracy and compression efficiency. We provide our code in the supplementary material for reproducibility.
Wen-Da Wei, Han-Bin Fang, Yang-Di Liu +3
Nanjing University, Nanjing, China · Tsinghua University, Beijing, China · Huazhong University of Science and Technology, Wuhan, China +1
Oct 15, 2026
Reducing memory use in large language model (LLM) training is an ongoing challenge. Activation checkpointing saves memory by recomputation, while lossy compression changes the stored activations. The reemergence of sparse
\relu and
\relu2 activations creates a third opportunity: losslessly compressing the activation cache itself. We present BitSparse, a GPU-resident lossless compression format for sparse activations. Compression and reconstruction run in custom Triton kernels fused into an autograd function, shrinking the activation cache while recovering it exactly for the backward pass. Bfloat16 (BF16) and FP8 activations are supported. Optional sign-bit packing further increases compression up to close to the theoretical limit. We analyze its compression ratio and compare it to the theoretical optimum. BitSparse lowers per-token VRAM by 31% and 30% Modded-NanoGPT and Nemotron-H respectively, with overheads of no more than 1% at practical sequence lengths. Additionally, a new preallocated buffer system avoids host synchronization during training when variable size tensors are allocated.
Max Zhu, Bohang Sun, Pietro Lio +1
University of Cambridge · University of Oxford
Feb 26, 2026
Activations have become the primary memory bottleneck in large-batch LLM training. However, existing compression methods fail to exploit the spectral structure of activations, resulting in slow convergence or limited compression. To address this, we bridge the relationship between the algorithm's fast convergence and the requirements for subspace projection, and show that an effective compression should yield an unbiased estimate of the original activation with low variance. We propose Principal-Random Subspace for LLM Activation Compression (PRAC), which novelly decomposes activations into two components: a principal subspace captured via SVD to retain dominant information, and a random subspace sampled from the orthogonal complement to approximate the tail. By introducing a precise scaling factor, we prove that PRAC yields an unbiased gradient estimator with minimum variance under certain conditions. Extensive experiments on pre-training and fine-tuning tasks demonstrate that PRAC achieves up to 36% total memory reduction with negligible performance degradation and minimal computational cost.
Yanyi Li, Yimu Zhang, Cong Fang
School of Intelligence Science and Technology, Peking University, China · Institute for Artificial Intelligence, Peking University, China