Quantization

GLM is a bilingual (English and Chinese) pre-trained transformer-based language model that follow the traditional architecture of decoder-only autoregressive language modeling. It leverages autoregressive blank infilling as its training objective.

Accordion is a gradient communication scheduling algorithm that is generic across models while imposing low computational overheads. Accordion inspects the change in the gradient norms to detect critical regimes and adjusts the communication schedule dynamically. Accordion works for both adjusting the gradient compression rate or the batch size without additional parameter tuning.

Traditional methods are based on block-wise regression. This framework, Enhanced Blockwise Classification (EBC), however, is based on the idea that aims to classify the count value within each block into several pre-defined bins. The enhancement comes from 3 aspects: discretization policy, label correction and loss function. Notice that the original block-wise classification concept was introduced by Liu et al. in Counting Objects by Blockwise Classification.

I-BERT is a quantized version of BERT that quantizes the entire inference with integer-only arithmetic. Based on lightweight integer only approximation methods for nonlinear operations, e.g., GELU, Softmax, and Layer Normalization, it performs an end-to-end integer-only BERT inference without any floating point calculation. In particular, GELU and Softmax are approximated with lightweight second-order polynomials, which can be evaluated with integer-only arithmetic. For LayerNorm, integer-only computation is performed by leveraging a known algorithm for integer calculation of square root.

Bi3D is a stereo depth estimation framework that estimates depth via a series of binary classifications. Rather than testing if objects are at a particular depth D, as existing stereo methods do, it classifies them as being closer or farther than D. It takes the stereo pair and a disparity $d\_{i}$ and produces a confidence map, which can be thresholded to yield the binary segmentation. To estimate depth on $N + 1$ quantization levels we run this network $N$ times and maximize the probability in Equation 8 (see paper). To estimate continuous depth, whether full or selective, we run the SegNet block of Bi3DNet for each disparity level and work directly on the confidence volume.

Neural Network Compression Framework, or NNCF, is a Python-based framework for neural network compression with fine-tuning. It leverages recent advances of various network compression methods and implements some of them, namely quantization, sparsity, filter pruning and binarization. These methods allow producing more hardware-friendly models that can be efficiently run on general-purpose hardware computation units (CPU, GPU) or specialized deep learning accelerators.

Ternary Weight Splitting is a ternarization approach used in BinaryBERT that exploits the flatness of ternary loss landscape as the optimization proxy of the binary model. We first train the half-sized ternary BERT to convergence, and then split both the latent full-precision weight $\mathbf{w}^{t}$ and quantized $\hat{\mathbf{w}}^{t}$ to their binary counterparts $\mathbf{w}\_{1}^{b}, \mathbf{w}\_{2}^{b}$ and $\hat{\mathbf{w}}\_{1}^{b}, \hat{\mathbf{w}}\_{2}^{b}$ via the TWS operator. To inherit the performance of the ternary model after splitting, the TWS operator requires the splitting equivalency (i.e., the same output given the same input): $$ \mathbf{w}^{t}=\mathbf{w}\_{1}^{b}+\mathbf{w}\_{2}^{b}, \quad \hat{\mathbf{w}}^{t}=\hat{\mathbf{w}}\_{1}^{b}+\hat{\mathbf{w}}\_{2}^{b} $$ While solution to the above equation is not unique, we constrain the latent full-precision weights after splitting $\mathbf{w}\_{1}^{b}, \mathbf{w}\_{2}^{b}$ to satisfy $\mathbf{w}^{t}=\mathbf{w}\_{1}^{b}+\mathbf{w}\_{2}^{b}$. See the paper for more details.

TernaryBERT is a Transformer-based model which ternarizes the weights of a pretrained BERT model to $\{-1,0,+1\}$, with different granularities for word embedding and weights in the Transformer layer. Instead of directly using knowledge distillation to compress a model, it is used to improve the performance of ternarized student model with the same size as the teacher model. In this way, we transfer the knowledge from the highly-accurate teacher model to the ternarized student model with smaller capacity.

Many communication-efficient variants of SGD use gradient quantization schemes. These schemes are often heuristic and fixed over the course of training. We empirically observe that the statistics of gradients of deep models change during the training. Motivated by this observation, we introduce two adaptive quantization schemes, ALQ and AMQ. In both schemes, processors update their compression schemes in parallel by efficiently computing sufficient statistics of a parametric distribution. We improve the validation accuracy by almost 2% on CIFAR-10 and 1% on ImageNet in challenging low-cost communication setups. Our adaptive methods are also significantly more robust to the choice of hyperparameters.

Aggregated Learning (AgrLearn) is a vector-quantization approach to learning neural network classifiers. It builds on an equivalence between IB learning and IB quantization and exploits the power of vector quantization, which is well known in information theory.

BAGUA is a communication framework whose design goal is to provide a system abstraction that is both flexible and modular to support state-of-the-art system relaxation techniques of distributed training. The abstraction goes beyond parameter server and Allreduce paradigms, and provides a collection of MPI-style collective operations to facilitate communications with different precision and centralization strategies.

BinaryBERT is a BERT-variant that applies quantization in the form of weight binarization. Specifically, ternary weight splitting is proposed which initializes BinaryBERT by equivalently splitting from a half-sized ternary network. To obtain BinaryBERT, we first train a half-sized ternary BERT model, and then apply a ternary weight splitting operator to obtain the latent full-precision and quantized weights as the initialization of the full-sized BinaryBERT. We then fine-tune BinaryBERT for further refinement.

MixLoRA is a type of PEFT method which construct a resource-efficient sparse MoE model based on LoRA.

As the size and complexity of models and datasets grow, so does the need for communication-efficient variants of stochastic gradient descent that can be deployed to perform parallel model training. One popular communication-compression method for data-parallel SGD is QSGD (Alistarh et al., 2017), which quantizes and encodes gradients to reduce communication costs. The baseline variant of QSGD provides strong theoretical guarantees, however, for practical purposes, the authors proposed a heuristic variant which we call QSGDinf, which demonstrated impressive empirical gains for distributed training of large neural networks. In this paper, we build on this work to propose a new gradient quantization scheme, and show that it has both stronger theoretical guarantees than QSGD, and matches and exceeds the empirical performance of the QSGDinf heuristic and of other compression methods.