Post Training Quantization With Dynamic Range Calibration

via “quantization with multiple precision formats and calibration strategies”

🤗 Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.

Unique: Implements a modular quantization system (src/transformers/quantization_config.py) that abstracts away backend-specific quantization details (bitsandbytes, GPTQ, AWQ) behind a unified QuantizationConfig interface, enabling seamless switching between quantization strategies

vs others: More accessible than standalone quantization libraries because it integrates quantization into model loading via config parameters, automatically handling weight conversion and calibration without requiring separate quantization pipelines

via “post-training quantization with dynamic range calibration”

Lightweight ML inference for mobile and edge devices.

Unique: Dynamic range calibration automatically profiles activation distributions across layers using representative data, computing per-layer or per-channel quantization scales that adapt to actual model behavior rather than using fixed ranges. Supports both symmetric (zero-point = 0) and asymmetric quantization with automatic selection per layer based on activation histogram analysis.

vs others: More automated than manual quantization-aware training (QAT) since it requires no retraining, and more accurate than simple min-max scaling because it uses distribution-aware calibration. Faster than QAT (minutes vs. hours) but typically yields 1-3% lower accuracy than QAT on complex models.

via “quantization with fp8 and low-precision inference”

High-throughput LLM serving engine — PagedAttention, continuous batching, OpenAI-compatible API.

Unique: Implements fused quantization kernels that perform dequantization and matrix multiplication in a single GPU operation, reducing memory bandwidth overhead vs separate dequant+compute steps

vs others: Achieves 4-8x memory reduction with 1-3% accuracy loss vs no quantization, outperforming naive INT8 quantization by using per-token scaling and mixed-precision strategies

via “quantization with accuracy preservation and layer-wise precision control”

Qualcomm's platform for optimizing AI models on Snapdragon edge devices.

Unique: Supports layer-wise precision control where sensitive layers (e.g., output layers) can remain in higher precision while others use INT8, optimizing the accuracy-latency tradeoff per layer rather than uniformly quantizing the entire model

vs others: More flexible than TensorFlow Lite's uniform INT8 quantization because it allows mixed-precision per layer, and more practical than quantization-aware training because it works on pre-trained models without retraining

via “calibration-driven per-channel scaling factor computation”

4-bit weight quantization for LLMs on consumer GPUs.

Unique: Computes scaling factors by analyzing actual activation patterns from calibration data rather than using weight statistics alone. This activation-aware approach identifies which weight channels are most important based on how often they are activated during inference, enabling selective protection of critical channels.

vs others: More accurate than weight-only quantization methods (GPTQ) because it accounts for activation patterns; more efficient than layer-wise quantization because per-channel factors provide finer-grained control without excessive overhead.

via “one-shot post-training quantization with calibration-free execution”

Toolkit for LLM quantization, pruning, and distillation.

Unique: Uses a modifier-based architecture where quantization logic is injected as PyTorch hooks into the model graph, enabling algorithm-agnostic calibration and composition of multiple compression techniques (quantization + pruning + distillation) in a single pipeline without model rewriting

vs others: Faster than AutoGPTQ or GPTQ-for-LLaMA because it abstracts algorithm selection and calibration into reusable modifiers, allowing parallel experimentation; more flexible than ONNX Runtime quantization because it preserves PyTorch semantics and integrates directly with vLLM

via “calibration-based quantization with sample-driven scale computation”

GPTQ-based LLM quantization with fast CUDA inference.

Unique: Implements Hessian-based scale computation from the GPTQ paper, using calibration samples to compute optimal per-group quantization scales that minimize reconstruction error. Supports configurable calibration dataset size and custom sample selection, enabling domain-specific quantization without retraining.

vs others: More accurate than static quantization (e.g., min-max scaling) because it uses Hessian information to weight important weights higher, and faster than QAT (quantization-aware training) because it requires only forward passes without backpropagation.

via “quantization-aware training (qat) with post-training quantization”

PyTorch-native LLM fine-tuning library.

Unique: Integrates PyTorch's native quantization APIs (torch.quantization) with torchtune recipes, allowing users to apply QAT via a single config flag (quantization_enabled: true) without modifying training code. For PTQ, torchtune provides a separate recipe that loads a pre-trained model, applies quantization with calibration data, and exports quantized weights.

vs others: More integrated than using PyTorch quantization directly because torchtune handles distributed training with quantization, checkpoint management, and metric logging, whereas raw PyTorch quantization requires manual integration with training loops.

via “model quantization for memory and latency reduction”

text-generation model by undefined. 1,60,37,172 downloads.

Unique: Supports both post-training quantization (no retraining) via bitsandbytes and quantization-aware training (better accuracy) via torch.quantization, with automatic calibration dataset selection for minimal accuracy loss

vs others: Faster and simpler than knowledge distillation (which requires training a smaller model), but less accurate than distillation for extreme compression — best for 2-4x size reduction, not 10x+

via “quantization with multiple precision formats and framework support”

Hugging Face's model library — thousands of pretrained transformers for NLP, vision, audio.

Unique: Integrates multiple quantization backends (bitsandbytes, GPTQ, AWQ) under a unified API where quantization method is specified via config object, enabling transparent switching between quantization schemes. Quantization is applied during model loading via load_in_8bit/load_in_4bit flags, avoiding explicit conversion code.

vs others: More convenient than manual quantization with bitsandbytes because quantization is applied automatically during model loading. More flexible than ONNX quantization because it supports multiple quantization methods and frameworks.

via “model quantization and compression for edge deployment”

fill-mask model by undefined. 5,92,18,905 downloads.

Unique: Post-training quantization via ONNX Runtime or PyTorch quantization APIs requires no retraining while achieving 4x model size reduction; supports multiple quantization schemes (symmetric, asymmetric, per-channel) for fine-grained accuracy-efficiency control

vs others: Simpler than quantization-aware training (no retraining required) and more portable than framework-specific quantization due to ONNX support

via “low-precision quantization with per-layer calibration and mixed-precision support”

OpenVINO™ is an open source toolkit for optimizing and deploying AI inference

Unique: Implements per-layer calibration with mixed-precision support, allowing different layers to use different precisions based on sensitivity analysis. The quantization pipeline is decoupled from the training process (post-training quantization only), making it applicable to any pre-trained model without retraining.

vs others: Provides more granular mixed-precision control than TensorFlow Lite's uniform quantization and supports INT8 quantization on a wider range of hardware than PyTorch's native quantization tools.

via “model quantization and compression for edge deployment”

automatic-speech-recognition model by undefined. 15,29,218 downloads.

Unique: Implements both post-training quantization (PTQ) for quick deployment and quantization-aware training (QAT) for minimal accuracy loss. Provides hardware-specific optimization paths (ONNX Runtime, TensorRT, CoreML) enabling deployment across diverse edge devices with automatic kernel selection for maximum performance.

vs others: Reduces model size by 50-75% compared to full precision with minimal accuracy loss (int8: <2% WER increase), enabling mobile deployment where cloud APIs are infeasible. More efficient than knowledge distillation for quick deployment, though distillation may achieve better accuracy-efficiency tradeoffs with additional training.

via “model quantization for edge deployment”

image-segmentation model by undefined. 1,55,904 downloads.

Unique: Supports standard PyTorch post-training quantization without model-specific modifications, enabling straightforward int8 deployment — though deformable attention operations may not quantize cleanly

vs others: Reduces model size 4x (500MB to 125MB) with minimal accuracy loss vs float32, enabling edge deployment, though 1-2% accuracy degradation and limited hardware support add deployment complexity

via “model quantization to int8 with minimal accuracy loss”

question-answering model by undefined. 56,200 downloads.

Unique: ONNX Runtime quantization uses symmetric int8 ranges with per-channel calibration, preserving accuracy better than asymmetric quantization; most mobile frameworks use simpler per-tensor quantization with 2-5% accuracy loss

vs others: 2-4x faster CPU inference and 75% smaller model size vs float32, with <3% accuracy loss on SQuAD (vs 5-10% for naive quantization)

via “quantization with post-training and dynamic quantization support”

Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.

Unique: Integrates multiple quantization backends (bitsandbytes, PyTorch native, GPTQ, AWQ) behind a unified QuantizationConfig interface, with automatic backend selection based on model type and hardware. Unlike standalone quantization libraries, Transformers' quantization is transparent to the user: quantized models are loaded identically to full-precision models, and inference code requires no changes.

vs others: More integrated than separate quantization libraries (bitsandbytes, GPTQ) because it handles model loading and inference automatically, and supports more quantization strategies (INT8, INT4, FP8, GPTQ, AWQ) in a single framework. However, less optimized than specialized quantization tools (e.g., TensorRT, ONNX Runtime) for production inference because it prioritizes ease of use over performance.

via “gptq quantization with calibration and per-layer configuration”

Optimum Library is an extension of the Hugging Face Transformers library, providing a framework to integrate third-party libraries from Hardware Partners and interface with their specific functionality.

Unique: Integrates Hugging Face datasets library for automatic calibration data loading and supports custom calibration datasets through flexible dataset interface. Per-layer quantization configuration allows fine-grained control over precision-accuracy tradeoffs, and quantization configs are serializable for reproducibility and transfer across model versions.

vs others: Provides integrated calibration dataset management and per-layer configuration control, whereas alternatives like bitsandbytes require manual calibration data handling and apply uniform quantization across all layers.

via “quantization with post-training and qat support via pt2e framework”

Tensors and Dynamic neural networks in Python with strong GPU acceleration

Unique: Integrates quantization with torch.export to generate portable quantized graphs, supporting both post-training quantization for quick optimization and QAT for accuracy recovery. PT2E framework enables backend-specific quantization strategies.

vs others: More flexible than TensorRT quantization because it supports arbitrary PyTorch models and multiple quantization schemes, while more accurate than simple INT8 conversion because it includes calibration and QAT support.

via “double quantization of quantization constants for nested compression”

* ⭐ 05/2023: [Voyager: An Open-Ended Embodied Agent with Large Language Models (Voyager)](https://arxiv.org/abs/2305.16291)

Unique: Introduces nested quantization where quantization constants themselves are quantized to 8-bit precision with separate scales, reducing constant overhead by 2-4x — prior quantization work treated constants as full-precision metadata, not subject to further compression

vs others: Reduces total model size by an additional 2-4% compared to single-level quantization, enabling 70B models to fit in 24GB memory where standard 4-bit quantization alone would require 28-32GB