SOTA low-bit LLM quantization (INT8/FP8/MXFP8/INT4/MXFP4/NVFP4) & sparsity; leading model compression techniques on PyTorch, TensorFlow, and ONNX Runtime

Low Precision Arithmetic Simulation in PyTorch

Stable and Efficient Reinforcement Learning for Trillion-Parameter LLMs

A script to convert floating-point CNN models into generalized low-precision ShiftCNN representation

Low Precision(quantized) Yolov5

CUDA/HIP header-only library for low-precision (16 bit, 8 bit) and vectorized GPU kernel development

JAX Scalify: end-to-end scaled arithmetics

Code for DNN feature map compression paper

Low Precision Floating Point Types in Fortran

Library of specialized algebra and performance primitives

NVFP4-DiT: Theory and Practice of 4-Bit Low-Precision Training and Inference for Audio-Guided Video Diffusion Transformers

Implemented post-training quantisation (PTQ) on transformer-based reasoning models using 8-bit and 4-bit weight quantisation (INT8, INT4) with frameworks like PyTorch and Hugging Face Transformers. Leveraged libraries such as bitsandbytes to reduce model size and accelerate inference, while evaluating performance degradation on reasoning tasks. Com

Block-scaled FP8 / FP4 / INT4 tensor primitive with Triton scaled-matmul at FP32 parity on H100. NumPy / PyTorch / MLX / JAX backends.

Shared neural-network primitives for the mantissa family on the mantissa C engine — engine binding, NumPy reference backend, and Dense/Flatten layers. Base package for cnn/mlp/auto-encoder. By Tekin Ertekin.

LinearCosine: Adding beats multiplying for lower-precision efficient cosine similarity

QuantLab-8bit is a reproducible benchmark of 8-bit quantization on compact vision backbones. It includes FP32 baselines, PTQ (dynamic & static), QAT, ONNX exports, parity checks, ORT CPU latency, and visual diagnostics.

CNN classifier in Python on the mantissa C engine — LeNet-5, VGG-style and AlexNet-style model zoo with paper citations, five classic image datasets, and honest speed/accuracy/RAM benchmarks vs PyTorch, TensorFlow and scikit-learn.

Precision-spectrum GEMM roofline on NVIDIA Blackwell (sm_120), measured in CUDA — FP32 down to FP4, with FP4 (nvfp4) ~20x FP32 via cuBLASLt, anchored to the computed peak.