D^2-MoE: Delta Decompression for MoE-based LLMs Compression

Decentralized personalized federated learning based on a conditional sparse-to-sparser scheme (TNNLS)

[AICCSA 2025] Official Implementation of the paper "Z-Pruner: Post-Training Pruning of Large Language Models for Efficiency without Retraining".

Code for Frequency Matters: Fast Model-Agnostic Data Curation for Pruning and Quantization

A custom PyTorch neural network that implements parameter pruning through an $L_1$-regularized, learnable gating mechanism. By wrapping standard linear weights in a trainable gate, the model actively learns to balance classification accuracy and network sparsity, demonstrating the trade-off through evaluated metrics on the CIFAR-10 dataset.

Developing efficient deep learning models for real-world use. Covers knowledge distillation, quantization, pruning, and more. Focused on reducing size and latency while preserving accuracy. Includes training pipelines, visualizations, and performance reports.

A GPU kernels for N:M structured sparse matrix multiplication (50% sparsity, 2:4 pattern) on NVIDIA RTX 4070.

Certain Finetuning , Inferencing, Optimization Tasks(Pruning, Distillation) for better LLM Performance