mqa

A powerful, open-source toolkit for audio analysis and playback. Verify lossless quality, detect AI-generated tracks, and explore your library with a built-in hi-res player and advanced EQ.

Master Quality Authenticated codec reverse engineering, Tool to identify MQA encoding and Master's Sample Rate

Collections of papers, databases, and codes targeted at point cloud quality assessment (PCQA), mesh quality assessment (MQA), 3D model quality assessment (3DQA).

Decoding Attention is specially optimized for MHA, MQA, GQA and MLA using CUDA core for the decoding stage of LLM inference.

MHA, MQA, GQA, MLA 相关原理及简要实现

Profile-based three dimensional convolutional neural network for protein model quality assessment

🚀 Accelerate attention mechanisms with FlashMLA, featuring optimized kernels for DeepSeek models, enhancing performance through sparse and dense attention.

A code deep-dive on one of the key innovations from Deepseek - Multihead Latent Attention (MLA)

A fast, lightweight cross-os toolkit for detecting, analyzing, organizing, and exporting MQA-encoded FLAC files.

Metadata Quality Stack is a comprehensive toolkit for analysing metadata quality. It implements the European Data Portal's MQA methodology. Docker Compose deployment and React web application.

Reference Flash Attention implementation in PyTorch with V1/V2, GQA/MQA, Triton kernels, benchmark and docs.

Docker Compose for Metadata Quality Assessment (MQA) on CKAN and European Data Portal catalogs

Web app for evaluating the quality of RDF metadata based on the EDP's MQA methodology. It supports DCAT-AP, DCAT-AP-ES and NTI-RISP (Spanish DCAT). Built with React and TypeScript. It is easily deployable to GitHub Pages.

High-Res Audio management, playback, and streaming app for Android. Features Music Center for real-time Audio Route Path telemetry, target output switching, and a high-performance bit-perfect DLNA/UPnP Media Server.

⚡ Optimize attention mechanisms with FlashMLA, a library of advanced sparse and dense kernels for DeepSeek models, improving performance and efficiency.

Empirical profiling of MHA, GQA, and MQA attention variants measuring KV-cache memory, decode throughput, and output fidelity. GQA-g2 halves KV-cache while preserving ~70% cosine similarity; MQA reduces cache by 92% but drops to ~27% similarity.

Modern LLM Attention from Scratch — MHA, GQA, MQA, RoPE, and KV-Cache implemented in pure PyTorch.

Create a local database of MQA albums from Tidal

buggy TIDAL Downloader