Thriving JU A correctness-gated platform for attention surgery, compression, and hardware-aware benchmarking
https://tahamajs.github.io/NeuralOps/ ↗ HEALTH ? 90 / Healthy
Thriving JU // readme
🧬 NeuralOps v2.0.5
A correctness-gated platform for attention surgery, compression, and hardware-aware benchmarking
131 registered attention mechanisms · 72 compression engines · 110 canonical CUDA modules (114 native implementations) · 26 Triton kernel modules (62 registered unified operators) · 94 comprehensive notebooks present in the checkout
Registry count contract: the current remote registry reports 131 attention mechanisms and 72 compression engines. Hardware source counts are reported separately.
📊 Release at a Glance
| Surface | Verified scope |
|---|---|
| Attention registry | 131 mechanisms across dense, linear, gated, SSM, sparse, and hybrid families |
| Compression registry | 72 engines covering SVD, quantization, pruning, and tensor decompositions |
| Hardware catalog inventory | 110 canonical CUDA modules, 114 native implementations, 26 Triton modules, 62 registered Triton operators |
| Notebook suite | 94 comprehensive interactive notebooks covering full-stack surgery, adaptation, benchmarks, distributed systems, and compiler graphs |
| Academic publications | Master’s Thesis (317 pages, 205 MB), Defense… |