Repository navigation
NeuralOps v1.5.0: FLA-style benchmark architecture, unified APIs, PEFT fixes, and verified fallback telemetry.
Verified benchmark ledger: 203 logical modules; 42 PASS_FAST, 161 PASS_SLOW, 0 fallback, 0 timeout. Backend telemetry: 6 CUDA, 1 Triton, 196 PyTorch.
The release includes the three-level CUDA -> Triton -> PyTorch fallback architecture, unified attention/training APIs, and kernel hardening.