Repository navigation
What is new
ONNX Weight Fusion Pass
Post-export optimization that fuses RMSNorm scales into adjacent MatMul weights and folds Euler step dt into output projections. Zero accuracy cost, fewer graph ops at runtime. Lifted from dexmal/realtime-vla (MIT).
Build Fix
Fixed duplicate ZIP entries in the wheel that caused PyPI upload rejection.
Install
pip install reflex-vla==0.11.1
Full changelog: v0.11.0...v0.11.1