Repository navigation
Tags
Tags
Release 0.2.1: compare graphs with graphs 0.2.0 compared capture=True against baselines without CUDA graphs, so its 12x mixed two effects: graph replay, and gradbridge doing the capture. Both examples now time every approach without and with graphs: - Newton's native loop, with the graph exactly as example_diffsim_ball captures it - a hand-written autograd.Function, with graphs captured by hand - gradbridge, with capture=True They also check that every approach produces the same gradient. With graphs on both sides the approaches match (Newton: 5.0 / 5.2 / 5.2 ms; projectile: 2.1 / 2.0 ms). The README and guide now claim parity with hand-captured graphs in one line, plus the cases gradbridge keeps correct, instead of a speedup.
Release 0.2.0: README shows measured value; standalone training example README gains a "What it buys you" section: training a PyTorch network through a Warp simulation with a hand-written bridge, gradbridge and capture=True, with numbers reproducible from two examples in the repo. examples/train_policy_through_warp.py needs only PyTorch and Warp. The release workflow also checks that _jax_op.cc ships in the wheel.