GRADSOLVE: fast exact gradients for ODE ensembles on GPUs
GRADSOLVE claims reverse-mode ODE gradients on NVIDIA GPUs without giving up fused-kernel solver speed.
The JAX library records accepted adaptive-solver steps, then replays them as a fixed-step chain to compute the exact discrete adjoint more cheaply. The paper targets low-dimensional ODE ensembles that are differentiated repeatedly against one recorded mesh, with Diffrax kept as a fallback. Reported gradient speedups over Diffrax’s checkpointed adjoint are 5.6x to 14.1x after recording, narrowing for large ensembles and reaching parity on stiff systems at tight accuracy. ArXiv · AI/CL/LG's note
The JAX library records accepted adaptive-solver steps, then replays them as a fixed-step chain to compute the exact discrete adjoint more cheaply. The paper targets low-dimensional ODE ensembles that are differentiated repeatedly against one recorded mesh, with Diffrax kept as a fallback. Reported gradient speedups over Diffrax’s checkpointed adjoint are 5.6x to 14.1x after recording, narrowing for large ensembles and reaching parity on stiff systems at tight accuracy. ArXiv · AI/CL/LG's note
score 5