v0.9.0

Performance optimizations (1.5-2x CUDA speedup)

• Pre-allocate tensors instead of list + torch.stack in rollout loop
• Direct Cholesky noise sampling replacing MultivariateNormal.rsample
• Diagonal covariance detection — element-wise ops instead of matmul (common case)
• M=1 fast path — skip .repeat() and variance when rollout_samples=1 (default)
• Lazy state/action storage when no terminal_state_cost
• torch.einsum for weighted perturbation sum
• Bypass handle_batch_input in hot path
• Resolve all branches at init time
• Cache KMPPI kernel matrix
• functorch.vmap → torch.vmap, torch.inverse → torch.linalg.inv

New features

• mppi.compile(**kwargs) — optional torch.compile for dynamics/cost (beneficial at large K)
• MPPI_Batched — N parallel environments in a single dynamics call

Testing

• 75 unit tests + CI on Python 3.9/3.12
• Solution quality regression tests
• Performance benchmarks

v0.7.3

Compatible with new gym and ray[tune] versions