trunk/3afa7072db9835cb0996cb5e16279b4a51c3750c: [Inductor] Bound incremental autotuning dispatches (#196501) (#196501)
- PyTorch: 2083 events in the last 90 days
- PyTorch: 2054th Release in the last 90 days
- Previous: earlier the same day · trunk/81f41c172b3ca4628d9de18f7df2a518a65b90c5: C++ helpers for SymInt-aware elementwise binary meta kernels (#198875)
What happened
Summary: Add a state-wide dispatch budget for incremental autotuning. incremental_autotune_max_dispatches defaults to 200 . Override it through torch.compile() options or TORCHINDUCTOR_INCREMENTAL_AUTOTUNE_MAX_DISPATCHES ; 0 disables the cap, and negative values are rejected. When an enabled budget is exhausted, freeze the current best launcher immediately. Pending timing events still drain but cannot change the sel…
Summary assembled by rule from the sources below