mstar.engine.resources.position.rope

mstar.engine.resources.position.rope#

The rope kernel, behind a custom op.

FlashInfer reaches its kernel through a TVM-FFI call dynamo can’t trace and can’t run on fake tensors. Called directly it breaks the graph once per layer, and each break makes the layer body a frame dynamo recompiles per layer_idx; behind an op with a registered fake the layer loop stays one graph.