AI Gets Faster: New Code Change Makes Chatbots Remember More

Dynamo's scan HOP now opts into input-input aliasing, because Dynamo already had supports_input_input_aliasing. That means combine_fn may now take several inputs that are views of one storage — for example K and V split from a fused QKV projection. Such calls were rejected before, so the caller had to clone.

Aliased inputs are accepted as long as none of the inputs sharing that storage is mutated: functionalization treats subgraph inputs as independent, so a write to one view would not be observed through another. Mutating one now raises its own graph break, GB8504, with a hint to clone one of the aliased inputs. Mutating an input that shares no storage is unaffected, and input-output and output-output aliasing stay rejected.

Because scan allows mutation of additional inputs under inference, a new SubgraphTracer.has_aliased_input_mutation check runs in check_aliasing_and_input_mutation for HOPs that set the flag, and _check_alias_and_mutation takes the same flag so functionalization agrees. It still rejects any input mutation, and the auto_functionalize path is only reached with Dynamo-set mutated_arg_indices, so the Dynamo guard covers it. Authored with the assistance of an AI coding assistant (Claude Code). Pull Request resolved: #199688, approved by ydwu4.