Fast-loading implementation sequence#

Date:

2026-08

Large-model startup follows one explicit four-document sequence:

  1. fix existing parser, external-data, and initializer-materialization defects in Model-loading bug fixes;

  2. implement Prepared and asynchronous execution;

  3. complete the ownership-aware cross-repository work in Using onnx-light fast loading in onnxruntime;

  4. connect adaptive I/O, model resolution, prepared tensors, and first-token overlap in Completing native fast model loading.

Parallel-for profiling may proceed alongside step 1, but its executor instrumentation must be stable before step 2 begins.

Assignable issue sequence#

Step

Issues

Order

  1. Bug fixes

#4608–#4610

#4608 -> #4609 -> #4610

  1. Prepared execution

#4613–#4617

#4613 -> #4614 -> #4615 -> #4616 -> #4617

  1. onnxruntime

#4611–#4612

#4611 -> #4612. #4612 is coordination for a PR in microsoft/onnxruntime and must not be assigned to an agent working only in xadupre/onnx-light.

  1. Native completion

#4618–#4623

#4618 -> #4619 -> #4620 -> #4621 -> #4622 -> #4623

Plans#