GLM-5.2-SQG-Coupled-H512-H128-K96Tail / MTP78_PREPARATION_BINDING_REPAIR.md
brandonmusic's picture
Add files using upload-large-folder tool
8aced41 verified
|
Raw History Blame Contribute Delete
1.29 kB

MTP78 preparation-binding orchestration repair

The MTP-native adapter completed all 256 layer-78 preparation permutations and wrote:

/workspace/sqg-run/mtp78/layer_078/preparation/preparation.json

Its SHA-256 was 548eb458518b478d51dce3405b8804c1ec29f3e44d6e1c09d4d626b3d1f2a826. The adapter's incremental one-work-unit branch published the command receipt but omitted the ordinary worker's final publish_preparation_binding(layout) orchestration step. On the next invocation the command receipt correctly prevented recomputation, while the supervisor-domain completion file remained absent, producing MTP78 stage has no ready atomic work unit.

With the campaign stopped, the existing run_layer_stage implementation was loaded and publish_preparation_binding(layout) was called for the MTP-native layout. It created:

/workspace/sqg-run/preparation/layer_078/preparation/preparation.json

The binding has schema glm52-b300-sealed-preparation-binding-v1, records numerical_recomputation: false, and binds the exact source path and SHA-256 above. No numerical tensor, permutation, profile, Hessian, encoded expert, or sealed source artifact was recomputed or changed.

The resulting binding SHA-256 is 09e4efc78d800fc2f9272efb28bda6c17ecb90838872c75c047406afb4f538c5.