|
Download tasks/format-code-task-000018/instruction.md from FineEnvs/MiMo-V2.6-RL-harbor-code: direct link, hf CLI and curl.
- Browser
- Download file 3.07 kB
-
https://huggingface.co/datasets/FineEnvs/MiMo-V2.6-RL-harbor-code/resolve/main/tasks/format-code-task-000018/instruction.md
- Command line
-
hf download hf://datasets/FineEnvs/MiMo-V2.6-RL-harbor-code/tasks/format-code-task-000018/instruction.md
-
curl -L -o instruction.md https://huggingface.co/datasets/FineEnvs/MiMo-V2.6-RL-harbor-code/resolve/main/tasks/format-code-task-000018/instruction.md
3.07 kB
| Fix the following issue: | |
| # Persist sublattice metadata atomically and track the root dispatch | |
| When the dispatcher runs a workflow that contains sublattices, each sublattice is | |
| itself dispatched and gets its own lattice record in the results database. Two | |
| things are currently awkward about how those records are created. | |
| First, the database column that links a sublattice's lattice record back to the | |
| parent electron that spawned it (`electron_id`) is only filled in *after* the | |
| sublattice has already been written to the database. There is a window during | |
| which a sublattice record exists with no link to its parent, so an interruption | |
| can leave it orphaned. We want the parent electron id to be written in the *same* | |
| insertion that creates the lattice record, not in a later update. | |
| Second, we are about to need a way to identify, for any (sub)lattice in a | |
| hierarchy, the dispatch id of the top-level ("root") workflow that ultimately | |
| kicked everything off. A nested sublattice three levels deep should still be able | |
| to report the dispatch id of the original top-level dispatch. | |
| Please make the following behavior available. | |
| ## `root_dispatch_id` on the result object | |
| The result object should expose a read-only `root_dispatch_id`. For a result | |
| object that represents a top-level dispatch, this equals its own dispatch id. A | |
| result object constructed for a given dispatch id reports that same id as its | |
| root dispatch id until it is told otherwise (see the factory below). | |
| ## Persisting a parent electron id atomically | |
| Persisting a result object should optionally accept the database id of the parent | |
| electron that spawned this workflow. When provided, that id must be stored on the | |
| lattice record in the very same transaction that first creates the lattice | |
| record — not written by a subsequent update. When it is not provided, the | |
| lattice record's `electron_id` stays null. Persisting a top-level workflow must | |
| continue to behave exactly as before (its `electron_id` is null). | |
| ## A factory for building a result object from a serialized lattice | |
| Add a convenience factory, importable as | |
| `covalent._results_manager.result.initialize_result_object`, that constructs and | |
| persists a result object from a JSON-serialized lattice. It takes the JSON | |
| lattice and, optionally, a parent result object and a parent electron id: | |
| ``` | |
| initialize_result_object(json_lattice, parent_result_object=None, parent_electron_id=None) | |
| ``` | |
| It must: | |
| - deserialize the lattice and assign the new result object a freshly generated, | |
| unique dispatch id (two calls never collide); | |
| - initialize the result object's nodes and persist it before returning, passing | |
| along the parent electron id so it lands atomically on the lattice record as | |
| described above; | |
| - when a parent result object is supplied (i.e. this lattice is a sublattice), | |
| make the new result object inherit the parent's `root_dispatch_id` rather than | |
| using its own dispatch id; | |
| - return the fully constructed result object. | |
| The returned result object's own dispatch id must still be distinct from its | |
| parent's. | |