Objective
Coordinate the complete migration process from the source repository to MindSpeed-MM FSDP2 by scheduling dedicated sub-skills and executing acceptance gates.
Applicable Scenarios
- The user requests migrating a new open-source model to MindSpeed-MM FSDP2.
- The user needs end-to-end migration orchestration covering model, data, configuration, and validation.
- The migration task requires sequential advancement across modules and a unified summary report.
Inapplicable Scenarios
- The task only involves model code adaptation.
- The task only involves dataset adaptation.
- The task only involves YAML/configuration adjustments.
- The task only involves post-implementation validation.
Input
source_repo_pathtarget_repo_pathmodel_identity(model name and modality)source_entrypoints(training, data, model entry points)runtime_assets(real runtime assets, such asmodel_path、dataset_path, optionalimage_root)constraints(editable paths, core changes prohibited, reuse first)acceptance(functional + reliability gates, and a single distributed E2E run passes)
Output
changeset_manifestcompatibility_matrix(model/data/configuration mapping)verification_evidencerisk_registernext_actions
K0 Knowledge Reserve Gate (must be completed before execution)
- Required reading documents:
../docs/MIGRATION_GUIDE_MINDSPEED_MM_FSDP2.md../docs/MIGRATION_GUIDE_MINDSPEED_MM_FSDP2_DEVELOPER.md
- Required directories to scan:
MindSpeed-MM/mindspeed_mm/fsdp/models/MindSpeed-MM/mindspeed_mm/fsdp/data/datasets/MindSpeed-MM/examples/fsdp2/
- Required understanding deliverables:
architecture_chain.md(minimal pipeline and key interface mapping)contract_matrix.yaml(model/data/config contract matrix)similar_case_selection.md(most similar cases, reusable points, risk points)preflight_risk_register.yaml(high/medium/low risks and mitigation actions)
- Stop if not passed:
- If any deliverable is missing or its content cannot be traced to code/documentation evidence, entering the implementation phase is prohibited
Mandatory Input Validation (must be completed before execution)
- Path and asset existence:
source_repo_path、target_repo_pathmust exist and be readableruntime_assets.model_path、runtime_assets.dataset_pathmust exist and be readable
- Entry point validity:
source_entrypoints.train、source_entrypoints.model、source_entrypoints.datasetmust exist
- Editable scope:
- Changed paths must fall within
constraints.editable_pathswithin
- Changed paths must fall within
- Failure if missing:
- If any validation fails, directly output the failure reason and repair suggestions, and do not enter the implementation phase
Workflow
- Complete the K0 knowledge reserve gate and deliver understanding deliverables.
- Discover the source repository and target repository entry points.
- Build the migration mapping and classify items as "reusable" and "must adapt".
- Scheduling:
mindspeed-fsdp2-model-migrationmindspeed-fsdp2-data-migrationmindspeed-fsdp2-config-migration
- Merge deliverables and run
mindspeed-fsdp2-verification。 - If validation fails, route back to the corresponding sub-skill for repair according to ownership.
- Produce the final migration summary report.
Phased Delivery Contract (to prevent execution deviation)
At the end of each phase, an "inspectable deliverable" must be delivered; otherwise, entering the next phase is prohibited:
- Discovery phase deliverables:
source_target_inventory.yamlentrypoint_checklist.md
- Mapping phase deliverables:
compatibility_matrix.yamlreuse_vs_adapt.md
- model/data/config phase deliverables:
- Each sub-skill separately
acceptance_checklist.md changeset_manifest.yaml
- Each sub-skill separately
- Validation phase deliverables:
verification_report.mdevidence.jsonfailed_cases.md(only on failure)
- Summary phase deliverables:
migration_report.mdrisk_register.yamlnext_actions.yaml
Failure Attribution and Backflow Rules (must be executed)
from_pretrained/_from_config/.losscontract issue -> modeldataset_type/ missing field / collate issue -> datamodel_id/dataset_type/training.plugin/ strict-extra layering -> config- When attribution is impossible:
- First output a minimal reproduction and evidence
- Then assign handling according to the "first failing gate", and do not attribute vaguely
Prohibited actions
- Do not skip the K0 knowledge preparation gate and implement directly
- Do not start code changes before completing prerequisite validation
- Do not skip any phase deliverable
- Do not claim "passed" without command evidence
- Do not write the cause of failure as "possibly/probably" without providing reproduction evidence
- Do not bypass
constraints.editable_pathsto modify the core framework
Mandatory rules
- Each sub-skill artifact must follow single-responsibility ownership.
- Do not claim completion without verification evidence.
- Enforce plugin/registration consistency:
model.model_iddata.dataset_param.dataset_typetraining.plugin
- Migration changes must be limited to the allowed editable paths.
Exit criteria
- Both the functional gate and the reliability gate pass.
- At least one distributed end-to-end training run succeeds.
- All conclusions in the final report are supported by evidence.
Progressive resources
checks/orchestration-checklist.mdchecks/knowledge-gates.mdchecks/preflight-gates.mdchecks/phase-delivery-gates.mdtemplates/migration-report-template.mdtemplates/fsdp2-migration-runbook.mdtemplates/fsdp2-migration-quickstart-zh.mdtemplates/knowledge-artifacts-template.mdtemplates/compatibility-matrix-template.mdtemplates/failure-routing-template.md
Read these files only when the current step requires a detailed checklist, Runbook process, Quickstart guide, or report format.
External references (key points of migration technology)
../docs/MIGRATION_GUIDE_MINDSPEED_MM_FSDP2.md../docs/MIGRATION_GUIDE_MINDSPEED_MM_FSDP2_DEVELOPER.md