mindspeed-fsdp2-verification

MindSpeed FSDP2 Verification

Objective

Verify the migration artifacts based on the functional and reliability gates, and produce evidence-backed acceptance results.

Applicable Scenarios

• The model, data, and config migration artifacts are ready and awaiting verification.
• The user requests end-to-end acceptance of the MindSpeed-MM FSDP2 migration.
• The migration task claims to be ready and needs evidence to prove it.

Inapplicable Scenarios

• The task is to implement model/data/config code.
• The task only does migration design without execution.
• The task focuses on performance benchmarking.

Inputs

• Migrated model/data/config artifacts
• Real runtime assets (model path, data path, optional image root path)
• Acceptance rules:
  • Enable functional gate
  • Enable reliability gate
  • Require one successful distributed E2E run
  • Do not enable performance gate
  • Do not enable checkpoint recovery gate

Outputs

• verification_report.md
• evidence.json
• failed_cases.md(produced only on failure)

Required Gates

1. Functional gate:
   • Model registration/build pipeline succeeds
   • Data registration/build pipeline succeeds
   • Config associations are consistent
   • At least one distributed end-to-end training run succeeds
2. Reliability gate:
   • Common misconfigurations have clear error behavior
   • Key compatibility checks pass (schema, signature, path type, registration)

Anti-Deviation Execution Protocol

1. Before verification, the following must be completed:
   • Confirm that verification uses real runtime assets and must not use placeholder paths
   • Confirm that the failure backflow template is ready
2. During verification, the following must be executed:
   • Record commands and exit codes separately for the functional gate and reliability gate
   • Execute at least one successful distributed E2E run
3. After verification, the following must be produced:
   • verification_report.md
   • evidence.json
   • failed_cases.md(if failed)

Prohibited Actions

• Prohibited to use print("check ...") instead of real checks
• Prohibited to record only "pass/fail" without attaching command evidence
• Prohibited to claim migration completion when the E2E has not executed successfully
• Prohibited to fail without providing responsible skill mapping and remediation actions

Explicit Exclusions

• This version does not perform performance gating.
• This version does not perform checkpoint save/restore gating.

Execution Rules

• Run verification commands and record exit codes.
• Without command evidence, pass/fail must not be claimed.
• When any gate fails, actionable failure analysis must be returned and mapped to responsible skills:
  • model -> mindspeed-fsdp2-model-migration
  • data -> mindspeed-fsdp2-data-migration
  • config -> mindspeed-fsdp2-config-migration

Exit Criteria

• All enabled gates pass and evidence records are complete.
• The final state is traceable to explicit command output.

Progressive Resources

• checks/verification-gates.md
• checks/verification-hard-gates.md
• templates/evidence-json-template.md
• templates/command-templates.md
• templates/failure-case-template.md
• examples/minimal-io.md
• examples/handoff-context.md

Read these files only when detailed criteria are needed for gate evaluation or evidence formatting.