mindspeed-fsdp2-data-migration

MindSpeed FSDP2 Data Migration

Objectives

Migrate the source dataset logic into the MindSpeed-MM FSDP2 data plugin structure, preserve preprocessing semantics, and ensure trainer/model input compatibility.

Applicable Scenarios

• The migration task requires creating a data plugin under mindspeed_mm/fsdp/data/ .
• The source preprocessing function needs to be reused or minimally adapted.
• Need to implement dataset_type registration,__getitem__ and collate_fn。

Non-Applicable Scenarios

• The task is model loading or model registration.
• The task only performs configuration mapping and does not involve data behavior changes.
• The task is the final verification gate execution.

Inputs

• Source dataset entry file
• Target dataset_type
• Expected model input fields
• Data and path format assumptions
• runtime_assets.dataset_path

Outputs

• Data plugin implementation
• Data migration checklist with pass/fail status
• Adaptation rationale for "non-direct-copy changes"

Mandatory Checks

1. Registration contract:
   • @data_register.register("<dataset_type>")
2. Data contract:
   • __getitem__ returns the fields required by the model forward pass
   • collate_fn can safely handle variable-length batches
3. Reuse contract:
   • When compatible, the source preprocessing function should be reused first
4. Robustness contract:
   • Safely handle path type variants (str、list[str]）
   • Raise clear errors for invalid schema assumptions

Bias-Prevention Execution Protocol

1. Must be completed before implementation:
   • Validate that runtime_assets.dataset_path exists and is readable
   • Sample and read the first N samples to confirm the schema fields exist
2. Must be produced after implementation:
   • dataset_schema_report.md
   • fields_contract_report.md
3. Stop on failure:
   • __getitem__ If fields are incomplete or collate crashes, entering the configuration/verification phase is prohibited

Prohibited Actions

• It is prohibited to assume the path type is only str
• It is prohibited to silently swallow schema errors
• It is prohibited to delete modality-critical fields (pixel_values、image_flags etc.) without declaring it
• It is prohibited to replace "real path sampling verification" with "sample data passing"

Implementation Rules

• First copy the core preprocessing logic from the source repository.
• The adaptation scope is limited to imports, registration, field mapping, and contract alignment.
• Modality-critical fields must not be silently discarded.
• Clearly record the reason each function cannot be directly reused.

Exit Criteria

• The dataset can be constructed via build_mm_dataset .
• The data fields and collate output meet model expectations.
• All conclusions in the data-side checklist are supported by evidence.

Progressive Resources

• checks/data-acceptance.md
• checks/data-contract-gates.md
• errors/data-error-dictionary.md
• errors/triage-severity.md
• examples/minimal-io.md

Read these files only when detailed criteria are needed for acceptance or data error triage.