Ascend/megatron-change-analyzer

Megatron Change Analyzer

Turn a raw Megatron change set into structured feature evolution findings that are ready for downstream implementation work. Focus on extracting the real upstream feature shape, not on writing MindSpeed code yet.

Core Rules

• Assume input is a normalized Megatron change-set, not an arbitrary pile of commits.
• Group multiple low-level commits into higher-level change events when they serve one feature.
• Prioritize new features, API shape changes, config or CLI schema changes, and behavior-affecting refactors.
• Tag each event with migration relevance, but do not edit MindSpeed here.
• Treat the upstream commit itself as implementation evidence, not just as a summary source.
• When a feature is represented by a specific commit or tightly related commit cluster, preserve that grouping so downstream steps can port the implementation with the commit as the primary reference.
• Treat megatron main findings as exploratory unless the user has supplied a strict target mapping.

Output Shape

Produce a structured report with one event per feature or migration-relevant change:

{
  "branch": "core_v0.15.3",
  "events": [
    {
      "title": "Add feature X",
      "kind": "new_feature",
      "commits": ["sha1", "sha2"],
      "primary_commit": "sha2",
      "areas": ["training", "config"],
      "breaking_risk": "medium",
      "migration_relevance": "high",
      "notes": "Short factual summary",
      "evidence": ["file/path.py", "arguments.py"],
      "upstream_changed_files": ["megatron/path_a.py", "megatron/path_b.py"],
      "implementation_units": [
        {
          "name": "Expose new config flag",
          "kind": "config_surface",
          "upstream_files": ["megatron/training/config/training_config.py"],
          "summary": "What this unit changes in upstream code"
        },
        {
          "name": "Add runtime behavior",
          "kind": "runtime_logic",
          "upstream_files": ["megatron/training/global_vars.py"],
          "summary": "What runtime path must be ported downstream"
        }
      ],
      "porting_notes": [
        "Facts the downstream mapper should preserve when porting the feature"
      ]
    }
  ]
}

Workflow

1. Read the normalized change-set.
2. Inspect commit titles, touched files, and diffs at a feature level.
3. Collapse related commits into one event when appropriate, but keep the primary implementation commit visible.
4. Label each event using the taxonomy in change-taxonomy.md.
5. Break each relevant event into implementation units: config and CLI exposure, runtime logic, wrappers or adaptors, tests and examples, lifecycle or cleanup behavior.
6. Separate events into: relevant for MindSpeed, probably already covered, not currently worth adaptation.
7. Hand off only the relevant subset to the impact mapper, including commit references, upstream changed files, and implementation units.

What To Highlight

• Newly exposed features or workflows
• Public API additions or signature changes
• CLI/config additions, removals, or renames
• Checkpoint or state format changes
• Parallelism and distributed execution changes
• Data pipeline or runtime behavior changes

What To Avoid

• Do not produce code patches here.
• Do not claim MindSpeed compatibility from Megatron evidence alone.
• Do not treat internal refactors as high-value migration work unless they change integration surfaces.
• Do not collapse a large feature into a one-line note if the upstream commit actually changes multiple integration surfaces. Preserve enough structure for downstream implementation.

References

• Read change-taxonomy.md before labeling events.
• Run build_feature_events.py when you need a deterministic first-pass event file that preserves primary commits, upstream changed files, and implementation units.
• Hand migration-relevant events to $megatron-impact-mapper.