Skip to content

Autotp z3#3

Open
jinyouzhi wants to merge 3 commits into
masterfrom
autotp_z3
Open

Autotp z3#3
jinyouzhi wants to merge 3 commits into
masterfrom
autotp_z3

Conversation

@jinyouzhi

Copy link
Copy Markdown
Owner

No description provided.

@jinyouzhi

Copy link
Copy Markdown
Owner Author

@copilot check the progress and give some guide

Copilot AI commented Jul 15, 2026

Copy link
Copy Markdown

@copilot check the progress and give some guide

Current progress on my changes is limited to 2 commits:

  • c9691dd: enables AutoTP with ZeRO-3 in deepspeed/runtime/engine.py and adds runtime guards for unsupported combinations (pipeline parallelism, param offload, MiCS, tp_overlap_comm).
  • 90e92be: formatting-only cleanup in the new test coverage.

Validation added in tests/unit/model_parallelism/test_autotp_training.py is a new TestAutoTPZero3Runtime distributed smoke test covering column-parallel and column->row-parallel flows.

Guide for next checks:

  1. Run python -m pytest -q tests/unit/model_parallelism/test_autotp_training.py -k 'TestAutoTPZero3Runtime' in a test environment with pytest installed.
  2. Add negative tests for each runtime guard if we want stronger coverage of the unsupported ZeRO-3 combinations.
  3. If CI exposes failures outside those guarded cases, the next place to inspect is ZeRO-3 parameter metadata/process-group assignment during AutoTP init and step.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants