[richardoo-707]

Problem and change

Fixes #6485. Stacked on #6477: base feat/moe-ep-layer, head fix/ep-default-process-group.

When distributed is initialized, constructing EPMoELayer with its default ep_group=None computes the WORLD size correctly but tags expert parameters with None. The existing get_ep_size(param) then raises an assertion. Resolve the default group to dist.group.WORLD before creating the default dispatcher or tagging parameters. Explicit subgroups and construction before distributed initialization retain their current behavior.

Add a CPU/Gloo regression for initialized and uninitialized construction, include it in both CPU batch manifests, and extend the existing four-GPU parameter-contract test to cover WORLD and explicit subgroups. No changes to gradient scaling, the getter API, checkpoint layout, dependencies, image configuration, or version.txt.

Validation

Tested commit: c1702d851d9086415e6a942224220ddfdca57e60 (parent 055853017e23c778402b2c8d965838cb09fb43e4).

Check	Result
New CPU regression with original parent layer.py mounted over the fixed snapshot	1 failed, 1 passed; exit 1, confirming the regression detects the original default-group mismatch
New CPU regression on the fixed snapshot	2 passed, 0 skipped; exit 0
Repository CPU gate	29 passed, 0 skipped; exit 0; includes both new regression cases
Targeted Black/isort and git diff --check	Passed
Four-GPU test_ep_moe_layer.py / Layer 3	Not run: all eight GPUs on the development server had active compute processes; no GPU was acquired

Commands inside the existing dependency image, with /workspace bound read-only to the exact source archive and zero visible GPUs:

python -m pytest -p no:cacheprovider -v -ra --tb=short \
  --junitxml=/results/cpu-after.xml tests/test_moe/test_ep_moe_layer_cpu.py

PYTEST_ADDOPTS='-p no:cacheprovider' COLOSSAL_RESULTS=/results/cpu-gate \
  bash .github/scripts/ci/run_cpu_unit_tests.sh

The targeted regression ran with FAST_TEST unset. The CPU gate used its normal FAST_MODE=1 / FAST_TEST=1 setting; these two new cases are not reduced by that setting. No new skip, xfail, ignore, tolerance change, or failure retry was added. One initial CPU-gate launch stopped before pytest because Windows git archive exported CRLF shell scripts; regenerating the unchanged commit with git -c core.autocrlf=false archive fixed the source packaging. Results above are from the corrected archive.

Environment: Ubuntu 22.04.5 LTS; Python 3.10.12; Torch 2.13.0+cu130 (CUDA runtime 13.0); Transformers 5.17.0; PEFT 0.21.2; pytest 7.4.4. GPU count 0, Gloo backend; no CUDA kernels/toolkit or flash-attn path exercised. Image: wangzhijian/colossalai-ci:torch2.13-cu130-tf5.17.0-peft0.21.2-20261022-091529, local image ID sha256:ae8ff34bf340a692803edca6e0f38a576fc1a2b206c01b4ef6f7c2a40dfd8c10. Containers had no network or GPU access and used the existing environment without installing dependencies.

The source import was verified as /workspace/colossalai/__init__.py. Source archive SHA256: 56af8c927ba8d094d728bb00124feb8e31544c2d49c38cc1d57ca2ee555c1fc1.

Logs/JUnit on server 30005: /mnt/beegfs/ricardoo/codex-ep-default-group-20261027/results/ (cpu-before.*, cpu-after.*, cpu-gate/cpu-local-1.{log,xml}). The reproducible red/green command wrapper is validate.sh in the same task directory.

Risk and rollback

The runtime change only makes default-group metadata explicit after distributed initialization; the communication group and numerical operations are unchanged. Revert this commit to roll back. No release or installation changes are needed. Keep as Draft pending the four-GPU regression; this validation does not claim a Layer 3 pass.