[JiaoliangYu]

#2776 discarded scheduled_tokens_gpu whenever use_ep was false, so pure TP lost the real-token count introduced by #2725. Restore the count for ungathered TP rows; the existing MoE methods forward it to AITER as num_local_tokens. DP/PCP-gathered rows still receive None, and pass_scheduled_tokens_gpu rejects row subsets that do not match the current pass.

Run the routed-row regression tests on CPU under AITER/Triton stubs. During this test module's import, bypass CUDA custom-op registration so fake enum defaults do not enter schema inference or leave stub kernels in PyTorch's process-wide registry. The TP/EP regression exercises the production _run_experts and pass_scheduled_tokens_gpu together: a complete local pass receives its count, while gathered rows and row subsets do not.

Dependency and backend scope: ROCm/aiter#7294 is still unmerged. Restoring num_local_tokens on older AITER with padded CUDA-graph shapes can reintroduce its atomic GEMM2 padding-sentinel regression. The stage2 paths covered by that dependency are FlyDSL, FlyDSL-v2, and native-MXFP4. The current target configuration does not use Opus GEMM2. Keep this PR in draft while #7294 is unmerged and the requested pure-TP decode performance comparison remains pending.

Validation:

• Python 3.12.15 with CPU PyTorch 2.15.0, without AITER or Triton: tests/test_pass_scheduled_tokens.py and tests/model_ops/test_dp_gather_routed_rows.py completed with 10 passed and no skips.
• Full non-GPU suite in a native Linux checkout, using the same exclusions as Pre Checkin: 10,698 passed, 1,139 skipped, 3 xfailed, with no failures.
• Pre Checkin on 9705a5e76 passed all checks. The CPU job reported 11,007 passed, 1,154 skipped, 3 xfailed and executed all seven routed-row tests successfully. The offline inference smoke test also passed.
• Black (--target-version py312), Ruff, and git diff --check passed for the changed file.
• GSM8K was not run. Pure-TP decode performance validation remains pending.

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2905 --add-label <label>

[valar-cop]

some changes needed..

pure-TP decode perf with aiter#7294 vs main, and gsm8k for a padded cudagraph bucket?

ci green.

[valar-cop]

some changes needed..

pure-TP decode perf with aiter#7294 vs main, and gsm8k for a padded cudagraph bucket on the pure-TP path?

ci green.