Fallback for unsupported Hybrid Engine policies

[LiRunGuo]

Summary

• register Hybrid Engine auxiliary Linear/Embedding/LayerNorm policies only when the model contains a supported complete transformer policy
• allow unsupported architectures such as Qwen2.5 to retain the native generate() fallback
• add CPU-only unit coverage for supported and unsupported policy registration

Fixes #8263.

Why

The generic wrappers are auxiliary pieces of a complete transformer injection policy. Registering them when no transformer layer matches creates a partial inference path. With a ZeRO-3-partitioned Qwen model, container construction can then reach _mark_uc_metadata() with a one-dimensional normalization weight and fail on weight.shape[1].

The change first builds the complete-policy map and checks it against the model's module classes. If no complete policy matches, the map remains empty, no partial containers are created, and Hybrid Engine leaves the model's native generation method intact. Supported model behavior is unchanged.

Validation

• pytest -q tests/unit/hybrid_engine/test_he_policy.py (2 passed)
• pre-commit run --files deepspeed/runtime/hybrid_engine.py tests/unit/hybrid_engine/test_he_policy.py
• Qwen2.5-0.5B / Qwen2.5-Math-7B OPSD completed a full prompt epoch and a separate 200-step run on 8 x MI250 with the fallback

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.

[delock]

Can you add a logger.warning here for not matching inference_policies? This would give user proper expectations.