Add warning for unsupported Hybrid Engine policies

Signed-off-by: LiRunGuo <li19107254665@gmail.com>

Author: LiRunGuo

deepspeed/runtime/hybrid_engine.py

@@ -125,6 +125,12 @@ def populate_all_inference_policies(self):
 
         has_transformer_policy = any(module.__class__ in self.inference_policies for module in self.module.modules())
         if not has_transformer_policy:
+            logger.warning(
+                "No compatible DeepSpeed inference policy found for model type %s. "
+                "Hybrid Engine inference acceleration is unavailable; rollout will "
+                "use the model's native generate() path.",
+                type(self.module).__name__,
+            )
             self.inference_policies = {}
             return
 

tests/unit/hybrid_engine/test_he_policy.py

@@ -29,13 +29,14 @@ def _make_engine(module):
     return engine
 
 
-def test_unsupported_model_uses_native_fallback(monkeypatch):
+def test_unsupported_model_uses_native_fallback(monkeypatch, caplog):
     monkeypatch.setattr(hybrid_engine, 'replace_policies', [SupportedPolicy])
     engine = _make_engine(nn.Sequential(UnsupportedLayer(), nn.Linear(2, 2), nn.LayerNorm(2)))
 
     engine.populate_all_inference_policies()
 
     assert engine.inference_policies == {}
+    assert "Hybrid Engine inference acceleration is unavailable" in caplog.text
 
 
 def test_supported_model_registers_auxiliary_policies(monkeypatch):