From 49171eddabbd9ed5dfa917c0481c10f33c8ea43e Mon Sep 17 00:00:00 2001
From: Guangjian <hiro20833@gmail.com>
Date: Thu, 10 Sep 2026 09:28:48 +0800
Subject: [PATCH] [Bugfix] Fail when a diffusion LoRA adapter binds no layer

DiffusionLoRAManager.add_adapter() returns True and activation completes even
when the adapter's target modules match nothing, so the output is identical to
the base model and the only trace is a DEBUG log. vLLM core fails fast on this
mismatch; the diffusion path dropped that guard.

bound_lora_names was only tracked when the pipeline supplies a
_validate_diffusion_lora_binding hook, so pipelines without it recorded nothing.
Track it unconditionally and raise from _bind_adapter_weights when zero layers
bound. _activate_adapter already resets every wrapper and re-raises, so the
rollback path is unchanged.

Closes #7190

Signed-off-by: Guangjian <hiro20833@gmail.com>
---
 tests/diffusion/lora/test_lora_manager.py | 44 +++++++++++++++++++++++
 vllm_omni/diffusion/lora/manager.py       | 13 +++++--
 2 files changed, 54 insertions(+), 3 deletions(-)

diff --git a/tests/diffusion/lora/test_lora_manager.py b/tests/diffusion/lora/test_lora_manager.py
index 0814bf15f1f..10b023bf0ee 100644
--- a/tests/diffusion/lora/test_lora_manager.py
+++ b/tests/diffusion/lora/test_lora_manager.py
@@ -411,6 +411,50 @@ def adapter(adapter_id: int):
     assert len(first.set_calls) == first_calls_after_success + 2
 
 
+def test_lora_manager_rejects_adapter_that_binds_no_layer():
+    """An adapter whose target modules match nothing must fail, not silently no-op."""
+    manager = DiffusionLoRAManager(
+        pipeline=torch.nn.Module(),
+        device=torch.device("cpu"),
+        dtype=torch.bfloat16,
+        max_cached_adapters=1,
+    )
+    layer = _DummyLoRALayer(n_slices=1, output_slices=(2,))
+    manager._lora_modules = {"transformer.blocks.0.attn.to_q": layer}
+
+    # Adapter-side name the engine does not expose, e.g. a diffusers-style
+    # checkpoint against a differently named engine layout.
+    mismatched = LoRALayerWeights(
+        module_name="unet.down_blocks.0.attn.to_q",
+        rank=2,
+        lora_alpha=2,
+        lora_a=torch.ones((2, 2)),
+        lora_b=torch.ones((2, 2)),
+    )
+    manager._registered_adapters = {
+        3: type(
+            "LM",
+            (),
+            {
+                "id": 3,
+                "loras": {"unet.down_blocks.0.attn.to_q": mismatched},
+                "get_lora": lambda self, key: self.loras.get(key),
+            },
+        )()
+    }
+
+    with pytest.raises(ValueError, match="applies to no layer") as excinfo:
+        manager._activate_adapter(3, scale=1.0)
+
+    # The message must name what was received so the mismatch is diagnosable.
+    assert "unet.down_blocks.0.attn.to_q" in str(excinfo.value)
+
+    # Nothing was bound and the adapter must not be left marked active.
+    assert manager._active_adapter_id is None
+    assert layer.set_calls == []
+    assert layer.reset_calls >= 1
+
+
 def _dummy_lora_request(adapter_id: int) -> LoRARequest:
     return LoRARequest(
         lora_name=f"adapter_{adapter_id}",
diff --git a/vllm_omni/diffusion/lora/manager.py b/vllm_omni/diffusion/lora/manager.py
index 2f6d4778056..11d19353158 100644
--- a/vllm_omni/diffusion/lora/manager.py
+++ b/vllm_omni/diffusion/lora/manager.py
@@ -556,9 +556,9 @@ def _is_active_at_scale(self, adapter_id: int, scale: float) -> bool:
 
     def _bind_adapter_weights(self, lora_model: LoRAModel, scale: float) -> None:
         binding_validator = getattr(self.pipeline, "_validate_diffusion_lora_binding", None)
-        lora_names_by_id = (
-            {id(weights): name for name, weights in lora_model.loras.items()} if callable(binding_validator) else {}
-        )
+        # Track bindings unconditionally. The zero-binding guard below needs this
+        # bookkeeping on every pipeline, not only the ones that supply a validator.
+        lora_names_by_id = {id(weights): name for name, weights in lora_model.loras.items()}
         bound_lora_names: set[str] = set()
 
         def _record_bound(weights: LoRALayerWeights | PackedLoRALayerWeights) -> None:
@@ -677,6 +677,13 @@ def _record_bound(weights: LoRALayerWeights | PackedLoRALayerWeights) -> None:
                 scale,
             )
 
+        if not bound_lora_names:
+            raise ValueError(
+                f"LoRA adapter {lora_model.id} applies to no layer: expected target modules in "
+                f"{sorted(self._expected_lora_modules)} but received {sorted(lora_model.loras)}. "
+                "Activating it would leave the base model unchanged."
+            )
+
         if callable(binding_validator):
             binding_validator(
                 lora_model=lora_model,