[bugfix]: run the base layer while a LoRA wrapper has no adapter

fastvideo/layers/lora/linear.py

@@ -163,8 +163,11 @@ class BaseLayerWithLoRA(nn.Module):
             lora_B = self.lora_B.to_local()
             lora_A = self.lora_A.to_local()
 
-        if self.side_branch and (self.lora_A is None or self._strength_is_zero):
-            # No adapter, or strength 0, is exactly the unadapted layer, down to the output dtype.
+        if self.lora_A is None or (self.side_branch and self._strength_is_zero):
+            # No adapter, or strength 0 on a side branch, is exactly the unadapted layer, down to the output dtype.
+            # A merge-mode wrapper reaches this branch while its first adapter is still loading or after a load
+            # failed: the pipeline wraps every linear before it reads the weights, so an unadapted request must
+            # not depend on an adapter that is not there.
             return self.base_layer(x)
 
         if not self.merged and not self.disable_lora:

fastvideo/tests/layers/test_lora_side_branch.py

@@ -54,6 +54,20 @@ def _packed_linear(seed: int = 0) -> ReplicatedLinear:
     return layer
 
 
+def test_a_wrapped_layer_without_an_adapter_runs_the_base_layer() -> None:
+    """A dense wrapper with no adapter is the base layer.
+
+    The pipeline wraps every linear when the first runtime adapter is set. A render that then runs without one, or a
+    load that failed part-way, must still produce the base model's output instead of failing on the missing factors.
+    """
+    base = _dense_linear()
+    wrapped = get_lora_layer(base)
+    assert not wrapped.side_branch and wrapped.lora_A is None
+    x = torch.randn(2, IN, dtype=torch.bfloat16)
+    with torch.no_grad():
+        assert torch.equal(wrapped(x)[0], base(x)[0])
+
+
 def _factors(seed: int = 1) -> tuple[torch.Tensor, torch.Tensor]:
     generator = torch.Generator().manual_seed(seed)
     return torch.randn(RANK, IN, generator=generator), torch.randn(OUT, RANK, generator=generator)