Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

ops: Handle CPU weight in VBAR caster (#12792)

rattus42e0e023eee6
diff --git a/comfy/ops.py b/comfy/ops.py
index 3e19cd1b6..06aa41d4f 100644
--- a/comfy/ops.py
+++ b/comfy/ops.py
@@ -80,6 +80,21 @@ def cast_to_input(weight, input, non_blocking=False, copy=True):
def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant):
+
+ #vbar doesn't support CPU weights, but some custom nodes have weird paths
+ #that might switch the layer to the CPU and expect it to work. We have to take
+ #a clone conservatively as we are mmapped and some SFT files are packed misaligned
+ #If you are a custom node author reading this, please move your layer to the GPU
+ #or declare your ModelPatcher as CPU in the first place.
+ if device is not None and device.type == "cpu":
+ weight = s.weight.to(dtype=dtype, copy=True)
+ if isinstance(weight, QuantizedTensor):
+ weight = weight.dequantize()
+ bias = None
+ if s.bias is not None:
+ bias = s.bias.to(dtype=bias_dtype, copy=True)
+ return weight, bias, (None, None, None)
+
offload_stream = None
xfer_dest = None