From 42e0e023eee6a19c1adb7bd3dc11c81ff6dcc9c8 Mon Sep 17 00:00:00 2001
From: rattus <46076784+rattus128@users.noreply.github.com>
Date: Thu, 5 Mar 2026 10:22:17 -0800
Subject: [PATCH] ops: Handle CPU weight in VBAR caster (#12792)

This shouldn't happen but custom nodes gets there. Handle it as best
we can.
---
 comfy/ops.py | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

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