From 123a7874a97c4a8b8f06d4b7c2b1a566b8f0d057 Mon Sep 17 00:00:00 2001
From: rattus <46076784+rattus128@users.noreply.github.com>
Date: Tue, 10 Feb 2026 10:38:28 -0800
Subject: [PATCH] ops: Fix vanilla-fp8 loaded lora quality (#12390)

This was missing the stochastic rounding required for fp8 downcast
to be consistent with model_patcher.patch_weight_to_device.

Missed in testing as I spend too much time with quantized tensors
and overlooked the simpler ones.
---
 comfy/ops.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/comfy/ops.py b/comfy/ops.py
index ea0d70702..33803b223 100644
--- a/comfy/ops.py
+++ b/comfy/ops.py
@@ -169,8 +169,8 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
                 if orig.dtype == dtype and len(fns) == 0:
                     #The layer actually wants our freshly saved QT
                     x = y
-            else:
-                y = x
+            elif update_weight:
+                y = comfy.float.stochastic_rounding(x, orig.dtype, seed = comfy.utils.string_to_seed(s.seed_key))
             if update_weight:
                 orig.copy_(y)
         for f in fns: