From 8fd49be7165c7395ce43ce8440b86ada1bda4da8 Mon Sep 17 00:00:00 2001
From: kijai <40791699+kijai@users.noreply.github.com>
Date: Tue, 25 Nov 2025 00:21:28 +0200
Subject: [PATCH] Fix fp8_scaled text encoder

---
 comfy/text_encoders/kandinsky5.py | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/comfy/text_encoders/kandinsky5.py b/comfy/text_encoders/kandinsky5.py
index c0c5341f0..8c25e8b4e 100644
--- a/comfy/text_encoders/kandinsky5.py
+++ b/comfy/text_encoders/kandinsky5.py
@@ -1,5 +1,7 @@
 from comfy import sd1_clip
-from .qwen_image import QwenImageTokenizer, QwenImageTEModel, Qwen25_7BVLIModel
+from .qwen_image import QwenImageTokenizer, QwenImageTEModel
+from .llama import Qwen25_7BVLI
+
 
 class Kandinsky5Tokenizer(QwenImageTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -17,6 +19,15 @@ class Kandinsky5Tokenizer(QwenImageTokenizer):
         return out
 
 
+class Qwen25_7BVLIModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="hidden", layer_idx=-1, dtype=None, attention_mask=True, model_options={}):
+        llama_scaled_fp8 = model_options.get("qwen_scaled_fp8", None)
+        if llama_scaled_fp8 is not None:
+            model_options = model_options.copy()
+            model_options["scaled_fp8"] = llama_scaled_fp8
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+
 class Kandinsky5TEModel(QwenImageTEModel):
     def __init__(self, device="cpu", dtype=None, model_options={}):
         super(QwenImageTEModel, self).__init__(device=device, dtype=dtype, name="qwen25_7b", clip_model=Qwen25_7BVLIModel, model_options=model_options)