From eb6597beba0b74c0c886d45eda4cdde67bd2d1bd Mon Sep 17 00:00:00 2001
From: VjumpKunG <vjumpkung@gmail.com>
Date: Fri, 26 Dec 2025 14:56:04 +0700
Subject: [PATCH] fixed open-clip-torch version

---
 requirements.txt                   | 2 +-
 video_to_video/modules/embedder.py | 5 ++---
 2 files changed, 3 insertions(+), 4 deletions(-)

diff --git a/requirements.txt b/requirements.txt
index ffcfb02..3786be3 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -3,7 +3,7 @@ av
 opencv-python
 easydict
 einops
-open-clip-torch
+open-clip-torch==2.20.0
 fairscale
 torchsde
 pytorch-lightning
diff --git a/video_to_video/modules/embedder.py b/video_to_video/modules/embedder.py
index da065a2..68431e0 100644
--- a/video_to_video/modules/embedder.py
+++ b/video_to_video/modules/embedder.py
@@ -1,6 +1,7 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
 
 import os
+from functools import wraps
 
 import numpy as np
 import open_clip
@@ -60,9 +61,7 @@ class FrozenOpenCLIPEmbedder(nn.Module):
         x = self.model.token_embedding(text)
         x = x + self.model.positional_embedding
         x = x.permute(1, 0, 2)
-        # Fix for PyTorch compatibility: use None for attn_mask to avoid shape mismatch errors
-        # The causal masking will be handled internally by the attention mechanism
-        x = self.text_transformer_forward(x, attn_mask=None)
+        x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
         x = x.permute(1, 0, 2)
         x = self.model.ln_final(x)
         return x