From 7d6a98bc626c56b3eb4c07eff5d7aeaf8f781c1d Mon Sep 17 00:00:00 2001
From: li-lizhe <147392333@qq.com>
Date: Sat, 12 Sep 2026 09:08:48 +0800
Subject: [PATCH] fix(framepack): use text_mask.device instead of hardcoded
 cuda

get_cu_seqlens creates a cu_seqlens tensor with device="cuda",
which crashes on non-CUDA devices (e.g. Ascend NPU raises
"Torch not compiled with CUDA enabled"). Create it on the same
device as the input text_mask instead, which is device-agnostic.
---
 modules/framepack/pipeline/hunyuan_video_packed.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/modules/framepack/pipeline/hunyuan_video_packed.py b/modules/framepack/pipeline/hunyuan_video_packed.py
index a5f171b5c..58ae8918a 100644
--- a/modules/framepack/pipeline/hunyuan_video_packed.py
+++ b/modules/framepack/pipeline/hunyuan_video_packed.py
@@ -74,7 +74,7 @@ def get_cu_seqlens(text_mask, img_len):
     text_len = text_mask.sum(dim=1)
     max_len = text_mask.shape[1] + img_len
 
-    cu_seqlens = torch.zeros([2 * batch_size + 1], dtype=torch.int32, device="cuda")
+    cu_seqlens = torch.zeros([2 * batch_size + 1], dtype=torch.int32, device=text_mask.device)
 
     for i in range(batch_size):
         s = text_len[i] + img_len