fix qwen2_vl flash_attn deepspeed (#3484)

Author: Jintao-Huang

README.md

@@ -26,7 +26,7 @@
 </p>
 
 <p align="center">
-        <a href="https://arxiv.org/abs/2408.05517">Paper</a> &nbsp ｜ <a href="https://swift.readthedocs.io/en/latest/">English Documentation</a> &nbsp ｜ &nbsp <a href="https://swift.readthedocs.io/zh-cn/latest/">中文文档</a> &nbsp
+        <a href="https://arxiv.org/abs/2408.05517">Paper</a> &nbsp ｜ <a href="https://swift.readthedocs.io/en/latest/">Swift3.x En Doc</a> &nbsp ｜ &nbsp <a href="https://swift.readthedocs.io/zh-cn/latest/">Swift3.x中文文档</a> &nbsp
 </p>
 <p align="center">
         <a href="https://swift2x-en.readthedocs.io/en/latest/">Swift2.x En Doc</a> &nbsp ｜ &nbsp <a href="https://swift2x.readthedocs.io/zh-cn/latest/">Swift2.x中文文档</a> &nbsp

README_CN.md

@@ -27,7 +27,7 @@
 </p>
 
 <p align="center">
-        <a href="https://arxiv.org/abs/2408.05517">论文</a> &nbsp ｜ <a href="https://swift.readthedocs.io/en/latest/">English Documentation</a> &nbsp ｜ &nbsp <a href="https://swift.readthedocs.io/zh-cn/latest/">中文文档</a> &nbsp
+        <a href="https://arxiv.org/abs/2408.05517">论文</a> &nbsp ｜ <a href="https://swift.readthedocs.io/en/latest/">Swift3.x En Doc</a> &nbsp ｜ &nbsp <a href="https://swift.readthedocs.io/zh-cn/latest/">Swift3.x中文文档</a> &nbsp
 </p>
 <p align="center">
         <a href="https://swift2x-en.readthedocs.io/en/latest/">Swift2.x En Doc</a> &nbsp ｜ &nbsp <a href="https://swift2x.readthedocs.io/zh-cn/latest/">Swift2.x中文文档</a> &nbsp

swift/llm/template/template/qwen.py

@@ -6,6 +6,7 @@
 import torch
 import torch.nn.functional as F
 
+from swift.llm import to_device
 from swift.utils import get_env_args, is_deepspeed_enabled
 from ..base import Template
 from ..constant import LLMTemplateType, MLLMTemplateType
@@ -284,9 +285,8 @@ def _post_encode(self, model, inputs: Dict[str, Any]) -> Dict[str, Any]:
                 images = [Image.new('RGB', (32, 32), (0, 0, 0))]
                 media_inputs = self.processor.image_processor(images=images, videos=None, return_tensors='pt')
                 device = input_ids.device
-                pixel_values = media_inputs['pixel_values'].to(device)
-
-                pixel_values = pixel_values.type(dtype)
+                media_inputs = to_device(media_inputs, device)
+                pixel_values = media_inputs['pixel_values'].type(dtype)
                 image_embeds = model.visual(pixel_values, grid_thw=media_inputs['image_grid_thw'])
                 inputs_embeds += image_embeds.mean() * 0.
         else: