biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[fix] fix qwen image_embeds input (#21049)

Browse Source 
Signed-off-by: h-avsha <avshalom.manevich@hcompany.ai>
This commit is contained in:
Avshalom Manevich
2025-07-16 17:17:20 +02:00
committed by GitHub
parent 18bdcf4113
commit a0f8a79646
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/models/qwen2_5_vl.py
View File

@@ -974,7 +974,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
grid_thw_list = grid_thw.tolist() grid_thw_list = grid_thw.tolist()
if image_input["type"] == "image_embeds": if image_input["type"] == "image_embeds":
image_embeds = image_input["image_embeds"] image_embeds = image_input["image_embeds"].type(self.visual.dtype)
else: else:
pixel_values = image_input["pixel_values"] pixel_values = image_input["pixel_values"]
image_embeds = self.visual(pixel_values, grid_thw=grid_thw_list) image_embeds = self.visual(pixel_values, grid_thw=grid_thw_list)
@@ -994,7 +994,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
grid_thw_list = grid_thw.tolist() grid_thw_list = grid_thw.tolist()
if video_input["type"] == "video_embeds": if video_input["type"] == "video_embeds":
video_embeds = video_input["video_embeds"] video_embeds = video_input["video_embeds"].type(self.visual.dtype)
else: else:
pixel_values_videos = video_input["pixel_values_videos"] pixel_values_videos = video_input["pixel_values_videos"]
video_embeds = self.visual(pixel_values_videos, video_embeds = self.visual(pixel_values_videos,