A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 01:00 UTC
[fix] fix qwen image_embeds input (#21049)
Signed-off-by: h-avsha <avshalom.manevich@hcompany.ai>
Avshalom Manevich committed
Jul 16, 2025 at 17:17 UTC
a0f8a7964694a6077689b242b5eca95de392d4bb
1 file changed
+2
-2
vllm/model_executor/models/qwen2_5_vl.py
+2
-2
@@ -974,7 +974,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
974
grid_thw_list = grid_thw.tolist()
975
976
if image_input["type"] == "image_embeds":
977
- image_embeds = image_input["image_embeds"]
977
+ image_embeds = image_input["image_embeds"].type(self.visual.dtype)
978
else:
979
pixel_values = image_input["pixel_values"]
980
image_embeds = self.visual(pixel_values, grid_thw=grid_thw_list)
@@ -994,7 +994,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
994
grid_thw_list = grid_thw.tolist()
995
996
if video_input["type"] == "video_embeds":
997
- video_embeds = video_input["video_embeds"]
997
+ video_embeds = video_input["video_embeds"].type(self.visual.dtype)
998
else:
999
pixel_values_videos = video_input["pixel_values_videos"]
1000
video_embeds = self.visual(pixel_values_videos,