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

[Model][QwenVL] Replace torch.repeat_interleave with faster np.repeat (#28964)

Browse Source 
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
This commit is contained in:
Lukas Geiger
2025-11-20 06:04:23 +00:00
committed by GitHub
parent 64192d5624
commit a9705a290a
3 changed files with 18 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
tests/models/multimodal/generation/test_qwen2_vl.py
View File

@@ -128,12 +128,7 @@ def batch_make_image_embeddings(
visual = model.visual
pixel_values_on_device = pixel_values.to(visual.device, dtype=visual.dtype)
image_grid_thw_on_device = image_grid_thw.to(
visual.device, dtype=torch.int64
)
return visual(
pixel_values_on_device, grid_thw=image_grid_thw_on_device
).cpu()
return visual(pixel_values_on_device, grid_thw=image_grid_thw).cpu()
image_embeds = torch.concat(llm.apply_model(get_image_embeds))
@@ -217,12 +212,7 @@ def batch_make_video_embeddings(
visual = model.visual
pixel_values_on_device = pixel_values.to(visual.device, dtype=visual.dtype)
video_grid_thw_on_device = video_grid_thw.to(
visual.device, dtype=torch.int64
)
return visual(
pixel_values_on_device, grid_thw=video_grid_thw_on_device
).cpu()
return visual(pixel_values_on_device, grid_thw=video_grid_thw).cpu()
video_embeds = torch.concat(llm.apply_model(get_image_embeds))