[Perf] Avoid more GPU<->CPU syncs on the model execution path (#51738)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-08-11 19:30:34 -07:00
committed by GitHub
parent 23f25aac14
commit f97e502969
15 changed files with 116 additions and 56 deletions
+3 -2
View File
@@ -266,9 +266,10 @@ class Phi3HDImageEmbedding(nn.Module):
)
batch_image_features_proj = []
image_sizes_list = image_sizes.tolist()
# need a for loop to process each image because of different image sizes
# (patch arrangement is different for each image)
for i, img_size in enumerate(image_sizes):
for i, img_size in enumerate(image_sizes_list):
h, w = img_size
h_crop = h // 336
w_crop = w // 336
@@ -426,7 +427,7 @@ class Phi3VMultiModalProcessor(BaseMultiModalProcessor[Phi3VProcessingInfo]):
) -> Mapping[str, MultiModalFieldConfig]:
return dict(
pixel_values=MultiModalFieldConfig.batched("image"),
image_sizes=MultiModalFieldConfig.batched("image"),
image_sizes=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
image_embeds=MultiModalFieldConfig.batched("image"),
)