[EPD] Remove duplicate image preprocessing in EPD and enable preprocess on GPU (#50390)

Signed-off-by: Tianyu Guo <guoty@inferact.ai>
This commit is contained in:
Tianyu Guo
2026-08-05 15:59:54 -07:00
committed by GitHub
parent 38ebd97bca
commit b30291cf0c
22 changed files with 904 additions and 45 deletions
+13 -8
View File
@@ -881,18 +881,23 @@ def _keye_field_config(
class KeyeMultiModalDataParser(MultiModalDataParser):
# The patch grid is what sizes the placeholder range.
embedding_fields = {
"image": {"image_embeds": "values", "image_grid_thw": "metadata"},
"video": {"video_embeds": "values", "video_grid_thw": "metadata"},
}
def _parse_image_data(
self,
data: dict[str, torch.Tensor] | ModalityData[ImageItem],
) -> ModalityDataItems[Any, Any] | None:
if isinstance(data, dict):
required, optional = self.embedding_field_sets("image")
return DictEmbeddingItems(
data,
modality="image",
required_fields={
"image_embeds",
"image_grid_thw",
},
required_fields=required,
optional_fields=optional,
fields_factory=_keye_field_config,
)
@@ -903,13 +908,12 @@ class KeyeMultiModalDataParser(MultiModalDataParser):
data: dict[str, torch.Tensor] | ModalityData[VideoItem],
) -> ModalityDataItems[Any, Any] | None:
if isinstance(data, dict):
required, optional = self.embedding_field_sets("video")
return DictEmbeddingItems(
data,
modality="video",
required_fields={
"video_embeds",
"video_grid_thw",
},
required_fields=required,
optional_fields=optional,
fields_factory=_keye_field_config,
)
@@ -929,6 +933,7 @@ class KeyeProcessingInfo(BaseProcessingInfo):
def get_data_parser(self):
return KeyeMultiModalDataParser(
expected_hidden_size=self._get_expected_hidden_size(),
embeds_from_ec_connector=self.embeds_from_ec_connector,
)
def get_supported_mm_limits(