[MM] Simplify _apply_hf_processor_main (#53275)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-08-22 01:49:37 -07:00
committed by GitHub
parent 7ca49fbe4b
commit 0b19ebcacd
90 changed files with 1848 additions and 1706 deletions
@@ -190,77 +190,71 @@ class HCXVisionDummyInputsBuilder(BaseDummyInputsBuilder[HCXVisionProcessingInfo
class HCXVisionMultiModalProcessor(BaseMultiModalProcessor[HCXVisionProcessingInfo]):
def _call_hf_processor(
def _apply_hf_processor_main(
self,
prompt: str,
mm_data: Mapping[str, object],
mm_kwargs: Mapping[str, object],
mm_items: MultiModalDataItems,
hf_processor_mm_kwargs: Mapping[str, object],
) -> BatchFeature:
valid_mm_items = mm_items.select(
{k for k, c in mm_items.get_all_counts().items() if c > 0}
)
mm_data, passthrough_data = self._get_hf_mm_data(valid_mm_items)
if not mm_data:
return BatchFeature(dict(passthrough_data))
for video_idx, video_arr in enumerate(mm_data.get("videos", [])):
if video_arr.dtype != np.uint8:
mm_data["videos"][video_idx] = video_arr.astype(np.uint8)
processed_outputs = self.info.ctx.call_hf_processor(
hf_processor=self.info.get_hf_processor(**mm_kwargs),
images = mm_data.get("images")
videos = mm_data.get("videos")
# batchify input as a single item
processed_data = self.info.ctx.call_hf_processor(
hf_processor=self.info.get_hf_processor(**hf_processor_mm_kwargs),
data=dict(
text=prompt,
images=None,
videos=None,
images=None if images is None else [images],
videos=None if videos is None else [videos],
),
) # text-only
)
if len(mm_data) > 0:
images = mm_data.get("images")
videos = mm_data.get("videos")
for k, v in processed_data.items():
if isinstance(v, list) and len(v) > 0:
assert len(v) == 1
processed_data[k] = v[0]
# batchify input as a single item
_processed_outputs = self.info.ctx.call_hf_processor(
hf_processor=self.info.get_hf_processor(**mm_kwargs),
data=dict(
text=None,
images=None if images is None else [images],
videos=None if videos is None else [videos],
),
) # mm-only
if images:
processed_data["image_sizes_images"] = torch.tensor(
processed_data["image_sizes_images"]
)
processed_data["vision_query_lengths_images"] = torch.tensor(
processed_data["vision_query_lengths_images"]
)
for k, v in _processed_outputs.items():
if isinstance(v, list) and len(v) > 0:
assert len(v) == 1
_processed_outputs[k] = v[0]
if images:
_processed_outputs["image_sizes_images"] = torch.tensor(
_processed_outputs["image_sizes_images"]
)
_processed_outputs["vision_query_lengths_images"] = torch.tensor(
_processed_outputs["vision_query_lengths_images"]
)
if videos:
_idx_per_video = [
0,
*accumulate(
get_num_combined_frames(len(video)) for video in videos
),
if videos:
_idx_per_video = [
0,
*accumulate(get_num_combined_frames(len(video)) for video in videos),
]
processed_data["pixel_values_videos"] = [
processed_data["pixel_values_videos"][
_idx_per_video[i] : _idx_per_video[i + 1]
]
_processed_outputs["pixel_values_videos"] = [
_processed_outputs["pixel_values_videos"][
for i in range(len(videos))
]
processed_data["vision_query_lengths_videos"] = [
torch.tensor(
processed_data["vision_query_lengths_videos"][
_idx_per_video[i] : _idx_per_video[i + 1]
]
for i in range(len(videos))
]
_processed_outputs["vision_query_lengths_videos"] = [
torch.tensor(
_processed_outputs["vision_query_lengths_videos"][
_idx_per_video[i] : _idx_per_video[i + 1]
]
)
for i in range(len(videos))
]
)
for i in range(len(videos))
]
processed_outputs.update(_processed_outputs)
processed_data.update(passthrough_data)
return processed_outputs
return processed_data
def _get_prompt_updates(
self,