[MM] Further cleanup _apply_hf_processor_main (#53610)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-09-16 10:21:00 -07:00
committed by GitHub
parent 6ca2b23e22
commit 0b9e018fb9
74 changed files with 1048 additions and 1323 deletions
+37 -33
View File
@@ -1556,23 +1556,26 @@ class LlavaOnevision2MultiModalDataParser(MultiModalDataParser):
class LlavaOnevision2MultiModalProcessor(
BaseMultiModalProcessor[LlavaOnevision2ProcessingInfo]
):
def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str:
return self.dummy_inputs.get_dummy_text(mm_counts)
def _apply_hf_processor_main(
self,
mm_items: MultiModalDataItems,
hf_processor_mm_kwargs: Mapping[str, object],
hf_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}
hf_data, hf_kwargs, passthrough_data = self._get_hf_mm_inputs(
mm_items, hf_kwargs
)
mm_data, passthrough_data = self._get_hf_mm_data(valid_mm_items)
if not mm_data:
return BatchFeature(dict(passthrough_data))
if not hf_data:
return self._finalize_hf_mm_data(hf_data, hf_kwargs, passthrough_data)
prompt_text = self.dummy_inputs.get_dummy_text(mm_items.get_all_counts())
prompt_text = hf_data.pop("text")
assert isinstance(prompt_text, str)
hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
merged_kwargs = self.info.ctx.get_merged_mm_kwargs(hf_processor_mm_kwargs)
hf_processor = self.info.get_hf_processor(**hf_kwargs)
merged_kwargs = self.info.ctx.get_merged_mm_kwargs(hf_kwargs)
merged_kwargs.setdefault("return_tensors", "pt")
call_kwargs = {
k: v
@@ -1589,17 +1592,17 @@ class LlavaOnevision2MultiModalProcessor(
"codec_config",
}
}
mm_data = dict(mm_data)
# Explicit None + length checks: ``mm_data[...]`` may be a list, numpy
# array, or tensor, and ``and <array>`` would raise on the ambiguous
# truth value of a multi-element array.
_videos = mm_data.get("videos")
_videos = hf_data.get("videos")
if _videos is not None and not isinstance(_videos, Sized):
raise TypeError(f"object of type '{type(_videos).__name__}' has no len()")
videos_present = _videos is not None and len(_videos) > 0
codec_video_paths = (
_extract_codec_video_paths(mm_data["videos"]) if videos_present else None
_extract_codec_video_paths(hf_data["videos"]) if videos_present else None
)
is_codec_marker = codec_video_paths is not None
# Fallback: caller passed video_backend=codec via mm_processor_kwargs
@@ -1613,7 +1616,7 @@ class LlavaOnevision2MultiModalProcessor(
and call_kwargs.get("video_backend") == "codec"
)
if is_codec_kwarg:
raw = mm_data["videos"]
raw = hf_data["videos"]
if isinstance(raw, str):
codec_video_paths = [raw]
elif isinstance(raw, (list, tuple)) and all(
@@ -1638,7 +1641,7 @@ class LlavaOnevision2MultiModalProcessor(
# Codec backend: HF processor consumes the path string directly and
# performs decode + canvas-packing internally. The dummy ndarray
# we attached during prepare_codec_video_input is discarded here.
mm_data["videos"] = (
hf_data["videos"] = (
codec_video_paths
if len(codec_video_paths) > 1
else codec_video_paths[0]
@@ -1648,11 +1651,9 @@ class LlavaOnevision2MultiModalProcessor(
# ``video_backend="codec"`` via mm_kwargs so the wrapped processor
# dispatches to its codec branch.
output = self.info.ctx.call_hf_processor(
self.info.get_hf_processor(
**{**hf_processor_mm_kwargs, "video_backend": "codec"}
),
dict(text=prompt_text, **mm_data),
{**hf_processor_mm_kwargs, "video_backend": "codec"},
self.info.get_hf_processor(**{**hf_kwargs, "video_backend": "codec"}),
dict(text=prompt_text, **hf_data),
{**hf_kwargs, "video_backend": "codec"},
)
data = dict(output)
processed_data = BatchFeature(
@@ -1660,8 +1661,9 @@ class LlavaOnevision2MultiModalProcessor(
data, codec_video_paths, hf_processor
)
)
processed_data.update(passthrough_data)
return processed_data
return self._finalize_hf_mm_data(
hf_data, hf_kwargs, passthrough_data, processed_data
)
# ---- Frame backend (registered LlavaOnevision2VideoBackend) ------
# Every non-codec video reaches here as a ``(frames_ndarray, metadata)``
@@ -1676,7 +1678,7 @@ class LlavaOnevision2MultiModalProcessor(
# provided by the backend's ``compute_frames_index_to_sample``; SSRF /
# local-file gating is enforced by the connector before decoding.
if videos_present:
timestamp_decimals = hf_processor_mm_kwargs.get(
timestamp_decimals = hf_kwargs.get(
"timestamp_decimals", _DEFAULT_TIMESTAMP_DECIMALS
)
if not isinstance(
@@ -1690,7 +1692,7 @@ class LlavaOnevision2MultiModalProcessor(
per_video_frames: list[list[Image.Image]] = []
per_video_timestamps: list[list[float]] = []
videos = mm_data["videos"]
videos = hf_data["videos"]
if not isinstance(videos, Iterable):
raise TypeError(f"Unsupported videos input: {type(videos)}")
for item in videos:
@@ -1714,7 +1716,7 @@ class LlavaOnevision2MultiModalProcessor(
# order) -- otherwise mixed image+video requests bind frames to the
# wrong placeholder. ``row_is_video`` labels each grid row so outputs
# can be split back into per-modality keys below.
merged_mm_data = dict(mm_data)
merged_mm_data = dict(hf_data)
existing_images = merged_mm_data.pop("images", None)
caller_images: list[Image.Image] = []
if existing_images:
@@ -1760,9 +1762,9 @@ class LlavaOnevision2MultiModalProcessor(
# **kwargs to the image processor, so passing the full merged kwarg
# set here is a no-op beyond return_tensors/padding.
output = self.info.ctx.call_hf_processor(
self.info.get_hf_processor(**hf_processor_mm_kwargs),
self.info.get_hf_processor(**hf_kwargs),
dict(text=new_prompt, **merged_mm_data),
hf_processor_mm_kwargs,
hf_kwargs,
)
data = dict(output)
@@ -1820,20 +1822,22 @@ class LlavaOnevision2MultiModalProcessor(
)
processed_data = BatchFeature(data)
processed_data.update(passthrough_data)
return processed_data
return self._finalize_hf_mm_data(
hf_data, hf_kwargs, passthrough_data, processed_data
)
# ---- Image-only / text-only call --------------------------------
# No videos present: delegate to ``ctx.call_hf_processor``, which
# runs the wrapped processor over the (possibly empty) image set and
# applies float-tensor dtype postprocessing automatically.
processed_data = self.info.ctx.call_hf_processor(
self.info.get_hf_processor(**hf_processor_mm_kwargs),
dict(text=prompt_text, **mm_data),
hf_processor_mm_kwargs,
self.info.get_hf_processor(**hf_kwargs),
dict(text=prompt_text, **hf_data),
hf_kwargs,
)
return self._finalize_hf_mm_data(
hf_data, hf_kwargs, passthrough_data, processed_data
)
processed_data.update(passthrough_data)
return processed_data
def _rename_codec_outputs_to_video(
self,