[Bugfix][MiniCPM-V] Route video_embeds to the shared vision parser (#54633)

Signed-off-by: subhash polisetti <subhashr161347@gmail.com>
This commit is contained in:
Subhash Polisetti
2026-09-01 09:16:07 +00:00
committed by GitHub
parent 5414b4e694
commit 4707679cd2
3 changed files with 75 additions and 8 deletions
+2 -7
View File
@@ -59,6 +59,7 @@ from .minicpmv import (
MiniCPMVMultiModalProcessor,
MiniCPMVProcessingInfo,
MiniCPMVVideoEmbeddingItems,
_image_kwargs_from_video,
)
from .module_mapping import MultiModelKeys
from .qwen3_5 import Qwen3_5ForCausalLM
@@ -1200,18 +1201,12 @@ class MiniCPMV4_6ForConditionalGeneration(
for t in use_vit_merger_tensors
)
# Split kwargs into image / video buckets (videos are processed via
# the same vision pipeline; their fields just carry a ``video_`` prefix).
image_kwargs = {
k: v
for k, v in kwargs.items()
if k in ("pixel_values", "image_embeds", "tgt_sizes")
}
video_kwargs = {
k.removeprefix("video_"): v
for k, v in kwargs.items()
if k.startswith("video_")
}
video_kwargs = _image_kwargs_from_video(kwargs)
multimodal_embeddings: tuple[torch.Tensor, ...] = ()