admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-05-24 04:59:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: route video_metadata through videos_kwargs for Qwen3/3.5 processors

Browse Source 
The Qwen3VLProcessor distributes kwargs to sub-processors via
_merge_kwargs. Flat kwargs like video_metadata and do_sample_frames
were not reaching the video processor, causing fps to default to 24
and producing shape mismatches.

Pass these kwargs explicitly under videos_kwargs so they reach
Qwen3VLVideoProcessor directly. Revert Qwen2VL to its simpler
original approach since its processor doesn't use videos_kwargs.

Made-with: Cursor
This commit is contained in:
Pepijn
2026-03-30 19:09:46 +02:00
parent 5f85b572d7
commit 2545f1a8ed
1 changed files with 22 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/lerobot/data_processing/data_annotations/vlm_annotations.py
+22 -20
View File
@@ -170,14 +170,11 @@ class Qwen2VL(BaseVLM):
] ]
text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = self.process_vision_info(messages, return_video_metadata=True) image_inputs, video_inputs = self.process_vision_info(messages)
videos, video_metadata = _unpack_video_inputs(video_inputs)
inputs = self.processor( inputs = self.processor(
text=[text], text=[text],
images=image_inputs, images=image_inputs,
videos=videos, videos=video_inputs,
video_metadata=video_metadata,
do_sample_frames=False,
padding=True, padding=True,
return_tensors="pt", return_tensors="pt",
).to(self.device) ).to(self.device)
@@ -222,20 +219,17 @@ class Qwen2VL(BaseVLM):
all_messages.append(messages) all_messages.append(messages)
all_texts = [] all_texts = []
all_video_tuples = [] all_video_inputs = []
for messages in all_messages: for messages in all_messages:
text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = self.process_vision_info(messages, return_video_metadata=True) image_inputs, video_inputs = self.process_vision_info(messages)
all_texts.append(text) all_texts.append(text)
all_video_tuples.extend(video_inputs or []) all_video_inputs.extend(video_inputs or [])
videos, video_metadata = _unpack_video_inputs(all_video_tuples or None)
inputs = self.processor( inputs = self.processor(
text=all_texts, text=all_texts,
videos=videos, videos=all_video_inputs or None,
video_metadata=video_metadata,
do_sample_frames=False,
padding=True, padding=True,
return_tensors="pt", return_tensors="pt",
).to(self.device) ).to(self.device)
@@ -350,8 +344,10 @@ class Qwen3VL(BaseVLM):
text=[text], text=[text],
images=image_inputs, images=image_inputs,
videos=videos, videos=videos,
video_metadata=video_metadata, videos_kwargs={
do_sample_frames=False, "video_metadata": video_metadata,
"do_sample_frames": False,
},
padding=True, padding=True,
return_tensors="pt", return_tensors="pt",
).to(self.device) ).to(self.device)
@@ -407,8 +403,10 @@ class Qwen3VL(BaseVLM):
inputs = self.processor( inputs = self.processor(
text=all_texts, text=all_texts,
videos=videos, videos=videos,
video_metadata=video_metadata, videos_kwargs={
do_sample_frames=False, "video_metadata": video_metadata,
"do_sample_frames": False,
},
padding=True, padding=True,
return_tensors="pt", return_tensors="pt",
).to(self.device) ).to(self.device)
@@ -517,8 +515,10 @@ class Qwen35VL(BaseVLM):
text=[text], text=[text],
images=image_inputs, images=image_inputs,
videos=videos, videos=videos,
video_metadata=video_metadata, videos_kwargs={
do_sample_frames=False, "video_metadata": video_metadata,
"do_sample_frames": False,
},
padding=True, padding=True,
return_tensors="pt", return_tensors="pt",
).to(self.device) ).to(self.device)
@@ -575,8 +575,10 @@ class Qwen35VL(BaseVLM):
inputs = self.processor( inputs = self.processor(
text=all_texts, text=all_texts,
videos=videos, videos=videos,
video_metadata=video_metadata, videos_kwargs={
do_sample_frames=False, "video_metadata": video_metadata,
"do_sample_frames": False,
},
padding=True, padding=True,
return_tensors="pt", return_tensors="pt",
).to(self.device) ).to(self.device)