Fix GLM-4.1V video placeholder token ID handling. (#49484)

Signed-off-by: aarushjain29 <Aarushi.Jain2@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Aarushi Jain
2026-07-24 12:35:16 -07:00
committed by GitHub
co-authored by Andreas Karatzas
parent 7e51939e25
commit c064fa52b6
2 changed files with 24 additions and 43 deletions
@@ -93,7 +93,6 @@ def test_processor_override(
limit_mm_per_prompt={"video": 1},
)
processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
tokenizer = processor.info.get_tokenizer()
hf_processor_mm_kwargs = {"fps": fps}
# Build the image str / prompt based on the number of images we pass
@@ -112,8 +111,8 @@ def test_processor_override(
# Ensure we have the right number of placeholders per num_crops size
hf_processor = processor.info.get_hf_processor(**hf_processor_mm_kwargs)
video_token_id = tokenizer.convert_tokens_to_ids(hf_processor.video_token)
video_tok_count = processed_inputs["prompt_token_ids"].count(video_token_id)
image_token_id = hf_processor.image_token_id
video_tok_count = processed_inputs["prompt_token_ids"].count(image_token_id)
grid_t, _, _ = processed_inputs["mm_kwargs"].get_data()["video_grid_thw"][0]
assert grid_t == expected_grid_t