Using QwenVL for multi-image inference, I'm facing a critical issue: CPU memory rises gradually with each inference until OOM. gc.collect() doesn't help. With tracemalloc, the top memory hog is in QwenVL's preprocessing (from transformers 4.51.0): pixel_values = np.array(pixel_values) # https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py#L226 This line converts a list of image patches (collected via extend() for multiple images) into a numpy array. Memory isn't released post-inference, causing accumulation. Using vllm 0.8.3. Any ideas why this persists or fixes? Continue reading...
