[Task] zero-shot-image-classification EP coverage investigation

Overview

Zero-shot image classification models (CLIP, SigLIP) encode images and text labels into a shared embedding space, enabling classification of arbitrary categories without task-specific fine-tuning. All 10 models in this task already pass on at least one EP.

This issue focuses on the 7 models that currently have incomplete EP coverage — primarily a VitisAI failure pattern shared across the CLIP family, plus one OV failure. The goal is to identify the root cause (graph compilation, op support, QDQ compatibility) and unblock the failing EPs.

Agent Scenarios

• Open-vocabulary visual search agent: classify images against a dynamic set of user-defined text labels without retraining, enabling flexible product tagging or content moderation
• Fashion / retail agent: match product images to free-form style descriptions (fashion-clip, marqo-fashionSigLIP) for recommendation and search
• Zero-shot content filtering agent: apply new content policy labels to images without model retraining, running on-device for privacy

EP Coverage Status

Reference: google/siglip-* and Marqo/marqo-fashionSigLIP already pass all 3 EPs.

Acceptance Criteria

• openai/clip-vit-base-patch32 passes all 3 EPs (investigate VitisAI failure)
• openai/clip-vit-large-patch14 passes all 3 EPs (investigate VitisAI failure)
• openai/clip-vit-large-patch14-336 passes all 3 EPs (investigate VitisAI failure)
• openai/clip-vit-base-patch16 passes all 3 EPs (investigate VitisAI failure)
• laion/CLIP-ViT-B-32-laion2B-s34B-b79K passes all 3 EPs (investigate VitisAI failure)
• patrickjohncyh/fashion-clip passes all 3 EPs (investigate VitisAI failure)
• laion/CLIP-ViT-H-14-laion2B-s32B-b79K passes all 3 EPs (investigate OV failure)