Target A100 40GB runtime (AWQ Int4 thinker, no offload)

AWQ Int4 + Thinker-only load to fit L4 24GB

Upload folder using huggingface_hub

L4 24GB: max_memory CPU offload + sdpa default for model load

Fix UnboundLocalError in attn_candidates dedup (transformers_qwen.py)

Upload OmniSub source