McClain/PlasmidLM

Add generate_compiled: CUDA graph + static KV cache for ~2.75x faster inference

Upload tokenization_plasmid_lm.py with huggingface_hub

Upload folder using huggingface_hub

Fix RoPE buffers zeroed by fast-init: recompute in _init_weights

Fix tokenizer: remove special token ID property overrides for transformers>=5.x compat

Fix _tied_weights_keys for transformers>=4.58 compatibility (dict, not list)

Add tokenizer_config.json for AutoTokenizer compatibility

Upload PlasmidLM pretrained checkpoint (v4, step 15000)

initial commit