Update models/initializer.py

Update models/initializer.py

Update models/initializer.py

Update models/engines/inference_engine.py

Update models/engines/inference_engine.py

Update models/engines/inference_engine.py

Update models/engines/inference_engine.py

Update models/engines/inference_engine.py

Update models/engines/inference_engine.py

Update models/engines/inference_engine.py

Update models/engines/inference_engine.py

Update models/engines/inference_engine.py

Update models/engines/inference_engine.py

Update models/llm_model.py

Update models/engines/inference_engine.py

Update models/engines/inference_engine.py

Update models/llm_model.py

Update models/engines/inference_engine.py

add decoding interface

fix: Modify code to ensure the input dtype/shape/key is what the ONNX model expects

fix: Modify code to ensure the input dtype/shape/key is what the ONNX model expects

fix: Modify code to ensure the input dtype/shape/key is what the ONNX model expects

fix: Modify code to ensure the input dtype/shape/key is what the ONNX model expects

fix(llm_model): align token chunking and prefix handling with engine

feat(initializer): store prefix cache as input_ids tensors

feat(inference engine): add input normalization and attention_mask support

update import block

refactor: move LLM model initialization from global scope to function-level for lazy loading

Edit : add model load method

Edit : add model load method

update initializer.py

fix: use weights_only=True in torch.load to safely load state_dict

Refactor model initialization to use hf_hub_download cache paths

Refactor load_llm: use AutoConfig and direct state_dict loading

Update initializer.py to use explicit Hub filenames

First codes update