{ "architectures": [ "IBMLModel" ], "base_url": "TinyLlama/TinyLlama_v1.1", "dtype": "float32", "ibml_num_heads": 4, "ibml_q1_size": 512, "ibml_q2_size": 16, "ibml_value_size": 128, "model_type": "ibml", "transformers_version": "4.57.0" }