{ "schema_version": 1, "labels": [ "backchannel", "barge_in" ], "max_length": 64, "pad_id": 1, "pad_token": "", "model_file": "model_int8.onnx", "calibrated": false, "threshold": 0.5, "quantization": "Dynamic signed int8, per-channel; MatMul/Gemm/Attention. Embeddings remain fp32.", "fused_operator_counts": { "EmbedLayerNormalization": 1, "Attention": 12, "MultiHeadAttention": 0, "Gelu": 0, "FastGelu": 0, "BiasGelu": 12, "GemmFastGelu": 0, "LayerNormalization": 0, "SimplifiedLayerNormalization": 0, "SkipLayerNormalization": 24, "SkipSimplifiedLayerNormalization": 0, "RotaryEmbedding": 0, "QOrderedAttention": 0, "QOrderedGelu": 0, "QOrderedLayerNormalization": 0, "QOrderedMatMul": 0 } }