| {"layers.0.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.0.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.0.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.0.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.0.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.0.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.0.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.1.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.1.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.1.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.1.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.1.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.1.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.1.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.2.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.2.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.2.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.2.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.2.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.2.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.2.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.3.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.3.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.3.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.3.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.3.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.3.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.3.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.4.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.4.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.4.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.4.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.4.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.4.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.4.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.5.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.5.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.5.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.5.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.5.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.5.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.5.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.6.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.6.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.6.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.6.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.6.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.6.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.6.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.7.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.7.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.7.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.7.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.7.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.7.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.7.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.8.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.8.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.8.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.8.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.8.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.8.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.8.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.9.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.9.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.9.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.9.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.9.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.9.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.9.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.10.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.10.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.10.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.10.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.10.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.10.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.10.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.11.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.11.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.11.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.11.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.11.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.11.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.11.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.12.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.12.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.12.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.12.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.12.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.12.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.12.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.13.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.13.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.13.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.13.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.13.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.13.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.13.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.14.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.14.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.14.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.14.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.14.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.14.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.14.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.15.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.15.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.15.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.15.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.15.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.15.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.15.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.16.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.16.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.16.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.16.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.16.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.16.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.16.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.17.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.17.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.17.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.17.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.17.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.17.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.17.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.18.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.18.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.18.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.18.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.18.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.18.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.18.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.19.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.19.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.19.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.19.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.19.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.19.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.19.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.20.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.20.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.20.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.20.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.20.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.20.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.20.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.21.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.21.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.21.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.21.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.21.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.21.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.21.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.22.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.22.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.22.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.22.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.22.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.22.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.22.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.23.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.23.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.23.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.23.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.23.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.23.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.23.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.24.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.24.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.24.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.24.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.24.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.24.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.24.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.25.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.25.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.25.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.25.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.25.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.25.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.25.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.26.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.26.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.26.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.26.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.26.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.26.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.26.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.27.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.27.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.27.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.27.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.27.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.27.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.27.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.28.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.28.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.28.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.28.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.28.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.28.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.28.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.29.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.29.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.29.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.29.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.29.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.29.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.29.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.30.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.30.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.30.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.30.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.30.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.30.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.30.mlp.down_proj": {"weights": "qint8", "activations": "none"}, "layers.31.self_attn.q_proj": {"weights": "qint8", "activations": "none"}, "layers.31.self_attn.k_proj": {"weights": "qint8", "activations": "none"}, "layers.31.self_attn.v_proj": {"weights": "qint8", "activations": "none"}, "layers.31.self_attn.o_proj": {"weights": "qint8", "activations": "none"}, "layers.31.mlp.gate_proj": {"weights": "qint8", "activations": "none"}, "layers.31.mlp.up_proj": {"weights": "qint8", "activations": "none"}, "layers.31.mlp.down_proj": {"weights": "qint8", "activations": "none"}} |