{ "_name_or_path": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "architectures": [ "Qwen2ForCausalLM" ], "attention_dropout": 0.0, "bos_token_id": 151643, "eos_token_id": 151643, "hidden_act": "silu", "hidden_size": 5120, "initializer_range": 0.02, "intermediate_size": 27648, "max_position_embeddings": 131072, "max_window_layers": 64, "model_type": "qwen2", "num_attention_heads": 40, "num_hidden_layers": 64, "num_key_value_heads": 8, "quantization_config": { "bits": 4, "group_size": 256, "hadamard_size": 512, "modules_to_not_convert": [ "lm_head" ], "p": 2, "quant_method": "higgs", "tune_metadata": { "model.layers.0.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.0.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.0.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.0.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.0.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.0.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.0.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.1.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.1.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.1.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.1.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.1.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.1.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.1.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.10.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.10.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.10.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.10.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.10.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.10.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.10.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.11.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.11.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.11.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.11.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.11.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.11.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.11.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.12.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.12.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.12.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.12.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.12.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.12.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.12.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.13.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.13.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.13.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.13.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.13.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.13.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.13.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.14.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.14.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.14.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.14.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.14.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.14.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.14.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.15.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.15.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.15.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.15.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.15.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.15.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.15.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.16.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.16.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.16.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.16.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.16.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.16.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.16.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.17.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.17.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.17.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.17.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.17.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.17.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.17.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.18.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.18.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.18.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.18.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.18.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.18.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.18.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.19.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.19.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.19.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.19.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.19.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.19.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.19.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.2.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.2.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.2.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.2.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.2.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.2.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.2.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.20.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.20.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.20.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.20.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.20.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.20.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.20.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.21.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.21.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.21.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.21.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.21.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.21.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.21.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.22.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.22.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.22.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.22.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.22.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.22.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.22.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.23.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.23.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.23.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.23.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.23.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.23.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.23.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.24.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.24.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.24.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.24.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.24.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.24.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.24.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.25.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.25.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.25.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.25.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.25.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.25.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.25.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.26.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.26.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.26.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.26.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.26.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.26.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.26.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.27.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.27.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.27.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.27.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.27.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.27.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.27.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.28.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.28.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.28.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.28.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.28.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.28.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.28.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.29.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.29.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.29.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.29.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.29.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.29.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.29.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.3.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.3.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.3.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.3.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.3.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.3.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.3.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.30.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.30.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.30.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.30.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.30.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.30.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.30.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.31.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.31.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.31.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.31.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.31.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.31.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.31.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.32.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.32.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.32.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.32.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.32.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.32.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.32.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.33.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.33.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.33.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.33.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.33.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.33.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.33.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.34.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.34.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.34.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.34.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.34.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.34.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.34.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.35.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.35.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.35.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.35.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.35.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.35.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.35.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.36.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.36.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.36.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.36.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.36.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.36.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.36.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.37.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.37.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.37.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.37.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.37.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.37.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.37.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.38.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.38.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.38.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.38.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.38.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.38.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.38.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.39.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.39.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.39.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.39.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.39.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.39.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.39.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.4.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.4.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.4.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.4.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.4.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.4.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.4.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.40.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.40.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.40.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.40.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.40.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.40.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.40.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.41.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.41.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.41.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.41.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.41.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.41.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.41.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.42.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.42.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.42.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.42.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.42.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.42.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.42.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.43.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.43.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.43.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.43.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.43.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.43.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.43.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.44.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.44.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.44.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.44.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.44.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.44.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.44.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.45.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.45.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.45.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.45.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.45.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.45.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.45.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.46.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.46.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.46.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.46.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.46.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.46.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.46.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:2", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.47.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.47.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.47.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.47.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.47.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.47.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.47.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.48.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.48.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.48.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.48.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.48.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.48.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.48.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.49.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.49.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.49.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.49.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.49.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.49.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.49.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.5.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.5.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.5.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.5.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.5.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.5.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.5.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.50.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.50.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.50.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.50.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.50.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.50.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.50.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.51.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.51.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.51.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.51.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.51.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.51.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.51.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.52.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.52.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.52.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.52.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.52.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.52.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.52.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.53.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.53.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.53.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.53.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.53.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.53.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.53.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.54.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.54.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.54.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.54.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.54.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.54.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.54.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.55.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.55.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.55.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.55.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.55.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.55.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.55.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.56.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.56.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.56.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.56.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.56.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.56.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.56.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.57.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.57.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.57.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.57.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.57.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.57.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.57.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.58.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.58.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.58.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.58.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.58.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.58.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.58.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.59.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.59.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.59.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.59.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.59.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.59.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.59.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.6.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.6.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.6.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.6.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.6.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.6.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.6.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:0", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.60.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.60.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.60.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.60.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.60.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.60.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.60.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.61.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.61.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.61.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.61.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.61.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.61.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.61.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.62.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.62.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.62.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.62.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.62.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.62.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.62.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.63.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.63.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.63.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.63.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.63.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.63.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.63.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:3", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.7.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.7.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.7.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.7.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.7.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.7.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.7.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.8.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.8.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.8.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.8.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.8.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.8.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.8.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.9.mlp.down_proj": { "K": 27648, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 95 }, "model.layers.9.mlp.gate_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.9.mlp.up_proj": { "K": 5120, "M": 1, "N": 27648, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 45 }, "model.layers.9.self_attn.k_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 }, "model.layers.9.self_attn.o_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.9.self_attn.q_proj": { "K": 5120, "M": 1, "N": 5120, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 90 }, "model.layers.9.self_attn.v_proj": { "K": 5120, "M": 1, "N": 1024, "device": "cuda:1", "dtype": "torch.float16", "group_size": 256, "num_bits": 4, "num_sms": 128, "template_id": 42 } } }, "rms_norm_eps": 1e-05, "rope_scaling": null, "rope_theta": 1000000.0, "sliding_window": null, "tie_word_embeddings": false, "torch_dtype": "float16", "transformers_version": "4.49.0.dev0", "use_cache": true, "use_sliding_window": false, "vocab_size": 152064 }