| { | |
| "bits": 4, | |
| "group_size": 128, | |
| "desc_act": true, | |
| "sym": true, | |
| "lm_head": false, | |
| "quant_method": "gptq", | |
| "checkpoint_format": "gptq", | |
| "pack_dtype": "int32", | |
| "meta": { | |
| "layer_config": { | |
| "model.layers.0.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.0.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.0.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.0.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.0.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.0.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.0.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.1.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.1.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.1.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.1.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.1.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.1.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.1.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.2.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.2.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.2.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.2.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.2.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.2.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.2.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.3.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.3.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.3.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.3.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.3.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.3.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.3.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.4.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.4.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.4.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.4.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.4.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.4.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.4.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.5.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.5.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.5.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.5.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.5.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.5.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.5.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.6.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.6.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.6.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.6.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.6.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.6.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.6.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.7.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.7.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.7.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.7.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.7.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.7.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.7.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.8.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.8.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.8.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.8.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.8.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.8.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.8.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.9.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.9.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.9.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.9.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.9.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.9.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.9.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.10.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.10.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.10.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.10.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.10.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.10.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.10.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.11.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.11.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.11.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.11.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.11.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.11.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.11.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.12.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.12.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.12.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.12.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.12.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.12.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.12.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.13.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.13.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.13.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.13.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.13.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.13.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.13.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.14.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.14.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.14.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.14.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.14.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.14.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.14.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.15.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.15.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.15.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.15.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.15.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.15.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.15.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.16.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.16.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.16.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.16.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.16.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.16.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.16.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.17.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.17.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.17.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.17.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.17.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.17.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.17.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.18.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.18.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.18.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.18.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.18.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.18.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.18.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.19.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.19.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.19.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.19.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.19.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.19.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.19.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.20.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.20.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.20.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.20.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.20.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.20.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.20.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.21.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.21.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.21.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.21.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.21.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.21.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.21.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.22.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.22.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.22.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.22.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.22.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.22.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.22.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.23.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.23.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.23.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.23.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.23.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.23.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.23.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.24.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.24.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.24.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.24.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.24.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.24.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.24.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.25.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.25.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.25.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.25.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.25.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.25.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.25.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.26.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.26.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.26.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.26.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.26.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.26.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.26.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.27.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.27.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.27.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.27.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.27.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.27.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.27.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.28.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.28.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.28.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.28.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.28.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.28.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.28.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.29.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.29.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.29.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.29.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.29.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.29.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.29.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.30.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.30.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.30.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.30.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.30.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.30.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.30.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.31.self_attn.q_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.31.self_attn.k_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.31.self_attn.v_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.31.self_attn.o_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.31.mlp.gate_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.31.mlp.up_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "model.layers.31.mlp.down_proj": { | |
| "data_type": "int", | |
| "bits": 4, | |
| "group_size": 128, | |
| "sym": true, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| }, | |
| "lm_head": { | |
| "data_type": "bfloat", | |
| "bits": 16, | |
| "group_size": null, | |
| "sym": null, | |
| "scale_dtype": "bfloat16", | |
| "act_bits": 16, | |
| "act_group_size": 128, | |
| "act_sym": true, | |
| "act_dynamic": true, | |
| "act_data_type": "int" | |
| } | |
| }, | |
| "batch_size": 8, | |
| "seqlen": 2048, | |
| "nsamples": 512, | |
| "amp": true, | |
| "lr_scheduler": null, | |
| "enable_quanted_input": true, | |
| "enable_minmax_tuning": true, | |
| "lr": 0.005, | |
| "minmax_lr": null, | |
| "low_gpu_mem_usage": true, | |
| "low_cache_mem_usage": true, | |
| "iters": 1000, | |
| "sampler": "rand", | |
| "seed": 42, | |
| "nblocks": 1, | |
| "gradient_accumulate_steps": 1, | |
| "not_use_best_mse": false, | |
| "dynamic_max_gap": -1, | |
| "data_type": "int", | |
| "scale_dtype": "bf16", | |
| "enable_torch_compile": true, | |
| "quantizer": [ | |
| "gptqmodel:1.9.0", | |
| "auto_round:0.4.6" | |
| ], | |
| "uri": "https://github.com/modelcloud/gptqmodel", | |
| "damp_percent": 0.01, | |
| "damp_auto_increment": 0.0025, | |
| "static_groups": false, | |
| "true_sequential": true, | |
| "mse": 0.0 | |
| } | |
| } |