2440 lines
68 KiB
JSON
2440 lines
68 KiB
JSON
|
{
|
||
|
|
"_attn_implementation_autoset": true,
|
||
|
|
"_name_or_path": "SeaLLMs/SeaLLMs-v3-7B",
|
||
|
|
"architectures": [
|
||
|
|
"Qwen2ForCausalLM"
|
||
|
|
],
|
||
|
|
"attention_dropout": 0.0,
|
||
|
|
"bos_token_id": 151643,
|
||
|
|
"eos_token_id": 151643,
|
||
|
|
"hidden_act": "silu",
|
||
|
|
"hidden_size": 3584,
|
||
|
|
"initializer_range": 0.02,
|
||
|
|
"intermediate_size": 18944,
|
||
|
|
"max_position_embeddings": 131072,
|
||
|
|
"max_window_layers": 28,
|
||
|
|
"model_type": "qwen2",
|
||
|
|
"num_attention_heads": 28,
|
||
|
|
"num_hidden_layers": 28,
|
||
|
|
"num_key_value_heads": 4,
|
||
|
|
"quantization_config": {
|
||
|
|
"bits": 4,
|
||
|
|
"checkpoint_format": "gptq",
|
||
|
|
"desc_act": true,
|
||
|
|
"group_size": 128,
|
||
|
|
"lm_head": false,
|
||
|
|
"meta": {
|
||
|
|
"amp": true,
|
||
|
|
"batch_size": 8,
|
||
|
|
"damp_auto_increment": 0.0025,
|
||
|
|
"damp_percent": 0.01,
|
||
|
|
"data_type": "int",
|
||
|
|
"dynamic_max_gap": -1,
|
||
|
|
"enable_minmax_tuning": true,
|
||
|
|
"enable_quanted_input": true,
|
||
|
|
"enable_torch_compile": true,
|
||
|
|
"gradient_accumulate_steps": 1,
|
||
|
|
"iters": 1000,
|
||
|
|
"layer_config": {
|
||
|
|
"lm_head": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 16,
|
||
|
|
"data_type": "bfloat",
|
||
|
|
"group_size": null,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": null
|
||
|
|
},
|
||
|
|
"model.layers.0.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.0.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.0.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.0.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.0.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.0.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.0.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.1.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.1.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.1.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.1.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.1.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.1.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.1.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.10.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.10.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.10.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.10.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.10.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.10.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.10.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.11.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.11.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.11.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.11.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.11.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.11.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.11.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.12.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.12.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.12.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.12.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.12.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.12.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.12.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.13.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.13.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.13.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.13.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.13.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.13.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.13.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.14.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.14.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.14.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.14.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.14.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.14.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.14.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.15.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.15.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.15.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.15.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.15.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.15.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.15.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.16.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.16.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.16.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.16.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.16.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.16.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.16.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.17.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.17.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.17.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.17.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.17.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.17.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.17.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.18.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.18.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.18.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.18.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.18.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.18.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.18.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.19.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.19.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.19.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.19.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.19.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.19.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.19.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.2.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.2.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.2.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.2.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.2.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.2.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.2.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.20.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.20.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.20.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.20.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.20.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.20.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.20.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.21.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.21.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.21.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.21.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.21.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.21.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.21.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.22.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.22.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.22.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.22.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.22.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.22.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.22.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.23.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.23.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.23.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.23.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.23.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.23.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.23.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.24.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.24.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.24.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.24.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.24.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.24.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.24.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.25.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.25.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.25.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.25.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.25.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.25.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.25.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.26.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.26.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.26.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.26.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.26.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.26.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.26.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.27.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.27.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.27.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.27.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.27.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.27.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.27.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.3.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.3.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.3.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.3.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.3.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.3.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.3.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.4.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.4.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.4.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.4.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.4.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.4.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.4.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.5.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.5.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.5.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.5.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.5.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.5.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.5.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.6.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.6.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.6.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.6.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.6.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.6.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.6.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.7.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.7.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.7.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.7.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.7.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.7.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.7.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.8.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.8.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.8.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.8.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.8.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.8.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.8.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.9.mlp.down_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.9.mlp.gate_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.9.mlp.up_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.9.self_attn.k_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.9.self_attn.o_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.9.self_attn.q_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"model.layers.9.self_attn.v_proj": {
|
||
|
|
"act_bits": 16,
|
||
|
|
"act_data_type": "int",
|
||
|
|
"act_dynamic": true,
|
||
|
|
"act_group_size": 128,
|
||
|
|
"act_sym": true,
|
||
|
|
"bits": 4,
|
||
|
|
"data_type": "int",
|
||
|
|
"group_size": 128,
|
||
|
|
"scale_dtype": "bfloat16",
|
||
|
|
"sym": true
|
||
|
|
}
|
||
|
|
},
|
||
|
|
"low_cache_mem_usage": true,
|
||
|
|
"low_gpu_mem_usage": true,
|
||
|
|
"lr": 0.005,
|
||
|
|
"lr_scheduler": null,
|
||
|
|
"minmax_lr": null,
|
||
|
|
"mse": 0.0,
|
||
|
|
"nblocks": 1,
|
||
|
|
"not_use_best_mse": false,
|
||
|
|
"nsamples": 512,
|
||
|
|
"quantizer": [
|
||
|
|
"gptqmodel:1.9.0",
|
||
|
|
"auto_round:0.4.6"
|
||
|
|
],
|
||
|
|
"sampler": "rand",
|
||
|
|
"scale_dtype": "bf16",
|
||
|
|
"seed": 42,
|
||
|
|
"seqlen": 2048,
|
||
|
|
"static_groups": false,
|
||
|
|
"true_sequential": true,
|
||
|
|
"uri": "https://github.com/modelcloud/gptqmodel"
|
||
|
|
},
|
||
|
|
"pack_dtype": "int32",
|
||
|
|
"quant_method": "gptq",
|
||
|
|
"sym": true
|
||
|
|
},
|
||
|
|
"rms_norm_eps": 1e-06,
|
||
|
|
"rope_scaling": null,
|
||
|
|
"rope_theta": 1000000.0,
|
||
|
|
"sliding_window": 131072,
|
||
|
|
"tie_word_embeddings": false,
|
||
|
|
"torch_dtype": "bfloat16",
|
||
|
|
"transformers_version": "4.49.0",
|
||
|
|
"use_cache": false,
|
||
|
|
"use_sliding_window": false,
|
||
|
|
"vocab_size": 152064
|
||
|
|
}
|