diff --git "a/config.json" "b/config.json" new file mode 100644--- /dev/null +++ "b/config.json" @@ -0,0 +1,5406 @@ +{ + "_name_or_path": "meta-llama/Meta-Llama-3-8B", + "architectures": [ + "LlamaForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "bos_token_id": 128000, + "eos_token_id": 128001, + "hidden_act": "silu", + "hidden_size": 4096, + "initializer_range": 0.02, + "intermediate_size": 14336, + "max_position_embeddings": 8192, + "model_type": "llama", + "num_attention_heads": 32, + "num_hidden_layers": 32, + "num_key_value_heads": 8, + "pretraining_tp": 1, + "quant_config": { + "model.layers.0.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.0.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.0.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.0.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.0.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.0.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.0.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.1.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.1.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.1.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.1.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.1.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.1.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.1.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.10.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.10.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.10.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.10.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.10.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.10.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.10.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.11.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.11.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.11.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.11.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.11.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.11.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.11.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.12.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.12.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.12.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.12.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.12.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.12.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.12.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.13.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.13.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.13.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.13.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.13.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.13.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.13.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.14.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.14.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.14.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.14.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.14.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.14.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.14.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.15.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.15.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.15.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.15.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.15.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.15.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.15.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.16.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.16.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.16.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.16.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.16.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.16.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.16.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.17.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.17.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.17.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.17.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.17.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.17.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.17.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.18.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.18.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.18.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.18.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.18.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.18.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.18.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.19.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.19.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.19.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.19.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.19.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.19.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.19.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.2.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.2.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.2.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.2.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.2.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.2.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.2.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.20.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.20.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.20.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.20.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.20.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.20.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.20.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.21.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.21.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.21.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.21.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.21.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.21.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.21.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.22.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.22.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.22.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.22.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.22.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.22.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.22.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.23.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.23.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.23.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.23.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.23.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.23.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.23.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.24.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.24.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.24.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.24.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.24.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.24.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.24.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.25.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.25.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.25.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.25.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.25.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.25.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.25.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.26.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.26.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.26.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.26.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.26.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.26.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.26.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.27.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.27.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.27.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.27.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.27.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.27.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.27.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.28.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.28.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.28.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.28.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.28.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.28.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.28.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.29.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.29.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.29.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.29.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.29.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.29.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.29.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.3.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.3.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.3.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.3.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.3.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.3.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.3.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.30.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.30.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.30.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.30.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.30.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.30.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.30.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.31.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.31.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.31.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.31.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.31.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.31.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.31.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.4.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.4.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.4.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.4.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.4.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.4.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.4.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.5.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.5.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.5.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.5.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.5.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.5.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.5.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.6.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.6.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.6.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.6.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.6.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.6.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.6.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.7.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.7.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.7.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.7.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.7.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.7.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.7.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.8.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.8.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.8.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.8.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.8.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.8.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.8.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.9.mlp.down_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 14336, + "in_features": 14336, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.9.mlp.gate_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.9.mlp.up_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 14336, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.9.self_attn.k_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.9.self_attn.o_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.9.self_attn.q_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 4096, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + }, + "model.layers.9.self_attn.v_proj": { + "bias": false, + "enable_norm": true, + "enable_perm": true, + "group_num": 1, + "group_size": 4096, + "in_features": 4096, + "indices_as_float": false, + "is_indice_packed": true, + "num_centroids": [ + -1, + 4096 + ], + "num_res_centroids": [ + -1, + -1 + ], + "out_features": 1024, + "outlier_size": 0, + "vector_lens": [ + -1, + 4 + ] + } + }, + "rms_norm_eps": 1e-05, + "rope_scaling": null, + "rope_theta": 500000.0, + "tie_word_embeddings": false, + "torch_dtype": "float16", + "transformers_version": "4.37.2", + "use_cache": true, + "vocab_size": 128256 +}