import json import numpy as np index = { 'call_candidate': None, 'exe_candidate': None, 'perf_candidates': [], } paths = [ '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_oss_120b_medium_v3_mem_0.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_oss_120b_medium_v3_mem_1.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_oss_120b_medium_v3_mem_2.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_oss_120b_medium_v3_mem_3.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_oss_120b_high_v3_mem_0.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_oss_120b_high_v3_mem_1.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_oss_120b_high_v3_mem_2.json', # '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8_mem_0.json', # '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8_mem_1.json', # '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8_mem_2.json', # '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8_mem_3.json', # '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8_mem_4.json', # '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8_mem_5.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_mem_0.json', # '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_mem_1.json', # '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_mem_2.json', # '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_mem_3.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_code_mem_0.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_code_mem_1.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_code_mem_2.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_code_mem_3.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_code_mem_4.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_oss_120b_high_v4_mem_0.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_oss_120b_high_v4_mem_1.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_30_mem_0.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_30_mem_1.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_30_mem_2.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1107_q8r_30_mem_3.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1114_q8r_code_mem_0.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1114_q8r_8b_mem_0.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1114_q8r_8b_mem_1.json', '/wekafs/zihao/exp/agent/jianghui_triton/GEAK-agent/outputs/outputs/1114_q8r_8b_mem_2.json', ] for path in paths: result = {i:[] for i in index} data = json.load(open(path)) for k, v in data.items(): for i in index: result[i].append(v[i]) print(f'[path]:') print(path) print('\n') for k, v in index.items(): print(f'[{k}]:') print(np.array([str(i) != str(v) for i in result[k]]).mean()) if k == 'perf_candidates': acc = [] for i in result[k]: if len(i) == 0: acc.append(1) else: acc.append(max(1, i[-1][1])) print(np.array(acc).mean()) print('\n') print('-'*100)