8
8
DEFAULT_PORT = 23333
9
9
10
10
11
- def write_to_summary (model_name , tp_num , result , msg , worker_id , work_dir = None ):
11
+ def write_to_summary (model_name , tp_num , result , msg , worker_id , backend_type , work_dir = None ):
12
12
status = '✅ PASS' if result else '❌ FAIL'
13
13
14
14
metrics = {}
@@ -39,7 +39,7 @@ def write_to_summary(model_name, tp_num, result, msg, worker_id, work_dir=None):
39
39
mmlu_value = metrics .get ('mmlu' , '' )
40
40
gsm8k_value = metrics .get ('gsm8k' , '' )
41
41
42
- summary_line = f'| { model_name } | TP{ tp_num } | { status } | { mmlu_value } | { gsm8k_value } |\n '
42
+ summary_line = f'| { model_name } | { backend_type } | TP{ tp_num } | { status } | { mmlu_value } | { gsm8k_value } |\n '
43
43
44
44
summary_file = os .environ .get ('GITHUB_STEP_SUMMARY' , None )
45
45
if summary_file :
@@ -49,17 +49,17 @@ def write_to_summary(model_name, tp_num, result, msg, worker_id, work_dir=None):
49
49
else :
50
50
with open (summary_file , 'r' ) as f :
51
51
first_lines = f .read (200 )
52
- if '| Model | TP | Status | mmlu | gsm8k |' not in first_lines :
52
+ if '| Model | Backend | TP | Status | mmlu | gsm8k |' not in first_lines :
53
53
write_header = True
54
54
55
55
with open (summary_file , 'a' ) as f :
56
56
if write_header :
57
57
f .write ('## Model Evaluation Results\n ' )
58
- f .write ('| Model | TP | Status | mmlu | gsm8k |\n ' )
59
- f .write ('|-------|----|--------|------|-------|\n ' )
58
+ f .write ('| Model | Backend | TP | Status | mmlu | gsm8k |\n ' )
59
+ f .write ('|-------|---------|---- |--------|------|-------|\n ' )
60
60
f .write (summary_line )
61
61
else :
62
- print (f'Summary: { model_name } | TP{ tp_num } | { status } | { mmlu_value } | { gsm8k_value } ' )
62
+ print (f'Summary: { model_name } | { backend_type } | TP{ tp_num } | { status } | { mmlu_value } | { gsm8k_value } ' )
63
63
64
64
65
65
def restful_test (config , run_id , prepare_environment , worker_id = 'gw0' , port = DEFAULT_PORT ):
@@ -181,7 +181,7 @@ def restful_test(config, run_id, prepare_environment, worker_id='gw0', port=DEFA
181
181
if error_lines :
182
182
final_msg += f'\n Log errors: { " | " .join (error_lines [:3 ])} '
183
183
184
- write_to_summary (model_name , tp_num , final_result , final_msg , worker_id , work_dir )
184
+ write_to_summary (model_name , tp_num , final_result , final_msg , worker_id , backend_type , work_dir )
185
185
186
186
return final_result , final_msg
187
187
@@ -193,10 +193,10 @@ def restful_test(config, run_id, prepare_environment, worker_id='gw0', port=DEFA
193
193
timeout_msg = (f'Evaluation timed out for { model_name } '
194
194
f'after 7200 seconds' )
195
195
if work_dir :
196
- write_to_summary (model_name , tp_num , False , timeout_msg , worker_id , work_dir )
196
+ write_to_summary (model_name , tp_num , False , timeout_msg , worker_id , backend_type , work_dir )
197
197
return False , timeout_msg
198
198
except Exception as e :
199
199
error_msg = f'Error during evaluation for { model_name } : { str (e )} '
200
200
if work_dir :
201
- write_to_summary (model_name , tp_num , False , error_msg , worker_id , work_dir )
201
+ write_to_summary (model_name , tp_num , False , error_msg , worker_id , backend_type , work_dir )
202
202
return False , error_msg
0 commit comments