jiangzhs/OpenCompass
1
0
Fork 0
mirror of https://github.com/open-compass/opencompass.git synced 2025-05-30 16:03:24 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

change max_task_size to dynamic

Browse Source
This commit is contained in:
Shiyao Ma 2024-05-22 11:28:09 +08:00
parent 31afe87026
commit db8d9a9798
1 changed files with 19 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
opencompass/partitioners/size.py
View File

@ -91,6 +91,25 @@ class SizePartitioner(BasePartitioner):
key=lambda x: self.get_cost(x), key=lambda x: self.get_cost(x),
reverse=True) reverse=True)
for model in comb['models']: for model in comb['models']:
# modified here, in order to maximize utilization of GPUs,
# change the CONSTANT max_task_size parameter to DYNAMIC based on model configuration
# if the model requires X gpu,
# available gpu is Y
# then there must be at least Y // X levels of parallelization, then max_task_size must produce at least Y // X tasks
X = model['run_cfg']['num_gpus']
Y = torch.cuda.device_count()
min_num_parallel_tasks = Y // X
if min_num_parallel_tasks > 1:
num_datasets = len(model_dataset_combinations[0]['datasets'])
total_estimated_size = num_datasets * \
min(
model['batch_size'],
sum([ele['num_repeats_per_file'] for ele in model_dataset_combinations[0]['datasets']]) //
len([ele['num_repeats_per_file'] for ele in model_dataset_combinations[0]['datasets']])
) * \
self.gen_task_coef
self.max_task_size = total_estimated_size // min_num_parallel_tasks - 1
chunks = [] # elements: tuple(size, dataset_chunk) chunks = [] # elements: tuple(size, dataset_chunk)
for dataset in comb['datasets']: for dataset in comb['datasets']:
filename = get_infer_output_path(model, dataset, out_dir) filename = get_infer_output_path(model, dataset, out_dir)