Unstructured Sparse Pruning

[rG223]

Hello,

I have several questions. Thank you in advance for your help!

(1) I was wondering if you could kindly provide example commands for running prune.py, owl_prune.py, and evo_prune_search.py. Some required parameters appear to lack default values, and I’m uncertain about the correct inputs to use.

(2) Could you please clarify how to run OWL and Uniform method?

(3) How do I set some parameters in lmeval.py? I run the following script:：

python lmeval.py  \
    --model_args "pretrained=mistralai/Mistral-7B-v0.3" \
    --tasks "arc_easy,arc_challenge,piqa,winogrande,hellaswag" \
    --drop_layer_config "/home/yerg/EvoPress/log/layer_drop_config.txt"

I got an error, UnicodeDecodeError: 'utf-8' codec can't decode byte 0xb5 in position 1: invalid start byte.
Traceback (most recent call last):
File "/mnt/data/home/yerg/EvoPress/lmeval.py", line 413, in
cli_evaluate()
File "/mnt/data/home/yerg/EvoPress/lmeval.py", line 360, in cli_evaluate
results = evaluator.simple_evaluate(
File "/home/yerg/anaconda3/envs/llm/lib/python3.10/site-packages/lm_eval/evaluator.py", line 124, in simple_evaluate
task_dict = lm_eval.tasks.get_task_dict(tasks)
File "/home/yerg/anaconda3/envs/llm/lib/python3.10/site-packages/lm_eval/tasks/init.py", line 250, in get_task_dict
task_name: get_task(task_name=task_element, config=config),
File "/home/yerg/anaconda3/envs/llm/lib/python3.10/site-packages/lm_eval/tasks/init.py", line 192, in get_task
return TASK_REGISTRY[task_name](config=config)
File "/home/yerg/anaconda3/envs/llm/lib/python3.10/site-packages/lm_eval/api/task.py", line 633, in init
self.download(self.config.dataset_kwargs)
File "/home/yerg/anaconda3/envs/llm/lib/python3.10/site-packages/lm_eval/api/task.py", line 730, in download
self.dataset = datasets.load_dataset(
File "/home/yerg/anaconda3/envs/llm/lib/python3.10/site-packages/datasets/load.py", line 2606, in load_dataset
builder_instance = load_dataset_builder(
File "/home/yerg/anaconda3/envs/llm/lib/python3.10/site-packages/datasets/load.py", line 2277, in load_dataset_builder
dataset_module = dataset_module_factory(
File "/home/yerg/anaconda3/envs/llm/lib/python3.10/site-packages/datasets/load.py", line 1923, in dataset_module_factory
raise e1 from None
File "/home/yerg/anaconda3/envs/llm/lib/python3.10/site-packages/datasets/load.py", line 1875, in dataset_module_factory
can_load_config_from_parquet_export = "DEFAULT_CONFIG_NAME" not in f.read()
File "/home/yerg/anaconda3/envs/llm/lib/python3.10/codecs.py", line 322, in decode
(result, consumed) = self._buffer_decode(data, self.errors, final)

[rG223]

The third problem can be solved by huggingface/datasets#6760.
Datasets package is needed to downgrade to 2.14.6 version for runing lmeval.py script.