wehub-resource-sync
65 lines
1.8 KiB
Python
65 lines
1.8 KiB
Python
import os
|
|
import argparse
|
|
from tqdm import tqdm
|
|
import MNN.llm as mnnllm
|
|
from datasets import load_dataset
|
|
import torch
|
|
import copy
|
|
|
|
def main(args):
|
|
|
|
model = mnnllm.create(args.mnn_path)
|
|
|
|
model.set_config({'all_logits': True, 'use_template': False})
|
|
|
|
model.set_config({'enable_debug': True})
|
|
model.load()
|
|
|
|
model.enable_collection_mode(1, args.output_path, args.target_sparsity)
|
|
|
|
eval_dataset = args.eval_dataset
|
|
dataset_parts = eval_dataset.split("/")
|
|
if len(dataset_parts) < 2:
|
|
raise ValueError("eval_dataset must be formatted as dataset/config or namespace/dataset/config.")
|
|
dataset_name = "/".join(dataset_parts[:-1])
|
|
dataset_dir = dataset_parts[-1]
|
|
|
|
dataset = load_dataset(dataset_name, dataset_dir, split="test")
|
|
input_ids = model.tokenizer_encode("\n\n".join(dataset["text"]))
|
|
input_ids = input_ids[:args.length]
|
|
|
|
_ = model.forward(input_ids)
|
|
|
|
|
|
|
|
if __name__ == "__main__":
|
|
parser = argparse.ArgumentParser(description="Get thresholds from MNN model.")
|
|
parser.add_argument(
|
|
"-m",
|
|
"--mnn-path",
|
|
type=str,
|
|
required=True,
|
|
help="mnn model path",
|
|
)
|
|
|
|
parser.add_argument(
|
|
"-d", "--eval_dataset", type=str, default='Salesforce/wikitext/wikitext-2-raw-v1', help="dataset, default is `Salesforce/wikitext/wikitext-2-raw-v1`."
|
|
)
|
|
|
|
parser.add_argument(
|
|
"-o", "--output-path", type=str, default='thresholds.json', help="output path, default is `thresholds.json`."
|
|
)
|
|
|
|
parser.add_argument(
|
|
"-t", "--target-sparsity", type=float, default=0.5, help="target sparsity, default is 0.5."
|
|
)
|
|
|
|
parser.add_argument(
|
|
"-l", "--length", type=int, default=512, help="length of samples, default is 512."
|
|
)
|
|
|
|
|
|
args = parser.parse_args()
|
|
|
|
main(args)
|