[Misc]add coding benchmark for speculative decoding (#15303)
Signed-off-by: CXIAAAAA <cxia0209@gmail.com>
This commit is contained in:
@@ -53,8 +53,9 @@ except ImportError:
|
||||
from argparse import ArgumentParser as FlexibleArgumentParser
|
||||
|
||||
from benchmark_dataset import (BurstGPTDataset, HuggingFaceDataset,
|
||||
RandomDataset, SampleRequest, ShareGPTDataset,
|
||||
SonnetDataset, VisionArenaDataset)
|
||||
InstructCoderDataset, RandomDataset,
|
||||
SampleRequest, ShareGPTDataset, SonnetDataset,
|
||||
VisionArenaDataset)
|
||||
from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
|
||||
|
||||
MILLISECONDS_TO_SECONDS_CONVERSION = 1000
|
||||
@@ -588,9 +589,14 @@ def main(args: argparse.Namespace):
|
||||
elif args.dataset_name == "hf":
|
||||
# Choose between VisionArenaDataset
|
||||
# and HuggingFaceDataset based on provided parameters.
|
||||
dataset_class = (VisionArenaDataset if args.dataset_path
|
||||
== VisionArenaDataset.VISION_ARENA_DATASET_PATH
|
||||
and args.hf_subset is None else HuggingFaceDataset)
|
||||
dataset_class = HuggingFaceDataset
|
||||
if args.dataset_path == VisionArenaDataset.VISION_ARENA_DATASET_PATH:
|
||||
assert args.hf_subset is None, "VisionArenaDataset needs hf_subset to be None." #noqa: E501
|
||||
dataset_class = VisionArenaDataset
|
||||
elif args.dataset_path == "likaixin/InstructCoder":
|
||||
dataset_class = InstructCoderDataset
|
||||
args.hf_split = "train"
|
||||
|
||||
input_requests = dataset_class(
|
||||
dataset_path=args.dataset_path,
|
||||
dataset_subset=args.hf_subset,
|
||||
|
||||
Reference in New Issue
Block a user