A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 11:00 UTC
[Bugfix] Add missing ASRDataset import and CLI args in benchmarks/throughput.py (#38114)
Signed-off-by: nemanjaudovic <nudovic@amd.com>
nemanjaudovic committed
Apr 8, 2026 at 15:53 UTC
1a2c17634eccc4e68d9e1ab654f702d55361c754
1 file changed
+22
vllm/benchmarks/throughput.py
+22
@@ -17,6 +17,7 @@ from transformers import AutoModelForCausalLM, PreTrainedTokenizerBase
17
18
from vllm.benchmarks.datasets import (
19
AIMODataset,
20
+ ASRDataset,
21
BurstGPTDataset,
22
ConversationDataset,
23
InstructCoderDataset,
@@ -414,6 +415,12 @@ def get_requests(args, tokenizer):
415
dataset_cls = AIMODataset
416
common_kwargs["dataset_subset"] = None
417
common_kwargs["dataset_split"] = "train"
418
+ elif args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS:
419
+ dataset_cls = ASRDataset
420
+ common_kwargs["dataset_subset"] = args.hf_subset
421
+ common_kwargs["dataset_split"] = args.hf_split
422
+ sample_kwargs["asr_min_audio_len_sec"] = args.asr_min_audio_len_sec
423
+ sample_kwargs["asr_max_audio_len_sec"] = args.asr_max_audio_len_sec
424
elif args.dataset_name == "prefix_repetition":
425
dataset_cls = PrefixRepetitionRandomDataset
426
sample_kwargs["prefix_len"] = args.prefix_repetition_prefix_len
@@ -557,6 +564,7 @@ def validate_args(args):
564
elif args.dataset_path in (
565
InstructCoderDataset.SUPPORTED_DATASET_PATHS
566
| AIMODataset.SUPPORTED_DATASET_PATHS
567
+ | ASRDataset.SUPPORTED_DATASET_PATHS
568
):
569
assert args.backend == "vllm", (
570
f"{args.dataset_path} needs to use vllm as the backend."
@@ -841,6 +849,20 @@ def add_cli_args(parser: argparse.ArgumentParser):
849
add_random_dataset_base_args(parser)
850
add_random_multimodal_dataset_args(parser)
851
852
+ # ASR dataset
853
+ parser.add_argument(
854
+ "--asr-min-audio-len-sec",
855
+ type=float,
856
+ default=0.0,
857
+ help="Minimum audio duration in seconds for ASR dataset filtering.",
858
+ )
859
+ parser.add_argument(
860
+ "--asr-max-audio-len-sec",
861
+ type=float,
862
+ default=float("inf"),
863
+ help="Maximum audio duration in seconds for ASR dataset filtering.",
864
+ )
865
+
866
parser = AsyncEngineArgs.add_cli_args(parser)
867
868