From 5b39b268f506150dbab38f6f6c04b7c843e37c07 Mon Sep 17 00:00:00 2001 From: pmaybank <113125070+pmaybank@users.noreply.github.com> Date: Wed, 29 Apr 2026 13:57:58 +0100 Subject: [PATCH] hf_name argument for vllm bench throughput CLI (#41012) Signed-off-by: Philip Maybank Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> --- vllm/benchmarks/throughput.py | 52 ++++++++++++++++++++++++++++++----- 1 file changed, 45 insertions(+), 7 deletions(-) diff --git a/vllm/benchmarks/throughput.py b/vllm/benchmarks/throughput.py index 42a8132ffe..1c3b728fdb 100644 --- a/vllm/benchmarks/throughput.py +++ b/vllm/benchmarks/throughput.py @@ -393,29 +393,48 @@ def get_requests(args, tokenizer): elif args.dataset_name == "hf": if args.output_len is not None: sample_kwargs["output_len"] = args.output_len - if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS: + common_kwargs["hf_name"] = args.hf_name + if ( + args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS + or args.hf_name in VisionArenaDataset.SUPPORTED_DATASET_PATHS + ): dataset_cls = VisionArenaDataset common_kwargs["dataset_subset"] = None common_kwargs["dataset_split"] = "train" sample_kwargs["enable_multimodal_chat"] = True - elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS: + elif ( + args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS + or args.hf_name in InstructCoderDataset.SUPPORTED_DATASET_PATHS + ): dataset_cls = InstructCoderDataset common_kwargs["dataset_split"] = "train" - elif args.dataset_path in MultiModalConversationDataset.SUPPORTED_DATASET_PATHS: + elif ( + args.dataset_path in MultiModalConversationDataset.SUPPORTED_DATASET_PATHS + or args.hf_name in MultiModalConversationDataset.SUPPORTED_DATASET_PATHS + ): dataset_cls = MultiModalConversationDataset common_kwargs["dataset_subset"] = args.hf_subset common_kwargs["dataset_split"] = args.hf_split sample_kwargs["enable_multimodal_chat"] = True - elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS: + elif ( + args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS + or args.hf_name in ConversationDataset.SUPPORTED_DATASET_PATHS + ): dataset_cls = ConversationDataset common_kwargs["dataset_subset"] = args.hf_subset common_kwargs["dataset_split"] = args.hf_split sample_kwargs["enable_multimodal_chat"] = True - elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS: + elif ( + args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS + or args.hf_name in AIMODataset.SUPPORTED_DATASET_PATHS + ): dataset_cls = AIMODataset common_kwargs["dataset_subset"] = None common_kwargs["dataset_split"] = "train" - elif args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS: + elif ( + args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS + or args.hf_name in ASRDataset.SUPPORTED_DATASET_PATHS + ): dataset_cls = ASRDataset common_kwargs["dataset_subset"] = args.hf_subset common_kwargs["dataset_split"] = args.hf_split @@ -557,6 +576,10 @@ def validate_args(args): VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys() | MultiModalConversationDataset.SUPPORTED_DATASET_PATHS | ConversationDataset.SUPPORTED_DATASET_PATHS + ) or args.hf_name in ( + VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys() + | MultiModalConversationDataset.SUPPORTED_DATASET_PATHS + | ConversationDataset.SUPPORTED_DATASET_PATHS ): assert args.backend == "vllm-chat", ( f"{args.dataset_path} needs to use vllm-chat as the backend." @@ -565,6 +588,10 @@ def validate_args(args): InstructCoderDataset.SUPPORTED_DATASET_PATHS | AIMODataset.SUPPORTED_DATASET_PATHS | ASRDataset.SUPPORTED_DATASET_PATHS + ) or args.hf_name in ( + InstructCoderDataset.SUPPORTED_DATASET_PATHS + | AIMODataset.SUPPORTED_DATASET_PATHS + | ASRDataset.SUPPORTED_DATASET_PATHS ): assert args.backend == "vllm", ( f"{args.dataset_path} needs to use vllm as the backend." @@ -795,7 +822,7 @@ def add_cli_args(parser: argparse.ArgumentParser): "context in a request (default: 0).", ) - # hf dtaset + # hf dataset parser.add_argument( "--hf-subset", type=str, @@ -808,6 +835,17 @@ def add_cli_args(parser: argparse.ArgumentParser): default=None, help="Split of the HF dataset.", ) + parser.add_argument( + "--hf-name", + type=str, + default=None, + help=( + "Name of the dataset on HuggingFace " + "(e.g., 'lmms-lab/LLaVA-OneVision-Data'). " + "Specify this when --dataset-path is a local filesystem path " + "so the benchmark can identify the correct dataset class." + ), + ) parser.add_argument( "--profile", action="store_true",