From ad60c6b76c41e3c3980313cda39a445fc08b94c5 Mon Sep 17 00:00:00 2001 From: nil0x9 Date: Wed, 21 Jan 2026 15:25:33 +0800 Subject: [PATCH] [Enhance] add prefetch_factor to DataloaderConfig --- xtuner/v1/datasets/config.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/xtuner/v1/datasets/config.py b/xtuner/v1/datasets/config.py index 724fc5b2f..e85c2cc34 100644 --- a/xtuner/v1/datasets/config.py +++ b/xtuner/v1/datasets/config.py @@ -287,6 +287,7 @@ class DataloaderConfig(BaseDataloaderConfig): int, Parameter(help="pack extra buffer size when pack_level is expand_soft model") ] = 100 num_workers: Annotated[int, Parameter(help="dataloader num workers")] = 0 + prefetch_factor: Annotated[int | None, Parameter(help="number of batches loaded in advance by each worker")] = None pad_token_id: Annotated[int | None, Parameter(help="padding token id")] = None tokenizer_hash: Annotated[str | None, Parameter(help="tokenizer hash")] = None @@ -431,6 +432,7 @@ def build( dataset, batch_size=micro_batch_size, num_workers=self.num_workers, + prefetch_factor=self.prefetch_factor, # Ensure to round up or drop last based on the `global_batch_size`, # if you want to replace a custom sampler. sampler=sampler,