setup defaults for dataloader to ensure GPU is kept busy (#2632) [skip ci]

This commit is contained in:
Wing Lian
2025-07-07 10:10:58 -04:00
committed by GitHub
parent 69cd49a7aa
commit 759cefb741
2 changed files with 29 additions and 0 deletions

View File

@@ -1114,3 +1114,17 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
raise ValueError("QAT is not supported on torch version < 2.6.0")
return data
@model_validator(mode="before")
@classmethod
def default_dataloader_opts(cls, data):
if (
data.get("dataloader_num_workers") is None
and data.get("dataloader_pin_memory") is None
and data.get("dataloader_prefetch_factor") is None
):
data["dataloader_num_workers"] = data.get("capabilities").get("n_gpu", 1)
data["dataloader_pin_memory"] = True
data["dataloader_prefetch_factor"] = 256
return data