From adc0c78dbc696d8553cb2bb3c249a2fa8113ef2d Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Wed, 18 Feb 2026 20:53:32 +0000 Subject: [PATCH] reduce dataset_num_proc to 1/4 of cpu_count --- studio/backend/core/training/trainer.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py index 1cd9f87e95..e93ba097d1 100644 --- a/studio/backend/core/training/trainer.py +++ b/studio/backend/core/training/trainer.py @@ -707,7 +707,7 @@ class UnslothTrainer: "output_dir": output_dir, "report_to": ["wandb"] if training_args.get('enable_wandb', False) else "none", "include_num_input_tokens_seen": True, # Enable token counting - "dataset_num_proc": max(1, os.cpu_count() // 3), + "dataset_num_proc": max(1, os.cpu_count() // 4), } # Add warmup parameter - use warmup_ratio if provided, otherwise warmup_steps