update

examples/tutorials/grpo/step_2_train_grpo_model_unsloth_ddp.py

@@ -18,13 +18,13 @@ torchrun --nproc_per_node=4 step_2_train_grpo_model_unsloth_ddp.py
 
 nohup torchrun --nproc_per_node=4 step_2_train_grpo_model_unsloth_ddp.py \
 --valid_dataset_size 500 \
---lora_rank 16 \
---num_train_epochs 1 \
+--lora_rank 32 \
+--num_train_epochs 2 \
 --per_device_batch_size 16 \
 --gradient_accumulation_steps 1 \
 --learning_rate 5e-6 \
 --num_generations 8 \
---steps_per_generation 8 \
+--steps_per_generation 4 \
 &
 
 kill -9 `ps -aef | grep 'step_2_train_grpo_model_unsloth_ddp.py' | grep -v grep | awk '{print $2}'`
@@ -133,7 +133,9 @@ def get_args():
     ),
 
     # 其他
-    parser.add_argument("--num_workers", default=0 if platform.system() == "Windows" else 2, type=int)
+    parser.add_argument("--num_workers",
+                        default=0 if platform.system() in ("Windows", "Darwin") else os.cpu_count() // 2,
+                        type=int)
     return parser.parse_args()