Skip to content

🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.

[fix] dist cleanup

jingyaogong42a4e8c86a0e
diff --git a/trainer/train_distill_reason.py b/trainer/train_distill_reason.py
index 3a10847..9162cca 100644
--- a/trainer/train_distill_reason.py
+++ b/trainer/train_distill_reason.py
@@ -176,3 +176,6 @@ if __name__ == "__main__":
else: # 默认从头开始
loader = DataLoader(train_ds, batch_size=args.batch_size, shuffle=(train_sampler is None), sampler=train_sampler, num_workers=args.num_workers, pin_memory=True)
train_epoch(epoch, loader, len(loader), tokenizer, lm_config, 0, wandb)
+
+ # ========== 9. 清理分布进程 ==========
+ if dist.is_initialized(): dist.destroy_process_group()
\ No newline at end of file
diff --git a/trainer/train_distillation.py b/trainer/train_distillation.py
index 24450a3..1deec83 100644
--- a/trainer/train_distillation.py
+++ b/trainer/train_distillation.py
@@ -226,3 +226,6 @@ if __name__ == "__main__":
else: # 默认从头开始
loader = DataLoader(train_ds, batch_size=args.batch_size, shuffle=(train_sampler is None), sampler=train_sampler, num_workers=args.num_workers, pin_memory=True)
train_epoch(epoch, loader, len(loader), teacher_model, lm_config_student, 0, wandb, args.alpha, args.temperature)
+
+ # ========== 9. 清理分布进程 ==========
+ if dist.is_initialized(): dist.destroy_process_group()
\ No newline at end of file
diff --git a/trainer/train_dpo.py b/trainer/train_dpo.py
index eccfaff..c4987df 100644
--- a/trainer/train_dpo.py
+++ b/trainer/train_dpo.py
@@ -211,3 +211,6 @@ if __name__ == "__main__":
else: # 默认从头开始
loader = DataLoader(train_ds, batch_size=args.batch_size, shuffle=(train_sampler is None), sampler=train_sampler, num_workers=args.num_workers, pin_memory=True)
train_epoch(epoch, loader, len(loader), ref_model, lm_config, 0, wandb, args.beta)
+
+ # ========== 9. 清理分布进程 ==========
+ if dist.is_initialized(): dist.destroy_process_group()
\ No newline at end of file
diff --git a/trainer/train_full_sft.py b/trainer/train_full_sft.py
index faec3f5..2ab8f6b 100644
--- a/trainer/train_full_sft.py
+++ b/trainer/train_full_sft.py
@@ -164,3 +164,6 @@ if __name__ == "__main__":
else: # 默认从头开始
loader = DataLoader(train_ds, batch_size=args.batch_size, shuffle=(train_sampler is None), sampler=train_sampler, num_workers=args.num_workers, pin_memory=True)
train_epoch(epoch, loader, len(loader), 0, wandb)
+
+ # ========== 9. 清理分布进程 ==========
+ if dist.is_initialized(): dist.destroy_process_group()
\ No newline at end of file
diff --git a/trainer/train_grpo.py b/trainer/train_grpo.py
index d5b30ed..1438bdd 100755
--- a/trainer/train_grpo.py
+++ b/trainer/train_grpo.py
@@ -294,3 +294,6 @@ if __name__ == "__main__":
drop_last=False, shuffle=(train_sampler is None),
num_workers=args.num_workers, sampler=train_sampler)
grpo_train_epoch(epoch, loader, len(loader), ref_model, reward_model, reward_tokenizer, 0, wandb)
+
+ # ========== 9. 清理分布进程 ==========
+ if dist.is_initialized(): dist.destroy_process_group()
\ No newline at end of file
diff --git a/trainer/train_lora.py b/trainer/train_lora.py
index dff0f19..2af9739 100644
--- a/trainer/train_lora.py
+++ b/trainer/train_lora.py
@@ -177,3 +177,6 @@ if __name__ == "__main__":
else: # 默认从头开始
loader = DataLoader(train_ds, batch_size=args.batch_size, shuffle=(train_sampler is None), sampler=train_sampler, num_workers=args.num_workers, pin_memory=True)
train_epoch(epoch, loader, len(loader), lora_params, 0, wandb)
+
+ # ========== 10. 清理分布进程 ==========
+ if dist.is_initialized(): dist.destroy_process_group()
\ No newline at end of file
diff --git a/trainer/train_ppo.py b/trainer/train_ppo.py
index fd094b3..d38836d 100644
--- a/trainer/train_ppo.py
+++ b/trainer/train_ppo.py
@@ -368,3 +368,6 @@ if __name__ == "__main__":
sampler=train_sampler, num_workers=args.num_workers, pin_memory=True)
ppo_train_epoch(epoch, loader, len(loader), old_actor_model, ref_model,
actor_scheduler, critic_scheduler, reward_model, reward_tokenizer, 0, wandb)
+
+ # ========== 9. 清理分布进程 ==========
+ if dist.is_initialized(): dist.destroy_process_group()
\ No newline at end of file
diff --git a/trainer/train_pretrain.py b/trainer/train_pretrain.py
index 4596041..29c4d88 100644
--- a/trainer/train_pretrain.py
+++ b/trainer/train_pretrain.py
@@ -163,3 +163,6 @@ if __name__ == "__main__":
else: # 默认从头开始
loader = DataLoader(train_ds, batch_size=args.batch_size, shuffle=(train_sampler is None), sampler=train_sampler, num_workers=args.num_workers, pin_memory=True)
train_epoch(epoch, loader, len(loader), 0, wandb)
+
+ # ========== 9. 清理分布进程 ==========
+ if dist.is_initialized(): dist.destroy_process_group()
\ No newline at end of file
diff --git a/trainer/train_spo.py b/trainer/train_spo.py
index a15a2ac..b81e6b5 100755
--- a/trainer/train_spo.py
+++ b/trainer/train_spo.py
@@ -344,3 +344,6 @@ if __name__ == "__main__":
drop_last=False, shuffle=(train_sampler is None),
num_workers=args.num_workers, sampler=train_sampler)
spo_train_epoch(epoch, loader, len(loader), ref_model, reward_model, reward_tokenizer, value_tracker, 0, wandb)
+
+ # ========== 9. 清理分布进程 ==========
+ if dist.is_initialized(): dist.destroy_process_group()
\ No newline at end of file