🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.
[fix] loss-issues-430
jingyaogong805744e60a13
diff --git a/trainer/train_distill_reason.py b/trainer/train_distill_reason.pyindex 9c1c9eb..525490e 100644--- a/trainer/train_distill_reason.py+++ b/trainer/train_distill_reason.py@@ -79,7 +79,7 @@ def train_epoch(epoch, wandb):if step % args.log_interval == 0 or step == iter_per_epoch - 1:spend_time = time.time() - start_timeLogger(- 'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.12f} epoch_Time:{}min:'.format(+ 'Epoch:[{}/{}]({}/{}) loss:{:.6f} lr:{:.12f} epoch_Time:{}min:'.format(epoch + 1,args.epochs,step,diff --git a/trainer/train_distillation.py b/trainer/train_distillation.pyindex 582e36d..377e5f3 100644--- a/trainer/train_distillation.py+++ b/trainer/train_distillation.py@@ -113,7 +113,7 @@ def train_epoch(epoch, wandb, alpha=0.0, temperature=1.0):if step % args.log_interval == 0 or step == iter_per_epoch - 1:spend_time = time.time() - start_timeLogger(- 'Epoch:[{}/{}]({}/{}) loss:{:.4f} lr:{:.12f} epoch_Time:{}min:'.format(+ 'Epoch:[{}/{}]({}/{}) loss:{:.6f} lr:{:.12f} epoch_Time:{}min:'.format(epoch,args.epochs - 1,step,diff --git a/trainer/train_dpo.py b/trainer/train_dpo.pyindex cdb9d92..4b72afa 100644--- a/trainer/train_dpo.py+++ b/trainer/train_dpo.py@@ -103,7 +103,7 @@ def train_epoch(epoch, wandb):if step % args.log_interval == 0 or step == iter_per_epoch - 1:spend_time = time.time() - start_timeLogger(- 'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.12f} epoch_Time:{}min:'.format(+ 'Epoch:[{}/{}]({}/{}) loss:{:.6f} lr:{:.12f} epoch_Time:{}min:'.format(epoch + 1,args.epochs,step,diff --git a/trainer/train_full_sft.py b/trainer/train_full_sft.pyindex bae0dff..7c601a4 100644--- a/trainer/train_full_sft.py+++ b/trainer/train_full_sft.py@@ -66,7 +66,7 @@ def train_epoch(epoch, wandb):if step % args.log_interval == 0 or step == iter_per_epoch - 1:spend_time = time.time() - start_timeLogger(- 'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.12f} epoch_Time:{}min:'.format(+ 'Epoch:[{}/{}]({}/{}) loss:{:.6f} lr:{:.12f} epoch_Time:{}min:'.format(epoch + 1,args.epochs,step,diff --git a/trainer/train_grpo.py b/trainer/train_grpo.pyindex 8aeb349..7c33a94 100755--- a/trainer/train_grpo.py+++ b/trainer/train_grpo.py@@ -169,7 +169,7 @@ def grpo_train_epoch(epoch, wandb):Logger(f'Epoch: {epoch}, Step: {step + 1}/{iter_per_epoch}, '- f'Actor Loss: {policy_loss_val:.4f}, Reward: {avg_reward_val:.4f}, '+ f'Actor Loss: {policy_loss_val:.6f}, Reward: {avg_reward_val:.6f}, 'f'Avg Response Len: {avg_len_val:.2f}, LR: {current_lr:.2e}')if wandb and (not ddp or dist.get_rank() == 0):diff --git a/trainer/train_lora.py b/trainer/train_lora.pyindex 4c73126..e2659f4 100644--- a/trainer/train_lora.py+++ b/trainer/train_lora.py@@ -67,7 +67,7 @@ def train_epoch(epoch, wandb):if step % args.log_interval == 0 or step == iter_per_epoch - 1:spend_time = time.time() - start_timeLogger(- 'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.12f} epoch_Time:{}min:'.format(+ 'Epoch:[{}/{}]({}/{}) loss:{:.6f} lr:{:.12f} epoch_Time:{}min:'.format(epoch + 1,args.epochs,step,diff --git a/trainer/train_ppo.py b/trainer/train_ppo.pyindex f8a8d38..6ddbb64 100644--- a/trainer/train_ppo.py+++ b/trainer/train_ppo.py@@ -193,8 +193,8 @@ def ppo_train_epoch(epoch: int, wandb_run, old_actor_model, ref_model, actor_sch})Logger(f"Epoch: {epoch}, Step: {step + 1}/{len(train_loader)}, "- f"Actor Loss: {actor_loss_val:.4f}, Critic Loss: {critic_loss_val:.4f}, "- f"Reward: {reward_val:.4f}, KL: {kl_val:.4f}, KL_ref: {kl_ref_val:.4f}, "+ f"Actor Loss: {actor_loss_val:.6f}, Critic Loss: {critic_loss_val:.6f}, "+ f"Reward: {reward_val:.6f}, KL: {kl_val:.6f}, KL_ref: {kl_ref_val:.6f}, "f"Avg Response Len: {avg_len_val:.2f}, Actor LR: {actor_lr:.2e}, Critic LR: {critic_lr:.2e}")if (step + 1) % args.update_old_actor_freq == 0:diff --git a/trainer/train_pretrain.py b/trainer/train_pretrain.pyindex ab6f68f..13b83fc 100644--- a/trainer/train_pretrain.py+++ b/trainer/train_pretrain.py@@ -66,7 +66,7 @@ def train_epoch(epoch, wandb):if step % args.log_interval == 0 or step == iter_per_epoch - 1:spend_time = time.time() - start_timeLogger(- 'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.12f} epoch_Time:{}min:'.format(+ 'Epoch:[{}/{}]({}/{}) loss:{:.6f} lr:{:.12f} epoch_Time:{}min:'.format(epoch + 1,args.epochs,step,diff --git a/trainer/train_spo.py b/trainer/train_spo.pyindex 7ea15bf..1b81791 100755--- a/trainer/train_spo.py+++ b/trainer/train_spo.py@@ -216,7 +216,7 @@ def spo_train_epoch(epoch, wandb, value_tracker):Logger(f'Epoch: {epoch}, Step: {step + 1}/{iter_per_epoch}, '- f'Actor Loss: {policy_loss_val:.4f}, Reward: {avg_reward_val:.4f}, '+ f'Actor Loss: {policy_loss_val:.6f}, Reward: {avg_reward_val:.6f}, 'f'Baseline: {avg_baseline_val:.4f}, KL: {kl_val:.4f}, Rho: {rho:.4f}, Avg Response Len: {avg_len_val:.2f}, LR: {current_lr:.2e}')if wandb and (not ddp or dist.get_rank() == 0):