🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.
[fix] align log/save last-step check and ETA with 1-indexed step
readlnhcf4b49a3482a
diff --git a/trainer/train_distillation.py b/trainer/train_distillation.pyindex e0d160f..dfbac0c 100644--- a/trainer/train_distillation.py+++ b/trainer/train_distillation.py@@ -98,13 +98,13 @@ def train_epoch(epoch, loader, iters, teacher_model, lm_config_student, start_stscaler.update()optimizer.zero_grad(set_to_none=True)- if step % args.log_interval == 0 or step == iters - 1:+ if step % args.log_interval == 0 or step == iters:spend_time = time.time() - start_timecurrent_loss = loss.item() * args.accumulation_stepscurrent_ce_loss = ce_loss_raw.item()current_aux_loss = res.aux_loss.item() if lm_config_student.use_moe else 0.0current_lr = optimizer.param_groups[-1]['lr']- eta_min = spend_time / (step + 1) * iters // 60 - spend_time // 60+ eta_min = spend_time / step * iters // 60 - spend_time // 60Logger(f'Epoch:[{epoch + 1}/{args.epochs}]({step}/{iters}), loss: {current_loss:.4f}, ce: {current_ce_loss:.4f}, aux_loss: {current_aux_loss:.4f}, distill: {distill_loss.item():.4f}, learning_rate: {current_lr:.8f}, epoch_time: {eta_min:.3f}min')@@ -118,7 +118,7 @@ def train_epoch(epoch, loader, iters, teacher_model, lm_config_student, start_st"epoch_time": eta_min})- if (step % args.save_interval == 0 or step == iters - 1) and is_main_process():+ if (step % args.save_interval == 0 or step == iters) and is_main_process():model.eval()moe_suffix = '_moe' if lm_config_student.use_moe else ''ckp = f'{args.save_dir}/{args.save_weight}_{lm_config_student.hidden_size}{moe_suffix}.pth'diff --git a/trainer/train_dpo.py b/trainer/train_dpo.pyindex 977953d..dcfc350 100644--- a/trainer/train_dpo.py+++ b/trainer/train_dpo.py@@ -92,19 +92,19 @@ def train_epoch(epoch, loader, iters, ref_model, lm_config, start_step=0, wandb=scaler.update()optimizer.zero_grad(set_to_none=True)- if step % args.log_interval == 0 or step == iters - 1:+ if step % args.log_interval == 0 or step == iters:spend_time = time.time() - start_timecurrent_loss = loss.item() * args.accumulation_stepscurrent_dpo_loss = dpo_loss_val.item()current_aux_loss = outputs.aux_loss.item()current_lr = optimizer.param_groups[-1]['lr']- eta_min = spend_time / (step + 1) * iters // 60 - spend_time // 60+ eta_min = spend_time / step * iters // 60 - spend_time // 60Logger(f'Epoch:[{epoch + 1}/{args.epochs}]({step}/{iters}), loss: {current_loss:.4f}, dpo_loss: {current_dpo_loss:.4f}, aux_loss: {current_aux_loss:.4f}, learning_rate: {current_lr:.8f}, epoch_time: {eta_min:.3f}min')if wandb: wandb.log({"loss": current_loss, "dpo_loss": current_dpo_loss, "aux_loss": current_aux_loss, "learning_rate": current_lr, "epoch_time": eta_min})- if (step % args.save_interval == 0 or step == iters - 1) and is_main_process():+ if (step % args.save_interval == 0 or step == iters) and is_main_process():model.eval()moe_suffix = '_moe' if lm_config.use_moe else ''ckp = f'{args.save_dir}/{args.save_weight}_{lm_config.hidden_size}{moe_suffix}.pth'diff --git a/trainer/train_full_sft.py b/trainer/train_full_sft.pyindex 16fadea..43b2dd2 100644--- a/trainer/train_full_sft.py+++ b/trainer/train_full_sft.py@@ -45,17 +45,17 @@ def train_epoch(epoch, loader, iters, start_step=0, wandb=None):optimizer.zero_grad(set_to_none=True)- if step % args.log_interval == 0 or step == iters - 1:+ if step % args.log_interval == 0 or step == iters:spend_time = time.time() - start_timecurrent_loss = loss.item() * args.accumulation_stepscurrent_aux_loss = res.aux_loss.item() if res.aux_loss is not None else 0.0current_logits_loss = current_loss - current_aux_losscurrent_lr = optimizer.param_groups[-1]['lr']- eta_min = spend_time / (step + 1) * iters // 60 - spend_time // 60+ eta_min = spend_time / step * iters // 60 - spend_time // 60Logger(f'Epoch:[{epoch + 1}/{args.epochs}]({step}/{iters}), loss: {current_loss:.4f}, logits_loss: {current_logits_loss:.4f}, aux_loss: {current_aux_loss:.4f}, lr: {current_lr:.8f}, epoch_time: {eta_min:.1f}min')if wandb: wandb.log({"loss": current_loss, "logits_loss": current_logits_loss, "aux_loss": current_aux_loss, "learning_rate": current_lr, "epoch_time": eta_min})- if (step % args.save_interval == 0 or step == iters - 1) and is_main_process():+ if (step % args.save_interval == 0 or step == iters) and is_main_process():model.eval()moe_suffix = '_moe' if lm_config.use_moe else ''ckp = f'{args.save_dir}/{args.save_weight}_{lm_config.hidden_size}{moe_suffix}.pth'diff --git a/trainer/train_grpo.py b/trainer/train_grpo.pyindex c7c122d..dd95e0b 100755--- a/trainer/train_grpo.py+++ b/trainer/train_grpo.py@@ -176,7 +176,7 @@ def grpo_train_epoch(epoch, loader, iters, ref_model, reward_model, reward_token"learning_rate": current_lr})- if (step % args.save_interval == 0 or step == iters - 1) and is_main_process():+ if (step % args.save_interval == 0 or step == iters) and is_main_process():model.eval()moe_suffix = '_moe' if lm_config.use_moe else ''ckp = f'{args.save_dir}/{args.save_weight}_{lm_config.hidden_size}{moe_suffix}.pth'diff --git a/trainer/train_lora.py b/trainer/train_lora.pyindex fb1b1b1..fc66cde 100644--- a/trainer/train_lora.py+++ b/trainer/train_lora.py@@ -44,17 +44,17 @@ def train_epoch(epoch, loader, iters, lora_params, start_step=0, wandb=None):scaler.update()optimizer.zero_grad(set_to_none=True)- if step % args.log_interval == 0 or step == iters - 1:+ if step % args.log_interval == 0 or step == iters:spend_time = time.time() - start_timecurrent_loss = loss.item() * args.accumulation_stepscurrent_aux_loss = res.aux_loss.item() if res.aux_loss is not None else 0.0current_logits_loss = current_loss - current_aux_losscurrent_lr = optimizer.param_groups[-1]['lr']- eta_min = spend_time / (step + 1) * iters // 60 - spend_time // 60+ eta_min = spend_time / step * iters // 60 - spend_time // 60Logger(f'Epoch:[{epoch + 1}/{args.epochs}]({step}/{iters}), loss: {current_loss:.4f}, logits_loss: {current_logits_loss:.4f}, aux_loss: {current_aux_loss:.4f}, lr: {current_lr:.8f}, epoch_time: {eta_min:.1f}min')if wandb: wandb.log({"loss": current_loss, "logits_loss": current_logits_loss, "aux_loss": current_aux_loss, "learning_rate": current_lr, "epoch_time": eta_min})- if (step % args.save_interval == 0 or step == iters - 1) and is_main_process():+ if (step % args.save_interval == 0 or step == iters) and is_main_process():model.eval()lora_save_path = f'{args.save_dir}/{args.lora_name}_{lm_config.hidden_size}.pth'# LoRA只保存LoRA权重diff --git a/trainer/train_ppo.py b/trainer/train_ppo.pyindex 4925651..f7e06f6 100644--- a/trainer/train_ppo.py+++ b/trainer/train_ppo.py@@ -226,7 +226,7 @@ def ppo_train_epoch(epoch, loader, iters, old_actor_model, ref_model, actor_scheold_actor_model.load_state_dict({k: v.detach().cpu() for k, v in state_dict.items()})old_actor_model.to(args.device)- if (step % args.save_interval == 0 or step == iters - 1) and is_main_process():+ if (step % args.save_interval == 0 or step == iters) and is_main_process():actor_model.eval()moe_suffix = '_moe' if lm_config.use_moe else ''ckp = f'{args.save_dir}/{args.save_weight}_{lm_config.hidden_size}{moe_suffix}.pth'diff --git a/trainer/train_pretrain.py b/trainer/train_pretrain.pyindex 1b6347d..a1b12b2 100644--- a/trainer/train_pretrain.py+++ b/trainer/train_pretrain.py@@ -45,17 +45,17 @@ def train_epoch(epoch, loader, iters, start_step=0, wandb=None):optimizer.zero_grad(set_to_none=True)- if step % args.log_interval == 0 or step == iters - 1:+ if step % args.log_interval == 0 or step == iters:spend_time = time.time() - start_timecurrent_loss = loss.item() * args.accumulation_stepscurrent_aux_loss = res.aux_loss.item() if res.aux_loss is not None else 0.0current_logits_loss = current_loss - current_aux_losscurrent_lr = optimizer.param_groups[-1]['lr']- eta_min = spend_time / (step + 1) * iters // 60 - spend_time // 60+ eta_min = spend_time / step * iters // 60 - spend_time // 60Logger(f'Epoch:[{epoch + 1}/{args.epochs}]({step}/{iters}), loss: {current_loss:.4f}, logits_loss: {current_logits_loss:.4f}, aux_loss: {current_aux_loss:.4f}, lr: {current_lr:.8f}, epoch_time: {eta_min:.1f}min')if wandb: wandb.log({"loss": current_loss, "logits_loss": current_logits_loss, "aux_loss": current_aux_loss, "learning_rate": current_lr, "epoch_time": eta_min})- if (step % args.save_interval == 0 or step == iters - 1) and is_main_process():+ if (step % args.save_interval == 0 or step == iters) and is_main_process():model.eval()moe_suffix = '_moe' if lm_config.use_moe else ''ckp = f'{args.save_dir}/{args.save_weight}_{lm_config.hidden_size}{moe_suffix}.pth'diff --git a/trainer/train_reason.py b/trainer/train_reason.pyindex c2bae0f..9edcdc1 100644--- a/trainer/train_reason.py+++ b/trainer/train_reason.py@@ -63,17 +63,17 @@ def train_epoch(epoch, loader, iters, tokenizer, lm_config, start_step=0, wandb=scaler.update()optimizer.zero_grad(set_to_none=True)- if step % args.log_interval == 0 or step == iters - 1:+ if step % args.log_interval == 0 or step == iters:spend_time = time.time() - start_timecurrent_loss = loss.item() * args.accumulation_stepscurrent_aux_loss = res.aux_loss.item() if res.aux_loss is not None else 0.0current_logits_loss = logits_loss.item()current_lr = optimizer.param_groups[-1]['lr']- eta_min = spend_time / (step + 1) * iters // 60 - spend_time // 60+ eta_min = spend_time / step * iters // 60 - spend_time // 60Logger(f'Epoch:[{epoch + 1}/{args.epochs}]({step}/{iters}), loss: {current_loss:.4f}, logits_loss: {current_logits_loss:.4f}, aux_loss: {current_aux_loss:.4f}, lr: {current_lr:.8f}, epoch_time: {eta_min:.1f}min')if wandb: wandb.log({"loss": current_loss, "logits_loss": current_logits_loss, "aux_loss": current_aux_loss, "learning_rate": current_lr, "epoch_time": eta_min})- if (step % args.save_interval == 0 or step == iters - 1) and is_main_process():+ if (step % args.save_interval == 0 or step == iters) and is_main_process():model.eval()moe_suffix = '_moe' if lm_config.use_moe else ''ckp = f'{args.save_dir}/{args.save_weight}_{lm_config.hidden_size}{moe_suffix}.pth'diff --git a/trainer/train_spo.py b/trainer/train_spo.pyindex 595bc50..6b89976 100755--- a/trainer/train_spo.py+++ b/trainer/train_spo.py@@ -224,7 +224,7 @@ def spo_train_epoch(epoch, loader, iters, ref_model, reward_model, reward_tokeni"learning_rate": current_lr})- if (step % args.save_interval == 0 or step == iters - 1) and is_main_process():+ if (step % args.save_interval == 0 or step == iters) and is_main_process():model.eval()moe_suffix = '_moe' if lm_config.use_moe else ''ckp = f'{args.save_dir}/{args.save_weight}_{lm_config.hidden_size}{moe_suffix}.pth'