🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.
[fix] graph-oom & ddp-pos_cis
jingyaogongfa7dff82916c
diff --git a/trainer/train_distill_reason.py b/trainer/train_distill_reason.pyindex 7bd6bce..ab7866c 100644--- a/trainer/train_distill_reason.py+++ b/trainer/train_distill_reason.py@@ -89,7 +89,7 @@ def train_epoch(epoch, wandb):spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60))if (wandb is not None) and (not ddp or dist.get_rank() == 0):- wandb.log({"loss": loss * args.accumulation_steps,+ wandb.log({"loss": loss.item() * args.accumulation_steps,"lr": optimizer.param_groups[-1]['lr'],"epoch_Time": spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60})@@ -208,7 +208,7 @@ if __name__ == "__main__":optimizer = optim.AdamW(model.parameters(), lr=args.learning_rate)if ddp:- model._ddp_params_and_buffers_to_ignore = {"pos_cis"}+ model._ddp_params_and_buffers_to_ignore = {"freqs_cos", "freqs_sin"}model = DistributedDataParallel(model, device_ids=[ddp_local_rank])iter_per_epoch = len(train_loader)diff --git a/trainer/train_distillation.py b/trainer/train_distillation.pyindex cccfee1..f4e2d41 100644--- a/trainer/train_distillation.py+++ b/trainer/train_distillation.py@@ -257,7 +257,7 @@ if __name__ == "__main__":optimizer = optim.AdamW(model.parameters(), lr=args.learning_rate)if ddp:- model._ddp_params_and_buffers_to_ignore = {"pos_cis"}+ model._ddp_params_and_buffers_to_ignore = {"freqs_cos", "freqs_sin"}model = DistributedDataParallel(model, device_ids=[ddp_local_rank])iter_per_epoch = len(train_loader)diff --git a/trainer/train_dpo.py b/trainer/train_dpo.pyindex db17164..c4acf36 100644--- a/trainer/train_dpo.py+++ b/trainer/train_dpo.py@@ -113,7 +113,7 @@ def train_epoch(epoch, wandb):spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60))if (wandb is not None) and (not ddp or dist.get_rank() == 0):- wandb.log({"loss": loss * args.accumulation_steps,+ wandb.log({"loss": loss.item() * args.accumulation_steps,"lr": optimizer.param_groups[-1]['lr'],"epoch_Time": spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60})@@ -240,7 +240,7 @@ if __name__ == "__main__":optimizer = optim.AdamW(model.parameters(), lr=args.learning_rate)if ddp:- model._ddp_params_and_buffers_to_ignore = {"pos_cis"}+ model._ddp_params_and_buffers_to_ignore = {"freqs_cos", "freqs_sin"}model = DistributedDataParallel(model, device_ids=[ddp_local_rank])iter_per_epoch = len(train_loader)diff --git a/trainer/train_full_sft.py b/trainer/train_full_sft.pyindex 49a3ff0..58f87aa 100644--- a/trainer/train_full_sft.py+++ b/trainer/train_full_sft.py@@ -76,7 +76,7 @@ def train_epoch(epoch, wandb):spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60))if (wandb is not None) and (not ddp or dist.get_rank() == 0):- wandb.log({"loss": loss * args.accumulation_steps,+ wandb.log({"loss": loss.item() * args.accumulation_steps,"lr": optimizer.param_groups[-1]['lr'],"epoch_Time": spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60})@@ -194,7 +194,7 @@ if __name__ == "__main__":optimizer = optim.AdamW(model.parameters(), lr=args.learning_rate)if ddp:- model._ddp_params_and_buffers_to_ignore = {"pos_cis"}+ model._ddp_params_and_buffers_to_ignore = {"freqs_cos", "freqs_sin"}model = DistributedDataParallel(model, device_ids=[ddp_local_rank])iter_per_epoch = len(train_loader)diff --git a/trainer/train_grpo.py b/trainer/train_grpo.pyindex a3e943c..0402f41 100755--- a/trainer/train_grpo.py+++ b/trainer/train_grpo.py@@ -309,7 +309,7 @@ if __name__ == "__main__":scheduler = CosineAnnealingLR(optimizer, T_max=total_optimizer_steps, eta_min=args.learning_rate / 10)if ddp:- model._ddp_params_and_buffers_to_ignore = {"pos_cis"}+ model._ddp_params_and_buffers_to_ignore = {"freqs_cos", "freqs_sin"}model = DistributedDataParallel(model, device_ids=[ddp_local_rank])for epoch in range(args.epochs):diff --git a/trainer/train_lora.py b/trainer/train_lora.pyindex 7d74b9f..5da1607 100644--- a/trainer/train_lora.py+++ b/trainer/train_lora.py@@ -77,7 +77,7 @@ def train_epoch(epoch, wandb):spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60))if (wandb is not None) and (not ddp or dist.get_rank() == 0):- wandb.log({"loss": loss * args.accumulation_steps,+ wandb.log({"loss": loss.item() * args.accumulation_steps,"lr": optimizer.param_groups[-1]['lr'],"epoch_Time": spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60})diff --git a/trainer/train_ppo.py b/trainer/train_ppo.pyindex dd67775..1132da9 100644--- a/trainer/train_ppo.py+++ b/trainer/train_ppo.py@@ -359,8 +359,8 @@ if __name__ == "__main__":# 如果使用分布式训练,包装模型if ddp:- actor_model._ddp_params_and_buffers_to_ignore = {"pos_cis"}- critic_model._ddp_params_and_buffers_to_ignore = {"pos_cis"}+ actor_model._ddp_params_and_buffers_to_ignore = {"freqs_cos", "freqs_sin"}+ critic_model._ddp_params_and_buffers_to_ignore = {"freqs_cos", "freqs_sin"}actor_model = DistributedDataParallel(actor_model, device_ids=[ddp_local_rank])critic_model = DistributedDataParallel(critic_model, device_ids=[ddp_local_rank])# old_actor_model 不需要DDP包装,因为它只在主进程上用于计算,并且不进行梯度更新diff --git a/trainer/train_pretrain.py b/trainer/train_pretrain.pyindex 6eed5a7..4166db5 100644--- a/trainer/train_pretrain.py+++ b/trainer/train_pretrain.py@@ -192,7 +192,7 @@ if __name__ == "__main__":optimizer = optim.AdamW(model.parameters(), lr=args.learning_rate)if ddp:- model._ddp_params_and_buffers_to_ignore = {"pos_cis"}+ model._ddp_params_and_buffers_to_ignore = {"freqs_cos", "freqs_sin"}model = DistributedDataParallel(model, device_ids=[ddp_local_rank])iter_per_epoch = len(train_loader)diff --git a/trainer/train_spo.py b/trainer/train_spo.pyindex 7e44d3a..8db67b9 100755--- a/trainer/train_spo.py+++ b/trainer/train_spo.py@@ -358,7 +358,7 @@ if __name__ == "__main__":scheduler = CosineAnnealingLR(optimizer, T_max=total_optimizer_steps, eta_min=args.learning_rate / 10)if ddp:- model._ddp_params_and_buffers_to_ignore = {"pos_cis"}+ model._ddp_params_and_buffers_to_ignore = {"freqs_cos", "freqs_sin"}model = DistributedDataParallel(model, device_ids=[ddp_local_rank])value_tracker = AutoAdaptiveValueTracker(rho_mode='kl', rho_const=0.9, D_half=0.06, clip_lower=0.5, clip_upper=0.96)