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.

Merge branch 'master' into wandb

jingyaogong5dd4e15aa25b
diff --cc 1-pretrain.py
index d6bfff6,50fee2a..ad1a45c
--- a/1-pretrain.py
+++ b/1-pretrain.py
@@@ -68,14 -67,15 +68,15 @@@ def train_epoch(epoch, wandb)
Logger(
'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.7f} epoch_Time:{}min:'.format(
epoch,
- epochs,
+ args.epochs,
step,
iter_per_epoch,
- loss.item() * accumulation_steps,
+ loss.item() * args.accumulation_steps,
optimizer.param_groups[-1]['lr'],
spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60))
- if wandb is not None:
+
+ if (wandb is not None) and (not ddp or dist.get_rank() == 0):
- wandb.log({"loss": loss.item() * 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})
@@@ -115,40 -125,40 +116,42 @@@ def init_distributed_mode()
DEVICE = f"cuda:{ddp_local_rank}"
torch.cuda.set_device(DEVICE)
+
# torchrun --nproc_per_node 2 1-pretrain.py
-# I/O
if __name__ == "__main__":
- # -----------------------------------------------------------------------------
+ parser = argparse.ArgumentParser(description="MiniMind Pretraining")
+ parser.add_argument("--out_dir", type=str, default="out", help="Output directory")
+ parser.add_argument("--epochs", type=int, default=20, help="Number of epochs")
+ parser.add_argument("--batch_size", type=int, default=64, help="Batch size")
+ parser.add_argument("--learning_rate", type=float, default=2e-4, help="Learning rate")
+ parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu", help="Device to use")
+ parser.add_argument("--dtype", type=str, default="bfloat16", help="Data type")
+ parser.add_argument("--use_wandb", action="store_true", help="Use Weights & Biases")
+ parser.add_argument("--wandb_project", type=str, default="MiniMind-Pretrain", help="Weights & Biases project name")
+ parser.add_argument("--num_workers", type=int, default=8, help="Number of workers for data loading")
+ parser.add_argument("--data_path", type=str, default="./dataset/pretrain_data.bin", help="Path to training data")
+ parser.add_argument("--ddp", action="store_true", help="Use DistributedDataParallel")
+ parser.add_argument("--accumulation_steps", type=int, default=8, help="Gradient accumulation steps")
+ parser.add_argument("--grad_clip", type=float, default=1.0, help="Gradient clipping threshold")
+ parser.add_argument("--warmup_iters", type=int, default=0, help="Number of warmup iterations")
+ parser.add_argument("--log_interval", type=int, default=100, help="Logging interval")
+ parser.add_argument("--save_interval", type=int, default=1000, help="Model saving interval")
+
+ args = parser.parse_args()
+
lm_config = LMConfig()
max_seq_len = lm_config.max_seq_len
- out_dir = 'out'
- epochs = 20
- batch_size = 64
- learning_rate = 2e-4
- device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
- dtype = 'bfloat16'
- save_dir = os.path.join(out_dir)
- os.makedirs(save_dir, exist_ok=True)
- os.makedirs(out_dir, exist_ok=True)
- tokens_per_iter = batch_size * max_seq_len
+ args.save_dir = os.path.join(args.out_dir)
+ os.makedirs(args.save_dir, exist_ok=True)
+ os.makedirs(args.out_dir, exist_ok=True)
+ tokens_per_iter = args.batch_size * max_seq_len
torch.manual_seed(1337)
- device_type = device if "cuda" in device else "cpu"
+ device_type = "cuda" if "cuda" in args.device else "cpu"
- use_wandb = False # 是否使用wandb
- wandb_project = "MiniMind-Pretrain"
- wandb_run_name = f"MiniMind-Pretrain-Epoch-{epochs}-BatchSize-{batch_size}-LearningRate-{learning_rate}"
- if use_wandb:
- import wandb
- wandb.init(project=wandb_project, name=wandb_run_name)
- else:
- wandb = None
+ args.wandb_run_name = f"MiniMind-Pretrain-Epoch-{args.epochs}-BatchSize-{args.batch_size}-LearningRate-{args.learning_rate}"
+
+ ctx = nullcontext() if device_type == "cpu" else torch.cuda.amp.autocast()
+
- ctx = (
- nullcontext()
- if device_type == "cpu"
- else torch.cuda.amp.autocast()
- )
ddp = int(os.environ.get("RANK", -1)) != -1 # is this a ddp run?
ddp_local_rank, DEVICE = 0, "cuda:0"
if ddp:
diff --cc 3-full_sft.py
index fc58bb3,c413de0..d82e662
--- a/3-full_sft.py
+++ b/3-full_sft.py
@@@ -58,30 -59,35 +58,31 @@@ def train_epoch(epoch, wandb)
loss_mask = loss_mask.view(-1)
loss = torch.sum(loss * loss_mask) / loss_mask.sum()
- # Backward pass
scaler.scale(loss).backward()
- # Unscale gradients and clip them
- scaler.unscale_(optimizer)
- torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
+ if (step + 1) % args.accumulation_steps == 0:
+ scaler.unscale_(optimizer)
+ torch.nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
- # Update parameters
- scaler.step(optimizer)
- scaler.update()
+ scaler.step(optimizer)
+ scaler.update()
- # Zero the gradients
- optimizer.zero_grad(set_to_none=True)
+ optimizer.zero_grad(set_to_none=True)
- # 打印日志
- if step % 100 == 0:
+ if step % args.log_interval == 0:
spend_time = time.time() - start_time
Logger(
- 'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.8f} epoch_Time:{}min:'.format(
+ 'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.7f} epoch_Time:{}min:'.format(
epoch,
- epochs,
+ args.epochs,
step,
iter_per_epoch,
- loss,
+ loss.item(),
optimizer.param_groups[-1]['lr'],
spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60))
- if wandb is not None:
- wandb.log({"loss": loss.item(),
+
+ if (wandb is not None) and (not ddp or dist.get_rank() == 0):
+ wandb.log({"loss": loss,
"lr": optimizer.param_groups[-1]['lr'],
"epoch_Time": spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60})
diff --cc 4-lora_sft.py
index 44d69ba,ab8ba31..e72f8ca
--- a/4-lora_sft.py
+++ b/4-lora_sft.py
@@@ -102,10 -92,10 +102,10 @@@ def find_all_linear_names(model)
def init_model():
- model_name_or_path = "./minimind"
- tokenizer_name_or_path = "./minimind"
+ model_name_or_path = "./minimind-v1-small"
+ tokenizer_name_or_path = "./minimind-v1-small"
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path, trust_remote_code=True, use_fast=False)
- model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True).to(device)
+ model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True).to(args.device)
target_modules = find_all_linear_names(model)
peft_config = LoraConfig(