feat(train): 支持分布式训练的优化器与调度器工厂配置

khaosz/config/train_config.py

@@ -1,4 +1,4 @@
-from torch import nn
+import torch.nn as nn
 from torch.utils.data import Dataset
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LRScheduler
@@ -30,6 +30,7 @@ class TrainConfig:
         default=None,
         metadata={"help": "Scheduler for training."}
     )
+    
     n_epoch: int = field(
         default=1,
         metadata={"help": "Number of epochs for training."}
@@ -104,6 +105,14 @@ class TrainConfig:
         default=None,
         metadata={"help": "Parallel function for training."}
     )
+    optimizer_factory: Optional[Callable[[nn.Module], Optimizer]] = field(
+        default=None,
+        metadata={"help": "Optimizer factory for training."}
+    )
+    scheduler_factory: Optional[Callable[[Optimizer], LRScheduler]] = field(
+        default=None,
+        metadata={"help": "Scheduler factory for training."}
+    )
 
     # others
     extra_kwargs: dict = field(
@@ -115,7 +124,17 @@ class TrainConfig:
         self.validate()
     
     def validate(self):
-        required_fields = ["model", "strategy", "dataset", "optimizer", "scheduler"]
+        required_fields = ["model", "strategy", "dataset"]
+        
         for field_name in required_fields:
             if getattr(self, field_name) is None:
                 raise ValueError(f"{field_name} is required.")
+        
+        factory_case = all([self.optimizer_factory, self.scheduler_factory])
+        argument_case = all([self.optimizer, self.scheduler])
+        self.nprocs = max(self.nprocs, 1)
+        
+        if self.nprocs > 1 and not factory_case:
+            raise ValueError("Distributed training requires optimizer and scheduler factories.")
+        elif self.nprocs == 1 and not argument_case:
+            raise ValueError("Single process training requires optimizer and scheduler arguments.")

khaosz/trainer/train_context.py

@@ -80,19 +80,27 @@ class TrainContextBuilder:
         return self
     
     def with_strategy(self) -> Self:
-        device = get_current_device()
         self._context.strategy = StrategyFactory.load(
             model=self.config.model,
             train_type=self.config.strategy,
-            device=device,
+            device=get_current_device(),
             **self.config.extra_kwargs
         )
         return self
     
     def with_parallel_fn(self) -> Self:
+        device = get_current_device()
+        self._context.model = self._context.model.to(device=device)
+        
+        if self.config.nprocs > 1:
+            
             fn = self.config.parallel_fn
-        if fn is not None:
+            optimizer_fn = self.config.optimizer_factory
+            scheduler_fn = self.config.scheduler_factory
+            
             self._context.model = fn(self._context.model)
+            self._context.optimizer = optimizer_fn(self._context.model.parameters())
+            self._context.scheduler = scheduler_fn(self._context.optimizer)
             
         return self
     

tools/train.py

@@ -2,9 +2,10 @@ import os
 import argparse
 import torch
 import torch.nn as nn
+import torch.optim as optim
 import torch.distributed.fsdp as fsdp
 
-from torch.optim import AdamW
+from functools import partial
 from khaosz.config import ModelParameter, TrainConfig, CosineScheduleConfig
 from khaosz.trainer import Trainer, SchedulerFactory
 from khaosz.data import DatasetLoader
@@ -26,7 +27,6 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--adamw_beta1", type=float, default=0.9, help="Beta values for AdamW optimizer.")
     parser.add_argument("--adamw_beta2", type=float, default=0.95, help="Beta values for AdamW optimizer.")
     parser.add_argument("--adamw_weight_decay", type=float, default=0.01, help="Weight decay for AdamW optimizer.")
-    parser.add_argument("--embdeding_lr_rate", type=float, default=1.0, help="The rate between the embedding layers lr rate and the max lr rate.")
     parser.add_argument("--random_seed", type=int, default=3407, help="Random seed for reproducibility.")
     parser.add_argument("--num_workers", type=int, default=4, help="Number of workers for data loading.")
     parser.add_argument("--no_pin_memory", action="store_false", dest="pin_memory", help="Disable pin memory")
@@ -59,6 +59,12 @@ def fsdp_wrap(model: nn.Module):
     )
     return fsdp_model
 
+def create_optimizer(model: nn.Module, **kwargs) -> optim.Optimizer:
+    return optim.AdamW(model.parameters(), **kwargs)
+
+def create_scheduler(optimizer: optim.Optimizer, **kwargs) -> optim.lr_scheduler.LRScheduler:
+    return SchedulerFactory.load(optimizer, **kwargs)
+
 def train(
     train_type: str,
     param_path: str,
@@ -77,7 +83,6 @@ def train(
     adamw_beta2: float,
     adamw_weight_decay: float,
     max_grad_norm: float,
-    embdeding_lr_rate: int,
     random_seed: int,
     num_workers: int,
     pin_memory: bool,
@@ -110,23 +115,19 @@ def train(
         stride=stride
     )
     
-    param_groups = [
-        {"params": [p for n, p in model.named_parameters() if "embed" in n], "lr": max_lr * embdeding_lr_rate},
-        {"params": [p for n, p in model.named_parameters() if "embed" not in n], "lr": max_lr}
-    ]
-
-    optimizer = AdamW(
-        param_groups,
-        betas=(adamw_beta1, adamw_beta2),
-        weight_decay=adamw_weight_decay
-    )
-    
     schedule_config = CosineScheduleConfig(
         warmup_steps=warmup_steps,
         total_steps=len(dataset) * n_epoch // (batch_size * nprocs), 
     )
     
-    scheduler = SchedulerFactory.load(optimizer, schedule_config)
+
+    optimizer_fn = partial(create_optimizer, **{"lr": max_lr, "betas": (adamw_beta1, adamw_beta2), "weight_decay": adamw_weight_decay})
+    scheduler_fn = partial(create_scheduler, **{"schedule_config": schedule_config})
+    optimizer, scheduler = None, None
+    
+    if nprocs == 1:
+        optimizer = optimizer_fn(model.parameters())
+        scheduler = scheduler_fn(optimizer)
     
     train_config = TrainConfig(
         model=model,
@@ -146,6 +147,8 @@ def train(
         num_workers=num_workers,
         pin_memory=pin_memory,
         nprocs=nprocs,
+        optimizer_factory=optimizer_fn,
+        scheduler_factory=scheduler_fn,
         extra_kwargs=kwargs,
         parallel_fn=fsdp_wrap
     )