fix: 修复一些运行时问题

.gitignore

@@ -1,15 +1,12 @@
-# cache
-__pycache__
-.pytest_cache
+# Ignore everything
+*
 
-# params
-params/*
+# Allow directories to be traversed
+!*/
 
-# vscode file
-.vscode
-
-# build file
-build
-*.egg-info
-
-*.ipynb
+# Allow specific file types and root files
+!*.py
+!*.md
+!*.png
+!LICENSE
+!pyproject.toml

khaosz/data/checkpoint.py

@@ -12,14 +12,12 @@ from khaosz.parallel.setup import get_rank
 class Checkpoint:
     def __init__(
         self,
-        optimizer_state_dict: Dict[str, Any],
-        scheduler_state_dict: Optional[Dict[str, Any]] = None,
+        state_dict: Dict[str, Any],
         epoch: int = 0,
         iteration: int = 0,
         metrics: Optional[Dict[str, list]] = None,
     ):
-        self.optimizer_state_dict = optimizer_state_dict
-        self.scheduler_state_dict = scheduler_state_dict
+        self.state_dict = state_dict
         self.epoch = epoch
         self.iteration = iteration
         self.metrics = metrics or {}
@@ -46,12 +44,8 @@ class Checkpoint:
             if save_metric_plot and self.metrics:
                 self._plot_metrics(str(save_path))
 
-        state_dict = {
-            "optimizer": self.optimizer_state_dict,
-            "scheduler": self.scheduler_state_dict
-        }
-        with open(save_path / f"state_dict_rank_{get_rank()}.pt", "wb") as f:
-            torch.save(state_dict, f)
+            with open(save_path / f"state_dict.pt", "wb") as f:
+                torch.save(self.state_dict, f)
 
     @classmethod
     def load(
@@ -72,7 +66,7 @@ class Checkpoint:
             dist.broadcast_object_list(meta_list, src=0)
             meta = meta_list[0]
 
-        with open(save_path / f"state_dict_rank_{get_rank()}.pt", "rb") as f:
+        with open(save_path / f"state_dict.pt", "rb") as f:
             state_dict = torch.load(f)
 
         return cls(

khaosz/data/dataset.py

@@ -1,4 +1,3 @@
-import h5py
 import torch
 import bisect
 
@@ -78,8 +77,10 @@ class BaseDataset(Dataset, ABC):
         self.fetcher = MultiSegmentFetcher(self.segments)
         
     def get_index(self, index: int) -> int:
-        begin_idx = min(index * self.stride, self.total_samples - self.window_size - 1)
-        end_idx = begin_idx + self.window_size
+        assert self.total_samples > self.window_size
+        
+        begin_idx = min(index * self.stride, self.total_samples - 1 - self.window_size)
+        end_idx = min(begin_idx + self.window_size, self.total_samples - 1)
         
         return begin_idx, end_idx
         
@@ -91,7 +92,7 @@ class BaseDataset(Dataset, ABC):
         assert self.total_samples is not None
         if self.total_samples <= self.window_size:
             return 0
-        return self.total_samples // self.stride + 1
+        return (self.total_samples - 1 - self.window_size) // self.stride + 1
     
 
 class SeqDataset(BaseDataset):

khaosz/data/file.py

@@ -2,6 +2,8 @@ import os
 import h5py
 import numpy as np
 import torch
+
+from pathlib import Path
 from torch import Tensor
 from typing import Dict, List, Tuple
 
@@ -17,10 +19,7 @@ def save_h5(file_path: str, tensor_group: Dict[str, List[Tensor]]):
                 arr = tensor.cpu().numpy()
                 dset = grp.create_dataset(
                     f'data_{idx}',
-                    data=arr,
-                    compression='gzip',
-                    compression_opts=4,
-                    shuffle=True
+                    data=arr
                 )
                 dset.attrs['numel'] = tensor.numel()
 
@@ -28,15 +27,19 @@ def load_h5(file_path: str) -> Tuple[Dict[str, List[Tensor]], int]:
     tensor_group: Dict[str, List[Tensor]] = {}
     total_samples = 0
 
-    with h5py.File(file_path, 'r') as f:
-        for key in f.keys():
-            grp = f[key]
-            dsets = []
-            for dset_name in grp.keys():
-                dset = grp[dset_name]
-                dsets.append(torch.from_numpy(dset[:]).share_memory_())
-                total_samples += dset.attrs.get('numel', np.prod(dset.shape))
-            tensor_group[key] = dsets
+    root_path = Path(file_path)
+    h5_files = list(root_path.rglob("*.h5")) + list(root_path.rglob("*.hdf5"))
+    
+    for h5_file in h5_files:
+        with h5py.File(h5_file, 'r') as f:
+            for key in f.keys():
+                grp = f[key]
+                dsets = []
+                for dset_name in grp.keys():
+                    dset = grp[dset_name]
+                    dsets.append(torch.from_numpy(dset[:]).share_memory_())
+                    total_samples += dset.attrs.get('numel', np.prod(dset.shape))
+                tensor_group[key] = dsets
 
     num_keys = max(len(tensor_group), 1)
     sample_per_key = total_samples // num_keys

khaosz/trainer/schedule.py

@@ -151,8 +151,8 @@ class SchedulerFactory:
     """
 
     @staticmethod
-    def load(optimizer, scedule_config: ScheduleConfig) -> BaseScheduler:
-        kwargs = scedule_config.get_kwargs()
+    def load(optimizer, schedule_config: ScheduleConfig) -> BaseScheduler:
+        kwargs = schedule_config.get_kwargs()
         schedule_type = kwargs.pop("schedule_type")
         
         if schedule_type == "cosine":

khaosz/trainer/train_callback.py

@@ -108,10 +108,10 @@ class CheckpointCallback(TrainCallback):
     
     def _save_checkpoint(self, context: 'TrainContext'):
         save_path = os.path.join(self.save_dir, f"epoch_{context.epoch}_iter_{context.iteration}")
-        state_dict = self.state_dict_fn(context.model) if self.state_dict_fn else context.optimizer.state_dict()
+        state_dict = self.state_dict_fn(context.model) if self.state_dict_fn else context.model.state_dict()
+        
         context.checkpoint = Checkpoint(
-            optimizer_state_dict=state_dict,
-            scheduler_state_dict=context.scheduler.state_dict() if context.scheduler else None,
+            state_dict=state_dict,
             epoch=context.epoch,
             iteration=context.iteration
         )

khaosz/trainer/train_context.py

@@ -53,15 +53,13 @@ class TrainContextBuilder:
     def with_checkpoint(self, checkpoint: Optional[Checkpoint]) -> Self:
         if checkpoint is None:
             checkpoint = Checkpoint(
-                optimizer_state_dict=self._context.optimizer.state_dict(),
-                scheduler_state_dict=self._context.scheduler.state_dict(),
+                state_dict=self._context.model.state_dict(),
             )
         else:
             # resume from the assigned checkpoint or assigned iteration
             self._context.epoch = max(checkpoint.epoch, self.config.start_epoch)
             self._context.iteration = max(checkpoint.iteration, self.config.start_batch)
-            self._context.optimizer.load_state_dict(checkpoint.optimizer_state_dict)
-            self._context.scheduler.load_state_dict(checkpoint.scheduler_state_dict)
+            self._context.model.load_state_dict(checkpoint.state_dict)
         
         self._context.checkpoint = checkpoint
         return self

tools/train.py

@@ -3,10 +3,8 @@ import argparse
 import torch
 import torch.nn as nn
 import torch.optim as optim
-import torch.distributed.fsdp as fsdp
+from torch.nn.parallel import DistributedDataParallel as DDP
 
-from torch.distributed.fsdp.api import StateDictType
-from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
 from typing import List, Optional
 from functools import partial
 from khaosz.config import ModelParameter, TrainConfig, CosineScheduleConfig
@@ -59,19 +57,16 @@ def parse_args() -> argparse.Namespace:
 
     return args
 
-def fsdp_wrap(model: nn.Module):
-    
-    fsdp_model = fsdp.FullyShardedDataParallel(
+def ddp_wrap(model: nn.Module):
+    local_rank = int(os.environ.get("LOCAL_RANK", 0))
+    model = model.to(device=f"cuda:{local_rank}", dtype=torch.bfloat16)
+    ddp_model = DDP(
         model,
-        sharding_strategy=fsdp.ShardingStrategy.SHARD_GRAD_OP,
-        mixed_precision=fsdp.MixedPrecision(
-            param_dtype=torch.bfloat16,
-            reduce_dtype=torch.bfloat16,
-            buffer_dtype=torch.bfloat16,
-        ),
-        backward_prefetch=fsdp.BackwardPrefetch.BACKWARD_PRE
+        device_ids=[local_rank],
+        output_device=local_rank,
+        find_unused_parameters=False
     )
-    return fsdp_model
+    return ddp_model
 
 def create_optimizer(model: nn.Module, **kwargs) -> optim.Optimizer:
     return optim.AdamW(model.parameters(), **kwargs)
@@ -79,12 +74,13 @@ def create_optimizer(model: nn.Module, **kwargs) -> optim.Optimizer:
 def create_scheduler(optimizer: optim.Optimizer, **kwargs) -> optim.lr_scheduler.LRScheduler:
     return SchedulerFactory.load(optimizer, **kwargs)
 
-def prepare_checkpoint(model: nn.Module, optimizer: optim.Optimizer) -> dict:
-    with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT):
-        model_state_dict = model.state_dict()
-        optim_state_dict = FSDP.optim_state_dict(model, optimizer)
+def prepare_checkpoint(model: nn.Module) -> dict:
+    if isinstance(model, torch.nn.parallel.DistributedDataParallel):
+        state_dict = model.module.state_dict()
+    else:
+        state_dict = model.state_dict()
+    return state_dict
 
-        return model_state_dict, optim_state_dict
 
 def train(
     train_type: str,
@@ -166,7 +162,7 @@ def train(
         num_workers=num_workers,
         pin_memory=pin_memory,
         nprocs=nprocs,
-        parallel_wrapper=fsdp_wrap,
+        parallel_wrapper=ddp_wrap,
         state_dict_fn=prepare_checkpoint,
         device_ids=device_ids,
         device_type=device_type,