[checkpoint] add ColoOptimizer checkpointing (#1316)

colossalai/nn/optimizer/colossalai_optimizer.py

@@ -1,6 +1,3 @@
-#!/usr/bin/env python
-# -*- encoding: utf-8 -*-
-
 import torch
 import torch.nn as nn
 from torch import Tensor

colossalai/utils/checkpoint/module_checkpoint.py

@@ -1,12 +1,15 @@
 import torch
 import torch.distributed as dist
 from colossalai.tensor import ColoTensor, DistSpecManager
+from colossalai.nn.optimizer import ColossalaiOptimizer
+from copy import copy
+from typing import Optional
 
 
 def save_checkpoint(dire: str,
                     epoch: int,
                     model: torch.nn.Module,
-                    optimizer: torch.optim.Optimizer = None,
+                    optimizer: Optional[ColossalaiOptimizer] = None,
                     lr_scheduler: torch.optim.lr_scheduler._LRScheduler = None,
                     *args,
                     **kwargs):
@@ -16,7 +19,7 @@ def save_checkpoint(dire: str,
         dire (str): directory to save the checkpoint files.
         epoch (int): the number of epoch
         model (torch.nn.Module): a torch module initialized by ColoInitContext
-        optimizer (torch.optim.Optimizer, optional): optimizers. Defaults to None.
+        optimizer (ColossalaiOptimizer, optional): optimizers. Defaults to None.
         lr_scheduler (torch.optim.lr_scheduler._LRScheduler, optional): lr schedule. Defaults to None.
     """
 
@@ -41,11 +44,21 @@ def save_checkpoint(dire: str,
     # delete the new dict
     del new_dict
 
+    optim_state_copy = copy(optimizer.state_dict())
+    for k, v in optim_state_copy['state'].items():
+        for n, t in v.items():
+            if isinstance(t, ColoTensor):
+                t.to_replicate_()
+    if dist.get_rank() == 0:
+        model_state = {'epoch': epoch, 'optim': optim_state_copy}
+        torch.save(model_state, dire + '/epoch_{}_optim.pth'.format(epoch))
+    del optim_state_copy
+
 
 def load_checkpoint(dire,
                     epoch: int,
                     model: torch.nn.Module,
-                    optimizer: torch.optim.Optimizer = None,
+                    optimizer: Optional[ColossalaiOptimizer] = None,
                     lr_scheduler: torch.optim.lr_scheduler._LRScheduler = None,
                     *args,
                     **kwargs):
@@ -56,7 +69,7 @@ def load_checkpoint(dire,
         epoch (int): _description_
         rank (int): _description_
         model (torch.nn.Module): _description_
-        optimizer (torch.optim.Optimizer, optional): _description_. Defaults to None.
+        optimizer (ColossalaiOptimizer, optional): _description_. Defaults to None.
         lr_scheduler (torch.optim.lr_scheduler._LRScheduler, optional): _description_. Defaults to None.
     """
 
@@ -74,3 +87,24 @@ def load_checkpoint(dire,
         for k, v in model.state_dict().items():
             if isinstance(v, ColoTensor):
                 v.set_tensor_spec(*mapping[k])
+
+    del mapping
+    mapping = dict()
+
+    for k, v in optimizer.state_dict()['state'].items():
+        for n, t in v.items():
+            if isinstance(t, ColoTensor):
+                mapping[(k, n)] = (t.dist_spec, t.compute_spec)
+                t.to_replicate_()
+
+    colo_checkpoint = torch.load(dire + '/epoch_{}_optim.pth'.format(epoch))
+    optimizer.load_state_dict(colo_checkpoint['optim'])
+
+    for k, v in optimizer.state_dict()['state'].items():
+        for n, t in v.items():
+            if isinstance(t, ColoTensor):
+                # skip key not in mapping.
+                # For Adam, if it dose not execute step() once, there will be not exp_avg and exp_avg_sq in optimizer
+                if (k, n) not in mapping:
+                    continue
+                t.set_tensor_spec(*mapping[(k, n)])

tests/test_utils/test_colo_checkpoint.py

@@ -77,6 +77,18 @@ def remove(path):
         raise ValueError("file {} is not a file or dir.".format(path))
 
 
+def compare_optims(optim1, optim2):
+    state1 = optim1.state_dict()['state']
+    state2 = optim2.state_dict()['state']
+    for k, p1 in state1.items():
+        if k not in state2:
+            continue
+        p2 = state2[k]
+        if isinstance(p1, ColoTensor):
+            assert isinstance(p2, ColoTensor)
+            assert torch.allclose(p1.to_replicate_(), p2.to_replicate_(), rtol=1e-3, atol=1e-1)
+
+
 def _run_checkpoint(model_name, init_spec_func, use_ddp, use_mp_reload, test_scheduler, pg):
     get_components_func = non_distributed_component_funcs.get_callable(model_name)
     model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
@@ -117,7 +129,10 @@ def _run_checkpoint(model_name, init_spec_func, use_ddp, use_mp_reload, test_sch
     model_reload = model_reload.cuda()
     model_reload.train()
 
-    colo_optimizer = ColossalaiOptimizer(torch.optim.SGD(model.named_parameters(), r=0.1))
+    opt_class = torch.optim.Adam
+    colo_optimizer = ColossalaiOptimizer(opt_class(model.parameters(), lr=0.1))
+    colo_optimizer_reload = ColossalaiOptimizer(opt_class(model_reload.parameters(), lr=0.1))
+    run_reload = False
 
     for i, (data, label) in enumerate(train_dataloader):
 
@@ -130,22 +145,35 @@ def _run_checkpoint(model_name, init_spec_func, use_ddp, use_mp_reload, test_sch
         # Bcast rank0 data to all processes
         if criterion:
             output = model(data)
+            output_reload = model_reload(data)
             loss = criterion(output, label)
+            loss_reload = criterion(output_reload, label)
         else:
-            output = model(data, label)
+            loss = model(data, label)
-            loss = output
+            loss_reload = model_reload(data, label)
 
         loss.backward()
-        colo_optimizer.step()
+        loss_reload.backward()
+
+        if run_reload:
+            colo_optimizer_reload.zero_grad()
+            if criterion:
+                output_reload = model_reload(data)
+                loss_reload = criterion(output_reload, label)
+            else:
+                loss_reload = model_reload(data, label)
+            loss_reload.backward()
+            colo_optimizer_reload.step()
 
         if i > 2:
             break
 
     if not os.path.isdir('./checkpoint') and rank == 0:
         os.mkdir('./checkpoint')
-    save_checkpoint('./checkpoint', 0, model, None, None)
+    save_checkpoint('./checkpoint', 0, model, colo_optimizer, None)
+    dist.barrier()
+    load_checkpoint('./checkpoint', 0, model_reload, colo_optimizer_reload, None)
     dist.barrier()
-    load_checkpoint('./checkpoint', 0, model_reload, None, None)
 
     # Since model is sharded, we merge them before param checking.
     for p in model.parameters():
@@ -155,7 +183,7 @@ def _run_checkpoint(model_name, init_spec_func, use_ddp, use_mp_reload, test_sch
         p.to_replicate_()
 
     check_param_equal(model, model_reload)
-
+    compare_optims(colo_optimizer, colo_optimizer_reload)
     if rank == 0:
         remove('./checkpoint')
 
@@ -163,7 +191,7 @@ def _run_checkpoint(model_name, init_spec_func, use_ddp, use_mp_reload, test_sch
 def run_dist(rank, world_size, port, use_ddp, use_mp_reload, test_scheduler):
     colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
     pg = ProcessGroup(tp_degree=world_size)
-    for model_name in ['bert', 'simple_net']:
+    for model_name in ['simple_net', 'bert']:
         _run_checkpoint(model_name,
                         init_1d_row_for_linear_weight_spec,
                         use_ddp,