[zero] check whether gradients have inf and nan in gpu (#712)

colossalai/zero/sharded_model/sharded_model_v2.py

@@ -148,6 +148,9 @@ class ShardedModelV2(nn.Module):
         self._cuda_margin_space = 0
         self.reuse_fp16_shard = reuse_fp16_shard
 
+        # record whether gradients have inf or nan
+        self.overflow_counter = 0
+
     def adjust_stateful_tensor_layout(self) -> None:
         self._stateful_tensor_mgr.adjust_layout()
 
@@ -345,6 +348,11 @@ class ShardedModelV2(nn.Module):
 
     # FIXME(ver217): refactor the below line when impl eviction policy
     def _save_grad(self, param: Parameter, grad: torch.Tensor):
+
+        # record whether we have overflow
+        self.overflow_counter += torch.isinf(grad).any().item()
+        self.overflow_counter += torch.isnan(grad).any().item()
+
         # move gradient to cpu
         if param.colo_attr.offload_grad:
             colo_model_data_move_to_cpu(grad)

colossalai/zero/sharded_optim/sharded_optim_v2.py

@@ -118,7 +118,7 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
                                              growth_interval=growth_interval,
                                              hysteresis=hysteresis,
                                              max_scale=max_scale)
-        self._found_overflow: Tensor = torch.FloatTensor([0]).to(torch.cuda.current_device())
+        self._found_overflow: Tensor = torch.IntTensor([0]).to(torch.cuda.current_device())
         self._logger = get_dist_logger("ShardedOptimizerV2")
 
         # Store fp32 param shards
@@ -210,20 +210,13 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
 
     def _check_overflow(self):
         # clear previous overflow record
-        self._found_overflow.fill_(0.0)
-
-        # check for overflow
-        for group in self.optim.param_groups:
-            for p in group['params']:
-                if has_inf_or_nan(p.grad):
-                    self._found_overflow.fill_(1.0)
-                    break
+        self._found_overflow.fill_(self.model.overflow_counter)
 
         # all-reduce across dp group
-        dist.all_reduce(self._found_overflow, op=dist.ReduceOp.MAX, group=self.dp_process_group)
+        dist.all_reduce(self._found_overflow, group=self.dp_process_group)
 
         # all-reduce over model parallel group
-        dist.all_reduce(self._found_overflow, op=dist.ReduceOp.MAX, group=self.mp_process_group)
+        dist.all_reduce(self._found_overflow, group=self.mp_process_group)
 
         return self._found_overflow.item() > 0
 
@@ -259,6 +252,7 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
                 else:
                     # release saved gradient
                     p.colo_attr.saved_grad.set_null()
+        self.model.overflow_counter = 0    # set overflow counter to zero
 
     def sync_grad(self):
         pass

tests/test_zero_data_parallel/test_found_inf.py

+from functools import partial
+
+import colossalai
+from colossalai.utils.cuda import get_current_device
+import pytest
+import torch
+import torch.multiprocessing as mp
+from colossalai.nn.optimizer import HybridAdam
+from colossalai.testing import parameterize, rerun_on_exception
+from colossalai.utils import free_port
+from colossalai.zero.init_ctx import ZeroInitContext
+from colossalai.zero.shard_utils import BucketTensorShardStrategy
+from colossalai.zero.sharded_model import ShardedModelV2
+from colossalai.zero.sharded_optim import ShardedOptimizerV2
+from colossalai.zero.sharded_optim._utils import has_inf_or_nan
+from tests.components_to_test.registry import non_distributed_component_funcs
+from tests.test_zero_data_parallel.test_sharded_optim_v2 import _run_step
+
+from common import CONFIG
+
+
+@parameterize("cpu_offload", [True, False])
+@parameterize("shard_strategy_class", [BucketTensorShardStrategy])
+@parameterize("gpu_margin_mem_ratio", [0.0, 0.7])
+def _run_test_found_inf(cpu_offload, shard_strategy_class, gpu_margin_mem_ratio):
+    test_models = ['repeated_computed_layers']
+    shard_strategy = shard_strategy_class()
+
+    for model_name in test_models:
+        get_components_func = non_distributed_component_funcs.get_callable(model_name)
+        model_builder, train_dataloader, _, optimizer_class, criterion = get_components_func()
+
+        with ZeroInitContext(
+                target_device=torch.device(f'cpu:0') if cpu_offload else torch.device(f'cuda:{get_current_device()}'),
+                shard_strategy=shard_strategy,
+                shard_param=True):
+            zero_model = model_builder(checkpoint=True)
+        zero_model = ShardedModelV2(
+            zero_model,
+            shard_strategy,
+            offload_config=dict(device='cpu') if cpu_offload else None,
+            use_memory_tracer=gpu_margin_mem_ratio > 0.0,
+            reuse_fp16_shard=True,
+        )
+
+        sharded_optim = HybridAdam(zero_model.parameters(), lr=1e-3)
+        sharded_optim = ShardedOptimizerV2(zero_model,
+                                           sharded_optim,
+                                           cpu_offload=cpu_offload,
+                                           gpu_margin_mem_ratio=gpu_margin_mem_ratio)
+
+        for i, (data, label) in enumerate(train_dataloader):
+            if i > 1:
+                break
+            assert zero_model.overflow_counter == 0
+            data, label = data.cuda(), label.cuda()
+            _run_step(zero_model, sharded_optim, data, label, criterion, False)
+            for param in zero_model.parameters():
+                assert not has_inf_or_nan(param.colo_attr.sharded_data_tensor.payload)
+
+
+def _run_dist(rank, world_size, port):
+    colossalai.launch(config=CONFIG, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
+    _run_test_found_inf()
+
+
+# use_cpuadam = True can be used with cpu_offload = False
+@pytest.mark.dist
+@pytest.mark.parametrize("world_size", [1, 2])
+@rerun_on_exception(exception_type=mp.ProcessRaisedException, pattern=".*Address already in use.*")
+def test_found_inf(world_size):
+    run_func = partial(_run_dist, world_size=world_size, port=free_port())
+    mp.spawn(run_func, nprocs=world_size)
+
+
+if __name__ == '__main__':
+    test_found_inf(world_size=2)