github
/
ColossalAI
mirror of https://github.com/hpcaitech/ColossalAI
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

[gemini] fix auto tensor placement policy (#775)

pull/781/head
ver217 2022-04-16 21:29:31 +08:00 committed by GitHub
parent 38102cf61a
commit 846406a07a
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 3 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
colossalai/gemini/tensor_placement_policy.py
View File

@ -46,7 +46,9 @@ class AutoTensorPlacementPolicy(TensorPlacementPolicy):
def __init__(self, mem_stats_collector: Optional[MemStatsCollector] = None) -> None: def __init__(self, mem_stats_collector: Optional[MemStatsCollector] = None) -> None:
super().__init__(None, mem_stats_collector=mem_stats_collector) super().__init__(None, mem_stats_collector=mem_stats_collector)
# model data will use 1-self._warmup_non_model_data_ratio CUDA memory in warmup phase # model data will use 1-self._warmup_non_model_data_ratio CUDA memory in warmup phase
# TODO(ver217): make these args configurable
self._warmup_non_model_data_ratio: float = 0.8 self._warmup_non_model_data_ratio: float = 0.8
self._steady_cuda_cap_ratio: float = 0.8
def evict_tensors(self, def evict_tensors(self,
hold_cuda_tensor_list: List[StatefulTensor], hold_cuda_tensor_list: List[StatefulTensor],
@ -63,6 +65,7 @@ class AutoTensorPlacementPolicy(TensorPlacementPolicy):
else: else:
# max non-model-data cuda memory consumption of this sampling moment and the next sampling moment. # max non-model-data cuda memory consumption of this sampling moment and the next sampling moment.
max_cuda_non_model_data_per_period = self.mem_stats_collector.next_period_non_model_data_usage('cuda') max_cuda_non_model_data_per_period = self.mem_stats_collector.next_period_non_model_data_usage('cuda')
cuda_capacity *= self._steady_cuda_cap_ratio
total_cuda_model_data = cuda_capacity - max_cuda_non_model_data_per_period total_cuda_model_data = cuda_capacity - max_cuda_non_model_data_per_period
avail_cuda_model_data = total_cuda_model_data - used_cuda_model_data avail_cuda_model_data = total_cuda_model_data - used_cuda_model_data
if avail_cuda_model_data < cuda_demand: if avail_cuda_model_data < cuda_demand: