github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-25 11:44:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

[autoparallel] fix forward memory calculation (#2062)

Browse Source
This commit is contained in:
Boyuan Yao
2022-12-04 15:00:16 +08:00
committed by GitHub
parent 44ea461890
commit 4b40fbd743
5 changed files with 29 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
colossalai/auto_parallel/meta_profiler/meta_registry/pooling.py
View File

@@ -110,7 +110,8 @@ def maxpool_meta_info(*args, **kwargs) -> Tuple[TrainCycleItem, TrainCycleItem,
# calculate memory cost
# NOTE: the index matrix will be discarded in backward phase
fwd_mem_cost = MemoryCost(activation=activation_size(output_tensor) + activation_size(index_matrix))
# NOTE: currently in SPMD solver we always believe that there will be a new tensor created in forward
fwd_mem_cost = MemoryCost(activation=activation_size([input_tensor, output_tensor, index_matrix]))
# temp memory for backward is the index matrix to be discarded
bwd_mem_cost = MemoryCost(activation=activation_size(input_tensor) - activation_size(index_matrix),