github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-16 06:30:41 +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

24
colossalai/auto_parallel/meta_profiler/meta_registry/conv.py
View File

@@ -96,19 +96,19 @@ def convnd_meta_info(*args, **kwargs) -> Tuple[TrainCycleItem, TrainCycleItem, L
# calculate memory cost
# TODO: use profiler to check conv temp memory
fwd_memory_cost = MemoryCost(activation=activation_size(output_tensor),
parameter=activation_size(weight_tensor) +
activation_size(bias_tensor) if has_bias else activation_size(weight_tensor),
temp=0,
buffer=0)
# NOTE: currently in SPMD solver we always believe that there will be a new tensor created in forward
fwd_memory_cost = MemoryCost(
activation=activation_size([input_tensor, output_tensor]),
parameter=activation_size([weight_tensor, bias_tensor]) if has_bias else activation_size(weight_tensor),
temp=0,
buffer=0)
bwd_memory_cost = MemoryCost(activation=activation_size(input_tensor) + activation_size(weight_tensor) +
activation_size(bias_tensor) if has_bias else activation_size(input_tensor) +
activation_size(weight_tensor),
parameter=activation_size(weight_tensor) +
activation_size(bias_tensor) if has_bias else activation_size(weight_tensor),
temp=0,
buffer=0)
bwd_memory_cost = MemoryCost(
activation=activation_size([input_tensor, weight_tensor, bias_tensor])
if has_bias else activation_size([input_tensor, weight_tensor]),
parameter=activation_size([weight_tensor, bias_tensor]) if has_bias else activation_size(weight_tensor),
temp=0,
buffer=0)
# total cost is the sum of forward and backward cost
total_cost = MemoryCost(activation=fwd_memory_cost.activation + bwd_memory_cost.activation,