github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-13 05:01:44 +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

13
colossalai/auto_parallel/meta_profiler/meta_registry/linear.py
View File

@@ -106,15 +106,15 @@ def linear_meta_info(*args, **kwargs) -> Tuple[TrainCycleItem, TrainCycleItem, L
# calculate memory cost
# NOTE: Linear don't have buffer and temp in forward and backward phase
# the forward activation cost is the size of output_tensor, parameter cost is the size of weight_tensor and bias_tensor
fwd_memory_cost = MemoryCost(activation=activation_size(output_tensor),
parameter=activation_size(weight_tensor) + activation_size(bias_tensor),
# 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]),
temp=0,
buffer=0)
# the backward activation cost is the size of input_tensor, weight_tensor and bias_tensor, parameter cost is 0
bwd_memory_cost = MemoryCost(activation=activation_size(input_tensor) + activation_size(weight_tensor) +
activation_size(bias_tensor),
parameter=activation_size(weight_tensor) + activation_size(bias_tensor),
bwd_memory_cost = MemoryCost(activation=activation_size([input_tensor, weight_tensor, bias_tensor]),
parameter=activation_size([weight_tensor, bias_tensor]),
temp=0,
buffer=0)
@@ -142,13 +142,14 @@ def linear_meta_info(*args, **kwargs) -> Tuple[TrainCycleItem, TrainCycleItem, L
# calculate memory cost
# NOTE: Linear don't have buffer and temp in forward and backward phase
# the forward activation cost is the size of output_tensor, parameter cost is the size of weight_tensor
# 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(output_tensor),
parameter=activation_size(weight_tensor),
temp=0,
buffer=0)
# the backward activation cost is the size of input_tensor and weight_tensor, parameter cost is 0
bwd_memory_cost = MemoryCost(activation=activation_size(input_tensor) + activation_size(weight_tensor),
bwd_memory_cost = MemoryCost(activation=activation_size([input_tensor, weight_tensor]),
parameter=activation_size(weight_tensor),
temp=0,
buffer=0)