github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-01 09:07:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

[autoparallel] modify comm nodes' memory cost in construct chain (#2263)

Browse Source 
* [autoparallel] align the data_ptr with the old version of auto activation checkpoint pipeline

* [autoparallel] using fwd_time and bwd_time instead of fwd_flop and bwd_flop

* [autoparallel] specifycomm nodes' memory cost in construct chain
This commit is contained in:
Boyuan Yao
2023-01-03 11:38:48 +08:00
committed by GitHub
parent 1ea99b869e
commit 5c2ef9fc76
3 changed files with 12 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
colossalai/fx/profiler/shard_utils.py
View File

@@ -100,7 +100,7 @@ def calculate_fwd_time(n: Node) -> float:
fwd_time (float): the result of `fwd_time`
"""
# TODO(super-dainiu): should divide the time by the number of GPUs as well as TFLOPs
return n.meta["fwd_flop"]
return n.meta["fwd_time"]
def calculate_bwd_time(n: Node) -> float:
@@ -111,4 +111,4 @@ def calculate_bwd_time(n: Node) -> float:
bwd_time (float): the result of `bwd_time`
"""
# TODO(super-dainiu): should divide the time by the number of GPUs as well as TFLOPs
return n.meta["bwd_flop"]
return n.meta["bwd_time"]