PaddleNLP icon indicating copy to clipboard operation
PaddleNLP copied to clipboard

[Performance] Optimize unified checkpoint save/load speed.

Open ZHUI opened this issue 10 months ago • 3 comments

PR types

PR changes

Description

For facebook/llama-7b

UC/PD Save/Load 对比 UC save PD save
testDP8 0 0
testPP4DP2 47.13 68.36
testPP4Sharding2 25.93 41.04
testPP8 27.04 39.68
testSharding2S1DP4 93.19 145.13
testSharding2S2DP4 90.62 147.15
testSharding4S1DP2 47.6 94.75
testSharding4S2DP2 44.12 95.68
testSharding8S1 25.94 64.37
testSharding8S2 23.95 65.77
testTP2PP4 27.92 35.71
testTP2Sharding4 27.78 46.31
testTP4DP2 53.74 51.98
testTP4PP2 27.32 30.55
testTP4Sharding2 28.35 31.31
testTP8 30.63 23.75

image

image

ZHUI avatar Mar 28 '24 07:03 ZHUI

Thanks for your contribution!

paddle-bot[bot] avatar Mar 28 '24 07:03 paddle-bot[bot]

CLA assistant check
All committers have signed the CLA.

CLAassistant avatar Mar 28 '24 07:03 CLAassistant

Codecov Report

Attention: Patch coverage is 73.49398% with 66 lines in your changes are missing coverage. Please review.

Project coverage is 55.42%. Comparing base (f29a7b9) to head (32132ae). Report is 23 commits behind head on develop.

Files Patch % Lines
paddlenlp/utils/safetensors.py 88.26% 23 Missing :warning:
paddlenlp/trainer/plugins/unified_checkpoint.py 6.25% 15 Missing :warning:
paddlenlp/transformers/model_utils.py 37.50% 15 Missing :warning:
paddlenlp/transformers/conversion_utils.py 0.00% 12 Missing :warning:
paddlenlp/trainer/trainer.py 0.00% 1 Missing :warning:
Additional details and impacted files
@@             Coverage Diff             @@
##           develop    #8204      +/-   ##
===========================================
+ Coverage    55.37%   55.42%   +0.04%     
===========================================
  Files          613      615       +2     
  Lines        95855    96235     +380     
===========================================
+ Hits         53083    53335     +252     
- Misses       42772    42900     +128     

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

codecov[bot] avatar Mar 28 '24 08:03 codecov[bot]