github
/
ColossalAI
mirror of https://github.com/hpcaitech/ColossalAI
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

hotfix

pull/5994/head
Tong Li 2024-08-13 09:35:03 +00:00
parent 22218d31e1
commit 2422341d03
2 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
applications/ColossalChat/tests/test_train.sh
View File

@ -30,7 +30,7 @@ MODEL_SAVE_PATH=$TEMP_DIR/rlhf_models
MODELS_DIR=$TEMP_DIR/models_config
# Skip those tests due to CI tests timeout
MODELS=('llama')
ADVANCED_PLUGINS=('zero2' 'sp_split_gather' 'sp_ring' 'sp_all_to_all' 'tp_zero2' '3d' 'gemini' 'gemini_auto' 'zero2_cpu' 'pp')
ADVANCED_PLUGINS=('zero2' 'sp_split_gather' 'sp_ring' 'sp_all_to_all' 'tp_zero2' '3d' 'gemini' 'gemini_auto' 'zero2_cpu' 'pp' 'tp_pp')
PLUGINS=('zero2' '3d' 'gemini' 'gemini_auto' 'zero2_cpu')
LORA_RANK=('0') # skip to reduce CI execution time, can pass all locally
LORA_CONFIG_ENABLE="--lora_config $BASE_DIR/examples/training_scripts/lora_config.json"

2
colossalai/booster/plugin/hybrid_parallel_plugin.py
View File

@ -1328,7 +1328,7 @@ class HybridParallelPlugin(PipelinePluginBase):
# run with gradients accumulation
if model.require_grad_sync == False or (
isinstance(optimizer, HybridParallelZeroOptimizer) and optimizer.require_grad_sync == False
):
) or not torch.is_grad_enabled():
return outputs
# Synchronize the grads of shared parameters of the model.