github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-02 09:38:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

[pipeline]: support arbitrary batch size in forward_only mode (#5201)

Browse Source 
* fix: remove drop last in val & test dataloader

* feat: add run_forward_only, support arbitrary bs

* chore: modify ci script
This commit is contained in:
Wenhao Chen
2024-01-02 23:41:12 +08:00
committed by GitHub
parent 02d2328a04
commit 3c0d82b19b
4 changed files with 293 additions and 202 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
examples/language/bert/data.py
View File

@@ -88,24 +88,21 @@ class GLUEDataBuilder:
)
def val_dataloader(self):
# TODO: drop_last is set to True for now to avoid error when using PP
# as the last batch may not be divisible by the number of microbatches
if len(self.eval_splits) == 1:
return self.plugin.prepare_dataloader(
self.dataset["validation"], batch_size=self.eval_batch_size, drop_last=True
)
return self.plugin.prepare_dataloader(self.dataset["validation"], batch_size=self.eval_batch_size)
elif len(self.eval_splits) > 1:
return [
self.plugin.prepare_dataloader(self.dataset[x], batch_size=self.eval_batch_size, drop_last=True)
self.plugin.prepare_dataloader(self.dataset[x], batch_size=self.eval_batch_size)
for x in self.eval_splits
]
def test_dataloader(self):
if len(self.eval_splits) == 1:
return self.plugin.prepare_dataloader(self.dataset["test"], batch_size=self.eval_batch_size, drop_last=True)
return self.plugin.prepare_dataloader(self.dataset["test"], batch_size=self.eval_batch_size)
elif len(self.eval_splits) > 1:
return [
self.plugin.prepare_dataloader(self.dataset[x], batch_size=self.eval_batch_size, drop_last=True)
self.plugin.prepare_dataloader(self.dataset[x], batch_size=self.eval_batch_size)
for x in self.eval_splits
]

13
examples/language/bert/test_ci.sh
View File

@@ -1,8 +1,17 @@
#!/bin/bash
set -xe
set -x
pip install -r requirements.txt
FAIL_LIMIT=3
for plugin in "torch_ddp" "torch_ddp_fp16" "gemini" "low_level_zero" "hybrid_parallel"; do
torchrun --standalone --nproc_per_node 4 finetune.py --target_f1 0.86 --plugin $plugin --model_type "bert"
for i in $(seq 1 $FAIL_LIMIT); do
torchrun --standalone --nproc_per_node 4 finetune.py --target_f1 0.86 --plugin $plugin --model_type "bert" && break
echo "Failed $i times"
if [ $i -eq $FAIL_LIMIT ]; then
echo "Failed $FAIL_LIMIT times, exiting"
exit 1
fi
done
done