github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-11 22:10:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' into sync/npu

Browse Source
This commit is contained in:
ver217
2024-01-18 12:05:21 +08:00
parent 9102d655ab 5d9a0ae75b
commit 148469348a
152 changed files with 8641 additions and 2138 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
examples/language/bert/finetune.py
View File

@@ -57,7 +57,7 @@ def evaluate_model(
def evaluate_subset(dataloader: DataLoader):
use_pipeline = isinstance(booster.plugin, HybridParallelPlugin) and booster.plugin.pp_size > 1
is_pp_last_stage = use_pipeline and booster.plugin.stage_manager.is_last_stage()
is_pp_last_device = use_pipeline and booster.plugin.stage_manager.is_last_stage(ignore_chunk=True)
accum_loss = torch.zeros(1, device=get_accelerator().get_current_device())
for batch in dataloader:
@@ -69,9 +69,10 @@ def evaluate_model(
current_pp_group_ranks = pg_mesh.get_ranks_in_group(pp_group)
current_rank = dist.get_rank()
batch = iter([batch])
outputs = booster.execute_pipeline(batch, model, criterion, return_loss=True, return_outputs=True)
if is_pp_last_stage:
if is_pp_last_device:
logits = outputs["outputs"]["logits"]
val_loss = outputs["loss"]
accum_loss.add_(val_loss)
@@ -135,8 +136,8 @@ def train_epoch(
coordinator: DistCoordinator,
):
use_pipeline = isinstance(booster.plugin, HybridParallelPlugin) and booster.plugin.pp_size > 1
is_pp_last_stage = use_pipeline and booster.plugin.stage_manager.is_last_stage()
print_flag = (not use_pipeline and coordinator.is_master()) or (use_pipeline and is_pp_last_stage)
is_pp_last_device = use_pipeline and booster.plugin.stage_manager.is_last_stage(ignore_chunk=True)
print_flag = (not use_pipeline and coordinator.is_master()) or (use_pipeline and is_pp_last_device)
total_step = len(train_dataloader)
model.train()
@@ -150,7 +151,7 @@ def train_epoch(
train_dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True
)
# Backward and optimize
if is_pp_last_stage:
if is_pp_last_device:
loss = outputs["loss"]
pbar.set_postfix({"loss": loss.item()})
else:
@@ -224,7 +225,9 @@ def main():
tp_size=1,
pp_size=2,
num_microbatches=None,
microbatch_size=1,
pp_style="interleaved",
num_model_chunks=2,
microbatch_size=16,
enable_all_optimization=True,
zero_stage=1,
precision="fp16",