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

add memory print

pull/456/head
yingtongxiong 2023-10-25 14:31:00 +08:00
parent 918dff7257
commit 363275b500
2 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
internlm/model/overlap_handler.py
View File

@ -316,7 +316,8 @@ class FSTPOverlapSchedulerHook(SchedulerHook):
self._overlap_handler.set_forward_mode(True) self._overlap_handler.set_forward_mode(True)
def after_forward(self, scheduler, outputs) -> None: def after_forward(self, scheduler, outputs) -> None:
pass print("after forward allocated memory: ", torch.cuda.memory_allocated() / 1024 / 1024 /1024, flush=True)
print("after forward max memory: ", torch.cuda.max_memory_allocated() / 1024 / 1024 / 1024, flush=True)
def before_criterion(self, scheduler, outputs, label) -> None: def before_criterion(self, scheduler, outputs, label) -> None:
pass pass

2
train.py
View File

@ -255,6 +255,8 @@ def main(args):
# update parameters, and returns (success_update, grad_norm) # update parameters, and returns (success_update, grad_norm)
trainer_result = trainer.step() trainer_result = trainer.step()
assert trainer_result is not None assert trainer_result is not None
print("after step: ", torch.cuda.memory_allocated() / 1024 / 1024 /1024, flush=True)
print("after step: ", torch.cuda.max_memory_allocated() / 1024 / 1024 / 1024, flush=True)
success_update, grad_norm_groups = trainer_result success_update, grad_norm_groups = trainer_result
if success_update: # update parameters successfully if success_update: # update parameters successfully