Skip to content

Allow turning off loss scaling wrt GAS + update tput calculator #2140

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jeffra merged 5 commits into from
Jul 26, 2022
Merged

Allow turning off loss scaling wrt GAS + update tput calculator #2140

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
ff75046
allow turning off gas loss scaling
jeffra May 11, 2022
291e4af
Merge branch 'master' into jeffra/fs-gas-fix-v2
jeffra Jun 23, 2022
514cab7
add curr samples per sec
jeffra Jun 24, 2022
8a0d639
Merge branch 'master' into jeffra/fs-gas-fix-v3
jeffra Jul 26, 2022
89da3a3
Merge branch 'master' into jeffra/fs-gas-fix-v3
jeffra Jul 26, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 10 additions & 2 deletions deepspeed/runtime/engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -229,6 +229,7 @@ def __init__(
global dist
from deepspeed import comm as dist
self._is_gradient_accumulation_boundary = None
self.scale_wrt_gas = None

# for debug purposes - can then debug print: debug_get_module_name(module)
debug_extract_module_and_param_names(model)
Expand Down Expand Up @@ -1693,7 +1694,11 @@ def allreduce_gradients(self, bucket_size=MEMORY_OPT_ALLREDUCE_SIZE):
self.buffered_allreduce_fallback(elements_per_buffer=bucket_size)

@instrument_w_nvtx
def backward(self, loss, allreduce_gradients=True, release_loss=False):
def backward(self,
loss,
allreduce_gradients=True,
release_loss=False,
scale_wrt_gas=True):
r"""Execute backward pass on the loss
Arguments:
loss: Torch tensor on which to execute backward propagation
Expand All @@ -1702,13 +1707,16 @@ def backward(self, loss, allreduce_gradients=True, release_loss=False):

see_memory_usage("Engine before backward", force=self.memory_breakdown())

if self.scale_wrt_gas is not None:
scale_wrt_gas = self.scale_wrt_gas

if not allreduce_gradients:
logger.warning(
f"Argument `allreduce_gradients` is deprecated, ignored, and will soon be removed"
)

# scale loss w.r.t. gradient accumulation if needed
if self.gradient_accumulation_steps() > 1:
if self.gradient_accumulation_steps() > 1 and scale_wrt_gas:
loss = self._scale_loss_by_gas(loss.float())

# Log training Loss
Expand Down
6 changes: 5 additions & 1 deletion deepspeed/utils/timer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -190,13 +190,17 @@ def stop(self, report_speed=True):
self.end_time = time.time()
duration = self.end_time - self.start_time
self.total_elapsed_time += duration

curr_samples_sec = (self.batch_size * self.num_workers) / duration

if self.local_step_count % self.steps_per_output == 0:
if report_speed:
self.logging(
"{}/{}, SamplesPerSec={}, MemAllocated={}GB, MaxMemAllocated={}GB"
"{}/{}, RunningAvgSamplesPerSec={}, CurrSamplesPerSec={}, MemAllocated={}GB, MaxMemAllocated={}GB"
.format(self.epoch_count,
self.local_step_count,
self.avg_samples_per_sec(),
curr_samples_sec,
round(torch.cuda.memory_allocated() / 1024**3,
2),
round(torch.cuda.max_memory_allocated() / 1024**3,
Expand Down