Skip to content

Cache FP8 weight and transpose only at the first micro-batch in each validation and test routine #7470

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
layalir merged 2 commits into from
Sep 22, 2023
Merged

Cache FP8 weight and transpose only at the first micro-batch in each validation and test routine #7470

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
063d526
Cache weight and transpose only in the first batch in all training, v…
erhoo82 Sep 20, 2023
9d5620d
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Sep 20, 2023
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
26 changes: 21 additions & 5 deletions nemo/collections/nlp/modules/common/megatron/transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1041,7 +1041,10 @@ def __init__(
reduce_amax=reduce_amax,
)

self.is_first_microbatch = True
self.is_first_train_microbatch = (
True # Is the current micro-batch the first micro-batch in a global-batch in training
)
self.is_prev_microbatch_training = True # Is the previous micro-batch in training mode
self.microbatch_count = 0 # transformer engine forward needs to know if it is working on the first microbatch
self.checkpoint_core_attention = (
activations_checkpoint_granularity == 'selective'
Expand Down Expand Up @@ -1247,6 +1250,12 @@ def custom_forward(*inputs):
attention_mask = inputs[1]
encoder_output = inputs[2]
enc_dec_attn_mask = inputs[3]
# Cache FP8 weight and transpose at (1) the first micro-batch in each global-batch
# in training, (2) the first micro-batch in each validation and test routine.
# The caching happens in TransformerEngine when passing `is_first_microbatch=True`.
is_first_microbatch = (self.is_first_train_microbatch and self.training) or (
self.is_prev_microbatch_training and not self.training
)
for index in range(start, end):
layer = self._get_layer(index)
hidden_states = layer(
Expand All @@ -1255,7 +1264,7 @@ def custom_forward(*inputs):
encoder_output=encoder_output,
enc_dec_attn_mask=enc_dec_attn_mask,
inference_params=None,
is_first_microbatch=self.is_first_microbatch,
is_first_microbatch=is_first_microbatch,
checkpoint_core_attention=False,
)

Expand Down Expand Up @@ -1531,14 +1540,20 @@ def forward(
else:
checkpoint_core_attention = False

# Cache FP8 weight and transpose at (1) the first micro-batch in each global-batch
# in training, (2) the first micro-batch in each validation and test routine.
# The caching happens in TransformerEngine when passing `is_first_microbatch=True`.
is_first_microbatch = (self.is_first_train_microbatch and self.training) or (
self.is_prev_microbatch_training and not self.training
)
if self.transformer_engine:
hidden_states = layer(
hidden_states,
attention_mask,
encoder_output=encoder_output,
enc_dec_attn_mask=enc_dec_attn_mask,
inference_params=self.inference_params,
is_first_microbatch=self.is_first_microbatch,
is_first_microbatch=is_first_microbatch,
checkpoint_core_attention=checkpoint_core_attention,
)
else:
Expand All @@ -1565,9 +1580,10 @@ def forward(
self.microbatch_count += 1
if self.microbatch_count % num_micro_batches == 0:
self.microbatch_count = 0
self.is_first_microbatch = True
self.is_first_train_microbatch = True
else:
self.is_first_microbatch = False
self.is_first_train_microbatch = False
self.is_prev_microbatch_training = self.training

output = hidden_states

Expand Down