Skip to content

Avoid calling set_save_original_input with FP8 delayed scaling #1860

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

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
+1 −1
Open

Avoid calling set_save_original_input with FP8 delayed scaling #1860

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2 changes: 1 addition & 1 deletion megatron/core/transformer/moe/shared_experts.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,7 +62,7 @@ def __init__(
else:
self.gate_weight = None

if self.config.fp8 and is_te_min_version("2.6.0dev0"):
if self.config.fp8 and self.config.fp8_recipe != 'delayed' and is_te_min_version("2.6.0dev0"):
# For fp8 training, the output of pre_mlp_layernorm is saved by router, and
# the shared expert linear_fc1 also saves the quantized tensor of this output.
# Here we set the linear_fc1 to save the original input tensors to avoid the extra
Expand Down