Skip to content

separate mtp head & loss for pp balance #4327

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
tianlef merged 6 commits into from
Apr 25, 2026
+7 −5
Merged

separate mtp head & loss for pp balance #4327

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
957c329
separate mtp head & loss for pp balance
Wennie396 Apr 21, 2026
023143d
separate mtp head & loss for pp balance
Wennie396 Apr 21, 2026
6a1d403
Merge branch 'mtpopt' of https://github.com/Wennie396/PaddleFormers i…
Wennie396 Apr 21, 2026
ada1c34
fix
Wennie396 Apr 21, 2026
8b209c6
update formers paddlefleet version
Wennie396 Apr 23, 2026
0ffe7e7
Merge branch 'develop' of https://github.com/PaddlePaddle/PaddleForme…
Wennie396 Apr 25, 2026
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
4 changes: 2 additions & 2 deletions paddleformers/trainer/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3452,7 +3452,7 @@ def _wrap_model(self, model, training=True):
else:

def _prepare_pipeline_inputs_func(inputs):
first_stage_keys = ["input_ids", "attention_mask", "position_ids"]
first_stage_keys = ["input_ids", "attention_mask", "position_ids", "labels"]
last_stage_keys = ["labels"]

def get_expected_keys(inputs, keys):
Expand All @@ -3471,7 +3471,7 @@ def get_expected_keys(inputs, keys):
inputs_batch = {key: [data.pop(key) for data in inputs] for key in keys}
if is_paddlefleet_available() and self.using_fleet_model:
first_stage_inputs_batch = inputs_batch
last_stage_inputs = first_stage_inputs_batch.pop("labels")
last_stage_inputs = first_stage_inputs_batch.get("labels")
outputs = (
first_stage_inputs_batch,
last_stage_inputs,
Expand Down
8 changes: 5 additions & 3 deletions paddleformers/transformers/gpt_provider.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -216,9 +216,11 @@ def provide(self, pre_process=None, post_process=None, vp_stage=None, loss_fn=No
"""

with model_init_device_context():
fleet_model = gpt_builder(
self, num_stages=pp_size, seg_method="layer:TransformerLayer|EmptyLayer", loss_fn=loss_fn
)
seg_method = "layer:TransformerLayer|EmptyLayer"
if self.separate_mtp_headloss:
seg_method = "layer:TransformerLayer|EmptyLayer|MultiTokenPredictionLayer"

fleet_model = gpt_builder(self, num_stages=pp_size, seg_method=seg_method, loss_fn=loss_fn)
# Convert original FleetGPTModel to our GPTModel to correctly inherit PretrainedModel methods
model = GPTModel.__new__(GPTModel)
# Manually copy all attributes
Expand Down
Loading