Skip to content

[bugfix] fix mcore-bridge vpp #6581

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
merged 2 commits into from
Nov 13, 2025
Merged

[bugfix] fix mcore-bridge vpp #6581

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d7a2450
fix bridge vpp
Jintao-Huang Nov 13, 2025
0c91437
fix cp kto
Jintao-Huang Nov 13, 2025
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
11 changes: 8 additions & 3 deletions swift/megatron/model/gpt_bridge.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -978,8 +978,9 @@ def _convert(self, mg_models, hf_state_dict, hf_prefix: str, to_mcore: bool, tqd
else:
yield from list(self._add_prefix(hf_state_dict, hf_prefix).items())
hf_state_dict = {}
for layer_idx in tqdm(
range(self.args.num_layers), dynamic_ncols=True, desc=tqdm_desc, disable=self.disable_tqmd):
layer_idx = 0
prog_bar = tqdm(range(self.args.num_layers), dynamic_ncols=True, desc=tqdm_desc, disable=self.disable_tqmd)
while layer_idx < self.args.num_layers:
lm_model = getattr(mg_model, 'language_model') if self.args.is_multimodal else mg_model
if len(lm_model.decoder.layers) > 0:
start_idx = lm_model.decoder.layers[0].layer_number - 1
Expand All @@ -990,16 +991,20 @@ def _convert(self, mg_models, hf_state_dict, hf_prefix: str, to_mcore: bool, tqd
mg_layer = lm_model.decoder.layers[layer_idx - start_idx]
else:
if to_mcore:
layer_idx += 1
prog_bar.update()
continue
else:
mg_layer = None
if not to_mcore and self.pp_size > 1:
has_model = torch.tensor([mg_layer is not None], dtype=torch.bool, device='cuda')
dist.all_reduce(has_model, group=self.pp_group)
if not has_model:
mg_model = next(mg_models)
mg_model = next(mg_models) # compat vpp
continue
res = self._set_layer_state(mg_layer, hf_state_dict, f'{self.hf_layers_prefix}.', layer_idx, to_mcore)
layer_idx += 1
prog_bar.update()
if to_mcore:
yield
else:
Expand Down
5 changes: 2 additions & 3 deletions swift/megatron/trainers/kto_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,7 +50,7 @@ def _kto_get_logps(self, output_tensor, data, is_KL: bool, is_ref: bool, length:
return self.get_logps(output, labels, packed_seq_params, packed_seq_params.num_samples)

def loss_func(self, output_tensor, *, data, kl_data, label):
length = data['packed_seq_params'].cu_seqlens_q[-1]
length = data['packed_seq_params'].cu_seqlens_q[-1] // self.args.context_parallel_size
policy_logps = self._kto_get_logps(output_tensor, data, False, False, length)
ref_logps = self._kto_get_logps(output_tensor, data, False, True, length)
if self.args.calculate_KL:
Expand Down Expand Up @@ -121,8 +121,7 @@ def forward_step(self, data_iterator, model):
data.pop('loss_scale', None)
kl_data.pop('loss_scale', None)

length = data['packed_seq_params'].cu_seqlens_q[-1]

length = data['packed_seq_params'].cu_seqlens_q[-1] // self.args.context_parallel_size
with torch.no_grad(), self.null_ref_context() as ref_models:
ref_model = ref_models[vp_stage or 0]
if self.args.calculate_KL:
Expand Down
Loading