Skip to content

[bugfix] LoRA + TP>1 + sequence_parallel: sum grads of the replicated LoRA factor over the TP group (exported adapter had last-layer linear_proj.lora_B == 0) #173

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

[bugfix] LoRA + TP>1 + sequence_parallel: sum grads of the replicated LoRA factor over the TP group (exported adapter had last-layer linear_proj.lora_B == 0) #173

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
12 changes: 12 additions & 0 deletions src/mcore_bridge/tuners/lora.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -279,6 +279,18 @@ def update_layer(self, adapter_name, r, *, lora_alpha, **kwargs):
lora.ub_overlap_ag_fprop = False
lora.ub_overlap_rs_dgrad = False

# With sequence parallelism the replicated (non-sharded) LoRA factor only sees this TP
# rank's sequence shard: for RowParallel targets lora_A reduce-scatters its output before
# lora_B, and for ColumnParallel targets lora_A consumes the sequence-sharded input. Its
# gradient must therefore be summed over the TP group. Megatron does this in
# finalize_model_grads for parameters flagged `sequence_parallel` (same as layernorm
# weights); without the flag each TP rank trains a different copy and export_weights
# saves rank 0 only (observed: last layer linear_proj.lora_B saved as all zeros).
if (self.tp_size > 1 and not isinstance(self.base_layer, TopKRouter)
and (getattr(self.config, 'sequence_parallel', False) or self.sequence_parallel)):
replicated = lora_b if self.is_parallel_a else lora_a
for p in replicated.parameters():
p.sequence_parallel = True
self.lora_A[adapter_name] = lora_a
self.lora_B[adapter_name] = lora_b
if hasattr(self, 'lora_bias'):
Expand Down