Skip to content

Test #26

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

Closed
Jintao-Huang wants to merge 3 commits into from
Closed

Test #26

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
4614ce3
Fix the multi-LoRA issue in Twinkle.
Jintao-Huang Apr 9, 2026
109b00b
update
Jintao-Huang Apr 10, 2026
6f6fa34
Merge branch 'main' into test
Jintao-Huang Apr 10, 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
1 change: 1 addition & 0 deletions README_zh.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,6 +102,7 @@ uv pip install -e . --torch-backend=auto

保存的模型，可以参考[模型卡片的示例代码](https://modelscope.cn/models/Qwen/Qwen3.5-35B-A3B)进行推理。


```python
# test env: transformers==5.2.0 megatron-core==0.16.1
import os
Expand Down
4 changes: 2 additions & 2 deletions src/mcore_bridge/bridge/gpt_bridge.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -267,7 +267,7 @@ def _set_module(self, mg_module, hf_state_dict, hf_prefix: str, to_mcore: bool):
new_state_dict = {}
for k, v in hf_state_dict.items():
if self._peft_format:
if '.lora_A.' in k or '.lora_B.' in k or '.modules_to_save.' in k:
if ('.lora_A.' in k or '.lora_B.' in k or '.modules_to_save.' in k) and f'{self._adapter_name}.' in k:
Copy link
Copy Markdown

@gemini-code-assist gemini-code-assist Bot Apr 10, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The check f'{self._adapter_name}.' in k is potentially too broad and could lead to incorrect matches if one adapter name is a prefix of another (e.g., default and default_new). Since keys are dot-separated, it is safer to include the leading dot in the check to ensure it matches the specific adapter component.

Suggested change
if ('.lora_A.' in k or '.lora_B.' in k or '.modules_to_save.' in k) and f'{self._adapter_name}.' in k:
if ('.lora_A.' in k or '.lora_B.' in k or '.modules_to_save.' in k) and f'.{self._adapter_name}.' in k:

k = k.replace(f'{self._adapter_name}.', '')
new_state_dict[k] = v
else:
Expand Down Expand Up @@ -1703,7 +1703,7 @@ def export_weights(
self.config = mg_models[0].config
with torch.no_grad():
for k, v in self._convert(mg_models, {}, hf_prefix, False, tqdm_desc=tqdm_desc):
if converter:
if converter and v is not None:
Copy link
Copy Markdown

@gemini-code-assist gemini-code-assist Bot Apr 10, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

Skipping the converter when v is None causes inconsistent key names across different ranks in a distributed environment. When only_master_rank=True is used, non-master ranks yield v=None. If the converter is skipped on these ranks, they will yield the original Megatron keys, while the master rank yields the converted (HuggingFace) keys. This mismatch can break downstream components that expect consistent keys across all ranks. The converter should be called even when v is None, and it is the responsibility of the converter implementation to handle None values.

Suggested change
if converter and v is not None:
if converter:

kv = converter(k, v)
if kv is None:
continue
Expand Down