Skip to content

Conversation

WYB27
Copy link
Collaborator

@WYB27 WYB27 commented Sep 24, 2025

改动:
支持SP下计算moe aux loss

Copy link

paddle-bot bot commented Sep 24, 2025

Thanks for your contribution!

@codecov-commenter
Copy link

codecov-commenter commented Sep 25, 2025

Codecov Report

❌ Patch coverage is 0% with 44 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@a103c70). Learn more about missing BASE report.

Files with missing lines Patch % Lines
paddleformers/transformers/glm4_moe/modeling.py 0.00% 23 Missing ⚠️
paddleformers/transformers/moe_gate.py 0.00% 20 Missing ⚠️
...ddleformers/transformers/glm4_moe/configuration.py 0.00% 1 Missing ⚠️
Additional details and impacted files
@@            Coverage Diff             @@
##             develop    #2682   +/-   ##
==========================================
  Coverage           ?   29.68%           
==========================================
  Files              ?      311           
  Lines              ?    54763           
  Branches           ?        0           
==========================================
  Hits               ?    16257           
  Misses             ?    38506           
  Partials           ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:
  • ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

# certain operations (like matmul) to FP16/BF16 for performance optimization. However, in scenarios where
# numerical stability is critical (e.g., RoPE init/compute), this conversion can lead to precision loss.
# Disabling auto_cast here ensures the matmul operation runs in the original precision (FP32) as intended.
with paddle.amp.auto_cast(False):
Copy link
Member

Choose a reason for hiding this comment

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

这个地方训推会不一致,不要修改

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants