Skip to content

[Ascend] fix AscendAttnMaskBuilder bug to support float16 models#14271

Merged
iforgetmyname merged 3 commits intosgl-project:mainfrom
MichelleWu351:patch-1
Dec 3, 2025
Merged

[Ascend] fix AscendAttnMaskBuilder bug to support float16 models#14271
iforgetmyname merged 3 commits intosgl-project:mainfrom
MichelleWu351:patch-1

Conversation

@MichelleWu351
Copy link
Contributor

@MichelleWu351 MichelleWu351 commented Dec 2, 2025

Motivation

The AscendAttnMaskBuilder class implements a system for generating, caching, and updating attention masks. However, there is a bug when using this class with Ascend devices and a model configured with the float16 data type. This pull request can fix the bug.

Modifications

Modify the generate_attn_mask fuction in the AscendAttnMaskBuilder class.

Accuracy Tests

Benchmarking and Profiling

Checklist

@gemini-code-assist
Copy link
Contributor

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

@github-actions github-actions bot added the npu label Dec 2, 2025
@MichelleWu351 MichelleWu351 changed the title bug fix to support float16 [Ascend] fix AscendAttnMaskBuilder fix to support float16 models Dec 2, 2025
@MichelleWu351 MichelleWu351 changed the title [Ascend] fix AscendAttnMaskBuilder fix to support float16 models [Ascend] fix AscendAttnMaskBuilder bug to support float16 models Dec 2, 2025
@iforgetmyname
Copy link
Collaborator

/tag-and-rerun-ci

@github-actions github-actions bot added the run-ci label Dec 2, 2025
@iforgetmyname iforgetmyname self-assigned this Dec 2, 2025
@ping1jing2 ping1jing2 self-assigned this Dec 2, 2025
@iforgetmyname iforgetmyname merged commit 443d7bc into sgl-project:main Dec 3, 2025
125 of 139 checks passed
yingluosanqian pushed a commit to yingluosanqian/sglang that referenced this pull request Dec 4, 2025
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

Comments