onnxruntime
Fix CUDA ONNX Attention: min_bias_align crash on SM<80 and MEA NaN for fully-masked batches
#27831

Merged

Fix CUDA ONNX Attention: min_bias_align crash on SM<80 and MEA NaN for fully-masked batches #27831

titaiwangms merged 1 commit into main from titaiwang/fix_cuda

yuslepukhin requested a review from

copilot-pull-request-reviewer 51 days ago

yuslepukhin removed review request from

copilot-pull-request-reviewer 51 days ago

yuslepukhin requested a review from

tianleiwu 51 days ago

yuslepukhin commented on 2026-03-24

github-actions commented on 2026-03-24

titaiwangms force pushed from a04b41a4 to 38fef79f 51 days ago

titaiwangms force pushed from 38fef79f to 07960375 50 days ago

titaiwangms changed the title ~~Fix TF32 misaligned address error in cuBLAS GEMM functions~~ Fix misaligned BiasLoader access in CUTLASS FMHA attention dispatch 50 days ago

titaiwangms assigned

tianleiwu 50 days ago

titaiwangms force pushed from 07960375 to 9a1bf882 50 days ago

titaiwangms force pushed from 9a1bf882 to 2cccef28 49 days ago

github-actions commented on 2026-03-26

titaiwangms commented on 2026-03-27

titaiwangms unassigned

tianleiwu 49 days ago

titaiwangms force pushed from 2cccef28 to 678c8408 49 days ago

titaiwangms changed the title ~~Fix misaligned BiasLoader access in CUTLASS FMHA attention dispatch~~ Fix ONNX Attention CUDA: bias alignment, unfused decode concat, and MEA NaN 49 days ago

titaiwangms requested a review from

copilot-pull-request-reviewer 49 days ago

copilot-pull-request-reviewer commented on 2026-03-27

titaiwangms force pushed from 678c8408 to ab43783a 49 days ago

titaiwangms requested a review from

copilot-pull-request-reviewer 49 days ago

copilot-pull-request-reviewer commented on 2026-03-27

titaiwangms marked this pull request as ready for review 48 days ago

titaiwangms closed this 48 days ago

titaiwangms reopened this 48 days ago

titaiwangms assigned

tianleiwu 48 days ago

yuslepukhin requested a review from

copilot-pull-request-reviewer 48 days ago

copilot-pull-request-reviewer commented on 2026-03-27

titaiwangms marked this pull request as draft 48 days ago

titaiwangms force pushed from ab43783a to 38f4ca42 48 days ago

titaiwangms requested a review from

copilot-pull-request-reviewer 48 days ago

copilot-pull-request-reviewer commented on 2026-03-28

titaiwangms changed the title ~~Fix ONNX Attention CUDA: bias alignment, unfused decode concat, and MEA NaN~~ Fix CUDA ONNX Attention: min_bias_align crash on SM<80 and MEA NaN for fully-masked batches 48 days ago

titaiwangms marked this pull request as ready for review 47 days ago

yuslepukhin requested a review from

copilot-pull-request-reviewer 45 days ago

copilot-pull-request-reviewer commented on 2026-03-30

titaiwangms force pushed from ca2fc349 to cedf1cd3 45 days ago

titaiwangms force pushed from cedf1cd3 to 2b73e0f7 45 days ago

titaiwangms added this to the 1.25.0 milestone 43 days ago

tianleiwu commented on 2026-04-01

titaiwangms force pushed from ee1429f3 to bde34a52 43 days ago

titaiwangms force pushed from bde34a52 to f23ade69 43 days ago

titaiwangms force pushed from f23ade69 to 09429ab8 43 days ago

titaiwangms force pushed from 09429ab8 to 40764f34 43 days ago

titaiwangms force pushed from 40764f34 to 86738d2f 43 days ago

tianleiwu commented on 2026-04-01

titaiwangms force pushed from 86738d2f to ed246ab1 42 days ago

titaiwangms force pushed from ed246ab1 to aeb40852 42 days ago

tianleiwu commented on 2026-04-02

Fix ONNX Attention: min_bias_align, UnfusedRunner decode, MEA NaN

1e59aa50

titaiwangms force pushed from aeb40852 to 1e59aa50 42 days ago

titaiwangms requested a review from

tianleiwu 41 days ago

tianleiwu approved these changes on 2026-04-04

titaiwangms merged eb706ed3 into main 41 days ago

titaiwangms deleted the titaiwang/fix_cuda branch 41 days ago

Reviewers

tianleiwu

github-actions

yuslepukhin

copilot-pull-request-reviewer

Assignees

tianleiwu

Labels

None yet

Milestone

1.25.0

onnxruntime Fix CUDA ONNX Attention: min_bias_align crash on SM<80 and MEA NaN for fully-masked batches #27831 Merged

Fix CUDA ONNX Attention: min_bias_align crash on SM<80 and MEA NaN for fully-masked batches #27831

onnxruntime
Fix CUDA ONNX Attention: min_bias_align crash on SM<80 and MEA NaN for fully-masked batches
#27831

Merged