🐛 Describe the bug

Encountered while testing for preview release/2.6 builds as part of #139175

python test_flex_attention.py -k "test_load_from_bias_seq_only_float16

Traceback

test_load_from_bias_seq_only_float16 (main.TestFlexAttention) ... python: /root/.triton/llvm/llvm-b5cc222d-ubuntu-x64/include/llvm/ADT/SmallVector.h:291: T& llvm::SmallVectorTemplateCommon<T, >::operator[](llvm::SmallVectorTemplateCommon<T, >::size_type) [with T = unsigned int; = void; llvm::SmallVectorTemplateCommon<T, >::reference = unsigned int&; llvm::SmallVectorTemplateCommon<T, >::size_type = long unsigned int]: Assertion `idx < size()' failed.

Reproducer:
https://gist.github.com/jataylo/dfd4179d26e8fd5019a81189949e6f3b

Versions

#139206

cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @hongxiayang @naromero77amd @bertmaher @int3 @davidberard98 @nmacchioni @chenyang78 @embg @peterbell10 @aakhundov