You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
RuntimeError: /opt/nvidia/TransformerEngine/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu:771 in function operator(): cuDNN Error: [cudnn_frontend] Error: No valid execution plans built.. For more information, enable cuDNN error logging by setting CUDNN_LOGERR_DBG=1 and CUDNN_LOGDEST_DBG=stderr in the environment.
With CUDNN_LOGERR_DBG=1 and CUDNN_LOGDEST_DBG=stderr:
Hi @wujingyue , it seems that you have solved the issue by yourself? :) The input x follows a bshd format, so you should set attn_input_format=bshd accordingly. Let me know if I misunderstood the intention of the bug - thanks!
Repro
FWIW, this is a simplified version of
https://github.com/NVIDIA/Fuser/blob/c154e90919c40bfe2202b432c2a38e106d1a5444/tests/python/test_transformer_engine.py#L50.
Errors
With CUDNN_LOGERR_DBG=1 and CUDNN_LOGDEST_DBG=stderr:
Environments
Notes
Setting attn_input_format to "bshd" works around the problem.
The text was updated successfully, but these errors were encountered: