Anerudhan Gopal
Anerudhan Gopal
- What card is this? `nvidia-smi` output will be great. - What is the cudnn version you are using? `python -c 'import cudnn; print(cudnn.backend_version_string())'`
You can try improvising on [this](https://github.com/NVIDIA/cudnn-frontend/tree/main/benchmark) Install FAV2 pip inside the container and go from there. Try out the latest container (24.07, just in case).
Hi @vedantroy, can you try if this persists in 1.5.2 release? We have identified some numerical issues in `1.6.0` that we will be addressing in `1.6.1` Thanks Anerudhan
Hi @Corendos / @steeve , From the sample, looks like there is a mismatch in us documenting the ragged offset and Q tensor. Looking at the multiple tensors, ``` "Q"...
Hi @yhyang201 Thanks for the question. In the case of Ragged offset where the sequences are padded together, the graph API dimensions of query, key, value, output are indeed [B,S,H,D],...
Will be addressed in 1.5.0
This has been addressed. Closing
Hi @gritukan Thanks for reporting this. We are investigating the root cause of this issue in the cuDNN. We will update this issue once we have a timeline for the...
Hi @gritukan , This issue is resolved in 9.6.0 release earlier this month. Can you please verify. Thanks, Anerudhan
Will be addressed in next release.