Jay Gu

Unfused Multihead attention TensorRT 9.2 is 2x slower than PyTorch 2.2 on GPU A100-SXM4-40GB

3

## Description When I'm comparing Multihead Attention between Torch2.2 and TensorRT 9.2 on A100-SXM4-40G, I found that for certain size the result engine does not use `_gemm_mha_v2` tactics. When not...

Engine build failure "cuda misaligned address" of TensorRT 10.1 when running fp16 group normalization with particular value of `num_groups`

4

## Description I believe this is regression of 10.1 on the normalization layer. It happens when FP16 mode is on and particular value of `num_groups`. When building a network with...