tianhaos comments

Repositories
Issues
Comments

Results 2 comments of


                                            tianhaos

[BUG]when use 'overlap_comm:True' w/ 'contiguous_gradients:True', grad_norm is NaN

Hi @whlook, I'm facing a similar issue with AMD GPUs. But it only appears when I use flash-attention3. Can you share your env and fa3 env setting? Maybe the memory...

[Bug] Running Quick Start Example in Windows gives Error: `'MLCEngine' object has no attribute '_ffi'`

The error might cause by lack of gcc and relative packages. After I tried "conda install gcc", it worked on my device(vulkan).