onesnep

Fused Attention Kernel with gfx1030?

2

I was glad to see [Flash Attention ported to ROCM](https://github.com/ROCmSoftwarePlatform/flash-attention), however currently compatibility is limited to gfx90a. I and many others would love to see this on other architectures. When...

Fix: Resolve torch_dtype correctly for AMP

10

# Description The previous logic in `resolve_dtype` incorrectly configured the model's `torch_dtype` for Automatic Mixed Precision (AMP) training. When `fp16: true` or `bf16: true` was set for mixed precision, the...