Re-applying G42 bias triton fix on 0.4.3#41

Merged

gshtras merged 3 commits intomainfrom G42_bias_triton_fix

Jun 6, 2024

+251-353

Commits on Jun 5, 2024

Using rocm_flash_attention that supports bias computed from alibi slopes; Using attn_fwd triton kernel from ROCm/triton main_perf that does not cause triton compolier to hang
gshtras
committed

Commits on Jun 6, 2024