[ROCm] Enable FP8 KV-cache and relax constraints for RDNA4 custom paged attention#34741

Open

laudney wants to merge 3 commits intovllm-project:mainfrom

mmonad:feat/rocm-rdna4-custom-attn-fp8kv

Commits on Mar 27, 2026

[ROCm] Enable FP8 KV-cache and relax constraints for RDNA4 custom paged attention
L.B.R.
committed
Fix gfx12 FP8 QK WMMA head-element misalignment in paged attention decode
L.B.R.
committed
Fix gfx11 behavior regression: scope gqa_ratio and block_size relaxations to gfx12 only
L.B.R.
committed